1ed29b0b4SJens Axboe // SPDX-License-Identifier: GPL-2.0 2ed29b0b4SJens Axboe /* 3ed29b0b4SJens Axboe * Shared application/kernel submission and completion ring pairs, for 4ed29b0b4SJens Axboe * supporting fast/efficient IO. 5ed29b0b4SJens Axboe * 6ed29b0b4SJens Axboe * A note on the read/write ordering memory barriers that are matched between 7ed29b0b4SJens Axboe * the application and kernel side. 8ed29b0b4SJens Axboe * 9ed29b0b4SJens Axboe * After the application reads the CQ ring tail, it must use an 10ed29b0b4SJens Axboe * appropriate smp_rmb() to pair with the smp_wmb() the kernel uses 11ed29b0b4SJens Axboe * before writing the tail (using smp_load_acquire to read the tail will 12ed29b0b4SJens Axboe * do). It also needs a smp_mb() before updating CQ head (ordering the 13ed29b0b4SJens Axboe * entry load(s) with the head store), pairing with an implicit barrier 14ed29b0b4SJens Axboe * through a control-dependency in io_get_cqe (smp_store_release to 15ed29b0b4SJens Axboe * store head will do). Failure to do so could lead to reading invalid 16ed29b0b4SJens Axboe * CQ entries. 17ed29b0b4SJens Axboe * 18ed29b0b4SJens Axboe * Likewise, the application must use an appropriate smp_wmb() before 19ed29b0b4SJens Axboe * writing the SQ tail (ordering SQ entry stores with the tail store), 20ed29b0b4SJens Axboe * which pairs with smp_load_acquire in io_get_sqring (smp_store_release 21ed29b0b4SJens Axboe * to store the tail will do). And it needs a barrier ordering the SQ 22ed29b0b4SJens Axboe * head load before writing new SQ entries (smp_load_acquire to read 23ed29b0b4SJens Axboe * head will do). 24ed29b0b4SJens Axboe * 25ed29b0b4SJens Axboe * When using the SQ poll thread (IORING_SETUP_SQPOLL), the application 26ed29b0b4SJens Axboe * needs to check the SQ flags for IORING_SQ_NEED_WAKEUP *after* 27ed29b0b4SJens Axboe * updating the SQ tail; a full memory barrier smp_mb() is needed 28ed29b0b4SJens Axboe * between. 29ed29b0b4SJens Axboe * 30ed29b0b4SJens Axboe * Also see the examples in the liburing library: 31ed29b0b4SJens Axboe * 32ed29b0b4SJens Axboe * git://git.kernel.dk/liburing 33ed29b0b4SJens Axboe * 34ed29b0b4SJens Axboe * io_uring also uses READ/WRITE_ONCE() for _any_ store or load that happens 35ed29b0b4SJens Axboe * from data shared between the kernel and application. This is done both 36ed29b0b4SJens Axboe * for ordering purposes, but also to ensure that once a value is loaded from 37ed29b0b4SJens Axboe * data that the application could potentially modify, it remains stable. 38ed29b0b4SJens Axboe * 39ed29b0b4SJens Axboe * Copyright (C) 2018-2019 Jens Axboe 40ed29b0b4SJens Axboe * Copyright (c) 2018-2019 Christoph Hellwig 41ed29b0b4SJens Axboe */ 42ed29b0b4SJens Axboe #include <linux/kernel.h> 43ed29b0b4SJens Axboe #include <linux/init.h> 44ed29b0b4SJens Axboe #include <linux/errno.h> 45ed29b0b4SJens Axboe #include <linux/syscalls.h> 46ed29b0b4SJens Axboe #include <net/compat.h> 47ed29b0b4SJens Axboe #include <linux/refcount.h> 48ed29b0b4SJens Axboe #include <linux/uio.h> 49ed29b0b4SJens Axboe #include <linux/bits.h> 50ed29b0b4SJens Axboe 51ed29b0b4SJens Axboe #include <linux/sched/signal.h> 52ed29b0b4SJens Axboe #include <linux/fs.h> 53ed29b0b4SJens Axboe #include <linux/file.h> 54ed29b0b4SJens Axboe #include <linux/fdtable.h> 55ed29b0b4SJens Axboe #include <linux/mm.h> 56ed29b0b4SJens Axboe #include <linux/mman.h> 57ed29b0b4SJens Axboe #include <linux/percpu.h> 58ed29b0b4SJens Axboe #include <linux/slab.h> 59ed29b0b4SJens Axboe #include <linux/bvec.h> 60ed29b0b4SJens Axboe #include <linux/net.h> 61ed29b0b4SJens Axboe #include <net/sock.h> 62ed29b0b4SJens Axboe #include <net/af_unix.h> 63ed29b0b4SJens Axboe #include <net/scm.h> 64ed29b0b4SJens Axboe #include <linux/anon_inodes.h> 65ed29b0b4SJens Axboe #include <linux/sched/mm.h> 66ed29b0b4SJens Axboe #include <linux/uaccess.h> 67ed29b0b4SJens Axboe #include <linux/nospec.h> 68ed29b0b4SJens Axboe #include <linux/highmem.h> 69ed29b0b4SJens Axboe #include <linux/fsnotify.h> 70ed29b0b4SJens Axboe #include <linux/fadvise.h> 71ed29b0b4SJens Axboe #include <linux/task_work.h> 72ed29b0b4SJens Axboe #include <linux/io_uring.h> 73ed29b0b4SJens Axboe #include <linux/audit.h> 74ed29b0b4SJens Axboe #include <linux/security.h> 75ed29b0b4SJens Axboe 76ed29b0b4SJens Axboe #define CREATE_TRACE_POINTS 77ed29b0b4SJens Axboe #include <trace/events/io_uring.h> 78ed29b0b4SJens Axboe 79ed29b0b4SJens Axboe #include <uapi/linux/io_uring.h> 80ed29b0b4SJens Axboe 81ed29b0b4SJens Axboe #include "io-wq.h" 82ed29b0b4SJens Axboe 83de23077eSJens Axboe #include "io_uring.h" 84329061d3SJens Axboe #include "opdef.h" 85e418bbc9SJens Axboe #include "refs.h" 86c9f06aa7SJens Axboe #include "tctx.h" 8717437f31SJens Axboe #include "sqpoll.h" 88a4ad4f74SJens Axboe #include "fdinfo.h" 893b77495aSJens Axboe #include "kbuf.h" 9073572984SJens Axboe #include "rsrc.h" 9138513c46SHao Xu #include "cancel.h" 9243e0bbbdSJens Axboe #include "net.h" 93eb42cebbSPavel Begunkov #include "notif.h" 94e27f928eSJens Axboe 9559915143SJens Axboe #include "timeout.h" 96329061d3SJens Axboe #include "poll.h" 979b797a37SJens Axboe #include "alloc_cache.h" 985e2a18d9SJens Axboe 99ed29b0b4SJens Axboe #define IORING_MAX_ENTRIES 32768 100ed29b0b4SJens Axboe #define IORING_MAX_CQ_ENTRIES (2 * IORING_MAX_ENTRIES) 101ed29b0b4SJens Axboe 102ed29b0b4SJens Axboe #define IORING_MAX_RESTRICTIONS (IORING_RESTRICTION_LAST + \ 103ed29b0b4SJens Axboe IORING_REGISTER_LAST + IORING_OP_LAST) 104ed29b0b4SJens Axboe 105ed29b0b4SJens Axboe #define SQE_COMMON_FLAGS (IOSQE_FIXED_FILE | IOSQE_IO_LINK | \ 106ed29b0b4SJens Axboe IOSQE_IO_HARDLINK | IOSQE_ASYNC) 107ed29b0b4SJens Axboe 108ed29b0b4SJens Axboe #define SQE_VALID_FLAGS (SQE_COMMON_FLAGS | IOSQE_BUFFER_SELECT | \ 109ed29b0b4SJens Axboe IOSQE_IO_DRAIN | IOSQE_CQE_SKIP_SUCCESS) 110ed29b0b4SJens Axboe 111ed29b0b4SJens Axboe #define IO_REQ_CLEAN_FLAGS (REQ_F_BUFFER_SELECTED | REQ_F_NEED_CLEANUP | \ 112ed29b0b4SJens Axboe REQ_F_POLLED | REQ_F_INFLIGHT | REQ_F_CREDS | \ 113ed29b0b4SJens Axboe REQ_F_ASYNC_DATA) 114ed29b0b4SJens Axboe 115ed29b0b4SJens Axboe #define IO_REQ_CLEAN_SLOW_FLAGS (REQ_F_REFCOUNT | REQ_F_LINK | REQ_F_HARDLINK |\ 116ed29b0b4SJens Axboe IO_REQ_CLEAN_FLAGS) 117ed29b0b4SJens Axboe 118ed29b0b4SJens Axboe #define IO_TCTX_REFS_CACHE_NR (1U << 10) 119ed29b0b4SJens Axboe 120ed29b0b4SJens Axboe #define IO_COMPL_BATCH 32 121ed29b0b4SJens Axboe #define IO_REQ_ALLOC_BATCH 8 122ed29b0b4SJens Axboe 123ed29b0b4SJens Axboe enum { 124ed29b0b4SJens Axboe IO_CHECK_CQ_OVERFLOW_BIT, 125ed29b0b4SJens Axboe IO_CHECK_CQ_DROPPED_BIT, 126ed29b0b4SJens Axboe }; 127ed29b0b4SJens Axboe 12821a091b9SDylan Yudaken enum { 12921a091b9SDylan Yudaken IO_EVENTFD_OP_SIGNAL_BIT, 13021a091b9SDylan Yudaken IO_EVENTFD_OP_FREE_BIT, 13121a091b9SDylan Yudaken }; 13221a091b9SDylan Yudaken 133ed29b0b4SJens Axboe struct io_defer_entry { 134ed29b0b4SJens Axboe struct list_head list; 135ed29b0b4SJens Axboe struct io_kiocb *req; 136ed29b0b4SJens Axboe u32 seq; 137ed29b0b4SJens Axboe }; 138ed29b0b4SJens Axboe 139ed29b0b4SJens Axboe /* requests with any of those set should undergo io_disarm_next() */ 140ed29b0b4SJens Axboe #define IO_DISARM_MASK (REQ_F_ARM_LTIMEOUT | REQ_F_LINK_TIMEOUT | REQ_F_FAIL) 141ed29b0b4SJens Axboe #define IO_REQ_LINK_FLAGS (REQ_F_LINK | REQ_F_HARDLINK) 142ed29b0b4SJens Axboe 143affa87dbSPavel Begunkov static bool io_uring_try_cancel_requests(struct io_ring_ctx *ctx, 144ed29b0b4SJens Axboe struct task_struct *task, 145ed29b0b4SJens Axboe bool cancel_all); 146ed29b0b4SJens Axboe 147ed29b0b4SJens Axboe static void io_dismantle_req(struct io_kiocb *req); 148ed29b0b4SJens Axboe static void io_clean_op(struct io_kiocb *req); 149ed29b0b4SJens Axboe static void io_queue_sqe(struct io_kiocb *req); 150c0e0d6baSDylan Yudaken static void io_move_task_work_from_local(struct io_ring_ctx *ctx); 151ed29b0b4SJens Axboe static void __io_submit_flush_completions(struct io_ring_ctx *ctx); 152ed29b0b4SJens Axboe 153ed29b0b4SJens Axboe static struct kmem_cache *req_cachep; 154ed29b0b4SJens Axboe 155ed29b0b4SJens Axboe struct sock *io_uring_get_socket(struct file *file) 156ed29b0b4SJens Axboe { 157ed29b0b4SJens Axboe #if defined(CONFIG_UNIX) 158cd40cae2SJens Axboe if (io_is_uring_fops(file)) { 159ed29b0b4SJens Axboe struct io_ring_ctx *ctx = file->private_data; 160ed29b0b4SJens Axboe 161ed29b0b4SJens Axboe return ctx->ring_sock->sk; 162ed29b0b4SJens Axboe } 163ed29b0b4SJens Axboe #endif 164ed29b0b4SJens Axboe return NULL; 165ed29b0b4SJens Axboe } 166ed29b0b4SJens Axboe EXPORT_SYMBOL(io_uring_get_socket); 167ed29b0b4SJens Axboe 168ed29b0b4SJens Axboe static inline void io_submit_flush_completions(struct io_ring_ctx *ctx) 169ed29b0b4SJens Axboe { 170931147ddSDylan Yudaken if (!wq_list_empty(&ctx->submit_state.compl_reqs) || 171931147ddSDylan Yudaken ctx->submit_state.cqes_count) 172ed29b0b4SJens Axboe __io_submit_flush_completions(ctx); 173ed29b0b4SJens Axboe } 174ed29b0b4SJens Axboe 175faf88ddeSPavel Begunkov static inline unsigned int __io_cqring_events(struct io_ring_ctx *ctx) 176faf88ddeSPavel Begunkov { 177faf88ddeSPavel Begunkov return ctx->cached_cq_tail - READ_ONCE(ctx->rings->cq.head); 178faf88ddeSPavel Begunkov } 179faf88ddeSPavel Begunkov 1800fc8c2acSDylan Yudaken static inline unsigned int __io_cqring_events_user(struct io_ring_ctx *ctx) 1810fc8c2acSDylan Yudaken { 1820fc8c2acSDylan Yudaken return READ_ONCE(ctx->rings->cq.tail) - READ_ONCE(ctx->rings->cq.head); 1830fc8c2acSDylan Yudaken } 1840fc8c2acSDylan Yudaken 185ed29b0b4SJens Axboe static bool io_match_linked(struct io_kiocb *head) 186ed29b0b4SJens Axboe { 187ed29b0b4SJens Axboe struct io_kiocb *req; 188ed29b0b4SJens Axboe 189ed29b0b4SJens Axboe io_for_each_link(req, head) { 190ed29b0b4SJens Axboe if (req->flags & REQ_F_INFLIGHT) 191ed29b0b4SJens Axboe return true; 192ed29b0b4SJens Axboe } 193ed29b0b4SJens Axboe return false; 194ed29b0b4SJens Axboe } 195ed29b0b4SJens Axboe 196ed29b0b4SJens Axboe /* 197ed29b0b4SJens Axboe * As io_match_task() but protected against racing with linked timeouts. 198ed29b0b4SJens Axboe * User must not hold timeout_lock. 199ed29b0b4SJens Axboe */ 200329061d3SJens Axboe bool io_match_task_safe(struct io_kiocb *head, struct task_struct *task, 201ed29b0b4SJens Axboe bool cancel_all) 202ed29b0b4SJens Axboe { 203ed29b0b4SJens Axboe bool matched; 204ed29b0b4SJens Axboe 205ed29b0b4SJens Axboe if (task && head->task != task) 206ed29b0b4SJens Axboe return false; 207ed29b0b4SJens Axboe if (cancel_all) 208ed29b0b4SJens Axboe return true; 209ed29b0b4SJens Axboe 210ed29b0b4SJens Axboe if (head->flags & REQ_F_LINK_TIMEOUT) { 211ed29b0b4SJens Axboe struct io_ring_ctx *ctx = head->ctx; 212ed29b0b4SJens Axboe 213ed29b0b4SJens Axboe /* protect against races with linked timeouts */ 214ed29b0b4SJens Axboe spin_lock_irq(&ctx->timeout_lock); 215ed29b0b4SJens Axboe matched = io_match_linked(head); 216ed29b0b4SJens Axboe spin_unlock_irq(&ctx->timeout_lock); 217ed29b0b4SJens Axboe } else { 218ed29b0b4SJens Axboe matched = io_match_linked(head); 219ed29b0b4SJens Axboe } 220ed29b0b4SJens Axboe return matched; 221ed29b0b4SJens Axboe } 222ed29b0b4SJens Axboe 223ed29b0b4SJens Axboe static inline void req_fail_link_node(struct io_kiocb *req, int res) 224ed29b0b4SJens Axboe { 225ed29b0b4SJens Axboe req_set_fail(req); 22697b388d7SJens Axboe io_req_set_res(req, res, 0); 227ed29b0b4SJens Axboe } 228ed29b0b4SJens Axboe 229ed29b0b4SJens Axboe static inline void io_req_add_to_cache(struct io_kiocb *req, struct io_ring_ctx *ctx) 230ed29b0b4SJens Axboe { 231ed29b0b4SJens Axboe wq_stack_add_head(&req->comp_list, &ctx->submit_state.free_list); 232ed29b0b4SJens Axboe } 233ed29b0b4SJens Axboe 234ed29b0b4SJens Axboe static __cold void io_ring_ctx_ref_free(struct percpu_ref *ref) 235ed29b0b4SJens Axboe { 236ed29b0b4SJens Axboe struct io_ring_ctx *ctx = container_of(ref, struct io_ring_ctx, refs); 237ed29b0b4SJens Axboe 238ed29b0b4SJens Axboe complete(&ctx->ref_comp); 239ed29b0b4SJens Axboe } 240ed29b0b4SJens Axboe 241ed29b0b4SJens Axboe static __cold void io_fallback_req_func(struct work_struct *work) 242ed29b0b4SJens Axboe { 243ed29b0b4SJens Axboe struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx, 244ed29b0b4SJens Axboe fallback_work.work); 245ed29b0b4SJens Axboe struct llist_node *node = llist_del_all(&ctx->fallback_llist); 246ed29b0b4SJens Axboe struct io_kiocb *req, *tmp; 247ed29b0b4SJens Axboe bool locked = false; 248ed29b0b4SJens Axboe 249ed29b0b4SJens Axboe percpu_ref_get(&ctx->refs); 2503218e5d3SPavel Begunkov llist_for_each_entry_safe(req, tmp, node, io_task_work.node) 251ed29b0b4SJens Axboe req->io_task_work.func(req, &locked); 252ed29b0b4SJens Axboe 253ed29b0b4SJens Axboe if (locked) { 254ed29b0b4SJens Axboe io_submit_flush_completions(ctx); 255ed29b0b4SJens Axboe mutex_unlock(&ctx->uring_lock); 256ed29b0b4SJens Axboe } 257ed29b0b4SJens Axboe percpu_ref_put(&ctx->refs); 258ed29b0b4SJens Axboe } 259ed29b0b4SJens Axboe 260e6f89be6SPavel Begunkov static int io_alloc_hash_table(struct io_hash_table *table, unsigned bits) 261e6f89be6SPavel Begunkov { 262e6f89be6SPavel Begunkov unsigned hash_buckets = 1U << bits; 263e6f89be6SPavel Begunkov size_t hash_size = hash_buckets * sizeof(table->hbs[0]); 264e6f89be6SPavel Begunkov 265e6f89be6SPavel Begunkov table->hbs = kmalloc(hash_size, GFP_KERNEL); 266e6f89be6SPavel Begunkov if (!table->hbs) 267e6f89be6SPavel Begunkov return -ENOMEM; 268e6f89be6SPavel Begunkov 269e6f89be6SPavel Begunkov table->hash_bits = bits; 270e6f89be6SPavel Begunkov init_hash_table(table, hash_buckets); 271e6f89be6SPavel Begunkov return 0; 272e6f89be6SPavel Begunkov } 273e6f89be6SPavel Begunkov 274ed29b0b4SJens Axboe static __cold struct io_ring_ctx *io_ring_ctx_alloc(struct io_uring_params *p) 275ed29b0b4SJens Axboe { 276ed29b0b4SJens Axboe struct io_ring_ctx *ctx; 277ed29b0b4SJens Axboe int hash_bits; 278ed29b0b4SJens Axboe 279ed29b0b4SJens Axboe ctx = kzalloc(sizeof(*ctx), GFP_KERNEL); 280ed29b0b4SJens Axboe if (!ctx) 281ed29b0b4SJens Axboe return NULL; 282ed29b0b4SJens Axboe 283ed29b0b4SJens Axboe xa_init(&ctx->io_bl_xa); 284ed29b0b4SJens Axboe 285ed29b0b4SJens Axboe /* 286ed29b0b4SJens Axboe * Use 5 bits less than the max cq entries, that should give us around 2874a07723fSPavel Begunkov * 32 entries per hash list if totally full and uniformly spread, but 2884a07723fSPavel Begunkov * don't keep too many buckets to not overconsume memory. 289ed29b0b4SJens Axboe */ 2904a07723fSPavel Begunkov hash_bits = ilog2(p->cq_entries) - 5; 2914a07723fSPavel Begunkov hash_bits = clamp(hash_bits, 1, 8); 292e6f89be6SPavel Begunkov if (io_alloc_hash_table(&ctx->cancel_table, hash_bits)) 293ed29b0b4SJens Axboe goto err; 2949ca9fb24SPavel Begunkov if (io_alloc_hash_table(&ctx->cancel_table_locked, hash_bits)) 2959ca9fb24SPavel Begunkov goto err; 29638513c46SHao Xu 297ed29b0b4SJens Axboe ctx->dummy_ubuf = kzalloc(sizeof(*ctx->dummy_ubuf), GFP_KERNEL); 298ed29b0b4SJens Axboe if (!ctx->dummy_ubuf) 299ed29b0b4SJens Axboe goto err; 300ed29b0b4SJens Axboe /* set invalid range, so io_import_fixed() fails meeting it */ 301ed29b0b4SJens Axboe ctx->dummy_ubuf->ubuf = -1UL; 302ed29b0b4SJens Axboe 303ed29b0b4SJens Axboe if (percpu_ref_init(&ctx->refs, io_ring_ctx_ref_free, 30448904229SMichal Koutný 0, GFP_KERNEL)) 305ed29b0b4SJens Axboe goto err; 306ed29b0b4SJens Axboe 307ed29b0b4SJens Axboe ctx->flags = p->flags; 308ed29b0b4SJens Axboe init_waitqueue_head(&ctx->sqo_sq_wait); 309ed29b0b4SJens Axboe INIT_LIST_HEAD(&ctx->sqd_list); 310ed29b0b4SJens Axboe INIT_LIST_HEAD(&ctx->cq_overflow_list); 311ed29b0b4SJens Axboe INIT_LIST_HEAD(&ctx->io_buffers_cache); 3129b797a37SJens Axboe io_alloc_cache_init(&ctx->apoll_cache); 31343e0bbbdSJens Axboe io_alloc_cache_init(&ctx->netmsg_cache); 314ed29b0b4SJens Axboe init_completion(&ctx->ref_comp); 315ed29b0b4SJens Axboe xa_init_flags(&ctx->personalities, XA_FLAGS_ALLOC1); 316ed29b0b4SJens Axboe mutex_init(&ctx->uring_lock); 317ed29b0b4SJens Axboe init_waitqueue_head(&ctx->cq_wait); 318ed29b0b4SJens Axboe spin_lock_init(&ctx->completion_lock); 319ed29b0b4SJens Axboe spin_lock_init(&ctx->timeout_lock); 320ed29b0b4SJens Axboe INIT_WQ_LIST(&ctx->iopoll_list); 321ed29b0b4SJens Axboe INIT_LIST_HEAD(&ctx->io_buffers_pages); 322ed29b0b4SJens Axboe INIT_LIST_HEAD(&ctx->io_buffers_comp); 323ed29b0b4SJens Axboe INIT_LIST_HEAD(&ctx->defer_list); 324ed29b0b4SJens Axboe INIT_LIST_HEAD(&ctx->timeout_list); 325ed29b0b4SJens Axboe INIT_LIST_HEAD(&ctx->ltimeout_list); 326ed29b0b4SJens Axboe spin_lock_init(&ctx->rsrc_ref_lock); 327ed29b0b4SJens Axboe INIT_LIST_HEAD(&ctx->rsrc_ref_list); 328ed29b0b4SJens Axboe INIT_DELAYED_WORK(&ctx->rsrc_put_work, io_rsrc_put_work); 329ed29b0b4SJens Axboe init_llist_head(&ctx->rsrc_put_llist); 330c0e0d6baSDylan Yudaken init_llist_head(&ctx->work_llist); 331ed29b0b4SJens Axboe INIT_LIST_HEAD(&ctx->tctx_list); 332ed29b0b4SJens Axboe ctx->submit_state.free_list.next = NULL; 333ed29b0b4SJens Axboe INIT_WQ_LIST(&ctx->locked_free_list); 334ed29b0b4SJens Axboe INIT_DELAYED_WORK(&ctx->fallback_work, io_fallback_req_func); 335ed29b0b4SJens Axboe INIT_WQ_LIST(&ctx->submit_state.compl_reqs); 336ed29b0b4SJens Axboe return ctx; 337ed29b0b4SJens Axboe err: 338ed29b0b4SJens Axboe kfree(ctx->dummy_ubuf); 339e6f89be6SPavel Begunkov kfree(ctx->cancel_table.hbs); 3409ca9fb24SPavel Begunkov kfree(ctx->cancel_table_locked.hbs); 341ed29b0b4SJens Axboe kfree(ctx->io_bl); 342ed29b0b4SJens Axboe xa_destroy(&ctx->io_bl_xa); 343ed29b0b4SJens Axboe kfree(ctx); 344ed29b0b4SJens Axboe return NULL; 345ed29b0b4SJens Axboe } 346ed29b0b4SJens Axboe 347ed29b0b4SJens Axboe static void io_account_cq_overflow(struct io_ring_ctx *ctx) 348ed29b0b4SJens Axboe { 349ed29b0b4SJens Axboe struct io_rings *r = ctx->rings; 350ed29b0b4SJens Axboe 351ed29b0b4SJens Axboe WRITE_ONCE(r->cq_overflow, READ_ONCE(r->cq_overflow) + 1); 352ed29b0b4SJens Axboe ctx->cq_extra--; 353ed29b0b4SJens Axboe } 354ed29b0b4SJens Axboe 355ed29b0b4SJens Axboe static bool req_need_defer(struct io_kiocb *req, u32 seq) 356ed29b0b4SJens Axboe { 357ed29b0b4SJens Axboe if (unlikely(req->flags & REQ_F_IO_DRAIN)) { 358ed29b0b4SJens Axboe struct io_ring_ctx *ctx = req->ctx; 359ed29b0b4SJens Axboe 360ed29b0b4SJens Axboe return seq + READ_ONCE(ctx->cq_extra) != ctx->cached_cq_tail; 361ed29b0b4SJens Axboe } 362ed29b0b4SJens Axboe 363ed29b0b4SJens Axboe return false; 364ed29b0b4SJens Axboe } 365ed29b0b4SJens Axboe 366ed29b0b4SJens Axboe static inline void io_req_track_inflight(struct io_kiocb *req) 367ed29b0b4SJens Axboe { 368ed29b0b4SJens Axboe if (!(req->flags & REQ_F_INFLIGHT)) { 369ed29b0b4SJens Axboe req->flags |= REQ_F_INFLIGHT; 370ed29b0b4SJens Axboe atomic_inc(&req->task->io_uring->inflight_tracked); 371ed29b0b4SJens Axboe } 372ed29b0b4SJens Axboe } 373ed29b0b4SJens Axboe 374ed29b0b4SJens Axboe static struct io_kiocb *__io_prep_linked_timeout(struct io_kiocb *req) 375ed29b0b4SJens Axboe { 376ed29b0b4SJens Axboe if (WARN_ON_ONCE(!req->link)) 377ed29b0b4SJens Axboe return NULL; 378ed29b0b4SJens Axboe 379ed29b0b4SJens Axboe req->flags &= ~REQ_F_ARM_LTIMEOUT; 380ed29b0b4SJens Axboe req->flags |= REQ_F_LINK_TIMEOUT; 381ed29b0b4SJens Axboe 382ed29b0b4SJens Axboe /* linked timeouts should have two refs once prep'ed */ 383ed29b0b4SJens Axboe io_req_set_refcount(req); 384ed29b0b4SJens Axboe __io_req_set_refcount(req->link, 2); 385ed29b0b4SJens Axboe return req->link; 386ed29b0b4SJens Axboe } 387ed29b0b4SJens Axboe 388ed29b0b4SJens Axboe static inline struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req) 389ed29b0b4SJens Axboe { 390ed29b0b4SJens Axboe if (likely(!(req->flags & REQ_F_ARM_LTIMEOUT))) 391ed29b0b4SJens Axboe return NULL; 392ed29b0b4SJens Axboe return __io_prep_linked_timeout(req); 393ed29b0b4SJens Axboe } 394ed29b0b4SJens Axboe 395ed29b0b4SJens Axboe static noinline void __io_arm_ltimeout(struct io_kiocb *req) 396ed29b0b4SJens Axboe { 397ed29b0b4SJens Axboe io_queue_linked_timeout(__io_prep_linked_timeout(req)); 398ed29b0b4SJens Axboe } 399ed29b0b4SJens Axboe 400ed29b0b4SJens Axboe static inline void io_arm_ltimeout(struct io_kiocb *req) 401ed29b0b4SJens Axboe { 402ed29b0b4SJens Axboe if (unlikely(req->flags & REQ_F_ARM_LTIMEOUT)) 403ed29b0b4SJens Axboe __io_arm_ltimeout(req); 404ed29b0b4SJens Axboe } 405ed29b0b4SJens Axboe 406ed29b0b4SJens Axboe static void io_prep_async_work(struct io_kiocb *req) 407ed29b0b4SJens Axboe { 408ed29b0b4SJens Axboe const struct io_op_def *def = &io_op_defs[req->opcode]; 409ed29b0b4SJens Axboe struct io_ring_ctx *ctx = req->ctx; 410ed29b0b4SJens Axboe 411ed29b0b4SJens Axboe if (!(req->flags & REQ_F_CREDS)) { 412ed29b0b4SJens Axboe req->flags |= REQ_F_CREDS; 413ed29b0b4SJens Axboe req->creds = get_current_cred(); 414ed29b0b4SJens Axboe } 415ed29b0b4SJens Axboe 416ed29b0b4SJens Axboe req->work.list.next = NULL; 417ed29b0b4SJens Axboe req->work.flags = 0; 418ed29b0b4SJens Axboe req->work.cancel_seq = atomic_read(&ctx->cancel_seq); 419ed29b0b4SJens Axboe if (req->flags & REQ_F_FORCE_ASYNC) 420ed29b0b4SJens Axboe req->work.flags |= IO_WQ_WORK_CONCURRENT; 421ed29b0b4SJens Axboe 422f6b543fdSJens Axboe if (req->file && !io_req_ffs_set(req)) 423f6b543fdSJens Axboe req->flags |= io_file_get_flags(req->file) << REQ_F_SUPPORT_NOWAIT_BIT; 424f6b543fdSJens Axboe 425ed29b0b4SJens Axboe if (req->flags & REQ_F_ISREG) { 426ed29b0b4SJens Axboe if (def->hash_reg_file || (ctx->flags & IORING_SETUP_IOPOLL)) 427ed29b0b4SJens Axboe io_wq_hash_work(&req->work, file_inode(req->file)); 428ed29b0b4SJens Axboe } else if (!req->file || !S_ISBLK(file_inode(req->file)->i_mode)) { 429ed29b0b4SJens Axboe if (def->unbound_nonreg_file) 430ed29b0b4SJens Axboe req->work.flags |= IO_WQ_WORK_UNBOUND; 431ed29b0b4SJens Axboe } 432ed29b0b4SJens Axboe } 433ed29b0b4SJens Axboe 434ed29b0b4SJens Axboe static void io_prep_async_link(struct io_kiocb *req) 435ed29b0b4SJens Axboe { 436ed29b0b4SJens Axboe struct io_kiocb *cur; 437ed29b0b4SJens Axboe 438ed29b0b4SJens Axboe if (req->flags & REQ_F_LINK_TIMEOUT) { 439ed29b0b4SJens Axboe struct io_ring_ctx *ctx = req->ctx; 440ed29b0b4SJens Axboe 441ed29b0b4SJens Axboe spin_lock_irq(&ctx->timeout_lock); 442ed29b0b4SJens Axboe io_for_each_link(cur, req) 443ed29b0b4SJens Axboe io_prep_async_work(cur); 444ed29b0b4SJens Axboe spin_unlock_irq(&ctx->timeout_lock); 445ed29b0b4SJens Axboe } else { 446ed29b0b4SJens Axboe io_for_each_link(cur, req) 447ed29b0b4SJens Axboe io_prep_async_work(cur); 448ed29b0b4SJens Axboe } 449ed29b0b4SJens Axboe } 450ed29b0b4SJens Axboe 451f3b44f92SJens Axboe void io_queue_iowq(struct io_kiocb *req, bool *dont_use) 452ed29b0b4SJens Axboe { 453ed29b0b4SJens Axboe struct io_kiocb *link = io_prep_linked_timeout(req); 454ed29b0b4SJens Axboe struct io_uring_task *tctx = req->task->io_uring; 455ed29b0b4SJens Axboe 456ed29b0b4SJens Axboe BUG_ON(!tctx); 457ed29b0b4SJens Axboe BUG_ON(!tctx->io_wq); 458ed29b0b4SJens Axboe 459ed29b0b4SJens Axboe /* init ->work of the whole link before punting */ 460ed29b0b4SJens Axboe io_prep_async_link(req); 461ed29b0b4SJens Axboe 462ed29b0b4SJens Axboe /* 463ed29b0b4SJens Axboe * Not expected to happen, but if we do have a bug where this _can_ 464ed29b0b4SJens Axboe * happen, catch it here and ensure the request is marked as 465ed29b0b4SJens Axboe * canceled. That will make io-wq go through the usual work cancel 466ed29b0b4SJens Axboe * procedure rather than attempt to run this request (or create a new 467ed29b0b4SJens Axboe * worker for it). 468ed29b0b4SJens Axboe */ 469ed29b0b4SJens Axboe if (WARN_ON_ONCE(!same_thread_group(req->task, current))) 470ed29b0b4SJens Axboe req->work.flags |= IO_WQ_WORK_CANCEL; 471ed29b0b4SJens Axboe 47248863ffdSPavel Begunkov trace_io_uring_queue_async_work(req, io_wq_is_hashed(&req->work)); 473ed29b0b4SJens Axboe io_wq_enqueue(tctx->io_wq, &req->work); 474ed29b0b4SJens Axboe if (link) 475ed29b0b4SJens Axboe io_queue_linked_timeout(link); 476ed29b0b4SJens Axboe } 477ed29b0b4SJens Axboe 478ed29b0b4SJens Axboe static __cold void io_queue_deferred(struct io_ring_ctx *ctx) 479ed29b0b4SJens Axboe { 480ed29b0b4SJens Axboe while (!list_empty(&ctx->defer_list)) { 481ed29b0b4SJens Axboe struct io_defer_entry *de = list_first_entry(&ctx->defer_list, 482ed29b0b4SJens Axboe struct io_defer_entry, list); 483ed29b0b4SJens Axboe 484ed29b0b4SJens Axboe if (req_need_defer(de->req, de->seq)) 485ed29b0b4SJens Axboe break; 486ed29b0b4SJens Axboe list_del_init(&de->list); 487ed29b0b4SJens Axboe io_req_task_queue(de->req); 488ed29b0b4SJens Axboe kfree(de); 489ed29b0b4SJens Axboe } 490ed29b0b4SJens Axboe } 491ed29b0b4SJens Axboe 49221a091b9SDylan Yudaken 49321a091b9SDylan Yudaken static void io_eventfd_ops(struct rcu_head *rcu) 494d8e9214fSDylan Yudaken { 495d8e9214fSDylan Yudaken struct io_ev_fd *ev_fd = container_of(rcu, struct io_ev_fd, rcu); 49621a091b9SDylan Yudaken int ops = atomic_xchg(&ev_fd->ops, 0); 497d8e9214fSDylan Yudaken 49821a091b9SDylan Yudaken if (ops & BIT(IO_EVENTFD_OP_SIGNAL_BIT)) 49944648532SJens Axboe eventfd_signal_mask(ev_fd->cq_ev_fd, 1, EPOLL_URING_WAKE); 50021a091b9SDylan Yudaken 50121a091b9SDylan Yudaken /* IO_EVENTFD_OP_FREE_BIT may not be set here depending on callback 50221a091b9SDylan Yudaken * ordering in a race but if references are 0 we know we have to free 50321a091b9SDylan Yudaken * it regardless. 50421a091b9SDylan Yudaken */ 50521a091b9SDylan Yudaken if (atomic_dec_and_test(&ev_fd->refs)) { 506d8e9214fSDylan Yudaken eventfd_ctx_put(ev_fd->cq_ev_fd); 507d8e9214fSDylan Yudaken kfree(ev_fd); 508d8e9214fSDylan Yudaken } 50921a091b9SDylan Yudaken } 510d8e9214fSDylan Yudaken 511ed29b0b4SJens Axboe static void io_eventfd_signal(struct io_ring_ctx *ctx) 512ed29b0b4SJens Axboe { 51321a091b9SDylan Yudaken struct io_ev_fd *ev_fd = NULL; 514ed29b0b4SJens Axboe 515ed29b0b4SJens Axboe rcu_read_lock(); 516ed29b0b4SJens Axboe /* 517ed29b0b4SJens Axboe * rcu_dereference ctx->io_ev_fd once and use it for both for checking 518ed29b0b4SJens Axboe * and eventfd_signal 519ed29b0b4SJens Axboe */ 520ed29b0b4SJens Axboe ev_fd = rcu_dereference(ctx->io_ev_fd); 521ed29b0b4SJens Axboe 522ed29b0b4SJens Axboe /* 523ed29b0b4SJens Axboe * Check again if ev_fd exists incase an io_eventfd_unregister call 524ed29b0b4SJens Axboe * completed between the NULL check of ctx->io_ev_fd at the start of 525ed29b0b4SJens Axboe * the function and rcu_read_lock. 526ed29b0b4SJens Axboe */ 527ed29b0b4SJens Axboe if (unlikely(!ev_fd)) 528ed29b0b4SJens Axboe goto out; 529ed29b0b4SJens Axboe if (READ_ONCE(ctx->rings->cq_flags) & IORING_CQ_EVENTFD_DISABLED) 530ed29b0b4SJens Axboe goto out; 53121a091b9SDylan Yudaken if (ev_fd->eventfd_async && !io_wq_current_is_worker()) 53221a091b9SDylan Yudaken goto out; 533ed29b0b4SJens Axboe 53421a091b9SDylan Yudaken if (likely(eventfd_signal_allowed())) { 53544648532SJens Axboe eventfd_signal_mask(ev_fd->cq_ev_fd, 1, EPOLL_URING_WAKE); 53621a091b9SDylan Yudaken } else { 53721a091b9SDylan Yudaken atomic_inc(&ev_fd->refs); 53821a091b9SDylan Yudaken if (!atomic_fetch_or(BIT(IO_EVENTFD_OP_SIGNAL_BIT), &ev_fd->ops)) 53921a091b9SDylan Yudaken call_rcu(&ev_fd->rcu, io_eventfd_ops); 54021a091b9SDylan Yudaken else 54121a091b9SDylan Yudaken atomic_dec(&ev_fd->refs); 54221a091b9SDylan Yudaken } 54321a091b9SDylan Yudaken 544ed29b0b4SJens Axboe out: 545ed29b0b4SJens Axboe rcu_read_unlock(); 546ed29b0b4SJens Axboe } 547ed29b0b4SJens Axboe 54821a091b9SDylan Yudaken static void io_eventfd_flush_signal(struct io_ring_ctx *ctx) 54921a091b9SDylan Yudaken { 55021a091b9SDylan Yudaken bool skip; 55121a091b9SDylan Yudaken 55221a091b9SDylan Yudaken spin_lock(&ctx->completion_lock); 55321a091b9SDylan Yudaken 55421a091b9SDylan Yudaken /* 55521a091b9SDylan Yudaken * Eventfd should only get triggered when at least one event has been 55621a091b9SDylan Yudaken * posted. Some applications rely on the eventfd notification count 55721a091b9SDylan Yudaken * only changing IFF a new CQE has been added to the CQ ring. There's 55821a091b9SDylan Yudaken * no depedency on 1:1 relationship between how many times this 55921a091b9SDylan Yudaken * function is called (and hence the eventfd count) and number of CQEs 56021a091b9SDylan Yudaken * posted to the CQ ring. 56121a091b9SDylan Yudaken */ 56221a091b9SDylan Yudaken skip = ctx->cached_cq_tail == ctx->evfd_last_cq_tail; 56321a091b9SDylan Yudaken ctx->evfd_last_cq_tail = ctx->cached_cq_tail; 56421a091b9SDylan Yudaken spin_unlock(&ctx->completion_lock); 56521a091b9SDylan Yudaken if (skip) 56621a091b9SDylan Yudaken return; 56721a091b9SDylan Yudaken 56821a091b9SDylan Yudaken io_eventfd_signal(ctx); 56921a091b9SDylan Yudaken } 57021a091b9SDylan Yudaken 571a830ffd2SPavel Begunkov void __io_commit_cqring_flush(struct io_ring_ctx *ctx) 572a830ffd2SPavel Begunkov { 573a830ffd2SPavel Begunkov if (ctx->off_timeout_used || ctx->drain_active) { 574a830ffd2SPavel Begunkov spin_lock(&ctx->completion_lock); 575a830ffd2SPavel Begunkov if (ctx->off_timeout_used) 576a830ffd2SPavel Begunkov io_flush_timeouts(ctx); 577a830ffd2SPavel Begunkov if (ctx->drain_active) 578a830ffd2SPavel Begunkov io_queue_deferred(ctx); 579a830ffd2SPavel Begunkov spin_unlock(&ctx->completion_lock); 580a830ffd2SPavel Begunkov } 581a830ffd2SPavel Begunkov if (ctx->has_evfd) 58221a091b9SDylan Yudaken io_eventfd_flush_signal(ctx); 583a830ffd2SPavel Begunkov } 584a830ffd2SPavel Begunkov 5856c16fe3cSJens Axboe void io_cq_unlock_post(struct io_ring_ctx *ctx) 58625399321SPavel Begunkov __releases(ctx->completion_lock) 58725399321SPavel Begunkov { 58825399321SPavel Begunkov io_commit_cqring(ctx); 58925399321SPavel Begunkov spin_unlock(&ctx->completion_lock); 59025399321SPavel Begunkov 5916c16fe3cSJens Axboe io_commit_cqring_flush(ctx); 5926c16fe3cSJens Axboe io_cqring_wake(ctx); 59325399321SPavel Begunkov } 59425399321SPavel Begunkov 595ed29b0b4SJens Axboe /* Returns true if there are no backlogged entries after the flush */ 596ed29b0b4SJens Axboe static bool __io_cqring_overflow_flush(struct io_ring_ctx *ctx, bool force) 597ed29b0b4SJens Axboe { 598305bef98SPavel Begunkov bool all_flushed; 599ed29b0b4SJens Axboe size_t cqe_size = sizeof(struct io_uring_cqe); 600ed29b0b4SJens Axboe 601ed29b0b4SJens Axboe if (!force && __io_cqring_events(ctx) == ctx->cq_entries) 602ed29b0b4SJens Axboe return false; 603ed29b0b4SJens Axboe 604ed29b0b4SJens Axboe if (ctx->flags & IORING_SETUP_CQE32) 605ed29b0b4SJens Axboe cqe_size <<= 1; 606ed29b0b4SJens Axboe 60725399321SPavel Begunkov io_cq_lock(ctx); 608ed29b0b4SJens Axboe while (!list_empty(&ctx->cq_overflow_list)) { 609aa1df3a3SPavel Begunkov struct io_uring_cqe *cqe = io_get_cqe_overflow(ctx, true); 610ed29b0b4SJens Axboe struct io_overflow_cqe *ocqe; 611ed29b0b4SJens Axboe 612ed29b0b4SJens Axboe if (!cqe && !force) 613ed29b0b4SJens Axboe break; 614ed29b0b4SJens Axboe ocqe = list_first_entry(&ctx->cq_overflow_list, 615ed29b0b4SJens Axboe struct io_overflow_cqe, list); 616ed29b0b4SJens Axboe if (cqe) 617ed29b0b4SJens Axboe memcpy(cqe, &ocqe->cqe, cqe_size); 618ed29b0b4SJens Axboe else 619ed29b0b4SJens Axboe io_account_cq_overflow(ctx); 620ed29b0b4SJens Axboe 621ed29b0b4SJens Axboe list_del(&ocqe->list); 622ed29b0b4SJens Axboe kfree(ocqe); 623ed29b0b4SJens Axboe } 624ed29b0b4SJens Axboe 625ed29b0b4SJens Axboe all_flushed = list_empty(&ctx->cq_overflow_list); 626ed29b0b4SJens Axboe if (all_flushed) { 627ed29b0b4SJens Axboe clear_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq); 628ed29b0b4SJens Axboe atomic_andnot(IORING_SQ_CQ_OVERFLOW, &ctx->rings->sq_flags); 629ed29b0b4SJens Axboe } 630ed29b0b4SJens Axboe 63125399321SPavel Begunkov io_cq_unlock_post(ctx); 632ed29b0b4SJens Axboe return all_flushed; 633ed29b0b4SJens Axboe } 634ed29b0b4SJens Axboe 635ed29b0b4SJens Axboe static bool io_cqring_overflow_flush(struct io_ring_ctx *ctx) 636ed29b0b4SJens Axboe { 637ed29b0b4SJens Axboe bool ret = true; 638ed29b0b4SJens Axboe 639ed29b0b4SJens Axboe if (test_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq)) { 640ed29b0b4SJens Axboe /* iopoll syncs against uring_lock, not completion_lock */ 641ed29b0b4SJens Axboe if (ctx->flags & IORING_SETUP_IOPOLL) 642ed29b0b4SJens Axboe mutex_lock(&ctx->uring_lock); 643ed29b0b4SJens Axboe ret = __io_cqring_overflow_flush(ctx, false); 644ed29b0b4SJens Axboe if (ctx->flags & IORING_SETUP_IOPOLL) 645ed29b0b4SJens Axboe mutex_unlock(&ctx->uring_lock); 646ed29b0b4SJens Axboe } 647ed29b0b4SJens Axboe 648ed29b0b4SJens Axboe return ret; 649ed29b0b4SJens Axboe } 650ed29b0b4SJens Axboe 651e70cb608SPavel Begunkov void __io_put_task(struct task_struct *task, int nr) 652ed29b0b4SJens Axboe { 653ed29b0b4SJens Axboe struct io_uring_task *tctx = task->io_uring; 654ed29b0b4SJens Axboe 655ed29b0b4SJens Axboe percpu_counter_sub(&tctx->inflight, nr); 656ed29b0b4SJens Axboe if (unlikely(atomic_read(&tctx->in_idle))) 657ed29b0b4SJens Axboe wake_up(&tctx->wait); 658ed29b0b4SJens Axboe put_task_struct_many(task, nr); 659ed29b0b4SJens Axboe } 660ed29b0b4SJens Axboe 66163809137SPavel Begunkov void io_task_refs_refill(struct io_uring_task *tctx) 662ed29b0b4SJens Axboe { 663ed29b0b4SJens Axboe unsigned int refill = -tctx->cached_refs + IO_TCTX_REFS_CACHE_NR; 664ed29b0b4SJens Axboe 665ed29b0b4SJens Axboe percpu_counter_add(&tctx->inflight, refill); 666ed29b0b4SJens Axboe refcount_add(refill, ¤t->usage); 667ed29b0b4SJens Axboe tctx->cached_refs += refill; 668ed29b0b4SJens Axboe } 669ed29b0b4SJens Axboe 670ed29b0b4SJens Axboe static __cold void io_uring_drop_tctx_refs(struct task_struct *task) 671ed29b0b4SJens Axboe { 672ed29b0b4SJens Axboe struct io_uring_task *tctx = task->io_uring; 673ed29b0b4SJens Axboe unsigned int refs = tctx->cached_refs; 674ed29b0b4SJens Axboe 675ed29b0b4SJens Axboe if (refs) { 676ed29b0b4SJens Axboe tctx->cached_refs = 0; 677ed29b0b4SJens Axboe percpu_counter_sub(&tctx->inflight, refs); 678ed29b0b4SJens Axboe put_task_struct_many(task, refs); 679ed29b0b4SJens Axboe } 680ed29b0b4SJens Axboe } 681ed29b0b4SJens Axboe 68268494a65SPavel Begunkov static bool io_cqring_event_overflow(struct io_ring_ctx *ctx, u64 user_data, 68368494a65SPavel Begunkov s32 res, u32 cflags, u64 extra1, u64 extra2) 684ed29b0b4SJens Axboe { 685ed29b0b4SJens Axboe struct io_overflow_cqe *ocqe; 686ed29b0b4SJens Axboe size_t ocq_size = sizeof(struct io_overflow_cqe); 687ed29b0b4SJens Axboe bool is_cqe32 = (ctx->flags & IORING_SETUP_CQE32); 688ed29b0b4SJens Axboe 689ed29b0b4SJens Axboe if (is_cqe32) 690ed29b0b4SJens Axboe ocq_size += sizeof(struct io_uring_cqe); 691ed29b0b4SJens Axboe 692ed29b0b4SJens Axboe ocqe = kmalloc(ocq_size, GFP_ATOMIC | __GFP_ACCOUNT); 693ed29b0b4SJens Axboe trace_io_uring_cqe_overflow(ctx, user_data, res, cflags, ocqe); 694ed29b0b4SJens Axboe if (!ocqe) { 695ed29b0b4SJens Axboe /* 696ed29b0b4SJens Axboe * If we're in ring overflow flush mode, or in task cancel mode, 697ed29b0b4SJens Axboe * or cannot allocate an overflow entry, then we need to drop it 698ed29b0b4SJens Axboe * on the floor. 699ed29b0b4SJens Axboe */ 700ed29b0b4SJens Axboe io_account_cq_overflow(ctx); 701ed29b0b4SJens Axboe set_bit(IO_CHECK_CQ_DROPPED_BIT, &ctx->check_cq); 702ed29b0b4SJens Axboe return false; 703ed29b0b4SJens Axboe } 704ed29b0b4SJens Axboe if (list_empty(&ctx->cq_overflow_list)) { 705ed29b0b4SJens Axboe set_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq); 706ed29b0b4SJens Axboe atomic_or(IORING_SQ_CQ_OVERFLOW, &ctx->rings->sq_flags); 707ed29b0b4SJens Axboe 708ed29b0b4SJens Axboe } 709ed29b0b4SJens Axboe ocqe->cqe.user_data = user_data; 710ed29b0b4SJens Axboe ocqe->cqe.res = res; 711ed29b0b4SJens Axboe ocqe->cqe.flags = cflags; 712ed29b0b4SJens Axboe if (is_cqe32) { 713ed29b0b4SJens Axboe ocqe->cqe.big_cqe[0] = extra1; 714ed29b0b4SJens Axboe ocqe->cqe.big_cqe[1] = extra2; 715ed29b0b4SJens Axboe } 716ed29b0b4SJens Axboe list_add_tail(&ocqe->list, &ctx->cq_overflow_list); 717ed29b0b4SJens Axboe return true; 718ed29b0b4SJens Axboe } 719ed29b0b4SJens Axboe 72068494a65SPavel Begunkov bool io_req_cqe_overflow(struct io_kiocb *req) 72168494a65SPavel Begunkov { 72268494a65SPavel Begunkov if (!(req->flags & REQ_F_CQE32_INIT)) { 72368494a65SPavel Begunkov req->extra1 = 0; 72468494a65SPavel Begunkov req->extra2 = 0; 72568494a65SPavel Begunkov } 72668494a65SPavel Begunkov return io_cqring_event_overflow(req->ctx, req->cqe.user_data, 72768494a65SPavel Begunkov req->cqe.res, req->cqe.flags, 72868494a65SPavel Begunkov req->extra1, req->extra2); 72968494a65SPavel Begunkov } 73068494a65SPavel Begunkov 731faf88ddeSPavel Begunkov /* 732faf88ddeSPavel Begunkov * writes to the cq entry need to come after reading head; the 733faf88ddeSPavel Begunkov * control dependency is enough as we're using WRITE_ONCE to 734faf88ddeSPavel Begunkov * fill the cq entry 735faf88ddeSPavel Begunkov */ 736aa1df3a3SPavel Begunkov struct io_uring_cqe *__io_get_cqe(struct io_ring_ctx *ctx, bool overflow) 737faf88ddeSPavel Begunkov { 738faf88ddeSPavel Begunkov struct io_rings *rings = ctx->rings; 739faf88ddeSPavel Begunkov unsigned int off = ctx->cached_cq_tail & (ctx->cq_entries - 1); 740faf88ddeSPavel Begunkov unsigned int free, queued, len; 741faf88ddeSPavel Begunkov 742aa1df3a3SPavel Begunkov /* 743aa1df3a3SPavel Begunkov * Posting into the CQ when there are pending overflowed CQEs may break 744aa1df3a3SPavel Begunkov * ordering guarantees, which will affect links, F_MORE users and more. 745aa1df3a3SPavel Begunkov * Force overflow the completion. 746aa1df3a3SPavel Begunkov */ 747aa1df3a3SPavel Begunkov if (!overflow && (ctx->check_cq & BIT(IO_CHECK_CQ_OVERFLOW_BIT))) 748aa1df3a3SPavel Begunkov return NULL; 749faf88ddeSPavel Begunkov 750faf88ddeSPavel Begunkov /* userspace may cheat modifying the tail, be safe and do min */ 751faf88ddeSPavel Begunkov queued = min(__io_cqring_events(ctx), ctx->cq_entries); 752faf88ddeSPavel Begunkov free = ctx->cq_entries - queued; 753faf88ddeSPavel Begunkov /* we need a contiguous range, limit based on the current array offset */ 754faf88ddeSPavel Begunkov len = min(free, ctx->cq_entries - off); 755faf88ddeSPavel Begunkov if (!len) 756faf88ddeSPavel Begunkov return NULL; 757faf88ddeSPavel Begunkov 758b3659a65SPavel Begunkov if (ctx->flags & IORING_SETUP_CQE32) { 759b3659a65SPavel Begunkov off <<= 1; 760b3659a65SPavel Begunkov len <<= 1; 761b3659a65SPavel Begunkov } 762b3659a65SPavel Begunkov 763faf88ddeSPavel Begunkov ctx->cqe_cached = &rings->cqes[off]; 764faf88ddeSPavel Begunkov ctx->cqe_sentinel = ctx->cqe_cached + len; 765b3659a65SPavel Begunkov 766b3659a65SPavel Begunkov ctx->cached_cq_tail++; 767faf88ddeSPavel Begunkov ctx->cqe_cached++; 768b3659a65SPavel Begunkov if (ctx->flags & IORING_SETUP_CQE32) 769b3659a65SPavel Begunkov ctx->cqe_cached++; 770b3659a65SPavel Begunkov return &rings->cqes[off]; 771faf88ddeSPavel Begunkov } 772faf88ddeSPavel Begunkov 773a77ab745SDylan Yudaken static bool io_fill_cqe_aux(struct io_ring_ctx *ctx, u64 user_data, s32 res, u32 cflags, 77452120f0fSDylan Yudaken bool allow_overflow) 775ed29b0b4SJens Axboe { 776ed29b0b4SJens Axboe struct io_uring_cqe *cqe; 777ed29b0b4SJens Axboe 778*2e2ef4a1SDylan Yudaken lockdep_assert_held(&ctx->completion_lock); 779*2e2ef4a1SDylan Yudaken 780ed29b0b4SJens Axboe ctx->cq_extra++; 781ed29b0b4SJens Axboe 782ed29b0b4SJens Axboe /* 783ed29b0b4SJens Axboe * If we can't get a cq entry, userspace overflowed the 784ed29b0b4SJens Axboe * submission (by quite a lot). Increment the overflow count in 785ed29b0b4SJens Axboe * the ring. 786ed29b0b4SJens Axboe */ 787ed29b0b4SJens Axboe cqe = io_get_cqe(ctx); 788ed29b0b4SJens Axboe if (likely(cqe)) { 789e0486f3fSDylan Yudaken trace_io_uring_complete(ctx, NULL, user_data, res, cflags, 0, 0); 790e0486f3fSDylan Yudaken 791ed29b0b4SJens Axboe WRITE_ONCE(cqe->user_data, user_data); 792ed29b0b4SJens Axboe WRITE_ONCE(cqe->res, res); 793ed29b0b4SJens Axboe WRITE_ONCE(cqe->flags, cflags); 794ed29b0b4SJens Axboe 795ed29b0b4SJens Axboe if (ctx->flags & IORING_SETUP_CQE32) { 796ed29b0b4SJens Axboe WRITE_ONCE(cqe->big_cqe[0], 0); 797ed29b0b4SJens Axboe WRITE_ONCE(cqe->big_cqe[1], 0); 798ed29b0b4SJens Axboe } 799ed29b0b4SJens Axboe return true; 800ed29b0b4SJens Axboe } 80152120f0fSDylan Yudaken 80252120f0fSDylan Yudaken if (allow_overflow) 803ed29b0b4SJens Axboe return io_cqring_event_overflow(ctx, user_data, res, cflags, 0, 0); 80452120f0fSDylan Yudaken 80552120f0fSDylan Yudaken return false; 806ed29b0b4SJens Axboe } 807ed29b0b4SJens Axboe 808931147ddSDylan Yudaken static void __io_flush_post_cqes(struct io_ring_ctx *ctx) 809931147ddSDylan Yudaken __must_hold(&ctx->uring_lock) 810931147ddSDylan Yudaken { 811931147ddSDylan Yudaken struct io_submit_state *state = &ctx->submit_state; 812931147ddSDylan Yudaken unsigned int i; 813931147ddSDylan Yudaken 814931147ddSDylan Yudaken lockdep_assert_held(&ctx->uring_lock); 815931147ddSDylan Yudaken for (i = 0; i < state->cqes_count; i++) { 816931147ddSDylan Yudaken struct io_uring_cqe *cqe = &state->cqes[i]; 817931147ddSDylan Yudaken 818931147ddSDylan Yudaken io_fill_cqe_aux(ctx, cqe->user_data, cqe->res, cqe->flags, true); 819931147ddSDylan Yudaken } 820931147ddSDylan Yudaken state->cqes_count = 0; 821931147ddSDylan Yudaken } 822931147ddSDylan Yudaken 823d245bca6SPavel Begunkov bool io_post_aux_cqe(struct io_ring_ctx *ctx, 82452120f0fSDylan Yudaken u64 user_data, s32 res, u32 cflags, 82552120f0fSDylan Yudaken bool allow_overflow) 826d245bca6SPavel Begunkov { 827d245bca6SPavel Begunkov bool filled; 828d245bca6SPavel Begunkov 82925399321SPavel Begunkov io_cq_lock(ctx); 83052120f0fSDylan Yudaken filled = io_fill_cqe_aux(ctx, user_data, res, cflags, allow_overflow); 83125399321SPavel Begunkov io_cq_unlock_post(ctx); 832d245bca6SPavel Begunkov return filled; 833d245bca6SPavel Begunkov } 834d245bca6SPavel Begunkov 8359b8c5475SDylan Yudaken bool io_aux_cqe(struct io_ring_ctx *ctx, bool defer, u64 user_data, s32 res, u32 cflags, 8369b8c5475SDylan Yudaken bool allow_overflow) 8379b8c5475SDylan Yudaken { 8389b8c5475SDylan Yudaken struct io_uring_cqe *cqe; 8399b8c5475SDylan Yudaken unsigned int length; 8409b8c5475SDylan Yudaken 8419b8c5475SDylan Yudaken if (!defer) 8429b8c5475SDylan Yudaken return io_post_aux_cqe(ctx, user_data, res, cflags, allow_overflow); 8439b8c5475SDylan Yudaken 8449b8c5475SDylan Yudaken length = ARRAY_SIZE(ctx->submit_state.cqes); 8459b8c5475SDylan Yudaken 8469b8c5475SDylan Yudaken lockdep_assert_held(&ctx->uring_lock); 8479b8c5475SDylan Yudaken 8489b8c5475SDylan Yudaken if (ctx->submit_state.cqes_count == length) { 8499b8c5475SDylan Yudaken io_cq_lock(ctx); 8509b8c5475SDylan Yudaken __io_flush_post_cqes(ctx); 8519b8c5475SDylan Yudaken /* no need to flush - flush is deferred */ 8529b8c5475SDylan Yudaken spin_unlock(&ctx->completion_lock); 8539b8c5475SDylan Yudaken } 8549b8c5475SDylan Yudaken 8559b8c5475SDylan Yudaken /* For defered completions this is not as strict as it is otherwise, 8569b8c5475SDylan Yudaken * however it's main job is to prevent unbounded posted completions, 8579b8c5475SDylan Yudaken * and in that it works just as well. 8589b8c5475SDylan Yudaken */ 8599b8c5475SDylan Yudaken if (!allow_overflow && test_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq)) 8609b8c5475SDylan Yudaken return false; 8619b8c5475SDylan Yudaken 8629b8c5475SDylan Yudaken cqe = &ctx->submit_state.cqes[ctx->submit_state.cqes_count++]; 8639b8c5475SDylan Yudaken cqe->user_data = user_data; 8649b8c5475SDylan Yudaken cqe->res = res; 8659b8c5475SDylan Yudaken cqe->flags = cflags; 8669b8c5475SDylan Yudaken return true; 8679b8c5475SDylan Yudaken } 8689b8c5475SDylan Yudaken 8691bec951cSPavel Begunkov static void __io_req_complete_post(struct io_kiocb *req) 870ed29b0b4SJens Axboe { 871fa18fa22SPavel Begunkov struct io_ring_ctx *ctx = req->ctx; 872fa18fa22SPavel Begunkov 873fa18fa22SPavel Begunkov io_cq_lock(ctx); 874fa18fa22SPavel Begunkov if (!(req->flags & REQ_F_CQE_SKIP)) 875fa18fa22SPavel Begunkov __io_fill_cqe_req(ctx, req); 876fa18fa22SPavel Begunkov 877ed29b0b4SJens Axboe /* 878ed29b0b4SJens Axboe * If we're the last reference to this request, add to our locked 879ed29b0b4SJens Axboe * free_list cache. 880ed29b0b4SJens Axboe */ 881ed29b0b4SJens Axboe if (req_ref_put_and_test(req)) { 882ed29b0b4SJens Axboe if (req->flags & IO_REQ_LINK_FLAGS) { 883ed29b0b4SJens Axboe if (req->flags & IO_DISARM_MASK) 884ed29b0b4SJens Axboe io_disarm_next(req); 885ed29b0b4SJens Axboe if (req->link) { 886ed29b0b4SJens Axboe io_req_task_queue(req->link); 887ed29b0b4SJens Axboe req->link = NULL; 888ed29b0b4SJens Axboe } 889ed29b0b4SJens Axboe } 890ed29b0b4SJens Axboe io_req_put_rsrc(req); 891ed29b0b4SJens Axboe /* 892ed29b0b4SJens Axboe * Selected buffer deallocation in io_clean_op() assumes that 893ed29b0b4SJens Axboe * we don't hold ->completion_lock. Clean them here to avoid 894ed29b0b4SJens Axboe * deadlocks. 895ed29b0b4SJens Axboe */ 896ed29b0b4SJens Axboe io_put_kbuf_comp(req); 897ed29b0b4SJens Axboe io_dismantle_req(req); 898ed29b0b4SJens Axboe io_put_task(req->task, 1); 899ed29b0b4SJens Axboe wq_list_add_head(&req->comp_list, &ctx->locked_free_list); 900ed29b0b4SJens Axboe ctx->locked_free_nr++; 901ed29b0b4SJens Axboe } 90225399321SPavel Begunkov io_cq_unlock_post(ctx); 903ed29b0b4SJens Axboe } 904ed29b0b4SJens Axboe 9051bec951cSPavel Begunkov void io_req_complete_post(struct io_kiocb *req, unsigned issue_flags) 906ed29b0b4SJens Axboe { 9071bec951cSPavel Begunkov if (!(issue_flags & IO_URING_F_UNLOCKED) || 9081bec951cSPavel Begunkov !(req->ctx->flags & IORING_SETUP_IOPOLL)) { 9091bec951cSPavel Begunkov __io_req_complete_post(req); 9101bec951cSPavel Begunkov } else { 9111bec951cSPavel Begunkov struct io_ring_ctx *ctx = req->ctx; 9121bec951cSPavel Begunkov 9131bec951cSPavel Begunkov mutex_lock(&ctx->uring_lock); 9141bec951cSPavel Begunkov __io_req_complete_post(req); 9151bec951cSPavel Begunkov mutex_unlock(&ctx->uring_lock); 9161bec951cSPavel Begunkov } 917ed29b0b4SJens Axboe } 918ed29b0b4SJens Axboe 919973fc83fSDylan Yudaken void io_req_defer_failed(struct io_kiocb *req, s32 res) 920e276ae34SPavel Begunkov __must_hold(&ctx->uring_lock) 921ed29b0b4SJens Axboe { 922a47b255eSPavel Begunkov const struct io_op_def *def = &io_op_defs[req->opcode]; 923a47b255eSPavel Begunkov 924e276ae34SPavel Begunkov lockdep_assert_held(&req->ctx->uring_lock); 925e276ae34SPavel Begunkov 926ed29b0b4SJens Axboe req_set_fail(req); 92797b388d7SJens Axboe io_req_set_res(req, res, io_put_kbuf(req, IO_URING_F_UNLOCKED)); 928a47b255eSPavel Begunkov if (def->fail) 929a47b255eSPavel Begunkov def->fail(req); 930973fc83fSDylan Yudaken io_req_complete_defer(req); 931ed29b0b4SJens Axboe } 932ed29b0b4SJens Axboe 933ed29b0b4SJens Axboe /* 934ed29b0b4SJens Axboe * Don't initialise the fields below on every allocation, but do that in 935ed29b0b4SJens Axboe * advance and keep them valid across allocations. 936ed29b0b4SJens Axboe */ 937ed29b0b4SJens Axboe static void io_preinit_req(struct io_kiocb *req, struct io_ring_ctx *ctx) 938ed29b0b4SJens Axboe { 939ed29b0b4SJens Axboe req->ctx = ctx; 940ed29b0b4SJens Axboe req->link = NULL; 941ed29b0b4SJens Axboe req->async_data = NULL; 942ed29b0b4SJens Axboe /* not necessary, but safer to zero */ 943ed29b0b4SJens Axboe req->cqe.res = 0; 944ed29b0b4SJens Axboe } 945ed29b0b4SJens Axboe 946ed29b0b4SJens Axboe static void io_flush_cached_locked_reqs(struct io_ring_ctx *ctx, 947ed29b0b4SJens Axboe struct io_submit_state *state) 948ed29b0b4SJens Axboe { 949ed29b0b4SJens Axboe spin_lock(&ctx->completion_lock); 950ed29b0b4SJens Axboe wq_list_splice(&ctx->locked_free_list, &state->free_list); 951ed29b0b4SJens Axboe ctx->locked_free_nr = 0; 952ed29b0b4SJens Axboe spin_unlock(&ctx->completion_lock); 953ed29b0b4SJens Axboe } 954ed29b0b4SJens Axboe 955ed29b0b4SJens Axboe /* 956ed29b0b4SJens Axboe * A request might get retired back into the request caches even before opcode 957ed29b0b4SJens Axboe * handlers and io_issue_sqe() are done with it, e.g. inline completion path. 958ed29b0b4SJens Axboe * Because of that, io_alloc_req() should be called only under ->uring_lock 959ed29b0b4SJens Axboe * and with extra caution to not get a request that is still worked on. 960ed29b0b4SJens Axboe */ 961bd1a3783SPavel Begunkov __cold bool __io_alloc_req_refill(struct io_ring_ctx *ctx) 962ed29b0b4SJens Axboe __must_hold(&ctx->uring_lock) 963ed29b0b4SJens Axboe { 964ed29b0b4SJens Axboe gfp_t gfp = GFP_KERNEL | __GFP_NOWARN; 965ed29b0b4SJens Axboe void *reqs[IO_REQ_ALLOC_BATCH]; 966ed29b0b4SJens Axboe int ret, i; 967ed29b0b4SJens Axboe 968ed29b0b4SJens Axboe /* 969ed29b0b4SJens Axboe * If we have more than a batch's worth of requests in our IRQ side 970ed29b0b4SJens Axboe * locked cache, grab the lock and move them over to our submission 971ed29b0b4SJens Axboe * side cache. 972ed29b0b4SJens Axboe */ 973ed29b0b4SJens Axboe if (data_race(ctx->locked_free_nr) > IO_COMPL_BATCH) { 974ed29b0b4SJens Axboe io_flush_cached_locked_reqs(ctx, &ctx->submit_state); 975ed29b0b4SJens Axboe if (!io_req_cache_empty(ctx)) 976ed29b0b4SJens Axboe return true; 977ed29b0b4SJens Axboe } 978ed29b0b4SJens Axboe 979ed29b0b4SJens Axboe ret = kmem_cache_alloc_bulk(req_cachep, gfp, ARRAY_SIZE(reqs), reqs); 980ed29b0b4SJens Axboe 981ed29b0b4SJens Axboe /* 982ed29b0b4SJens Axboe * Bulk alloc is all-or-nothing. If we fail to get a batch, 983ed29b0b4SJens Axboe * retry single alloc to be on the safe side. 984ed29b0b4SJens Axboe */ 985ed29b0b4SJens Axboe if (unlikely(ret <= 0)) { 986ed29b0b4SJens Axboe reqs[0] = kmem_cache_alloc(req_cachep, gfp); 987ed29b0b4SJens Axboe if (!reqs[0]) 988ed29b0b4SJens Axboe return false; 989ed29b0b4SJens Axboe ret = 1; 990ed29b0b4SJens Axboe } 991ed29b0b4SJens Axboe 992ed29b0b4SJens Axboe percpu_ref_get_many(&ctx->refs, ret); 993ed29b0b4SJens Axboe for (i = 0; i < ret; i++) { 994ed29b0b4SJens Axboe struct io_kiocb *req = reqs[i]; 995ed29b0b4SJens Axboe 996ed29b0b4SJens Axboe io_preinit_req(req, ctx); 997ed29b0b4SJens Axboe io_req_add_to_cache(req, ctx); 998ed29b0b4SJens Axboe } 999ed29b0b4SJens Axboe return true; 1000ed29b0b4SJens Axboe } 1001ed29b0b4SJens Axboe 1002ed29b0b4SJens Axboe static inline void io_dismantle_req(struct io_kiocb *req) 1003ed29b0b4SJens Axboe { 1004ed29b0b4SJens Axboe unsigned int flags = req->flags; 1005ed29b0b4SJens Axboe 1006ed29b0b4SJens Axboe if (unlikely(flags & IO_REQ_CLEAN_FLAGS)) 1007ed29b0b4SJens Axboe io_clean_op(req); 1008ed29b0b4SJens Axboe if (!(flags & REQ_F_FIXED_FILE)) 1009ed29b0b4SJens Axboe io_put_file(req->file); 1010ed29b0b4SJens Axboe } 1011ed29b0b4SJens Axboe 101259915143SJens Axboe __cold void io_free_req(struct io_kiocb *req) 1013ed29b0b4SJens Axboe { 1014ed29b0b4SJens Axboe struct io_ring_ctx *ctx = req->ctx; 1015ed29b0b4SJens Axboe 1016ed29b0b4SJens Axboe io_req_put_rsrc(req); 1017ed29b0b4SJens Axboe io_dismantle_req(req); 1018ed29b0b4SJens Axboe io_put_task(req->task, 1); 1019ed29b0b4SJens Axboe 1020ed29b0b4SJens Axboe spin_lock(&ctx->completion_lock); 1021ed29b0b4SJens Axboe wq_list_add_head(&req->comp_list, &ctx->locked_free_list); 1022ed29b0b4SJens Axboe ctx->locked_free_nr++; 1023ed29b0b4SJens Axboe spin_unlock(&ctx->completion_lock); 1024ed29b0b4SJens Axboe } 1025ed29b0b4SJens Axboe 1026ed29b0b4SJens Axboe static void __io_req_find_next_prep(struct io_kiocb *req) 1027ed29b0b4SJens Axboe { 1028ed29b0b4SJens Axboe struct io_ring_ctx *ctx = req->ctx; 1029ed29b0b4SJens Axboe 103025399321SPavel Begunkov io_cq_lock(ctx); 1031305bef98SPavel Begunkov io_disarm_next(req); 103225399321SPavel Begunkov io_cq_unlock_post(ctx); 1033ed29b0b4SJens Axboe } 1034ed29b0b4SJens Axboe 1035ed29b0b4SJens Axboe static inline struct io_kiocb *io_req_find_next(struct io_kiocb *req) 1036ed29b0b4SJens Axboe { 1037ed29b0b4SJens Axboe struct io_kiocb *nxt; 1038ed29b0b4SJens Axboe 1039ed29b0b4SJens Axboe /* 1040ed29b0b4SJens Axboe * If LINK is set, we have dependent requests in this chain. If we 1041ed29b0b4SJens Axboe * didn't fail this request, queue the first one up, moving any other 1042ed29b0b4SJens Axboe * dependencies to the next request. In case of failure, fail the rest 1043ed29b0b4SJens Axboe * of the chain. 1044ed29b0b4SJens Axboe */ 1045ed29b0b4SJens Axboe if (unlikely(req->flags & IO_DISARM_MASK)) 1046ed29b0b4SJens Axboe __io_req_find_next_prep(req); 1047ed29b0b4SJens Axboe nxt = req->link; 1048ed29b0b4SJens Axboe req->link = NULL; 1049ed29b0b4SJens Axboe return nxt; 1050ed29b0b4SJens Axboe } 1051ed29b0b4SJens Axboe 1052ed29b0b4SJens Axboe static void ctx_flush_and_put(struct io_ring_ctx *ctx, bool *locked) 1053ed29b0b4SJens Axboe { 1054ed29b0b4SJens Axboe if (!ctx) 1055ed29b0b4SJens Axboe return; 1056ed29b0b4SJens Axboe if (ctx->flags & IORING_SETUP_TASKRUN_FLAG) 1057ed29b0b4SJens Axboe atomic_andnot(IORING_SQ_TASKRUN, &ctx->rings->sq_flags); 1058ed29b0b4SJens Axboe if (*locked) { 1059ed29b0b4SJens Axboe io_submit_flush_completions(ctx); 1060ed29b0b4SJens Axboe mutex_unlock(&ctx->uring_lock); 1061ed29b0b4SJens Axboe *locked = false; 1062ed29b0b4SJens Axboe } 1063ed29b0b4SJens Axboe percpu_ref_put(&ctx->refs); 1064ed29b0b4SJens Axboe } 1065ed29b0b4SJens Axboe 1066c6dd763cSDylan Yudaken static unsigned int handle_tw_list(struct llist_node *node, 10673a0c037bSDylan Yudaken struct io_ring_ctx **ctx, bool *locked, 10683a0c037bSDylan Yudaken struct llist_node *last) 1069ed29b0b4SJens Axboe { 1070c6dd763cSDylan Yudaken unsigned int count = 0; 1071c6dd763cSDylan Yudaken 10723a0c037bSDylan Yudaken while (node != last) { 1073f88262e6SDylan Yudaken struct llist_node *next = node->next; 1074ed29b0b4SJens Axboe struct io_kiocb *req = container_of(node, struct io_kiocb, 1075ed29b0b4SJens Axboe io_task_work.node); 1076ed29b0b4SJens Axboe 1077ed29b0b4SJens Axboe prefetch(container_of(next, struct io_kiocb, io_task_work.node)); 1078ed29b0b4SJens Axboe 1079ed29b0b4SJens Axboe if (req->ctx != *ctx) { 1080ed29b0b4SJens Axboe ctx_flush_and_put(*ctx, locked); 1081ed29b0b4SJens Axboe *ctx = req->ctx; 1082ed29b0b4SJens Axboe /* if not contended, grab and improve batching */ 1083ed29b0b4SJens Axboe *locked = mutex_trylock(&(*ctx)->uring_lock); 1084ed29b0b4SJens Axboe percpu_ref_get(&(*ctx)->refs); 1085ed29b0b4SJens Axboe } 1086ed29b0b4SJens Axboe req->io_task_work.func(req, locked); 1087ed29b0b4SJens Axboe node = next; 1088c6dd763cSDylan Yudaken count++; 10893a0c037bSDylan Yudaken } 1090c6dd763cSDylan Yudaken 1091c6dd763cSDylan Yudaken return count; 1092ed29b0b4SJens Axboe } 1093ed29b0b4SJens Axboe 1094923d1592SDylan Yudaken /** 1095923d1592SDylan Yudaken * io_llist_xchg - swap all entries in a lock-less list 1096923d1592SDylan Yudaken * @head: the head of lock-less list to delete all entries 1097923d1592SDylan Yudaken * @new: new entry as the head of the list 1098923d1592SDylan Yudaken * 1099923d1592SDylan Yudaken * If list is empty, return NULL, otherwise, return the pointer to the first entry. 1100923d1592SDylan Yudaken * The order of entries returned is from the newest to the oldest added one. 1101923d1592SDylan Yudaken */ 1102923d1592SDylan Yudaken static inline struct llist_node *io_llist_xchg(struct llist_head *head, 1103923d1592SDylan Yudaken struct llist_node *new) 1104923d1592SDylan Yudaken { 1105923d1592SDylan Yudaken return xchg(&head->first, new); 1106923d1592SDylan Yudaken } 1107923d1592SDylan Yudaken 1108923d1592SDylan Yudaken /** 1109923d1592SDylan Yudaken * io_llist_cmpxchg - possibly swap all entries in a lock-less list 1110923d1592SDylan Yudaken * @head: the head of lock-less list to delete all entries 1111923d1592SDylan Yudaken * @old: expected old value of the first entry of the list 1112923d1592SDylan Yudaken * @new: new entry as the head of the list 1113923d1592SDylan Yudaken * 1114923d1592SDylan Yudaken * perform a cmpxchg on the first entry of the list. 1115923d1592SDylan Yudaken */ 1116923d1592SDylan Yudaken 1117923d1592SDylan Yudaken static inline struct llist_node *io_llist_cmpxchg(struct llist_head *head, 1118923d1592SDylan Yudaken struct llist_node *old, 1119923d1592SDylan Yudaken struct llist_node *new) 1120923d1592SDylan Yudaken { 1121923d1592SDylan Yudaken return cmpxchg(&head->first, old, new); 1122923d1592SDylan Yudaken } 1123923d1592SDylan Yudaken 1124c9f06aa7SJens Axboe void tctx_task_work(struct callback_head *cb) 1125ed29b0b4SJens Axboe { 1126ed29b0b4SJens Axboe bool uring_locked = false; 1127ed29b0b4SJens Axboe struct io_ring_ctx *ctx = NULL; 1128ed29b0b4SJens Axboe struct io_uring_task *tctx = container_of(cb, struct io_uring_task, 1129ed29b0b4SJens Axboe task_work); 11303a0c037bSDylan Yudaken struct llist_node fake = {}; 11313a0c037bSDylan Yudaken struct llist_node *node = io_llist_xchg(&tctx->task_list, &fake); 1132c6dd763cSDylan Yudaken unsigned int loops = 1; 1133c6dd763cSDylan Yudaken unsigned int count = handle_tw_list(node, &ctx, &uring_locked, NULL); 1134ed29b0b4SJens Axboe 11353a0c037bSDylan Yudaken node = io_llist_cmpxchg(&tctx->task_list, &fake, NULL); 11363a0c037bSDylan Yudaken while (node != &fake) { 1137c6dd763cSDylan Yudaken loops++; 11383a0c037bSDylan Yudaken node = io_llist_xchg(&tctx->task_list, &fake); 1139c6dd763cSDylan Yudaken count += handle_tw_list(node, &ctx, &uring_locked, &fake); 11403a0c037bSDylan Yudaken node = io_llist_cmpxchg(&tctx->task_list, &fake, NULL); 1141ed29b0b4SJens Axboe } 1142ed29b0b4SJens Axboe 1143ed29b0b4SJens Axboe ctx_flush_and_put(ctx, &uring_locked); 1144ed29b0b4SJens Axboe 1145ed29b0b4SJens Axboe /* relaxed read is enough as only the task itself sets ->in_idle */ 1146ed29b0b4SJens Axboe if (unlikely(atomic_read(&tctx->in_idle))) 1147ed29b0b4SJens Axboe io_uring_drop_tctx_refs(current); 1148c6dd763cSDylan Yudaken 1149c6dd763cSDylan Yudaken trace_io_uring_task_work_run(tctx, count, loops); 1150ed29b0b4SJens Axboe } 1151ed29b0b4SJens Axboe 1152d7593606SPavel Begunkov static __cold void io_fallback_tw(struct io_uring_task *tctx) 1153d7593606SPavel Begunkov { 1154d7593606SPavel Begunkov struct llist_node *node = llist_del_all(&tctx->task_list); 1155d7593606SPavel Begunkov struct io_kiocb *req; 1156d7593606SPavel Begunkov 1157d7593606SPavel Begunkov while (node) { 1158d7593606SPavel Begunkov req = container_of(node, struct io_kiocb, io_task_work.node); 1159d7593606SPavel Begunkov node = node->next; 1160d7593606SPavel Begunkov if (llist_add(&req->io_task_work.node, 1161d7593606SPavel Begunkov &req->ctx->fallback_llist)) 1162d7593606SPavel Begunkov schedule_delayed_work(&req->ctx->fallback_work, 1); 1163d7593606SPavel Begunkov } 1164d7593606SPavel Begunkov } 1165d7593606SPavel Begunkov 1166c0e0d6baSDylan Yudaken static void io_req_local_work_add(struct io_kiocb *req) 1167c0e0d6baSDylan Yudaken { 1168c0e0d6baSDylan Yudaken struct io_ring_ctx *ctx = req->ctx; 1169c0e0d6baSDylan Yudaken 1170c0e0d6baSDylan Yudaken if (!llist_add(&req->io_task_work.node, &ctx->work_llist)) 1171c0e0d6baSDylan Yudaken return; 1172fc86f9d3SPavel Begunkov /* need it for the following io_cqring_wake() */ 1173fc86f9d3SPavel Begunkov smp_mb__after_atomic(); 1174c0e0d6baSDylan Yudaken 1175c0e0d6baSDylan Yudaken if (unlikely(atomic_read(&req->task->io_uring->in_idle))) { 1176c0e0d6baSDylan Yudaken io_move_task_work_from_local(ctx); 1177c0e0d6baSDylan Yudaken return; 1178c0e0d6baSDylan Yudaken } 1179c0e0d6baSDylan Yudaken 1180c0e0d6baSDylan Yudaken if (ctx->flags & IORING_SETUP_TASKRUN_FLAG) 1181c0e0d6baSDylan Yudaken atomic_or(IORING_SQ_TASKRUN, &ctx->rings->sq_flags); 1182c0e0d6baSDylan Yudaken 118321a091b9SDylan Yudaken if (ctx->has_evfd) 118421a091b9SDylan Yudaken io_eventfd_signal(ctx); 1185fc86f9d3SPavel Begunkov __io_cqring_wake(ctx); 1186c0e0d6baSDylan Yudaken } 1187c0e0d6baSDylan Yudaken 1188e52d2e58SPavel Begunkov void __io_req_task_work_add(struct io_kiocb *req, bool allow_local) 1189ed29b0b4SJens Axboe { 1190c34398a8SDylan Yudaken struct io_uring_task *tctx = req->task->io_uring; 1191ed29b0b4SJens Axboe struct io_ring_ctx *ctx = req->ctx; 1192ed29b0b4SJens Axboe 1193c0e0d6baSDylan Yudaken if (allow_local && ctx->flags & IORING_SETUP_DEFER_TASKRUN) { 1194c0e0d6baSDylan Yudaken io_req_local_work_add(req); 1195c0e0d6baSDylan Yudaken return; 1196c0e0d6baSDylan Yudaken } 1197ed29b0b4SJens Axboe 1198ed29b0b4SJens Axboe /* task_work already pending, we're done */ 119932d91f05SDylan Yudaken if (!llist_add(&req->io_task_work.node, &tctx->task_list)) 1200ed29b0b4SJens Axboe return; 1201ed29b0b4SJens Axboe 1202ed29b0b4SJens Axboe if (ctx->flags & IORING_SETUP_TASKRUN_FLAG) 1203ed29b0b4SJens Axboe atomic_or(IORING_SQ_TASKRUN, &ctx->rings->sq_flags); 1204ed29b0b4SJens Axboe 1205ed29b0b4SJens Axboe if (likely(!task_work_add(req->task, &tctx->task_work, ctx->notify_method))) 1206ed29b0b4SJens Axboe return; 1207ed29b0b4SJens Axboe 1208d7593606SPavel Begunkov io_fallback_tw(tctx); 1209ed29b0b4SJens Axboe } 1210ed29b0b4SJens Axboe 1211c0e0d6baSDylan Yudaken static void __cold io_move_task_work_from_local(struct io_ring_ctx *ctx) 1212c0e0d6baSDylan Yudaken { 1213c0e0d6baSDylan Yudaken struct llist_node *node; 1214c0e0d6baSDylan Yudaken 1215c0e0d6baSDylan Yudaken node = llist_del_all(&ctx->work_llist); 1216c0e0d6baSDylan Yudaken while (node) { 1217c0e0d6baSDylan Yudaken struct io_kiocb *req = container_of(node, struct io_kiocb, 1218c0e0d6baSDylan Yudaken io_task_work.node); 1219c0e0d6baSDylan Yudaken 1220c0e0d6baSDylan Yudaken node = node->next; 1221c0e0d6baSDylan Yudaken __io_req_task_work_add(req, false); 1222c0e0d6baSDylan Yudaken } 1223c0e0d6baSDylan Yudaken } 1224c0e0d6baSDylan Yudaken 1225b3026767SDylan Yudaken int __io_run_local_work(struct io_ring_ctx *ctx, bool *locked) 1226c0e0d6baSDylan Yudaken { 1227c0e0d6baSDylan Yudaken struct llist_node *node; 1228c0e0d6baSDylan Yudaken struct llist_node fake; 1229c0e0d6baSDylan Yudaken struct llist_node *current_final = NULL; 1230c0e0d6baSDylan Yudaken int ret; 1231f75d5036SDylan Yudaken unsigned int loops = 1; 1232c0e0d6baSDylan Yudaken 12336567506bSPavel Begunkov if (unlikely(ctx->submitter_task != current)) 1234c0e0d6baSDylan Yudaken return -EEXIST; 1235c0e0d6baSDylan Yudaken 1236c0e0d6baSDylan Yudaken node = io_llist_xchg(&ctx->work_llist, &fake); 1237c0e0d6baSDylan Yudaken ret = 0; 1238c0e0d6baSDylan Yudaken again: 1239c0e0d6baSDylan Yudaken while (node != current_final) { 1240c0e0d6baSDylan Yudaken struct llist_node *next = node->next; 1241c0e0d6baSDylan Yudaken struct io_kiocb *req = container_of(node, struct io_kiocb, 1242c0e0d6baSDylan Yudaken io_task_work.node); 1243c0e0d6baSDylan Yudaken prefetch(container_of(next, struct io_kiocb, io_task_work.node)); 1244b3026767SDylan Yudaken req->io_task_work.func(req, locked); 1245c0e0d6baSDylan Yudaken ret++; 1246c0e0d6baSDylan Yudaken node = next; 1247c0e0d6baSDylan Yudaken } 1248c0e0d6baSDylan Yudaken 1249c0e0d6baSDylan Yudaken if (ctx->flags & IORING_SETUP_TASKRUN_FLAG) 1250c0e0d6baSDylan Yudaken atomic_andnot(IORING_SQ_TASKRUN, &ctx->rings->sq_flags); 1251c0e0d6baSDylan Yudaken 1252c0e0d6baSDylan Yudaken node = io_llist_cmpxchg(&ctx->work_llist, &fake, NULL); 1253c0e0d6baSDylan Yudaken if (node != &fake) { 1254f75d5036SDylan Yudaken loops++; 1255c0e0d6baSDylan Yudaken current_final = &fake; 1256c0e0d6baSDylan Yudaken node = io_llist_xchg(&ctx->work_llist, &fake); 1257c0e0d6baSDylan Yudaken goto again; 1258c0e0d6baSDylan Yudaken } 1259c0e0d6baSDylan Yudaken 1260b3026767SDylan Yudaken if (*locked) 1261c0e0d6baSDylan Yudaken io_submit_flush_completions(ctx); 1262f75d5036SDylan Yudaken trace_io_uring_local_work_run(ctx, ret, loops); 1263c0e0d6baSDylan Yudaken return ret; 12648ac5d85aSJens Axboe 12658ac5d85aSJens Axboe } 12668ac5d85aSJens Axboe 12678ac5d85aSJens Axboe int io_run_local_work(struct io_ring_ctx *ctx) 12688ac5d85aSJens Axboe { 12698ac5d85aSJens Axboe bool locked; 12708ac5d85aSJens Axboe int ret; 12718ac5d85aSJens Axboe 12727924fdfeSPavel Begunkov if (llist_empty(&ctx->work_llist)) 12737924fdfeSPavel Begunkov return 0; 12747924fdfeSPavel Begunkov 1275ec7fd256SJens Axboe __set_current_state(TASK_RUNNING); 12768ac5d85aSJens Axboe locked = mutex_trylock(&ctx->uring_lock); 1277b3026767SDylan Yudaken ret = __io_run_local_work(ctx, &locked); 12788ac5d85aSJens Axboe if (locked) 12798ac5d85aSJens Axboe mutex_unlock(&ctx->uring_lock); 12808ac5d85aSJens Axboe 12818ac5d85aSJens Axboe return ret; 1282c0e0d6baSDylan Yudaken } 1283c0e0d6baSDylan Yudaken 1284ed29b0b4SJens Axboe static void io_req_task_cancel(struct io_kiocb *req, bool *locked) 1285ed29b0b4SJens Axboe { 1286ed29b0b4SJens Axboe io_tw_lock(req->ctx, locked); 1287973fc83fSDylan Yudaken io_req_defer_failed(req, req->cqe.res); 1288ed29b0b4SJens Axboe } 1289ed29b0b4SJens Axboe 1290329061d3SJens Axboe void io_req_task_submit(struct io_kiocb *req, bool *locked) 1291ed29b0b4SJens Axboe { 1292ed29b0b4SJens Axboe io_tw_lock(req->ctx, locked); 1293ed29b0b4SJens Axboe /* req->task == current here, checking PF_EXITING is safe */ 1294ed29b0b4SJens Axboe if (likely(!(req->task->flags & PF_EXITING))) 1295ed29b0b4SJens Axboe io_queue_sqe(req); 1296ed29b0b4SJens Axboe else 1297973fc83fSDylan Yudaken io_req_defer_failed(req, -EFAULT); 1298ed29b0b4SJens Axboe } 1299ed29b0b4SJens Axboe 130059915143SJens Axboe void io_req_task_queue_fail(struct io_kiocb *req, int ret) 1301ed29b0b4SJens Axboe { 130297b388d7SJens Axboe io_req_set_res(req, ret, 0); 1303ed29b0b4SJens Axboe req->io_task_work.func = io_req_task_cancel; 1304ed29b0b4SJens Axboe io_req_task_work_add(req); 1305ed29b0b4SJens Axboe } 1306ed29b0b4SJens Axboe 1307f3b44f92SJens Axboe void io_req_task_queue(struct io_kiocb *req) 1308ed29b0b4SJens Axboe { 1309ed29b0b4SJens Axboe req->io_task_work.func = io_req_task_submit; 1310ed29b0b4SJens Axboe io_req_task_work_add(req); 1311ed29b0b4SJens Axboe } 1312ed29b0b4SJens Axboe 131359915143SJens Axboe void io_queue_next(struct io_kiocb *req) 1314ed29b0b4SJens Axboe { 1315ed29b0b4SJens Axboe struct io_kiocb *nxt = io_req_find_next(req); 1316ed29b0b4SJens Axboe 1317ed29b0b4SJens Axboe if (nxt) 1318ed29b0b4SJens Axboe io_req_task_queue(nxt); 1319ed29b0b4SJens Axboe } 1320ed29b0b4SJens Axboe 1321f3b44f92SJens Axboe void io_free_batch_list(struct io_ring_ctx *ctx, struct io_wq_work_node *node) 1322ed29b0b4SJens Axboe __must_hold(&ctx->uring_lock) 1323ed29b0b4SJens Axboe { 1324ed29b0b4SJens Axboe struct task_struct *task = NULL; 1325ed29b0b4SJens Axboe int task_refs = 0; 1326ed29b0b4SJens Axboe 1327ed29b0b4SJens Axboe do { 1328ed29b0b4SJens Axboe struct io_kiocb *req = container_of(node, struct io_kiocb, 1329ed29b0b4SJens Axboe comp_list); 1330ed29b0b4SJens Axboe 1331ed29b0b4SJens Axboe if (unlikely(req->flags & IO_REQ_CLEAN_SLOW_FLAGS)) { 1332ed29b0b4SJens Axboe if (req->flags & REQ_F_REFCOUNT) { 1333ed29b0b4SJens Axboe node = req->comp_list.next; 1334ed29b0b4SJens Axboe if (!req_ref_put_and_test(req)) 1335ed29b0b4SJens Axboe continue; 1336ed29b0b4SJens Axboe } 1337ed29b0b4SJens Axboe if ((req->flags & REQ_F_POLLED) && req->apoll) { 1338ed29b0b4SJens Axboe struct async_poll *apoll = req->apoll; 1339ed29b0b4SJens Axboe 1340ed29b0b4SJens Axboe if (apoll->double_poll) 1341ed29b0b4SJens Axboe kfree(apoll->double_poll); 13429731bc98SJens Axboe if (!io_alloc_cache_put(&ctx->apoll_cache, &apoll->cache)) 13439731bc98SJens Axboe kfree(apoll); 1344ed29b0b4SJens Axboe req->flags &= ~REQ_F_POLLED; 1345ed29b0b4SJens Axboe } 1346ed29b0b4SJens Axboe if (req->flags & IO_REQ_LINK_FLAGS) 1347ed29b0b4SJens Axboe io_queue_next(req); 1348ed29b0b4SJens Axboe if (unlikely(req->flags & IO_REQ_CLEAN_FLAGS)) 1349ed29b0b4SJens Axboe io_clean_op(req); 1350ed29b0b4SJens Axboe } 1351ed29b0b4SJens Axboe if (!(req->flags & REQ_F_FIXED_FILE)) 1352ed29b0b4SJens Axboe io_put_file(req->file); 1353ed29b0b4SJens Axboe 1354ed29b0b4SJens Axboe io_req_put_rsrc_locked(req, ctx); 1355ed29b0b4SJens Axboe 1356ed29b0b4SJens Axboe if (req->task != task) { 1357ed29b0b4SJens Axboe if (task) 1358ed29b0b4SJens Axboe io_put_task(task, task_refs); 1359ed29b0b4SJens Axboe task = req->task; 1360ed29b0b4SJens Axboe task_refs = 0; 1361ed29b0b4SJens Axboe } 1362ed29b0b4SJens Axboe task_refs++; 1363ed29b0b4SJens Axboe node = req->comp_list.next; 1364ed29b0b4SJens Axboe io_req_add_to_cache(req, ctx); 1365ed29b0b4SJens Axboe } while (node); 1366ed29b0b4SJens Axboe 1367ed29b0b4SJens Axboe if (task) 1368ed29b0b4SJens Axboe io_put_task(task, task_refs); 1369ed29b0b4SJens Axboe } 1370ed29b0b4SJens Axboe 1371ed29b0b4SJens Axboe static void __io_submit_flush_completions(struct io_ring_ctx *ctx) 1372ed29b0b4SJens Axboe __must_hold(&ctx->uring_lock) 1373ed29b0b4SJens Axboe { 1374ed29b0b4SJens Axboe struct io_wq_work_node *node, *prev; 1375ed29b0b4SJens Axboe struct io_submit_state *state = &ctx->submit_state; 1376ed29b0b4SJens Axboe 1377e9a88428SPavel Begunkov io_cq_lock(ctx); 1378931147ddSDylan Yudaken /* must come first to preserve CQE ordering in failure cases */ 1379931147ddSDylan Yudaken if (state->cqes_count) 1380931147ddSDylan Yudaken __io_flush_post_cqes(ctx); 1381ed29b0b4SJens Axboe wq_list_for_each(node, prev, &state->compl_reqs) { 1382ed29b0b4SJens Axboe struct io_kiocb *req = container_of(node, struct io_kiocb, 1383ed29b0b4SJens Axboe comp_list); 1384ed29b0b4SJens Axboe 1385ed29b0b4SJens Axboe if (!(req->flags & REQ_F_CQE_SKIP)) 1386ed29b0b4SJens Axboe __io_fill_cqe_req(ctx, req); 1387ed29b0b4SJens Axboe } 13886c16fe3cSJens Axboe io_cq_unlock_post(ctx); 1389ed29b0b4SJens Axboe 1390931147ddSDylan Yudaken if (!wq_list_empty(&ctx->submit_state.compl_reqs)) { 1391ed29b0b4SJens Axboe io_free_batch_list(ctx, state->compl_reqs.first); 1392ed29b0b4SJens Axboe INIT_WQ_LIST(&state->compl_reqs); 1393ed29b0b4SJens Axboe } 1394931147ddSDylan Yudaken } 1395ed29b0b4SJens Axboe 1396ed29b0b4SJens Axboe /* 1397ed29b0b4SJens Axboe * Drop reference to request, return next in chain (if there is one) if this 1398ed29b0b4SJens Axboe * was the last reference to this request. 1399ed29b0b4SJens Axboe */ 1400ed29b0b4SJens Axboe static inline struct io_kiocb *io_put_req_find_next(struct io_kiocb *req) 1401ed29b0b4SJens Axboe { 1402ed29b0b4SJens Axboe struct io_kiocb *nxt = NULL; 1403ed29b0b4SJens Axboe 1404ed29b0b4SJens Axboe if (req_ref_put_and_test(req)) { 1405ed29b0b4SJens Axboe if (unlikely(req->flags & IO_REQ_LINK_FLAGS)) 1406ed29b0b4SJens Axboe nxt = io_req_find_next(req); 1407ed29b0b4SJens Axboe io_free_req(req); 1408ed29b0b4SJens Axboe } 1409ed29b0b4SJens Axboe return nxt; 1410ed29b0b4SJens Axboe } 1411ed29b0b4SJens Axboe 1412ed29b0b4SJens Axboe static unsigned io_cqring_events(struct io_ring_ctx *ctx) 1413ed29b0b4SJens Axboe { 1414ed29b0b4SJens Axboe /* See comment at the top of this file */ 1415ed29b0b4SJens Axboe smp_rmb(); 1416ed29b0b4SJens Axboe return __io_cqring_events(ctx); 1417ed29b0b4SJens Axboe } 1418ed29b0b4SJens Axboe 1419ed29b0b4SJens Axboe /* 1420ed29b0b4SJens Axboe * We can't just wait for polled events to come to us, we have to actively 1421ed29b0b4SJens Axboe * find and complete them. 1422ed29b0b4SJens Axboe */ 1423ed29b0b4SJens Axboe static __cold void io_iopoll_try_reap_events(struct io_ring_ctx *ctx) 1424ed29b0b4SJens Axboe { 1425ed29b0b4SJens Axboe if (!(ctx->flags & IORING_SETUP_IOPOLL)) 1426ed29b0b4SJens Axboe return; 1427ed29b0b4SJens Axboe 1428ed29b0b4SJens Axboe mutex_lock(&ctx->uring_lock); 1429ed29b0b4SJens Axboe while (!wq_list_empty(&ctx->iopoll_list)) { 1430ed29b0b4SJens Axboe /* let it sleep and repeat later if can't complete a request */ 1431ed29b0b4SJens Axboe if (io_do_iopoll(ctx, true) == 0) 1432ed29b0b4SJens Axboe break; 1433ed29b0b4SJens Axboe /* 1434ed29b0b4SJens Axboe * Ensure we allow local-to-the-cpu processing to take place, 1435ed29b0b4SJens Axboe * in this case we need to ensure that we reap all events. 1436ed29b0b4SJens Axboe * Also let task_work, etc. to progress by releasing the mutex 1437ed29b0b4SJens Axboe */ 1438ed29b0b4SJens Axboe if (need_resched()) { 1439ed29b0b4SJens Axboe mutex_unlock(&ctx->uring_lock); 1440ed29b0b4SJens Axboe cond_resched(); 1441ed29b0b4SJens Axboe mutex_lock(&ctx->uring_lock); 1442ed29b0b4SJens Axboe } 1443ed29b0b4SJens Axboe } 1444ed29b0b4SJens Axboe mutex_unlock(&ctx->uring_lock); 1445ed29b0b4SJens Axboe } 1446ed29b0b4SJens Axboe 1447ed29b0b4SJens Axboe static int io_iopoll_check(struct io_ring_ctx *ctx, long min) 1448ed29b0b4SJens Axboe { 1449ed29b0b4SJens Axboe unsigned int nr_events = 0; 1450ed29b0b4SJens Axboe int ret = 0; 1451ed29b0b4SJens Axboe unsigned long check_cq; 1452ed29b0b4SJens Axboe 145376de6749SPavel Begunkov if (!io_allowed_run_tw(ctx)) 145476de6749SPavel Begunkov return -EEXIST; 145576de6749SPavel Begunkov 14563a08576bSPavel Begunkov check_cq = READ_ONCE(ctx->check_cq); 14573a08576bSPavel Begunkov if (unlikely(check_cq)) { 14583a08576bSPavel Begunkov if (check_cq & BIT(IO_CHECK_CQ_OVERFLOW_BIT)) 14593a08576bSPavel Begunkov __io_cqring_overflow_flush(ctx, false); 14603a08576bSPavel Begunkov /* 14613a08576bSPavel Begunkov * Similarly do not spin if we have not informed the user of any 14623a08576bSPavel Begunkov * dropped CQE. 14633a08576bSPavel Begunkov */ 14643a08576bSPavel Begunkov if (check_cq & BIT(IO_CHECK_CQ_DROPPED_BIT)) 14653a08576bSPavel Begunkov return -EBADR; 14663a08576bSPavel Begunkov } 1467ed29b0b4SJens Axboe /* 1468ed29b0b4SJens Axboe * Don't enter poll loop if we already have events pending. 1469ed29b0b4SJens Axboe * If we do, we can potentially be spinning for commands that 1470ed29b0b4SJens Axboe * already triggered a CQE (eg in error). 1471ed29b0b4SJens Axboe */ 1472ed29b0b4SJens Axboe if (io_cqring_events(ctx)) 1473ed29b0b4SJens Axboe return 0; 1474ed29b0b4SJens Axboe 1475ed29b0b4SJens Axboe do { 1476ed29b0b4SJens Axboe /* 1477ed29b0b4SJens Axboe * If a submit got punted to a workqueue, we can have the 1478ed29b0b4SJens Axboe * application entering polling for a command before it gets 1479ed29b0b4SJens Axboe * issued. That app will hold the uring_lock for the duration 1480ed29b0b4SJens Axboe * of the poll right here, so we need to take a breather every 1481ed29b0b4SJens Axboe * now and then to ensure that the issue has a chance to add 1482ed29b0b4SJens Axboe * the poll to the issued list. Otherwise we can spin here 1483ed29b0b4SJens Axboe * forever, while the workqueue is stuck trying to acquire the 1484ed29b0b4SJens Axboe * very same mutex. 1485ed29b0b4SJens Axboe */ 1486dac6a0eaSJens Axboe if (wq_list_empty(&ctx->iopoll_list) || 1487dac6a0eaSJens Axboe io_task_work_pending(ctx)) { 1488ed29b0b4SJens Axboe u32 tail = ctx->cached_cq_tail; 1489ed29b0b4SJens Axboe 14908de11cdcSDylan Yudaken (void) io_run_local_work_locked(ctx); 14911f8d5bbeSPavel Begunkov 14921f8d5bbeSPavel Begunkov if (task_work_pending(current) || 14931f8d5bbeSPavel Begunkov wq_list_empty(&ctx->iopoll_list)) { 1494ed29b0b4SJens Axboe mutex_unlock(&ctx->uring_lock); 1495ed29b0b4SJens Axboe io_run_task_work(); 1496ed29b0b4SJens Axboe mutex_lock(&ctx->uring_lock); 14971f8d5bbeSPavel Begunkov } 1498ed29b0b4SJens Axboe /* some requests don't go through iopoll_list */ 1499ed29b0b4SJens Axboe if (tail != ctx->cached_cq_tail || 1500ed29b0b4SJens Axboe wq_list_empty(&ctx->iopoll_list)) 1501ed29b0b4SJens Axboe break; 1502ed29b0b4SJens Axboe } 1503ed29b0b4SJens Axboe ret = io_do_iopoll(ctx, !min); 1504ed29b0b4SJens Axboe if (ret < 0) 1505ed29b0b4SJens Axboe break; 1506ed29b0b4SJens Axboe nr_events += ret; 1507ed29b0b4SJens Axboe ret = 0; 1508ed29b0b4SJens Axboe } while (nr_events < min && !need_resched()); 1509ed29b0b4SJens Axboe 1510ed29b0b4SJens Axboe return ret; 1511ed29b0b4SJens Axboe } 15127012c815SPavel Begunkov 15137012c815SPavel Begunkov void io_req_task_complete(struct io_kiocb *req, bool *locked) 1514ed29b0b4SJens Axboe { 15157012c815SPavel Begunkov if (*locked) 15169da070b1SPavel Begunkov io_req_complete_defer(req); 15177012c815SPavel Begunkov else 15181bec951cSPavel Begunkov io_req_complete_post_tw(req, locked); 1519ed29b0b4SJens Axboe } 1520ed29b0b4SJens Axboe 1521ed29b0b4SJens Axboe /* 1522ed29b0b4SJens Axboe * After the iocb has been issued, it's safe to be found on the poll list. 1523ed29b0b4SJens Axboe * Adding the kiocb to the list AFTER submission ensures that we don't 1524ed29b0b4SJens Axboe * find it from a io_do_iopoll() thread before the issuer is done 1525ed29b0b4SJens Axboe * accessing the kiocb cookie. 1526ed29b0b4SJens Axboe */ 1527ed29b0b4SJens Axboe static void io_iopoll_req_issued(struct io_kiocb *req, unsigned int issue_flags) 1528ed29b0b4SJens Axboe { 1529ed29b0b4SJens Axboe struct io_ring_ctx *ctx = req->ctx; 1530ed29b0b4SJens Axboe const bool needs_lock = issue_flags & IO_URING_F_UNLOCKED; 1531ed29b0b4SJens Axboe 1532ed29b0b4SJens Axboe /* workqueue context doesn't hold uring_lock, grab it now */ 1533ed29b0b4SJens Axboe if (unlikely(needs_lock)) 1534ed29b0b4SJens Axboe mutex_lock(&ctx->uring_lock); 1535ed29b0b4SJens Axboe 1536ed29b0b4SJens Axboe /* 1537ed29b0b4SJens Axboe * Track whether we have multiple files in our lists. This will impact 1538ed29b0b4SJens Axboe * how we do polling eventually, not spinning if we're on potentially 1539ed29b0b4SJens Axboe * different devices. 1540ed29b0b4SJens Axboe */ 1541ed29b0b4SJens Axboe if (wq_list_empty(&ctx->iopoll_list)) { 1542ed29b0b4SJens Axboe ctx->poll_multi_queue = false; 1543ed29b0b4SJens Axboe } else if (!ctx->poll_multi_queue) { 1544ed29b0b4SJens Axboe struct io_kiocb *list_req; 1545ed29b0b4SJens Axboe 1546ed29b0b4SJens Axboe list_req = container_of(ctx->iopoll_list.first, struct io_kiocb, 1547ed29b0b4SJens Axboe comp_list); 1548ed29b0b4SJens Axboe if (list_req->file != req->file) 1549ed29b0b4SJens Axboe ctx->poll_multi_queue = true; 1550ed29b0b4SJens Axboe } 1551ed29b0b4SJens Axboe 1552ed29b0b4SJens Axboe /* 1553ed29b0b4SJens Axboe * For fast devices, IO may have already completed. If it has, add 1554ed29b0b4SJens Axboe * it to the front so we find it first. 1555ed29b0b4SJens Axboe */ 1556ed29b0b4SJens Axboe if (READ_ONCE(req->iopoll_completed)) 1557ed29b0b4SJens Axboe wq_list_add_head(&req->comp_list, &ctx->iopoll_list); 1558ed29b0b4SJens Axboe else 1559ed29b0b4SJens Axboe wq_list_add_tail(&req->comp_list, &ctx->iopoll_list); 1560ed29b0b4SJens Axboe 1561ed29b0b4SJens Axboe if (unlikely(needs_lock)) { 1562ed29b0b4SJens Axboe /* 1563ed29b0b4SJens Axboe * If IORING_SETUP_SQPOLL is enabled, sqes are either handle 1564ed29b0b4SJens Axboe * in sq thread task context or in io worker task context. If 1565ed29b0b4SJens Axboe * current task context is sq thread, we don't need to check 1566ed29b0b4SJens Axboe * whether should wake up sq thread. 1567ed29b0b4SJens Axboe */ 1568ed29b0b4SJens Axboe if ((ctx->flags & IORING_SETUP_SQPOLL) && 1569ed29b0b4SJens Axboe wq_has_sleeper(&ctx->sq_data->wait)) 1570ed29b0b4SJens Axboe wake_up(&ctx->sq_data->wait); 1571ed29b0b4SJens Axboe 1572ed29b0b4SJens Axboe mutex_unlock(&ctx->uring_lock); 1573ed29b0b4SJens Axboe } 1574ed29b0b4SJens Axboe } 1575ed29b0b4SJens Axboe 1576ed29b0b4SJens Axboe static bool io_bdev_nowait(struct block_device *bdev) 1577ed29b0b4SJens Axboe { 1578568ec936SChristoph Hellwig return !bdev || bdev_nowait(bdev); 1579ed29b0b4SJens Axboe } 1580ed29b0b4SJens Axboe 1581ed29b0b4SJens Axboe /* 1582ed29b0b4SJens Axboe * If we tracked the file through the SCM inflight mechanism, we could support 1583ed29b0b4SJens Axboe * any file. For now, just ensure that anything potentially problematic is done 1584ed29b0b4SJens Axboe * inline. 1585ed29b0b4SJens Axboe */ 1586ed29b0b4SJens Axboe static bool __io_file_supports_nowait(struct file *file, umode_t mode) 1587ed29b0b4SJens Axboe { 1588ed29b0b4SJens Axboe if (S_ISBLK(mode)) { 1589ed29b0b4SJens Axboe if (IS_ENABLED(CONFIG_BLOCK) && 1590ed29b0b4SJens Axboe io_bdev_nowait(I_BDEV(file->f_mapping->host))) 1591ed29b0b4SJens Axboe return true; 1592ed29b0b4SJens Axboe return false; 1593ed29b0b4SJens Axboe } 1594ed29b0b4SJens Axboe if (S_ISSOCK(mode)) 1595ed29b0b4SJens Axboe return true; 1596ed29b0b4SJens Axboe if (S_ISREG(mode)) { 1597ed29b0b4SJens Axboe if (IS_ENABLED(CONFIG_BLOCK) && 1598ed29b0b4SJens Axboe io_bdev_nowait(file->f_inode->i_sb->s_bdev) && 1599e5550a14SJens Axboe !io_is_uring_fops(file)) 1600ed29b0b4SJens Axboe return true; 1601ed29b0b4SJens Axboe return false; 1602ed29b0b4SJens Axboe } 1603ed29b0b4SJens Axboe 1604ed29b0b4SJens Axboe /* any ->read/write should understand O_NONBLOCK */ 1605ed29b0b4SJens Axboe if (file->f_flags & O_NONBLOCK) 1606ed29b0b4SJens Axboe return true; 1607ed29b0b4SJens Axboe return file->f_mode & FMODE_NOWAIT; 1608ed29b0b4SJens Axboe } 1609ed29b0b4SJens Axboe 1610ed29b0b4SJens Axboe /* 1611ed29b0b4SJens Axboe * If we tracked the file through the SCM inflight mechanism, we could support 1612ed29b0b4SJens Axboe * any file. For now, just ensure that anything potentially problematic is done 1613ed29b0b4SJens Axboe * inline. 1614ed29b0b4SJens Axboe */ 1615a4ad4f74SJens Axboe unsigned int io_file_get_flags(struct file *file) 1616ed29b0b4SJens Axboe { 1617ed29b0b4SJens Axboe umode_t mode = file_inode(file)->i_mode; 1618ed29b0b4SJens Axboe unsigned int res = 0; 1619ed29b0b4SJens Axboe 1620ed29b0b4SJens Axboe if (S_ISREG(mode)) 1621ed29b0b4SJens Axboe res |= FFS_ISREG; 1622ed29b0b4SJens Axboe if (__io_file_supports_nowait(file, mode)) 1623ed29b0b4SJens Axboe res |= FFS_NOWAIT; 1624ed29b0b4SJens Axboe return res; 1625ed29b0b4SJens Axboe } 1626ed29b0b4SJens Axboe 162799f15d8dSJens Axboe bool io_alloc_async_data(struct io_kiocb *req) 1628ed29b0b4SJens Axboe { 1629ed29b0b4SJens Axboe WARN_ON_ONCE(!io_op_defs[req->opcode].async_size); 1630ed29b0b4SJens Axboe req->async_data = kmalloc(io_op_defs[req->opcode].async_size, GFP_KERNEL); 1631ed29b0b4SJens Axboe if (req->async_data) { 1632ed29b0b4SJens Axboe req->flags |= REQ_F_ASYNC_DATA; 1633ed29b0b4SJens Axboe return false; 1634ed29b0b4SJens Axboe } 1635ed29b0b4SJens Axboe return true; 1636ed29b0b4SJens Axboe } 1637ed29b0b4SJens Axboe 1638f3b44f92SJens Axboe int io_req_prep_async(struct io_kiocb *req) 1639ed29b0b4SJens Axboe { 1640ed29b0b4SJens Axboe const struct io_op_def *def = &io_op_defs[req->opcode]; 1641ed29b0b4SJens Axboe 1642ed29b0b4SJens Axboe /* assign early for deferred execution for non-fixed file */ 1643ed29b0b4SJens Axboe if (def->needs_file && !(req->flags & REQ_F_FIXED_FILE)) 1644ed29b0b4SJens Axboe req->file = io_file_get_normal(req, req->cqe.fd); 1645dc919cafSJens Axboe if (!def->prep_async) 1646ed29b0b4SJens Axboe return 0; 1647ed29b0b4SJens Axboe if (WARN_ON_ONCE(req_has_async_data(req))) 1648ed29b0b4SJens Axboe return -EFAULT; 164959169439SPavel Begunkov if (!io_op_defs[req->opcode].manual_alloc) { 1650ed29b0b4SJens Axboe if (io_alloc_async_data(req)) 1651ed29b0b4SJens Axboe return -EAGAIN; 165259169439SPavel Begunkov } 1653dc919cafSJens Axboe return def->prep_async(req); 1654ed29b0b4SJens Axboe } 1655ed29b0b4SJens Axboe 1656ed29b0b4SJens Axboe static u32 io_get_sequence(struct io_kiocb *req) 1657ed29b0b4SJens Axboe { 1658ed29b0b4SJens Axboe u32 seq = req->ctx->cached_sq_head; 1659ed29b0b4SJens Axboe struct io_kiocb *cur; 1660ed29b0b4SJens Axboe 1661ed29b0b4SJens Axboe /* need original cached_sq_head, but it was increased for each req */ 1662ed29b0b4SJens Axboe io_for_each_link(cur, req) 1663ed29b0b4SJens Axboe seq--; 1664ed29b0b4SJens Axboe return seq; 1665ed29b0b4SJens Axboe } 1666ed29b0b4SJens Axboe 1667ed29b0b4SJens Axboe static __cold void io_drain_req(struct io_kiocb *req) 1668e276ae34SPavel Begunkov __must_hold(&ctx->uring_lock) 1669ed29b0b4SJens Axboe { 1670ed29b0b4SJens Axboe struct io_ring_ctx *ctx = req->ctx; 1671ed29b0b4SJens Axboe struct io_defer_entry *de; 1672ed29b0b4SJens Axboe int ret; 1673ed29b0b4SJens Axboe u32 seq = io_get_sequence(req); 1674ed29b0b4SJens Axboe 1675ed29b0b4SJens Axboe /* Still need defer if there is pending req in defer list. */ 1676ed29b0b4SJens Axboe spin_lock(&ctx->completion_lock); 1677ed29b0b4SJens Axboe if (!req_need_defer(req, seq) && list_empty_careful(&ctx->defer_list)) { 1678ed29b0b4SJens Axboe spin_unlock(&ctx->completion_lock); 1679ed29b0b4SJens Axboe queue: 1680ed29b0b4SJens Axboe ctx->drain_active = false; 1681ed29b0b4SJens Axboe io_req_task_queue(req); 1682ed29b0b4SJens Axboe return; 1683ed29b0b4SJens Axboe } 1684ed29b0b4SJens Axboe spin_unlock(&ctx->completion_lock); 1685ed29b0b4SJens Axboe 1686ed29b0b4SJens Axboe ret = io_req_prep_async(req); 1687ed29b0b4SJens Axboe if (ret) { 1688ed29b0b4SJens Axboe fail: 1689973fc83fSDylan Yudaken io_req_defer_failed(req, ret); 1690ed29b0b4SJens Axboe return; 1691ed29b0b4SJens Axboe } 1692ed29b0b4SJens Axboe io_prep_async_link(req); 1693ed29b0b4SJens Axboe de = kmalloc(sizeof(*de), GFP_KERNEL); 1694ed29b0b4SJens Axboe if (!de) { 1695ed29b0b4SJens Axboe ret = -ENOMEM; 1696ed29b0b4SJens Axboe goto fail; 1697ed29b0b4SJens Axboe } 1698ed29b0b4SJens Axboe 1699ed29b0b4SJens Axboe spin_lock(&ctx->completion_lock); 1700ed29b0b4SJens Axboe if (!req_need_defer(req, seq) && list_empty(&ctx->defer_list)) { 1701ed29b0b4SJens Axboe spin_unlock(&ctx->completion_lock); 1702ed29b0b4SJens Axboe kfree(de); 1703ed29b0b4SJens Axboe goto queue; 1704ed29b0b4SJens Axboe } 1705ed29b0b4SJens Axboe 170648863ffdSPavel Begunkov trace_io_uring_defer(req); 1707ed29b0b4SJens Axboe de->req = req; 1708ed29b0b4SJens Axboe de->seq = seq; 1709ed29b0b4SJens Axboe list_add_tail(&de->list, &ctx->defer_list); 1710ed29b0b4SJens Axboe spin_unlock(&ctx->completion_lock); 1711ed29b0b4SJens Axboe } 1712ed29b0b4SJens Axboe 1713ed29b0b4SJens Axboe static void io_clean_op(struct io_kiocb *req) 1714ed29b0b4SJens Axboe { 1715ed29b0b4SJens Axboe if (req->flags & REQ_F_BUFFER_SELECTED) { 1716ed29b0b4SJens Axboe spin_lock(&req->ctx->completion_lock); 1717ed29b0b4SJens Axboe io_put_kbuf_comp(req); 1718ed29b0b4SJens Axboe spin_unlock(&req->ctx->completion_lock); 1719ed29b0b4SJens Axboe } 1720ed29b0b4SJens Axboe 1721ed29b0b4SJens Axboe if (req->flags & REQ_F_NEED_CLEANUP) { 17224d4c9cffSJens Axboe const struct io_op_def *def = &io_op_defs[req->opcode]; 1723ed29b0b4SJens Axboe 17244d4c9cffSJens Axboe if (def->cleanup) 17254d4c9cffSJens Axboe def->cleanup(req); 1726ed29b0b4SJens Axboe } 1727ed29b0b4SJens Axboe if ((req->flags & REQ_F_POLLED) && req->apoll) { 1728ed29b0b4SJens Axboe kfree(req->apoll->double_poll); 1729ed29b0b4SJens Axboe kfree(req->apoll); 1730ed29b0b4SJens Axboe req->apoll = NULL; 1731ed29b0b4SJens Axboe } 1732ed29b0b4SJens Axboe if (req->flags & REQ_F_INFLIGHT) { 1733ed29b0b4SJens Axboe struct io_uring_task *tctx = req->task->io_uring; 1734ed29b0b4SJens Axboe 1735ed29b0b4SJens Axboe atomic_dec(&tctx->inflight_tracked); 1736ed29b0b4SJens Axboe } 1737ed29b0b4SJens Axboe if (req->flags & REQ_F_CREDS) 1738ed29b0b4SJens Axboe put_cred(req->creds); 1739ed29b0b4SJens Axboe if (req->flags & REQ_F_ASYNC_DATA) { 1740ed29b0b4SJens Axboe kfree(req->async_data); 1741ed29b0b4SJens Axboe req->async_data = NULL; 1742ed29b0b4SJens Axboe } 1743ed29b0b4SJens Axboe req->flags &= ~IO_REQ_CLEAN_FLAGS; 1744ed29b0b4SJens Axboe } 1745ed29b0b4SJens Axboe 1746ed29b0b4SJens Axboe static bool io_assign_file(struct io_kiocb *req, unsigned int issue_flags) 1747ed29b0b4SJens Axboe { 1748ed29b0b4SJens Axboe if (req->file || !io_op_defs[req->opcode].needs_file) 1749ed29b0b4SJens Axboe return true; 1750ed29b0b4SJens Axboe 1751ed29b0b4SJens Axboe if (req->flags & REQ_F_FIXED_FILE) 1752ed29b0b4SJens Axboe req->file = io_file_get_fixed(req, req->cqe.fd, issue_flags); 1753ed29b0b4SJens Axboe else 1754ed29b0b4SJens Axboe req->file = io_file_get_normal(req, req->cqe.fd); 1755ed29b0b4SJens Axboe 1756ed29b0b4SJens Axboe return !!req->file; 1757ed29b0b4SJens Axboe } 1758ed29b0b4SJens Axboe 1759ed29b0b4SJens Axboe static int io_issue_sqe(struct io_kiocb *req, unsigned int issue_flags) 1760ed29b0b4SJens Axboe { 1761ed29b0b4SJens Axboe const struct io_op_def *def = &io_op_defs[req->opcode]; 1762ed29b0b4SJens Axboe const struct cred *creds = NULL; 1763ed29b0b4SJens Axboe int ret; 1764ed29b0b4SJens Axboe 1765ed29b0b4SJens Axboe if (unlikely(!io_assign_file(req, issue_flags))) 1766ed29b0b4SJens Axboe return -EBADF; 1767ed29b0b4SJens Axboe 1768ed29b0b4SJens Axboe if (unlikely((req->flags & REQ_F_CREDS) && req->creds != current_cred())) 1769ed29b0b4SJens Axboe creds = override_creds(req->creds); 1770ed29b0b4SJens Axboe 1771ed29b0b4SJens Axboe if (!def->audit_skip) 1772ed29b0b4SJens Axboe audit_uring_entry(req->opcode); 1773ed29b0b4SJens Axboe 1774ed29b0b4SJens Axboe ret = def->issue(req, issue_flags); 1775ed29b0b4SJens Axboe 1776ed29b0b4SJens Axboe if (!def->audit_skip) 1777ed29b0b4SJens Axboe audit_uring_exit(!ret, ret); 1778ed29b0b4SJens Axboe 1779ed29b0b4SJens Axboe if (creds) 1780ed29b0b4SJens Axboe revert_creds(creds); 178197b388d7SJens Axboe 178275d7b3aeSPavel Begunkov if (ret == IOU_OK) { 178375d7b3aeSPavel Begunkov if (issue_flags & IO_URING_F_COMPLETE_DEFER) 17849da070b1SPavel Begunkov io_req_complete_defer(req); 178575d7b3aeSPavel Begunkov else 17861bec951cSPavel Begunkov io_req_complete_post(req, issue_flags); 178775d7b3aeSPavel Begunkov } else if (ret != IOU_ISSUE_SKIP_COMPLETE) 1788ed29b0b4SJens Axboe return ret; 178997b388d7SJens Axboe 1790ed29b0b4SJens Axboe /* If the op doesn't have a file, we're not polling for it */ 1791ed29b0b4SJens Axboe if ((req->ctx->flags & IORING_SETUP_IOPOLL) && req->file) 1792ed29b0b4SJens Axboe io_iopoll_req_issued(req, issue_flags); 1793ed29b0b4SJens Axboe 1794ed29b0b4SJens Axboe return 0; 1795ed29b0b4SJens Axboe } 1796ed29b0b4SJens Axboe 1797329061d3SJens Axboe int io_poll_issue(struct io_kiocb *req, bool *locked) 1798329061d3SJens Axboe { 1799329061d3SJens Axboe io_tw_lock(req->ctx, locked); 1800329061d3SJens Axboe if (unlikely(req->task->flags & PF_EXITING)) 1801329061d3SJens Axboe return -EFAULT; 180291482864SPavel Begunkov return io_issue_sqe(req, IO_URING_F_NONBLOCK|IO_URING_F_MULTISHOT); 1803329061d3SJens Axboe } 1804329061d3SJens Axboe 1805c9f06aa7SJens Axboe struct io_wq_work *io_wq_free_work(struct io_wq_work *work) 1806ed29b0b4SJens Axboe { 1807ed29b0b4SJens Axboe struct io_kiocb *req = container_of(work, struct io_kiocb, work); 1808ed29b0b4SJens Axboe 1809ed29b0b4SJens Axboe req = io_put_req_find_next(req); 1810ed29b0b4SJens Axboe return req ? &req->work : NULL; 1811ed29b0b4SJens Axboe } 1812ed29b0b4SJens Axboe 1813c9f06aa7SJens Axboe void io_wq_submit_work(struct io_wq_work *work) 1814ed29b0b4SJens Axboe { 1815ed29b0b4SJens Axboe struct io_kiocb *req = container_of(work, struct io_kiocb, work); 1816ed29b0b4SJens Axboe const struct io_op_def *def = &io_op_defs[req->opcode]; 1817ed29b0b4SJens Axboe unsigned int issue_flags = IO_URING_F_UNLOCKED; 1818ed29b0b4SJens Axboe bool needs_poll = false; 1819ed29b0b4SJens Axboe int ret = 0, err = -ECANCELED; 1820ed29b0b4SJens Axboe 182123a6c9acSLin Ma /* one will be dropped by ->io_wq_free_work() after returning to io-wq */ 1822ed29b0b4SJens Axboe if (!(req->flags & REQ_F_REFCOUNT)) 1823ed29b0b4SJens Axboe __io_req_set_refcount(req, 2); 1824ed29b0b4SJens Axboe else 1825ed29b0b4SJens Axboe req_ref_get(req); 1826ed29b0b4SJens Axboe 1827ed29b0b4SJens Axboe io_arm_ltimeout(req); 1828ed29b0b4SJens Axboe 1829ed29b0b4SJens Axboe /* either cancelled or io-wq is dying, so don't touch tctx->iowq */ 1830ed29b0b4SJens Axboe if (work->flags & IO_WQ_WORK_CANCEL) { 1831ed29b0b4SJens Axboe fail: 1832ed29b0b4SJens Axboe io_req_task_queue_fail(req, err); 1833ed29b0b4SJens Axboe return; 1834ed29b0b4SJens Axboe } 1835ed29b0b4SJens Axboe if (!io_assign_file(req, issue_flags)) { 1836ed29b0b4SJens Axboe err = -EBADF; 1837ed29b0b4SJens Axboe work->flags |= IO_WQ_WORK_CANCEL; 1838ed29b0b4SJens Axboe goto fail; 1839ed29b0b4SJens Axboe } 1840ed29b0b4SJens Axboe 1841ed29b0b4SJens Axboe if (req->flags & REQ_F_FORCE_ASYNC) { 1842ed29b0b4SJens Axboe bool opcode_poll = def->pollin || def->pollout; 1843ed29b0b4SJens Axboe 1844ed29b0b4SJens Axboe if (opcode_poll && file_can_poll(req->file)) { 1845ed29b0b4SJens Axboe needs_poll = true; 1846ed29b0b4SJens Axboe issue_flags |= IO_URING_F_NONBLOCK; 1847ed29b0b4SJens Axboe } 1848ed29b0b4SJens Axboe } 1849ed29b0b4SJens Axboe 1850ed29b0b4SJens Axboe do { 1851ed29b0b4SJens Axboe ret = io_issue_sqe(req, issue_flags); 1852ed29b0b4SJens Axboe if (ret != -EAGAIN) 1853ed29b0b4SJens Axboe break; 1854ed29b0b4SJens Axboe /* 1855ed29b0b4SJens Axboe * We can get EAGAIN for iopolled IO even though we're 1856ed29b0b4SJens Axboe * forcing a sync submission from here, since we can't 1857ed29b0b4SJens Axboe * wait for request slots on the block side. 1858ed29b0b4SJens Axboe */ 1859ed29b0b4SJens Axboe if (!needs_poll) { 1860ed29b0b4SJens Axboe if (!(req->ctx->flags & IORING_SETUP_IOPOLL)) 1861ed29b0b4SJens Axboe break; 1862ed29b0b4SJens Axboe cond_resched(); 1863ed29b0b4SJens Axboe continue; 1864ed29b0b4SJens Axboe } 1865ed29b0b4SJens Axboe 1866ed29b0b4SJens Axboe if (io_arm_poll_handler(req, issue_flags) == IO_APOLL_OK) 1867ed29b0b4SJens Axboe return; 1868ed29b0b4SJens Axboe /* aborted or ready, in either case retry blocking */ 1869ed29b0b4SJens Axboe needs_poll = false; 1870ed29b0b4SJens Axboe issue_flags &= ~IO_URING_F_NONBLOCK; 1871ed29b0b4SJens Axboe } while (1); 1872ed29b0b4SJens Axboe 1873ed29b0b4SJens Axboe /* avoid locking problems by failing it from a clean context */ 187497b388d7SJens Axboe if (ret < 0) 1875ed29b0b4SJens Axboe io_req_task_queue_fail(req, ret); 1876ed29b0b4SJens Axboe } 1877ed29b0b4SJens Axboe 1878531113bbSJens Axboe inline struct file *io_file_get_fixed(struct io_kiocb *req, int fd, 1879ed29b0b4SJens Axboe unsigned int issue_flags) 1880ed29b0b4SJens Axboe { 1881ed29b0b4SJens Axboe struct io_ring_ctx *ctx = req->ctx; 1882ed29b0b4SJens Axboe struct file *file = NULL; 1883ed29b0b4SJens Axboe unsigned long file_ptr; 1884ed29b0b4SJens Axboe 1885ed29b0b4SJens Axboe io_ring_submit_lock(ctx, issue_flags); 1886ed29b0b4SJens Axboe 1887ed29b0b4SJens Axboe if (unlikely((unsigned int)fd >= ctx->nr_user_files)) 1888ed29b0b4SJens Axboe goto out; 1889ed29b0b4SJens Axboe fd = array_index_nospec(fd, ctx->nr_user_files); 1890ed29b0b4SJens Axboe file_ptr = io_fixed_file_slot(&ctx->file_table, fd)->file_ptr; 1891ed29b0b4SJens Axboe file = (struct file *) (file_ptr & FFS_MASK); 1892ed29b0b4SJens Axboe file_ptr &= ~FFS_MASK; 1893ed29b0b4SJens Axboe /* mask in overlapping REQ_F and FFS bits */ 1894ed29b0b4SJens Axboe req->flags |= (file_ptr << REQ_F_SUPPORT_NOWAIT_BIT); 1895ed29b0b4SJens Axboe io_req_set_rsrc_node(req, ctx, 0); 1896ed29b0b4SJens Axboe out: 1897ed29b0b4SJens Axboe io_ring_submit_unlock(ctx, issue_flags); 1898ed29b0b4SJens Axboe return file; 1899ed29b0b4SJens Axboe } 1900ed29b0b4SJens Axboe 1901531113bbSJens Axboe struct file *io_file_get_normal(struct io_kiocb *req, int fd) 1902ed29b0b4SJens Axboe { 1903ed29b0b4SJens Axboe struct file *file = fget(fd); 1904ed29b0b4SJens Axboe 190548863ffdSPavel Begunkov trace_io_uring_file_get(req, fd); 1906ed29b0b4SJens Axboe 1907ed29b0b4SJens Axboe /* we don't allow fixed io_uring files */ 1908e5550a14SJens Axboe if (file && io_is_uring_fops(file)) 1909ed29b0b4SJens Axboe io_req_track_inflight(req); 1910ed29b0b4SJens Axboe return file; 1911ed29b0b4SJens Axboe } 1912ed29b0b4SJens Axboe 1913ed29b0b4SJens Axboe static void io_queue_async(struct io_kiocb *req, int ret) 1914ed29b0b4SJens Axboe __must_hold(&req->ctx->uring_lock) 1915ed29b0b4SJens Axboe { 1916ed29b0b4SJens Axboe struct io_kiocb *linked_timeout; 1917ed29b0b4SJens Axboe 1918ed29b0b4SJens Axboe if (ret != -EAGAIN || (req->flags & REQ_F_NOWAIT)) { 1919973fc83fSDylan Yudaken io_req_defer_failed(req, ret); 1920ed29b0b4SJens Axboe return; 1921ed29b0b4SJens Axboe } 1922ed29b0b4SJens Axboe 1923ed29b0b4SJens Axboe linked_timeout = io_prep_linked_timeout(req); 1924ed29b0b4SJens Axboe 1925ed29b0b4SJens Axboe switch (io_arm_poll_handler(req, 0)) { 1926ed29b0b4SJens Axboe case IO_APOLL_READY: 1927336d28a8SPavel Begunkov io_kbuf_recycle(req, 0); 1928ed29b0b4SJens Axboe io_req_task_queue(req); 1929ed29b0b4SJens Axboe break; 1930ed29b0b4SJens Axboe case IO_APOLL_ABORTED: 1931ed29b0b4SJens Axboe io_kbuf_recycle(req, 0); 1932ed29b0b4SJens Axboe io_queue_iowq(req, NULL); 1933ed29b0b4SJens Axboe break; 1934ed29b0b4SJens Axboe case IO_APOLL_OK: 1935ed29b0b4SJens Axboe break; 1936ed29b0b4SJens Axboe } 1937ed29b0b4SJens Axboe 1938ed29b0b4SJens Axboe if (linked_timeout) 1939ed29b0b4SJens Axboe io_queue_linked_timeout(linked_timeout); 1940ed29b0b4SJens Axboe } 1941ed29b0b4SJens Axboe 1942ed29b0b4SJens Axboe static inline void io_queue_sqe(struct io_kiocb *req) 1943ed29b0b4SJens Axboe __must_hold(&req->ctx->uring_lock) 1944ed29b0b4SJens Axboe { 1945ed29b0b4SJens Axboe int ret; 1946ed29b0b4SJens Axboe 1947ed29b0b4SJens Axboe ret = io_issue_sqe(req, IO_URING_F_NONBLOCK|IO_URING_F_COMPLETE_DEFER); 1948ed29b0b4SJens Axboe 1949ed29b0b4SJens Axboe /* 1950ed29b0b4SJens Axboe * We async punt it if the file wasn't marked NOWAIT, or if the file 1951ed29b0b4SJens Axboe * doesn't support non-blocking read/write attempts 1952ed29b0b4SJens Axboe */ 1953ed29b0b4SJens Axboe if (likely(!ret)) 1954ed29b0b4SJens Axboe io_arm_ltimeout(req); 1955ed29b0b4SJens Axboe else 1956ed29b0b4SJens Axboe io_queue_async(req, ret); 1957ed29b0b4SJens Axboe } 1958ed29b0b4SJens Axboe 1959ed29b0b4SJens Axboe static void io_queue_sqe_fallback(struct io_kiocb *req) 1960ed29b0b4SJens Axboe __must_hold(&req->ctx->uring_lock) 1961ed29b0b4SJens Axboe { 1962ed29b0b4SJens Axboe if (unlikely(req->flags & REQ_F_FAIL)) { 1963ed29b0b4SJens Axboe /* 1964ed29b0b4SJens Axboe * We don't submit, fail them all, for that replace hardlinks 1965ed29b0b4SJens Axboe * with normal links. Extra REQ_F_LINK is tolerated. 1966ed29b0b4SJens Axboe */ 1967ed29b0b4SJens Axboe req->flags &= ~REQ_F_HARDLINK; 1968ed29b0b4SJens Axboe req->flags |= REQ_F_LINK; 1969973fc83fSDylan Yudaken io_req_defer_failed(req, req->cqe.res); 1970ed29b0b4SJens Axboe } else if (unlikely(req->ctx->drain_active)) { 1971ed29b0b4SJens Axboe io_drain_req(req); 1972ed29b0b4SJens Axboe } else { 1973ed29b0b4SJens Axboe int ret = io_req_prep_async(req); 1974ed29b0b4SJens Axboe 1975ed29b0b4SJens Axboe if (unlikely(ret)) 1976973fc83fSDylan Yudaken io_req_defer_failed(req, ret); 1977ed29b0b4SJens Axboe else 1978ed29b0b4SJens Axboe io_queue_iowq(req, NULL); 1979ed29b0b4SJens Axboe } 1980ed29b0b4SJens Axboe } 1981ed29b0b4SJens Axboe 1982ed29b0b4SJens Axboe /* 1983ed29b0b4SJens Axboe * Check SQE restrictions (opcode and flags). 1984ed29b0b4SJens Axboe * 1985ed29b0b4SJens Axboe * Returns 'true' if SQE is allowed, 'false' otherwise. 1986ed29b0b4SJens Axboe */ 1987ed29b0b4SJens Axboe static inline bool io_check_restriction(struct io_ring_ctx *ctx, 1988ed29b0b4SJens Axboe struct io_kiocb *req, 1989ed29b0b4SJens Axboe unsigned int sqe_flags) 1990ed29b0b4SJens Axboe { 1991ed29b0b4SJens Axboe if (!test_bit(req->opcode, ctx->restrictions.sqe_op)) 1992ed29b0b4SJens Axboe return false; 1993ed29b0b4SJens Axboe 1994ed29b0b4SJens Axboe if ((sqe_flags & ctx->restrictions.sqe_flags_required) != 1995ed29b0b4SJens Axboe ctx->restrictions.sqe_flags_required) 1996ed29b0b4SJens Axboe return false; 1997ed29b0b4SJens Axboe 1998ed29b0b4SJens Axboe if (sqe_flags & ~(ctx->restrictions.sqe_flags_allowed | 1999ed29b0b4SJens Axboe ctx->restrictions.sqe_flags_required)) 2000ed29b0b4SJens Axboe return false; 2001ed29b0b4SJens Axboe 2002ed29b0b4SJens Axboe return true; 2003ed29b0b4SJens Axboe } 2004ed29b0b4SJens Axboe 2005ed29b0b4SJens Axboe static void io_init_req_drain(struct io_kiocb *req) 2006ed29b0b4SJens Axboe { 2007ed29b0b4SJens Axboe struct io_ring_ctx *ctx = req->ctx; 2008ed29b0b4SJens Axboe struct io_kiocb *head = ctx->submit_state.link.head; 2009ed29b0b4SJens Axboe 2010ed29b0b4SJens Axboe ctx->drain_active = true; 2011ed29b0b4SJens Axboe if (head) { 2012ed29b0b4SJens Axboe /* 2013ed29b0b4SJens Axboe * If we need to drain a request in the middle of a link, drain 2014ed29b0b4SJens Axboe * the head request and the next request/link after the current 2015ed29b0b4SJens Axboe * link. Considering sequential execution of links, 2016ed29b0b4SJens Axboe * REQ_F_IO_DRAIN will be maintained for every request of our 2017ed29b0b4SJens Axboe * link. 2018ed29b0b4SJens Axboe */ 2019ed29b0b4SJens Axboe head->flags |= REQ_F_IO_DRAIN | REQ_F_FORCE_ASYNC; 2020ed29b0b4SJens Axboe ctx->drain_next = true; 2021ed29b0b4SJens Axboe } 2022ed29b0b4SJens Axboe } 2023ed29b0b4SJens Axboe 2024ed29b0b4SJens Axboe static int io_init_req(struct io_ring_ctx *ctx, struct io_kiocb *req, 2025ed29b0b4SJens Axboe const struct io_uring_sqe *sqe) 2026ed29b0b4SJens Axboe __must_hold(&ctx->uring_lock) 2027ed29b0b4SJens Axboe { 2028ed29b0b4SJens Axboe const struct io_op_def *def; 2029ed29b0b4SJens Axboe unsigned int sqe_flags; 2030ed29b0b4SJens Axboe int personality; 2031ed29b0b4SJens Axboe u8 opcode; 2032ed29b0b4SJens Axboe 2033ed29b0b4SJens Axboe /* req is partially pre-initialised, see io_preinit_req() */ 2034ed29b0b4SJens Axboe req->opcode = opcode = READ_ONCE(sqe->opcode); 2035ed29b0b4SJens Axboe /* same numerical values with corresponding REQ_F_*, safe to copy */ 2036ed29b0b4SJens Axboe req->flags = sqe_flags = READ_ONCE(sqe->flags); 2037ed29b0b4SJens Axboe req->cqe.user_data = READ_ONCE(sqe->user_data); 2038ed29b0b4SJens Axboe req->file = NULL; 2039ed29b0b4SJens Axboe req->rsrc_node = NULL; 2040ed29b0b4SJens Axboe req->task = current; 2041ed29b0b4SJens Axboe 2042ed29b0b4SJens Axboe if (unlikely(opcode >= IORING_OP_LAST)) { 2043ed29b0b4SJens Axboe req->opcode = 0; 2044ed29b0b4SJens Axboe return -EINVAL; 2045ed29b0b4SJens Axboe } 2046ed29b0b4SJens Axboe def = &io_op_defs[opcode]; 2047ed29b0b4SJens Axboe if (unlikely(sqe_flags & ~SQE_COMMON_FLAGS)) { 2048ed29b0b4SJens Axboe /* enforce forwards compatibility on users */ 2049ed29b0b4SJens Axboe if (sqe_flags & ~SQE_VALID_FLAGS) 2050ed29b0b4SJens Axboe return -EINVAL; 2051ed29b0b4SJens Axboe if (sqe_flags & IOSQE_BUFFER_SELECT) { 2052ed29b0b4SJens Axboe if (!def->buffer_select) 2053ed29b0b4SJens Axboe return -EOPNOTSUPP; 2054ed29b0b4SJens Axboe req->buf_index = READ_ONCE(sqe->buf_group); 2055ed29b0b4SJens Axboe } 2056ed29b0b4SJens Axboe if (sqe_flags & IOSQE_CQE_SKIP_SUCCESS) 2057ed29b0b4SJens Axboe ctx->drain_disabled = true; 2058ed29b0b4SJens Axboe if (sqe_flags & IOSQE_IO_DRAIN) { 2059ed29b0b4SJens Axboe if (ctx->drain_disabled) 2060ed29b0b4SJens Axboe return -EOPNOTSUPP; 2061ed29b0b4SJens Axboe io_init_req_drain(req); 2062ed29b0b4SJens Axboe } 2063ed29b0b4SJens Axboe } 2064ed29b0b4SJens Axboe if (unlikely(ctx->restricted || ctx->drain_active || ctx->drain_next)) { 2065ed29b0b4SJens Axboe if (ctx->restricted && !io_check_restriction(ctx, req, sqe_flags)) 2066ed29b0b4SJens Axboe return -EACCES; 2067ed29b0b4SJens Axboe /* knock it to the slow queue path, will be drained there */ 2068ed29b0b4SJens Axboe if (ctx->drain_active) 2069ed29b0b4SJens Axboe req->flags |= REQ_F_FORCE_ASYNC; 2070ed29b0b4SJens Axboe /* if there is no link, we're at "next" request and need to drain */ 2071ed29b0b4SJens Axboe if (unlikely(ctx->drain_next) && !ctx->submit_state.link.head) { 2072ed29b0b4SJens Axboe ctx->drain_next = false; 2073ed29b0b4SJens Axboe ctx->drain_active = true; 2074ed29b0b4SJens Axboe req->flags |= REQ_F_IO_DRAIN | REQ_F_FORCE_ASYNC; 2075ed29b0b4SJens Axboe } 2076ed29b0b4SJens Axboe } 2077ed29b0b4SJens Axboe 2078ed29b0b4SJens Axboe if (!def->ioprio && sqe->ioprio) 2079ed29b0b4SJens Axboe return -EINVAL; 2080ed29b0b4SJens Axboe if (!def->iopoll && (ctx->flags & IORING_SETUP_IOPOLL)) 2081ed29b0b4SJens Axboe return -EINVAL; 2082ed29b0b4SJens Axboe 2083ed29b0b4SJens Axboe if (def->needs_file) { 2084ed29b0b4SJens Axboe struct io_submit_state *state = &ctx->submit_state; 2085ed29b0b4SJens Axboe 2086ed29b0b4SJens Axboe req->cqe.fd = READ_ONCE(sqe->fd); 2087ed29b0b4SJens Axboe 2088ed29b0b4SJens Axboe /* 2089ed29b0b4SJens Axboe * Plug now if we have more than 2 IO left after this, and the 2090ed29b0b4SJens Axboe * target is potentially a read/write to block based storage. 2091ed29b0b4SJens Axboe */ 2092ed29b0b4SJens Axboe if (state->need_plug && def->plug) { 2093ed29b0b4SJens Axboe state->plug_started = true; 2094ed29b0b4SJens Axboe state->need_plug = false; 2095ed29b0b4SJens Axboe blk_start_plug_nr_ios(&state->plug, state->submit_nr); 2096ed29b0b4SJens Axboe } 2097ed29b0b4SJens Axboe } 2098ed29b0b4SJens Axboe 2099ed29b0b4SJens Axboe personality = READ_ONCE(sqe->personality); 2100ed29b0b4SJens Axboe if (personality) { 2101ed29b0b4SJens Axboe int ret; 2102ed29b0b4SJens Axboe 2103ed29b0b4SJens Axboe req->creds = xa_load(&ctx->personalities, personality); 2104ed29b0b4SJens Axboe if (!req->creds) 2105ed29b0b4SJens Axboe return -EINVAL; 2106ed29b0b4SJens Axboe get_cred(req->creds); 2107ed29b0b4SJens Axboe ret = security_uring_override_creds(req->creds); 2108ed29b0b4SJens Axboe if (ret) { 2109ed29b0b4SJens Axboe put_cred(req->creds); 2110ed29b0b4SJens Axboe return ret; 2111ed29b0b4SJens Axboe } 2112ed29b0b4SJens Axboe req->flags |= REQ_F_CREDS; 2113ed29b0b4SJens Axboe } 2114ed29b0b4SJens Axboe 2115ed29b0b4SJens Axboe return def->prep(req, sqe); 2116ed29b0b4SJens Axboe } 2117ed29b0b4SJens Axboe 2118ed29b0b4SJens Axboe static __cold int io_submit_fail_init(const struct io_uring_sqe *sqe, 2119ed29b0b4SJens Axboe struct io_kiocb *req, int ret) 2120ed29b0b4SJens Axboe { 2121ed29b0b4SJens Axboe struct io_ring_ctx *ctx = req->ctx; 2122ed29b0b4SJens Axboe struct io_submit_link *link = &ctx->submit_state.link; 2123ed29b0b4SJens Axboe struct io_kiocb *head = link->head; 2124ed29b0b4SJens Axboe 212548863ffdSPavel Begunkov trace_io_uring_req_failed(sqe, req, ret); 2126ed29b0b4SJens Axboe 2127ed29b0b4SJens Axboe /* 2128ed29b0b4SJens Axboe * Avoid breaking links in the middle as it renders links with SQPOLL 2129ed29b0b4SJens Axboe * unusable. Instead of failing eagerly, continue assembling the link if 2130ed29b0b4SJens Axboe * applicable and mark the head with REQ_F_FAIL. The link flushing code 2131ed29b0b4SJens Axboe * should find the flag and handle the rest. 2132ed29b0b4SJens Axboe */ 2133ed29b0b4SJens Axboe req_fail_link_node(req, ret); 2134ed29b0b4SJens Axboe if (head && !(head->flags & REQ_F_FAIL)) 2135ed29b0b4SJens Axboe req_fail_link_node(head, -ECANCELED); 2136ed29b0b4SJens Axboe 2137ed29b0b4SJens Axboe if (!(req->flags & IO_REQ_LINK_FLAGS)) { 2138ed29b0b4SJens Axboe if (head) { 2139ed29b0b4SJens Axboe link->last->link = req; 2140ed29b0b4SJens Axboe link->head = NULL; 2141ed29b0b4SJens Axboe req = head; 2142ed29b0b4SJens Axboe } 2143ed29b0b4SJens Axboe io_queue_sqe_fallback(req); 2144ed29b0b4SJens Axboe return ret; 2145ed29b0b4SJens Axboe } 2146ed29b0b4SJens Axboe 2147ed29b0b4SJens Axboe if (head) 2148ed29b0b4SJens Axboe link->last->link = req; 2149ed29b0b4SJens Axboe else 2150ed29b0b4SJens Axboe link->head = req; 2151ed29b0b4SJens Axboe link->last = req; 2152ed29b0b4SJens Axboe return 0; 2153ed29b0b4SJens Axboe } 2154ed29b0b4SJens Axboe 2155ed29b0b4SJens Axboe static inline int io_submit_sqe(struct io_ring_ctx *ctx, struct io_kiocb *req, 2156ed29b0b4SJens Axboe const struct io_uring_sqe *sqe) 2157ed29b0b4SJens Axboe __must_hold(&ctx->uring_lock) 2158ed29b0b4SJens Axboe { 2159ed29b0b4SJens Axboe struct io_submit_link *link = &ctx->submit_state.link; 2160ed29b0b4SJens Axboe int ret; 2161ed29b0b4SJens Axboe 2162ed29b0b4SJens Axboe ret = io_init_req(ctx, req, sqe); 2163ed29b0b4SJens Axboe if (unlikely(ret)) 2164ed29b0b4SJens Axboe return io_submit_fail_init(sqe, req, ret); 2165ed29b0b4SJens Axboe 2166ed29b0b4SJens Axboe /* don't need @sqe from now on */ 216748863ffdSPavel Begunkov trace_io_uring_submit_sqe(req, true); 2168ed29b0b4SJens Axboe 2169ed29b0b4SJens Axboe /* 2170ed29b0b4SJens Axboe * If we already have a head request, queue this one for async 2171ed29b0b4SJens Axboe * submittal once the head completes. If we don't have a head but 2172ed29b0b4SJens Axboe * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be 2173ed29b0b4SJens Axboe * submitted sync once the chain is complete. If none of those 2174ed29b0b4SJens Axboe * conditions are true (normal request), then just queue it. 2175ed29b0b4SJens Axboe */ 2176ed29b0b4SJens Axboe if (unlikely(link->head)) { 2177ed29b0b4SJens Axboe ret = io_req_prep_async(req); 2178ed29b0b4SJens Axboe if (unlikely(ret)) 2179ed29b0b4SJens Axboe return io_submit_fail_init(sqe, req, ret); 2180ed29b0b4SJens Axboe 218148863ffdSPavel Begunkov trace_io_uring_link(req, link->head); 2182ed29b0b4SJens Axboe link->last->link = req; 2183ed29b0b4SJens Axboe link->last = req; 2184ed29b0b4SJens Axboe 2185ed29b0b4SJens Axboe if (req->flags & IO_REQ_LINK_FLAGS) 2186ed29b0b4SJens Axboe return 0; 2187ed29b0b4SJens Axboe /* last request of the link, flush it */ 2188ed29b0b4SJens Axboe req = link->head; 2189ed29b0b4SJens Axboe link->head = NULL; 2190ed29b0b4SJens Axboe if (req->flags & (REQ_F_FORCE_ASYNC | REQ_F_FAIL)) 2191ed29b0b4SJens Axboe goto fallback; 2192ed29b0b4SJens Axboe 2193ed29b0b4SJens Axboe } else if (unlikely(req->flags & (IO_REQ_LINK_FLAGS | 2194ed29b0b4SJens Axboe REQ_F_FORCE_ASYNC | REQ_F_FAIL))) { 2195ed29b0b4SJens Axboe if (req->flags & IO_REQ_LINK_FLAGS) { 2196ed29b0b4SJens Axboe link->head = req; 2197ed29b0b4SJens Axboe link->last = req; 2198ed29b0b4SJens Axboe } else { 2199ed29b0b4SJens Axboe fallback: 2200ed29b0b4SJens Axboe io_queue_sqe_fallback(req); 2201ed29b0b4SJens Axboe } 2202ed29b0b4SJens Axboe return 0; 2203ed29b0b4SJens Axboe } 2204ed29b0b4SJens Axboe 2205ed29b0b4SJens Axboe io_queue_sqe(req); 2206ed29b0b4SJens Axboe return 0; 2207ed29b0b4SJens Axboe } 2208ed29b0b4SJens Axboe 2209ed29b0b4SJens Axboe /* 2210ed29b0b4SJens Axboe * Batched submission is done, ensure local IO is flushed out. 2211ed29b0b4SJens Axboe */ 2212ed29b0b4SJens Axboe static void io_submit_state_end(struct io_ring_ctx *ctx) 2213ed29b0b4SJens Axboe { 2214ed29b0b4SJens Axboe struct io_submit_state *state = &ctx->submit_state; 2215ed29b0b4SJens Axboe 2216ed29b0b4SJens Axboe if (unlikely(state->link.head)) 2217ed29b0b4SJens Axboe io_queue_sqe_fallback(state->link.head); 2218ed29b0b4SJens Axboe /* flush only after queuing links as they can generate completions */ 2219ed29b0b4SJens Axboe io_submit_flush_completions(ctx); 2220ed29b0b4SJens Axboe if (state->plug_started) 2221ed29b0b4SJens Axboe blk_finish_plug(&state->plug); 2222ed29b0b4SJens Axboe } 2223ed29b0b4SJens Axboe 2224ed29b0b4SJens Axboe /* 2225ed29b0b4SJens Axboe * Start submission side cache. 2226ed29b0b4SJens Axboe */ 2227ed29b0b4SJens Axboe static void io_submit_state_start(struct io_submit_state *state, 2228ed29b0b4SJens Axboe unsigned int max_ios) 2229ed29b0b4SJens Axboe { 2230ed29b0b4SJens Axboe state->plug_started = false; 2231ed29b0b4SJens Axboe state->need_plug = max_ios > 2; 2232ed29b0b4SJens Axboe state->submit_nr = max_ios; 2233ed29b0b4SJens Axboe /* set only head, no need to init link_last in advance */ 2234ed29b0b4SJens Axboe state->link.head = NULL; 2235ed29b0b4SJens Axboe } 2236ed29b0b4SJens Axboe 2237ed29b0b4SJens Axboe static void io_commit_sqring(struct io_ring_ctx *ctx) 2238ed29b0b4SJens Axboe { 2239ed29b0b4SJens Axboe struct io_rings *rings = ctx->rings; 2240ed29b0b4SJens Axboe 2241ed29b0b4SJens Axboe /* 2242ed29b0b4SJens Axboe * Ensure any loads from the SQEs are done at this point, 2243ed29b0b4SJens Axboe * since once we write the new head, the application could 2244ed29b0b4SJens Axboe * write new data to them. 2245ed29b0b4SJens Axboe */ 2246ed29b0b4SJens Axboe smp_store_release(&rings->sq.head, ctx->cached_sq_head); 2247ed29b0b4SJens Axboe } 2248ed29b0b4SJens Axboe 2249ed29b0b4SJens Axboe /* 2250ed29b0b4SJens Axboe * Fetch an sqe, if one is available. Note this returns a pointer to memory 2251ed29b0b4SJens Axboe * that is mapped by userspace. This means that care needs to be taken to 2252ed29b0b4SJens Axboe * ensure that reads are stable, as we cannot rely on userspace always 2253ed29b0b4SJens Axboe * being a good citizen. If members of the sqe are validated and then later 2254ed29b0b4SJens Axboe * used, it's important that those reads are done through READ_ONCE() to 2255ed29b0b4SJens Axboe * prevent a re-load down the line. 2256ed29b0b4SJens Axboe */ 2257ed29b0b4SJens Axboe static const struct io_uring_sqe *io_get_sqe(struct io_ring_ctx *ctx) 2258ed29b0b4SJens Axboe { 2259ed29b0b4SJens Axboe unsigned head, mask = ctx->sq_entries - 1; 2260ed29b0b4SJens Axboe unsigned sq_idx = ctx->cached_sq_head++ & mask; 2261ed29b0b4SJens Axboe 2262ed29b0b4SJens Axboe /* 2263ed29b0b4SJens Axboe * The cached sq head (or cq tail) serves two purposes: 2264ed29b0b4SJens Axboe * 2265ed29b0b4SJens Axboe * 1) allows us to batch the cost of updating the user visible 2266ed29b0b4SJens Axboe * head updates. 2267ed29b0b4SJens Axboe * 2) allows the kernel side to track the head on its own, even 2268ed29b0b4SJens Axboe * though the application is the one updating it. 2269ed29b0b4SJens Axboe */ 2270ed29b0b4SJens Axboe head = READ_ONCE(ctx->sq_array[sq_idx]); 2271ed29b0b4SJens Axboe if (likely(head < ctx->sq_entries)) { 2272ed29b0b4SJens Axboe /* double index for 128-byte SQEs, twice as long */ 2273ed29b0b4SJens Axboe if (ctx->flags & IORING_SETUP_SQE128) 2274ed29b0b4SJens Axboe head <<= 1; 2275ed29b0b4SJens Axboe return &ctx->sq_sqes[head]; 2276ed29b0b4SJens Axboe } 2277ed29b0b4SJens Axboe 2278ed29b0b4SJens Axboe /* drop invalid entries */ 2279ed29b0b4SJens Axboe ctx->cq_extra--; 2280ed29b0b4SJens Axboe WRITE_ONCE(ctx->rings->sq_dropped, 2281ed29b0b4SJens Axboe READ_ONCE(ctx->rings->sq_dropped) + 1); 2282ed29b0b4SJens Axboe return NULL; 2283ed29b0b4SJens Axboe } 2284ed29b0b4SJens Axboe 228517437f31SJens Axboe int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr) 2286ed29b0b4SJens Axboe __must_hold(&ctx->uring_lock) 2287ed29b0b4SJens Axboe { 2288ed29b0b4SJens Axboe unsigned int entries = io_sqring_entries(ctx); 2289ed29b0b4SJens Axboe unsigned int left; 2290ed29b0b4SJens Axboe int ret; 2291ed29b0b4SJens Axboe 2292ed29b0b4SJens Axboe if (unlikely(!entries)) 2293ed29b0b4SJens Axboe return 0; 2294ed29b0b4SJens Axboe /* make sure SQ entry isn't read before tail */ 2295ed29b0b4SJens Axboe ret = left = min3(nr, ctx->sq_entries, entries); 2296ed29b0b4SJens Axboe io_get_task_refs(left); 2297ed29b0b4SJens Axboe io_submit_state_start(&ctx->submit_state, left); 2298ed29b0b4SJens Axboe 2299ed29b0b4SJens Axboe do { 2300ed29b0b4SJens Axboe const struct io_uring_sqe *sqe; 2301ed29b0b4SJens Axboe struct io_kiocb *req; 2302ed29b0b4SJens Axboe 2303ed29b0b4SJens Axboe if (unlikely(!io_alloc_req_refill(ctx))) 2304ed29b0b4SJens Axboe break; 2305ed29b0b4SJens Axboe req = io_alloc_req(ctx); 2306ed29b0b4SJens Axboe sqe = io_get_sqe(ctx); 2307ed29b0b4SJens Axboe if (unlikely(!sqe)) { 2308ed29b0b4SJens Axboe io_req_add_to_cache(req, ctx); 2309ed29b0b4SJens Axboe break; 2310ed29b0b4SJens Axboe } 2311ed29b0b4SJens Axboe 2312ed29b0b4SJens Axboe /* 2313ed29b0b4SJens Axboe * Continue submitting even for sqe failure if the 2314ed29b0b4SJens Axboe * ring was setup with IORING_SETUP_SUBMIT_ALL 2315ed29b0b4SJens Axboe */ 2316ed29b0b4SJens Axboe if (unlikely(io_submit_sqe(ctx, req, sqe)) && 2317ed29b0b4SJens Axboe !(ctx->flags & IORING_SETUP_SUBMIT_ALL)) { 2318ed29b0b4SJens Axboe left--; 2319ed29b0b4SJens Axboe break; 2320ed29b0b4SJens Axboe } 2321ed29b0b4SJens Axboe } while (--left); 2322ed29b0b4SJens Axboe 2323ed29b0b4SJens Axboe if (unlikely(left)) { 2324ed29b0b4SJens Axboe ret -= left; 2325ed29b0b4SJens Axboe /* try again if it submitted nothing and can't allocate a req */ 2326ed29b0b4SJens Axboe if (!ret && io_req_cache_empty(ctx)) 2327ed29b0b4SJens Axboe ret = -EAGAIN; 2328ed29b0b4SJens Axboe current->io_uring->cached_refs += left; 2329ed29b0b4SJens Axboe } 2330ed29b0b4SJens Axboe 2331ed29b0b4SJens Axboe io_submit_state_end(ctx); 2332ed29b0b4SJens Axboe /* Commit SQ ring head once we've consumed and submitted all SQEs */ 2333ed29b0b4SJens Axboe io_commit_sqring(ctx); 2334ed29b0b4SJens Axboe return ret; 2335ed29b0b4SJens Axboe } 2336ed29b0b4SJens Axboe 2337ed29b0b4SJens Axboe struct io_wait_queue { 2338ed29b0b4SJens Axboe struct wait_queue_entry wq; 2339ed29b0b4SJens Axboe struct io_ring_ctx *ctx; 2340ed29b0b4SJens Axboe unsigned cq_tail; 2341ed29b0b4SJens Axboe unsigned nr_timeouts; 2342ed29b0b4SJens Axboe }; 2343ed29b0b4SJens Axboe 2344b4c98d59SDylan Yudaken static inline bool io_has_work(struct io_ring_ctx *ctx) 2345b4c98d59SDylan Yudaken { 2346c0e0d6baSDylan Yudaken return test_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq) || 2347c0e0d6baSDylan Yudaken ((ctx->flags & IORING_SETUP_DEFER_TASKRUN) && 2348c0e0d6baSDylan Yudaken !llist_empty(&ctx->work_llist)); 2349b4c98d59SDylan Yudaken } 2350b4c98d59SDylan Yudaken 2351ed29b0b4SJens Axboe static inline bool io_should_wake(struct io_wait_queue *iowq) 2352ed29b0b4SJens Axboe { 2353ed29b0b4SJens Axboe struct io_ring_ctx *ctx = iowq->ctx; 23540fc8c2acSDylan Yudaken int dist = READ_ONCE(ctx->rings->cq.tail) - (int) iowq->cq_tail; 2355ed29b0b4SJens Axboe 2356ed29b0b4SJens Axboe /* 2357ed29b0b4SJens Axboe * Wake up if we have enough events, or if a timeout occurred since we 2358ed29b0b4SJens Axboe * started waiting. For timeouts, we always want to return to userspace, 2359ed29b0b4SJens Axboe * regardless of event count. 2360ed29b0b4SJens Axboe */ 2361ed29b0b4SJens Axboe return dist >= 0 || atomic_read(&ctx->cq_timeouts) != iowq->nr_timeouts; 2362ed29b0b4SJens Axboe } 2363ed29b0b4SJens Axboe 2364ed29b0b4SJens Axboe static int io_wake_function(struct wait_queue_entry *curr, unsigned int mode, 2365ed29b0b4SJens Axboe int wake_flags, void *key) 2366ed29b0b4SJens Axboe { 2367ed29b0b4SJens Axboe struct io_wait_queue *iowq = container_of(curr, struct io_wait_queue, 2368ed29b0b4SJens Axboe wq); 2369b4c98d59SDylan Yudaken struct io_ring_ctx *ctx = iowq->ctx; 2370ed29b0b4SJens Axboe 2371ed29b0b4SJens Axboe /* 2372ed29b0b4SJens Axboe * Cannot safely flush overflowed CQEs from here, ensure we wake up 2373ed29b0b4SJens Axboe * the task, and the next invocation will do it. 2374ed29b0b4SJens Axboe */ 2375b4c98d59SDylan Yudaken if (io_should_wake(iowq) || io_has_work(ctx)) 2376ed29b0b4SJens Axboe return autoremove_wake_function(curr, mode, wake_flags, key); 2377ed29b0b4SJens Axboe return -1; 2378ed29b0b4SJens Axboe } 2379ed29b0b4SJens Axboe 2380c0e0d6baSDylan Yudaken int io_run_task_work_sig(struct io_ring_ctx *ctx) 2381ed29b0b4SJens Axboe { 2382c0e0d6baSDylan Yudaken if (io_run_task_work_ctx(ctx) > 0) 2383ed29b0b4SJens Axboe return 1; 2384ed29b0b4SJens Axboe if (task_sigpending(current)) 2385ed29b0b4SJens Axboe return -EINTR; 2386ed29b0b4SJens Axboe return 0; 2387ed29b0b4SJens Axboe } 2388ed29b0b4SJens Axboe 2389ed29b0b4SJens Axboe /* when returns >0, the caller should retry */ 2390ed29b0b4SJens Axboe static inline int io_cqring_wait_schedule(struct io_ring_ctx *ctx, 2391ed29b0b4SJens Axboe struct io_wait_queue *iowq, 2392ed29b0b4SJens Axboe ktime_t timeout) 2393ed29b0b4SJens Axboe { 2394ed29b0b4SJens Axboe int ret; 2395ed29b0b4SJens Axboe unsigned long check_cq; 2396ed29b0b4SJens Axboe 2397ed29b0b4SJens Axboe /* make sure we run task_work before checking for signals */ 2398c0e0d6baSDylan Yudaken ret = io_run_task_work_sig(ctx); 2399ed29b0b4SJens Axboe if (ret || io_should_wake(iowq)) 2400ed29b0b4SJens Axboe return ret; 24013a08576bSPavel Begunkov 2402ed29b0b4SJens Axboe check_cq = READ_ONCE(ctx->check_cq); 24033a08576bSPavel Begunkov if (unlikely(check_cq)) { 2404ed29b0b4SJens Axboe /* let the caller flush overflows, retry */ 2405ed29b0b4SJens Axboe if (check_cq & BIT(IO_CHECK_CQ_OVERFLOW_BIT)) 2406ed29b0b4SJens Axboe return 1; 24073a08576bSPavel Begunkov if (check_cq & BIT(IO_CHECK_CQ_DROPPED_BIT)) 2408ed29b0b4SJens Axboe return -EBADR; 24093a08576bSPavel Begunkov } 2410ed29b0b4SJens Axboe if (!schedule_hrtimeout(&timeout, HRTIMER_MODE_ABS)) 2411ed29b0b4SJens Axboe return -ETIME; 2412ed29b0b4SJens Axboe return 1; 2413ed29b0b4SJens Axboe } 2414ed29b0b4SJens Axboe 2415ed29b0b4SJens Axboe /* 2416ed29b0b4SJens Axboe * Wait until events become available, if we don't already have some. The 2417ed29b0b4SJens Axboe * application must reap them itself, as they reside on the shared cq ring. 2418ed29b0b4SJens Axboe */ 2419ed29b0b4SJens Axboe static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events, 2420ed29b0b4SJens Axboe const sigset_t __user *sig, size_t sigsz, 2421ed29b0b4SJens Axboe struct __kernel_timespec __user *uts) 2422ed29b0b4SJens Axboe { 2423ed29b0b4SJens Axboe struct io_wait_queue iowq; 2424ed29b0b4SJens Axboe struct io_rings *rings = ctx->rings; 2425ed29b0b4SJens Axboe ktime_t timeout = KTIME_MAX; 2426ed29b0b4SJens Axboe int ret; 2427ed29b0b4SJens Axboe 242876de6749SPavel Begunkov if (!io_allowed_run_tw(ctx)) 242976de6749SPavel Begunkov return -EEXIST; 243076de6749SPavel Begunkov 2431ed29b0b4SJens Axboe do { 2432c0e0d6baSDylan Yudaken /* always run at least 1 task work to process local work */ 2433c0e0d6baSDylan Yudaken ret = io_run_task_work_ctx(ctx); 2434c0e0d6baSDylan Yudaken if (ret < 0) 2435c0e0d6baSDylan Yudaken return ret; 2436ed29b0b4SJens Axboe io_cqring_overflow_flush(ctx); 2437aa1df3a3SPavel Begunkov 24380fc8c2acSDylan Yudaken /* if user messes with these they will just get an early return */ 24390fc8c2acSDylan Yudaken if (__io_cqring_events_user(ctx) >= min_events) 2440ed29b0b4SJens Axboe return 0; 2441c0e0d6baSDylan Yudaken } while (ret > 0); 2442ed29b0b4SJens Axboe 2443ed29b0b4SJens Axboe if (sig) { 2444ed29b0b4SJens Axboe #ifdef CONFIG_COMPAT 2445ed29b0b4SJens Axboe if (in_compat_syscall()) 2446ed29b0b4SJens Axboe ret = set_compat_user_sigmask((const compat_sigset_t __user *)sig, 2447ed29b0b4SJens Axboe sigsz); 2448ed29b0b4SJens Axboe else 2449ed29b0b4SJens Axboe #endif 2450ed29b0b4SJens Axboe ret = set_user_sigmask(sig, sigsz); 2451ed29b0b4SJens Axboe 2452ed29b0b4SJens Axboe if (ret) 2453ed29b0b4SJens Axboe return ret; 2454ed29b0b4SJens Axboe } 2455ed29b0b4SJens Axboe 2456ed29b0b4SJens Axboe if (uts) { 2457ed29b0b4SJens Axboe struct timespec64 ts; 2458ed29b0b4SJens Axboe 2459ed29b0b4SJens Axboe if (get_timespec64(&ts, uts)) 2460ed29b0b4SJens Axboe return -EFAULT; 2461ed29b0b4SJens Axboe timeout = ktime_add_ns(timespec64_to_ktime(ts), ktime_get_ns()); 2462ed29b0b4SJens Axboe } 2463ed29b0b4SJens Axboe 2464ed29b0b4SJens Axboe init_waitqueue_func_entry(&iowq.wq, io_wake_function); 2465ed29b0b4SJens Axboe iowq.wq.private = current; 2466ed29b0b4SJens Axboe INIT_LIST_HEAD(&iowq.wq.entry); 2467ed29b0b4SJens Axboe iowq.ctx = ctx; 2468ed29b0b4SJens Axboe iowq.nr_timeouts = atomic_read(&ctx->cq_timeouts); 2469ed29b0b4SJens Axboe iowq.cq_tail = READ_ONCE(ctx->rings->cq.head) + min_events; 2470ed29b0b4SJens Axboe 2471ed29b0b4SJens Axboe trace_io_uring_cqring_wait(ctx, min_events); 2472ed29b0b4SJens Axboe do { 2473ed29b0b4SJens Axboe /* if we can't even flush overflow, don't wait for more */ 2474ed29b0b4SJens Axboe if (!io_cqring_overflow_flush(ctx)) { 2475ed29b0b4SJens Axboe ret = -EBUSY; 2476ed29b0b4SJens Axboe break; 2477ed29b0b4SJens Axboe } 2478ed29b0b4SJens Axboe prepare_to_wait_exclusive(&ctx->cq_wait, &iowq.wq, 2479ed29b0b4SJens Axboe TASK_INTERRUPTIBLE); 2480ed29b0b4SJens Axboe ret = io_cqring_wait_schedule(ctx, &iowq, timeout); 2481ed29b0b4SJens Axboe cond_resched(); 2482ed29b0b4SJens Axboe } while (ret > 0); 2483ed29b0b4SJens Axboe 2484ed29b0b4SJens Axboe finish_wait(&ctx->cq_wait, &iowq.wq); 2485ed29b0b4SJens Axboe restore_saved_sigmask_unless(ret == -EINTR); 2486ed29b0b4SJens Axboe 2487ed29b0b4SJens Axboe return READ_ONCE(rings->cq.head) == READ_ONCE(rings->cq.tail) ? ret : 0; 2488ed29b0b4SJens Axboe } 2489ed29b0b4SJens Axboe 2490ed29b0b4SJens Axboe static void io_mem_free(void *ptr) 2491ed29b0b4SJens Axboe { 2492ed29b0b4SJens Axboe struct page *page; 2493ed29b0b4SJens Axboe 2494ed29b0b4SJens Axboe if (!ptr) 2495ed29b0b4SJens Axboe return; 2496ed29b0b4SJens Axboe 2497ed29b0b4SJens Axboe page = virt_to_head_page(ptr); 2498ed29b0b4SJens Axboe if (put_page_testzero(page)) 2499ed29b0b4SJens Axboe free_compound_page(page); 2500ed29b0b4SJens Axboe } 2501ed29b0b4SJens Axboe 2502ed29b0b4SJens Axboe static void *io_mem_alloc(size_t size) 2503ed29b0b4SJens Axboe { 2504ed29b0b4SJens Axboe gfp_t gfp = GFP_KERNEL_ACCOUNT | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP; 2505ed29b0b4SJens Axboe 2506ed29b0b4SJens Axboe return (void *) __get_free_pages(gfp, get_order(size)); 2507ed29b0b4SJens Axboe } 2508ed29b0b4SJens Axboe 2509ed29b0b4SJens Axboe static unsigned long rings_size(struct io_ring_ctx *ctx, unsigned int sq_entries, 2510ed29b0b4SJens Axboe unsigned int cq_entries, size_t *sq_offset) 2511ed29b0b4SJens Axboe { 2512ed29b0b4SJens Axboe struct io_rings *rings; 2513ed29b0b4SJens Axboe size_t off, sq_array_size; 2514ed29b0b4SJens Axboe 2515ed29b0b4SJens Axboe off = struct_size(rings, cqes, cq_entries); 2516ed29b0b4SJens Axboe if (off == SIZE_MAX) 2517ed29b0b4SJens Axboe return SIZE_MAX; 2518ed29b0b4SJens Axboe if (ctx->flags & IORING_SETUP_CQE32) { 2519ed29b0b4SJens Axboe if (check_shl_overflow(off, 1, &off)) 2520ed29b0b4SJens Axboe return SIZE_MAX; 2521ed29b0b4SJens Axboe } 2522ed29b0b4SJens Axboe 2523ed29b0b4SJens Axboe #ifdef CONFIG_SMP 2524ed29b0b4SJens Axboe off = ALIGN(off, SMP_CACHE_BYTES); 2525ed29b0b4SJens Axboe if (off == 0) 2526ed29b0b4SJens Axboe return SIZE_MAX; 2527ed29b0b4SJens Axboe #endif 2528ed29b0b4SJens Axboe 2529ed29b0b4SJens Axboe if (sq_offset) 2530ed29b0b4SJens Axboe *sq_offset = off; 2531ed29b0b4SJens Axboe 2532ed29b0b4SJens Axboe sq_array_size = array_size(sizeof(u32), sq_entries); 2533ed29b0b4SJens Axboe if (sq_array_size == SIZE_MAX) 2534ed29b0b4SJens Axboe return SIZE_MAX; 2535ed29b0b4SJens Axboe 2536ed29b0b4SJens Axboe if (check_add_overflow(off, sq_array_size, &off)) 2537ed29b0b4SJens Axboe return SIZE_MAX; 2538ed29b0b4SJens Axboe 2539ed29b0b4SJens Axboe return off; 2540ed29b0b4SJens Axboe } 2541ed29b0b4SJens Axboe 2542ed29b0b4SJens Axboe static int io_eventfd_register(struct io_ring_ctx *ctx, void __user *arg, 2543ed29b0b4SJens Axboe unsigned int eventfd_async) 2544ed29b0b4SJens Axboe { 2545ed29b0b4SJens Axboe struct io_ev_fd *ev_fd; 2546ed29b0b4SJens Axboe __s32 __user *fds = arg; 2547ed29b0b4SJens Axboe int fd; 2548ed29b0b4SJens Axboe 2549ed29b0b4SJens Axboe ev_fd = rcu_dereference_protected(ctx->io_ev_fd, 2550ed29b0b4SJens Axboe lockdep_is_held(&ctx->uring_lock)); 2551ed29b0b4SJens Axboe if (ev_fd) 2552ed29b0b4SJens Axboe return -EBUSY; 2553ed29b0b4SJens Axboe 2554ed29b0b4SJens Axboe if (copy_from_user(&fd, fds, sizeof(*fds))) 2555ed29b0b4SJens Axboe return -EFAULT; 2556ed29b0b4SJens Axboe 2557ed29b0b4SJens Axboe ev_fd = kmalloc(sizeof(*ev_fd), GFP_KERNEL); 2558ed29b0b4SJens Axboe if (!ev_fd) 2559ed29b0b4SJens Axboe return -ENOMEM; 2560ed29b0b4SJens Axboe 2561ed29b0b4SJens Axboe ev_fd->cq_ev_fd = eventfd_ctx_fdget(fd); 2562ed29b0b4SJens Axboe if (IS_ERR(ev_fd->cq_ev_fd)) { 2563ed29b0b4SJens Axboe int ret = PTR_ERR(ev_fd->cq_ev_fd); 2564ed29b0b4SJens Axboe kfree(ev_fd); 2565ed29b0b4SJens Axboe return ret; 2566ed29b0b4SJens Axboe } 2567305bef98SPavel Begunkov 2568305bef98SPavel Begunkov spin_lock(&ctx->completion_lock); 2569305bef98SPavel Begunkov ctx->evfd_last_cq_tail = ctx->cached_cq_tail; 2570305bef98SPavel Begunkov spin_unlock(&ctx->completion_lock); 2571305bef98SPavel Begunkov 2572ed29b0b4SJens Axboe ev_fd->eventfd_async = eventfd_async; 2573ed29b0b4SJens Axboe ctx->has_evfd = true; 2574ed29b0b4SJens Axboe rcu_assign_pointer(ctx->io_ev_fd, ev_fd); 257521a091b9SDylan Yudaken atomic_set(&ev_fd->refs, 1); 257621a091b9SDylan Yudaken atomic_set(&ev_fd->ops, 0); 2577ed29b0b4SJens Axboe return 0; 2578ed29b0b4SJens Axboe } 2579ed29b0b4SJens Axboe 2580ed29b0b4SJens Axboe static int io_eventfd_unregister(struct io_ring_ctx *ctx) 2581ed29b0b4SJens Axboe { 2582ed29b0b4SJens Axboe struct io_ev_fd *ev_fd; 2583ed29b0b4SJens Axboe 2584ed29b0b4SJens Axboe ev_fd = rcu_dereference_protected(ctx->io_ev_fd, 2585ed29b0b4SJens Axboe lockdep_is_held(&ctx->uring_lock)); 2586ed29b0b4SJens Axboe if (ev_fd) { 2587ed29b0b4SJens Axboe ctx->has_evfd = false; 2588ed29b0b4SJens Axboe rcu_assign_pointer(ctx->io_ev_fd, NULL); 258921a091b9SDylan Yudaken if (!atomic_fetch_or(BIT(IO_EVENTFD_OP_FREE_BIT), &ev_fd->ops)) 259021a091b9SDylan Yudaken call_rcu(&ev_fd->rcu, io_eventfd_ops); 2591ed29b0b4SJens Axboe return 0; 2592ed29b0b4SJens Axboe } 2593ed29b0b4SJens Axboe 2594ed29b0b4SJens Axboe return -ENXIO; 2595ed29b0b4SJens Axboe } 2596ed29b0b4SJens Axboe 2597ed29b0b4SJens Axboe static void io_req_caches_free(struct io_ring_ctx *ctx) 2598ed29b0b4SJens Axboe { 2599ed29b0b4SJens Axboe int nr = 0; 2600ed29b0b4SJens Axboe 2601ed29b0b4SJens Axboe mutex_lock(&ctx->uring_lock); 260234f0bc42SPavel Begunkov io_flush_cached_locked_reqs(ctx, &ctx->submit_state); 2603ed29b0b4SJens Axboe 2604ed29b0b4SJens Axboe while (!io_req_cache_empty(ctx)) { 260534f0bc42SPavel Begunkov struct io_kiocb *req = io_alloc_req(ctx); 2606ed29b0b4SJens Axboe 2607ed29b0b4SJens Axboe kmem_cache_free(req_cachep, req); 2608ed29b0b4SJens Axboe nr++; 2609ed29b0b4SJens Axboe } 2610ed29b0b4SJens Axboe if (nr) 2611ed29b0b4SJens Axboe percpu_ref_put_many(&ctx->refs, nr); 2612ed29b0b4SJens Axboe mutex_unlock(&ctx->uring_lock); 2613ed29b0b4SJens Axboe } 2614ed29b0b4SJens Axboe 2615ed29b0b4SJens Axboe static __cold void io_ring_ctx_free(struct io_ring_ctx *ctx) 2616ed29b0b4SJens Axboe { 2617ed29b0b4SJens Axboe io_sq_thread_finish(ctx); 2618ed29b0b4SJens Axboe io_rsrc_refs_drop(ctx); 2619ed29b0b4SJens Axboe /* __io_rsrc_put_work() may need uring_lock to progress, wait w/o it */ 2620ed29b0b4SJens Axboe io_wait_rsrc_data(ctx->buf_data); 2621ed29b0b4SJens Axboe io_wait_rsrc_data(ctx->file_data); 2622ed29b0b4SJens Axboe 2623ed29b0b4SJens Axboe mutex_lock(&ctx->uring_lock); 2624ed29b0b4SJens Axboe if (ctx->buf_data) 2625ed29b0b4SJens Axboe __io_sqe_buffers_unregister(ctx); 2626ed29b0b4SJens Axboe if (ctx->file_data) 2627ed29b0b4SJens Axboe __io_sqe_files_unregister(ctx); 2628ed29b0b4SJens Axboe if (ctx->rings) 2629ed29b0b4SJens Axboe __io_cqring_overflow_flush(ctx, true); 2630ed29b0b4SJens Axboe io_eventfd_unregister(ctx); 26319b797a37SJens Axboe io_alloc_cache_free(&ctx->apoll_cache, io_apoll_cache_free); 263243e0bbbdSJens Axboe io_alloc_cache_free(&ctx->netmsg_cache, io_netmsg_cache_free); 2633ed29b0b4SJens Axboe mutex_unlock(&ctx->uring_lock); 2634ed29b0b4SJens Axboe io_destroy_buffers(ctx); 2635ed29b0b4SJens Axboe if (ctx->sq_creds) 2636ed29b0b4SJens Axboe put_cred(ctx->sq_creds); 263797bbdc06SPavel Begunkov if (ctx->submitter_task) 263897bbdc06SPavel Begunkov put_task_struct(ctx->submitter_task); 2639ed29b0b4SJens Axboe 2640ed29b0b4SJens Axboe /* there are no registered resources left, nobody uses it */ 2641ed29b0b4SJens Axboe if (ctx->rsrc_node) 2642ed29b0b4SJens Axboe io_rsrc_node_destroy(ctx->rsrc_node); 2643ed29b0b4SJens Axboe if (ctx->rsrc_backup_node) 2644ed29b0b4SJens Axboe io_rsrc_node_destroy(ctx->rsrc_backup_node); 2645ed29b0b4SJens Axboe flush_delayed_work(&ctx->rsrc_put_work); 2646ed29b0b4SJens Axboe flush_delayed_work(&ctx->fallback_work); 2647ed29b0b4SJens Axboe 2648ed29b0b4SJens Axboe WARN_ON_ONCE(!list_empty(&ctx->rsrc_ref_list)); 2649ed29b0b4SJens Axboe WARN_ON_ONCE(!llist_empty(&ctx->rsrc_put_llist)); 2650ed29b0b4SJens Axboe 2651ed29b0b4SJens Axboe #if defined(CONFIG_UNIX) 2652ed29b0b4SJens Axboe if (ctx->ring_sock) { 2653ed29b0b4SJens Axboe ctx->ring_sock->file = NULL; /* so that iput() is called */ 2654ed29b0b4SJens Axboe sock_release(ctx->ring_sock); 2655ed29b0b4SJens Axboe } 2656ed29b0b4SJens Axboe #endif 2657ed29b0b4SJens Axboe WARN_ON_ONCE(!list_empty(&ctx->ltimeout_list)); 2658ed29b0b4SJens Axboe 265942b6419dSPavel Begunkov if (ctx->mm_account) { 266042b6419dSPavel Begunkov mmdrop(ctx->mm_account); 266142b6419dSPavel Begunkov ctx->mm_account = NULL; 266242b6419dSPavel Begunkov } 2663ed29b0b4SJens Axboe io_mem_free(ctx->rings); 2664ed29b0b4SJens Axboe io_mem_free(ctx->sq_sqes); 2665ed29b0b4SJens Axboe 2666ed29b0b4SJens Axboe percpu_ref_exit(&ctx->refs); 2667ed29b0b4SJens Axboe free_uid(ctx->user); 2668ed29b0b4SJens Axboe io_req_caches_free(ctx); 2669ed29b0b4SJens Axboe if (ctx->hash_map) 2670ed29b0b4SJens Axboe io_wq_put_hash(ctx->hash_map); 2671e6f89be6SPavel Begunkov kfree(ctx->cancel_table.hbs); 26729ca9fb24SPavel Begunkov kfree(ctx->cancel_table_locked.hbs); 2673ed29b0b4SJens Axboe kfree(ctx->dummy_ubuf); 2674ed29b0b4SJens Axboe kfree(ctx->io_bl); 2675ed29b0b4SJens Axboe xa_destroy(&ctx->io_bl_xa); 2676ed29b0b4SJens Axboe kfree(ctx); 2677ed29b0b4SJens Axboe } 2678ed29b0b4SJens Axboe 2679ed29b0b4SJens Axboe static __poll_t io_uring_poll(struct file *file, poll_table *wait) 2680ed29b0b4SJens Axboe { 2681ed29b0b4SJens Axboe struct io_ring_ctx *ctx = file->private_data; 2682ed29b0b4SJens Axboe __poll_t mask = 0; 2683ed29b0b4SJens Axboe 2684ed29b0b4SJens Axboe poll_wait(file, &ctx->cq_wait, wait); 2685ed29b0b4SJens Axboe /* 2686ed29b0b4SJens Axboe * synchronizes with barrier from wq_has_sleeper call in 2687ed29b0b4SJens Axboe * io_commit_cqring 2688ed29b0b4SJens Axboe */ 2689ed29b0b4SJens Axboe smp_rmb(); 2690ed29b0b4SJens Axboe if (!io_sqring_full(ctx)) 2691ed29b0b4SJens Axboe mask |= EPOLLOUT | EPOLLWRNORM; 2692ed29b0b4SJens Axboe 2693ed29b0b4SJens Axboe /* 2694ed29b0b4SJens Axboe * Don't flush cqring overflow list here, just do a simple check. 2695ed29b0b4SJens Axboe * Otherwise there could possible be ABBA deadlock: 2696ed29b0b4SJens Axboe * CPU0 CPU1 2697ed29b0b4SJens Axboe * ---- ---- 2698ed29b0b4SJens Axboe * lock(&ctx->uring_lock); 2699ed29b0b4SJens Axboe * lock(&ep->mtx); 2700ed29b0b4SJens Axboe * lock(&ctx->uring_lock); 2701ed29b0b4SJens Axboe * lock(&ep->mtx); 2702ed29b0b4SJens Axboe * 2703ed29b0b4SJens Axboe * Users may get EPOLLIN meanwhile seeing nothing in cqring, this 2704ed29b0b4SJens Axboe * pushs them to do the flush. 2705ed29b0b4SJens Axboe */ 2706b4c98d59SDylan Yudaken 2707b4c98d59SDylan Yudaken if (io_cqring_events(ctx) || io_has_work(ctx)) 2708ed29b0b4SJens Axboe mask |= EPOLLIN | EPOLLRDNORM; 2709ed29b0b4SJens Axboe 2710ed29b0b4SJens Axboe return mask; 2711ed29b0b4SJens Axboe } 2712ed29b0b4SJens Axboe 2713ed29b0b4SJens Axboe static int io_unregister_personality(struct io_ring_ctx *ctx, unsigned id) 2714ed29b0b4SJens Axboe { 2715ed29b0b4SJens Axboe const struct cred *creds; 2716ed29b0b4SJens Axboe 2717ed29b0b4SJens Axboe creds = xa_erase(&ctx->personalities, id); 2718ed29b0b4SJens Axboe if (creds) { 2719ed29b0b4SJens Axboe put_cred(creds); 2720ed29b0b4SJens Axboe return 0; 2721ed29b0b4SJens Axboe } 2722ed29b0b4SJens Axboe 2723ed29b0b4SJens Axboe return -EINVAL; 2724ed29b0b4SJens Axboe } 2725ed29b0b4SJens Axboe 2726ed29b0b4SJens Axboe struct io_tctx_exit { 2727ed29b0b4SJens Axboe struct callback_head task_work; 2728ed29b0b4SJens Axboe struct completion completion; 2729ed29b0b4SJens Axboe struct io_ring_ctx *ctx; 2730ed29b0b4SJens Axboe }; 2731ed29b0b4SJens Axboe 2732ed29b0b4SJens Axboe static __cold void io_tctx_exit_cb(struct callback_head *cb) 2733ed29b0b4SJens Axboe { 2734ed29b0b4SJens Axboe struct io_uring_task *tctx = current->io_uring; 2735ed29b0b4SJens Axboe struct io_tctx_exit *work; 2736ed29b0b4SJens Axboe 2737ed29b0b4SJens Axboe work = container_of(cb, struct io_tctx_exit, task_work); 2738ed29b0b4SJens Axboe /* 2739ed29b0b4SJens Axboe * When @in_idle, we're in cancellation and it's racy to remove the 2740ed29b0b4SJens Axboe * node. It'll be removed by the end of cancellation, just ignore it. 2741ed29b0b4SJens Axboe */ 2742ed29b0b4SJens Axboe if (!atomic_read(&tctx->in_idle)) 2743ed29b0b4SJens Axboe io_uring_del_tctx_node((unsigned long)work->ctx); 2744ed29b0b4SJens Axboe complete(&work->completion); 2745ed29b0b4SJens Axboe } 2746ed29b0b4SJens Axboe 2747ed29b0b4SJens Axboe static __cold bool io_cancel_ctx_cb(struct io_wq_work *work, void *data) 2748ed29b0b4SJens Axboe { 2749ed29b0b4SJens Axboe struct io_kiocb *req = container_of(work, struct io_kiocb, work); 2750ed29b0b4SJens Axboe 2751ed29b0b4SJens Axboe return req->ctx == data; 2752ed29b0b4SJens Axboe } 2753ed29b0b4SJens Axboe 2754ed29b0b4SJens Axboe static __cold void io_ring_exit_work(struct work_struct *work) 2755ed29b0b4SJens Axboe { 2756ed29b0b4SJens Axboe struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx, exit_work); 2757ed29b0b4SJens Axboe unsigned long timeout = jiffies + HZ * 60 * 5; 2758ed29b0b4SJens Axboe unsigned long interval = HZ / 20; 2759ed29b0b4SJens Axboe struct io_tctx_exit exit; 2760ed29b0b4SJens Axboe struct io_tctx_node *node; 2761ed29b0b4SJens Axboe int ret; 2762ed29b0b4SJens Axboe 2763ed29b0b4SJens Axboe /* 2764ed29b0b4SJens Axboe * If we're doing polled IO and end up having requests being 2765ed29b0b4SJens Axboe * submitted async (out-of-line), then completions can come in while 2766ed29b0b4SJens Axboe * we're waiting for refs to drop. We need to reap these manually, 2767ed29b0b4SJens Axboe * as nobody else will be looking for them. 2768ed29b0b4SJens Axboe */ 2769ed29b0b4SJens Axboe do { 2770c0e0d6baSDylan Yudaken if (ctx->flags & IORING_SETUP_DEFER_TASKRUN) 2771c0e0d6baSDylan Yudaken io_move_task_work_from_local(ctx); 2772c0e0d6baSDylan Yudaken 2773affa87dbSPavel Begunkov while (io_uring_try_cancel_requests(ctx, NULL, true)) 2774affa87dbSPavel Begunkov cond_resched(); 2775affa87dbSPavel Begunkov 2776ed29b0b4SJens Axboe if (ctx->sq_data) { 2777ed29b0b4SJens Axboe struct io_sq_data *sqd = ctx->sq_data; 2778ed29b0b4SJens Axboe struct task_struct *tsk; 2779ed29b0b4SJens Axboe 2780ed29b0b4SJens Axboe io_sq_thread_park(sqd); 2781ed29b0b4SJens Axboe tsk = sqd->thread; 2782ed29b0b4SJens Axboe if (tsk && tsk->io_uring && tsk->io_uring->io_wq) 2783ed29b0b4SJens Axboe io_wq_cancel_cb(tsk->io_uring->io_wq, 2784ed29b0b4SJens Axboe io_cancel_ctx_cb, ctx, true); 2785ed29b0b4SJens Axboe io_sq_thread_unpark(sqd); 2786ed29b0b4SJens Axboe } 2787ed29b0b4SJens Axboe 2788ed29b0b4SJens Axboe io_req_caches_free(ctx); 2789ed29b0b4SJens Axboe 2790ed29b0b4SJens Axboe if (WARN_ON_ONCE(time_after(jiffies, timeout))) { 2791ed29b0b4SJens Axboe /* there is little hope left, don't run it too often */ 2792ed29b0b4SJens Axboe interval = HZ * 60; 2793ed29b0b4SJens Axboe } 2794ed29b0b4SJens Axboe } while (!wait_for_completion_timeout(&ctx->ref_comp, interval)); 2795ed29b0b4SJens Axboe 2796ed29b0b4SJens Axboe init_completion(&exit.completion); 2797ed29b0b4SJens Axboe init_task_work(&exit.task_work, io_tctx_exit_cb); 2798ed29b0b4SJens Axboe exit.ctx = ctx; 2799ed29b0b4SJens Axboe /* 2800ed29b0b4SJens Axboe * Some may use context even when all refs and requests have been put, 2801ed29b0b4SJens Axboe * and they are free to do so while still holding uring_lock or 2802ed29b0b4SJens Axboe * completion_lock, see io_req_task_submit(). Apart from other work, 2803ed29b0b4SJens Axboe * this lock/unlock section also waits them to finish. 2804ed29b0b4SJens Axboe */ 2805ed29b0b4SJens Axboe mutex_lock(&ctx->uring_lock); 2806ed29b0b4SJens Axboe while (!list_empty(&ctx->tctx_list)) { 2807ed29b0b4SJens Axboe WARN_ON_ONCE(time_after(jiffies, timeout)); 2808ed29b0b4SJens Axboe 2809ed29b0b4SJens Axboe node = list_first_entry(&ctx->tctx_list, struct io_tctx_node, 2810ed29b0b4SJens Axboe ctx_node); 2811ed29b0b4SJens Axboe /* don't spin on a single task if cancellation failed */ 2812ed29b0b4SJens Axboe list_rotate_left(&ctx->tctx_list); 2813ed29b0b4SJens Axboe ret = task_work_add(node->task, &exit.task_work, TWA_SIGNAL); 2814ed29b0b4SJens Axboe if (WARN_ON_ONCE(ret)) 2815ed29b0b4SJens Axboe continue; 2816ed29b0b4SJens Axboe 2817ed29b0b4SJens Axboe mutex_unlock(&ctx->uring_lock); 2818ed29b0b4SJens Axboe wait_for_completion(&exit.completion); 2819ed29b0b4SJens Axboe mutex_lock(&ctx->uring_lock); 2820ed29b0b4SJens Axboe } 2821ed29b0b4SJens Axboe mutex_unlock(&ctx->uring_lock); 2822ed29b0b4SJens Axboe spin_lock(&ctx->completion_lock); 2823ed29b0b4SJens Axboe spin_unlock(&ctx->completion_lock); 2824ed29b0b4SJens Axboe 2825ed29b0b4SJens Axboe io_ring_ctx_free(ctx); 2826ed29b0b4SJens Axboe } 2827ed29b0b4SJens Axboe 2828ed29b0b4SJens Axboe static __cold void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx) 2829ed29b0b4SJens Axboe { 2830ed29b0b4SJens Axboe unsigned long index; 2831ed29b0b4SJens Axboe struct creds *creds; 2832ed29b0b4SJens Axboe 2833ed29b0b4SJens Axboe mutex_lock(&ctx->uring_lock); 2834ed29b0b4SJens Axboe percpu_ref_kill(&ctx->refs); 2835ed29b0b4SJens Axboe if (ctx->rings) 2836ed29b0b4SJens Axboe __io_cqring_overflow_flush(ctx, true); 2837ed29b0b4SJens Axboe xa_for_each(&ctx->personalities, index, creds) 2838ed29b0b4SJens Axboe io_unregister_personality(ctx, index); 28399ca9fb24SPavel Begunkov if (ctx->rings) 28409ca9fb24SPavel Begunkov io_poll_remove_all(ctx, NULL, true); 2841ed29b0b4SJens Axboe mutex_unlock(&ctx->uring_lock); 2842ed29b0b4SJens Axboe 284302bac94bSPavel Begunkov /* 284402bac94bSPavel Begunkov * If we failed setting up the ctx, we might not have any rings 284502bac94bSPavel Begunkov * and therefore did not submit any requests 284602bac94bSPavel Begunkov */ 284702bac94bSPavel Begunkov if (ctx->rings) 2848ed29b0b4SJens Axboe io_kill_timeouts(ctx, NULL, true); 2849ed29b0b4SJens Axboe 2850ed29b0b4SJens Axboe INIT_WORK(&ctx->exit_work, io_ring_exit_work); 2851ed29b0b4SJens Axboe /* 2852ed29b0b4SJens Axboe * Use system_unbound_wq to avoid spawning tons of event kworkers 2853ed29b0b4SJens Axboe * if we're exiting a ton of rings at the same time. It just adds 2854ed29b0b4SJens Axboe * noise and overhead, there's no discernable change in runtime 2855ed29b0b4SJens Axboe * over using system_wq. 2856ed29b0b4SJens Axboe */ 2857ed29b0b4SJens Axboe queue_work(system_unbound_wq, &ctx->exit_work); 2858ed29b0b4SJens Axboe } 2859ed29b0b4SJens Axboe 2860ed29b0b4SJens Axboe static int io_uring_release(struct inode *inode, struct file *file) 2861ed29b0b4SJens Axboe { 2862ed29b0b4SJens Axboe struct io_ring_ctx *ctx = file->private_data; 2863ed29b0b4SJens Axboe 2864ed29b0b4SJens Axboe file->private_data = NULL; 2865ed29b0b4SJens Axboe io_ring_ctx_wait_and_kill(ctx); 2866ed29b0b4SJens Axboe return 0; 2867ed29b0b4SJens Axboe } 2868ed29b0b4SJens Axboe 2869ed29b0b4SJens Axboe struct io_task_cancel { 2870ed29b0b4SJens Axboe struct task_struct *task; 2871ed29b0b4SJens Axboe bool all; 2872ed29b0b4SJens Axboe }; 2873ed29b0b4SJens Axboe 2874ed29b0b4SJens Axboe static bool io_cancel_task_cb(struct io_wq_work *work, void *data) 2875ed29b0b4SJens Axboe { 2876ed29b0b4SJens Axboe struct io_kiocb *req = container_of(work, struct io_kiocb, work); 2877ed29b0b4SJens Axboe struct io_task_cancel *cancel = data; 2878ed29b0b4SJens Axboe 2879ed29b0b4SJens Axboe return io_match_task_safe(req, cancel->task, cancel->all); 2880ed29b0b4SJens Axboe } 2881ed29b0b4SJens Axboe 2882ed29b0b4SJens Axboe static __cold bool io_cancel_defer_files(struct io_ring_ctx *ctx, 2883ed29b0b4SJens Axboe struct task_struct *task, 2884ed29b0b4SJens Axboe bool cancel_all) 2885ed29b0b4SJens Axboe { 2886ed29b0b4SJens Axboe struct io_defer_entry *de; 2887ed29b0b4SJens Axboe LIST_HEAD(list); 2888ed29b0b4SJens Axboe 2889ed29b0b4SJens Axboe spin_lock(&ctx->completion_lock); 2890ed29b0b4SJens Axboe list_for_each_entry_reverse(de, &ctx->defer_list, list) { 2891ed29b0b4SJens Axboe if (io_match_task_safe(de->req, task, cancel_all)) { 2892ed29b0b4SJens Axboe list_cut_position(&list, &ctx->defer_list, &de->list); 2893ed29b0b4SJens Axboe break; 2894ed29b0b4SJens Axboe } 2895ed29b0b4SJens Axboe } 2896ed29b0b4SJens Axboe spin_unlock(&ctx->completion_lock); 2897ed29b0b4SJens Axboe if (list_empty(&list)) 2898ed29b0b4SJens Axboe return false; 2899ed29b0b4SJens Axboe 2900ed29b0b4SJens Axboe while (!list_empty(&list)) { 2901ed29b0b4SJens Axboe de = list_first_entry(&list, struct io_defer_entry, list); 2902ed29b0b4SJens Axboe list_del_init(&de->list); 2903e276ae34SPavel Begunkov io_req_task_queue_fail(de->req, -ECANCELED); 2904ed29b0b4SJens Axboe kfree(de); 2905ed29b0b4SJens Axboe } 2906ed29b0b4SJens Axboe return true; 2907ed29b0b4SJens Axboe } 2908ed29b0b4SJens Axboe 2909ed29b0b4SJens Axboe static __cold bool io_uring_try_cancel_iowq(struct io_ring_ctx *ctx) 2910ed29b0b4SJens Axboe { 2911ed29b0b4SJens Axboe struct io_tctx_node *node; 2912ed29b0b4SJens Axboe enum io_wq_cancel cret; 2913ed29b0b4SJens Axboe bool ret = false; 2914ed29b0b4SJens Axboe 2915ed29b0b4SJens Axboe mutex_lock(&ctx->uring_lock); 2916ed29b0b4SJens Axboe list_for_each_entry(node, &ctx->tctx_list, ctx_node) { 2917ed29b0b4SJens Axboe struct io_uring_task *tctx = node->task->io_uring; 2918ed29b0b4SJens Axboe 2919ed29b0b4SJens Axboe /* 2920ed29b0b4SJens Axboe * io_wq will stay alive while we hold uring_lock, because it's 2921ed29b0b4SJens Axboe * killed after ctx nodes, which requires to take the lock. 2922ed29b0b4SJens Axboe */ 2923ed29b0b4SJens Axboe if (!tctx || !tctx->io_wq) 2924ed29b0b4SJens Axboe continue; 2925ed29b0b4SJens Axboe cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_ctx_cb, ctx, true); 2926ed29b0b4SJens Axboe ret |= (cret != IO_WQ_CANCEL_NOTFOUND); 2927ed29b0b4SJens Axboe } 2928ed29b0b4SJens Axboe mutex_unlock(&ctx->uring_lock); 2929ed29b0b4SJens Axboe 2930ed29b0b4SJens Axboe return ret; 2931ed29b0b4SJens Axboe } 2932ed29b0b4SJens Axboe 2933affa87dbSPavel Begunkov static __cold bool io_uring_try_cancel_requests(struct io_ring_ctx *ctx, 2934ed29b0b4SJens Axboe struct task_struct *task, 2935ed29b0b4SJens Axboe bool cancel_all) 2936ed29b0b4SJens Axboe { 2937ed29b0b4SJens Axboe struct io_task_cancel cancel = { .task = task, .all = cancel_all, }; 2938ed29b0b4SJens Axboe struct io_uring_task *tctx = task ? task->io_uring : NULL; 2939affa87dbSPavel Begunkov enum io_wq_cancel cret; 2940affa87dbSPavel Begunkov bool ret = false; 2941ed29b0b4SJens Axboe 2942ed29b0b4SJens Axboe /* failed during ring init, it couldn't have issued any requests */ 2943ed29b0b4SJens Axboe if (!ctx->rings) 2944affa87dbSPavel Begunkov return false; 2945ed29b0b4SJens Axboe 2946ed29b0b4SJens Axboe if (!task) { 2947ed29b0b4SJens Axboe ret |= io_uring_try_cancel_iowq(ctx); 2948ed29b0b4SJens Axboe } else if (tctx && tctx->io_wq) { 2949ed29b0b4SJens Axboe /* 2950ed29b0b4SJens Axboe * Cancels requests of all rings, not only @ctx, but 2951ed29b0b4SJens Axboe * it's fine as the task is in exit/exec. 2952ed29b0b4SJens Axboe */ 2953ed29b0b4SJens Axboe cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_task_cb, 2954ed29b0b4SJens Axboe &cancel, true); 2955ed29b0b4SJens Axboe ret |= (cret != IO_WQ_CANCEL_NOTFOUND); 2956ed29b0b4SJens Axboe } 2957ed29b0b4SJens Axboe 2958ed29b0b4SJens Axboe /* SQPOLL thread does its own polling */ 2959ed29b0b4SJens Axboe if ((!(ctx->flags & IORING_SETUP_SQPOLL) && cancel_all) || 2960ed29b0b4SJens Axboe (ctx->sq_data && ctx->sq_data->thread == current)) { 2961ed29b0b4SJens Axboe while (!wq_list_empty(&ctx->iopoll_list)) { 2962ed29b0b4SJens Axboe io_iopoll_try_reap_events(ctx); 2963ed29b0b4SJens Axboe ret = true; 2964ed29b0b4SJens Axboe } 2965ed29b0b4SJens Axboe } 2966ed29b0b4SJens Axboe 2967c0e0d6baSDylan Yudaken if (ctx->flags & IORING_SETUP_DEFER_TASKRUN) 2968c0e0d6baSDylan Yudaken ret |= io_run_local_work(ctx) > 0; 2969ed29b0b4SJens Axboe ret |= io_cancel_defer_files(ctx, task, cancel_all); 29709ca9fb24SPavel Begunkov mutex_lock(&ctx->uring_lock); 2971ed29b0b4SJens Axboe ret |= io_poll_remove_all(ctx, task, cancel_all); 29729ca9fb24SPavel Begunkov mutex_unlock(&ctx->uring_lock); 2973ed29b0b4SJens Axboe ret |= io_kill_timeouts(ctx, task, cancel_all); 2974ed29b0b4SJens Axboe if (task) 2975c0e0d6baSDylan Yudaken ret |= io_run_task_work() > 0; 2976affa87dbSPavel Begunkov return ret; 2977ed29b0b4SJens Axboe } 2978ed29b0b4SJens Axboe 2979ed29b0b4SJens Axboe static s64 tctx_inflight(struct io_uring_task *tctx, bool tracked) 2980ed29b0b4SJens Axboe { 2981ed29b0b4SJens Axboe if (tracked) 2982ed29b0b4SJens Axboe return atomic_read(&tctx->inflight_tracked); 2983ed29b0b4SJens Axboe return percpu_counter_sum(&tctx->inflight); 2984ed29b0b4SJens Axboe } 2985ed29b0b4SJens Axboe 2986ed29b0b4SJens Axboe /* 2987ed29b0b4SJens Axboe * Find any io_uring ctx that this task has registered or done IO on, and cancel 2988ed29b0b4SJens Axboe * requests. @sqd should be not-null IFF it's an SQPOLL thread cancellation. 2989ed29b0b4SJens Axboe */ 299017437f31SJens Axboe __cold void io_uring_cancel_generic(bool cancel_all, struct io_sq_data *sqd) 2991ed29b0b4SJens Axboe { 2992ed29b0b4SJens Axboe struct io_uring_task *tctx = current->io_uring; 2993ed29b0b4SJens Axboe struct io_ring_ctx *ctx; 2994ed29b0b4SJens Axboe s64 inflight; 2995ed29b0b4SJens Axboe DEFINE_WAIT(wait); 2996ed29b0b4SJens Axboe 2997ed29b0b4SJens Axboe WARN_ON_ONCE(sqd && sqd->thread != current); 2998ed29b0b4SJens Axboe 2999ed29b0b4SJens Axboe if (!current->io_uring) 3000ed29b0b4SJens Axboe return; 3001ed29b0b4SJens Axboe if (tctx->io_wq) 3002ed29b0b4SJens Axboe io_wq_exit_start(tctx->io_wq); 3003ed29b0b4SJens Axboe 3004ed29b0b4SJens Axboe atomic_inc(&tctx->in_idle); 3005ed29b0b4SJens Axboe do { 3006affa87dbSPavel Begunkov bool loop = false; 3007affa87dbSPavel Begunkov 3008ed29b0b4SJens Axboe io_uring_drop_tctx_refs(current); 3009ed29b0b4SJens Axboe /* read completions before cancelations */ 3010ed29b0b4SJens Axboe inflight = tctx_inflight(tctx, !cancel_all); 3011ed29b0b4SJens Axboe if (!inflight) 3012ed29b0b4SJens Axboe break; 3013ed29b0b4SJens Axboe 3014ed29b0b4SJens Axboe if (!sqd) { 3015ed29b0b4SJens Axboe struct io_tctx_node *node; 3016ed29b0b4SJens Axboe unsigned long index; 3017ed29b0b4SJens Axboe 3018ed29b0b4SJens Axboe xa_for_each(&tctx->xa, index, node) { 3019ed29b0b4SJens Axboe /* sqpoll task will cancel all its requests */ 3020ed29b0b4SJens Axboe if (node->ctx->sq_data) 3021ed29b0b4SJens Axboe continue; 3022affa87dbSPavel Begunkov loop |= io_uring_try_cancel_requests(node->ctx, 3023affa87dbSPavel Begunkov current, cancel_all); 3024ed29b0b4SJens Axboe } 3025ed29b0b4SJens Axboe } else { 3026ed29b0b4SJens Axboe list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) 3027affa87dbSPavel Begunkov loop |= io_uring_try_cancel_requests(ctx, 3028affa87dbSPavel Begunkov current, 3029ed29b0b4SJens Axboe cancel_all); 3030ed29b0b4SJens Axboe } 3031ed29b0b4SJens Axboe 3032affa87dbSPavel Begunkov if (loop) { 3033affa87dbSPavel Begunkov cond_resched(); 3034affa87dbSPavel Begunkov continue; 3035affa87dbSPavel Begunkov } 3036affa87dbSPavel Begunkov 3037ed29b0b4SJens Axboe prepare_to_wait(&tctx->wait, &wait, TASK_INTERRUPTIBLE); 3038ed29b0b4SJens Axboe io_run_task_work(); 3039ed29b0b4SJens Axboe io_uring_drop_tctx_refs(current); 3040ed29b0b4SJens Axboe 3041ed29b0b4SJens Axboe /* 3042ed29b0b4SJens Axboe * If we've seen completions, retry without waiting. This 3043ed29b0b4SJens Axboe * avoids a race where a completion comes in before we did 3044ed29b0b4SJens Axboe * prepare_to_wait(). 3045ed29b0b4SJens Axboe */ 3046ed29b0b4SJens Axboe if (inflight == tctx_inflight(tctx, !cancel_all)) 3047ed29b0b4SJens Axboe schedule(); 3048ed29b0b4SJens Axboe finish_wait(&tctx->wait, &wait); 3049ed29b0b4SJens Axboe } while (1); 3050ed29b0b4SJens Axboe 3051ed29b0b4SJens Axboe io_uring_clean_tctx(tctx); 3052ed29b0b4SJens Axboe if (cancel_all) { 3053ed29b0b4SJens Axboe /* 3054ed29b0b4SJens Axboe * We shouldn't run task_works after cancel, so just leave 3055ed29b0b4SJens Axboe * ->in_idle set for normal exit. 3056ed29b0b4SJens Axboe */ 3057ed29b0b4SJens Axboe atomic_dec(&tctx->in_idle); 3058ed29b0b4SJens Axboe /* for exec all current's requests should be gone, kill tctx */ 3059ed29b0b4SJens Axboe __io_uring_free(current); 3060ed29b0b4SJens Axboe } 3061ed29b0b4SJens Axboe } 3062ed29b0b4SJens Axboe 3063ed29b0b4SJens Axboe void __io_uring_cancel(bool cancel_all) 3064ed29b0b4SJens Axboe { 3065ed29b0b4SJens Axboe io_uring_cancel_generic(cancel_all, NULL); 3066ed29b0b4SJens Axboe } 3067ed29b0b4SJens Axboe 3068ed29b0b4SJens Axboe static void *io_uring_validate_mmap_request(struct file *file, 3069ed29b0b4SJens Axboe loff_t pgoff, size_t sz) 3070ed29b0b4SJens Axboe { 3071ed29b0b4SJens Axboe struct io_ring_ctx *ctx = file->private_data; 3072ed29b0b4SJens Axboe loff_t offset = pgoff << PAGE_SHIFT; 3073ed29b0b4SJens Axboe struct page *page; 3074ed29b0b4SJens Axboe void *ptr; 3075ed29b0b4SJens Axboe 3076ed29b0b4SJens Axboe switch (offset) { 3077ed29b0b4SJens Axboe case IORING_OFF_SQ_RING: 3078ed29b0b4SJens Axboe case IORING_OFF_CQ_RING: 3079ed29b0b4SJens Axboe ptr = ctx->rings; 3080ed29b0b4SJens Axboe break; 3081ed29b0b4SJens Axboe case IORING_OFF_SQES: 3082ed29b0b4SJens Axboe ptr = ctx->sq_sqes; 3083ed29b0b4SJens Axboe break; 3084ed29b0b4SJens Axboe default: 3085ed29b0b4SJens Axboe return ERR_PTR(-EINVAL); 3086ed29b0b4SJens Axboe } 3087ed29b0b4SJens Axboe 3088ed29b0b4SJens Axboe page = virt_to_head_page(ptr); 3089ed29b0b4SJens Axboe if (sz > page_size(page)) 3090ed29b0b4SJens Axboe return ERR_PTR(-EINVAL); 3091ed29b0b4SJens Axboe 3092ed29b0b4SJens Axboe return ptr; 3093ed29b0b4SJens Axboe } 3094ed29b0b4SJens Axboe 3095ed29b0b4SJens Axboe #ifdef CONFIG_MMU 3096ed29b0b4SJens Axboe 3097ed29b0b4SJens Axboe static __cold int io_uring_mmap(struct file *file, struct vm_area_struct *vma) 3098ed29b0b4SJens Axboe { 3099ed29b0b4SJens Axboe size_t sz = vma->vm_end - vma->vm_start; 3100ed29b0b4SJens Axboe unsigned long pfn; 3101ed29b0b4SJens Axboe void *ptr; 3102ed29b0b4SJens Axboe 3103ed29b0b4SJens Axboe ptr = io_uring_validate_mmap_request(file, vma->vm_pgoff, sz); 3104ed29b0b4SJens Axboe if (IS_ERR(ptr)) 3105ed29b0b4SJens Axboe return PTR_ERR(ptr); 3106ed29b0b4SJens Axboe 3107ed29b0b4SJens Axboe pfn = virt_to_phys(ptr) >> PAGE_SHIFT; 3108ed29b0b4SJens Axboe return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot); 3109ed29b0b4SJens Axboe } 3110ed29b0b4SJens Axboe 3111ed29b0b4SJens Axboe #else /* !CONFIG_MMU */ 3112ed29b0b4SJens Axboe 3113ed29b0b4SJens Axboe static int io_uring_mmap(struct file *file, struct vm_area_struct *vma) 3114ed29b0b4SJens Axboe { 3115ed29b0b4SJens Axboe return vma->vm_flags & (VM_SHARED | VM_MAYSHARE) ? 0 : -EINVAL; 3116ed29b0b4SJens Axboe } 3117ed29b0b4SJens Axboe 3118ed29b0b4SJens Axboe static unsigned int io_uring_nommu_mmap_capabilities(struct file *file) 3119ed29b0b4SJens Axboe { 3120ed29b0b4SJens Axboe return NOMMU_MAP_DIRECT | NOMMU_MAP_READ | NOMMU_MAP_WRITE; 3121ed29b0b4SJens Axboe } 3122ed29b0b4SJens Axboe 3123ed29b0b4SJens Axboe static unsigned long io_uring_nommu_get_unmapped_area(struct file *file, 3124ed29b0b4SJens Axboe unsigned long addr, unsigned long len, 3125ed29b0b4SJens Axboe unsigned long pgoff, unsigned long flags) 3126ed29b0b4SJens Axboe { 3127ed29b0b4SJens Axboe void *ptr; 3128ed29b0b4SJens Axboe 3129ed29b0b4SJens Axboe ptr = io_uring_validate_mmap_request(file, pgoff, len); 3130ed29b0b4SJens Axboe if (IS_ERR(ptr)) 3131ed29b0b4SJens Axboe return PTR_ERR(ptr); 3132ed29b0b4SJens Axboe 3133ed29b0b4SJens Axboe return (unsigned long) ptr; 3134ed29b0b4SJens Axboe } 3135ed29b0b4SJens Axboe 3136ed29b0b4SJens Axboe #endif /* !CONFIG_MMU */ 3137ed29b0b4SJens Axboe 3138ed29b0b4SJens Axboe static int io_validate_ext_arg(unsigned flags, const void __user *argp, size_t argsz) 3139ed29b0b4SJens Axboe { 3140ed29b0b4SJens Axboe if (flags & IORING_ENTER_EXT_ARG) { 3141ed29b0b4SJens Axboe struct io_uring_getevents_arg arg; 3142ed29b0b4SJens Axboe 3143ed29b0b4SJens Axboe if (argsz != sizeof(arg)) 3144ed29b0b4SJens Axboe return -EINVAL; 3145ed29b0b4SJens Axboe if (copy_from_user(&arg, argp, sizeof(arg))) 3146ed29b0b4SJens Axboe return -EFAULT; 3147ed29b0b4SJens Axboe } 3148ed29b0b4SJens Axboe return 0; 3149ed29b0b4SJens Axboe } 3150ed29b0b4SJens Axboe 3151ed29b0b4SJens Axboe static int io_get_ext_arg(unsigned flags, const void __user *argp, size_t *argsz, 3152ed29b0b4SJens Axboe struct __kernel_timespec __user **ts, 3153ed29b0b4SJens Axboe const sigset_t __user **sig) 3154ed29b0b4SJens Axboe { 3155ed29b0b4SJens Axboe struct io_uring_getevents_arg arg; 3156ed29b0b4SJens Axboe 3157ed29b0b4SJens Axboe /* 3158ed29b0b4SJens Axboe * If EXT_ARG isn't set, then we have no timespec and the argp pointer 3159ed29b0b4SJens Axboe * is just a pointer to the sigset_t. 3160ed29b0b4SJens Axboe */ 3161ed29b0b4SJens Axboe if (!(flags & IORING_ENTER_EXT_ARG)) { 3162ed29b0b4SJens Axboe *sig = (const sigset_t __user *) argp; 3163ed29b0b4SJens Axboe *ts = NULL; 3164ed29b0b4SJens Axboe return 0; 3165ed29b0b4SJens Axboe } 3166ed29b0b4SJens Axboe 3167ed29b0b4SJens Axboe /* 3168ed29b0b4SJens Axboe * EXT_ARG is set - ensure we agree on the size of it and copy in our 3169ed29b0b4SJens Axboe * timespec and sigset_t pointers if good. 3170ed29b0b4SJens Axboe */ 3171ed29b0b4SJens Axboe if (*argsz != sizeof(arg)) 3172ed29b0b4SJens Axboe return -EINVAL; 3173ed29b0b4SJens Axboe if (copy_from_user(&arg, argp, sizeof(arg))) 3174ed29b0b4SJens Axboe return -EFAULT; 3175ed29b0b4SJens Axboe if (arg.pad) 3176ed29b0b4SJens Axboe return -EINVAL; 3177ed29b0b4SJens Axboe *sig = u64_to_user_ptr(arg.sigmask); 3178ed29b0b4SJens Axboe *argsz = arg.sigmask_sz; 3179ed29b0b4SJens Axboe *ts = u64_to_user_ptr(arg.ts); 3180ed29b0b4SJens Axboe return 0; 3181ed29b0b4SJens Axboe } 3182ed29b0b4SJens Axboe 3183ed29b0b4SJens Axboe SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit, 3184ed29b0b4SJens Axboe u32, min_complete, u32, flags, const void __user *, argp, 3185ed29b0b4SJens Axboe size_t, argsz) 3186ed29b0b4SJens Axboe { 3187ed29b0b4SJens Axboe struct io_ring_ctx *ctx; 3188ed29b0b4SJens Axboe struct fd f; 3189ed29b0b4SJens Axboe long ret; 3190ed29b0b4SJens Axboe 3191ed29b0b4SJens Axboe if (unlikely(flags & ~(IORING_ENTER_GETEVENTS | IORING_ENTER_SQ_WAKEUP | 3192ed29b0b4SJens Axboe IORING_ENTER_SQ_WAIT | IORING_ENTER_EXT_ARG | 3193ed29b0b4SJens Axboe IORING_ENTER_REGISTERED_RING))) 3194ed29b0b4SJens Axboe return -EINVAL; 3195ed29b0b4SJens Axboe 3196ed29b0b4SJens Axboe /* 3197ed29b0b4SJens Axboe * Ring fd has been registered via IORING_REGISTER_RING_FDS, we 3198ed29b0b4SJens Axboe * need only dereference our task private array to find it. 3199ed29b0b4SJens Axboe */ 3200ed29b0b4SJens Axboe if (flags & IORING_ENTER_REGISTERED_RING) { 3201ed29b0b4SJens Axboe struct io_uring_task *tctx = current->io_uring; 3202ed29b0b4SJens Axboe 32033273c440SPavel Begunkov if (unlikely(!tctx || fd >= IO_RINGFD_REG_MAX)) 3204ed29b0b4SJens Axboe return -EINVAL; 3205ed29b0b4SJens Axboe fd = array_index_nospec(fd, IO_RINGFD_REG_MAX); 3206ed29b0b4SJens Axboe f.file = tctx->registered_rings[fd]; 3207ed29b0b4SJens Axboe f.flags = 0; 3208ed29b0b4SJens Axboe if (unlikely(!f.file)) 3209ed29b0b4SJens Axboe return -EBADF; 32103273c440SPavel Begunkov } else { 32113273c440SPavel Begunkov f = fdget(fd); 32123273c440SPavel Begunkov if (unlikely(!f.file)) 32133273c440SPavel Begunkov return -EBADF; 3214ed29b0b4SJens Axboe ret = -EOPNOTSUPP; 3215e5550a14SJens Axboe if (unlikely(!io_is_uring_fops(f.file))) 3216fbb8bb02SPavel Begunkov goto out; 32173273c440SPavel Begunkov } 3218ed29b0b4SJens Axboe 3219ed29b0b4SJens Axboe ctx = f.file->private_data; 3220ed29b0b4SJens Axboe ret = -EBADFD; 3221ed29b0b4SJens Axboe if (unlikely(ctx->flags & IORING_SETUP_R_DISABLED)) 3222ed29b0b4SJens Axboe goto out; 3223ed29b0b4SJens Axboe 3224ed29b0b4SJens Axboe /* 3225ed29b0b4SJens Axboe * For SQ polling, the thread will do all submissions and completions. 3226ed29b0b4SJens Axboe * Just return the requested submit count, and wake the thread if 3227ed29b0b4SJens Axboe * we were asked to. 3228ed29b0b4SJens Axboe */ 3229ed29b0b4SJens Axboe ret = 0; 3230ed29b0b4SJens Axboe if (ctx->flags & IORING_SETUP_SQPOLL) { 3231ed29b0b4SJens Axboe io_cqring_overflow_flush(ctx); 3232ed29b0b4SJens Axboe 3233ed29b0b4SJens Axboe if (unlikely(ctx->sq_data->thread == NULL)) { 3234ed29b0b4SJens Axboe ret = -EOWNERDEAD; 3235ed29b0b4SJens Axboe goto out; 3236ed29b0b4SJens Axboe } 3237ed29b0b4SJens Axboe if (flags & IORING_ENTER_SQ_WAKEUP) 3238ed29b0b4SJens Axboe wake_up(&ctx->sq_data->wait); 3239ed29b0b4SJens Axboe if (flags & IORING_ENTER_SQ_WAIT) { 3240ed29b0b4SJens Axboe ret = io_sqpoll_wait_sq(ctx); 3241ed29b0b4SJens Axboe if (ret) 3242ed29b0b4SJens Axboe goto out; 3243ed29b0b4SJens Axboe } 3244ed29b0b4SJens Axboe ret = to_submit; 3245ed29b0b4SJens Axboe } else if (to_submit) { 3246ed29b0b4SJens Axboe ret = io_uring_add_tctx_node(ctx); 3247ed29b0b4SJens Axboe if (unlikely(ret)) 3248ed29b0b4SJens Axboe goto out; 3249ed29b0b4SJens Axboe 3250ed29b0b4SJens Axboe mutex_lock(&ctx->uring_lock); 3251ed29b0b4SJens Axboe ret = io_submit_sqes(ctx, to_submit); 3252ed29b0b4SJens Axboe if (ret != to_submit) { 3253ed29b0b4SJens Axboe mutex_unlock(&ctx->uring_lock); 3254ed29b0b4SJens Axboe goto out; 3255ed29b0b4SJens Axboe } 325644f87745SPavel Begunkov if (flags & IORING_ENTER_GETEVENTS) { 325744f87745SPavel Begunkov if (ctx->syscall_iopoll) 3258ed29b0b4SJens Axboe goto iopoll_locked; 325944f87745SPavel Begunkov /* 326044f87745SPavel Begunkov * Ignore errors, we'll soon call io_cqring_wait() and 326144f87745SPavel Begunkov * it should handle ownership problems if any. 326244f87745SPavel Begunkov */ 326344f87745SPavel Begunkov if (ctx->flags & IORING_SETUP_DEFER_TASKRUN) 326444f87745SPavel Begunkov (void)io_run_local_work_locked(ctx); 326544f87745SPavel Begunkov } 3266ed29b0b4SJens Axboe mutex_unlock(&ctx->uring_lock); 3267ed29b0b4SJens Axboe } 3268c0e0d6baSDylan Yudaken 3269ed29b0b4SJens Axboe if (flags & IORING_ENTER_GETEVENTS) { 3270ed29b0b4SJens Axboe int ret2; 3271c0e0d6baSDylan Yudaken 3272ed29b0b4SJens Axboe if (ctx->syscall_iopoll) { 3273ed29b0b4SJens Axboe /* 3274ed29b0b4SJens Axboe * We disallow the app entering submit/complete with 3275ed29b0b4SJens Axboe * polling, but we still need to lock the ring to 3276ed29b0b4SJens Axboe * prevent racing with polled issue that got punted to 3277ed29b0b4SJens Axboe * a workqueue. 3278ed29b0b4SJens Axboe */ 3279ed29b0b4SJens Axboe mutex_lock(&ctx->uring_lock); 3280ed29b0b4SJens Axboe iopoll_locked: 3281ed29b0b4SJens Axboe ret2 = io_validate_ext_arg(flags, argp, argsz); 3282ed29b0b4SJens Axboe if (likely(!ret2)) { 3283ed29b0b4SJens Axboe min_complete = min(min_complete, 3284ed29b0b4SJens Axboe ctx->cq_entries); 3285ed29b0b4SJens Axboe ret2 = io_iopoll_check(ctx, min_complete); 3286ed29b0b4SJens Axboe } 3287ed29b0b4SJens Axboe mutex_unlock(&ctx->uring_lock); 3288ed29b0b4SJens Axboe } else { 3289ed29b0b4SJens Axboe const sigset_t __user *sig; 3290ed29b0b4SJens Axboe struct __kernel_timespec __user *ts; 3291ed29b0b4SJens Axboe 3292ed29b0b4SJens Axboe ret2 = io_get_ext_arg(flags, argp, &argsz, &ts, &sig); 3293ed29b0b4SJens Axboe if (likely(!ret2)) { 3294ed29b0b4SJens Axboe min_complete = min(min_complete, 3295ed29b0b4SJens Axboe ctx->cq_entries); 3296ed29b0b4SJens Axboe ret2 = io_cqring_wait(ctx, min_complete, sig, 3297ed29b0b4SJens Axboe argsz, ts); 3298ed29b0b4SJens Axboe } 3299ed29b0b4SJens Axboe } 3300ed29b0b4SJens Axboe 3301ed29b0b4SJens Axboe if (!ret) { 3302ed29b0b4SJens Axboe ret = ret2; 3303ed29b0b4SJens Axboe 3304ed29b0b4SJens Axboe /* 3305ed29b0b4SJens Axboe * EBADR indicates that one or more CQE were dropped. 3306ed29b0b4SJens Axboe * Once the user has been informed we can clear the bit 3307ed29b0b4SJens Axboe * as they are obviously ok with those drops. 3308ed29b0b4SJens Axboe */ 3309ed29b0b4SJens Axboe if (unlikely(ret2 == -EBADR)) 3310ed29b0b4SJens Axboe clear_bit(IO_CHECK_CQ_DROPPED_BIT, 3311ed29b0b4SJens Axboe &ctx->check_cq); 3312ed29b0b4SJens Axboe } 3313ed29b0b4SJens Axboe } 3314ed29b0b4SJens Axboe out: 3315ed29b0b4SJens Axboe fdput(f); 3316ed29b0b4SJens Axboe return ret; 3317ed29b0b4SJens Axboe } 3318ed29b0b4SJens Axboe 3319ed29b0b4SJens Axboe static const struct file_operations io_uring_fops = { 3320ed29b0b4SJens Axboe .release = io_uring_release, 3321ed29b0b4SJens Axboe .mmap = io_uring_mmap, 3322ed29b0b4SJens Axboe #ifndef CONFIG_MMU 3323ed29b0b4SJens Axboe .get_unmapped_area = io_uring_nommu_get_unmapped_area, 3324ed29b0b4SJens Axboe .mmap_capabilities = io_uring_nommu_mmap_capabilities, 3325ed29b0b4SJens Axboe #endif 3326ed29b0b4SJens Axboe .poll = io_uring_poll, 3327ed29b0b4SJens Axboe #ifdef CONFIG_PROC_FS 3328ed29b0b4SJens Axboe .show_fdinfo = io_uring_show_fdinfo, 3329ed29b0b4SJens Axboe #endif 3330ed29b0b4SJens Axboe }; 3331ed29b0b4SJens Axboe 333292ac8beaSJens Axboe bool io_is_uring_fops(struct file *file) 333392ac8beaSJens Axboe { 333492ac8beaSJens Axboe return file->f_op == &io_uring_fops; 333592ac8beaSJens Axboe } 333692ac8beaSJens Axboe 3337ed29b0b4SJens Axboe static __cold int io_allocate_scq_urings(struct io_ring_ctx *ctx, 3338ed29b0b4SJens Axboe struct io_uring_params *p) 3339ed29b0b4SJens Axboe { 3340ed29b0b4SJens Axboe struct io_rings *rings; 3341ed29b0b4SJens Axboe size_t size, sq_array_offset; 3342ed29b0b4SJens Axboe 3343ed29b0b4SJens Axboe /* make sure these are sane, as we already accounted them */ 3344ed29b0b4SJens Axboe ctx->sq_entries = p->sq_entries; 3345ed29b0b4SJens Axboe ctx->cq_entries = p->cq_entries; 3346ed29b0b4SJens Axboe 3347ed29b0b4SJens Axboe size = rings_size(ctx, p->sq_entries, p->cq_entries, &sq_array_offset); 3348ed29b0b4SJens Axboe if (size == SIZE_MAX) 3349ed29b0b4SJens Axboe return -EOVERFLOW; 3350ed29b0b4SJens Axboe 3351ed29b0b4SJens Axboe rings = io_mem_alloc(size); 3352ed29b0b4SJens Axboe if (!rings) 3353ed29b0b4SJens Axboe return -ENOMEM; 3354ed29b0b4SJens Axboe 3355ed29b0b4SJens Axboe ctx->rings = rings; 3356ed29b0b4SJens Axboe ctx->sq_array = (u32 *)((char *)rings + sq_array_offset); 3357ed29b0b4SJens Axboe rings->sq_ring_mask = p->sq_entries - 1; 3358ed29b0b4SJens Axboe rings->cq_ring_mask = p->cq_entries - 1; 3359ed29b0b4SJens Axboe rings->sq_ring_entries = p->sq_entries; 3360ed29b0b4SJens Axboe rings->cq_ring_entries = p->cq_entries; 3361ed29b0b4SJens Axboe 3362ed29b0b4SJens Axboe if (p->flags & IORING_SETUP_SQE128) 3363ed29b0b4SJens Axboe size = array_size(2 * sizeof(struct io_uring_sqe), p->sq_entries); 3364ed29b0b4SJens Axboe else 3365ed29b0b4SJens Axboe size = array_size(sizeof(struct io_uring_sqe), p->sq_entries); 3366ed29b0b4SJens Axboe if (size == SIZE_MAX) { 3367ed29b0b4SJens Axboe io_mem_free(ctx->rings); 3368ed29b0b4SJens Axboe ctx->rings = NULL; 3369ed29b0b4SJens Axboe return -EOVERFLOW; 3370ed29b0b4SJens Axboe } 3371ed29b0b4SJens Axboe 3372ed29b0b4SJens Axboe ctx->sq_sqes = io_mem_alloc(size); 3373ed29b0b4SJens Axboe if (!ctx->sq_sqes) { 3374ed29b0b4SJens Axboe io_mem_free(ctx->rings); 3375ed29b0b4SJens Axboe ctx->rings = NULL; 3376ed29b0b4SJens Axboe return -ENOMEM; 3377ed29b0b4SJens Axboe } 3378ed29b0b4SJens Axboe 3379ed29b0b4SJens Axboe return 0; 3380ed29b0b4SJens Axboe } 3381ed29b0b4SJens Axboe 3382ed29b0b4SJens Axboe static int io_uring_install_fd(struct io_ring_ctx *ctx, struct file *file) 3383ed29b0b4SJens Axboe { 3384ed29b0b4SJens Axboe int ret, fd; 3385ed29b0b4SJens Axboe 3386ed29b0b4SJens Axboe fd = get_unused_fd_flags(O_RDWR | O_CLOEXEC); 3387ed29b0b4SJens Axboe if (fd < 0) 3388ed29b0b4SJens Axboe return fd; 3389ed29b0b4SJens Axboe 339097c96e9fSDylan Yudaken ret = __io_uring_add_tctx_node(ctx); 3391ed29b0b4SJens Axboe if (ret) { 3392ed29b0b4SJens Axboe put_unused_fd(fd); 3393ed29b0b4SJens Axboe return ret; 3394ed29b0b4SJens Axboe } 3395ed29b0b4SJens Axboe fd_install(fd, file); 3396ed29b0b4SJens Axboe return fd; 3397ed29b0b4SJens Axboe } 3398ed29b0b4SJens Axboe 3399ed29b0b4SJens Axboe /* 3400ed29b0b4SJens Axboe * Allocate an anonymous fd, this is what constitutes the application 3401ed29b0b4SJens Axboe * visible backing of an io_uring instance. The application mmaps this 3402ed29b0b4SJens Axboe * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled, 3403ed29b0b4SJens Axboe * we have to tie this fd to a socket for file garbage collection purposes. 3404ed29b0b4SJens Axboe */ 3405ed29b0b4SJens Axboe static struct file *io_uring_get_file(struct io_ring_ctx *ctx) 3406ed29b0b4SJens Axboe { 3407ed29b0b4SJens Axboe struct file *file; 3408ed29b0b4SJens Axboe #if defined(CONFIG_UNIX) 3409ed29b0b4SJens Axboe int ret; 3410ed29b0b4SJens Axboe 3411ed29b0b4SJens Axboe ret = sock_create_kern(&init_net, PF_UNIX, SOCK_RAW, IPPROTO_IP, 3412ed29b0b4SJens Axboe &ctx->ring_sock); 3413ed29b0b4SJens Axboe if (ret) 3414ed29b0b4SJens Axboe return ERR_PTR(ret); 3415ed29b0b4SJens Axboe #endif 3416ed29b0b4SJens Axboe 3417ed29b0b4SJens Axboe file = anon_inode_getfile_secure("[io_uring]", &io_uring_fops, ctx, 3418ed29b0b4SJens Axboe O_RDWR | O_CLOEXEC, NULL); 3419ed29b0b4SJens Axboe #if defined(CONFIG_UNIX) 3420ed29b0b4SJens Axboe if (IS_ERR(file)) { 3421ed29b0b4SJens Axboe sock_release(ctx->ring_sock); 3422ed29b0b4SJens Axboe ctx->ring_sock = NULL; 3423ed29b0b4SJens Axboe } else { 3424ed29b0b4SJens Axboe ctx->ring_sock->file = file; 3425ed29b0b4SJens Axboe } 3426ed29b0b4SJens Axboe #endif 3427ed29b0b4SJens Axboe return file; 3428ed29b0b4SJens Axboe } 3429ed29b0b4SJens Axboe 3430ed29b0b4SJens Axboe static __cold int io_uring_create(unsigned entries, struct io_uring_params *p, 3431ed29b0b4SJens Axboe struct io_uring_params __user *params) 3432ed29b0b4SJens Axboe { 3433ed29b0b4SJens Axboe struct io_ring_ctx *ctx; 3434ed29b0b4SJens Axboe struct file *file; 3435ed29b0b4SJens Axboe int ret; 3436ed29b0b4SJens Axboe 3437ed29b0b4SJens Axboe if (!entries) 3438ed29b0b4SJens Axboe return -EINVAL; 3439ed29b0b4SJens Axboe if (entries > IORING_MAX_ENTRIES) { 3440ed29b0b4SJens Axboe if (!(p->flags & IORING_SETUP_CLAMP)) 3441ed29b0b4SJens Axboe return -EINVAL; 3442ed29b0b4SJens Axboe entries = IORING_MAX_ENTRIES; 3443ed29b0b4SJens Axboe } 3444ed29b0b4SJens Axboe 3445ed29b0b4SJens Axboe /* 3446ed29b0b4SJens Axboe * Use twice as many entries for the CQ ring. It's possible for the 3447ed29b0b4SJens Axboe * application to drive a higher depth than the size of the SQ ring, 3448ed29b0b4SJens Axboe * since the sqes are only used at submission time. This allows for 3449ed29b0b4SJens Axboe * some flexibility in overcommitting a bit. If the application has 3450ed29b0b4SJens Axboe * set IORING_SETUP_CQSIZE, it will have passed in the desired number 3451ed29b0b4SJens Axboe * of CQ ring entries manually. 3452ed29b0b4SJens Axboe */ 3453ed29b0b4SJens Axboe p->sq_entries = roundup_pow_of_two(entries); 3454ed29b0b4SJens Axboe if (p->flags & IORING_SETUP_CQSIZE) { 3455ed29b0b4SJens Axboe /* 3456ed29b0b4SJens Axboe * If IORING_SETUP_CQSIZE is set, we do the same roundup 3457ed29b0b4SJens Axboe * to a power-of-two, if it isn't already. We do NOT impose 3458ed29b0b4SJens Axboe * any cq vs sq ring sizing. 3459ed29b0b4SJens Axboe */ 3460ed29b0b4SJens Axboe if (!p->cq_entries) 3461ed29b0b4SJens Axboe return -EINVAL; 3462ed29b0b4SJens Axboe if (p->cq_entries > IORING_MAX_CQ_ENTRIES) { 3463ed29b0b4SJens Axboe if (!(p->flags & IORING_SETUP_CLAMP)) 3464ed29b0b4SJens Axboe return -EINVAL; 3465ed29b0b4SJens Axboe p->cq_entries = IORING_MAX_CQ_ENTRIES; 3466ed29b0b4SJens Axboe } 3467ed29b0b4SJens Axboe p->cq_entries = roundup_pow_of_two(p->cq_entries); 3468ed29b0b4SJens Axboe if (p->cq_entries < p->sq_entries) 3469ed29b0b4SJens Axboe return -EINVAL; 3470ed29b0b4SJens Axboe } else { 3471ed29b0b4SJens Axboe p->cq_entries = 2 * p->sq_entries; 3472ed29b0b4SJens Axboe } 3473ed29b0b4SJens Axboe 3474ed29b0b4SJens Axboe ctx = io_ring_ctx_alloc(p); 3475ed29b0b4SJens Axboe if (!ctx) 3476ed29b0b4SJens Axboe return -ENOMEM; 3477ed29b0b4SJens Axboe 3478ed29b0b4SJens Axboe /* 3479ed29b0b4SJens Axboe * When SETUP_IOPOLL and SETUP_SQPOLL are both enabled, user 3480ed29b0b4SJens Axboe * space applications don't need to do io completion events 3481ed29b0b4SJens Axboe * polling again, they can rely on io_sq_thread to do polling 3482ed29b0b4SJens Axboe * work, which can reduce cpu usage and uring_lock contention. 3483ed29b0b4SJens Axboe */ 3484ed29b0b4SJens Axboe if (ctx->flags & IORING_SETUP_IOPOLL && 3485ed29b0b4SJens Axboe !(ctx->flags & IORING_SETUP_SQPOLL)) 3486ed29b0b4SJens Axboe ctx->syscall_iopoll = 1; 3487ed29b0b4SJens Axboe 3488ed29b0b4SJens Axboe ctx->compat = in_compat_syscall(); 3489ed29b0b4SJens Axboe if (!capable(CAP_IPC_LOCK)) 3490ed29b0b4SJens Axboe ctx->user = get_uid(current_user()); 3491ed29b0b4SJens Axboe 3492ed29b0b4SJens Axboe /* 3493ed29b0b4SJens Axboe * For SQPOLL, we just need a wakeup, always. For !SQPOLL, if 3494ed29b0b4SJens Axboe * COOP_TASKRUN is set, then IPIs are never needed by the app. 3495ed29b0b4SJens Axboe */ 3496ed29b0b4SJens Axboe ret = -EINVAL; 3497ed29b0b4SJens Axboe if (ctx->flags & IORING_SETUP_SQPOLL) { 3498ed29b0b4SJens Axboe /* IPI related flags don't make sense with SQPOLL */ 3499ed29b0b4SJens Axboe if (ctx->flags & (IORING_SETUP_COOP_TASKRUN | 3500c0e0d6baSDylan Yudaken IORING_SETUP_TASKRUN_FLAG | 3501c0e0d6baSDylan Yudaken IORING_SETUP_DEFER_TASKRUN)) 3502ed29b0b4SJens Axboe goto err; 3503ed29b0b4SJens Axboe ctx->notify_method = TWA_SIGNAL_NO_IPI; 3504ed29b0b4SJens Axboe } else if (ctx->flags & IORING_SETUP_COOP_TASKRUN) { 3505ed29b0b4SJens Axboe ctx->notify_method = TWA_SIGNAL_NO_IPI; 3506ed29b0b4SJens Axboe } else { 3507c0e0d6baSDylan Yudaken if (ctx->flags & IORING_SETUP_TASKRUN_FLAG && 3508c0e0d6baSDylan Yudaken !(ctx->flags & IORING_SETUP_DEFER_TASKRUN)) 3509ed29b0b4SJens Axboe goto err; 3510ed29b0b4SJens Axboe ctx->notify_method = TWA_SIGNAL; 3511ed29b0b4SJens Axboe } 3512ed29b0b4SJens Axboe 3513ed29b0b4SJens Axboe /* 3514c0e0d6baSDylan Yudaken * For DEFER_TASKRUN we require the completion task to be the same as the 3515c0e0d6baSDylan Yudaken * submission task. This implies that there is only one submitter, so enforce 3516c0e0d6baSDylan Yudaken * that. 3517c0e0d6baSDylan Yudaken */ 3518c0e0d6baSDylan Yudaken if (ctx->flags & IORING_SETUP_DEFER_TASKRUN && 3519c0e0d6baSDylan Yudaken !(ctx->flags & IORING_SETUP_SINGLE_ISSUER)) { 3520c0e0d6baSDylan Yudaken goto err; 3521c0e0d6baSDylan Yudaken } 3522c0e0d6baSDylan Yudaken 3523c0e0d6baSDylan Yudaken /* 3524ed29b0b4SJens Axboe * This is just grabbed for accounting purposes. When a process exits, 3525ed29b0b4SJens Axboe * the mm is exited and dropped before the files, hence we need to hang 3526ed29b0b4SJens Axboe * on to this mm purely for the purposes of being able to unaccount 3527ed29b0b4SJens Axboe * memory (locked/pinned vm). It's not used for anything else. 3528ed29b0b4SJens Axboe */ 3529ed29b0b4SJens Axboe mmgrab(current->mm); 3530ed29b0b4SJens Axboe ctx->mm_account = current->mm; 3531ed29b0b4SJens Axboe 3532ed29b0b4SJens Axboe ret = io_allocate_scq_urings(ctx, p); 3533ed29b0b4SJens Axboe if (ret) 3534ed29b0b4SJens Axboe goto err; 3535ed29b0b4SJens Axboe 3536ed29b0b4SJens Axboe ret = io_sq_offload_create(ctx, p); 3537ed29b0b4SJens Axboe if (ret) 3538ed29b0b4SJens Axboe goto err; 3539ed29b0b4SJens Axboe /* always set a rsrc node */ 3540ed29b0b4SJens Axboe ret = io_rsrc_node_switch_start(ctx); 3541ed29b0b4SJens Axboe if (ret) 3542ed29b0b4SJens Axboe goto err; 3543ed29b0b4SJens Axboe io_rsrc_node_switch(ctx, NULL); 3544ed29b0b4SJens Axboe 3545ed29b0b4SJens Axboe memset(&p->sq_off, 0, sizeof(p->sq_off)); 3546ed29b0b4SJens Axboe p->sq_off.head = offsetof(struct io_rings, sq.head); 3547ed29b0b4SJens Axboe p->sq_off.tail = offsetof(struct io_rings, sq.tail); 3548ed29b0b4SJens Axboe p->sq_off.ring_mask = offsetof(struct io_rings, sq_ring_mask); 3549ed29b0b4SJens Axboe p->sq_off.ring_entries = offsetof(struct io_rings, sq_ring_entries); 3550ed29b0b4SJens Axboe p->sq_off.flags = offsetof(struct io_rings, sq_flags); 3551ed29b0b4SJens Axboe p->sq_off.dropped = offsetof(struct io_rings, sq_dropped); 3552ed29b0b4SJens Axboe p->sq_off.array = (char *)ctx->sq_array - (char *)ctx->rings; 3553ed29b0b4SJens Axboe 3554ed29b0b4SJens Axboe memset(&p->cq_off, 0, sizeof(p->cq_off)); 3555ed29b0b4SJens Axboe p->cq_off.head = offsetof(struct io_rings, cq.head); 3556ed29b0b4SJens Axboe p->cq_off.tail = offsetof(struct io_rings, cq.tail); 3557ed29b0b4SJens Axboe p->cq_off.ring_mask = offsetof(struct io_rings, cq_ring_mask); 3558ed29b0b4SJens Axboe p->cq_off.ring_entries = offsetof(struct io_rings, cq_ring_entries); 3559ed29b0b4SJens Axboe p->cq_off.overflow = offsetof(struct io_rings, cq_overflow); 3560ed29b0b4SJens Axboe p->cq_off.cqes = offsetof(struct io_rings, cqes); 3561ed29b0b4SJens Axboe p->cq_off.flags = offsetof(struct io_rings, cq_flags); 3562ed29b0b4SJens Axboe 3563ed29b0b4SJens Axboe p->features = IORING_FEAT_SINGLE_MMAP | IORING_FEAT_NODROP | 3564ed29b0b4SJens Axboe IORING_FEAT_SUBMIT_STABLE | IORING_FEAT_RW_CUR_POS | 3565ed29b0b4SJens Axboe IORING_FEAT_CUR_PERSONALITY | IORING_FEAT_FAST_POLL | 3566ed29b0b4SJens Axboe IORING_FEAT_POLL_32BITS | IORING_FEAT_SQPOLL_NONFIXED | 3567ed29b0b4SJens Axboe IORING_FEAT_EXT_ARG | IORING_FEAT_NATIVE_WORKERS | 3568ed29b0b4SJens Axboe IORING_FEAT_RSRC_TAGS | IORING_FEAT_CQE_SKIP | 3569ed29b0b4SJens Axboe IORING_FEAT_LINKED_FILE; 3570ed29b0b4SJens Axboe 3571ed29b0b4SJens Axboe if (copy_to_user(params, p, sizeof(*p))) { 3572ed29b0b4SJens Axboe ret = -EFAULT; 3573ed29b0b4SJens Axboe goto err; 3574ed29b0b4SJens Axboe } 3575ed29b0b4SJens Axboe 35767cae596bSDylan Yudaken if (ctx->flags & IORING_SETUP_SINGLE_ISSUER 35777cae596bSDylan Yudaken && !(ctx->flags & IORING_SETUP_R_DISABLED)) 35787cae596bSDylan Yudaken ctx->submitter_task = get_task_struct(current); 35797cae596bSDylan Yudaken 3580ed29b0b4SJens Axboe file = io_uring_get_file(ctx); 3581ed29b0b4SJens Axboe if (IS_ERR(file)) { 3582ed29b0b4SJens Axboe ret = PTR_ERR(file); 3583ed29b0b4SJens Axboe goto err; 3584ed29b0b4SJens Axboe } 3585ed29b0b4SJens Axboe 3586ed29b0b4SJens Axboe /* 3587ed29b0b4SJens Axboe * Install ring fd as the very last thing, so we don't risk someone 3588ed29b0b4SJens Axboe * having closed it before we finish setup 3589ed29b0b4SJens Axboe */ 3590ed29b0b4SJens Axboe ret = io_uring_install_fd(ctx, file); 3591ed29b0b4SJens Axboe if (ret < 0) { 3592ed29b0b4SJens Axboe /* fput will clean it up */ 3593ed29b0b4SJens Axboe fput(file); 3594ed29b0b4SJens Axboe return ret; 3595ed29b0b4SJens Axboe } 3596ed29b0b4SJens Axboe 3597ed29b0b4SJens Axboe trace_io_uring_create(ret, ctx, p->sq_entries, p->cq_entries, p->flags); 3598ed29b0b4SJens Axboe return ret; 3599ed29b0b4SJens Axboe err: 3600ed29b0b4SJens Axboe io_ring_ctx_wait_and_kill(ctx); 3601ed29b0b4SJens Axboe return ret; 3602ed29b0b4SJens Axboe } 3603ed29b0b4SJens Axboe 3604ed29b0b4SJens Axboe /* 3605ed29b0b4SJens Axboe * Sets up an aio uring context, and returns the fd. Applications asks for a 3606ed29b0b4SJens Axboe * ring size, we return the actual sq/cq ring sizes (among other things) in the 3607ed29b0b4SJens Axboe * params structure passed in. 3608ed29b0b4SJens Axboe */ 3609ed29b0b4SJens Axboe static long io_uring_setup(u32 entries, struct io_uring_params __user *params) 3610ed29b0b4SJens Axboe { 3611ed29b0b4SJens Axboe struct io_uring_params p; 3612ed29b0b4SJens Axboe int i; 3613ed29b0b4SJens Axboe 3614ed29b0b4SJens Axboe if (copy_from_user(&p, params, sizeof(p))) 3615ed29b0b4SJens Axboe return -EFAULT; 3616ed29b0b4SJens Axboe for (i = 0; i < ARRAY_SIZE(p.resv); i++) { 3617ed29b0b4SJens Axboe if (p.resv[i]) 3618ed29b0b4SJens Axboe return -EINVAL; 3619ed29b0b4SJens Axboe } 3620ed29b0b4SJens Axboe 3621ed29b0b4SJens Axboe if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL | 3622ed29b0b4SJens Axboe IORING_SETUP_SQ_AFF | IORING_SETUP_CQSIZE | 3623ed29b0b4SJens Axboe IORING_SETUP_CLAMP | IORING_SETUP_ATTACH_WQ | 3624ed29b0b4SJens Axboe IORING_SETUP_R_DISABLED | IORING_SETUP_SUBMIT_ALL | 3625ed29b0b4SJens Axboe IORING_SETUP_COOP_TASKRUN | IORING_SETUP_TASKRUN_FLAG | 362697bbdc06SPavel Begunkov IORING_SETUP_SQE128 | IORING_SETUP_CQE32 | 3627c0e0d6baSDylan Yudaken IORING_SETUP_SINGLE_ISSUER | IORING_SETUP_DEFER_TASKRUN)) 3628ed29b0b4SJens Axboe return -EINVAL; 3629ed29b0b4SJens Axboe 3630ed29b0b4SJens Axboe return io_uring_create(entries, &p, params); 3631ed29b0b4SJens Axboe } 3632ed29b0b4SJens Axboe 3633ed29b0b4SJens Axboe SYSCALL_DEFINE2(io_uring_setup, u32, entries, 3634ed29b0b4SJens Axboe struct io_uring_params __user *, params) 3635ed29b0b4SJens Axboe { 3636ed29b0b4SJens Axboe return io_uring_setup(entries, params); 3637ed29b0b4SJens Axboe } 3638ed29b0b4SJens Axboe 3639ed29b0b4SJens Axboe static __cold int io_probe(struct io_ring_ctx *ctx, void __user *arg, 3640ed29b0b4SJens Axboe unsigned nr_args) 3641ed29b0b4SJens Axboe { 3642ed29b0b4SJens Axboe struct io_uring_probe *p; 3643ed29b0b4SJens Axboe size_t size; 3644ed29b0b4SJens Axboe int i, ret; 3645ed29b0b4SJens Axboe 3646ed29b0b4SJens Axboe size = struct_size(p, ops, nr_args); 3647ed29b0b4SJens Axboe if (size == SIZE_MAX) 3648ed29b0b4SJens Axboe return -EOVERFLOW; 3649ed29b0b4SJens Axboe p = kzalloc(size, GFP_KERNEL); 3650ed29b0b4SJens Axboe if (!p) 3651ed29b0b4SJens Axboe return -ENOMEM; 3652ed29b0b4SJens Axboe 3653ed29b0b4SJens Axboe ret = -EFAULT; 3654ed29b0b4SJens Axboe if (copy_from_user(p, arg, size)) 3655ed29b0b4SJens Axboe goto out; 3656ed29b0b4SJens Axboe ret = -EINVAL; 3657ed29b0b4SJens Axboe if (memchr_inv(p, 0, size)) 3658ed29b0b4SJens Axboe goto out; 3659ed29b0b4SJens Axboe 3660ed29b0b4SJens Axboe p->last_op = IORING_OP_LAST - 1; 3661ed29b0b4SJens Axboe if (nr_args > IORING_OP_LAST) 3662ed29b0b4SJens Axboe nr_args = IORING_OP_LAST; 3663ed29b0b4SJens Axboe 3664ed29b0b4SJens Axboe for (i = 0; i < nr_args; i++) { 3665ed29b0b4SJens Axboe p->ops[i].op = i; 3666ed29b0b4SJens Axboe if (!io_op_defs[i].not_supported) 3667ed29b0b4SJens Axboe p->ops[i].flags = IO_URING_OP_SUPPORTED; 3668ed29b0b4SJens Axboe } 3669ed29b0b4SJens Axboe p->ops_len = i; 3670ed29b0b4SJens Axboe 3671ed29b0b4SJens Axboe ret = 0; 3672ed29b0b4SJens Axboe if (copy_to_user(arg, p, size)) 3673ed29b0b4SJens Axboe ret = -EFAULT; 3674ed29b0b4SJens Axboe out: 3675ed29b0b4SJens Axboe kfree(p); 3676ed29b0b4SJens Axboe return ret; 3677ed29b0b4SJens Axboe } 3678ed29b0b4SJens Axboe 3679ed29b0b4SJens Axboe static int io_register_personality(struct io_ring_ctx *ctx) 3680ed29b0b4SJens Axboe { 3681ed29b0b4SJens Axboe const struct cred *creds; 3682ed29b0b4SJens Axboe u32 id; 3683ed29b0b4SJens Axboe int ret; 3684ed29b0b4SJens Axboe 3685ed29b0b4SJens Axboe creds = get_current_cred(); 3686ed29b0b4SJens Axboe 3687ed29b0b4SJens Axboe ret = xa_alloc_cyclic(&ctx->personalities, &id, (void *)creds, 3688ed29b0b4SJens Axboe XA_LIMIT(0, USHRT_MAX), &ctx->pers_next, GFP_KERNEL); 3689ed29b0b4SJens Axboe if (ret < 0) { 3690ed29b0b4SJens Axboe put_cred(creds); 3691ed29b0b4SJens Axboe return ret; 3692ed29b0b4SJens Axboe } 3693ed29b0b4SJens Axboe return id; 3694ed29b0b4SJens Axboe } 3695ed29b0b4SJens Axboe 3696ed29b0b4SJens Axboe static __cold int io_register_restrictions(struct io_ring_ctx *ctx, 3697ed29b0b4SJens Axboe void __user *arg, unsigned int nr_args) 3698ed29b0b4SJens Axboe { 3699ed29b0b4SJens Axboe struct io_uring_restriction *res; 3700ed29b0b4SJens Axboe size_t size; 3701ed29b0b4SJens Axboe int i, ret; 3702ed29b0b4SJens Axboe 3703ed29b0b4SJens Axboe /* Restrictions allowed only if rings started disabled */ 3704ed29b0b4SJens Axboe if (!(ctx->flags & IORING_SETUP_R_DISABLED)) 3705ed29b0b4SJens Axboe return -EBADFD; 3706ed29b0b4SJens Axboe 3707ed29b0b4SJens Axboe /* We allow only a single restrictions registration */ 3708ed29b0b4SJens Axboe if (ctx->restrictions.registered) 3709ed29b0b4SJens Axboe return -EBUSY; 3710ed29b0b4SJens Axboe 3711ed29b0b4SJens Axboe if (!arg || nr_args > IORING_MAX_RESTRICTIONS) 3712ed29b0b4SJens Axboe return -EINVAL; 3713ed29b0b4SJens Axboe 3714ed29b0b4SJens Axboe size = array_size(nr_args, sizeof(*res)); 3715ed29b0b4SJens Axboe if (size == SIZE_MAX) 3716ed29b0b4SJens Axboe return -EOVERFLOW; 3717ed29b0b4SJens Axboe 3718ed29b0b4SJens Axboe res = memdup_user(arg, size); 3719ed29b0b4SJens Axboe if (IS_ERR(res)) 3720ed29b0b4SJens Axboe return PTR_ERR(res); 3721ed29b0b4SJens Axboe 3722ed29b0b4SJens Axboe ret = 0; 3723ed29b0b4SJens Axboe 3724ed29b0b4SJens Axboe for (i = 0; i < nr_args; i++) { 3725ed29b0b4SJens Axboe switch (res[i].opcode) { 3726ed29b0b4SJens Axboe case IORING_RESTRICTION_REGISTER_OP: 3727ed29b0b4SJens Axboe if (res[i].register_op >= IORING_REGISTER_LAST) { 3728ed29b0b4SJens Axboe ret = -EINVAL; 3729ed29b0b4SJens Axboe goto out; 3730ed29b0b4SJens Axboe } 3731ed29b0b4SJens Axboe 3732ed29b0b4SJens Axboe __set_bit(res[i].register_op, 3733ed29b0b4SJens Axboe ctx->restrictions.register_op); 3734ed29b0b4SJens Axboe break; 3735ed29b0b4SJens Axboe case IORING_RESTRICTION_SQE_OP: 3736ed29b0b4SJens Axboe if (res[i].sqe_op >= IORING_OP_LAST) { 3737ed29b0b4SJens Axboe ret = -EINVAL; 3738ed29b0b4SJens Axboe goto out; 3739ed29b0b4SJens Axboe } 3740ed29b0b4SJens Axboe 3741ed29b0b4SJens Axboe __set_bit(res[i].sqe_op, ctx->restrictions.sqe_op); 3742ed29b0b4SJens Axboe break; 3743ed29b0b4SJens Axboe case IORING_RESTRICTION_SQE_FLAGS_ALLOWED: 3744ed29b0b4SJens Axboe ctx->restrictions.sqe_flags_allowed = res[i].sqe_flags; 3745ed29b0b4SJens Axboe break; 3746ed29b0b4SJens Axboe case IORING_RESTRICTION_SQE_FLAGS_REQUIRED: 3747ed29b0b4SJens Axboe ctx->restrictions.sqe_flags_required = res[i].sqe_flags; 3748ed29b0b4SJens Axboe break; 3749ed29b0b4SJens Axboe default: 3750ed29b0b4SJens Axboe ret = -EINVAL; 3751ed29b0b4SJens Axboe goto out; 3752ed29b0b4SJens Axboe } 3753ed29b0b4SJens Axboe } 3754ed29b0b4SJens Axboe 3755ed29b0b4SJens Axboe out: 3756ed29b0b4SJens Axboe /* Reset all restrictions if an error happened */ 3757ed29b0b4SJens Axboe if (ret != 0) 3758ed29b0b4SJens Axboe memset(&ctx->restrictions, 0, sizeof(ctx->restrictions)); 3759ed29b0b4SJens Axboe else 3760ed29b0b4SJens Axboe ctx->restrictions.registered = true; 3761ed29b0b4SJens Axboe 3762ed29b0b4SJens Axboe kfree(res); 3763ed29b0b4SJens Axboe return ret; 3764ed29b0b4SJens Axboe } 3765ed29b0b4SJens Axboe 3766ed29b0b4SJens Axboe static int io_register_enable_rings(struct io_ring_ctx *ctx) 3767ed29b0b4SJens Axboe { 3768ed29b0b4SJens Axboe if (!(ctx->flags & IORING_SETUP_R_DISABLED)) 3769ed29b0b4SJens Axboe return -EBADFD; 3770ed29b0b4SJens Axboe 37717cae596bSDylan Yudaken if (ctx->flags & IORING_SETUP_SINGLE_ISSUER && !ctx->submitter_task) 37727cae596bSDylan Yudaken ctx->submitter_task = get_task_struct(current); 37737cae596bSDylan Yudaken 3774ed29b0b4SJens Axboe if (ctx->restrictions.registered) 3775ed29b0b4SJens Axboe ctx->restricted = 1; 3776ed29b0b4SJens Axboe 3777ed29b0b4SJens Axboe ctx->flags &= ~IORING_SETUP_R_DISABLED; 3778ed29b0b4SJens Axboe if (ctx->sq_data && wq_has_sleeper(&ctx->sq_data->wait)) 3779ed29b0b4SJens Axboe wake_up(&ctx->sq_data->wait); 3780ed29b0b4SJens Axboe return 0; 3781ed29b0b4SJens Axboe } 3782ed29b0b4SJens Axboe 3783ed29b0b4SJens Axboe static __cold int io_register_iowq_aff(struct io_ring_ctx *ctx, 3784ed29b0b4SJens Axboe void __user *arg, unsigned len) 3785ed29b0b4SJens Axboe { 3786ed29b0b4SJens Axboe struct io_uring_task *tctx = current->io_uring; 3787ed29b0b4SJens Axboe cpumask_var_t new_mask; 3788ed29b0b4SJens Axboe int ret; 3789ed29b0b4SJens Axboe 3790ed29b0b4SJens Axboe if (!tctx || !tctx->io_wq) 3791ed29b0b4SJens Axboe return -EINVAL; 3792ed29b0b4SJens Axboe 3793ed29b0b4SJens Axboe if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) 3794ed29b0b4SJens Axboe return -ENOMEM; 3795ed29b0b4SJens Axboe 3796ed29b0b4SJens Axboe cpumask_clear(new_mask); 3797ed29b0b4SJens Axboe if (len > cpumask_size()) 3798ed29b0b4SJens Axboe len = cpumask_size(); 3799ed29b0b4SJens Axboe 3800ed29b0b4SJens Axboe if (in_compat_syscall()) { 3801ed29b0b4SJens Axboe ret = compat_get_bitmap(cpumask_bits(new_mask), 3802ed29b0b4SJens Axboe (const compat_ulong_t __user *)arg, 3803ed29b0b4SJens Axboe len * 8 /* CHAR_BIT */); 3804ed29b0b4SJens Axboe } else { 3805ed29b0b4SJens Axboe ret = copy_from_user(new_mask, arg, len); 3806ed29b0b4SJens Axboe } 3807ed29b0b4SJens Axboe 3808ed29b0b4SJens Axboe if (ret) { 3809ed29b0b4SJens Axboe free_cpumask_var(new_mask); 3810ed29b0b4SJens Axboe return -EFAULT; 3811ed29b0b4SJens Axboe } 3812ed29b0b4SJens Axboe 3813ed29b0b4SJens Axboe ret = io_wq_cpu_affinity(tctx->io_wq, new_mask); 3814ed29b0b4SJens Axboe free_cpumask_var(new_mask); 3815ed29b0b4SJens Axboe return ret; 3816ed29b0b4SJens Axboe } 3817ed29b0b4SJens Axboe 3818ed29b0b4SJens Axboe static __cold int io_unregister_iowq_aff(struct io_ring_ctx *ctx) 3819ed29b0b4SJens Axboe { 3820ed29b0b4SJens Axboe struct io_uring_task *tctx = current->io_uring; 3821ed29b0b4SJens Axboe 3822ed29b0b4SJens Axboe if (!tctx || !tctx->io_wq) 3823ed29b0b4SJens Axboe return -EINVAL; 3824ed29b0b4SJens Axboe 3825ed29b0b4SJens Axboe return io_wq_cpu_affinity(tctx->io_wq, NULL); 3826ed29b0b4SJens Axboe } 3827ed29b0b4SJens Axboe 3828ed29b0b4SJens Axboe static __cold int io_register_iowq_max_workers(struct io_ring_ctx *ctx, 3829ed29b0b4SJens Axboe void __user *arg) 3830ed29b0b4SJens Axboe __must_hold(&ctx->uring_lock) 3831ed29b0b4SJens Axboe { 3832ed29b0b4SJens Axboe struct io_tctx_node *node; 3833ed29b0b4SJens Axboe struct io_uring_task *tctx = NULL; 3834ed29b0b4SJens Axboe struct io_sq_data *sqd = NULL; 3835ed29b0b4SJens Axboe __u32 new_count[2]; 3836ed29b0b4SJens Axboe int i, ret; 3837ed29b0b4SJens Axboe 3838ed29b0b4SJens Axboe if (copy_from_user(new_count, arg, sizeof(new_count))) 3839ed29b0b4SJens Axboe return -EFAULT; 3840ed29b0b4SJens Axboe for (i = 0; i < ARRAY_SIZE(new_count); i++) 3841ed29b0b4SJens Axboe if (new_count[i] > INT_MAX) 3842ed29b0b4SJens Axboe return -EINVAL; 3843ed29b0b4SJens Axboe 3844ed29b0b4SJens Axboe if (ctx->flags & IORING_SETUP_SQPOLL) { 3845ed29b0b4SJens Axboe sqd = ctx->sq_data; 3846ed29b0b4SJens Axboe if (sqd) { 3847ed29b0b4SJens Axboe /* 3848ed29b0b4SJens Axboe * Observe the correct sqd->lock -> ctx->uring_lock 3849ed29b0b4SJens Axboe * ordering. Fine to drop uring_lock here, we hold 3850ed29b0b4SJens Axboe * a ref to the ctx. 3851ed29b0b4SJens Axboe */ 3852ed29b0b4SJens Axboe refcount_inc(&sqd->refs); 3853ed29b0b4SJens Axboe mutex_unlock(&ctx->uring_lock); 3854ed29b0b4SJens Axboe mutex_lock(&sqd->lock); 3855ed29b0b4SJens Axboe mutex_lock(&ctx->uring_lock); 3856ed29b0b4SJens Axboe if (sqd->thread) 3857ed29b0b4SJens Axboe tctx = sqd->thread->io_uring; 3858ed29b0b4SJens Axboe } 3859ed29b0b4SJens Axboe } else { 3860ed29b0b4SJens Axboe tctx = current->io_uring; 3861ed29b0b4SJens Axboe } 3862ed29b0b4SJens Axboe 3863ed29b0b4SJens Axboe BUILD_BUG_ON(sizeof(new_count) != sizeof(ctx->iowq_limits)); 3864ed29b0b4SJens Axboe 3865ed29b0b4SJens Axboe for (i = 0; i < ARRAY_SIZE(new_count); i++) 3866ed29b0b4SJens Axboe if (new_count[i]) 3867ed29b0b4SJens Axboe ctx->iowq_limits[i] = new_count[i]; 3868ed29b0b4SJens Axboe ctx->iowq_limits_set = true; 3869ed29b0b4SJens Axboe 3870ed29b0b4SJens Axboe if (tctx && tctx->io_wq) { 3871ed29b0b4SJens Axboe ret = io_wq_max_workers(tctx->io_wq, new_count); 3872ed29b0b4SJens Axboe if (ret) 3873ed29b0b4SJens Axboe goto err; 3874ed29b0b4SJens Axboe } else { 3875ed29b0b4SJens Axboe memset(new_count, 0, sizeof(new_count)); 3876ed29b0b4SJens Axboe } 3877ed29b0b4SJens Axboe 3878ed29b0b4SJens Axboe if (sqd) { 3879ed29b0b4SJens Axboe mutex_unlock(&sqd->lock); 3880ed29b0b4SJens Axboe io_put_sq_data(sqd); 3881ed29b0b4SJens Axboe } 3882ed29b0b4SJens Axboe 3883ed29b0b4SJens Axboe if (copy_to_user(arg, new_count, sizeof(new_count))) 3884ed29b0b4SJens Axboe return -EFAULT; 3885ed29b0b4SJens Axboe 3886ed29b0b4SJens Axboe /* that's it for SQPOLL, only the SQPOLL task creates requests */ 3887ed29b0b4SJens Axboe if (sqd) 3888ed29b0b4SJens Axboe return 0; 3889ed29b0b4SJens Axboe 3890ed29b0b4SJens Axboe /* now propagate the restriction to all registered users */ 3891ed29b0b4SJens Axboe list_for_each_entry(node, &ctx->tctx_list, ctx_node) { 3892ed29b0b4SJens Axboe struct io_uring_task *tctx = node->task->io_uring; 3893ed29b0b4SJens Axboe 3894ed29b0b4SJens Axboe if (WARN_ON_ONCE(!tctx->io_wq)) 3895ed29b0b4SJens Axboe continue; 3896ed29b0b4SJens Axboe 3897ed29b0b4SJens Axboe for (i = 0; i < ARRAY_SIZE(new_count); i++) 3898ed29b0b4SJens Axboe new_count[i] = ctx->iowq_limits[i]; 3899ed29b0b4SJens Axboe /* ignore errors, it always returns zero anyway */ 3900ed29b0b4SJens Axboe (void)io_wq_max_workers(tctx->io_wq, new_count); 3901ed29b0b4SJens Axboe } 3902ed29b0b4SJens Axboe return 0; 3903ed29b0b4SJens Axboe err: 3904ed29b0b4SJens Axboe if (sqd) { 3905ed29b0b4SJens Axboe mutex_unlock(&sqd->lock); 3906ed29b0b4SJens Axboe io_put_sq_data(sqd); 3907ed29b0b4SJens Axboe } 3908ed29b0b4SJens Axboe return ret; 3909ed29b0b4SJens Axboe } 3910ed29b0b4SJens Axboe 3911ed29b0b4SJens Axboe static int __io_uring_register(struct io_ring_ctx *ctx, unsigned opcode, 3912ed29b0b4SJens Axboe void __user *arg, unsigned nr_args) 3913ed29b0b4SJens Axboe __releases(ctx->uring_lock) 3914ed29b0b4SJens Axboe __acquires(ctx->uring_lock) 3915ed29b0b4SJens Axboe { 3916ed29b0b4SJens Axboe int ret; 3917ed29b0b4SJens Axboe 3918ed29b0b4SJens Axboe /* 3919fbb8bb02SPavel Begunkov * We don't quiesce the refs for register anymore and so it can't be 3920fbb8bb02SPavel Begunkov * dying as we're holding a file ref here. 3921ed29b0b4SJens Axboe */ 3922fbb8bb02SPavel Begunkov if (WARN_ON_ONCE(percpu_ref_is_dying(&ctx->refs))) 3923ed29b0b4SJens Axboe return -ENXIO; 3924ed29b0b4SJens Axboe 3925d7cce96cSPavel Begunkov if (ctx->submitter_task && ctx->submitter_task != current) 3926d7cce96cSPavel Begunkov return -EEXIST; 3927d7cce96cSPavel Begunkov 3928ed29b0b4SJens Axboe if (ctx->restricted) { 3929ed29b0b4SJens Axboe if (opcode >= IORING_REGISTER_LAST) 3930ed29b0b4SJens Axboe return -EINVAL; 3931ed29b0b4SJens Axboe opcode = array_index_nospec(opcode, IORING_REGISTER_LAST); 3932ed29b0b4SJens Axboe if (!test_bit(opcode, ctx->restrictions.register_op)) 3933ed29b0b4SJens Axboe return -EACCES; 3934ed29b0b4SJens Axboe } 3935ed29b0b4SJens Axboe 3936ed29b0b4SJens Axboe switch (opcode) { 3937ed29b0b4SJens Axboe case IORING_REGISTER_BUFFERS: 3938ed29b0b4SJens Axboe ret = -EFAULT; 3939ed29b0b4SJens Axboe if (!arg) 3940ed29b0b4SJens Axboe break; 3941ed29b0b4SJens Axboe ret = io_sqe_buffers_register(ctx, arg, nr_args, NULL); 3942ed29b0b4SJens Axboe break; 3943ed29b0b4SJens Axboe case IORING_UNREGISTER_BUFFERS: 3944ed29b0b4SJens Axboe ret = -EINVAL; 3945ed29b0b4SJens Axboe if (arg || nr_args) 3946ed29b0b4SJens Axboe break; 3947ed29b0b4SJens Axboe ret = io_sqe_buffers_unregister(ctx); 3948ed29b0b4SJens Axboe break; 3949ed29b0b4SJens Axboe case IORING_REGISTER_FILES: 3950ed29b0b4SJens Axboe ret = -EFAULT; 3951ed29b0b4SJens Axboe if (!arg) 3952ed29b0b4SJens Axboe break; 3953ed29b0b4SJens Axboe ret = io_sqe_files_register(ctx, arg, nr_args, NULL); 3954ed29b0b4SJens Axboe break; 3955ed29b0b4SJens Axboe case IORING_UNREGISTER_FILES: 3956ed29b0b4SJens Axboe ret = -EINVAL; 3957ed29b0b4SJens Axboe if (arg || nr_args) 3958ed29b0b4SJens Axboe break; 3959ed29b0b4SJens Axboe ret = io_sqe_files_unregister(ctx); 3960ed29b0b4SJens Axboe break; 3961ed29b0b4SJens Axboe case IORING_REGISTER_FILES_UPDATE: 3962ed29b0b4SJens Axboe ret = io_register_files_update(ctx, arg, nr_args); 3963ed29b0b4SJens Axboe break; 3964ed29b0b4SJens Axboe case IORING_REGISTER_EVENTFD: 3965ed29b0b4SJens Axboe ret = -EINVAL; 3966ed29b0b4SJens Axboe if (nr_args != 1) 3967ed29b0b4SJens Axboe break; 3968ed29b0b4SJens Axboe ret = io_eventfd_register(ctx, arg, 0); 3969ed29b0b4SJens Axboe break; 3970ed29b0b4SJens Axboe case IORING_REGISTER_EVENTFD_ASYNC: 3971ed29b0b4SJens Axboe ret = -EINVAL; 3972ed29b0b4SJens Axboe if (nr_args != 1) 3973ed29b0b4SJens Axboe break; 3974ed29b0b4SJens Axboe ret = io_eventfd_register(ctx, arg, 1); 3975ed29b0b4SJens Axboe break; 3976ed29b0b4SJens Axboe case IORING_UNREGISTER_EVENTFD: 3977ed29b0b4SJens Axboe ret = -EINVAL; 3978ed29b0b4SJens Axboe if (arg || nr_args) 3979ed29b0b4SJens Axboe break; 3980ed29b0b4SJens Axboe ret = io_eventfd_unregister(ctx); 3981ed29b0b4SJens Axboe break; 3982ed29b0b4SJens Axboe case IORING_REGISTER_PROBE: 3983ed29b0b4SJens Axboe ret = -EINVAL; 3984ed29b0b4SJens Axboe if (!arg || nr_args > 256) 3985ed29b0b4SJens Axboe break; 3986ed29b0b4SJens Axboe ret = io_probe(ctx, arg, nr_args); 3987ed29b0b4SJens Axboe break; 3988ed29b0b4SJens Axboe case IORING_REGISTER_PERSONALITY: 3989ed29b0b4SJens Axboe ret = -EINVAL; 3990ed29b0b4SJens Axboe if (arg || nr_args) 3991ed29b0b4SJens Axboe break; 3992ed29b0b4SJens Axboe ret = io_register_personality(ctx); 3993ed29b0b4SJens Axboe break; 3994ed29b0b4SJens Axboe case IORING_UNREGISTER_PERSONALITY: 3995ed29b0b4SJens Axboe ret = -EINVAL; 3996ed29b0b4SJens Axboe if (arg) 3997ed29b0b4SJens Axboe break; 3998ed29b0b4SJens Axboe ret = io_unregister_personality(ctx, nr_args); 3999ed29b0b4SJens Axboe break; 4000ed29b0b4SJens Axboe case IORING_REGISTER_ENABLE_RINGS: 4001ed29b0b4SJens Axboe ret = -EINVAL; 4002ed29b0b4SJens Axboe if (arg || nr_args) 4003ed29b0b4SJens Axboe break; 4004ed29b0b4SJens Axboe ret = io_register_enable_rings(ctx); 4005ed29b0b4SJens Axboe break; 4006ed29b0b4SJens Axboe case IORING_REGISTER_RESTRICTIONS: 4007ed29b0b4SJens Axboe ret = io_register_restrictions(ctx, arg, nr_args); 4008ed29b0b4SJens Axboe break; 4009ed29b0b4SJens Axboe case IORING_REGISTER_FILES2: 4010ed29b0b4SJens Axboe ret = io_register_rsrc(ctx, arg, nr_args, IORING_RSRC_FILE); 4011ed29b0b4SJens Axboe break; 4012ed29b0b4SJens Axboe case IORING_REGISTER_FILES_UPDATE2: 4013ed29b0b4SJens Axboe ret = io_register_rsrc_update(ctx, arg, nr_args, 4014ed29b0b4SJens Axboe IORING_RSRC_FILE); 4015ed29b0b4SJens Axboe break; 4016ed29b0b4SJens Axboe case IORING_REGISTER_BUFFERS2: 4017ed29b0b4SJens Axboe ret = io_register_rsrc(ctx, arg, nr_args, IORING_RSRC_BUFFER); 4018ed29b0b4SJens Axboe break; 4019ed29b0b4SJens Axboe case IORING_REGISTER_BUFFERS_UPDATE: 4020ed29b0b4SJens Axboe ret = io_register_rsrc_update(ctx, arg, nr_args, 4021ed29b0b4SJens Axboe IORING_RSRC_BUFFER); 4022ed29b0b4SJens Axboe break; 4023ed29b0b4SJens Axboe case IORING_REGISTER_IOWQ_AFF: 4024ed29b0b4SJens Axboe ret = -EINVAL; 4025ed29b0b4SJens Axboe if (!arg || !nr_args) 4026ed29b0b4SJens Axboe break; 4027ed29b0b4SJens Axboe ret = io_register_iowq_aff(ctx, arg, nr_args); 4028ed29b0b4SJens Axboe break; 4029ed29b0b4SJens Axboe case IORING_UNREGISTER_IOWQ_AFF: 4030ed29b0b4SJens Axboe ret = -EINVAL; 4031ed29b0b4SJens Axboe if (arg || nr_args) 4032ed29b0b4SJens Axboe break; 4033ed29b0b4SJens Axboe ret = io_unregister_iowq_aff(ctx); 4034ed29b0b4SJens Axboe break; 4035ed29b0b4SJens Axboe case IORING_REGISTER_IOWQ_MAX_WORKERS: 4036ed29b0b4SJens Axboe ret = -EINVAL; 4037ed29b0b4SJens Axboe if (!arg || nr_args != 2) 4038ed29b0b4SJens Axboe break; 4039ed29b0b4SJens Axboe ret = io_register_iowq_max_workers(ctx, arg); 4040ed29b0b4SJens Axboe break; 4041ed29b0b4SJens Axboe case IORING_REGISTER_RING_FDS: 4042ed29b0b4SJens Axboe ret = io_ringfd_register(ctx, arg, nr_args); 4043ed29b0b4SJens Axboe break; 4044ed29b0b4SJens Axboe case IORING_UNREGISTER_RING_FDS: 4045ed29b0b4SJens Axboe ret = io_ringfd_unregister(ctx, arg, nr_args); 4046ed29b0b4SJens Axboe break; 4047ed29b0b4SJens Axboe case IORING_REGISTER_PBUF_RING: 4048ed29b0b4SJens Axboe ret = -EINVAL; 4049ed29b0b4SJens Axboe if (!arg || nr_args != 1) 4050ed29b0b4SJens Axboe break; 4051ed29b0b4SJens Axboe ret = io_register_pbuf_ring(ctx, arg); 4052ed29b0b4SJens Axboe break; 4053ed29b0b4SJens Axboe case IORING_UNREGISTER_PBUF_RING: 4054ed29b0b4SJens Axboe ret = -EINVAL; 4055ed29b0b4SJens Axboe if (!arg || nr_args != 1) 4056ed29b0b4SJens Axboe break; 4057ed29b0b4SJens Axboe ret = io_unregister_pbuf_ring(ctx, arg); 4058ed29b0b4SJens Axboe break; 405978a861b9SJens Axboe case IORING_REGISTER_SYNC_CANCEL: 406078a861b9SJens Axboe ret = -EINVAL; 406178a861b9SJens Axboe if (!arg || nr_args != 1) 406278a861b9SJens Axboe break; 406378a861b9SJens Axboe ret = io_sync_cancel(ctx, arg); 406478a861b9SJens Axboe break; 40656e73dffbSPavel Begunkov case IORING_REGISTER_FILE_ALLOC_RANGE: 40666e73dffbSPavel Begunkov ret = -EINVAL; 40676e73dffbSPavel Begunkov if (!arg || nr_args) 40686e73dffbSPavel Begunkov break; 40696e73dffbSPavel Begunkov ret = io_register_file_alloc_range(ctx, arg); 40706e73dffbSPavel Begunkov break; 4071ed29b0b4SJens Axboe default: 4072ed29b0b4SJens Axboe ret = -EINVAL; 4073ed29b0b4SJens Axboe break; 4074ed29b0b4SJens Axboe } 4075ed29b0b4SJens Axboe 4076ed29b0b4SJens Axboe return ret; 4077ed29b0b4SJens Axboe } 4078ed29b0b4SJens Axboe 4079ed29b0b4SJens Axboe SYSCALL_DEFINE4(io_uring_register, unsigned int, fd, unsigned int, opcode, 4080ed29b0b4SJens Axboe void __user *, arg, unsigned int, nr_args) 4081ed29b0b4SJens Axboe { 4082ed29b0b4SJens Axboe struct io_ring_ctx *ctx; 4083ed29b0b4SJens Axboe long ret = -EBADF; 4084ed29b0b4SJens Axboe struct fd f; 4085ed29b0b4SJens Axboe 4086ed29b0b4SJens Axboe f = fdget(fd); 4087ed29b0b4SJens Axboe if (!f.file) 4088ed29b0b4SJens Axboe return -EBADF; 4089ed29b0b4SJens Axboe 4090ed29b0b4SJens Axboe ret = -EOPNOTSUPP; 4091e5550a14SJens Axboe if (!io_is_uring_fops(f.file)) 4092ed29b0b4SJens Axboe goto out_fput; 4093ed29b0b4SJens Axboe 4094ed29b0b4SJens Axboe ctx = f.file->private_data; 4095ed29b0b4SJens Axboe 4096ed29b0b4SJens Axboe mutex_lock(&ctx->uring_lock); 4097ed29b0b4SJens Axboe ret = __io_uring_register(ctx, opcode, arg, nr_args); 4098ed29b0b4SJens Axboe mutex_unlock(&ctx->uring_lock); 4099ed29b0b4SJens Axboe trace_io_uring_register(ctx, opcode, ctx->nr_user_files, ctx->nr_user_bufs, ret); 4100ed29b0b4SJens Axboe out_fput: 4101ed29b0b4SJens Axboe fdput(f); 4102ed29b0b4SJens Axboe return ret; 4103ed29b0b4SJens Axboe } 4104ed29b0b4SJens Axboe 4105ed29b0b4SJens Axboe static int __init io_uring_init(void) 4106ed29b0b4SJens Axboe { 41079c71d39aSStefan Metzmacher #define __BUILD_BUG_VERIFY_OFFSET_SIZE(stype, eoffset, esize, ename) do { \ 4108ed29b0b4SJens Axboe BUILD_BUG_ON(offsetof(stype, ename) != eoffset); \ 41099c71d39aSStefan Metzmacher BUILD_BUG_ON(sizeof_field(stype, ename) != esize); \ 4110ed29b0b4SJens Axboe } while (0) 4111ed29b0b4SJens Axboe 4112ed29b0b4SJens Axboe #define BUILD_BUG_SQE_ELEM(eoffset, etype, ename) \ 41139c71d39aSStefan Metzmacher __BUILD_BUG_VERIFY_OFFSET_SIZE(struct io_uring_sqe, eoffset, sizeof(etype), ename) 41149c71d39aSStefan Metzmacher #define BUILD_BUG_SQE_ELEM_SIZE(eoffset, esize, ename) \ 41159c71d39aSStefan Metzmacher __BUILD_BUG_VERIFY_OFFSET_SIZE(struct io_uring_sqe, eoffset, esize, ename) 4116ed29b0b4SJens Axboe BUILD_BUG_ON(sizeof(struct io_uring_sqe) != 64); 4117ed29b0b4SJens Axboe BUILD_BUG_SQE_ELEM(0, __u8, opcode); 4118ed29b0b4SJens Axboe BUILD_BUG_SQE_ELEM(1, __u8, flags); 4119ed29b0b4SJens Axboe BUILD_BUG_SQE_ELEM(2, __u16, ioprio); 4120ed29b0b4SJens Axboe BUILD_BUG_SQE_ELEM(4, __s32, fd); 4121ed29b0b4SJens Axboe BUILD_BUG_SQE_ELEM(8, __u64, off); 4122ed29b0b4SJens Axboe BUILD_BUG_SQE_ELEM(8, __u64, addr2); 41239c71d39aSStefan Metzmacher BUILD_BUG_SQE_ELEM(8, __u32, cmd_op); 41249c71d39aSStefan Metzmacher BUILD_BUG_SQE_ELEM(12, __u32, __pad1); 4125ed29b0b4SJens Axboe BUILD_BUG_SQE_ELEM(16, __u64, addr); 4126ed29b0b4SJens Axboe BUILD_BUG_SQE_ELEM(16, __u64, splice_off_in); 4127ed29b0b4SJens Axboe BUILD_BUG_SQE_ELEM(24, __u32, len); 4128ed29b0b4SJens Axboe BUILD_BUG_SQE_ELEM(28, __kernel_rwf_t, rw_flags); 4129ed29b0b4SJens Axboe BUILD_BUG_SQE_ELEM(28, /* compat */ int, rw_flags); 4130ed29b0b4SJens Axboe BUILD_BUG_SQE_ELEM(28, /* compat */ __u32, rw_flags); 4131ed29b0b4SJens Axboe BUILD_BUG_SQE_ELEM(28, __u32, fsync_flags); 4132ed29b0b4SJens Axboe BUILD_BUG_SQE_ELEM(28, /* compat */ __u16, poll_events); 4133ed29b0b4SJens Axboe BUILD_BUG_SQE_ELEM(28, __u32, poll32_events); 4134ed29b0b4SJens Axboe BUILD_BUG_SQE_ELEM(28, __u32, sync_range_flags); 4135ed29b0b4SJens Axboe BUILD_BUG_SQE_ELEM(28, __u32, msg_flags); 4136ed29b0b4SJens Axboe BUILD_BUG_SQE_ELEM(28, __u32, timeout_flags); 4137ed29b0b4SJens Axboe BUILD_BUG_SQE_ELEM(28, __u32, accept_flags); 4138ed29b0b4SJens Axboe BUILD_BUG_SQE_ELEM(28, __u32, cancel_flags); 4139ed29b0b4SJens Axboe BUILD_BUG_SQE_ELEM(28, __u32, open_flags); 4140ed29b0b4SJens Axboe BUILD_BUG_SQE_ELEM(28, __u32, statx_flags); 4141ed29b0b4SJens Axboe BUILD_BUG_SQE_ELEM(28, __u32, fadvise_advice); 4142ed29b0b4SJens Axboe BUILD_BUG_SQE_ELEM(28, __u32, splice_flags); 41439c71d39aSStefan Metzmacher BUILD_BUG_SQE_ELEM(28, __u32, rename_flags); 41449c71d39aSStefan Metzmacher BUILD_BUG_SQE_ELEM(28, __u32, unlink_flags); 41459c71d39aSStefan Metzmacher BUILD_BUG_SQE_ELEM(28, __u32, hardlink_flags); 41469c71d39aSStefan Metzmacher BUILD_BUG_SQE_ELEM(28, __u32, xattr_flags); 41479c71d39aSStefan Metzmacher BUILD_BUG_SQE_ELEM(28, __u32, msg_ring_flags); 4148ed29b0b4SJens Axboe BUILD_BUG_SQE_ELEM(32, __u64, user_data); 4149ed29b0b4SJens Axboe BUILD_BUG_SQE_ELEM(40, __u16, buf_index); 4150ed29b0b4SJens Axboe BUILD_BUG_SQE_ELEM(40, __u16, buf_group); 4151ed29b0b4SJens Axboe BUILD_BUG_SQE_ELEM(42, __u16, personality); 4152ed29b0b4SJens Axboe BUILD_BUG_SQE_ELEM(44, __s32, splice_fd_in); 4153ed29b0b4SJens Axboe BUILD_BUG_SQE_ELEM(44, __u32, file_index); 4154b48c312bSPavel Begunkov BUILD_BUG_SQE_ELEM(44, __u16, addr_len); 4155b48c312bSPavel Begunkov BUILD_BUG_SQE_ELEM(46, __u16, __pad3[0]); 4156ed29b0b4SJens Axboe BUILD_BUG_SQE_ELEM(48, __u64, addr3); 41579c71d39aSStefan Metzmacher BUILD_BUG_SQE_ELEM_SIZE(48, 0, cmd); 41589c71d39aSStefan Metzmacher BUILD_BUG_SQE_ELEM(56, __u64, __pad2); 4159ed29b0b4SJens Axboe 4160ed29b0b4SJens Axboe BUILD_BUG_ON(sizeof(struct io_uring_files_update) != 4161ed29b0b4SJens Axboe sizeof(struct io_uring_rsrc_update)); 4162ed29b0b4SJens Axboe BUILD_BUG_ON(sizeof(struct io_uring_rsrc_update) > 4163ed29b0b4SJens Axboe sizeof(struct io_uring_rsrc_update2)); 4164ed29b0b4SJens Axboe 4165ed29b0b4SJens Axboe /* ->buf_index is u16 */ 4166ed29b0b4SJens Axboe BUILD_BUG_ON(offsetof(struct io_uring_buf_ring, bufs) != 0); 4167ed29b0b4SJens Axboe BUILD_BUG_ON(offsetof(struct io_uring_buf, resv) != 4168ed29b0b4SJens Axboe offsetof(struct io_uring_buf_ring, tail)); 4169ed29b0b4SJens Axboe 4170ed29b0b4SJens Axboe /* should fit into one byte */ 4171ed29b0b4SJens Axboe BUILD_BUG_ON(SQE_VALID_FLAGS >= (1 << 8)); 4172ed29b0b4SJens Axboe BUILD_BUG_ON(SQE_COMMON_FLAGS >= (1 << 8)); 4173ed29b0b4SJens Axboe BUILD_BUG_ON((SQE_VALID_FLAGS | SQE_COMMON_FLAGS) != SQE_VALID_FLAGS); 4174ed29b0b4SJens Axboe 4175ed29b0b4SJens Axboe BUILD_BUG_ON(__REQ_F_LAST_BIT > 8 * sizeof(int)); 4176ed29b0b4SJens Axboe 4177ed29b0b4SJens Axboe BUILD_BUG_ON(sizeof(atomic_t) != sizeof(u32)); 4178ed29b0b4SJens Axboe 4179d9b57aa3SJens Axboe io_uring_optable_init(); 4180ed29b0b4SJens Axboe 4181ed29b0b4SJens Axboe req_cachep = KMEM_CACHE(io_kiocb, SLAB_HWCACHE_ALIGN | SLAB_PANIC | 4182ed29b0b4SJens Axboe SLAB_ACCOUNT); 4183ed29b0b4SJens Axboe return 0; 4184ed29b0b4SJens Axboe }; 4185ed29b0b4SJens Axboe __initcall(io_uring_init); 4186