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 { 170ed29b0b4SJens Axboe if (!wq_list_empty(&ctx->submit_state.compl_reqs)) 171ed29b0b4SJens Axboe __io_submit_flush_completions(ctx); 172ed29b0b4SJens Axboe } 173ed29b0b4SJens Axboe 174faf88ddeSPavel Begunkov static inline unsigned int __io_cqring_events(struct io_ring_ctx *ctx) 175faf88ddeSPavel Begunkov { 176faf88ddeSPavel Begunkov return ctx->cached_cq_tail - READ_ONCE(ctx->rings->cq.head); 177faf88ddeSPavel Begunkov } 178faf88ddeSPavel Begunkov 1790fc8c2acSDylan Yudaken static inline unsigned int __io_cqring_events_user(struct io_ring_ctx *ctx) 1800fc8c2acSDylan Yudaken { 1810fc8c2acSDylan Yudaken return READ_ONCE(ctx->rings->cq.tail) - READ_ONCE(ctx->rings->cq.head); 1820fc8c2acSDylan Yudaken } 1830fc8c2acSDylan Yudaken 184ed29b0b4SJens Axboe static bool io_match_linked(struct io_kiocb *head) 185ed29b0b4SJens Axboe { 186ed29b0b4SJens Axboe struct io_kiocb *req; 187ed29b0b4SJens Axboe 188ed29b0b4SJens Axboe io_for_each_link(req, head) { 189ed29b0b4SJens Axboe if (req->flags & REQ_F_INFLIGHT) 190ed29b0b4SJens Axboe return true; 191ed29b0b4SJens Axboe } 192ed29b0b4SJens Axboe return false; 193ed29b0b4SJens Axboe } 194ed29b0b4SJens Axboe 195ed29b0b4SJens Axboe /* 196ed29b0b4SJens Axboe * As io_match_task() but protected against racing with linked timeouts. 197ed29b0b4SJens Axboe * User must not hold timeout_lock. 198ed29b0b4SJens Axboe */ 199329061d3SJens Axboe bool io_match_task_safe(struct io_kiocb *head, struct task_struct *task, 200ed29b0b4SJens Axboe bool cancel_all) 201ed29b0b4SJens Axboe { 202ed29b0b4SJens Axboe bool matched; 203ed29b0b4SJens Axboe 204ed29b0b4SJens Axboe if (task && head->task != task) 205ed29b0b4SJens Axboe return false; 206ed29b0b4SJens Axboe if (cancel_all) 207ed29b0b4SJens Axboe return true; 208ed29b0b4SJens Axboe 209ed29b0b4SJens Axboe if (head->flags & REQ_F_LINK_TIMEOUT) { 210ed29b0b4SJens Axboe struct io_ring_ctx *ctx = head->ctx; 211ed29b0b4SJens Axboe 212ed29b0b4SJens Axboe /* protect against races with linked timeouts */ 213ed29b0b4SJens Axboe spin_lock_irq(&ctx->timeout_lock); 214ed29b0b4SJens Axboe matched = io_match_linked(head); 215ed29b0b4SJens Axboe spin_unlock_irq(&ctx->timeout_lock); 216ed29b0b4SJens Axboe } else { 217ed29b0b4SJens Axboe matched = io_match_linked(head); 218ed29b0b4SJens Axboe } 219ed29b0b4SJens Axboe return matched; 220ed29b0b4SJens Axboe } 221ed29b0b4SJens Axboe 222ed29b0b4SJens Axboe static inline void req_fail_link_node(struct io_kiocb *req, int res) 223ed29b0b4SJens Axboe { 224ed29b0b4SJens Axboe req_set_fail(req); 22597b388d7SJens Axboe io_req_set_res(req, res, 0); 226ed29b0b4SJens Axboe } 227ed29b0b4SJens Axboe 228ed29b0b4SJens Axboe static inline void io_req_add_to_cache(struct io_kiocb *req, struct io_ring_ctx *ctx) 229ed29b0b4SJens Axboe { 230ed29b0b4SJens Axboe wq_stack_add_head(&req->comp_list, &ctx->submit_state.free_list); 231ed29b0b4SJens Axboe } 232ed29b0b4SJens Axboe 233ed29b0b4SJens Axboe static __cold void io_ring_ctx_ref_free(struct percpu_ref *ref) 234ed29b0b4SJens Axboe { 235ed29b0b4SJens Axboe struct io_ring_ctx *ctx = container_of(ref, struct io_ring_ctx, refs); 236ed29b0b4SJens Axboe 237ed29b0b4SJens Axboe complete(&ctx->ref_comp); 238ed29b0b4SJens Axboe } 239ed29b0b4SJens Axboe 240ed29b0b4SJens Axboe static __cold void io_fallback_req_func(struct work_struct *work) 241ed29b0b4SJens Axboe { 242ed29b0b4SJens Axboe struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx, 243ed29b0b4SJens Axboe fallback_work.work); 244ed29b0b4SJens Axboe struct llist_node *node = llist_del_all(&ctx->fallback_llist); 245ed29b0b4SJens Axboe struct io_kiocb *req, *tmp; 246ed29b0b4SJens Axboe bool locked = false; 247ed29b0b4SJens Axboe 248ed29b0b4SJens Axboe percpu_ref_get(&ctx->refs); 2493218e5d3SPavel Begunkov llist_for_each_entry_safe(req, tmp, node, io_task_work.node) 250ed29b0b4SJens Axboe req->io_task_work.func(req, &locked); 251ed29b0b4SJens Axboe 252ed29b0b4SJens Axboe if (locked) { 253ed29b0b4SJens Axboe io_submit_flush_completions(ctx); 254ed29b0b4SJens Axboe mutex_unlock(&ctx->uring_lock); 255ed29b0b4SJens Axboe } 256ed29b0b4SJens Axboe percpu_ref_put(&ctx->refs); 257ed29b0b4SJens Axboe } 258ed29b0b4SJens Axboe 259e6f89be6SPavel Begunkov static int io_alloc_hash_table(struct io_hash_table *table, unsigned bits) 260e6f89be6SPavel Begunkov { 261e6f89be6SPavel Begunkov unsigned hash_buckets = 1U << bits; 262e6f89be6SPavel Begunkov size_t hash_size = hash_buckets * sizeof(table->hbs[0]); 263e6f89be6SPavel Begunkov 264e6f89be6SPavel Begunkov table->hbs = kmalloc(hash_size, GFP_KERNEL); 265e6f89be6SPavel Begunkov if (!table->hbs) 266e6f89be6SPavel Begunkov return -ENOMEM; 267e6f89be6SPavel Begunkov 268e6f89be6SPavel Begunkov table->hash_bits = bits; 269e6f89be6SPavel Begunkov init_hash_table(table, hash_buckets); 270e6f89be6SPavel Begunkov return 0; 271e6f89be6SPavel Begunkov } 272e6f89be6SPavel Begunkov 273ed29b0b4SJens Axboe static __cold struct io_ring_ctx *io_ring_ctx_alloc(struct io_uring_params *p) 274ed29b0b4SJens Axboe { 275ed29b0b4SJens Axboe struct io_ring_ctx *ctx; 276ed29b0b4SJens Axboe int hash_bits; 277ed29b0b4SJens Axboe 278ed29b0b4SJens Axboe ctx = kzalloc(sizeof(*ctx), GFP_KERNEL); 279ed29b0b4SJens Axboe if (!ctx) 280ed29b0b4SJens Axboe return NULL; 281ed29b0b4SJens Axboe 282ed29b0b4SJens Axboe xa_init(&ctx->io_bl_xa); 283ed29b0b4SJens Axboe 284ed29b0b4SJens Axboe /* 285ed29b0b4SJens Axboe * Use 5 bits less than the max cq entries, that should give us around 2864a07723fSPavel Begunkov * 32 entries per hash list if totally full and uniformly spread, but 2874a07723fSPavel Begunkov * don't keep too many buckets to not overconsume memory. 288ed29b0b4SJens Axboe */ 2894a07723fSPavel Begunkov hash_bits = ilog2(p->cq_entries) - 5; 2904a07723fSPavel Begunkov hash_bits = clamp(hash_bits, 1, 8); 291e6f89be6SPavel Begunkov if (io_alloc_hash_table(&ctx->cancel_table, hash_bits)) 292ed29b0b4SJens Axboe goto err; 2939ca9fb24SPavel Begunkov if (io_alloc_hash_table(&ctx->cancel_table_locked, hash_bits)) 2949ca9fb24SPavel Begunkov goto err; 29538513c46SHao Xu 296ed29b0b4SJens Axboe ctx->dummy_ubuf = kzalloc(sizeof(*ctx->dummy_ubuf), GFP_KERNEL); 297ed29b0b4SJens Axboe if (!ctx->dummy_ubuf) 298ed29b0b4SJens Axboe goto err; 299ed29b0b4SJens Axboe /* set invalid range, so io_import_fixed() fails meeting it */ 300ed29b0b4SJens Axboe ctx->dummy_ubuf->ubuf = -1UL; 301ed29b0b4SJens Axboe 302ed29b0b4SJens Axboe if (percpu_ref_init(&ctx->refs, io_ring_ctx_ref_free, 30348904229SMichal Koutný 0, GFP_KERNEL)) 304ed29b0b4SJens Axboe goto err; 305ed29b0b4SJens Axboe 306ed29b0b4SJens Axboe ctx->flags = p->flags; 307ed29b0b4SJens Axboe init_waitqueue_head(&ctx->sqo_sq_wait); 308ed29b0b4SJens Axboe INIT_LIST_HEAD(&ctx->sqd_list); 309ed29b0b4SJens Axboe INIT_LIST_HEAD(&ctx->cq_overflow_list); 310ed29b0b4SJens Axboe INIT_LIST_HEAD(&ctx->io_buffers_cache); 3119b797a37SJens Axboe io_alloc_cache_init(&ctx->apoll_cache); 31243e0bbbdSJens Axboe io_alloc_cache_init(&ctx->netmsg_cache); 313ed29b0b4SJens Axboe init_completion(&ctx->ref_comp); 314ed29b0b4SJens Axboe xa_init_flags(&ctx->personalities, XA_FLAGS_ALLOC1); 315ed29b0b4SJens Axboe mutex_init(&ctx->uring_lock); 316ed29b0b4SJens Axboe init_waitqueue_head(&ctx->cq_wait); 317ed29b0b4SJens Axboe spin_lock_init(&ctx->completion_lock); 318ed29b0b4SJens Axboe spin_lock_init(&ctx->timeout_lock); 319ed29b0b4SJens Axboe INIT_WQ_LIST(&ctx->iopoll_list); 320ed29b0b4SJens Axboe INIT_LIST_HEAD(&ctx->io_buffers_pages); 321ed29b0b4SJens Axboe INIT_LIST_HEAD(&ctx->io_buffers_comp); 322ed29b0b4SJens Axboe INIT_LIST_HEAD(&ctx->defer_list); 323ed29b0b4SJens Axboe INIT_LIST_HEAD(&ctx->timeout_list); 324ed29b0b4SJens Axboe INIT_LIST_HEAD(&ctx->ltimeout_list); 325ed29b0b4SJens Axboe spin_lock_init(&ctx->rsrc_ref_lock); 326ed29b0b4SJens Axboe INIT_LIST_HEAD(&ctx->rsrc_ref_list); 327ed29b0b4SJens Axboe INIT_DELAYED_WORK(&ctx->rsrc_put_work, io_rsrc_put_work); 328ed29b0b4SJens Axboe init_llist_head(&ctx->rsrc_put_llist); 329c0e0d6baSDylan Yudaken init_llist_head(&ctx->work_llist); 330ed29b0b4SJens Axboe INIT_LIST_HEAD(&ctx->tctx_list); 331ed29b0b4SJens Axboe ctx->submit_state.free_list.next = NULL; 332ed29b0b4SJens Axboe INIT_WQ_LIST(&ctx->locked_free_list); 333ed29b0b4SJens Axboe INIT_DELAYED_WORK(&ctx->fallback_work, io_fallback_req_func); 334ed29b0b4SJens Axboe INIT_WQ_LIST(&ctx->submit_state.compl_reqs); 335ed29b0b4SJens Axboe return ctx; 336ed29b0b4SJens Axboe err: 337ed29b0b4SJens Axboe kfree(ctx->dummy_ubuf); 338e6f89be6SPavel Begunkov kfree(ctx->cancel_table.hbs); 3399ca9fb24SPavel Begunkov kfree(ctx->cancel_table_locked.hbs); 340ed29b0b4SJens Axboe kfree(ctx->io_bl); 341ed29b0b4SJens Axboe xa_destroy(&ctx->io_bl_xa); 342ed29b0b4SJens Axboe kfree(ctx); 343ed29b0b4SJens Axboe return NULL; 344ed29b0b4SJens Axboe } 345ed29b0b4SJens Axboe 346ed29b0b4SJens Axboe static void io_account_cq_overflow(struct io_ring_ctx *ctx) 347ed29b0b4SJens Axboe { 348ed29b0b4SJens Axboe struct io_rings *r = ctx->rings; 349ed29b0b4SJens Axboe 350ed29b0b4SJens Axboe WRITE_ONCE(r->cq_overflow, READ_ONCE(r->cq_overflow) + 1); 351ed29b0b4SJens Axboe ctx->cq_extra--; 352ed29b0b4SJens Axboe } 353ed29b0b4SJens Axboe 354ed29b0b4SJens Axboe static bool req_need_defer(struct io_kiocb *req, u32 seq) 355ed29b0b4SJens Axboe { 356ed29b0b4SJens Axboe if (unlikely(req->flags & REQ_F_IO_DRAIN)) { 357ed29b0b4SJens Axboe struct io_ring_ctx *ctx = req->ctx; 358ed29b0b4SJens Axboe 359ed29b0b4SJens Axboe return seq + READ_ONCE(ctx->cq_extra) != ctx->cached_cq_tail; 360ed29b0b4SJens Axboe } 361ed29b0b4SJens Axboe 362ed29b0b4SJens Axboe return false; 363ed29b0b4SJens Axboe } 364ed29b0b4SJens Axboe 365ed29b0b4SJens Axboe static inline void io_req_track_inflight(struct io_kiocb *req) 366ed29b0b4SJens Axboe { 367ed29b0b4SJens Axboe if (!(req->flags & REQ_F_INFLIGHT)) { 368ed29b0b4SJens Axboe req->flags |= REQ_F_INFLIGHT; 369ed29b0b4SJens Axboe atomic_inc(&req->task->io_uring->inflight_tracked); 370ed29b0b4SJens Axboe } 371ed29b0b4SJens Axboe } 372ed29b0b4SJens Axboe 373ed29b0b4SJens Axboe static struct io_kiocb *__io_prep_linked_timeout(struct io_kiocb *req) 374ed29b0b4SJens Axboe { 375ed29b0b4SJens Axboe if (WARN_ON_ONCE(!req->link)) 376ed29b0b4SJens Axboe return NULL; 377ed29b0b4SJens Axboe 378ed29b0b4SJens Axboe req->flags &= ~REQ_F_ARM_LTIMEOUT; 379ed29b0b4SJens Axboe req->flags |= REQ_F_LINK_TIMEOUT; 380ed29b0b4SJens Axboe 381ed29b0b4SJens Axboe /* linked timeouts should have two refs once prep'ed */ 382ed29b0b4SJens Axboe io_req_set_refcount(req); 383ed29b0b4SJens Axboe __io_req_set_refcount(req->link, 2); 384ed29b0b4SJens Axboe return req->link; 385ed29b0b4SJens Axboe } 386ed29b0b4SJens Axboe 387ed29b0b4SJens Axboe static inline struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req) 388ed29b0b4SJens Axboe { 389ed29b0b4SJens Axboe if (likely(!(req->flags & REQ_F_ARM_LTIMEOUT))) 390ed29b0b4SJens Axboe return NULL; 391ed29b0b4SJens Axboe return __io_prep_linked_timeout(req); 392ed29b0b4SJens Axboe } 393ed29b0b4SJens Axboe 394ed29b0b4SJens Axboe static noinline void __io_arm_ltimeout(struct io_kiocb *req) 395ed29b0b4SJens Axboe { 396ed29b0b4SJens Axboe io_queue_linked_timeout(__io_prep_linked_timeout(req)); 397ed29b0b4SJens Axboe } 398ed29b0b4SJens Axboe 399ed29b0b4SJens Axboe static inline void io_arm_ltimeout(struct io_kiocb *req) 400ed29b0b4SJens Axboe { 401ed29b0b4SJens Axboe if (unlikely(req->flags & REQ_F_ARM_LTIMEOUT)) 402ed29b0b4SJens Axboe __io_arm_ltimeout(req); 403ed29b0b4SJens Axboe } 404ed29b0b4SJens Axboe 405ed29b0b4SJens Axboe static void io_prep_async_work(struct io_kiocb *req) 406ed29b0b4SJens Axboe { 407ed29b0b4SJens Axboe const struct io_op_def *def = &io_op_defs[req->opcode]; 408ed29b0b4SJens Axboe struct io_ring_ctx *ctx = req->ctx; 409ed29b0b4SJens Axboe 410ed29b0b4SJens Axboe if (!(req->flags & REQ_F_CREDS)) { 411ed29b0b4SJens Axboe req->flags |= REQ_F_CREDS; 412ed29b0b4SJens Axboe req->creds = get_current_cred(); 413ed29b0b4SJens Axboe } 414ed29b0b4SJens Axboe 415ed29b0b4SJens Axboe req->work.list.next = NULL; 416ed29b0b4SJens Axboe req->work.flags = 0; 417ed29b0b4SJens Axboe req->work.cancel_seq = atomic_read(&ctx->cancel_seq); 418ed29b0b4SJens Axboe if (req->flags & REQ_F_FORCE_ASYNC) 419ed29b0b4SJens Axboe req->work.flags |= IO_WQ_WORK_CONCURRENT; 420ed29b0b4SJens Axboe 421f6b543fdSJens Axboe if (req->file && !io_req_ffs_set(req)) 422f6b543fdSJens Axboe req->flags |= io_file_get_flags(req->file) << REQ_F_SUPPORT_NOWAIT_BIT; 423f6b543fdSJens Axboe 424ed29b0b4SJens Axboe if (req->flags & REQ_F_ISREG) { 425ed29b0b4SJens Axboe if (def->hash_reg_file || (ctx->flags & IORING_SETUP_IOPOLL)) 426ed29b0b4SJens Axboe io_wq_hash_work(&req->work, file_inode(req->file)); 427ed29b0b4SJens Axboe } else if (!req->file || !S_ISBLK(file_inode(req->file)->i_mode)) { 428ed29b0b4SJens Axboe if (def->unbound_nonreg_file) 429ed29b0b4SJens Axboe req->work.flags |= IO_WQ_WORK_UNBOUND; 430ed29b0b4SJens Axboe } 431ed29b0b4SJens Axboe } 432ed29b0b4SJens Axboe 433ed29b0b4SJens Axboe static void io_prep_async_link(struct io_kiocb *req) 434ed29b0b4SJens Axboe { 435ed29b0b4SJens Axboe struct io_kiocb *cur; 436ed29b0b4SJens Axboe 437ed29b0b4SJens Axboe if (req->flags & REQ_F_LINK_TIMEOUT) { 438ed29b0b4SJens Axboe struct io_ring_ctx *ctx = req->ctx; 439ed29b0b4SJens Axboe 440ed29b0b4SJens Axboe spin_lock_irq(&ctx->timeout_lock); 441ed29b0b4SJens Axboe io_for_each_link(cur, req) 442ed29b0b4SJens Axboe io_prep_async_work(cur); 443ed29b0b4SJens Axboe spin_unlock_irq(&ctx->timeout_lock); 444ed29b0b4SJens Axboe } else { 445ed29b0b4SJens Axboe io_for_each_link(cur, req) 446ed29b0b4SJens Axboe io_prep_async_work(cur); 447ed29b0b4SJens Axboe } 448ed29b0b4SJens Axboe } 449ed29b0b4SJens Axboe 450f3b44f92SJens Axboe void io_queue_iowq(struct io_kiocb *req, bool *dont_use) 451ed29b0b4SJens Axboe { 452ed29b0b4SJens Axboe struct io_kiocb *link = io_prep_linked_timeout(req); 453ed29b0b4SJens Axboe struct io_uring_task *tctx = req->task->io_uring; 454ed29b0b4SJens Axboe 455ed29b0b4SJens Axboe BUG_ON(!tctx); 456ed29b0b4SJens Axboe BUG_ON(!tctx->io_wq); 457ed29b0b4SJens Axboe 458ed29b0b4SJens Axboe /* init ->work of the whole link before punting */ 459ed29b0b4SJens Axboe io_prep_async_link(req); 460ed29b0b4SJens Axboe 461ed29b0b4SJens Axboe /* 462ed29b0b4SJens Axboe * Not expected to happen, but if we do have a bug where this _can_ 463ed29b0b4SJens Axboe * happen, catch it here and ensure the request is marked as 464ed29b0b4SJens Axboe * canceled. That will make io-wq go through the usual work cancel 465ed29b0b4SJens Axboe * procedure rather than attempt to run this request (or create a new 466ed29b0b4SJens Axboe * worker for it). 467ed29b0b4SJens Axboe */ 468ed29b0b4SJens Axboe if (WARN_ON_ONCE(!same_thread_group(req->task, current))) 469ed29b0b4SJens Axboe req->work.flags |= IO_WQ_WORK_CANCEL; 470ed29b0b4SJens Axboe 47148863ffdSPavel Begunkov trace_io_uring_queue_async_work(req, io_wq_is_hashed(&req->work)); 472ed29b0b4SJens Axboe io_wq_enqueue(tctx->io_wq, &req->work); 473ed29b0b4SJens Axboe if (link) 474ed29b0b4SJens Axboe io_queue_linked_timeout(link); 475ed29b0b4SJens Axboe } 476ed29b0b4SJens Axboe 477ed29b0b4SJens Axboe static __cold void io_queue_deferred(struct io_ring_ctx *ctx) 478ed29b0b4SJens Axboe { 479ed29b0b4SJens Axboe while (!list_empty(&ctx->defer_list)) { 480ed29b0b4SJens Axboe struct io_defer_entry *de = list_first_entry(&ctx->defer_list, 481ed29b0b4SJens Axboe struct io_defer_entry, list); 482ed29b0b4SJens Axboe 483ed29b0b4SJens Axboe if (req_need_defer(de->req, de->seq)) 484ed29b0b4SJens Axboe break; 485ed29b0b4SJens Axboe list_del_init(&de->list); 486ed29b0b4SJens Axboe io_req_task_queue(de->req); 487ed29b0b4SJens Axboe kfree(de); 488ed29b0b4SJens Axboe } 489ed29b0b4SJens Axboe } 490ed29b0b4SJens Axboe 49121a091b9SDylan Yudaken 49221a091b9SDylan Yudaken static void io_eventfd_ops(struct rcu_head *rcu) 493d8e9214fSDylan Yudaken { 494d8e9214fSDylan Yudaken struct io_ev_fd *ev_fd = container_of(rcu, struct io_ev_fd, rcu); 49521a091b9SDylan Yudaken int ops = atomic_xchg(&ev_fd->ops, 0); 496d8e9214fSDylan Yudaken 49721a091b9SDylan Yudaken if (ops & BIT(IO_EVENTFD_OP_SIGNAL_BIT)) 49844648532SJens Axboe eventfd_signal_mask(ev_fd->cq_ev_fd, 1, EPOLL_URING_WAKE); 49921a091b9SDylan Yudaken 50021a091b9SDylan Yudaken /* IO_EVENTFD_OP_FREE_BIT may not be set here depending on callback 50121a091b9SDylan Yudaken * ordering in a race but if references are 0 we know we have to free 50221a091b9SDylan Yudaken * it regardless. 50321a091b9SDylan Yudaken */ 50421a091b9SDylan Yudaken if (atomic_dec_and_test(&ev_fd->refs)) { 505d8e9214fSDylan Yudaken eventfd_ctx_put(ev_fd->cq_ev_fd); 506d8e9214fSDylan Yudaken kfree(ev_fd); 507d8e9214fSDylan Yudaken } 50821a091b9SDylan Yudaken } 509d8e9214fSDylan Yudaken 510ed29b0b4SJens Axboe static void io_eventfd_signal(struct io_ring_ctx *ctx) 511ed29b0b4SJens Axboe { 51221a091b9SDylan Yudaken struct io_ev_fd *ev_fd = NULL; 513ed29b0b4SJens Axboe 514ed29b0b4SJens Axboe rcu_read_lock(); 515ed29b0b4SJens Axboe /* 516ed29b0b4SJens Axboe * rcu_dereference ctx->io_ev_fd once and use it for both for checking 517ed29b0b4SJens Axboe * and eventfd_signal 518ed29b0b4SJens Axboe */ 519ed29b0b4SJens Axboe ev_fd = rcu_dereference(ctx->io_ev_fd); 520ed29b0b4SJens Axboe 521ed29b0b4SJens Axboe /* 522ed29b0b4SJens Axboe * Check again if ev_fd exists incase an io_eventfd_unregister call 523ed29b0b4SJens Axboe * completed between the NULL check of ctx->io_ev_fd at the start of 524ed29b0b4SJens Axboe * the function and rcu_read_lock. 525ed29b0b4SJens Axboe */ 526ed29b0b4SJens Axboe if (unlikely(!ev_fd)) 527ed29b0b4SJens Axboe goto out; 528ed29b0b4SJens Axboe if (READ_ONCE(ctx->rings->cq_flags) & IORING_CQ_EVENTFD_DISABLED) 529ed29b0b4SJens Axboe goto out; 53021a091b9SDylan Yudaken if (ev_fd->eventfd_async && !io_wq_current_is_worker()) 53121a091b9SDylan Yudaken goto out; 532ed29b0b4SJens Axboe 53321a091b9SDylan Yudaken if (likely(eventfd_signal_allowed())) { 53444648532SJens Axboe eventfd_signal_mask(ev_fd->cq_ev_fd, 1, EPOLL_URING_WAKE); 53521a091b9SDylan Yudaken } else { 53621a091b9SDylan Yudaken atomic_inc(&ev_fd->refs); 53721a091b9SDylan Yudaken if (!atomic_fetch_or(BIT(IO_EVENTFD_OP_SIGNAL_BIT), &ev_fd->ops)) 53821a091b9SDylan Yudaken call_rcu(&ev_fd->rcu, io_eventfd_ops); 53921a091b9SDylan Yudaken else 54021a091b9SDylan Yudaken atomic_dec(&ev_fd->refs); 54121a091b9SDylan Yudaken } 54221a091b9SDylan Yudaken 543ed29b0b4SJens Axboe out: 544ed29b0b4SJens Axboe rcu_read_unlock(); 545ed29b0b4SJens Axboe } 546ed29b0b4SJens Axboe 54721a091b9SDylan Yudaken static void io_eventfd_flush_signal(struct io_ring_ctx *ctx) 54821a091b9SDylan Yudaken { 54921a091b9SDylan Yudaken bool skip; 55021a091b9SDylan Yudaken 55121a091b9SDylan Yudaken spin_lock(&ctx->completion_lock); 55221a091b9SDylan Yudaken 55321a091b9SDylan Yudaken /* 55421a091b9SDylan Yudaken * Eventfd should only get triggered when at least one event has been 55521a091b9SDylan Yudaken * posted. Some applications rely on the eventfd notification count 55621a091b9SDylan Yudaken * only changing IFF a new CQE has been added to the CQ ring. There's 55721a091b9SDylan Yudaken * no depedency on 1:1 relationship between how many times this 55821a091b9SDylan Yudaken * function is called (and hence the eventfd count) and number of CQEs 55921a091b9SDylan Yudaken * posted to the CQ ring. 56021a091b9SDylan Yudaken */ 56121a091b9SDylan Yudaken skip = ctx->cached_cq_tail == ctx->evfd_last_cq_tail; 56221a091b9SDylan Yudaken ctx->evfd_last_cq_tail = ctx->cached_cq_tail; 56321a091b9SDylan Yudaken spin_unlock(&ctx->completion_lock); 56421a091b9SDylan Yudaken if (skip) 56521a091b9SDylan Yudaken return; 56621a091b9SDylan Yudaken 56721a091b9SDylan Yudaken io_eventfd_signal(ctx); 56821a091b9SDylan Yudaken } 56921a091b9SDylan Yudaken 570a830ffd2SPavel Begunkov void __io_commit_cqring_flush(struct io_ring_ctx *ctx) 571a830ffd2SPavel Begunkov { 572a830ffd2SPavel Begunkov if (ctx->off_timeout_used || ctx->drain_active) { 573a830ffd2SPavel Begunkov spin_lock(&ctx->completion_lock); 574a830ffd2SPavel Begunkov if (ctx->off_timeout_used) 575a830ffd2SPavel Begunkov io_flush_timeouts(ctx); 576a830ffd2SPavel Begunkov if (ctx->drain_active) 577a830ffd2SPavel Begunkov io_queue_deferred(ctx); 578a830ffd2SPavel Begunkov spin_unlock(&ctx->completion_lock); 579a830ffd2SPavel Begunkov } 580a830ffd2SPavel Begunkov if (ctx->has_evfd) 58121a091b9SDylan Yudaken io_eventfd_flush_signal(ctx); 582a830ffd2SPavel Begunkov } 583a830ffd2SPavel Begunkov 5846c16fe3cSJens Axboe void io_cq_unlock_post(struct io_ring_ctx *ctx) 58525399321SPavel Begunkov __releases(ctx->completion_lock) 58625399321SPavel Begunkov { 58725399321SPavel Begunkov io_commit_cqring(ctx); 58825399321SPavel Begunkov spin_unlock(&ctx->completion_lock); 58925399321SPavel Begunkov 5906c16fe3cSJens Axboe io_commit_cqring_flush(ctx); 5916c16fe3cSJens Axboe io_cqring_wake(ctx); 59225399321SPavel Begunkov } 59325399321SPavel Begunkov 594ed29b0b4SJens Axboe /* Returns true if there are no backlogged entries after the flush */ 595ed29b0b4SJens Axboe static bool __io_cqring_overflow_flush(struct io_ring_ctx *ctx, bool force) 596ed29b0b4SJens Axboe { 597305bef98SPavel Begunkov bool all_flushed; 598ed29b0b4SJens Axboe size_t cqe_size = sizeof(struct io_uring_cqe); 599ed29b0b4SJens Axboe 600ed29b0b4SJens Axboe if (!force && __io_cqring_events(ctx) == ctx->cq_entries) 601ed29b0b4SJens Axboe return false; 602ed29b0b4SJens Axboe 603ed29b0b4SJens Axboe if (ctx->flags & IORING_SETUP_CQE32) 604ed29b0b4SJens Axboe cqe_size <<= 1; 605ed29b0b4SJens Axboe 60625399321SPavel Begunkov io_cq_lock(ctx); 607ed29b0b4SJens Axboe while (!list_empty(&ctx->cq_overflow_list)) { 608aa1df3a3SPavel Begunkov struct io_uring_cqe *cqe = io_get_cqe_overflow(ctx, true); 609ed29b0b4SJens Axboe struct io_overflow_cqe *ocqe; 610ed29b0b4SJens Axboe 611ed29b0b4SJens Axboe if (!cqe && !force) 612ed29b0b4SJens Axboe break; 613ed29b0b4SJens Axboe ocqe = list_first_entry(&ctx->cq_overflow_list, 614ed29b0b4SJens Axboe struct io_overflow_cqe, list); 615ed29b0b4SJens Axboe if (cqe) 616ed29b0b4SJens Axboe memcpy(cqe, &ocqe->cqe, cqe_size); 617ed29b0b4SJens Axboe else 618ed29b0b4SJens Axboe io_account_cq_overflow(ctx); 619ed29b0b4SJens Axboe 620ed29b0b4SJens Axboe list_del(&ocqe->list); 621ed29b0b4SJens Axboe kfree(ocqe); 622ed29b0b4SJens Axboe } 623ed29b0b4SJens Axboe 624ed29b0b4SJens Axboe all_flushed = list_empty(&ctx->cq_overflow_list); 625ed29b0b4SJens Axboe if (all_flushed) { 626ed29b0b4SJens Axboe clear_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq); 627ed29b0b4SJens Axboe atomic_andnot(IORING_SQ_CQ_OVERFLOW, &ctx->rings->sq_flags); 628ed29b0b4SJens Axboe } 629ed29b0b4SJens Axboe 63025399321SPavel Begunkov io_cq_unlock_post(ctx); 631ed29b0b4SJens Axboe return all_flushed; 632ed29b0b4SJens Axboe } 633ed29b0b4SJens Axboe 634ed29b0b4SJens Axboe static bool io_cqring_overflow_flush(struct io_ring_ctx *ctx) 635ed29b0b4SJens Axboe { 636ed29b0b4SJens Axboe bool ret = true; 637ed29b0b4SJens Axboe 638ed29b0b4SJens Axboe if (test_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq)) { 639ed29b0b4SJens Axboe /* iopoll syncs against uring_lock, not completion_lock */ 640ed29b0b4SJens Axboe if (ctx->flags & IORING_SETUP_IOPOLL) 641ed29b0b4SJens Axboe mutex_lock(&ctx->uring_lock); 642ed29b0b4SJens Axboe ret = __io_cqring_overflow_flush(ctx, false); 643ed29b0b4SJens Axboe if (ctx->flags & IORING_SETUP_IOPOLL) 644ed29b0b4SJens Axboe mutex_unlock(&ctx->uring_lock); 645ed29b0b4SJens Axboe } 646ed29b0b4SJens Axboe 647ed29b0b4SJens Axboe return ret; 648ed29b0b4SJens Axboe } 649ed29b0b4SJens Axboe 650e70cb608SPavel Begunkov void __io_put_task(struct task_struct *task, int nr) 651ed29b0b4SJens Axboe { 652ed29b0b4SJens Axboe struct io_uring_task *tctx = task->io_uring; 653ed29b0b4SJens Axboe 654ed29b0b4SJens Axboe percpu_counter_sub(&tctx->inflight, nr); 655ed29b0b4SJens Axboe if (unlikely(atomic_read(&tctx->in_idle))) 656ed29b0b4SJens Axboe wake_up(&tctx->wait); 657ed29b0b4SJens Axboe put_task_struct_many(task, nr); 658ed29b0b4SJens Axboe } 659ed29b0b4SJens Axboe 66063809137SPavel Begunkov void io_task_refs_refill(struct io_uring_task *tctx) 661ed29b0b4SJens Axboe { 662ed29b0b4SJens Axboe unsigned int refill = -tctx->cached_refs + IO_TCTX_REFS_CACHE_NR; 663ed29b0b4SJens Axboe 664ed29b0b4SJens Axboe percpu_counter_add(&tctx->inflight, refill); 665ed29b0b4SJens Axboe refcount_add(refill, ¤t->usage); 666ed29b0b4SJens Axboe tctx->cached_refs += refill; 667ed29b0b4SJens Axboe } 668ed29b0b4SJens Axboe 669ed29b0b4SJens Axboe static __cold void io_uring_drop_tctx_refs(struct task_struct *task) 670ed29b0b4SJens Axboe { 671ed29b0b4SJens Axboe struct io_uring_task *tctx = task->io_uring; 672ed29b0b4SJens Axboe unsigned int refs = tctx->cached_refs; 673ed29b0b4SJens Axboe 674ed29b0b4SJens Axboe if (refs) { 675ed29b0b4SJens Axboe tctx->cached_refs = 0; 676ed29b0b4SJens Axboe percpu_counter_sub(&tctx->inflight, refs); 677ed29b0b4SJens Axboe put_task_struct_many(task, refs); 678ed29b0b4SJens Axboe } 679ed29b0b4SJens Axboe } 680ed29b0b4SJens Axboe 68168494a65SPavel Begunkov static bool io_cqring_event_overflow(struct io_ring_ctx *ctx, u64 user_data, 68268494a65SPavel Begunkov s32 res, u32 cflags, u64 extra1, u64 extra2) 683ed29b0b4SJens Axboe { 684ed29b0b4SJens Axboe struct io_overflow_cqe *ocqe; 685ed29b0b4SJens Axboe size_t ocq_size = sizeof(struct io_overflow_cqe); 686ed29b0b4SJens Axboe bool is_cqe32 = (ctx->flags & IORING_SETUP_CQE32); 687ed29b0b4SJens Axboe 688ed29b0b4SJens Axboe if (is_cqe32) 689ed29b0b4SJens Axboe ocq_size += sizeof(struct io_uring_cqe); 690ed29b0b4SJens Axboe 691ed29b0b4SJens Axboe ocqe = kmalloc(ocq_size, GFP_ATOMIC | __GFP_ACCOUNT); 692ed29b0b4SJens Axboe trace_io_uring_cqe_overflow(ctx, user_data, res, cflags, ocqe); 693ed29b0b4SJens Axboe if (!ocqe) { 694ed29b0b4SJens Axboe /* 695ed29b0b4SJens Axboe * If we're in ring overflow flush mode, or in task cancel mode, 696ed29b0b4SJens Axboe * or cannot allocate an overflow entry, then we need to drop it 697ed29b0b4SJens Axboe * on the floor. 698ed29b0b4SJens Axboe */ 699ed29b0b4SJens Axboe io_account_cq_overflow(ctx); 700ed29b0b4SJens Axboe set_bit(IO_CHECK_CQ_DROPPED_BIT, &ctx->check_cq); 701ed29b0b4SJens Axboe return false; 702ed29b0b4SJens Axboe } 703ed29b0b4SJens Axboe if (list_empty(&ctx->cq_overflow_list)) { 704ed29b0b4SJens Axboe set_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq); 705ed29b0b4SJens Axboe atomic_or(IORING_SQ_CQ_OVERFLOW, &ctx->rings->sq_flags); 706ed29b0b4SJens Axboe 707ed29b0b4SJens Axboe } 708ed29b0b4SJens Axboe ocqe->cqe.user_data = user_data; 709ed29b0b4SJens Axboe ocqe->cqe.res = res; 710ed29b0b4SJens Axboe ocqe->cqe.flags = cflags; 711ed29b0b4SJens Axboe if (is_cqe32) { 712ed29b0b4SJens Axboe ocqe->cqe.big_cqe[0] = extra1; 713ed29b0b4SJens Axboe ocqe->cqe.big_cqe[1] = extra2; 714ed29b0b4SJens Axboe } 715ed29b0b4SJens Axboe list_add_tail(&ocqe->list, &ctx->cq_overflow_list); 716ed29b0b4SJens Axboe return true; 717ed29b0b4SJens Axboe } 718ed29b0b4SJens Axboe 71968494a65SPavel Begunkov bool io_req_cqe_overflow(struct io_kiocb *req) 72068494a65SPavel Begunkov { 72168494a65SPavel Begunkov if (!(req->flags & REQ_F_CQE32_INIT)) { 72268494a65SPavel Begunkov req->extra1 = 0; 72368494a65SPavel Begunkov req->extra2 = 0; 72468494a65SPavel Begunkov } 72568494a65SPavel Begunkov return io_cqring_event_overflow(req->ctx, req->cqe.user_data, 72668494a65SPavel Begunkov req->cqe.res, req->cqe.flags, 72768494a65SPavel Begunkov req->extra1, req->extra2); 72868494a65SPavel Begunkov } 72968494a65SPavel Begunkov 730faf88ddeSPavel Begunkov /* 731faf88ddeSPavel Begunkov * writes to the cq entry need to come after reading head; the 732faf88ddeSPavel Begunkov * control dependency is enough as we're using WRITE_ONCE to 733faf88ddeSPavel Begunkov * fill the cq entry 734faf88ddeSPavel Begunkov */ 735aa1df3a3SPavel Begunkov struct io_uring_cqe *__io_get_cqe(struct io_ring_ctx *ctx, bool overflow) 736faf88ddeSPavel Begunkov { 737faf88ddeSPavel Begunkov struct io_rings *rings = ctx->rings; 738faf88ddeSPavel Begunkov unsigned int off = ctx->cached_cq_tail & (ctx->cq_entries - 1); 739faf88ddeSPavel Begunkov unsigned int free, queued, len; 740faf88ddeSPavel Begunkov 741aa1df3a3SPavel Begunkov /* 742aa1df3a3SPavel Begunkov * Posting into the CQ when there are pending overflowed CQEs may break 743aa1df3a3SPavel Begunkov * ordering guarantees, which will affect links, F_MORE users and more. 744aa1df3a3SPavel Begunkov * Force overflow the completion. 745aa1df3a3SPavel Begunkov */ 746aa1df3a3SPavel Begunkov if (!overflow && (ctx->check_cq & BIT(IO_CHECK_CQ_OVERFLOW_BIT))) 747aa1df3a3SPavel Begunkov return NULL; 748faf88ddeSPavel Begunkov 749faf88ddeSPavel Begunkov /* userspace may cheat modifying the tail, be safe and do min */ 750faf88ddeSPavel Begunkov queued = min(__io_cqring_events(ctx), ctx->cq_entries); 751faf88ddeSPavel Begunkov free = ctx->cq_entries - queued; 752faf88ddeSPavel Begunkov /* we need a contiguous range, limit based on the current array offset */ 753faf88ddeSPavel Begunkov len = min(free, ctx->cq_entries - off); 754faf88ddeSPavel Begunkov if (!len) 755faf88ddeSPavel Begunkov return NULL; 756faf88ddeSPavel Begunkov 757b3659a65SPavel Begunkov if (ctx->flags & IORING_SETUP_CQE32) { 758b3659a65SPavel Begunkov off <<= 1; 759b3659a65SPavel Begunkov len <<= 1; 760b3659a65SPavel Begunkov } 761b3659a65SPavel Begunkov 762faf88ddeSPavel Begunkov ctx->cqe_cached = &rings->cqes[off]; 763faf88ddeSPavel Begunkov ctx->cqe_sentinel = ctx->cqe_cached + len; 764b3659a65SPavel Begunkov 765b3659a65SPavel Begunkov ctx->cached_cq_tail++; 766faf88ddeSPavel Begunkov ctx->cqe_cached++; 767b3659a65SPavel Begunkov if (ctx->flags & IORING_SETUP_CQE32) 768b3659a65SPavel Begunkov ctx->cqe_cached++; 769b3659a65SPavel Begunkov return &rings->cqes[off]; 770faf88ddeSPavel Begunkov } 771faf88ddeSPavel Begunkov 772eb42cebbSPavel Begunkov bool io_fill_cqe_aux(struct io_ring_ctx *ctx, u64 user_data, s32 res, u32 cflags, 77352120f0fSDylan Yudaken bool allow_overflow) 774ed29b0b4SJens Axboe { 775ed29b0b4SJens Axboe struct io_uring_cqe *cqe; 776ed29b0b4SJens Axboe 777ed29b0b4SJens Axboe ctx->cq_extra++; 778ed29b0b4SJens Axboe 779ed29b0b4SJens Axboe /* 780ed29b0b4SJens Axboe * If we can't get a cq entry, userspace overflowed the 781ed29b0b4SJens Axboe * submission (by quite a lot). Increment the overflow count in 782ed29b0b4SJens Axboe * the ring. 783ed29b0b4SJens Axboe */ 784ed29b0b4SJens Axboe cqe = io_get_cqe(ctx); 785ed29b0b4SJens Axboe if (likely(cqe)) { 786e0486f3fSDylan Yudaken trace_io_uring_complete(ctx, NULL, user_data, res, cflags, 0, 0); 787e0486f3fSDylan Yudaken 788ed29b0b4SJens Axboe WRITE_ONCE(cqe->user_data, user_data); 789ed29b0b4SJens Axboe WRITE_ONCE(cqe->res, res); 790ed29b0b4SJens Axboe WRITE_ONCE(cqe->flags, cflags); 791ed29b0b4SJens Axboe 792ed29b0b4SJens Axboe if (ctx->flags & IORING_SETUP_CQE32) { 793ed29b0b4SJens Axboe WRITE_ONCE(cqe->big_cqe[0], 0); 794ed29b0b4SJens Axboe WRITE_ONCE(cqe->big_cqe[1], 0); 795ed29b0b4SJens Axboe } 796ed29b0b4SJens Axboe return true; 797ed29b0b4SJens Axboe } 79852120f0fSDylan Yudaken 79952120f0fSDylan Yudaken if (allow_overflow) 800ed29b0b4SJens Axboe return io_cqring_event_overflow(ctx, user_data, res, cflags, 0, 0); 80152120f0fSDylan Yudaken 80252120f0fSDylan Yudaken return false; 803ed29b0b4SJens Axboe } 804ed29b0b4SJens Axboe 805d245bca6SPavel Begunkov bool io_post_aux_cqe(struct io_ring_ctx *ctx, 80652120f0fSDylan Yudaken u64 user_data, s32 res, u32 cflags, 80752120f0fSDylan Yudaken bool allow_overflow) 808d245bca6SPavel Begunkov { 809d245bca6SPavel Begunkov bool filled; 810d245bca6SPavel Begunkov 81125399321SPavel Begunkov io_cq_lock(ctx); 81252120f0fSDylan Yudaken filled = io_fill_cqe_aux(ctx, user_data, res, cflags, allow_overflow); 81325399321SPavel Begunkov io_cq_unlock_post(ctx); 814d245bca6SPavel Begunkov return filled; 815d245bca6SPavel Begunkov } 816d245bca6SPavel Begunkov 817*fa18fa22SPavel Begunkov void io_req_complete_post(struct io_kiocb *req) 818ed29b0b4SJens Axboe { 819*fa18fa22SPavel Begunkov struct io_ring_ctx *ctx = req->ctx; 820*fa18fa22SPavel Begunkov 821*fa18fa22SPavel Begunkov io_cq_lock(ctx); 822*fa18fa22SPavel Begunkov if (!(req->flags & REQ_F_CQE_SKIP)) 823*fa18fa22SPavel Begunkov __io_fill_cqe_req(ctx, req); 824*fa18fa22SPavel Begunkov 825ed29b0b4SJens Axboe /* 826ed29b0b4SJens Axboe * If we're the last reference to this request, add to our locked 827ed29b0b4SJens Axboe * free_list cache. 828ed29b0b4SJens Axboe */ 829ed29b0b4SJens Axboe if (req_ref_put_and_test(req)) { 830ed29b0b4SJens Axboe if (req->flags & IO_REQ_LINK_FLAGS) { 831ed29b0b4SJens Axboe if (req->flags & IO_DISARM_MASK) 832ed29b0b4SJens Axboe io_disarm_next(req); 833ed29b0b4SJens Axboe if (req->link) { 834ed29b0b4SJens Axboe io_req_task_queue(req->link); 835ed29b0b4SJens Axboe req->link = NULL; 836ed29b0b4SJens Axboe } 837ed29b0b4SJens Axboe } 838ed29b0b4SJens Axboe io_req_put_rsrc(req); 839ed29b0b4SJens Axboe /* 840ed29b0b4SJens Axboe * Selected buffer deallocation in io_clean_op() assumes that 841ed29b0b4SJens Axboe * we don't hold ->completion_lock. Clean them here to avoid 842ed29b0b4SJens Axboe * deadlocks. 843ed29b0b4SJens Axboe */ 844ed29b0b4SJens Axboe io_put_kbuf_comp(req); 845ed29b0b4SJens Axboe io_dismantle_req(req); 846ed29b0b4SJens Axboe io_put_task(req->task, 1); 847ed29b0b4SJens Axboe wq_list_add_head(&req->comp_list, &ctx->locked_free_list); 848ed29b0b4SJens Axboe ctx->locked_free_nr++; 849ed29b0b4SJens Axboe } 85025399321SPavel Begunkov io_cq_unlock_post(ctx); 851ed29b0b4SJens Axboe } 852ed29b0b4SJens Axboe 85399f15d8dSJens Axboe inline void __io_req_complete(struct io_kiocb *req, unsigned issue_flags) 854ed29b0b4SJens Axboe { 85597b388d7SJens Axboe io_req_complete_post(req); 856ed29b0b4SJens Axboe } 857ed29b0b4SJens Axboe 858329061d3SJens Axboe void io_req_complete_failed(struct io_kiocb *req, s32 res) 859e276ae34SPavel Begunkov __must_hold(&ctx->uring_lock) 860ed29b0b4SJens Axboe { 861a47b255eSPavel Begunkov const struct io_op_def *def = &io_op_defs[req->opcode]; 862a47b255eSPavel Begunkov 863e276ae34SPavel Begunkov lockdep_assert_held(&req->ctx->uring_lock); 864e276ae34SPavel Begunkov 865ed29b0b4SJens Axboe req_set_fail(req); 86697b388d7SJens Axboe io_req_set_res(req, res, io_put_kbuf(req, IO_URING_F_UNLOCKED)); 867a47b255eSPavel Begunkov if (def->fail) 868a47b255eSPavel Begunkov def->fail(req); 86997b388d7SJens Axboe io_req_complete_post(req); 870ed29b0b4SJens Axboe } 871ed29b0b4SJens Axboe 872ed29b0b4SJens Axboe /* 873ed29b0b4SJens Axboe * Don't initialise the fields below on every allocation, but do that in 874ed29b0b4SJens Axboe * advance and keep them valid across allocations. 875ed29b0b4SJens Axboe */ 876ed29b0b4SJens Axboe static void io_preinit_req(struct io_kiocb *req, struct io_ring_ctx *ctx) 877ed29b0b4SJens Axboe { 878ed29b0b4SJens Axboe req->ctx = ctx; 879ed29b0b4SJens Axboe req->link = NULL; 880ed29b0b4SJens Axboe req->async_data = NULL; 881ed29b0b4SJens Axboe /* not necessary, but safer to zero */ 882ed29b0b4SJens Axboe req->cqe.res = 0; 883ed29b0b4SJens Axboe } 884ed29b0b4SJens Axboe 885ed29b0b4SJens Axboe static void io_flush_cached_locked_reqs(struct io_ring_ctx *ctx, 886ed29b0b4SJens Axboe struct io_submit_state *state) 887ed29b0b4SJens Axboe { 888ed29b0b4SJens Axboe spin_lock(&ctx->completion_lock); 889ed29b0b4SJens Axboe wq_list_splice(&ctx->locked_free_list, &state->free_list); 890ed29b0b4SJens Axboe ctx->locked_free_nr = 0; 891ed29b0b4SJens Axboe spin_unlock(&ctx->completion_lock); 892ed29b0b4SJens Axboe } 893ed29b0b4SJens Axboe 894ed29b0b4SJens Axboe /* 895ed29b0b4SJens Axboe * A request might get retired back into the request caches even before opcode 896ed29b0b4SJens Axboe * handlers and io_issue_sqe() are done with it, e.g. inline completion path. 897ed29b0b4SJens Axboe * Because of that, io_alloc_req() should be called only under ->uring_lock 898ed29b0b4SJens Axboe * and with extra caution to not get a request that is still worked on. 899ed29b0b4SJens Axboe */ 900bd1a3783SPavel Begunkov __cold bool __io_alloc_req_refill(struct io_ring_ctx *ctx) 901ed29b0b4SJens Axboe __must_hold(&ctx->uring_lock) 902ed29b0b4SJens Axboe { 903ed29b0b4SJens Axboe gfp_t gfp = GFP_KERNEL | __GFP_NOWARN; 904ed29b0b4SJens Axboe void *reqs[IO_REQ_ALLOC_BATCH]; 905ed29b0b4SJens Axboe int ret, i; 906ed29b0b4SJens Axboe 907ed29b0b4SJens Axboe /* 908ed29b0b4SJens Axboe * If we have more than a batch's worth of requests in our IRQ side 909ed29b0b4SJens Axboe * locked cache, grab the lock and move them over to our submission 910ed29b0b4SJens Axboe * side cache. 911ed29b0b4SJens Axboe */ 912ed29b0b4SJens Axboe if (data_race(ctx->locked_free_nr) > IO_COMPL_BATCH) { 913ed29b0b4SJens Axboe io_flush_cached_locked_reqs(ctx, &ctx->submit_state); 914ed29b0b4SJens Axboe if (!io_req_cache_empty(ctx)) 915ed29b0b4SJens Axboe return true; 916ed29b0b4SJens Axboe } 917ed29b0b4SJens Axboe 918ed29b0b4SJens Axboe ret = kmem_cache_alloc_bulk(req_cachep, gfp, ARRAY_SIZE(reqs), reqs); 919ed29b0b4SJens Axboe 920ed29b0b4SJens Axboe /* 921ed29b0b4SJens Axboe * Bulk alloc is all-or-nothing. If we fail to get a batch, 922ed29b0b4SJens Axboe * retry single alloc to be on the safe side. 923ed29b0b4SJens Axboe */ 924ed29b0b4SJens Axboe if (unlikely(ret <= 0)) { 925ed29b0b4SJens Axboe reqs[0] = kmem_cache_alloc(req_cachep, gfp); 926ed29b0b4SJens Axboe if (!reqs[0]) 927ed29b0b4SJens Axboe return false; 928ed29b0b4SJens Axboe ret = 1; 929ed29b0b4SJens Axboe } 930ed29b0b4SJens Axboe 931ed29b0b4SJens Axboe percpu_ref_get_many(&ctx->refs, ret); 932ed29b0b4SJens Axboe for (i = 0; i < ret; i++) { 933ed29b0b4SJens Axboe struct io_kiocb *req = reqs[i]; 934ed29b0b4SJens Axboe 935ed29b0b4SJens Axboe io_preinit_req(req, ctx); 936ed29b0b4SJens Axboe io_req_add_to_cache(req, ctx); 937ed29b0b4SJens Axboe } 938ed29b0b4SJens Axboe return true; 939ed29b0b4SJens Axboe } 940ed29b0b4SJens Axboe 941ed29b0b4SJens Axboe static inline void io_dismantle_req(struct io_kiocb *req) 942ed29b0b4SJens Axboe { 943ed29b0b4SJens Axboe unsigned int flags = req->flags; 944ed29b0b4SJens Axboe 945ed29b0b4SJens Axboe if (unlikely(flags & IO_REQ_CLEAN_FLAGS)) 946ed29b0b4SJens Axboe io_clean_op(req); 947ed29b0b4SJens Axboe if (!(flags & REQ_F_FIXED_FILE)) 948ed29b0b4SJens Axboe io_put_file(req->file); 949ed29b0b4SJens Axboe } 950ed29b0b4SJens Axboe 95159915143SJens Axboe __cold void io_free_req(struct io_kiocb *req) 952ed29b0b4SJens Axboe { 953ed29b0b4SJens Axboe struct io_ring_ctx *ctx = req->ctx; 954ed29b0b4SJens Axboe 955ed29b0b4SJens Axboe io_req_put_rsrc(req); 956ed29b0b4SJens Axboe io_dismantle_req(req); 957ed29b0b4SJens Axboe io_put_task(req->task, 1); 958ed29b0b4SJens Axboe 959ed29b0b4SJens Axboe spin_lock(&ctx->completion_lock); 960ed29b0b4SJens Axboe wq_list_add_head(&req->comp_list, &ctx->locked_free_list); 961ed29b0b4SJens Axboe ctx->locked_free_nr++; 962ed29b0b4SJens Axboe spin_unlock(&ctx->completion_lock); 963ed29b0b4SJens Axboe } 964ed29b0b4SJens Axboe 965ed29b0b4SJens Axboe static void __io_req_find_next_prep(struct io_kiocb *req) 966ed29b0b4SJens Axboe { 967ed29b0b4SJens Axboe struct io_ring_ctx *ctx = req->ctx; 968ed29b0b4SJens Axboe 96925399321SPavel Begunkov io_cq_lock(ctx); 970305bef98SPavel Begunkov io_disarm_next(req); 97125399321SPavel Begunkov io_cq_unlock_post(ctx); 972ed29b0b4SJens Axboe } 973ed29b0b4SJens Axboe 974ed29b0b4SJens Axboe static inline struct io_kiocb *io_req_find_next(struct io_kiocb *req) 975ed29b0b4SJens Axboe { 976ed29b0b4SJens Axboe struct io_kiocb *nxt; 977ed29b0b4SJens Axboe 978ed29b0b4SJens Axboe /* 979ed29b0b4SJens Axboe * If LINK is set, we have dependent requests in this chain. If we 980ed29b0b4SJens Axboe * didn't fail this request, queue the first one up, moving any other 981ed29b0b4SJens Axboe * dependencies to the next request. In case of failure, fail the rest 982ed29b0b4SJens Axboe * of the chain. 983ed29b0b4SJens Axboe */ 984ed29b0b4SJens Axboe if (unlikely(req->flags & IO_DISARM_MASK)) 985ed29b0b4SJens Axboe __io_req_find_next_prep(req); 986ed29b0b4SJens Axboe nxt = req->link; 987ed29b0b4SJens Axboe req->link = NULL; 988ed29b0b4SJens Axboe return nxt; 989ed29b0b4SJens Axboe } 990ed29b0b4SJens Axboe 991ed29b0b4SJens Axboe static void ctx_flush_and_put(struct io_ring_ctx *ctx, bool *locked) 992ed29b0b4SJens Axboe { 993ed29b0b4SJens Axboe if (!ctx) 994ed29b0b4SJens Axboe return; 995ed29b0b4SJens Axboe if (ctx->flags & IORING_SETUP_TASKRUN_FLAG) 996ed29b0b4SJens Axboe atomic_andnot(IORING_SQ_TASKRUN, &ctx->rings->sq_flags); 997ed29b0b4SJens Axboe if (*locked) { 998ed29b0b4SJens Axboe io_submit_flush_completions(ctx); 999ed29b0b4SJens Axboe mutex_unlock(&ctx->uring_lock); 1000ed29b0b4SJens Axboe *locked = false; 1001ed29b0b4SJens Axboe } 1002ed29b0b4SJens Axboe percpu_ref_put(&ctx->refs); 1003ed29b0b4SJens Axboe } 1004ed29b0b4SJens Axboe 1005c6dd763cSDylan Yudaken static unsigned int handle_tw_list(struct llist_node *node, 10063a0c037bSDylan Yudaken struct io_ring_ctx **ctx, bool *locked, 10073a0c037bSDylan Yudaken struct llist_node *last) 1008ed29b0b4SJens Axboe { 1009c6dd763cSDylan Yudaken unsigned int count = 0; 1010c6dd763cSDylan Yudaken 10113a0c037bSDylan Yudaken while (node != last) { 1012f88262e6SDylan Yudaken struct llist_node *next = node->next; 1013ed29b0b4SJens Axboe struct io_kiocb *req = container_of(node, struct io_kiocb, 1014ed29b0b4SJens Axboe io_task_work.node); 1015ed29b0b4SJens Axboe 1016ed29b0b4SJens Axboe prefetch(container_of(next, struct io_kiocb, io_task_work.node)); 1017ed29b0b4SJens Axboe 1018ed29b0b4SJens Axboe if (req->ctx != *ctx) { 1019ed29b0b4SJens Axboe ctx_flush_and_put(*ctx, locked); 1020ed29b0b4SJens Axboe *ctx = req->ctx; 1021ed29b0b4SJens Axboe /* if not contended, grab and improve batching */ 1022ed29b0b4SJens Axboe *locked = mutex_trylock(&(*ctx)->uring_lock); 1023ed29b0b4SJens Axboe percpu_ref_get(&(*ctx)->refs); 1024ed29b0b4SJens Axboe } 1025ed29b0b4SJens Axboe req->io_task_work.func(req, locked); 1026ed29b0b4SJens Axboe node = next; 1027c6dd763cSDylan Yudaken count++; 10283a0c037bSDylan Yudaken } 1029c6dd763cSDylan Yudaken 1030c6dd763cSDylan Yudaken return count; 1031ed29b0b4SJens Axboe } 1032ed29b0b4SJens Axboe 1033923d1592SDylan Yudaken /** 1034923d1592SDylan Yudaken * io_llist_xchg - swap all entries in a lock-less list 1035923d1592SDylan Yudaken * @head: the head of lock-less list to delete all entries 1036923d1592SDylan Yudaken * @new: new entry as the head of the list 1037923d1592SDylan Yudaken * 1038923d1592SDylan Yudaken * If list is empty, return NULL, otherwise, return the pointer to the first entry. 1039923d1592SDylan Yudaken * The order of entries returned is from the newest to the oldest added one. 1040923d1592SDylan Yudaken */ 1041923d1592SDylan Yudaken static inline struct llist_node *io_llist_xchg(struct llist_head *head, 1042923d1592SDylan Yudaken struct llist_node *new) 1043923d1592SDylan Yudaken { 1044923d1592SDylan Yudaken return xchg(&head->first, new); 1045923d1592SDylan Yudaken } 1046923d1592SDylan Yudaken 1047923d1592SDylan Yudaken /** 1048923d1592SDylan Yudaken * io_llist_cmpxchg - possibly swap all entries in a lock-less list 1049923d1592SDylan Yudaken * @head: the head of lock-less list to delete all entries 1050923d1592SDylan Yudaken * @old: expected old value of the first entry of the list 1051923d1592SDylan Yudaken * @new: new entry as the head of the list 1052923d1592SDylan Yudaken * 1053923d1592SDylan Yudaken * perform a cmpxchg on the first entry of the list. 1054923d1592SDylan Yudaken */ 1055923d1592SDylan Yudaken 1056923d1592SDylan Yudaken static inline struct llist_node *io_llist_cmpxchg(struct llist_head *head, 1057923d1592SDylan Yudaken struct llist_node *old, 1058923d1592SDylan Yudaken struct llist_node *new) 1059923d1592SDylan Yudaken { 1060923d1592SDylan Yudaken return cmpxchg(&head->first, old, new); 1061923d1592SDylan Yudaken } 1062923d1592SDylan Yudaken 1063c9f06aa7SJens Axboe void tctx_task_work(struct callback_head *cb) 1064ed29b0b4SJens Axboe { 1065ed29b0b4SJens Axboe bool uring_locked = false; 1066ed29b0b4SJens Axboe struct io_ring_ctx *ctx = NULL; 1067ed29b0b4SJens Axboe struct io_uring_task *tctx = container_of(cb, struct io_uring_task, 1068ed29b0b4SJens Axboe task_work); 10693a0c037bSDylan Yudaken struct llist_node fake = {}; 10703a0c037bSDylan Yudaken struct llist_node *node = io_llist_xchg(&tctx->task_list, &fake); 1071c6dd763cSDylan Yudaken unsigned int loops = 1; 1072c6dd763cSDylan Yudaken unsigned int count = handle_tw_list(node, &ctx, &uring_locked, NULL); 1073ed29b0b4SJens Axboe 10743a0c037bSDylan Yudaken node = io_llist_cmpxchg(&tctx->task_list, &fake, NULL); 10753a0c037bSDylan Yudaken while (node != &fake) { 1076c6dd763cSDylan Yudaken loops++; 10773a0c037bSDylan Yudaken node = io_llist_xchg(&tctx->task_list, &fake); 1078c6dd763cSDylan Yudaken count += handle_tw_list(node, &ctx, &uring_locked, &fake); 10793a0c037bSDylan Yudaken node = io_llist_cmpxchg(&tctx->task_list, &fake, NULL); 1080ed29b0b4SJens Axboe } 1081ed29b0b4SJens Axboe 1082ed29b0b4SJens Axboe ctx_flush_and_put(ctx, &uring_locked); 1083ed29b0b4SJens Axboe 1084ed29b0b4SJens Axboe /* relaxed read is enough as only the task itself sets ->in_idle */ 1085ed29b0b4SJens Axboe if (unlikely(atomic_read(&tctx->in_idle))) 1086ed29b0b4SJens Axboe io_uring_drop_tctx_refs(current); 1087c6dd763cSDylan Yudaken 1088c6dd763cSDylan Yudaken trace_io_uring_task_work_run(tctx, count, loops); 1089ed29b0b4SJens Axboe } 1090ed29b0b4SJens Axboe 1091d7593606SPavel Begunkov static __cold void io_fallback_tw(struct io_uring_task *tctx) 1092d7593606SPavel Begunkov { 1093d7593606SPavel Begunkov struct llist_node *node = llist_del_all(&tctx->task_list); 1094d7593606SPavel Begunkov struct io_kiocb *req; 1095d7593606SPavel Begunkov 1096d7593606SPavel Begunkov while (node) { 1097d7593606SPavel Begunkov req = container_of(node, struct io_kiocb, io_task_work.node); 1098d7593606SPavel Begunkov node = node->next; 1099d7593606SPavel Begunkov if (llist_add(&req->io_task_work.node, 1100d7593606SPavel Begunkov &req->ctx->fallback_llist)) 1101d7593606SPavel Begunkov schedule_delayed_work(&req->ctx->fallback_work, 1); 1102d7593606SPavel Begunkov } 1103d7593606SPavel Begunkov } 1104d7593606SPavel Begunkov 1105c0e0d6baSDylan Yudaken static void io_req_local_work_add(struct io_kiocb *req) 1106c0e0d6baSDylan Yudaken { 1107c0e0d6baSDylan Yudaken struct io_ring_ctx *ctx = req->ctx; 1108c0e0d6baSDylan Yudaken 1109c0e0d6baSDylan Yudaken if (!llist_add(&req->io_task_work.node, &ctx->work_llist)) 1110c0e0d6baSDylan Yudaken return; 1111fc86f9d3SPavel Begunkov /* need it for the following io_cqring_wake() */ 1112fc86f9d3SPavel Begunkov smp_mb__after_atomic(); 1113c0e0d6baSDylan Yudaken 1114c0e0d6baSDylan Yudaken if (unlikely(atomic_read(&req->task->io_uring->in_idle))) { 1115c0e0d6baSDylan Yudaken io_move_task_work_from_local(ctx); 1116c0e0d6baSDylan Yudaken return; 1117c0e0d6baSDylan Yudaken } 1118c0e0d6baSDylan Yudaken 1119c0e0d6baSDylan Yudaken if (ctx->flags & IORING_SETUP_TASKRUN_FLAG) 1120c0e0d6baSDylan Yudaken atomic_or(IORING_SQ_TASKRUN, &ctx->rings->sq_flags); 1121c0e0d6baSDylan Yudaken 112221a091b9SDylan Yudaken if (ctx->has_evfd) 112321a091b9SDylan Yudaken io_eventfd_signal(ctx); 1124fc86f9d3SPavel Begunkov __io_cqring_wake(ctx); 1125c0e0d6baSDylan Yudaken } 1126c0e0d6baSDylan Yudaken 1127e52d2e58SPavel Begunkov void __io_req_task_work_add(struct io_kiocb *req, bool allow_local) 1128ed29b0b4SJens Axboe { 1129c34398a8SDylan Yudaken struct io_uring_task *tctx = req->task->io_uring; 1130ed29b0b4SJens Axboe struct io_ring_ctx *ctx = req->ctx; 1131ed29b0b4SJens Axboe 1132c0e0d6baSDylan Yudaken if (allow_local && ctx->flags & IORING_SETUP_DEFER_TASKRUN) { 1133c0e0d6baSDylan Yudaken io_req_local_work_add(req); 1134c0e0d6baSDylan Yudaken return; 1135c0e0d6baSDylan Yudaken } 1136ed29b0b4SJens Axboe 1137ed29b0b4SJens Axboe /* task_work already pending, we're done */ 113832d91f05SDylan Yudaken if (!llist_add(&req->io_task_work.node, &tctx->task_list)) 1139ed29b0b4SJens Axboe return; 1140ed29b0b4SJens Axboe 1141ed29b0b4SJens Axboe if (ctx->flags & IORING_SETUP_TASKRUN_FLAG) 1142ed29b0b4SJens Axboe atomic_or(IORING_SQ_TASKRUN, &ctx->rings->sq_flags); 1143ed29b0b4SJens Axboe 1144ed29b0b4SJens Axboe if (likely(!task_work_add(req->task, &tctx->task_work, ctx->notify_method))) 1145ed29b0b4SJens Axboe return; 1146ed29b0b4SJens Axboe 1147d7593606SPavel Begunkov io_fallback_tw(tctx); 1148ed29b0b4SJens Axboe } 1149ed29b0b4SJens Axboe 1150c0e0d6baSDylan Yudaken static void __cold io_move_task_work_from_local(struct io_ring_ctx *ctx) 1151c0e0d6baSDylan Yudaken { 1152c0e0d6baSDylan Yudaken struct llist_node *node; 1153c0e0d6baSDylan Yudaken 1154c0e0d6baSDylan Yudaken node = llist_del_all(&ctx->work_llist); 1155c0e0d6baSDylan Yudaken while (node) { 1156c0e0d6baSDylan Yudaken struct io_kiocb *req = container_of(node, struct io_kiocb, 1157c0e0d6baSDylan Yudaken io_task_work.node); 1158c0e0d6baSDylan Yudaken 1159c0e0d6baSDylan Yudaken node = node->next; 1160c0e0d6baSDylan Yudaken __io_req_task_work_add(req, false); 1161c0e0d6baSDylan Yudaken } 1162c0e0d6baSDylan Yudaken } 1163c0e0d6baSDylan Yudaken 1164b3026767SDylan Yudaken int __io_run_local_work(struct io_ring_ctx *ctx, bool *locked) 1165c0e0d6baSDylan Yudaken { 1166c0e0d6baSDylan Yudaken struct llist_node *node; 1167c0e0d6baSDylan Yudaken struct llist_node fake; 1168c0e0d6baSDylan Yudaken struct llist_node *current_final = NULL; 1169c0e0d6baSDylan Yudaken int ret; 1170f75d5036SDylan Yudaken unsigned int loops = 1; 1171c0e0d6baSDylan Yudaken 11726567506bSPavel Begunkov if (unlikely(ctx->submitter_task != current)) 1173c0e0d6baSDylan Yudaken return -EEXIST; 1174c0e0d6baSDylan Yudaken 1175c0e0d6baSDylan Yudaken node = io_llist_xchg(&ctx->work_llist, &fake); 1176c0e0d6baSDylan Yudaken ret = 0; 1177c0e0d6baSDylan Yudaken again: 1178c0e0d6baSDylan Yudaken while (node != current_final) { 1179c0e0d6baSDylan Yudaken struct llist_node *next = node->next; 1180c0e0d6baSDylan Yudaken struct io_kiocb *req = container_of(node, struct io_kiocb, 1181c0e0d6baSDylan Yudaken io_task_work.node); 1182c0e0d6baSDylan Yudaken prefetch(container_of(next, struct io_kiocb, io_task_work.node)); 1183b3026767SDylan Yudaken req->io_task_work.func(req, locked); 1184c0e0d6baSDylan Yudaken ret++; 1185c0e0d6baSDylan Yudaken node = next; 1186c0e0d6baSDylan Yudaken } 1187c0e0d6baSDylan Yudaken 1188c0e0d6baSDylan Yudaken if (ctx->flags & IORING_SETUP_TASKRUN_FLAG) 1189c0e0d6baSDylan Yudaken atomic_andnot(IORING_SQ_TASKRUN, &ctx->rings->sq_flags); 1190c0e0d6baSDylan Yudaken 1191c0e0d6baSDylan Yudaken node = io_llist_cmpxchg(&ctx->work_llist, &fake, NULL); 1192c0e0d6baSDylan Yudaken if (node != &fake) { 1193f75d5036SDylan Yudaken loops++; 1194c0e0d6baSDylan Yudaken current_final = &fake; 1195c0e0d6baSDylan Yudaken node = io_llist_xchg(&ctx->work_llist, &fake); 1196c0e0d6baSDylan Yudaken goto again; 1197c0e0d6baSDylan Yudaken } 1198c0e0d6baSDylan Yudaken 1199b3026767SDylan Yudaken if (*locked) 1200c0e0d6baSDylan Yudaken io_submit_flush_completions(ctx); 1201f75d5036SDylan Yudaken trace_io_uring_local_work_run(ctx, ret, loops); 1202c0e0d6baSDylan Yudaken return ret; 12038ac5d85aSJens Axboe 12048ac5d85aSJens Axboe } 12058ac5d85aSJens Axboe 12068ac5d85aSJens Axboe int io_run_local_work(struct io_ring_ctx *ctx) 12078ac5d85aSJens Axboe { 12088ac5d85aSJens Axboe bool locked; 12098ac5d85aSJens Axboe int ret; 12108ac5d85aSJens Axboe 12117924fdfeSPavel Begunkov if (llist_empty(&ctx->work_llist)) 12127924fdfeSPavel Begunkov return 0; 12137924fdfeSPavel Begunkov 1214ec7fd256SJens Axboe __set_current_state(TASK_RUNNING); 12158ac5d85aSJens Axboe locked = mutex_trylock(&ctx->uring_lock); 1216b3026767SDylan Yudaken ret = __io_run_local_work(ctx, &locked); 12178ac5d85aSJens Axboe if (locked) 12188ac5d85aSJens Axboe mutex_unlock(&ctx->uring_lock); 12198ac5d85aSJens Axboe 12208ac5d85aSJens Axboe return ret; 1221c0e0d6baSDylan Yudaken } 1222c0e0d6baSDylan Yudaken 1223ed29b0b4SJens Axboe static void io_req_task_cancel(struct io_kiocb *req, bool *locked) 1224ed29b0b4SJens Axboe { 1225ed29b0b4SJens Axboe /* not needed for normal modes, but SQPOLL depends on it */ 1226ed29b0b4SJens Axboe io_tw_lock(req->ctx, locked); 1227ed29b0b4SJens Axboe io_req_complete_failed(req, req->cqe.res); 1228ed29b0b4SJens Axboe } 1229ed29b0b4SJens Axboe 1230329061d3SJens Axboe void io_req_task_submit(struct io_kiocb *req, bool *locked) 1231ed29b0b4SJens Axboe { 1232ed29b0b4SJens Axboe io_tw_lock(req->ctx, locked); 1233ed29b0b4SJens Axboe /* req->task == current here, checking PF_EXITING is safe */ 1234ed29b0b4SJens Axboe if (likely(!(req->task->flags & PF_EXITING))) 1235ed29b0b4SJens Axboe io_queue_sqe(req); 1236ed29b0b4SJens Axboe else 1237ed29b0b4SJens Axboe io_req_complete_failed(req, -EFAULT); 1238ed29b0b4SJens Axboe } 1239ed29b0b4SJens Axboe 124059915143SJens Axboe void io_req_task_queue_fail(struct io_kiocb *req, int ret) 1241ed29b0b4SJens Axboe { 124297b388d7SJens Axboe io_req_set_res(req, ret, 0); 1243ed29b0b4SJens Axboe req->io_task_work.func = io_req_task_cancel; 1244ed29b0b4SJens Axboe io_req_task_work_add(req); 1245ed29b0b4SJens Axboe } 1246ed29b0b4SJens Axboe 1247f3b44f92SJens Axboe void io_req_task_queue(struct io_kiocb *req) 1248ed29b0b4SJens Axboe { 1249ed29b0b4SJens Axboe req->io_task_work.func = io_req_task_submit; 1250ed29b0b4SJens Axboe io_req_task_work_add(req); 1251ed29b0b4SJens Axboe } 1252ed29b0b4SJens Axboe 125359915143SJens Axboe void io_queue_next(struct io_kiocb *req) 1254ed29b0b4SJens Axboe { 1255ed29b0b4SJens Axboe struct io_kiocb *nxt = io_req_find_next(req); 1256ed29b0b4SJens Axboe 1257ed29b0b4SJens Axboe if (nxt) 1258ed29b0b4SJens Axboe io_req_task_queue(nxt); 1259ed29b0b4SJens Axboe } 1260ed29b0b4SJens Axboe 1261f3b44f92SJens Axboe void io_free_batch_list(struct io_ring_ctx *ctx, struct io_wq_work_node *node) 1262ed29b0b4SJens Axboe __must_hold(&ctx->uring_lock) 1263ed29b0b4SJens Axboe { 1264ed29b0b4SJens Axboe struct task_struct *task = NULL; 1265ed29b0b4SJens Axboe int task_refs = 0; 1266ed29b0b4SJens Axboe 1267ed29b0b4SJens Axboe do { 1268ed29b0b4SJens Axboe struct io_kiocb *req = container_of(node, struct io_kiocb, 1269ed29b0b4SJens Axboe comp_list); 1270ed29b0b4SJens Axboe 1271ed29b0b4SJens Axboe if (unlikely(req->flags & IO_REQ_CLEAN_SLOW_FLAGS)) { 1272ed29b0b4SJens Axboe if (req->flags & REQ_F_REFCOUNT) { 1273ed29b0b4SJens Axboe node = req->comp_list.next; 1274ed29b0b4SJens Axboe if (!req_ref_put_and_test(req)) 1275ed29b0b4SJens Axboe continue; 1276ed29b0b4SJens Axboe } 1277ed29b0b4SJens Axboe if ((req->flags & REQ_F_POLLED) && req->apoll) { 1278ed29b0b4SJens Axboe struct async_poll *apoll = req->apoll; 1279ed29b0b4SJens Axboe 1280ed29b0b4SJens Axboe if (apoll->double_poll) 1281ed29b0b4SJens Axboe kfree(apoll->double_poll); 12829731bc98SJens Axboe if (!io_alloc_cache_put(&ctx->apoll_cache, &apoll->cache)) 12839731bc98SJens Axboe kfree(apoll); 1284ed29b0b4SJens Axboe req->flags &= ~REQ_F_POLLED; 1285ed29b0b4SJens Axboe } 1286ed29b0b4SJens Axboe if (req->flags & IO_REQ_LINK_FLAGS) 1287ed29b0b4SJens Axboe io_queue_next(req); 1288ed29b0b4SJens Axboe if (unlikely(req->flags & IO_REQ_CLEAN_FLAGS)) 1289ed29b0b4SJens Axboe io_clean_op(req); 1290ed29b0b4SJens Axboe } 1291ed29b0b4SJens Axboe if (!(req->flags & REQ_F_FIXED_FILE)) 1292ed29b0b4SJens Axboe io_put_file(req->file); 1293ed29b0b4SJens Axboe 1294ed29b0b4SJens Axboe io_req_put_rsrc_locked(req, ctx); 1295ed29b0b4SJens Axboe 1296ed29b0b4SJens Axboe if (req->task != task) { 1297ed29b0b4SJens Axboe if (task) 1298ed29b0b4SJens Axboe io_put_task(task, task_refs); 1299ed29b0b4SJens Axboe task = req->task; 1300ed29b0b4SJens Axboe task_refs = 0; 1301ed29b0b4SJens Axboe } 1302ed29b0b4SJens Axboe task_refs++; 1303ed29b0b4SJens Axboe node = req->comp_list.next; 1304ed29b0b4SJens Axboe io_req_add_to_cache(req, ctx); 1305ed29b0b4SJens Axboe } while (node); 1306ed29b0b4SJens Axboe 1307ed29b0b4SJens Axboe if (task) 1308ed29b0b4SJens Axboe io_put_task(task, task_refs); 1309ed29b0b4SJens Axboe } 1310ed29b0b4SJens Axboe 1311ed29b0b4SJens Axboe static void __io_submit_flush_completions(struct io_ring_ctx *ctx) 1312ed29b0b4SJens Axboe __must_hold(&ctx->uring_lock) 1313ed29b0b4SJens Axboe { 1314ed29b0b4SJens Axboe struct io_wq_work_node *node, *prev; 1315ed29b0b4SJens Axboe struct io_submit_state *state = &ctx->submit_state; 1316ed29b0b4SJens Axboe 1317e9a88428SPavel Begunkov io_cq_lock(ctx); 1318ed29b0b4SJens Axboe wq_list_for_each(node, prev, &state->compl_reqs) { 1319ed29b0b4SJens Axboe struct io_kiocb *req = container_of(node, struct io_kiocb, 1320ed29b0b4SJens Axboe comp_list); 1321ed29b0b4SJens Axboe 1322ed29b0b4SJens Axboe if (!(req->flags & REQ_F_CQE_SKIP)) 1323ed29b0b4SJens Axboe __io_fill_cqe_req(ctx, req); 1324ed29b0b4SJens Axboe } 13256c16fe3cSJens Axboe io_cq_unlock_post(ctx); 1326ed29b0b4SJens Axboe 1327ed29b0b4SJens Axboe io_free_batch_list(ctx, state->compl_reqs.first); 1328ed29b0b4SJens Axboe INIT_WQ_LIST(&state->compl_reqs); 1329ed29b0b4SJens Axboe } 1330ed29b0b4SJens Axboe 1331ed29b0b4SJens Axboe /* 1332ed29b0b4SJens Axboe * Drop reference to request, return next in chain (if there is one) if this 1333ed29b0b4SJens Axboe * was the last reference to this request. 1334ed29b0b4SJens Axboe */ 1335ed29b0b4SJens Axboe static inline struct io_kiocb *io_put_req_find_next(struct io_kiocb *req) 1336ed29b0b4SJens Axboe { 1337ed29b0b4SJens Axboe struct io_kiocb *nxt = NULL; 1338ed29b0b4SJens Axboe 1339ed29b0b4SJens Axboe if (req_ref_put_and_test(req)) { 1340ed29b0b4SJens Axboe if (unlikely(req->flags & IO_REQ_LINK_FLAGS)) 1341ed29b0b4SJens Axboe nxt = io_req_find_next(req); 1342ed29b0b4SJens Axboe io_free_req(req); 1343ed29b0b4SJens Axboe } 1344ed29b0b4SJens Axboe return nxt; 1345ed29b0b4SJens Axboe } 1346ed29b0b4SJens Axboe 1347ed29b0b4SJens Axboe static unsigned io_cqring_events(struct io_ring_ctx *ctx) 1348ed29b0b4SJens Axboe { 1349ed29b0b4SJens Axboe /* See comment at the top of this file */ 1350ed29b0b4SJens Axboe smp_rmb(); 1351ed29b0b4SJens Axboe return __io_cqring_events(ctx); 1352ed29b0b4SJens Axboe } 1353ed29b0b4SJens Axboe 1354ed29b0b4SJens Axboe /* 1355ed29b0b4SJens Axboe * We can't just wait for polled events to come to us, we have to actively 1356ed29b0b4SJens Axboe * find and complete them. 1357ed29b0b4SJens Axboe */ 1358ed29b0b4SJens Axboe static __cold void io_iopoll_try_reap_events(struct io_ring_ctx *ctx) 1359ed29b0b4SJens Axboe { 1360ed29b0b4SJens Axboe if (!(ctx->flags & IORING_SETUP_IOPOLL)) 1361ed29b0b4SJens Axboe return; 1362ed29b0b4SJens Axboe 1363ed29b0b4SJens Axboe mutex_lock(&ctx->uring_lock); 1364ed29b0b4SJens Axboe while (!wq_list_empty(&ctx->iopoll_list)) { 1365ed29b0b4SJens Axboe /* let it sleep and repeat later if can't complete a request */ 1366ed29b0b4SJens Axboe if (io_do_iopoll(ctx, true) == 0) 1367ed29b0b4SJens Axboe break; 1368ed29b0b4SJens Axboe /* 1369ed29b0b4SJens Axboe * Ensure we allow local-to-the-cpu processing to take place, 1370ed29b0b4SJens Axboe * in this case we need to ensure that we reap all events. 1371ed29b0b4SJens Axboe * Also let task_work, etc. to progress by releasing the mutex 1372ed29b0b4SJens Axboe */ 1373ed29b0b4SJens Axboe if (need_resched()) { 1374ed29b0b4SJens Axboe mutex_unlock(&ctx->uring_lock); 1375ed29b0b4SJens Axboe cond_resched(); 1376ed29b0b4SJens Axboe mutex_lock(&ctx->uring_lock); 1377ed29b0b4SJens Axboe } 1378ed29b0b4SJens Axboe } 1379ed29b0b4SJens Axboe mutex_unlock(&ctx->uring_lock); 1380ed29b0b4SJens Axboe } 1381ed29b0b4SJens Axboe 1382ed29b0b4SJens Axboe static int io_iopoll_check(struct io_ring_ctx *ctx, long min) 1383ed29b0b4SJens Axboe { 1384ed29b0b4SJens Axboe unsigned int nr_events = 0; 1385ed29b0b4SJens Axboe int ret = 0; 1386ed29b0b4SJens Axboe unsigned long check_cq; 1387ed29b0b4SJens Axboe 138876de6749SPavel Begunkov if (!io_allowed_run_tw(ctx)) 138976de6749SPavel Begunkov return -EEXIST; 139076de6749SPavel Begunkov 13913a08576bSPavel Begunkov check_cq = READ_ONCE(ctx->check_cq); 13923a08576bSPavel Begunkov if (unlikely(check_cq)) { 13933a08576bSPavel Begunkov if (check_cq & BIT(IO_CHECK_CQ_OVERFLOW_BIT)) 13943a08576bSPavel Begunkov __io_cqring_overflow_flush(ctx, false); 13953a08576bSPavel Begunkov /* 13963a08576bSPavel Begunkov * Similarly do not spin if we have not informed the user of any 13973a08576bSPavel Begunkov * dropped CQE. 13983a08576bSPavel Begunkov */ 13993a08576bSPavel Begunkov if (check_cq & BIT(IO_CHECK_CQ_DROPPED_BIT)) 14003a08576bSPavel Begunkov return -EBADR; 14013a08576bSPavel Begunkov } 1402ed29b0b4SJens Axboe /* 1403ed29b0b4SJens Axboe * Don't enter poll loop if we already have events pending. 1404ed29b0b4SJens Axboe * If we do, we can potentially be spinning for commands that 1405ed29b0b4SJens Axboe * already triggered a CQE (eg in error). 1406ed29b0b4SJens Axboe */ 1407ed29b0b4SJens Axboe if (io_cqring_events(ctx)) 1408ed29b0b4SJens Axboe return 0; 1409ed29b0b4SJens Axboe 1410ed29b0b4SJens Axboe do { 1411ed29b0b4SJens Axboe /* 1412ed29b0b4SJens Axboe * If a submit got punted to a workqueue, we can have the 1413ed29b0b4SJens Axboe * application entering polling for a command before it gets 1414ed29b0b4SJens Axboe * issued. That app will hold the uring_lock for the duration 1415ed29b0b4SJens Axboe * of the poll right here, so we need to take a breather every 1416ed29b0b4SJens Axboe * now and then to ensure that the issue has a chance to add 1417ed29b0b4SJens Axboe * the poll to the issued list. Otherwise we can spin here 1418ed29b0b4SJens Axboe * forever, while the workqueue is stuck trying to acquire the 1419ed29b0b4SJens Axboe * very same mutex. 1420ed29b0b4SJens Axboe */ 1421dac6a0eaSJens Axboe if (wq_list_empty(&ctx->iopoll_list) || 1422dac6a0eaSJens Axboe io_task_work_pending(ctx)) { 1423ed29b0b4SJens Axboe u32 tail = ctx->cached_cq_tail; 1424ed29b0b4SJens Axboe 14258de11cdcSDylan Yudaken (void) io_run_local_work_locked(ctx); 14261f8d5bbeSPavel Begunkov 14271f8d5bbeSPavel Begunkov if (task_work_pending(current) || 14281f8d5bbeSPavel Begunkov wq_list_empty(&ctx->iopoll_list)) { 1429ed29b0b4SJens Axboe mutex_unlock(&ctx->uring_lock); 1430ed29b0b4SJens Axboe io_run_task_work(); 1431ed29b0b4SJens Axboe mutex_lock(&ctx->uring_lock); 14321f8d5bbeSPavel Begunkov } 1433ed29b0b4SJens Axboe /* some requests don't go through iopoll_list */ 1434ed29b0b4SJens Axboe if (tail != ctx->cached_cq_tail || 1435ed29b0b4SJens Axboe wq_list_empty(&ctx->iopoll_list)) 1436ed29b0b4SJens Axboe break; 1437ed29b0b4SJens Axboe } 1438ed29b0b4SJens Axboe ret = io_do_iopoll(ctx, !min); 1439ed29b0b4SJens Axboe if (ret < 0) 1440ed29b0b4SJens Axboe break; 1441ed29b0b4SJens Axboe nr_events += ret; 1442ed29b0b4SJens Axboe ret = 0; 1443ed29b0b4SJens Axboe } while (nr_events < min && !need_resched()); 1444ed29b0b4SJens Axboe 1445ed29b0b4SJens Axboe return ret; 1446ed29b0b4SJens Axboe } 14477012c815SPavel Begunkov 14487012c815SPavel Begunkov void io_req_task_complete(struct io_kiocb *req, bool *locked) 1449ed29b0b4SJens Axboe { 14507012c815SPavel Begunkov if (*locked) 14519da070b1SPavel Begunkov io_req_complete_defer(req); 14527012c815SPavel Begunkov else 14537012c815SPavel Begunkov io_req_complete_post(req); 1454ed29b0b4SJens Axboe } 1455ed29b0b4SJens Axboe 1456ed29b0b4SJens Axboe /* 1457ed29b0b4SJens Axboe * After the iocb has been issued, it's safe to be found on the poll list. 1458ed29b0b4SJens Axboe * Adding the kiocb to the list AFTER submission ensures that we don't 1459ed29b0b4SJens Axboe * find it from a io_do_iopoll() thread before the issuer is done 1460ed29b0b4SJens Axboe * accessing the kiocb cookie. 1461ed29b0b4SJens Axboe */ 1462ed29b0b4SJens Axboe static void io_iopoll_req_issued(struct io_kiocb *req, unsigned int issue_flags) 1463ed29b0b4SJens Axboe { 1464ed29b0b4SJens Axboe struct io_ring_ctx *ctx = req->ctx; 1465ed29b0b4SJens Axboe const bool needs_lock = issue_flags & IO_URING_F_UNLOCKED; 1466ed29b0b4SJens Axboe 1467ed29b0b4SJens Axboe /* workqueue context doesn't hold uring_lock, grab it now */ 1468ed29b0b4SJens Axboe if (unlikely(needs_lock)) 1469ed29b0b4SJens Axboe mutex_lock(&ctx->uring_lock); 1470ed29b0b4SJens Axboe 1471ed29b0b4SJens Axboe /* 1472ed29b0b4SJens Axboe * Track whether we have multiple files in our lists. This will impact 1473ed29b0b4SJens Axboe * how we do polling eventually, not spinning if we're on potentially 1474ed29b0b4SJens Axboe * different devices. 1475ed29b0b4SJens Axboe */ 1476ed29b0b4SJens Axboe if (wq_list_empty(&ctx->iopoll_list)) { 1477ed29b0b4SJens Axboe ctx->poll_multi_queue = false; 1478ed29b0b4SJens Axboe } else if (!ctx->poll_multi_queue) { 1479ed29b0b4SJens Axboe struct io_kiocb *list_req; 1480ed29b0b4SJens Axboe 1481ed29b0b4SJens Axboe list_req = container_of(ctx->iopoll_list.first, struct io_kiocb, 1482ed29b0b4SJens Axboe comp_list); 1483ed29b0b4SJens Axboe if (list_req->file != req->file) 1484ed29b0b4SJens Axboe ctx->poll_multi_queue = true; 1485ed29b0b4SJens Axboe } 1486ed29b0b4SJens Axboe 1487ed29b0b4SJens Axboe /* 1488ed29b0b4SJens Axboe * For fast devices, IO may have already completed. If it has, add 1489ed29b0b4SJens Axboe * it to the front so we find it first. 1490ed29b0b4SJens Axboe */ 1491ed29b0b4SJens Axboe if (READ_ONCE(req->iopoll_completed)) 1492ed29b0b4SJens Axboe wq_list_add_head(&req->comp_list, &ctx->iopoll_list); 1493ed29b0b4SJens Axboe else 1494ed29b0b4SJens Axboe wq_list_add_tail(&req->comp_list, &ctx->iopoll_list); 1495ed29b0b4SJens Axboe 1496ed29b0b4SJens Axboe if (unlikely(needs_lock)) { 1497ed29b0b4SJens Axboe /* 1498ed29b0b4SJens Axboe * If IORING_SETUP_SQPOLL is enabled, sqes are either handle 1499ed29b0b4SJens Axboe * in sq thread task context or in io worker task context. If 1500ed29b0b4SJens Axboe * current task context is sq thread, we don't need to check 1501ed29b0b4SJens Axboe * whether should wake up sq thread. 1502ed29b0b4SJens Axboe */ 1503ed29b0b4SJens Axboe if ((ctx->flags & IORING_SETUP_SQPOLL) && 1504ed29b0b4SJens Axboe wq_has_sleeper(&ctx->sq_data->wait)) 1505ed29b0b4SJens Axboe wake_up(&ctx->sq_data->wait); 1506ed29b0b4SJens Axboe 1507ed29b0b4SJens Axboe mutex_unlock(&ctx->uring_lock); 1508ed29b0b4SJens Axboe } 1509ed29b0b4SJens Axboe } 1510ed29b0b4SJens Axboe 1511ed29b0b4SJens Axboe static bool io_bdev_nowait(struct block_device *bdev) 1512ed29b0b4SJens Axboe { 1513568ec936SChristoph Hellwig return !bdev || bdev_nowait(bdev); 1514ed29b0b4SJens Axboe } 1515ed29b0b4SJens Axboe 1516ed29b0b4SJens Axboe /* 1517ed29b0b4SJens Axboe * If we tracked the file through the SCM inflight mechanism, we could support 1518ed29b0b4SJens Axboe * any file. For now, just ensure that anything potentially problematic is done 1519ed29b0b4SJens Axboe * inline. 1520ed29b0b4SJens Axboe */ 1521ed29b0b4SJens Axboe static bool __io_file_supports_nowait(struct file *file, umode_t mode) 1522ed29b0b4SJens Axboe { 1523ed29b0b4SJens Axboe if (S_ISBLK(mode)) { 1524ed29b0b4SJens Axboe if (IS_ENABLED(CONFIG_BLOCK) && 1525ed29b0b4SJens Axboe io_bdev_nowait(I_BDEV(file->f_mapping->host))) 1526ed29b0b4SJens Axboe return true; 1527ed29b0b4SJens Axboe return false; 1528ed29b0b4SJens Axboe } 1529ed29b0b4SJens Axboe if (S_ISSOCK(mode)) 1530ed29b0b4SJens Axboe return true; 1531ed29b0b4SJens Axboe if (S_ISREG(mode)) { 1532ed29b0b4SJens Axboe if (IS_ENABLED(CONFIG_BLOCK) && 1533ed29b0b4SJens Axboe io_bdev_nowait(file->f_inode->i_sb->s_bdev) && 1534e5550a14SJens Axboe !io_is_uring_fops(file)) 1535ed29b0b4SJens Axboe return true; 1536ed29b0b4SJens Axboe return false; 1537ed29b0b4SJens Axboe } 1538ed29b0b4SJens Axboe 1539ed29b0b4SJens Axboe /* any ->read/write should understand O_NONBLOCK */ 1540ed29b0b4SJens Axboe if (file->f_flags & O_NONBLOCK) 1541ed29b0b4SJens Axboe return true; 1542ed29b0b4SJens Axboe return file->f_mode & FMODE_NOWAIT; 1543ed29b0b4SJens Axboe } 1544ed29b0b4SJens Axboe 1545ed29b0b4SJens Axboe /* 1546ed29b0b4SJens Axboe * If we tracked the file through the SCM inflight mechanism, we could support 1547ed29b0b4SJens Axboe * any file. For now, just ensure that anything potentially problematic is done 1548ed29b0b4SJens Axboe * inline. 1549ed29b0b4SJens Axboe */ 1550a4ad4f74SJens Axboe unsigned int io_file_get_flags(struct file *file) 1551ed29b0b4SJens Axboe { 1552ed29b0b4SJens Axboe umode_t mode = file_inode(file)->i_mode; 1553ed29b0b4SJens Axboe unsigned int res = 0; 1554ed29b0b4SJens Axboe 1555ed29b0b4SJens Axboe if (S_ISREG(mode)) 1556ed29b0b4SJens Axboe res |= FFS_ISREG; 1557ed29b0b4SJens Axboe if (__io_file_supports_nowait(file, mode)) 1558ed29b0b4SJens Axboe res |= FFS_NOWAIT; 1559ed29b0b4SJens Axboe return res; 1560ed29b0b4SJens Axboe } 1561ed29b0b4SJens Axboe 156299f15d8dSJens Axboe bool io_alloc_async_data(struct io_kiocb *req) 1563ed29b0b4SJens Axboe { 1564ed29b0b4SJens Axboe WARN_ON_ONCE(!io_op_defs[req->opcode].async_size); 1565ed29b0b4SJens Axboe req->async_data = kmalloc(io_op_defs[req->opcode].async_size, GFP_KERNEL); 1566ed29b0b4SJens Axboe if (req->async_data) { 1567ed29b0b4SJens Axboe req->flags |= REQ_F_ASYNC_DATA; 1568ed29b0b4SJens Axboe return false; 1569ed29b0b4SJens Axboe } 1570ed29b0b4SJens Axboe return true; 1571ed29b0b4SJens Axboe } 1572ed29b0b4SJens Axboe 1573f3b44f92SJens Axboe int io_req_prep_async(struct io_kiocb *req) 1574ed29b0b4SJens Axboe { 1575ed29b0b4SJens Axboe const struct io_op_def *def = &io_op_defs[req->opcode]; 1576ed29b0b4SJens Axboe 1577ed29b0b4SJens Axboe /* assign early for deferred execution for non-fixed file */ 1578ed29b0b4SJens Axboe if (def->needs_file && !(req->flags & REQ_F_FIXED_FILE)) 1579ed29b0b4SJens Axboe req->file = io_file_get_normal(req, req->cqe.fd); 1580dc919cafSJens Axboe if (!def->prep_async) 1581ed29b0b4SJens Axboe return 0; 1582ed29b0b4SJens Axboe if (WARN_ON_ONCE(req_has_async_data(req))) 1583ed29b0b4SJens Axboe return -EFAULT; 158459169439SPavel Begunkov if (!io_op_defs[req->opcode].manual_alloc) { 1585ed29b0b4SJens Axboe if (io_alloc_async_data(req)) 1586ed29b0b4SJens Axboe return -EAGAIN; 158759169439SPavel Begunkov } 1588dc919cafSJens Axboe return def->prep_async(req); 1589ed29b0b4SJens Axboe } 1590ed29b0b4SJens Axboe 1591ed29b0b4SJens Axboe static u32 io_get_sequence(struct io_kiocb *req) 1592ed29b0b4SJens Axboe { 1593ed29b0b4SJens Axboe u32 seq = req->ctx->cached_sq_head; 1594ed29b0b4SJens Axboe struct io_kiocb *cur; 1595ed29b0b4SJens Axboe 1596ed29b0b4SJens Axboe /* need original cached_sq_head, but it was increased for each req */ 1597ed29b0b4SJens Axboe io_for_each_link(cur, req) 1598ed29b0b4SJens Axboe seq--; 1599ed29b0b4SJens Axboe return seq; 1600ed29b0b4SJens Axboe } 1601ed29b0b4SJens Axboe 1602ed29b0b4SJens Axboe static __cold void io_drain_req(struct io_kiocb *req) 1603e276ae34SPavel Begunkov __must_hold(&ctx->uring_lock) 1604ed29b0b4SJens Axboe { 1605ed29b0b4SJens Axboe struct io_ring_ctx *ctx = req->ctx; 1606ed29b0b4SJens Axboe struct io_defer_entry *de; 1607ed29b0b4SJens Axboe int ret; 1608ed29b0b4SJens Axboe u32 seq = io_get_sequence(req); 1609ed29b0b4SJens Axboe 1610ed29b0b4SJens Axboe /* Still need defer if there is pending req in defer list. */ 1611ed29b0b4SJens Axboe spin_lock(&ctx->completion_lock); 1612ed29b0b4SJens Axboe if (!req_need_defer(req, seq) && list_empty_careful(&ctx->defer_list)) { 1613ed29b0b4SJens Axboe spin_unlock(&ctx->completion_lock); 1614ed29b0b4SJens Axboe queue: 1615ed29b0b4SJens Axboe ctx->drain_active = false; 1616ed29b0b4SJens Axboe io_req_task_queue(req); 1617ed29b0b4SJens Axboe return; 1618ed29b0b4SJens Axboe } 1619ed29b0b4SJens Axboe spin_unlock(&ctx->completion_lock); 1620ed29b0b4SJens Axboe 1621ed29b0b4SJens Axboe ret = io_req_prep_async(req); 1622ed29b0b4SJens Axboe if (ret) { 1623ed29b0b4SJens Axboe fail: 1624ed29b0b4SJens Axboe io_req_complete_failed(req, ret); 1625ed29b0b4SJens Axboe return; 1626ed29b0b4SJens Axboe } 1627ed29b0b4SJens Axboe io_prep_async_link(req); 1628ed29b0b4SJens Axboe de = kmalloc(sizeof(*de), GFP_KERNEL); 1629ed29b0b4SJens Axboe if (!de) { 1630ed29b0b4SJens Axboe ret = -ENOMEM; 1631ed29b0b4SJens Axboe goto fail; 1632ed29b0b4SJens Axboe } 1633ed29b0b4SJens Axboe 1634ed29b0b4SJens Axboe spin_lock(&ctx->completion_lock); 1635ed29b0b4SJens Axboe if (!req_need_defer(req, seq) && list_empty(&ctx->defer_list)) { 1636ed29b0b4SJens Axboe spin_unlock(&ctx->completion_lock); 1637ed29b0b4SJens Axboe kfree(de); 1638ed29b0b4SJens Axboe goto queue; 1639ed29b0b4SJens Axboe } 1640ed29b0b4SJens Axboe 164148863ffdSPavel Begunkov trace_io_uring_defer(req); 1642ed29b0b4SJens Axboe de->req = req; 1643ed29b0b4SJens Axboe de->seq = seq; 1644ed29b0b4SJens Axboe list_add_tail(&de->list, &ctx->defer_list); 1645ed29b0b4SJens Axboe spin_unlock(&ctx->completion_lock); 1646ed29b0b4SJens Axboe } 1647ed29b0b4SJens Axboe 1648ed29b0b4SJens Axboe static void io_clean_op(struct io_kiocb *req) 1649ed29b0b4SJens Axboe { 1650ed29b0b4SJens Axboe if (req->flags & REQ_F_BUFFER_SELECTED) { 1651ed29b0b4SJens Axboe spin_lock(&req->ctx->completion_lock); 1652ed29b0b4SJens Axboe io_put_kbuf_comp(req); 1653ed29b0b4SJens Axboe spin_unlock(&req->ctx->completion_lock); 1654ed29b0b4SJens Axboe } 1655ed29b0b4SJens Axboe 1656ed29b0b4SJens Axboe if (req->flags & REQ_F_NEED_CLEANUP) { 16574d4c9cffSJens Axboe const struct io_op_def *def = &io_op_defs[req->opcode]; 1658ed29b0b4SJens Axboe 16594d4c9cffSJens Axboe if (def->cleanup) 16604d4c9cffSJens Axboe def->cleanup(req); 1661ed29b0b4SJens Axboe } 1662ed29b0b4SJens Axboe if ((req->flags & REQ_F_POLLED) && req->apoll) { 1663ed29b0b4SJens Axboe kfree(req->apoll->double_poll); 1664ed29b0b4SJens Axboe kfree(req->apoll); 1665ed29b0b4SJens Axboe req->apoll = NULL; 1666ed29b0b4SJens Axboe } 1667ed29b0b4SJens Axboe if (req->flags & REQ_F_INFLIGHT) { 1668ed29b0b4SJens Axboe struct io_uring_task *tctx = req->task->io_uring; 1669ed29b0b4SJens Axboe 1670ed29b0b4SJens Axboe atomic_dec(&tctx->inflight_tracked); 1671ed29b0b4SJens Axboe } 1672ed29b0b4SJens Axboe if (req->flags & REQ_F_CREDS) 1673ed29b0b4SJens Axboe put_cred(req->creds); 1674ed29b0b4SJens Axboe if (req->flags & REQ_F_ASYNC_DATA) { 1675ed29b0b4SJens Axboe kfree(req->async_data); 1676ed29b0b4SJens Axboe req->async_data = NULL; 1677ed29b0b4SJens Axboe } 1678ed29b0b4SJens Axboe req->flags &= ~IO_REQ_CLEAN_FLAGS; 1679ed29b0b4SJens Axboe } 1680ed29b0b4SJens Axboe 1681ed29b0b4SJens Axboe static bool io_assign_file(struct io_kiocb *req, unsigned int issue_flags) 1682ed29b0b4SJens Axboe { 1683ed29b0b4SJens Axboe if (req->file || !io_op_defs[req->opcode].needs_file) 1684ed29b0b4SJens Axboe return true; 1685ed29b0b4SJens Axboe 1686ed29b0b4SJens Axboe if (req->flags & REQ_F_FIXED_FILE) 1687ed29b0b4SJens Axboe req->file = io_file_get_fixed(req, req->cqe.fd, issue_flags); 1688ed29b0b4SJens Axboe else 1689ed29b0b4SJens Axboe req->file = io_file_get_normal(req, req->cqe.fd); 1690ed29b0b4SJens Axboe 1691ed29b0b4SJens Axboe return !!req->file; 1692ed29b0b4SJens Axboe } 1693ed29b0b4SJens Axboe 1694ed29b0b4SJens Axboe static int io_issue_sqe(struct io_kiocb *req, unsigned int issue_flags) 1695ed29b0b4SJens Axboe { 1696ed29b0b4SJens Axboe const struct io_op_def *def = &io_op_defs[req->opcode]; 1697ed29b0b4SJens Axboe const struct cred *creds = NULL; 1698ed29b0b4SJens Axboe int ret; 1699ed29b0b4SJens Axboe 1700ed29b0b4SJens Axboe if (unlikely(!io_assign_file(req, issue_flags))) 1701ed29b0b4SJens Axboe return -EBADF; 1702ed29b0b4SJens Axboe 1703ed29b0b4SJens Axboe if (unlikely((req->flags & REQ_F_CREDS) && req->creds != current_cred())) 1704ed29b0b4SJens Axboe creds = override_creds(req->creds); 1705ed29b0b4SJens Axboe 1706ed29b0b4SJens Axboe if (!def->audit_skip) 1707ed29b0b4SJens Axboe audit_uring_entry(req->opcode); 1708ed29b0b4SJens Axboe 1709ed29b0b4SJens Axboe ret = def->issue(req, issue_flags); 1710ed29b0b4SJens Axboe 1711ed29b0b4SJens Axboe if (!def->audit_skip) 1712ed29b0b4SJens Axboe audit_uring_exit(!ret, ret); 1713ed29b0b4SJens Axboe 1714ed29b0b4SJens Axboe if (creds) 1715ed29b0b4SJens Axboe revert_creds(creds); 171697b388d7SJens Axboe 171775d7b3aeSPavel Begunkov if (ret == IOU_OK) { 171875d7b3aeSPavel Begunkov if (issue_flags & IO_URING_F_COMPLETE_DEFER) 17199da070b1SPavel Begunkov io_req_complete_defer(req); 172075d7b3aeSPavel Begunkov else 172175d7b3aeSPavel Begunkov io_req_complete_post(req); 172275d7b3aeSPavel Begunkov } else if (ret != IOU_ISSUE_SKIP_COMPLETE) 1723ed29b0b4SJens Axboe return ret; 172497b388d7SJens Axboe 1725ed29b0b4SJens Axboe /* If the op doesn't have a file, we're not polling for it */ 1726ed29b0b4SJens Axboe if ((req->ctx->flags & IORING_SETUP_IOPOLL) && req->file) 1727ed29b0b4SJens Axboe io_iopoll_req_issued(req, issue_flags); 1728ed29b0b4SJens Axboe 1729ed29b0b4SJens Axboe return 0; 1730ed29b0b4SJens Axboe } 1731ed29b0b4SJens Axboe 1732329061d3SJens Axboe int io_poll_issue(struct io_kiocb *req, bool *locked) 1733329061d3SJens Axboe { 1734329061d3SJens Axboe io_tw_lock(req->ctx, locked); 1735329061d3SJens Axboe if (unlikely(req->task->flags & PF_EXITING)) 1736329061d3SJens Axboe return -EFAULT; 173791482864SPavel Begunkov return io_issue_sqe(req, IO_URING_F_NONBLOCK|IO_URING_F_MULTISHOT); 1738329061d3SJens Axboe } 1739329061d3SJens Axboe 1740c9f06aa7SJens Axboe struct io_wq_work *io_wq_free_work(struct io_wq_work *work) 1741ed29b0b4SJens Axboe { 1742ed29b0b4SJens Axboe struct io_kiocb *req = container_of(work, struct io_kiocb, work); 1743ed29b0b4SJens Axboe 1744ed29b0b4SJens Axboe req = io_put_req_find_next(req); 1745ed29b0b4SJens Axboe return req ? &req->work : NULL; 1746ed29b0b4SJens Axboe } 1747ed29b0b4SJens Axboe 1748c9f06aa7SJens Axboe void io_wq_submit_work(struct io_wq_work *work) 1749ed29b0b4SJens Axboe { 1750ed29b0b4SJens Axboe struct io_kiocb *req = container_of(work, struct io_kiocb, work); 1751ed29b0b4SJens Axboe const struct io_op_def *def = &io_op_defs[req->opcode]; 1752ed29b0b4SJens Axboe unsigned int issue_flags = IO_URING_F_UNLOCKED; 1753ed29b0b4SJens Axboe bool needs_poll = false; 1754ed29b0b4SJens Axboe int ret = 0, err = -ECANCELED; 1755ed29b0b4SJens Axboe 175623a6c9acSLin Ma /* one will be dropped by ->io_wq_free_work() after returning to io-wq */ 1757ed29b0b4SJens Axboe if (!(req->flags & REQ_F_REFCOUNT)) 1758ed29b0b4SJens Axboe __io_req_set_refcount(req, 2); 1759ed29b0b4SJens Axboe else 1760ed29b0b4SJens Axboe req_ref_get(req); 1761ed29b0b4SJens Axboe 1762ed29b0b4SJens Axboe io_arm_ltimeout(req); 1763ed29b0b4SJens Axboe 1764ed29b0b4SJens Axboe /* either cancelled or io-wq is dying, so don't touch tctx->iowq */ 1765ed29b0b4SJens Axboe if (work->flags & IO_WQ_WORK_CANCEL) { 1766ed29b0b4SJens Axboe fail: 1767ed29b0b4SJens Axboe io_req_task_queue_fail(req, err); 1768ed29b0b4SJens Axboe return; 1769ed29b0b4SJens Axboe } 1770ed29b0b4SJens Axboe if (!io_assign_file(req, issue_flags)) { 1771ed29b0b4SJens Axboe err = -EBADF; 1772ed29b0b4SJens Axboe work->flags |= IO_WQ_WORK_CANCEL; 1773ed29b0b4SJens Axboe goto fail; 1774ed29b0b4SJens Axboe } 1775ed29b0b4SJens Axboe 1776ed29b0b4SJens Axboe if (req->flags & REQ_F_FORCE_ASYNC) { 1777ed29b0b4SJens Axboe bool opcode_poll = def->pollin || def->pollout; 1778ed29b0b4SJens Axboe 1779ed29b0b4SJens Axboe if (opcode_poll && file_can_poll(req->file)) { 1780ed29b0b4SJens Axboe needs_poll = true; 1781ed29b0b4SJens Axboe issue_flags |= IO_URING_F_NONBLOCK; 1782ed29b0b4SJens Axboe } 1783ed29b0b4SJens Axboe } 1784ed29b0b4SJens Axboe 1785ed29b0b4SJens Axboe do { 1786ed29b0b4SJens Axboe ret = io_issue_sqe(req, issue_flags); 1787ed29b0b4SJens Axboe if (ret != -EAGAIN) 1788ed29b0b4SJens Axboe break; 1789ed29b0b4SJens Axboe /* 1790ed29b0b4SJens Axboe * We can get EAGAIN for iopolled IO even though we're 1791ed29b0b4SJens Axboe * forcing a sync submission from here, since we can't 1792ed29b0b4SJens Axboe * wait for request slots on the block side. 1793ed29b0b4SJens Axboe */ 1794ed29b0b4SJens Axboe if (!needs_poll) { 1795ed29b0b4SJens Axboe if (!(req->ctx->flags & IORING_SETUP_IOPOLL)) 1796ed29b0b4SJens Axboe break; 1797ed29b0b4SJens Axboe cond_resched(); 1798ed29b0b4SJens Axboe continue; 1799ed29b0b4SJens Axboe } 1800ed29b0b4SJens Axboe 1801ed29b0b4SJens Axboe if (io_arm_poll_handler(req, issue_flags) == IO_APOLL_OK) 1802ed29b0b4SJens Axboe return; 1803ed29b0b4SJens Axboe /* aborted or ready, in either case retry blocking */ 1804ed29b0b4SJens Axboe needs_poll = false; 1805ed29b0b4SJens Axboe issue_flags &= ~IO_URING_F_NONBLOCK; 1806ed29b0b4SJens Axboe } while (1); 1807ed29b0b4SJens Axboe 1808ed29b0b4SJens Axboe /* avoid locking problems by failing it from a clean context */ 180997b388d7SJens Axboe if (ret < 0) 1810ed29b0b4SJens Axboe io_req_task_queue_fail(req, ret); 1811ed29b0b4SJens Axboe } 1812ed29b0b4SJens Axboe 1813531113bbSJens Axboe inline struct file *io_file_get_fixed(struct io_kiocb *req, int fd, 1814ed29b0b4SJens Axboe unsigned int issue_flags) 1815ed29b0b4SJens Axboe { 1816ed29b0b4SJens Axboe struct io_ring_ctx *ctx = req->ctx; 1817ed29b0b4SJens Axboe struct file *file = NULL; 1818ed29b0b4SJens Axboe unsigned long file_ptr; 1819ed29b0b4SJens Axboe 1820ed29b0b4SJens Axboe io_ring_submit_lock(ctx, issue_flags); 1821ed29b0b4SJens Axboe 1822ed29b0b4SJens Axboe if (unlikely((unsigned int)fd >= ctx->nr_user_files)) 1823ed29b0b4SJens Axboe goto out; 1824ed29b0b4SJens Axboe fd = array_index_nospec(fd, ctx->nr_user_files); 1825ed29b0b4SJens Axboe file_ptr = io_fixed_file_slot(&ctx->file_table, fd)->file_ptr; 1826ed29b0b4SJens Axboe file = (struct file *) (file_ptr & FFS_MASK); 1827ed29b0b4SJens Axboe file_ptr &= ~FFS_MASK; 1828ed29b0b4SJens Axboe /* mask in overlapping REQ_F and FFS bits */ 1829ed29b0b4SJens Axboe req->flags |= (file_ptr << REQ_F_SUPPORT_NOWAIT_BIT); 1830ed29b0b4SJens Axboe io_req_set_rsrc_node(req, ctx, 0); 1831ed29b0b4SJens Axboe out: 1832ed29b0b4SJens Axboe io_ring_submit_unlock(ctx, issue_flags); 1833ed29b0b4SJens Axboe return file; 1834ed29b0b4SJens Axboe } 1835ed29b0b4SJens Axboe 1836531113bbSJens Axboe struct file *io_file_get_normal(struct io_kiocb *req, int fd) 1837ed29b0b4SJens Axboe { 1838ed29b0b4SJens Axboe struct file *file = fget(fd); 1839ed29b0b4SJens Axboe 184048863ffdSPavel Begunkov trace_io_uring_file_get(req, fd); 1841ed29b0b4SJens Axboe 1842ed29b0b4SJens Axboe /* we don't allow fixed io_uring files */ 1843e5550a14SJens Axboe if (file && io_is_uring_fops(file)) 1844ed29b0b4SJens Axboe io_req_track_inflight(req); 1845ed29b0b4SJens Axboe return file; 1846ed29b0b4SJens Axboe } 1847ed29b0b4SJens Axboe 1848ed29b0b4SJens Axboe static void io_queue_async(struct io_kiocb *req, int ret) 1849ed29b0b4SJens Axboe __must_hold(&req->ctx->uring_lock) 1850ed29b0b4SJens Axboe { 1851ed29b0b4SJens Axboe struct io_kiocb *linked_timeout; 1852ed29b0b4SJens Axboe 1853ed29b0b4SJens Axboe if (ret != -EAGAIN || (req->flags & REQ_F_NOWAIT)) { 1854ed29b0b4SJens Axboe io_req_complete_failed(req, ret); 1855ed29b0b4SJens Axboe return; 1856ed29b0b4SJens Axboe } 1857ed29b0b4SJens Axboe 1858ed29b0b4SJens Axboe linked_timeout = io_prep_linked_timeout(req); 1859ed29b0b4SJens Axboe 1860ed29b0b4SJens Axboe switch (io_arm_poll_handler(req, 0)) { 1861ed29b0b4SJens Axboe case IO_APOLL_READY: 1862336d28a8SPavel Begunkov io_kbuf_recycle(req, 0); 1863ed29b0b4SJens Axboe io_req_task_queue(req); 1864ed29b0b4SJens Axboe break; 1865ed29b0b4SJens Axboe case IO_APOLL_ABORTED: 1866ed29b0b4SJens Axboe io_kbuf_recycle(req, 0); 1867ed29b0b4SJens Axboe io_queue_iowq(req, NULL); 1868ed29b0b4SJens Axboe break; 1869ed29b0b4SJens Axboe case IO_APOLL_OK: 1870ed29b0b4SJens Axboe break; 1871ed29b0b4SJens Axboe } 1872ed29b0b4SJens Axboe 1873ed29b0b4SJens Axboe if (linked_timeout) 1874ed29b0b4SJens Axboe io_queue_linked_timeout(linked_timeout); 1875ed29b0b4SJens Axboe } 1876ed29b0b4SJens Axboe 1877ed29b0b4SJens Axboe static inline void io_queue_sqe(struct io_kiocb *req) 1878ed29b0b4SJens Axboe __must_hold(&req->ctx->uring_lock) 1879ed29b0b4SJens Axboe { 1880ed29b0b4SJens Axboe int ret; 1881ed29b0b4SJens Axboe 1882ed29b0b4SJens Axboe ret = io_issue_sqe(req, IO_URING_F_NONBLOCK|IO_URING_F_COMPLETE_DEFER); 1883ed29b0b4SJens Axboe 1884ed29b0b4SJens Axboe /* 1885ed29b0b4SJens Axboe * We async punt it if the file wasn't marked NOWAIT, or if the file 1886ed29b0b4SJens Axboe * doesn't support non-blocking read/write attempts 1887ed29b0b4SJens Axboe */ 1888ed29b0b4SJens Axboe if (likely(!ret)) 1889ed29b0b4SJens Axboe io_arm_ltimeout(req); 1890ed29b0b4SJens Axboe else 1891ed29b0b4SJens Axboe io_queue_async(req, ret); 1892ed29b0b4SJens Axboe } 1893ed29b0b4SJens Axboe 1894ed29b0b4SJens Axboe static void io_queue_sqe_fallback(struct io_kiocb *req) 1895ed29b0b4SJens Axboe __must_hold(&req->ctx->uring_lock) 1896ed29b0b4SJens Axboe { 1897ed29b0b4SJens Axboe if (unlikely(req->flags & REQ_F_FAIL)) { 1898ed29b0b4SJens Axboe /* 1899ed29b0b4SJens Axboe * We don't submit, fail them all, for that replace hardlinks 1900ed29b0b4SJens Axboe * with normal links. Extra REQ_F_LINK is tolerated. 1901ed29b0b4SJens Axboe */ 1902ed29b0b4SJens Axboe req->flags &= ~REQ_F_HARDLINK; 1903ed29b0b4SJens Axboe req->flags |= REQ_F_LINK; 1904ed29b0b4SJens Axboe io_req_complete_failed(req, req->cqe.res); 1905ed29b0b4SJens Axboe } else if (unlikely(req->ctx->drain_active)) { 1906ed29b0b4SJens Axboe io_drain_req(req); 1907ed29b0b4SJens Axboe } else { 1908ed29b0b4SJens Axboe int ret = io_req_prep_async(req); 1909ed29b0b4SJens Axboe 1910ed29b0b4SJens Axboe if (unlikely(ret)) 1911ed29b0b4SJens Axboe io_req_complete_failed(req, ret); 1912ed29b0b4SJens Axboe else 1913ed29b0b4SJens Axboe io_queue_iowq(req, NULL); 1914ed29b0b4SJens Axboe } 1915ed29b0b4SJens Axboe } 1916ed29b0b4SJens Axboe 1917ed29b0b4SJens Axboe /* 1918ed29b0b4SJens Axboe * Check SQE restrictions (opcode and flags). 1919ed29b0b4SJens Axboe * 1920ed29b0b4SJens Axboe * Returns 'true' if SQE is allowed, 'false' otherwise. 1921ed29b0b4SJens Axboe */ 1922ed29b0b4SJens Axboe static inline bool io_check_restriction(struct io_ring_ctx *ctx, 1923ed29b0b4SJens Axboe struct io_kiocb *req, 1924ed29b0b4SJens Axboe unsigned int sqe_flags) 1925ed29b0b4SJens Axboe { 1926ed29b0b4SJens Axboe if (!test_bit(req->opcode, ctx->restrictions.sqe_op)) 1927ed29b0b4SJens Axboe return false; 1928ed29b0b4SJens Axboe 1929ed29b0b4SJens Axboe if ((sqe_flags & ctx->restrictions.sqe_flags_required) != 1930ed29b0b4SJens Axboe ctx->restrictions.sqe_flags_required) 1931ed29b0b4SJens Axboe return false; 1932ed29b0b4SJens Axboe 1933ed29b0b4SJens Axboe if (sqe_flags & ~(ctx->restrictions.sqe_flags_allowed | 1934ed29b0b4SJens Axboe ctx->restrictions.sqe_flags_required)) 1935ed29b0b4SJens Axboe return false; 1936ed29b0b4SJens Axboe 1937ed29b0b4SJens Axboe return true; 1938ed29b0b4SJens Axboe } 1939ed29b0b4SJens Axboe 1940ed29b0b4SJens Axboe static void io_init_req_drain(struct io_kiocb *req) 1941ed29b0b4SJens Axboe { 1942ed29b0b4SJens Axboe struct io_ring_ctx *ctx = req->ctx; 1943ed29b0b4SJens Axboe struct io_kiocb *head = ctx->submit_state.link.head; 1944ed29b0b4SJens Axboe 1945ed29b0b4SJens Axboe ctx->drain_active = true; 1946ed29b0b4SJens Axboe if (head) { 1947ed29b0b4SJens Axboe /* 1948ed29b0b4SJens Axboe * If we need to drain a request in the middle of a link, drain 1949ed29b0b4SJens Axboe * the head request and the next request/link after the current 1950ed29b0b4SJens Axboe * link. Considering sequential execution of links, 1951ed29b0b4SJens Axboe * REQ_F_IO_DRAIN will be maintained for every request of our 1952ed29b0b4SJens Axboe * link. 1953ed29b0b4SJens Axboe */ 1954ed29b0b4SJens Axboe head->flags |= REQ_F_IO_DRAIN | REQ_F_FORCE_ASYNC; 1955ed29b0b4SJens Axboe ctx->drain_next = true; 1956ed29b0b4SJens Axboe } 1957ed29b0b4SJens Axboe } 1958ed29b0b4SJens Axboe 1959ed29b0b4SJens Axboe static int io_init_req(struct io_ring_ctx *ctx, struct io_kiocb *req, 1960ed29b0b4SJens Axboe const struct io_uring_sqe *sqe) 1961ed29b0b4SJens Axboe __must_hold(&ctx->uring_lock) 1962ed29b0b4SJens Axboe { 1963ed29b0b4SJens Axboe const struct io_op_def *def; 1964ed29b0b4SJens Axboe unsigned int sqe_flags; 1965ed29b0b4SJens Axboe int personality; 1966ed29b0b4SJens Axboe u8 opcode; 1967ed29b0b4SJens Axboe 1968ed29b0b4SJens Axboe /* req is partially pre-initialised, see io_preinit_req() */ 1969ed29b0b4SJens Axboe req->opcode = opcode = READ_ONCE(sqe->opcode); 1970ed29b0b4SJens Axboe /* same numerical values with corresponding REQ_F_*, safe to copy */ 1971ed29b0b4SJens Axboe req->flags = sqe_flags = READ_ONCE(sqe->flags); 1972ed29b0b4SJens Axboe req->cqe.user_data = READ_ONCE(sqe->user_data); 1973ed29b0b4SJens Axboe req->file = NULL; 1974ed29b0b4SJens Axboe req->rsrc_node = NULL; 1975ed29b0b4SJens Axboe req->task = current; 1976ed29b0b4SJens Axboe 1977ed29b0b4SJens Axboe if (unlikely(opcode >= IORING_OP_LAST)) { 1978ed29b0b4SJens Axboe req->opcode = 0; 1979ed29b0b4SJens Axboe return -EINVAL; 1980ed29b0b4SJens Axboe } 1981ed29b0b4SJens Axboe def = &io_op_defs[opcode]; 1982ed29b0b4SJens Axboe if (unlikely(sqe_flags & ~SQE_COMMON_FLAGS)) { 1983ed29b0b4SJens Axboe /* enforce forwards compatibility on users */ 1984ed29b0b4SJens Axboe if (sqe_flags & ~SQE_VALID_FLAGS) 1985ed29b0b4SJens Axboe return -EINVAL; 1986ed29b0b4SJens Axboe if (sqe_flags & IOSQE_BUFFER_SELECT) { 1987ed29b0b4SJens Axboe if (!def->buffer_select) 1988ed29b0b4SJens Axboe return -EOPNOTSUPP; 1989ed29b0b4SJens Axboe req->buf_index = READ_ONCE(sqe->buf_group); 1990ed29b0b4SJens Axboe } 1991ed29b0b4SJens Axboe if (sqe_flags & IOSQE_CQE_SKIP_SUCCESS) 1992ed29b0b4SJens Axboe ctx->drain_disabled = true; 1993ed29b0b4SJens Axboe if (sqe_flags & IOSQE_IO_DRAIN) { 1994ed29b0b4SJens Axboe if (ctx->drain_disabled) 1995ed29b0b4SJens Axboe return -EOPNOTSUPP; 1996ed29b0b4SJens Axboe io_init_req_drain(req); 1997ed29b0b4SJens Axboe } 1998ed29b0b4SJens Axboe } 1999ed29b0b4SJens Axboe if (unlikely(ctx->restricted || ctx->drain_active || ctx->drain_next)) { 2000ed29b0b4SJens Axboe if (ctx->restricted && !io_check_restriction(ctx, req, sqe_flags)) 2001ed29b0b4SJens Axboe return -EACCES; 2002ed29b0b4SJens Axboe /* knock it to the slow queue path, will be drained there */ 2003ed29b0b4SJens Axboe if (ctx->drain_active) 2004ed29b0b4SJens Axboe req->flags |= REQ_F_FORCE_ASYNC; 2005ed29b0b4SJens Axboe /* if there is no link, we're at "next" request and need to drain */ 2006ed29b0b4SJens Axboe if (unlikely(ctx->drain_next) && !ctx->submit_state.link.head) { 2007ed29b0b4SJens Axboe ctx->drain_next = false; 2008ed29b0b4SJens Axboe ctx->drain_active = true; 2009ed29b0b4SJens Axboe req->flags |= REQ_F_IO_DRAIN | REQ_F_FORCE_ASYNC; 2010ed29b0b4SJens Axboe } 2011ed29b0b4SJens Axboe } 2012ed29b0b4SJens Axboe 2013ed29b0b4SJens Axboe if (!def->ioprio && sqe->ioprio) 2014ed29b0b4SJens Axboe return -EINVAL; 2015ed29b0b4SJens Axboe if (!def->iopoll && (ctx->flags & IORING_SETUP_IOPOLL)) 2016ed29b0b4SJens Axboe return -EINVAL; 2017ed29b0b4SJens Axboe 2018ed29b0b4SJens Axboe if (def->needs_file) { 2019ed29b0b4SJens Axboe struct io_submit_state *state = &ctx->submit_state; 2020ed29b0b4SJens Axboe 2021ed29b0b4SJens Axboe req->cqe.fd = READ_ONCE(sqe->fd); 2022ed29b0b4SJens Axboe 2023ed29b0b4SJens Axboe /* 2024ed29b0b4SJens Axboe * Plug now if we have more than 2 IO left after this, and the 2025ed29b0b4SJens Axboe * target is potentially a read/write to block based storage. 2026ed29b0b4SJens Axboe */ 2027ed29b0b4SJens Axboe if (state->need_plug && def->plug) { 2028ed29b0b4SJens Axboe state->plug_started = true; 2029ed29b0b4SJens Axboe state->need_plug = false; 2030ed29b0b4SJens Axboe blk_start_plug_nr_ios(&state->plug, state->submit_nr); 2031ed29b0b4SJens Axboe } 2032ed29b0b4SJens Axboe } 2033ed29b0b4SJens Axboe 2034ed29b0b4SJens Axboe personality = READ_ONCE(sqe->personality); 2035ed29b0b4SJens Axboe if (personality) { 2036ed29b0b4SJens Axboe int ret; 2037ed29b0b4SJens Axboe 2038ed29b0b4SJens Axboe req->creds = xa_load(&ctx->personalities, personality); 2039ed29b0b4SJens Axboe if (!req->creds) 2040ed29b0b4SJens Axboe return -EINVAL; 2041ed29b0b4SJens Axboe get_cred(req->creds); 2042ed29b0b4SJens Axboe ret = security_uring_override_creds(req->creds); 2043ed29b0b4SJens Axboe if (ret) { 2044ed29b0b4SJens Axboe put_cred(req->creds); 2045ed29b0b4SJens Axboe return ret; 2046ed29b0b4SJens Axboe } 2047ed29b0b4SJens Axboe req->flags |= REQ_F_CREDS; 2048ed29b0b4SJens Axboe } 2049ed29b0b4SJens Axboe 2050ed29b0b4SJens Axboe return def->prep(req, sqe); 2051ed29b0b4SJens Axboe } 2052ed29b0b4SJens Axboe 2053ed29b0b4SJens Axboe static __cold int io_submit_fail_init(const struct io_uring_sqe *sqe, 2054ed29b0b4SJens Axboe struct io_kiocb *req, int ret) 2055ed29b0b4SJens Axboe { 2056ed29b0b4SJens Axboe struct io_ring_ctx *ctx = req->ctx; 2057ed29b0b4SJens Axboe struct io_submit_link *link = &ctx->submit_state.link; 2058ed29b0b4SJens Axboe struct io_kiocb *head = link->head; 2059ed29b0b4SJens Axboe 206048863ffdSPavel Begunkov trace_io_uring_req_failed(sqe, req, ret); 2061ed29b0b4SJens Axboe 2062ed29b0b4SJens Axboe /* 2063ed29b0b4SJens Axboe * Avoid breaking links in the middle as it renders links with SQPOLL 2064ed29b0b4SJens Axboe * unusable. Instead of failing eagerly, continue assembling the link if 2065ed29b0b4SJens Axboe * applicable and mark the head with REQ_F_FAIL. The link flushing code 2066ed29b0b4SJens Axboe * should find the flag and handle the rest. 2067ed29b0b4SJens Axboe */ 2068ed29b0b4SJens Axboe req_fail_link_node(req, ret); 2069ed29b0b4SJens Axboe if (head && !(head->flags & REQ_F_FAIL)) 2070ed29b0b4SJens Axboe req_fail_link_node(head, -ECANCELED); 2071ed29b0b4SJens Axboe 2072ed29b0b4SJens Axboe if (!(req->flags & IO_REQ_LINK_FLAGS)) { 2073ed29b0b4SJens Axboe if (head) { 2074ed29b0b4SJens Axboe link->last->link = req; 2075ed29b0b4SJens Axboe link->head = NULL; 2076ed29b0b4SJens Axboe req = head; 2077ed29b0b4SJens Axboe } 2078ed29b0b4SJens Axboe io_queue_sqe_fallback(req); 2079ed29b0b4SJens Axboe return ret; 2080ed29b0b4SJens Axboe } 2081ed29b0b4SJens Axboe 2082ed29b0b4SJens Axboe if (head) 2083ed29b0b4SJens Axboe link->last->link = req; 2084ed29b0b4SJens Axboe else 2085ed29b0b4SJens Axboe link->head = req; 2086ed29b0b4SJens Axboe link->last = req; 2087ed29b0b4SJens Axboe return 0; 2088ed29b0b4SJens Axboe } 2089ed29b0b4SJens Axboe 2090ed29b0b4SJens Axboe static inline int io_submit_sqe(struct io_ring_ctx *ctx, struct io_kiocb *req, 2091ed29b0b4SJens Axboe const struct io_uring_sqe *sqe) 2092ed29b0b4SJens Axboe __must_hold(&ctx->uring_lock) 2093ed29b0b4SJens Axboe { 2094ed29b0b4SJens Axboe struct io_submit_link *link = &ctx->submit_state.link; 2095ed29b0b4SJens Axboe int ret; 2096ed29b0b4SJens Axboe 2097ed29b0b4SJens Axboe ret = io_init_req(ctx, req, sqe); 2098ed29b0b4SJens Axboe if (unlikely(ret)) 2099ed29b0b4SJens Axboe return io_submit_fail_init(sqe, req, ret); 2100ed29b0b4SJens Axboe 2101ed29b0b4SJens Axboe /* don't need @sqe from now on */ 210248863ffdSPavel Begunkov trace_io_uring_submit_sqe(req, true); 2103ed29b0b4SJens Axboe 2104ed29b0b4SJens Axboe /* 2105ed29b0b4SJens Axboe * If we already have a head request, queue this one for async 2106ed29b0b4SJens Axboe * submittal once the head completes. If we don't have a head but 2107ed29b0b4SJens Axboe * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be 2108ed29b0b4SJens Axboe * submitted sync once the chain is complete. If none of those 2109ed29b0b4SJens Axboe * conditions are true (normal request), then just queue it. 2110ed29b0b4SJens Axboe */ 2111ed29b0b4SJens Axboe if (unlikely(link->head)) { 2112ed29b0b4SJens Axboe ret = io_req_prep_async(req); 2113ed29b0b4SJens Axboe if (unlikely(ret)) 2114ed29b0b4SJens Axboe return io_submit_fail_init(sqe, req, ret); 2115ed29b0b4SJens Axboe 211648863ffdSPavel Begunkov trace_io_uring_link(req, link->head); 2117ed29b0b4SJens Axboe link->last->link = req; 2118ed29b0b4SJens Axboe link->last = req; 2119ed29b0b4SJens Axboe 2120ed29b0b4SJens Axboe if (req->flags & IO_REQ_LINK_FLAGS) 2121ed29b0b4SJens Axboe return 0; 2122ed29b0b4SJens Axboe /* last request of the link, flush it */ 2123ed29b0b4SJens Axboe req = link->head; 2124ed29b0b4SJens Axboe link->head = NULL; 2125ed29b0b4SJens Axboe if (req->flags & (REQ_F_FORCE_ASYNC | REQ_F_FAIL)) 2126ed29b0b4SJens Axboe goto fallback; 2127ed29b0b4SJens Axboe 2128ed29b0b4SJens Axboe } else if (unlikely(req->flags & (IO_REQ_LINK_FLAGS | 2129ed29b0b4SJens Axboe REQ_F_FORCE_ASYNC | REQ_F_FAIL))) { 2130ed29b0b4SJens Axboe if (req->flags & IO_REQ_LINK_FLAGS) { 2131ed29b0b4SJens Axboe link->head = req; 2132ed29b0b4SJens Axboe link->last = req; 2133ed29b0b4SJens Axboe } else { 2134ed29b0b4SJens Axboe fallback: 2135ed29b0b4SJens Axboe io_queue_sqe_fallback(req); 2136ed29b0b4SJens Axboe } 2137ed29b0b4SJens Axboe return 0; 2138ed29b0b4SJens Axboe } 2139ed29b0b4SJens Axboe 2140ed29b0b4SJens Axboe io_queue_sqe(req); 2141ed29b0b4SJens Axboe return 0; 2142ed29b0b4SJens Axboe } 2143ed29b0b4SJens Axboe 2144ed29b0b4SJens Axboe /* 2145ed29b0b4SJens Axboe * Batched submission is done, ensure local IO is flushed out. 2146ed29b0b4SJens Axboe */ 2147ed29b0b4SJens Axboe static void io_submit_state_end(struct io_ring_ctx *ctx) 2148ed29b0b4SJens Axboe { 2149ed29b0b4SJens Axboe struct io_submit_state *state = &ctx->submit_state; 2150ed29b0b4SJens Axboe 2151ed29b0b4SJens Axboe if (unlikely(state->link.head)) 2152ed29b0b4SJens Axboe io_queue_sqe_fallback(state->link.head); 2153ed29b0b4SJens Axboe /* flush only after queuing links as they can generate completions */ 2154ed29b0b4SJens Axboe io_submit_flush_completions(ctx); 2155ed29b0b4SJens Axboe if (state->plug_started) 2156ed29b0b4SJens Axboe blk_finish_plug(&state->plug); 2157ed29b0b4SJens Axboe } 2158ed29b0b4SJens Axboe 2159ed29b0b4SJens Axboe /* 2160ed29b0b4SJens Axboe * Start submission side cache. 2161ed29b0b4SJens Axboe */ 2162ed29b0b4SJens Axboe static void io_submit_state_start(struct io_submit_state *state, 2163ed29b0b4SJens Axboe unsigned int max_ios) 2164ed29b0b4SJens Axboe { 2165ed29b0b4SJens Axboe state->plug_started = false; 2166ed29b0b4SJens Axboe state->need_plug = max_ios > 2; 2167ed29b0b4SJens Axboe state->submit_nr = max_ios; 2168ed29b0b4SJens Axboe /* set only head, no need to init link_last in advance */ 2169ed29b0b4SJens Axboe state->link.head = NULL; 2170ed29b0b4SJens Axboe } 2171ed29b0b4SJens Axboe 2172ed29b0b4SJens Axboe static void io_commit_sqring(struct io_ring_ctx *ctx) 2173ed29b0b4SJens Axboe { 2174ed29b0b4SJens Axboe struct io_rings *rings = ctx->rings; 2175ed29b0b4SJens Axboe 2176ed29b0b4SJens Axboe /* 2177ed29b0b4SJens Axboe * Ensure any loads from the SQEs are done at this point, 2178ed29b0b4SJens Axboe * since once we write the new head, the application could 2179ed29b0b4SJens Axboe * write new data to them. 2180ed29b0b4SJens Axboe */ 2181ed29b0b4SJens Axboe smp_store_release(&rings->sq.head, ctx->cached_sq_head); 2182ed29b0b4SJens Axboe } 2183ed29b0b4SJens Axboe 2184ed29b0b4SJens Axboe /* 2185ed29b0b4SJens Axboe * Fetch an sqe, if one is available. Note this returns a pointer to memory 2186ed29b0b4SJens Axboe * that is mapped by userspace. This means that care needs to be taken to 2187ed29b0b4SJens Axboe * ensure that reads are stable, as we cannot rely on userspace always 2188ed29b0b4SJens Axboe * being a good citizen. If members of the sqe are validated and then later 2189ed29b0b4SJens Axboe * used, it's important that those reads are done through READ_ONCE() to 2190ed29b0b4SJens Axboe * prevent a re-load down the line. 2191ed29b0b4SJens Axboe */ 2192ed29b0b4SJens Axboe static const struct io_uring_sqe *io_get_sqe(struct io_ring_ctx *ctx) 2193ed29b0b4SJens Axboe { 2194ed29b0b4SJens Axboe unsigned head, mask = ctx->sq_entries - 1; 2195ed29b0b4SJens Axboe unsigned sq_idx = ctx->cached_sq_head++ & mask; 2196ed29b0b4SJens Axboe 2197ed29b0b4SJens Axboe /* 2198ed29b0b4SJens Axboe * The cached sq head (or cq tail) serves two purposes: 2199ed29b0b4SJens Axboe * 2200ed29b0b4SJens Axboe * 1) allows us to batch the cost of updating the user visible 2201ed29b0b4SJens Axboe * head updates. 2202ed29b0b4SJens Axboe * 2) allows the kernel side to track the head on its own, even 2203ed29b0b4SJens Axboe * though the application is the one updating it. 2204ed29b0b4SJens Axboe */ 2205ed29b0b4SJens Axboe head = READ_ONCE(ctx->sq_array[sq_idx]); 2206ed29b0b4SJens Axboe if (likely(head < ctx->sq_entries)) { 2207ed29b0b4SJens Axboe /* double index for 128-byte SQEs, twice as long */ 2208ed29b0b4SJens Axboe if (ctx->flags & IORING_SETUP_SQE128) 2209ed29b0b4SJens Axboe head <<= 1; 2210ed29b0b4SJens Axboe return &ctx->sq_sqes[head]; 2211ed29b0b4SJens Axboe } 2212ed29b0b4SJens Axboe 2213ed29b0b4SJens Axboe /* drop invalid entries */ 2214ed29b0b4SJens Axboe ctx->cq_extra--; 2215ed29b0b4SJens Axboe WRITE_ONCE(ctx->rings->sq_dropped, 2216ed29b0b4SJens Axboe READ_ONCE(ctx->rings->sq_dropped) + 1); 2217ed29b0b4SJens Axboe return NULL; 2218ed29b0b4SJens Axboe } 2219ed29b0b4SJens Axboe 222017437f31SJens Axboe int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr) 2221ed29b0b4SJens Axboe __must_hold(&ctx->uring_lock) 2222ed29b0b4SJens Axboe { 2223ed29b0b4SJens Axboe unsigned int entries = io_sqring_entries(ctx); 2224ed29b0b4SJens Axboe unsigned int left; 2225ed29b0b4SJens Axboe int ret; 2226ed29b0b4SJens Axboe 2227ed29b0b4SJens Axboe if (unlikely(!entries)) 2228ed29b0b4SJens Axboe return 0; 2229ed29b0b4SJens Axboe /* make sure SQ entry isn't read before tail */ 2230ed29b0b4SJens Axboe ret = left = min3(nr, ctx->sq_entries, entries); 2231ed29b0b4SJens Axboe io_get_task_refs(left); 2232ed29b0b4SJens Axboe io_submit_state_start(&ctx->submit_state, left); 2233ed29b0b4SJens Axboe 2234ed29b0b4SJens Axboe do { 2235ed29b0b4SJens Axboe const struct io_uring_sqe *sqe; 2236ed29b0b4SJens Axboe struct io_kiocb *req; 2237ed29b0b4SJens Axboe 2238ed29b0b4SJens Axboe if (unlikely(!io_alloc_req_refill(ctx))) 2239ed29b0b4SJens Axboe break; 2240ed29b0b4SJens Axboe req = io_alloc_req(ctx); 2241ed29b0b4SJens Axboe sqe = io_get_sqe(ctx); 2242ed29b0b4SJens Axboe if (unlikely(!sqe)) { 2243ed29b0b4SJens Axboe io_req_add_to_cache(req, ctx); 2244ed29b0b4SJens Axboe break; 2245ed29b0b4SJens Axboe } 2246ed29b0b4SJens Axboe 2247ed29b0b4SJens Axboe /* 2248ed29b0b4SJens Axboe * Continue submitting even for sqe failure if the 2249ed29b0b4SJens Axboe * ring was setup with IORING_SETUP_SUBMIT_ALL 2250ed29b0b4SJens Axboe */ 2251ed29b0b4SJens Axboe if (unlikely(io_submit_sqe(ctx, req, sqe)) && 2252ed29b0b4SJens Axboe !(ctx->flags & IORING_SETUP_SUBMIT_ALL)) { 2253ed29b0b4SJens Axboe left--; 2254ed29b0b4SJens Axboe break; 2255ed29b0b4SJens Axboe } 2256ed29b0b4SJens Axboe } while (--left); 2257ed29b0b4SJens Axboe 2258ed29b0b4SJens Axboe if (unlikely(left)) { 2259ed29b0b4SJens Axboe ret -= left; 2260ed29b0b4SJens Axboe /* try again if it submitted nothing and can't allocate a req */ 2261ed29b0b4SJens Axboe if (!ret && io_req_cache_empty(ctx)) 2262ed29b0b4SJens Axboe ret = -EAGAIN; 2263ed29b0b4SJens Axboe current->io_uring->cached_refs += left; 2264ed29b0b4SJens Axboe } 2265ed29b0b4SJens Axboe 2266ed29b0b4SJens Axboe io_submit_state_end(ctx); 2267ed29b0b4SJens Axboe /* Commit SQ ring head once we've consumed and submitted all SQEs */ 2268ed29b0b4SJens Axboe io_commit_sqring(ctx); 2269ed29b0b4SJens Axboe return ret; 2270ed29b0b4SJens Axboe } 2271ed29b0b4SJens Axboe 2272ed29b0b4SJens Axboe struct io_wait_queue { 2273ed29b0b4SJens Axboe struct wait_queue_entry wq; 2274ed29b0b4SJens Axboe struct io_ring_ctx *ctx; 2275ed29b0b4SJens Axboe unsigned cq_tail; 2276ed29b0b4SJens Axboe unsigned nr_timeouts; 2277ed29b0b4SJens Axboe }; 2278ed29b0b4SJens Axboe 2279b4c98d59SDylan Yudaken static inline bool io_has_work(struct io_ring_ctx *ctx) 2280b4c98d59SDylan Yudaken { 2281c0e0d6baSDylan Yudaken return test_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq) || 2282c0e0d6baSDylan Yudaken ((ctx->flags & IORING_SETUP_DEFER_TASKRUN) && 2283c0e0d6baSDylan Yudaken !llist_empty(&ctx->work_llist)); 2284b4c98d59SDylan Yudaken } 2285b4c98d59SDylan Yudaken 2286ed29b0b4SJens Axboe static inline bool io_should_wake(struct io_wait_queue *iowq) 2287ed29b0b4SJens Axboe { 2288ed29b0b4SJens Axboe struct io_ring_ctx *ctx = iowq->ctx; 22890fc8c2acSDylan Yudaken int dist = READ_ONCE(ctx->rings->cq.tail) - (int) iowq->cq_tail; 2290ed29b0b4SJens Axboe 2291ed29b0b4SJens Axboe /* 2292ed29b0b4SJens Axboe * Wake up if we have enough events, or if a timeout occurred since we 2293ed29b0b4SJens Axboe * started waiting. For timeouts, we always want to return to userspace, 2294ed29b0b4SJens Axboe * regardless of event count. 2295ed29b0b4SJens Axboe */ 2296ed29b0b4SJens Axboe return dist >= 0 || atomic_read(&ctx->cq_timeouts) != iowq->nr_timeouts; 2297ed29b0b4SJens Axboe } 2298ed29b0b4SJens Axboe 2299ed29b0b4SJens Axboe static int io_wake_function(struct wait_queue_entry *curr, unsigned int mode, 2300ed29b0b4SJens Axboe int wake_flags, void *key) 2301ed29b0b4SJens Axboe { 2302ed29b0b4SJens Axboe struct io_wait_queue *iowq = container_of(curr, struct io_wait_queue, 2303ed29b0b4SJens Axboe wq); 2304b4c98d59SDylan Yudaken struct io_ring_ctx *ctx = iowq->ctx; 2305ed29b0b4SJens Axboe 2306ed29b0b4SJens Axboe /* 2307ed29b0b4SJens Axboe * Cannot safely flush overflowed CQEs from here, ensure we wake up 2308ed29b0b4SJens Axboe * the task, and the next invocation will do it. 2309ed29b0b4SJens Axboe */ 2310b4c98d59SDylan Yudaken if (io_should_wake(iowq) || io_has_work(ctx)) 2311ed29b0b4SJens Axboe return autoremove_wake_function(curr, mode, wake_flags, key); 2312ed29b0b4SJens Axboe return -1; 2313ed29b0b4SJens Axboe } 2314ed29b0b4SJens Axboe 2315c0e0d6baSDylan Yudaken int io_run_task_work_sig(struct io_ring_ctx *ctx) 2316ed29b0b4SJens Axboe { 2317c0e0d6baSDylan Yudaken if (io_run_task_work_ctx(ctx) > 0) 2318ed29b0b4SJens Axboe return 1; 2319ed29b0b4SJens Axboe if (task_sigpending(current)) 2320ed29b0b4SJens Axboe return -EINTR; 2321ed29b0b4SJens Axboe return 0; 2322ed29b0b4SJens Axboe } 2323ed29b0b4SJens Axboe 2324ed29b0b4SJens Axboe /* when returns >0, the caller should retry */ 2325ed29b0b4SJens Axboe static inline int io_cqring_wait_schedule(struct io_ring_ctx *ctx, 2326ed29b0b4SJens Axboe struct io_wait_queue *iowq, 2327ed29b0b4SJens Axboe ktime_t timeout) 2328ed29b0b4SJens Axboe { 2329ed29b0b4SJens Axboe int ret; 2330ed29b0b4SJens Axboe unsigned long check_cq; 2331ed29b0b4SJens Axboe 2332ed29b0b4SJens Axboe /* make sure we run task_work before checking for signals */ 2333c0e0d6baSDylan Yudaken ret = io_run_task_work_sig(ctx); 2334ed29b0b4SJens Axboe if (ret || io_should_wake(iowq)) 2335ed29b0b4SJens Axboe return ret; 23363a08576bSPavel Begunkov 2337ed29b0b4SJens Axboe check_cq = READ_ONCE(ctx->check_cq); 23383a08576bSPavel Begunkov if (unlikely(check_cq)) { 2339ed29b0b4SJens Axboe /* let the caller flush overflows, retry */ 2340ed29b0b4SJens Axboe if (check_cq & BIT(IO_CHECK_CQ_OVERFLOW_BIT)) 2341ed29b0b4SJens Axboe return 1; 23423a08576bSPavel Begunkov if (check_cq & BIT(IO_CHECK_CQ_DROPPED_BIT)) 2343ed29b0b4SJens Axboe return -EBADR; 23443a08576bSPavel Begunkov } 2345ed29b0b4SJens Axboe if (!schedule_hrtimeout(&timeout, HRTIMER_MODE_ABS)) 2346ed29b0b4SJens Axboe return -ETIME; 2347ed29b0b4SJens Axboe return 1; 2348ed29b0b4SJens Axboe } 2349ed29b0b4SJens Axboe 2350ed29b0b4SJens Axboe /* 2351ed29b0b4SJens Axboe * Wait until events become available, if we don't already have some. The 2352ed29b0b4SJens Axboe * application must reap them itself, as they reside on the shared cq ring. 2353ed29b0b4SJens Axboe */ 2354ed29b0b4SJens Axboe static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events, 2355ed29b0b4SJens Axboe const sigset_t __user *sig, size_t sigsz, 2356ed29b0b4SJens Axboe struct __kernel_timespec __user *uts) 2357ed29b0b4SJens Axboe { 2358ed29b0b4SJens Axboe struct io_wait_queue iowq; 2359ed29b0b4SJens Axboe struct io_rings *rings = ctx->rings; 2360ed29b0b4SJens Axboe ktime_t timeout = KTIME_MAX; 2361ed29b0b4SJens Axboe int ret; 2362ed29b0b4SJens Axboe 236376de6749SPavel Begunkov if (!io_allowed_run_tw(ctx)) 236476de6749SPavel Begunkov return -EEXIST; 236576de6749SPavel Begunkov 2366ed29b0b4SJens Axboe do { 2367c0e0d6baSDylan Yudaken /* always run at least 1 task work to process local work */ 2368c0e0d6baSDylan Yudaken ret = io_run_task_work_ctx(ctx); 2369c0e0d6baSDylan Yudaken if (ret < 0) 2370c0e0d6baSDylan Yudaken return ret; 2371ed29b0b4SJens Axboe io_cqring_overflow_flush(ctx); 2372aa1df3a3SPavel Begunkov 23730fc8c2acSDylan Yudaken /* if user messes with these they will just get an early return */ 23740fc8c2acSDylan Yudaken if (__io_cqring_events_user(ctx) >= min_events) 2375ed29b0b4SJens Axboe return 0; 2376c0e0d6baSDylan Yudaken } while (ret > 0); 2377ed29b0b4SJens Axboe 2378ed29b0b4SJens Axboe if (sig) { 2379ed29b0b4SJens Axboe #ifdef CONFIG_COMPAT 2380ed29b0b4SJens Axboe if (in_compat_syscall()) 2381ed29b0b4SJens Axboe ret = set_compat_user_sigmask((const compat_sigset_t __user *)sig, 2382ed29b0b4SJens Axboe sigsz); 2383ed29b0b4SJens Axboe else 2384ed29b0b4SJens Axboe #endif 2385ed29b0b4SJens Axboe ret = set_user_sigmask(sig, sigsz); 2386ed29b0b4SJens Axboe 2387ed29b0b4SJens Axboe if (ret) 2388ed29b0b4SJens Axboe return ret; 2389ed29b0b4SJens Axboe } 2390ed29b0b4SJens Axboe 2391ed29b0b4SJens Axboe if (uts) { 2392ed29b0b4SJens Axboe struct timespec64 ts; 2393ed29b0b4SJens Axboe 2394ed29b0b4SJens Axboe if (get_timespec64(&ts, uts)) 2395ed29b0b4SJens Axboe return -EFAULT; 2396ed29b0b4SJens Axboe timeout = ktime_add_ns(timespec64_to_ktime(ts), ktime_get_ns()); 2397ed29b0b4SJens Axboe } 2398ed29b0b4SJens Axboe 2399ed29b0b4SJens Axboe init_waitqueue_func_entry(&iowq.wq, io_wake_function); 2400ed29b0b4SJens Axboe iowq.wq.private = current; 2401ed29b0b4SJens Axboe INIT_LIST_HEAD(&iowq.wq.entry); 2402ed29b0b4SJens Axboe iowq.ctx = ctx; 2403ed29b0b4SJens Axboe iowq.nr_timeouts = atomic_read(&ctx->cq_timeouts); 2404ed29b0b4SJens Axboe iowq.cq_tail = READ_ONCE(ctx->rings->cq.head) + min_events; 2405ed29b0b4SJens Axboe 2406ed29b0b4SJens Axboe trace_io_uring_cqring_wait(ctx, min_events); 2407ed29b0b4SJens Axboe do { 2408ed29b0b4SJens Axboe /* if we can't even flush overflow, don't wait for more */ 2409ed29b0b4SJens Axboe if (!io_cqring_overflow_flush(ctx)) { 2410ed29b0b4SJens Axboe ret = -EBUSY; 2411ed29b0b4SJens Axboe break; 2412ed29b0b4SJens Axboe } 2413ed29b0b4SJens Axboe prepare_to_wait_exclusive(&ctx->cq_wait, &iowq.wq, 2414ed29b0b4SJens Axboe TASK_INTERRUPTIBLE); 2415ed29b0b4SJens Axboe ret = io_cqring_wait_schedule(ctx, &iowq, timeout); 2416ed29b0b4SJens Axboe cond_resched(); 2417ed29b0b4SJens Axboe } while (ret > 0); 2418ed29b0b4SJens Axboe 2419ed29b0b4SJens Axboe finish_wait(&ctx->cq_wait, &iowq.wq); 2420ed29b0b4SJens Axboe restore_saved_sigmask_unless(ret == -EINTR); 2421ed29b0b4SJens Axboe 2422ed29b0b4SJens Axboe return READ_ONCE(rings->cq.head) == READ_ONCE(rings->cq.tail) ? ret : 0; 2423ed29b0b4SJens Axboe } 2424ed29b0b4SJens Axboe 2425ed29b0b4SJens Axboe static void io_mem_free(void *ptr) 2426ed29b0b4SJens Axboe { 2427ed29b0b4SJens Axboe struct page *page; 2428ed29b0b4SJens Axboe 2429ed29b0b4SJens Axboe if (!ptr) 2430ed29b0b4SJens Axboe return; 2431ed29b0b4SJens Axboe 2432ed29b0b4SJens Axboe page = virt_to_head_page(ptr); 2433ed29b0b4SJens Axboe if (put_page_testzero(page)) 2434ed29b0b4SJens Axboe free_compound_page(page); 2435ed29b0b4SJens Axboe } 2436ed29b0b4SJens Axboe 2437ed29b0b4SJens Axboe static void *io_mem_alloc(size_t size) 2438ed29b0b4SJens Axboe { 2439ed29b0b4SJens Axboe gfp_t gfp = GFP_KERNEL_ACCOUNT | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP; 2440ed29b0b4SJens Axboe 2441ed29b0b4SJens Axboe return (void *) __get_free_pages(gfp, get_order(size)); 2442ed29b0b4SJens Axboe } 2443ed29b0b4SJens Axboe 2444ed29b0b4SJens Axboe static unsigned long rings_size(struct io_ring_ctx *ctx, unsigned int sq_entries, 2445ed29b0b4SJens Axboe unsigned int cq_entries, size_t *sq_offset) 2446ed29b0b4SJens Axboe { 2447ed29b0b4SJens Axboe struct io_rings *rings; 2448ed29b0b4SJens Axboe size_t off, sq_array_size; 2449ed29b0b4SJens Axboe 2450ed29b0b4SJens Axboe off = struct_size(rings, cqes, cq_entries); 2451ed29b0b4SJens Axboe if (off == SIZE_MAX) 2452ed29b0b4SJens Axboe return SIZE_MAX; 2453ed29b0b4SJens Axboe if (ctx->flags & IORING_SETUP_CQE32) { 2454ed29b0b4SJens Axboe if (check_shl_overflow(off, 1, &off)) 2455ed29b0b4SJens Axboe return SIZE_MAX; 2456ed29b0b4SJens Axboe } 2457ed29b0b4SJens Axboe 2458ed29b0b4SJens Axboe #ifdef CONFIG_SMP 2459ed29b0b4SJens Axboe off = ALIGN(off, SMP_CACHE_BYTES); 2460ed29b0b4SJens Axboe if (off == 0) 2461ed29b0b4SJens Axboe return SIZE_MAX; 2462ed29b0b4SJens Axboe #endif 2463ed29b0b4SJens Axboe 2464ed29b0b4SJens Axboe if (sq_offset) 2465ed29b0b4SJens Axboe *sq_offset = off; 2466ed29b0b4SJens Axboe 2467ed29b0b4SJens Axboe sq_array_size = array_size(sizeof(u32), sq_entries); 2468ed29b0b4SJens Axboe if (sq_array_size == SIZE_MAX) 2469ed29b0b4SJens Axboe return SIZE_MAX; 2470ed29b0b4SJens Axboe 2471ed29b0b4SJens Axboe if (check_add_overflow(off, sq_array_size, &off)) 2472ed29b0b4SJens Axboe return SIZE_MAX; 2473ed29b0b4SJens Axboe 2474ed29b0b4SJens Axboe return off; 2475ed29b0b4SJens Axboe } 2476ed29b0b4SJens Axboe 2477ed29b0b4SJens Axboe static int io_eventfd_register(struct io_ring_ctx *ctx, void __user *arg, 2478ed29b0b4SJens Axboe unsigned int eventfd_async) 2479ed29b0b4SJens Axboe { 2480ed29b0b4SJens Axboe struct io_ev_fd *ev_fd; 2481ed29b0b4SJens Axboe __s32 __user *fds = arg; 2482ed29b0b4SJens Axboe int fd; 2483ed29b0b4SJens Axboe 2484ed29b0b4SJens Axboe ev_fd = rcu_dereference_protected(ctx->io_ev_fd, 2485ed29b0b4SJens Axboe lockdep_is_held(&ctx->uring_lock)); 2486ed29b0b4SJens Axboe if (ev_fd) 2487ed29b0b4SJens Axboe return -EBUSY; 2488ed29b0b4SJens Axboe 2489ed29b0b4SJens Axboe if (copy_from_user(&fd, fds, sizeof(*fds))) 2490ed29b0b4SJens Axboe return -EFAULT; 2491ed29b0b4SJens Axboe 2492ed29b0b4SJens Axboe ev_fd = kmalloc(sizeof(*ev_fd), GFP_KERNEL); 2493ed29b0b4SJens Axboe if (!ev_fd) 2494ed29b0b4SJens Axboe return -ENOMEM; 2495ed29b0b4SJens Axboe 2496ed29b0b4SJens Axboe ev_fd->cq_ev_fd = eventfd_ctx_fdget(fd); 2497ed29b0b4SJens Axboe if (IS_ERR(ev_fd->cq_ev_fd)) { 2498ed29b0b4SJens Axboe int ret = PTR_ERR(ev_fd->cq_ev_fd); 2499ed29b0b4SJens Axboe kfree(ev_fd); 2500ed29b0b4SJens Axboe return ret; 2501ed29b0b4SJens Axboe } 2502305bef98SPavel Begunkov 2503305bef98SPavel Begunkov spin_lock(&ctx->completion_lock); 2504305bef98SPavel Begunkov ctx->evfd_last_cq_tail = ctx->cached_cq_tail; 2505305bef98SPavel Begunkov spin_unlock(&ctx->completion_lock); 2506305bef98SPavel Begunkov 2507ed29b0b4SJens Axboe ev_fd->eventfd_async = eventfd_async; 2508ed29b0b4SJens Axboe ctx->has_evfd = true; 2509ed29b0b4SJens Axboe rcu_assign_pointer(ctx->io_ev_fd, ev_fd); 251021a091b9SDylan Yudaken atomic_set(&ev_fd->refs, 1); 251121a091b9SDylan Yudaken atomic_set(&ev_fd->ops, 0); 2512ed29b0b4SJens Axboe return 0; 2513ed29b0b4SJens Axboe } 2514ed29b0b4SJens Axboe 2515ed29b0b4SJens Axboe static int io_eventfd_unregister(struct io_ring_ctx *ctx) 2516ed29b0b4SJens Axboe { 2517ed29b0b4SJens Axboe struct io_ev_fd *ev_fd; 2518ed29b0b4SJens Axboe 2519ed29b0b4SJens Axboe ev_fd = rcu_dereference_protected(ctx->io_ev_fd, 2520ed29b0b4SJens Axboe lockdep_is_held(&ctx->uring_lock)); 2521ed29b0b4SJens Axboe if (ev_fd) { 2522ed29b0b4SJens Axboe ctx->has_evfd = false; 2523ed29b0b4SJens Axboe rcu_assign_pointer(ctx->io_ev_fd, NULL); 252421a091b9SDylan Yudaken if (!atomic_fetch_or(BIT(IO_EVENTFD_OP_FREE_BIT), &ev_fd->ops)) 252521a091b9SDylan Yudaken call_rcu(&ev_fd->rcu, io_eventfd_ops); 2526ed29b0b4SJens Axboe return 0; 2527ed29b0b4SJens Axboe } 2528ed29b0b4SJens Axboe 2529ed29b0b4SJens Axboe return -ENXIO; 2530ed29b0b4SJens Axboe } 2531ed29b0b4SJens Axboe 2532ed29b0b4SJens Axboe static void io_req_caches_free(struct io_ring_ctx *ctx) 2533ed29b0b4SJens Axboe { 2534ed29b0b4SJens Axboe int nr = 0; 2535ed29b0b4SJens Axboe 2536ed29b0b4SJens Axboe mutex_lock(&ctx->uring_lock); 253734f0bc42SPavel Begunkov io_flush_cached_locked_reqs(ctx, &ctx->submit_state); 2538ed29b0b4SJens Axboe 2539ed29b0b4SJens Axboe while (!io_req_cache_empty(ctx)) { 254034f0bc42SPavel Begunkov struct io_kiocb *req = io_alloc_req(ctx); 2541ed29b0b4SJens Axboe 2542ed29b0b4SJens Axboe kmem_cache_free(req_cachep, req); 2543ed29b0b4SJens Axboe nr++; 2544ed29b0b4SJens Axboe } 2545ed29b0b4SJens Axboe if (nr) 2546ed29b0b4SJens Axboe percpu_ref_put_many(&ctx->refs, nr); 2547ed29b0b4SJens Axboe mutex_unlock(&ctx->uring_lock); 2548ed29b0b4SJens Axboe } 2549ed29b0b4SJens Axboe 2550ed29b0b4SJens Axboe static __cold void io_ring_ctx_free(struct io_ring_ctx *ctx) 2551ed29b0b4SJens Axboe { 2552ed29b0b4SJens Axboe io_sq_thread_finish(ctx); 2553ed29b0b4SJens Axboe io_rsrc_refs_drop(ctx); 2554ed29b0b4SJens Axboe /* __io_rsrc_put_work() may need uring_lock to progress, wait w/o it */ 2555ed29b0b4SJens Axboe io_wait_rsrc_data(ctx->buf_data); 2556ed29b0b4SJens Axboe io_wait_rsrc_data(ctx->file_data); 2557ed29b0b4SJens Axboe 2558ed29b0b4SJens Axboe mutex_lock(&ctx->uring_lock); 2559ed29b0b4SJens Axboe if (ctx->buf_data) 2560ed29b0b4SJens Axboe __io_sqe_buffers_unregister(ctx); 2561ed29b0b4SJens Axboe if (ctx->file_data) 2562ed29b0b4SJens Axboe __io_sqe_files_unregister(ctx); 2563ed29b0b4SJens Axboe if (ctx->rings) 2564ed29b0b4SJens Axboe __io_cqring_overflow_flush(ctx, true); 2565ed29b0b4SJens Axboe io_eventfd_unregister(ctx); 25669b797a37SJens Axboe io_alloc_cache_free(&ctx->apoll_cache, io_apoll_cache_free); 256743e0bbbdSJens Axboe io_alloc_cache_free(&ctx->netmsg_cache, io_netmsg_cache_free); 2568ed29b0b4SJens Axboe mutex_unlock(&ctx->uring_lock); 2569ed29b0b4SJens Axboe io_destroy_buffers(ctx); 2570ed29b0b4SJens Axboe if (ctx->sq_creds) 2571ed29b0b4SJens Axboe put_cred(ctx->sq_creds); 257297bbdc06SPavel Begunkov if (ctx->submitter_task) 257397bbdc06SPavel Begunkov put_task_struct(ctx->submitter_task); 2574ed29b0b4SJens Axboe 2575ed29b0b4SJens Axboe /* there are no registered resources left, nobody uses it */ 2576ed29b0b4SJens Axboe if (ctx->rsrc_node) 2577ed29b0b4SJens Axboe io_rsrc_node_destroy(ctx->rsrc_node); 2578ed29b0b4SJens Axboe if (ctx->rsrc_backup_node) 2579ed29b0b4SJens Axboe io_rsrc_node_destroy(ctx->rsrc_backup_node); 2580ed29b0b4SJens Axboe flush_delayed_work(&ctx->rsrc_put_work); 2581ed29b0b4SJens Axboe flush_delayed_work(&ctx->fallback_work); 2582ed29b0b4SJens Axboe 2583ed29b0b4SJens Axboe WARN_ON_ONCE(!list_empty(&ctx->rsrc_ref_list)); 2584ed29b0b4SJens Axboe WARN_ON_ONCE(!llist_empty(&ctx->rsrc_put_llist)); 2585ed29b0b4SJens Axboe 2586ed29b0b4SJens Axboe #if defined(CONFIG_UNIX) 2587ed29b0b4SJens Axboe if (ctx->ring_sock) { 2588ed29b0b4SJens Axboe ctx->ring_sock->file = NULL; /* so that iput() is called */ 2589ed29b0b4SJens Axboe sock_release(ctx->ring_sock); 2590ed29b0b4SJens Axboe } 2591ed29b0b4SJens Axboe #endif 2592ed29b0b4SJens Axboe WARN_ON_ONCE(!list_empty(&ctx->ltimeout_list)); 2593ed29b0b4SJens Axboe 259442b6419dSPavel Begunkov if (ctx->mm_account) { 259542b6419dSPavel Begunkov mmdrop(ctx->mm_account); 259642b6419dSPavel Begunkov ctx->mm_account = NULL; 259742b6419dSPavel Begunkov } 2598ed29b0b4SJens Axboe io_mem_free(ctx->rings); 2599ed29b0b4SJens Axboe io_mem_free(ctx->sq_sqes); 2600ed29b0b4SJens Axboe 2601ed29b0b4SJens Axboe percpu_ref_exit(&ctx->refs); 2602ed29b0b4SJens Axboe free_uid(ctx->user); 2603ed29b0b4SJens Axboe io_req_caches_free(ctx); 2604ed29b0b4SJens Axboe if (ctx->hash_map) 2605ed29b0b4SJens Axboe io_wq_put_hash(ctx->hash_map); 2606e6f89be6SPavel Begunkov kfree(ctx->cancel_table.hbs); 26079ca9fb24SPavel Begunkov kfree(ctx->cancel_table_locked.hbs); 2608ed29b0b4SJens Axboe kfree(ctx->dummy_ubuf); 2609ed29b0b4SJens Axboe kfree(ctx->io_bl); 2610ed29b0b4SJens Axboe xa_destroy(&ctx->io_bl_xa); 2611ed29b0b4SJens Axboe kfree(ctx); 2612ed29b0b4SJens Axboe } 2613ed29b0b4SJens Axboe 2614ed29b0b4SJens Axboe static __poll_t io_uring_poll(struct file *file, poll_table *wait) 2615ed29b0b4SJens Axboe { 2616ed29b0b4SJens Axboe struct io_ring_ctx *ctx = file->private_data; 2617ed29b0b4SJens Axboe __poll_t mask = 0; 2618ed29b0b4SJens Axboe 2619ed29b0b4SJens Axboe poll_wait(file, &ctx->cq_wait, wait); 2620ed29b0b4SJens Axboe /* 2621ed29b0b4SJens Axboe * synchronizes with barrier from wq_has_sleeper call in 2622ed29b0b4SJens Axboe * io_commit_cqring 2623ed29b0b4SJens Axboe */ 2624ed29b0b4SJens Axboe smp_rmb(); 2625ed29b0b4SJens Axboe if (!io_sqring_full(ctx)) 2626ed29b0b4SJens Axboe mask |= EPOLLOUT | EPOLLWRNORM; 2627ed29b0b4SJens Axboe 2628ed29b0b4SJens Axboe /* 2629ed29b0b4SJens Axboe * Don't flush cqring overflow list here, just do a simple check. 2630ed29b0b4SJens Axboe * Otherwise there could possible be ABBA deadlock: 2631ed29b0b4SJens Axboe * CPU0 CPU1 2632ed29b0b4SJens Axboe * ---- ---- 2633ed29b0b4SJens Axboe * lock(&ctx->uring_lock); 2634ed29b0b4SJens Axboe * lock(&ep->mtx); 2635ed29b0b4SJens Axboe * lock(&ctx->uring_lock); 2636ed29b0b4SJens Axboe * lock(&ep->mtx); 2637ed29b0b4SJens Axboe * 2638ed29b0b4SJens Axboe * Users may get EPOLLIN meanwhile seeing nothing in cqring, this 2639ed29b0b4SJens Axboe * pushs them to do the flush. 2640ed29b0b4SJens Axboe */ 2641b4c98d59SDylan Yudaken 2642b4c98d59SDylan Yudaken if (io_cqring_events(ctx) || io_has_work(ctx)) 2643ed29b0b4SJens Axboe mask |= EPOLLIN | EPOLLRDNORM; 2644ed29b0b4SJens Axboe 2645ed29b0b4SJens Axboe return mask; 2646ed29b0b4SJens Axboe } 2647ed29b0b4SJens Axboe 2648ed29b0b4SJens Axboe static int io_unregister_personality(struct io_ring_ctx *ctx, unsigned id) 2649ed29b0b4SJens Axboe { 2650ed29b0b4SJens Axboe const struct cred *creds; 2651ed29b0b4SJens Axboe 2652ed29b0b4SJens Axboe creds = xa_erase(&ctx->personalities, id); 2653ed29b0b4SJens Axboe if (creds) { 2654ed29b0b4SJens Axboe put_cred(creds); 2655ed29b0b4SJens Axboe return 0; 2656ed29b0b4SJens Axboe } 2657ed29b0b4SJens Axboe 2658ed29b0b4SJens Axboe return -EINVAL; 2659ed29b0b4SJens Axboe } 2660ed29b0b4SJens Axboe 2661ed29b0b4SJens Axboe struct io_tctx_exit { 2662ed29b0b4SJens Axboe struct callback_head task_work; 2663ed29b0b4SJens Axboe struct completion completion; 2664ed29b0b4SJens Axboe struct io_ring_ctx *ctx; 2665ed29b0b4SJens Axboe }; 2666ed29b0b4SJens Axboe 2667ed29b0b4SJens Axboe static __cold void io_tctx_exit_cb(struct callback_head *cb) 2668ed29b0b4SJens Axboe { 2669ed29b0b4SJens Axboe struct io_uring_task *tctx = current->io_uring; 2670ed29b0b4SJens Axboe struct io_tctx_exit *work; 2671ed29b0b4SJens Axboe 2672ed29b0b4SJens Axboe work = container_of(cb, struct io_tctx_exit, task_work); 2673ed29b0b4SJens Axboe /* 2674ed29b0b4SJens Axboe * When @in_idle, we're in cancellation and it's racy to remove the 2675ed29b0b4SJens Axboe * node. It'll be removed by the end of cancellation, just ignore it. 2676ed29b0b4SJens Axboe */ 2677ed29b0b4SJens Axboe if (!atomic_read(&tctx->in_idle)) 2678ed29b0b4SJens Axboe io_uring_del_tctx_node((unsigned long)work->ctx); 2679ed29b0b4SJens Axboe complete(&work->completion); 2680ed29b0b4SJens Axboe } 2681ed29b0b4SJens Axboe 2682ed29b0b4SJens Axboe static __cold bool io_cancel_ctx_cb(struct io_wq_work *work, void *data) 2683ed29b0b4SJens Axboe { 2684ed29b0b4SJens Axboe struct io_kiocb *req = container_of(work, struct io_kiocb, work); 2685ed29b0b4SJens Axboe 2686ed29b0b4SJens Axboe return req->ctx == data; 2687ed29b0b4SJens Axboe } 2688ed29b0b4SJens Axboe 2689ed29b0b4SJens Axboe static __cold void io_ring_exit_work(struct work_struct *work) 2690ed29b0b4SJens Axboe { 2691ed29b0b4SJens Axboe struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx, exit_work); 2692ed29b0b4SJens Axboe unsigned long timeout = jiffies + HZ * 60 * 5; 2693ed29b0b4SJens Axboe unsigned long interval = HZ / 20; 2694ed29b0b4SJens Axboe struct io_tctx_exit exit; 2695ed29b0b4SJens Axboe struct io_tctx_node *node; 2696ed29b0b4SJens Axboe int ret; 2697ed29b0b4SJens Axboe 2698ed29b0b4SJens Axboe /* 2699ed29b0b4SJens Axboe * If we're doing polled IO and end up having requests being 2700ed29b0b4SJens Axboe * submitted async (out-of-line), then completions can come in while 2701ed29b0b4SJens Axboe * we're waiting for refs to drop. We need to reap these manually, 2702ed29b0b4SJens Axboe * as nobody else will be looking for them. 2703ed29b0b4SJens Axboe */ 2704ed29b0b4SJens Axboe do { 2705c0e0d6baSDylan Yudaken if (ctx->flags & IORING_SETUP_DEFER_TASKRUN) 2706c0e0d6baSDylan Yudaken io_move_task_work_from_local(ctx); 2707c0e0d6baSDylan Yudaken 2708affa87dbSPavel Begunkov while (io_uring_try_cancel_requests(ctx, NULL, true)) 2709affa87dbSPavel Begunkov cond_resched(); 2710affa87dbSPavel Begunkov 2711ed29b0b4SJens Axboe if (ctx->sq_data) { 2712ed29b0b4SJens Axboe struct io_sq_data *sqd = ctx->sq_data; 2713ed29b0b4SJens Axboe struct task_struct *tsk; 2714ed29b0b4SJens Axboe 2715ed29b0b4SJens Axboe io_sq_thread_park(sqd); 2716ed29b0b4SJens Axboe tsk = sqd->thread; 2717ed29b0b4SJens Axboe if (tsk && tsk->io_uring && tsk->io_uring->io_wq) 2718ed29b0b4SJens Axboe io_wq_cancel_cb(tsk->io_uring->io_wq, 2719ed29b0b4SJens Axboe io_cancel_ctx_cb, ctx, true); 2720ed29b0b4SJens Axboe io_sq_thread_unpark(sqd); 2721ed29b0b4SJens Axboe } 2722ed29b0b4SJens Axboe 2723ed29b0b4SJens Axboe io_req_caches_free(ctx); 2724ed29b0b4SJens Axboe 2725ed29b0b4SJens Axboe if (WARN_ON_ONCE(time_after(jiffies, timeout))) { 2726ed29b0b4SJens Axboe /* there is little hope left, don't run it too often */ 2727ed29b0b4SJens Axboe interval = HZ * 60; 2728ed29b0b4SJens Axboe } 2729ed29b0b4SJens Axboe } while (!wait_for_completion_timeout(&ctx->ref_comp, interval)); 2730ed29b0b4SJens Axboe 2731ed29b0b4SJens Axboe init_completion(&exit.completion); 2732ed29b0b4SJens Axboe init_task_work(&exit.task_work, io_tctx_exit_cb); 2733ed29b0b4SJens Axboe exit.ctx = ctx; 2734ed29b0b4SJens Axboe /* 2735ed29b0b4SJens Axboe * Some may use context even when all refs and requests have been put, 2736ed29b0b4SJens Axboe * and they are free to do so while still holding uring_lock or 2737ed29b0b4SJens Axboe * completion_lock, see io_req_task_submit(). Apart from other work, 2738ed29b0b4SJens Axboe * this lock/unlock section also waits them to finish. 2739ed29b0b4SJens Axboe */ 2740ed29b0b4SJens Axboe mutex_lock(&ctx->uring_lock); 2741ed29b0b4SJens Axboe while (!list_empty(&ctx->tctx_list)) { 2742ed29b0b4SJens Axboe WARN_ON_ONCE(time_after(jiffies, timeout)); 2743ed29b0b4SJens Axboe 2744ed29b0b4SJens Axboe node = list_first_entry(&ctx->tctx_list, struct io_tctx_node, 2745ed29b0b4SJens Axboe ctx_node); 2746ed29b0b4SJens Axboe /* don't spin on a single task if cancellation failed */ 2747ed29b0b4SJens Axboe list_rotate_left(&ctx->tctx_list); 2748ed29b0b4SJens Axboe ret = task_work_add(node->task, &exit.task_work, TWA_SIGNAL); 2749ed29b0b4SJens Axboe if (WARN_ON_ONCE(ret)) 2750ed29b0b4SJens Axboe continue; 2751ed29b0b4SJens Axboe 2752ed29b0b4SJens Axboe mutex_unlock(&ctx->uring_lock); 2753ed29b0b4SJens Axboe wait_for_completion(&exit.completion); 2754ed29b0b4SJens Axboe mutex_lock(&ctx->uring_lock); 2755ed29b0b4SJens Axboe } 2756ed29b0b4SJens Axboe mutex_unlock(&ctx->uring_lock); 2757ed29b0b4SJens Axboe spin_lock(&ctx->completion_lock); 2758ed29b0b4SJens Axboe spin_unlock(&ctx->completion_lock); 2759ed29b0b4SJens Axboe 2760ed29b0b4SJens Axboe io_ring_ctx_free(ctx); 2761ed29b0b4SJens Axboe } 2762ed29b0b4SJens Axboe 2763ed29b0b4SJens Axboe static __cold void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx) 2764ed29b0b4SJens Axboe { 2765ed29b0b4SJens Axboe unsigned long index; 2766ed29b0b4SJens Axboe struct creds *creds; 2767ed29b0b4SJens Axboe 2768ed29b0b4SJens Axboe mutex_lock(&ctx->uring_lock); 2769ed29b0b4SJens Axboe percpu_ref_kill(&ctx->refs); 2770ed29b0b4SJens Axboe if (ctx->rings) 2771ed29b0b4SJens Axboe __io_cqring_overflow_flush(ctx, true); 2772ed29b0b4SJens Axboe xa_for_each(&ctx->personalities, index, creds) 2773ed29b0b4SJens Axboe io_unregister_personality(ctx, index); 27749ca9fb24SPavel Begunkov if (ctx->rings) 27759ca9fb24SPavel Begunkov io_poll_remove_all(ctx, NULL, true); 2776ed29b0b4SJens Axboe mutex_unlock(&ctx->uring_lock); 2777ed29b0b4SJens Axboe 277802bac94bSPavel Begunkov /* 277902bac94bSPavel Begunkov * If we failed setting up the ctx, we might not have any rings 278002bac94bSPavel Begunkov * and therefore did not submit any requests 278102bac94bSPavel Begunkov */ 278202bac94bSPavel Begunkov if (ctx->rings) 2783ed29b0b4SJens Axboe io_kill_timeouts(ctx, NULL, true); 2784ed29b0b4SJens Axboe 2785ed29b0b4SJens Axboe INIT_WORK(&ctx->exit_work, io_ring_exit_work); 2786ed29b0b4SJens Axboe /* 2787ed29b0b4SJens Axboe * Use system_unbound_wq to avoid spawning tons of event kworkers 2788ed29b0b4SJens Axboe * if we're exiting a ton of rings at the same time. It just adds 2789ed29b0b4SJens Axboe * noise and overhead, there's no discernable change in runtime 2790ed29b0b4SJens Axboe * over using system_wq. 2791ed29b0b4SJens Axboe */ 2792ed29b0b4SJens Axboe queue_work(system_unbound_wq, &ctx->exit_work); 2793ed29b0b4SJens Axboe } 2794ed29b0b4SJens Axboe 2795ed29b0b4SJens Axboe static int io_uring_release(struct inode *inode, struct file *file) 2796ed29b0b4SJens Axboe { 2797ed29b0b4SJens Axboe struct io_ring_ctx *ctx = file->private_data; 2798ed29b0b4SJens Axboe 2799ed29b0b4SJens Axboe file->private_data = NULL; 2800ed29b0b4SJens Axboe io_ring_ctx_wait_and_kill(ctx); 2801ed29b0b4SJens Axboe return 0; 2802ed29b0b4SJens Axboe } 2803ed29b0b4SJens Axboe 2804ed29b0b4SJens Axboe struct io_task_cancel { 2805ed29b0b4SJens Axboe struct task_struct *task; 2806ed29b0b4SJens Axboe bool all; 2807ed29b0b4SJens Axboe }; 2808ed29b0b4SJens Axboe 2809ed29b0b4SJens Axboe static bool io_cancel_task_cb(struct io_wq_work *work, void *data) 2810ed29b0b4SJens Axboe { 2811ed29b0b4SJens Axboe struct io_kiocb *req = container_of(work, struct io_kiocb, work); 2812ed29b0b4SJens Axboe struct io_task_cancel *cancel = data; 2813ed29b0b4SJens Axboe 2814ed29b0b4SJens Axboe return io_match_task_safe(req, cancel->task, cancel->all); 2815ed29b0b4SJens Axboe } 2816ed29b0b4SJens Axboe 2817ed29b0b4SJens Axboe static __cold bool io_cancel_defer_files(struct io_ring_ctx *ctx, 2818ed29b0b4SJens Axboe struct task_struct *task, 2819ed29b0b4SJens Axboe bool cancel_all) 2820ed29b0b4SJens Axboe { 2821ed29b0b4SJens Axboe struct io_defer_entry *de; 2822ed29b0b4SJens Axboe LIST_HEAD(list); 2823ed29b0b4SJens Axboe 2824ed29b0b4SJens Axboe spin_lock(&ctx->completion_lock); 2825ed29b0b4SJens Axboe list_for_each_entry_reverse(de, &ctx->defer_list, list) { 2826ed29b0b4SJens Axboe if (io_match_task_safe(de->req, task, cancel_all)) { 2827ed29b0b4SJens Axboe list_cut_position(&list, &ctx->defer_list, &de->list); 2828ed29b0b4SJens Axboe break; 2829ed29b0b4SJens Axboe } 2830ed29b0b4SJens Axboe } 2831ed29b0b4SJens Axboe spin_unlock(&ctx->completion_lock); 2832ed29b0b4SJens Axboe if (list_empty(&list)) 2833ed29b0b4SJens Axboe return false; 2834ed29b0b4SJens Axboe 2835ed29b0b4SJens Axboe while (!list_empty(&list)) { 2836ed29b0b4SJens Axboe de = list_first_entry(&list, struct io_defer_entry, list); 2837ed29b0b4SJens Axboe list_del_init(&de->list); 2838e276ae34SPavel Begunkov io_req_task_queue_fail(de->req, -ECANCELED); 2839ed29b0b4SJens Axboe kfree(de); 2840ed29b0b4SJens Axboe } 2841ed29b0b4SJens Axboe return true; 2842ed29b0b4SJens Axboe } 2843ed29b0b4SJens Axboe 2844ed29b0b4SJens Axboe static __cold bool io_uring_try_cancel_iowq(struct io_ring_ctx *ctx) 2845ed29b0b4SJens Axboe { 2846ed29b0b4SJens Axboe struct io_tctx_node *node; 2847ed29b0b4SJens Axboe enum io_wq_cancel cret; 2848ed29b0b4SJens Axboe bool ret = false; 2849ed29b0b4SJens Axboe 2850ed29b0b4SJens Axboe mutex_lock(&ctx->uring_lock); 2851ed29b0b4SJens Axboe list_for_each_entry(node, &ctx->tctx_list, ctx_node) { 2852ed29b0b4SJens Axboe struct io_uring_task *tctx = node->task->io_uring; 2853ed29b0b4SJens Axboe 2854ed29b0b4SJens Axboe /* 2855ed29b0b4SJens Axboe * io_wq will stay alive while we hold uring_lock, because it's 2856ed29b0b4SJens Axboe * killed after ctx nodes, which requires to take the lock. 2857ed29b0b4SJens Axboe */ 2858ed29b0b4SJens Axboe if (!tctx || !tctx->io_wq) 2859ed29b0b4SJens Axboe continue; 2860ed29b0b4SJens Axboe cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_ctx_cb, ctx, true); 2861ed29b0b4SJens Axboe ret |= (cret != IO_WQ_CANCEL_NOTFOUND); 2862ed29b0b4SJens Axboe } 2863ed29b0b4SJens Axboe mutex_unlock(&ctx->uring_lock); 2864ed29b0b4SJens Axboe 2865ed29b0b4SJens Axboe return ret; 2866ed29b0b4SJens Axboe } 2867ed29b0b4SJens Axboe 2868affa87dbSPavel Begunkov static __cold bool io_uring_try_cancel_requests(struct io_ring_ctx *ctx, 2869ed29b0b4SJens Axboe struct task_struct *task, 2870ed29b0b4SJens Axboe bool cancel_all) 2871ed29b0b4SJens Axboe { 2872ed29b0b4SJens Axboe struct io_task_cancel cancel = { .task = task, .all = cancel_all, }; 2873ed29b0b4SJens Axboe struct io_uring_task *tctx = task ? task->io_uring : NULL; 2874affa87dbSPavel Begunkov enum io_wq_cancel cret; 2875affa87dbSPavel Begunkov bool ret = false; 2876ed29b0b4SJens Axboe 2877ed29b0b4SJens Axboe /* failed during ring init, it couldn't have issued any requests */ 2878ed29b0b4SJens Axboe if (!ctx->rings) 2879affa87dbSPavel Begunkov return false; 2880ed29b0b4SJens Axboe 2881ed29b0b4SJens Axboe if (!task) { 2882ed29b0b4SJens Axboe ret |= io_uring_try_cancel_iowq(ctx); 2883ed29b0b4SJens Axboe } else if (tctx && tctx->io_wq) { 2884ed29b0b4SJens Axboe /* 2885ed29b0b4SJens Axboe * Cancels requests of all rings, not only @ctx, but 2886ed29b0b4SJens Axboe * it's fine as the task is in exit/exec. 2887ed29b0b4SJens Axboe */ 2888ed29b0b4SJens Axboe cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_task_cb, 2889ed29b0b4SJens Axboe &cancel, true); 2890ed29b0b4SJens Axboe ret |= (cret != IO_WQ_CANCEL_NOTFOUND); 2891ed29b0b4SJens Axboe } 2892ed29b0b4SJens Axboe 2893ed29b0b4SJens Axboe /* SQPOLL thread does its own polling */ 2894ed29b0b4SJens Axboe if ((!(ctx->flags & IORING_SETUP_SQPOLL) && cancel_all) || 2895ed29b0b4SJens Axboe (ctx->sq_data && ctx->sq_data->thread == current)) { 2896ed29b0b4SJens Axboe while (!wq_list_empty(&ctx->iopoll_list)) { 2897ed29b0b4SJens Axboe io_iopoll_try_reap_events(ctx); 2898ed29b0b4SJens Axboe ret = true; 2899ed29b0b4SJens Axboe } 2900ed29b0b4SJens Axboe } 2901ed29b0b4SJens Axboe 2902c0e0d6baSDylan Yudaken if (ctx->flags & IORING_SETUP_DEFER_TASKRUN) 2903c0e0d6baSDylan Yudaken ret |= io_run_local_work(ctx) > 0; 2904ed29b0b4SJens Axboe ret |= io_cancel_defer_files(ctx, task, cancel_all); 29059ca9fb24SPavel Begunkov mutex_lock(&ctx->uring_lock); 2906ed29b0b4SJens Axboe ret |= io_poll_remove_all(ctx, task, cancel_all); 29079ca9fb24SPavel Begunkov mutex_unlock(&ctx->uring_lock); 2908ed29b0b4SJens Axboe ret |= io_kill_timeouts(ctx, task, cancel_all); 2909ed29b0b4SJens Axboe if (task) 2910c0e0d6baSDylan Yudaken ret |= io_run_task_work() > 0; 2911affa87dbSPavel Begunkov return ret; 2912ed29b0b4SJens Axboe } 2913ed29b0b4SJens Axboe 2914ed29b0b4SJens Axboe static s64 tctx_inflight(struct io_uring_task *tctx, bool tracked) 2915ed29b0b4SJens Axboe { 2916ed29b0b4SJens Axboe if (tracked) 2917ed29b0b4SJens Axboe return atomic_read(&tctx->inflight_tracked); 2918ed29b0b4SJens Axboe return percpu_counter_sum(&tctx->inflight); 2919ed29b0b4SJens Axboe } 2920ed29b0b4SJens Axboe 2921ed29b0b4SJens Axboe /* 2922ed29b0b4SJens Axboe * Find any io_uring ctx that this task has registered or done IO on, and cancel 2923ed29b0b4SJens Axboe * requests. @sqd should be not-null IFF it's an SQPOLL thread cancellation. 2924ed29b0b4SJens Axboe */ 292517437f31SJens Axboe __cold void io_uring_cancel_generic(bool cancel_all, struct io_sq_data *sqd) 2926ed29b0b4SJens Axboe { 2927ed29b0b4SJens Axboe struct io_uring_task *tctx = current->io_uring; 2928ed29b0b4SJens Axboe struct io_ring_ctx *ctx; 2929ed29b0b4SJens Axboe s64 inflight; 2930ed29b0b4SJens Axboe DEFINE_WAIT(wait); 2931ed29b0b4SJens Axboe 2932ed29b0b4SJens Axboe WARN_ON_ONCE(sqd && sqd->thread != current); 2933ed29b0b4SJens Axboe 2934ed29b0b4SJens Axboe if (!current->io_uring) 2935ed29b0b4SJens Axboe return; 2936ed29b0b4SJens Axboe if (tctx->io_wq) 2937ed29b0b4SJens Axboe io_wq_exit_start(tctx->io_wq); 2938ed29b0b4SJens Axboe 2939ed29b0b4SJens Axboe atomic_inc(&tctx->in_idle); 2940ed29b0b4SJens Axboe do { 2941affa87dbSPavel Begunkov bool loop = false; 2942affa87dbSPavel Begunkov 2943ed29b0b4SJens Axboe io_uring_drop_tctx_refs(current); 2944ed29b0b4SJens Axboe /* read completions before cancelations */ 2945ed29b0b4SJens Axboe inflight = tctx_inflight(tctx, !cancel_all); 2946ed29b0b4SJens Axboe if (!inflight) 2947ed29b0b4SJens Axboe break; 2948ed29b0b4SJens Axboe 2949ed29b0b4SJens Axboe if (!sqd) { 2950ed29b0b4SJens Axboe struct io_tctx_node *node; 2951ed29b0b4SJens Axboe unsigned long index; 2952ed29b0b4SJens Axboe 2953ed29b0b4SJens Axboe xa_for_each(&tctx->xa, index, node) { 2954ed29b0b4SJens Axboe /* sqpoll task will cancel all its requests */ 2955ed29b0b4SJens Axboe if (node->ctx->sq_data) 2956ed29b0b4SJens Axboe continue; 2957affa87dbSPavel Begunkov loop |= io_uring_try_cancel_requests(node->ctx, 2958affa87dbSPavel Begunkov current, cancel_all); 2959ed29b0b4SJens Axboe } 2960ed29b0b4SJens Axboe } else { 2961ed29b0b4SJens Axboe list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) 2962affa87dbSPavel Begunkov loop |= io_uring_try_cancel_requests(ctx, 2963affa87dbSPavel Begunkov current, 2964ed29b0b4SJens Axboe cancel_all); 2965ed29b0b4SJens Axboe } 2966ed29b0b4SJens Axboe 2967affa87dbSPavel Begunkov if (loop) { 2968affa87dbSPavel Begunkov cond_resched(); 2969affa87dbSPavel Begunkov continue; 2970affa87dbSPavel Begunkov } 2971affa87dbSPavel Begunkov 2972ed29b0b4SJens Axboe prepare_to_wait(&tctx->wait, &wait, TASK_INTERRUPTIBLE); 2973ed29b0b4SJens Axboe io_run_task_work(); 2974ed29b0b4SJens Axboe io_uring_drop_tctx_refs(current); 2975ed29b0b4SJens Axboe 2976ed29b0b4SJens Axboe /* 2977ed29b0b4SJens Axboe * If we've seen completions, retry without waiting. This 2978ed29b0b4SJens Axboe * avoids a race where a completion comes in before we did 2979ed29b0b4SJens Axboe * prepare_to_wait(). 2980ed29b0b4SJens Axboe */ 2981ed29b0b4SJens Axboe if (inflight == tctx_inflight(tctx, !cancel_all)) 2982ed29b0b4SJens Axboe schedule(); 2983ed29b0b4SJens Axboe finish_wait(&tctx->wait, &wait); 2984ed29b0b4SJens Axboe } while (1); 2985ed29b0b4SJens Axboe 2986ed29b0b4SJens Axboe io_uring_clean_tctx(tctx); 2987ed29b0b4SJens Axboe if (cancel_all) { 2988ed29b0b4SJens Axboe /* 2989ed29b0b4SJens Axboe * We shouldn't run task_works after cancel, so just leave 2990ed29b0b4SJens Axboe * ->in_idle set for normal exit. 2991ed29b0b4SJens Axboe */ 2992ed29b0b4SJens Axboe atomic_dec(&tctx->in_idle); 2993ed29b0b4SJens Axboe /* for exec all current's requests should be gone, kill tctx */ 2994ed29b0b4SJens Axboe __io_uring_free(current); 2995ed29b0b4SJens Axboe } 2996ed29b0b4SJens Axboe } 2997ed29b0b4SJens Axboe 2998ed29b0b4SJens Axboe void __io_uring_cancel(bool cancel_all) 2999ed29b0b4SJens Axboe { 3000ed29b0b4SJens Axboe io_uring_cancel_generic(cancel_all, NULL); 3001ed29b0b4SJens Axboe } 3002ed29b0b4SJens Axboe 3003ed29b0b4SJens Axboe static void *io_uring_validate_mmap_request(struct file *file, 3004ed29b0b4SJens Axboe loff_t pgoff, size_t sz) 3005ed29b0b4SJens Axboe { 3006ed29b0b4SJens Axboe struct io_ring_ctx *ctx = file->private_data; 3007ed29b0b4SJens Axboe loff_t offset = pgoff << PAGE_SHIFT; 3008ed29b0b4SJens Axboe struct page *page; 3009ed29b0b4SJens Axboe void *ptr; 3010ed29b0b4SJens Axboe 3011ed29b0b4SJens Axboe switch (offset) { 3012ed29b0b4SJens Axboe case IORING_OFF_SQ_RING: 3013ed29b0b4SJens Axboe case IORING_OFF_CQ_RING: 3014ed29b0b4SJens Axboe ptr = ctx->rings; 3015ed29b0b4SJens Axboe break; 3016ed29b0b4SJens Axboe case IORING_OFF_SQES: 3017ed29b0b4SJens Axboe ptr = ctx->sq_sqes; 3018ed29b0b4SJens Axboe break; 3019ed29b0b4SJens Axboe default: 3020ed29b0b4SJens Axboe return ERR_PTR(-EINVAL); 3021ed29b0b4SJens Axboe } 3022ed29b0b4SJens Axboe 3023ed29b0b4SJens Axboe page = virt_to_head_page(ptr); 3024ed29b0b4SJens Axboe if (sz > page_size(page)) 3025ed29b0b4SJens Axboe return ERR_PTR(-EINVAL); 3026ed29b0b4SJens Axboe 3027ed29b0b4SJens Axboe return ptr; 3028ed29b0b4SJens Axboe } 3029ed29b0b4SJens Axboe 3030ed29b0b4SJens Axboe #ifdef CONFIG_MMU 3031ed29b0b4SJens Axboe 3032ed29b0b4SJens Axboe static __cold int io_uring_mmap(struct file *file, struct vm_area_struct *vma) 3033ed29b0b4SJens Axboe { 3034ed29b0b4SJens Axboe size_t sz = vma->vm_end - vma->vm_start; 3035ed29b0b4SJens Axboe unsigned long pfn; 3036ed29b0b4SJens Axboe void *ptr; 3037ed29b0b4SJens Axboe 3038ed29b0b4SJens Axboe ptr = io_uring_validate_mmap_request(file, vma->vm_pgoff, sz); 3039ed29b0b4SJens Axboe if (IS_ERR(ptr)) 3040ed29b0b4SJens Axboe return PTR_ERR(ptr); 3041ed29b0b4SJens Axboe 3042ed29b0b4SJens Axboe pfn = virt_to_phys(ptr) >> PAGE_SHIFT; 3043ed29b0b4SJens Axboe return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot); 3044ed29b0b4SJens Axboe } 3045ed29b0b4SJens Axboe 3046ed29b0b4SJens Axboe #else /* !CONFIG_MMU */ 3047ed29b0b4SJens Axboe 3048ed29b0b4SJens Axboe static int io_uring_mmap(struct file *file, struct vm_area_struct *vma) 3049ed29b0b4SJens Axboe { 3050ed29b0b4SJens Axboe return vma->vm_flags & (VM_SHARED | VM_MAYSHARE) ? 0 : -EINVAL; 3051ed29b0b4SJens Axboe } 3052ed29b0b4SJens Axboe 3053ed29b0b4SJens Axboe static unsigned int io_uring_nommu_mmap_capabilities(struct file *file) 3054ed29b0b4SJens Axboe { 3055ed29b0b4SJens Axboe return NOMMU_MAP_DIRECT | NOMMU_MAP_READ | NOMMU_MAP_WRITE; 3056ed29b0b4SJens Axboe } 3057ed29b0b4SJens Axboe 3058ed29b0b4SJens Axboe static unsigned long io_uring_nommu_get_unmapped_area(struct file *file, 3059ed29b0b4SJens Axboe unsigned long addr, unsigned long len, 3060ed29b0b4SJens Axboe unsigned long pgoff, unsigned long flags) 3061ed29b0b4SJens Axboe { 3062ed29b0b4SJens Axboe void *ptr; 3063ed29b0b4SJens Axboe 3064ed29b0b4SJens Axboe ptr = io_uring_validate_mmap_request(file, pgoff, len); 3065ed29b0b4SJens Axboe if (IS_ERR(ptr)) 3066ed29b0b4SJens Axboe return PTR_ERR(ptr); 3067ed29b0b4SJens Axboe 3068ed29b0b4SJens Axboe return (unsigned long) ptr; 3069ed29b0b4SJens Axboe } 3070ed29b0b4SJens Axboe 3071ed29b0b4SJens Axboe #endif /* !CONFIG_MMU */ 3072ed29b0b4SJens Axboe 3073ed29b0b4SJens Axboe static int io_validate_ext_arg(unsigned flags, const void __user *argp, size_t argsz) 3074ed29b0b4SJens Axboe { 3075ed29b0b4SJens Axboe if (flags & IORING_ENTER_EXT_ARG) { 3076ed29b0b4SJens Axboe struct io_uring_getevents_arg arg; 3077ed29b0b4SJens Axboe 3078ed29b0b4SJens Axboe if (argsz != sizeof(arg)) 3079ed29b0b4SJens Axboe return -EINVAL; 3080ed29b0b4SJens Axboe if (copy_from_user(&arg, argp, sizeof(arg))) 3081ed29b0b4SJens Axboe return -EFAULT; 3082ed29b0b4SJens Axboe } 3083ed29b0b4SJens Axboe return 0; 3084ed29b0b4SJens Axboe } 3085ed29b0b4SJens Axboe 3086ed29b0b4SJens Axboe static int io_get_ext_arg(unsigned flags, const void __user *argp, size_t *argsz, 3087ed29b0b4SJens Axboe struct __kernel_timespec __user **ts, 3088ed29b0b4SJens Axboe const sigset_t __user **sig) 3089ed29b0b4SJens Axboe { 3090ed29b0b4SJens Axboe struct io_uring_getevents_arg arg; 3091ed29b0b4SJens Axboe 3092ed29b0b4SJens Axboe /* 3093ed29b0b4SJens Axboe * If EXT_ARG isn't set, then we have no timespec and the argp pointer 3094ed29b0b4SJens Axboe * is just a pointer to the sigset_t. 3095ed29b0b4SJens Axboe */ 3096ed29b0b4SJens Axboe if (!(flags & IORING_ENTER_EXT_ARG)) { 3097ed29b0b4SJens Axboe *sig = (const sigset_t __user *) argp; 3098ed29b0b4SJens Axboe *ts = NULL; 3099ed29b0b4SJens Axboe return 0; 3100ed29b0b4SJens Axboe } 3101ed29b0b4SJens Axboe 3102ed29b0b4SJens Axboe /* 3103ed29b0b4SJens Axboe * EXT_ARG is set - ensure we agree on the size of it and copy in our 3104ed29b0b4SJens Axboe * timespec and sigset_t pointers if good. 3105ed29b0b4SJens Axboe */ 3106ed29b0b4SJens Axboe if (*argsz != sizeof(arg)) 3107ed29b0b4SJens Axboe return -EINVAL; 3108ed29b0b4SJens Axboe if (copy_from_user(&arg, argp, sizeof(arg))) 3109ed29b0b4SJens Axboe return -EFAULT; 3110ed29b0b4SJens Axboe if (arg.pad) 3111ed29b0b4SJens Axboe return -EINVAL; 3112ed29b0b4SJens Axboe *sig = u64_to_user_ptr(arg.sigmask); 3113ed29b0b4SJens Axboe *argsz = arg.sigmask_sz; 3114ed29b0b4SJens Axboe *ts = u64_to_user_ptr(arg.ts); 3115ed29b0b4SJens Axboe return 0; 3116ed29b0b4SJens Axboe } 3117ed29b0b4SJens Axboe 3118ed29b0b4SJens Axboe SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit, 3119ed29b0b4SJens Axboe u32, min_complete, u32, flags, const void __user *, argp, 3120ed29b0b4SJens Axboe size_t, argsz) 3121ed29b0b4SJens Axboe { 3122ed29b0b4SJens Axboe struct io_ring_ctx *ctx; 3123ed29b0b4SJens Axboe struct fd f; 3124ed29b0b4SJens Axboe long ret; 3125ed29b0b4SJens Axboe 3126ed29b0b4SJens Axboe if (unlikely(flags & ~(IORING_ENTER_GETEVENTS | IORING_ENTER_SQ_WAKEUP | 3127ed29b0b4SJens Axboe IORING_ENTER_SQ_WAIT | IORING_ENTER_EXT_ARG | 3128ed29b0b4SJens Axboe IORING_ENTER_REGISTERED_RING))) 3129ed29b0b4SJens Axboe return -EINVAL; 3130ed29b0b4SJens Axboe 3131ed29b0b4SJens Axboe /* 3132ed29b0b4SJens Axboe * Ring fd has been registered via IORING_REGISTER_RING_FDS, we 3133ed29b0b4SJens Axboe * need only dereference our task private array to find it. 3134ed29b0b4SJens Axboe */ 3135ed29b0b4SJens Axboe if (flags & IORING_ENTER_REGISTERED_RING) { 3136ed29b0b4SJens Axboe struct io_uring_task *tctx = current->io_uring; 3137ed29b0b4SJens Axboe 31383273c440SPavel Begunkov if (unlikely(!tctx || fd >= IO_RINGFD_REG_MAX)) 3139ed29b0b4SJens Axboe return -EINVAL; 3140ed29b0b4SJens Axboe fd = array_index_nospec(fd, IO_RINGFD_REG_MAX); 3141ed29b0b4SJens Axboe f.file = tctx->registered_rings[fd]; 3142ed29b0b4SJens Axboe f.flags = 0; 3143ed29b0b4SJens Axboe if (unlikely(!f.file)) 3144ed29b0b4SJens Axboe return -EBADF; 31453273c440SPavel Begunkov } else { 31463273c440SPavel Begunkov f = fdget(fd); 31473273c440SPavel Begunkov if (unlikely(!f.file)) 31483273c440SPavel Begunkov return -EBADF; 3149ed29b0b4SJens Axboe ret = -EOPNOTSUPP; 3150e5550a14SJens Axboe if (unlikely(!io_is_uring_fops(f.file))) 3151fbb8bb02SPavel Begunkov goto out; 31523273c440SPavel Begunkov } 3153ed29b0b4SJens Axboe 3154ed29b0b4SJens Axboe ctx = f.file->private_data; 3155ed29b0b4SJens Axboe ret = -EBADFD; 3156ed29b0b4SJens Axboe if (unlikely(ctx->flags & IORING_SETUP_R_DISABLED)) 3157ed29b0b4SJens Axboe goto out; 3158ed29b0b4SJens Axboe 3159ed29b0b4SJens Axboe /* 3160ed29b0b4SJens Axboe * For SQ polling, the thread will do all submissions and completions. 3161ed29b0b4SJens Axboe * Just return the requested submit count, and wake the thread if 3162ed29b0b4SJens Axboe * we were asked to. 3163ed29b0b4SJens Axboe */ 3164ed29b0b4SJens Axboe ret = 0; 3165ed29b0b4SJens Axboe if (ctx->flags & IORING_SETUP_SQPOLL) { 3166ed29b0b4SJens Axboe io_cqring_overflow_flush(ctx); 3167ed29b0b4SJens Axboe 3168ed29b0b4SJens Axboe if (unlikely(ctx->sq_data->thread == NULL)) { 3169ed29b0b4SJens Axboe ret = -EOWNERDEAD; 3170ed29b0b4SJens Axboe goto out; 3171ed29b0b4SJens Axboe } 3172ed29b0b4SJens Axboe if (flags & IORING_ENTER_SQ_WAKEUP) 3173ed29b0b4SJens Axboe wake_up(&ctx->sq_data->wait); 3174ed29b0b4SJens Axboe if (flags & IORING_ENTER_SQ_WAIT) { 3175ed29b0b4SJens Axboe ret = io_sqpoll_wait_sq(ctx); 3176ed29b0b4SJens Axboe if (ret) 3177ed29b0b4SJens Axboe goto out; 3178ed29b0b4SJens Axboe } 3179ed29b0b4SJens Axboe ret = to_submit; 3180ed29b0b4SJens Axboe } else if (to_submit) { 3181ed29b0b4SJens Axboe ret = io_uring_add_tctx_node(ctx); 3182ed29b0b4SJens Axboe if (unlikely(ret)) 3183ed29b0b4SJens Axboe goto out; 3184ed29b0b4SJens Axboe 3185ed29b0b4SJens Axboe mutex_lock(&ctx->uring_lock); 3186ed29b0b4SJens Axboe ret = io_submit_sqes(ctx, to_submit); 3187ed29b0b4SJens Axboe if (ret != to_submit) { 3188ed29b0b4SJens Axboe mutex_unlock(&ctx->uring_lock); 3189ed29b0b4SJens Axboe goto out; 3190ed29b0b4SJens Axboe } 319144f87745SPavel Begunkov if (flags & IORING_ENTER_GETEVENTS) { 319244f87745SPavel Begunkov if (ctx->syscall_iopoll) 3193ed29b0b4SJens Axboe goto iopoll_locked; 319444f87745SPavel Begunkov /* 319544f87745SPavel Begunkov * Ignore errors, we'll soon call io_cqring_wait() and 319644f87745SPavel Begunkov * it should handle ownership problems if any. 319744f87745SPavel Begunkov */ 319844f87745SPavel Begunkov if (ctx->flags & IORING_SETUP_DEFER_TASKRUN) 319944f87745SPavel Begunkov (void)io_run_local_work_locked(ctx); 320044f87745SPavel Begunkov } 3201ed29b0b4SJens Axboe mutex_unlock(&ctx->uring_lock); 3202ed29b0b4SJens Axboe } 3203c0e0d6baSDylan Yudaken 3204ed29b0b4SJens Axboe if (flags & IORING_ENTER_GETEVENTS) { 3205ed29b0b4SJens Axboe int ret2; 3206c0e0d6baSDylan Yudaken 3207ed29b0b4SJens Axboe if (ctx->syscall_iopoll) { 3208ed29b0b4SJens Axboe /* 3209ed29b0b4SJens Axboe * We disallow the app entering submit/complete with 3210ed29b0b4SJens Axboe * polling, but we still need to lock the ring to 3211ed29b0b4SJens Axboe * prevent racing with polled issue that got punted to 3212ed29b0b4SJens Axboe * a workqueue. 3213ed29b0b4SJens Axboe */ 3214ed29b0b4SJens Axboe mutex_lock(&ctx->uring_lock); 3215ed29b0b4SJens Axboe iopoll_locked: 3216ed29b0b4SJens Axboe ret2 = io_validate_ext_arg(flags, argp, argsz); 3217ed29b0b4SJens Axboe if (likely(!ret2)) { 3218ed29b0b4SJens Axboe min_complete = min(min_complete, 3219ed29b0b4SJens Axboe ctx->cq_entries); 3220ed29b0b4SJens Axboe ret2 = io_iopoll_check(ctx, min_complete); 3221ed29b0b4SJens Axboe } 3222ed29b0b4SJens Axboe mutex_unlock(&ctx->uring_lock); 3223ed29b0b4SJens Axboe } else { 3224ed29b0b4SJens Axboe const sigset_t __user *sig; 3225ed29b0b4SJens Axboe struct __kernel_timespec __user *ts; 3226ed29b0b4SJens Axboe 3227ed29b0b4SJens Axboe ret2 = io_get_ext_arg(flags, argp, &argsz, &ts, &sig); 3228ed29b0b4SJens Axboe if (likely(!ret2)) { 3229ed29b0b4SJens Axboe min_complete = min(min_complete, 3230ed29b0b4SJens Axboe ctx->cq_entries); 3231ed29b0b4SJens Axboe ret2 = io_cqring_wait(ctx, min_complete, sig, 3232ed29b0b4SJens Axboe argsz, ts); 3233ed29b0b4SJens Axboe } 3234ed29b0b4SJens Axboe } 3235ed29b0b4SJens Axboe 3236ed29b0b4SJens Axboe if (!ret) { 3237ed29b0b4SJens Axboe ret = ret2; 3238ed29b0b4SJens Axboe 3239ed29b0b4SJens Axboe /* 3240ed29b0b4SJens Axboe * EBADR indicates that one or more CQE were dropped. 3241ed29b0b4SJens Axboe * Once the user has been informed we can clear the bit 3242ed29b0b4SJens Axboe * as they are obviously ok with those drops. 3243ed29b0b4SJens Axboe */ 3244ed29b0b4SJens Axboe if (unlikely(ret2 == -EBADR)) 3245ed29b0b4SJens Axboe clear_bit(IO_CHECK_CQ_DROPPED_BIT, 3246ed29b0b4SJens Axboe &ctx->check_cq); 3247ed29b0b4SJens Axboe } 3248ed29b0b4SJens Axboe } 3249ed29b0b4SJens Axboe out: 3250ed29b0b4SJens Axboe fdput(f); 3251ed29b0b4SJens Axboe return ret; 3252ed29b0b4SJens Axboe } 3253ed29b0b4SJens Axboe 3254ed29b0b4SJens Axboe static const struct file_operations io_uring_fops = { 3255ed29b0b4SJens Axboe .release = io_uring_release, 3256ed29b0b4SJens Axboe .mmap = io_uring_mmap, 3257ed29b0b4SJens Axboe #ifndef CONFIG_MMU 3258ed29b0b4SJens Axboe .get_unmapped_area = io_uring_nommu_get_unmapped_area, 3259ed29b0b4SJens Axboe .mmap_capabilities = io_uring_nommu_mmap_capabilities, 3260ed29b0b4SJens Axboe #endif 3261ed29b0b4SJens Axboe .poll = io_uring_poll, 3262ed29b0b4SJens Axboe #ifdef CONFIG_PROC_FS 3263ed29b0b4SJens Axboe .show_fdinfo = io_uring_show_fdinfo, 3264ed29b0b4SJens Axboe #endif 3265ed29b0b4SJens Axboe }; 3266ed29b0b4SJens Axboe 326792ac8beaSJens Axboe bool io_is_uring_fops(struct file *file) 326892ac8beaSJens Axboe { 326992ac8beaSJens Axboe return file->f_op == &io_uring_fops; 327092ac8beaSJens Axboe } 327192ac8beaSJens Axboe 3272ed29b0b4SJens Axboe static __cold int io_allocate_scq_urings(struct io_ring_ctx *ctx, 3273ed29b0b4SJens Axboe struct io_uring_params *p) 3274ed29b0b4SJens Axboe { 3275ed29b0b4SJens Axboe struct io_rings *rings; 3276ed29b0b4SJens Axboe size_t size, sq_array_offset; 3277ed29b0b4SJens Axboe 3278ed29b0b4SJens Axboe /* make sure these are sane, as we already accounted them */ 3279ed29b0b4SJens Axboe ctx->sq_entries = p->sq_entries; 3280ed29b0b4SJens Axboe ctx->cq_entries = p->cq_entries; 3281ed29b0b4SJens Axboe 3282ed29b0b4SJens Axboe size = rings_size(ctx, p->sq_entries, p->cq_entries, &sq_array_offset); 3283ed29b0b4SJens Axboe if (size == SIZE_MAX) 3284ed29b0b4SJens Axboe return -EOVERFLOW; 3285ed29b0b4SJens Axboe 3286ed29b0b4SJens Axboe rings = io_mem_alloc(size); 3287ed29b0b4SJens Axboe if (!rings) 3288ed29b0b4SJens Axboe return -ENOMEM; 3289ed29b0b4SJens Axboe 3290ed29b0b4SJens Axboe ctx->rings = rings; 3291ed29b0b4SJens Axboe ctx->sq_array = (u32 *)((char *)rings + sq_array_offset); 3292ed29b0b4SJens Axboe rings->sq_ring_mask = p->sq_entries - 1; 3293ed29b0b4SJens Axboe rings->cq_ring_mask = p->cq_entries - 1; 3294ed29b0b4SJens Axboe rings->sq_ring_entries = p->sq_entries; 3295ed29b0b4SJens Axboe rings->cq_ring_entries = p->cq_entries; 3296ed29b0b4SJens Axboe 3297ed29b0b4SJens Axboe if (p->flags & IORING_SETUP_SQE128) 3298ed29b0b4SJens Axboe size = array_size(2 * sizeof(struct io_uring_sqe), p->sq_entries); 3299ed29b0b4SJens Axboe else 3300ed29b0b4SJens Axboe size = array_size(sizeof(struct io_uring_sqe), p->sq_entries); 3301ed29b0b4SJens Axboe if (size == SIZE_MAX) { 3302ed29b0b4SJens Axboe io_mem_free(ctx->rings); 3303ed29b0b4SJens Axboe ctx->rings = NULL; 3304ed29b0b4SJens Axboe return -EOVERFLOW; 3305ed29b0b4SJens Axboe } 3306ed29b0b4SJens Axboe 3307ed29b0b4SJens Axboe ctx->sq_sqes = io_mem_alloc(size); 3308ed29b0b4SJens Axboe if (!ctx->sq_sqes) { 3309ed29b0b4SJens Axboe io_mem_free(ctx->rings); 3310ed29b0b4SJens Axboe ctx->rings = NULL; 3311ed29b0b4SJens Axboe return -ENOMEM; 3312ed29b0b4SJens Axboe } 3313ed29b0b4SJens Axboe 3314ed29b0b4SJens Axboe return 0; 3315ed29b0b4SJens Axboe } 3316ed29b0b4SJens Axboe 3317ed29b0b4SJens Axboe static int io_uring_install_fd(struct io_ring_ctx *ctx, struct file *file) 3318ed29b0b4SJens Axboe { 3319ed29b0b4SJens Axboe int ret, fd; 3320ed29b0b4SJens Axboe 3321ed29b0b4SJens Axboe fd = get_unused_fd_flags(O_RDWR | O_CLOEXEC); 3322ed29b0b4SJens Axboe if (fd < 0) 3323ed29b0b4SJens Axboe return fd; 3324ed29b0b4SJens Axboe 332597c96e9fSDylan Yudaken ret = __io_uring_add_tctx_node(ctx); 3326ed29b0b4SJens Axboe if (ret) { 3327ed29b0b4SJens Axboe put_unused_fd(fd); 3328ed29b0b4SJens Axboe return ret; 3329ed29b0b4SJens Axboe } 3330ed29b0b4SJens Axboe fd_install(fd, file); 3331ed29b0b4SJens Axboe return fd; 3332ed29b0b4SJens Axboe } 3333ed29b0b4SJens Axboe 3334ed29b0b4SJens Axboe /* 3335ed29b0b4SJens Axboe * Allocate an anonymous fd, this is what constitutes the application 3336ed29b0b4SJens Axboe * visible backing of an io_uring instance. The application mmaps this 3337ed29b0b4SJens Axboe * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled, 3338ed29b0b4SJens Axboe * we have to tie this fd to a socket for file garbage collection purposes. 3339ed29b0b4SJens Axboe */ 3340ed29b0b4SJens Axboe static struct file *io_uring_get_file(struct io_ring_ctx *ctx) 3341ed29b0b4SJens Axboe { 3342ed29b0b4SJens Axboe struct file *file; 3343ed29b0b4SJens Axboe #if defined(CONFIG_UNIX) 3344ed29b0b4SJens Axboe int ret; 3345ed29b0b4SJens Axboe 3346ed29b0b4SJens Axboe ret = sock_create_kern(&init_net, PF_UNIX, SOCK_RAW, IPPROTO_IP, 3347ed29b0b4SJens Axboe &ctx->ring_sock); 3348ed29b0b4SJens Axboe if (ret) 3349ed29b0b4SJens Axboe return ERR_PTR(ret); 3350ed29b0b4SJens Axboe #endif 3351ed29b0b4SJens Axboe 3352ed29b0b4SJens Axboe file = anon_inode_getfile_secure("[io_uring]", &io_uring_fops, ctx, 3353ed29b0b4SJens Axboe O_RDWR | O_CLOEXEC, NULL); 3354ed29b0b4SJens Axboe #if defined(CONFIG_UNIX) 3355ed29b0b4SJens Axboe if (IS_ERR(file)) { 3356ed29b0b4SJens Axboe sock_release(ctx->ring_sock); 3357ed29b0b4SJens Axboe ctx->ring_sock = NULL; 3358ed29b0b4SJens Axboe } else { 3359ed29b0b4SJens Axboe ctx->ring_sock->file = file; 3360ed29b0b4SJens Axboe } 3361ed29b0b4SJens Axboe #endif 3362ed29b0b4SJens Axboe return file; 3363ed29b0b4SJens Axboe } 3364ed29b0b4SJens Axboe 3365ed29b0b4SJens Axboe static __cold int io_uring_create(unsigned entries, struct io_uring_params *p, 3366ed29b0b4SJens Axboe struct io_uring_params __user *params) 3367ed29b0b4SJens Axboe { 3368ed29b0b4SJens Axboe struct io_ring_ctx *ctx; 3369ed29b0b4SJens Axboe struct file *file; 3370ed29b0b4SJens Axboe int ret; 3371ed29b0b4SJens Axboe 3372ed29b0b4SJens Axboe if (!entries) 3373ed29b0b4SJens Axboe return -EINVAL; 3374ed29b0b4SJens Axboe if (entries > IORING_MAX_ENTRIES) { 3375ed29b0b4SJens Axboe if (!(p->flags & IORING_SETUP_CLAMP)) 3376ed29b0b4SJens Axboe return -EINVAL; 3377ed29b0b4SJens Axboe entries = IORING_MAX_ENTRIES; 3378ed29b0b4SJens Axboe } 3379ed29b0b4SJens Axboe 3380ed29b0b4SJens Axboe /* 3381ed29b0b4SJens Axboe * Use twice as many entries for the CQ ring. It's possible for the 3382ed29b0b4SJens Axboe * application to drive a higher depth than the size of the SQ ring, 3383ed29b0b4SJens Axboe * since the sqes are only used at submission time. This allows for 3384ed29b0b4SJens Axboe * some flexibility in overcommitting a bit. If the application has 3385ed29b0b4SJens Axboe * set IORING_SETUP_CQSIZE, it will have passed in the desired number 3386ed29b0b4SJens Axboe * of CQ ring entries manually. 3387ed29b0b4SJens Axboe */ 3388ed29b0b4SJens Axboe p->sq_entries = roundup_pow_of_two(entries); 3389ed29b0b4SJens Axboe if (p->flags & IORING_SETUP_CQSIZE) { 3390ed29b0b4SJens Axboe /* 3391ed29b0b4SJens Axboe * If IORING_SETUP_CQSIZE is set, we do the same roundup 3392ed29b0b4SJens Axboe * to a power-of-two, if it isn't already. We do NOT impose 3393ed29b0b4SJens Axboe * any cq vs sq ring sizing. 3394ed29b0b4SJens Axboe */ 3395ed29b0b4SJens Axboe if (!p->cq_entries) 3396ed29b0b4SJens Axboe return -EINVAL; 3397ed29b0b4SJens Axboe if (p->cq_entries > IORING_MAX_CQ_ENTRIES) { 3398ed29b0b4SJens Axboe if (!(p->flags & IORING_SETUP_CLAMP)) 3399ed29b0b4SJens Axboe return -EINVAL; 3400ed29b0b4SJens Axboe p->cq_entries = IORING_MAX_CQ_ENTRIES; 3401ed29b0b4SJens Axboe } 3402ed29b0b4SJens Axboe p->cq_entries = roundup_pow_of_two(p->cq_entries); 3403ed29b0b4SJens Axboe if (p->cq_entries < p->sq_entries) 3404ed29b0b4SJens Axboe return -EINVAL; 3405ed29b0b4SJens Axboe } else { 3406ed29b0b4SJens Axboe p->cq_entries = 2 * p->sq_entries; 3407ed29b0b4SJens Axboe } 3408ed29b0b4SJens Axboe 3409ed29b0b4SJens Axboe ctx = io_ring_ctx_alloc(p); 3410ed29b0b4SJens Axboe if (!ctx) 3411ed29b0b4SJens Axboe return -ENOMEM; 3412ed29b0b4SJens Axboe 3413ed29b0b4SJens Axboe /* 3414ed29b0b4SJens Axboe * When SETUP_IOPOLL and SETUP_SQPOLL are both enabled, user 3415ed29b0b4SJens Axboe * space applications don't need to do io completion events 3416ed29b0b4SJens Axboe * polling again, they can rely on io_sq_thread to do polling 3417ed29b0b4SJens Axboe * work, which can reduce cpu usage and uring_lock contention. 3418ed29b0b4SJens Axboe */ 3419ed29b0b4SJens Axboe if (ctx->flags & IORING_SETUP_IOPOLL && 3420ed29b0b4SJens Axboe !(ctx->flags & IORING_SETUP_SQPOLL)) 3421ed29b0b4SJens Axboe ctx->syscall_iopoll = 1; 3422ed29b0b4SJens Axboe 3423ed29b0b4SJens Axboe ctx->compat = in_compat_syscall(); 3424ed29b0b4SJens Axboe if (!capable(CAP_IPC_LOCK)) 3425ed29b0b4SJens Axboe ctx->user = get_uid(current_user()); 3426ed29b0b4SJens Axboe 3427ed29b0b4SJens Axboe /* 3428ed29b0b4SJens Axboe * For SQPOLL, we just need a wakeup, always. For !SQPOLL, if 3429ed29b0b4SJens Axboe * COOP_TASKRUN is set, then IPIs are never needed by the app. 3430ed29b0b4SJens Axboe */ 3431ed29b0b4SJens Axboe ret = -EINVAL; 3432ed29b0b4SJens Axboe if (ctx->flags & IORING_SETUP_SQPOLL) { 3433ed29b0b4SJens Axboe /* IPI related flags don't make sense with SQPOLL */ 3434ed29b0b4SJens Axboe if (ctx->flags & (IORING_SETUP_COOP_TASKRUN | 3435c0e0d6baSDylan Yudaken IORING_SETUP_TASKRUN_FLAG | 3436c0e0d6baSDylan Yudaken IORING_SETUP_DEFER_TASKRUN)) 3437ed29b0b4SJens Axboe goto err; 3438ed29b0b4SJens Axboe ctx->notify_method = TWA_SIGNAL_NO_IPI; 3439ed29b0b4SJens Axboe } else if (ctx->flags & IORING_SETUP_COOP_TASKRUN) { 3440ed29b0b4SJens Axboe ctx->notify_method = TWA_SIGNAL_NO_IPI; 3441ed29b0b4SJens Axboe } else { 3442c0e0d6baSDylan Yudaken if (ctx->flags & IORING_SETUP_TASKRUN_FLAG && 3443c0e0d6baSDylan Yudaken !(ctx->flags & IORING_SETUP_DEFER_TASKRUN)) 3444ed29b0b4SJens Axboe goto err; 3445ed29b0b4SJens Axboe ctx->notify_method = TWA_SIGNAL; 3446ed29b0b4SJens Axboe } 3447ed29b0b4SJens Axboe 3448ed29b0b4SJens Axboe /* 3449c0e0d6baSDylan Yudaken * For DEFER_TASKRUN we require the completion task to be the same as the 3450c0e0d6baSDylan Yudaken * submission task. This implies that there is only one submitter, so enforce 3451c0e0d6baSDylan Yudaken * that. 3452c0e0d6baSDylan Yudaken */ 3453c0e0d6baSDylan Yudaken if (ctx->flags & IORING_SETUP_DEFER_TASKRUN && 3454c0e0d6baSDylan Yudaken !(ctx->flags & IORING_SETUP_SINGLE_ISSUER)) { 3455c0e0d6baSDylan Yudaken goto err; 3456c0e0d6baSDylan Yudaken } 3457c0e0d6baSDylan Yudaken 3458c0e0d6baSDylan Yudaken /* 3459ed29b0b4SJens Axboe * This is just grabbed for accounting purposes. When a process exits, 3460ed29b0b4SJens Axboe * the mm is exited and dropped before the files, hence we need to hang 3461ed29b0b4SJens Axboe * on to this mm purely for the purposes of being able to unaccount 3462ed29b0b4SJens Axboe * memory (locked/pinned vm). It's not used for anything else. 3463ed29b0b4SJens Axboe */ 3464ed29b0b4SJens Axboe mmgrab(current->mm); 3465ed29b0b4SJens Axboe ctx->mm_account = current->mm; 3466ed29b0b4SJens Axboe 3467ed29b0b4SJens Axboe ret = io_allocate_scq_urings(ctx, p); 3468ed29b0b4SJens Axboe if (ret) 3469ed29b0b4SJens Axboe goto err; 3470ed29b0b4SJens Axboe 3471ed29b0b4SJens Axboe ret = io_sq_offload_create(ctx, p); 3472ed29b0b4SJens Axboe if (ret) 3473ed29b0b4SJens Axboe goto err; 3474ed29b0b4SJens Axboe /* always set a rsrc node */ 3475ed29b0b4SJens Axboe ret = io_rsrc_node_switch_start(ctx); 3476ed29b0b4SJens Axboe if (ret) 3477ed29b0b4SJens Axboe goto err; 3478ed29b0b4SJens Axboe io_rsrc_node_switch(ctx, NULL); 3479ed29b0b4SJens Axboe 3480ed29b0b4SJens Axboe memset(&p->sq_off, 0, sizeof(p->sq_off)); 3481ed29b0b4SJens Axboe p->sq_off.head = offsetof(struct io_rings, sq.head); 3482ed29b0b4SJens Axboe p->sq_off.tail = offsetof(struct io_rings, sq.tail); 3483ed29b0b4SJens Axboe p->sq_off.ring_mask = offsetof(struct io_rings, sq_ring_mask); 3484ed29b0b4SJens Axboe p->sq_off.ring_entries = offsetof(struct io_rings, sq_ring_entries); 3485ed29b0b4SJens Axboe p->sq_off.flags = offsetof(struct io_rings, sq_flags); 3486ed29b0b4SJens Axboe p->sq_off.dropped = offsetof(struct io_rings, sq_dropped); 3487ed29b0b4SJens Axboe p->sq_off.array = (char *)ctx->sq_array - (char *)ctx->rings; 3488ed29b0b4SJens Axboe 3489ed29b0b4SJens Axboe memset(&p->cq_off, 0, sizeof(p->cq_off)); 3490ed29b0b4SJens Axboe p->cq_off.head = offsetof(struct io_rings, cq.head); 3491ed29b0b4SJens Axboe p->cq_off.tail = offsetof(struct io_rings, cq.tail); 3492ed29b0b4SJens Axboe p->cq_off.ring_mask = offsetof(struct io_rings, cq_ring_mask); 3493ed29b0b4SJens Axboe p->cq_off.ring_entries = offsetof(struct io_rings, cq_ring_entries); 3494ed29b0b4SJens Axboe p->cq_off.overflow = offsetof(struct io_rings, cq_overflow); 3495ed29b0b4SJens Axboe p->cq_off.cqes = offsetof(struct io_rings, cqes); 3496ed29b0b4SJens Axboe p->cq_off.flags = offsetof(struct io_rings, cq_flags); 3497ed29b0b4SJens Axboe 3498ed29b0b4SJens Axboe p->features = IORING_FEAT_SINGLE_MMAP | IORING_FEAT_NODROP | 3499ed29b0b4SJens Axboe IORING_FEAT_SUBMIT_STABLE | IORING_FEAT_RW_CUR_POS | 3500ed29b0b4SJens Axboe IORING_FEAT_CUR_PERSONALITY | IORING_FEAT_FAST_POLL | 3501ed29b0b4SJens Axboe IORING_FEAT_POLL_32BITS | IORING_FEAT_SQPOLL_NONFIXED | 3502ed29b0b4SJens Axboe IORING_FEAT_EXT_ARG | IORING_FEAT_NATIVE_WORKERS | 3503ed29b0b4SJens Axboe IORING_FEAT_RSRC_TAGS | IORING_FEAT_CQE_SKIP | 3504ed29b0b4SJens Axboe IORING_FEAT_LINKED_FILE; 3505ed29b0b4SJens Axboe 3506ed29b0b4SJens Axboe if (copy_to_user(params, p, sizeof(*p))) { 3507ed29b0b4SJens Axboe ret = -EFAULT; 3508ed29b0b4SJens Axboe goto err; 3509ed29b0b4SJens Axboe } 3510ed29b0b4SJens Axboe 35117cae596bSDylan Yudaken if (ctx->flags & IORING_SETUP_SINGLE_ISSUER 35127cae596bSDylan Yudaken && !(ctx->flags & IORING_SETUP_R_DISABLED)) 35137cae596bSDylan Yudaken ctx->submitter_task = get_task_struct(current); 35147cae596bSDylan Yudaken 3515ed29b0b4SJens Axboe file = io_uring_get_file(ctx); 3516ed29b0b4SJens Axboe if (IS_ERR(file)) { 3517ed29b0b4SJens Axboe ret = PTR_ERR(file); 3518ed29b0b4SJens Axboe goto err; 3519ed29b0b4SJens Axboe } 3520ed29b0b4SJens Axboe 3521ed29b0b4SJens Axboe /* 3522ed29b0b4SJens Axboe * Install ring fd as the very last thing, so we don't risk someone 3523ed29b0b4SJens Axboe * having closed it before we finish setup 3524ed29b0b4SJens Axboe */ 3525ed29b0b4SJens Axboe ret = io_uring_install_fd(ctx, file); 3526ed29b0b4SJens Axboe if (ret < 0) { 3527ed29b0b4SJens Axboe /* fput will clean it up */ 3528ed29b0b4SJens Axboe fput(file); 3529ed29b0b4SJens Axboe return ret; 3530ed29b0b4SJens Axboe } 3531ed29b0b4SJens Axboe 3532ed29b0b4SJens Axboe trace_io_uring_create(ret, ctx, p->sq_entries, p->cq_entries, p->flags); 3533ed29b0b4SJens Axboe return ret; 3534ed29b0b4SJens Axboe err: 3535ed29b0b4SJens Axboe io_ring_ctx_wait_and_kill(ctx); 3536ed29b0b4SJens Axboe return ret; 3537ed29b0b4SJens Axboe } 3538ed29b0b4SJens Axboe 3539ed29b0b4SJens Axboe /* 3540ed29b0b4SJens Axboe * Sets up an aio uring context, and returns the fd. Applications asks for a 3541ed29b0b4SJens Axboe * ring size, we return the actual sq/cq ring sizes (among other things) in the 3542ed29b0b4SJens Axboe * params structure passed in. 3543ed29b0b4SJens Axboe */ 3544ed29b0b4SJens Axboe static long io_uring_setup(u32 entries, struct io_uring_params __user *params) 3545ed29b0b4SJens Axboe { 3546ed29b0b4SJens Axboe struct io_uring_params p; 3547ed29b0b4SJens Axboe int i; 3548ed29b0b4SJens Axboe 3549ed29b0b4SJens Axboe if (copy_from_user(&p, params, sizeof(p))) 3550ed29b0b4SJens Axboe return -EFAULT; 3551ed29b0b4SJens Axboe for (i = 0; i < ARRAY_SIZE(p.resv); i++) { 3552ed29b0b4SJens Axboe if (p.resv[i]) 3553ed29b0b4SJens Axboe return -EINVAL; 3554ed29b0b4SJens Axboe } 3555ed29b0b4SJens Axboe 3556ed29b0b4SJens Axboe if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL | 3557ed29b0b4SJens Axboe IORING_SETUP_SQ_AFF | IORING_SETUP_CQSIZE | 3558ed29b0b4SJens Axboe IORING_SETUP_CLAMP | IORING_SETUP_ATTACH_WQ | 3559ed29b0b4SJens Axboe IORING_SETUP_R_DISABLED | IORING_SETUP_SUBMIT_ALL | 3560ed29b0b4SJens Axboe IORING_SETUP_COOP_TASKRUN | IORING_SETUP_TASKRUN_FLAG | 356197bbdc06SPavel Begunkov IORING_SETUP_SQE128 | IORING_SETUP_CQE32 | 3562c0e0d6baSDylan Yudaken IORING_SETUP_SINGLE_ISSUER | IORING_SETUP_DEFER_TASKRUN)) 3563ed29b0b4SJens Axboe return -EINVAL; 3564ed29b0b4SJens Axboe 3565ed29b0b4SJens Axboe return io_uring_create(entries, &p, params); 3566ed29b0b4SJens Axboe } 3567ed29b0b4SJens Axboe 3568ed29b0b4SJens Axboe SYSCALL_DEFINE2(io_uring_setup, u32, entries, 3569ed29b0b4SJens Axboe struct io_uring_params __user *, params) 3570ed29b0b4SJens Axboe { 3571ed29b0b4SJens Axboe return io_uring_setup(entries, params); 3572ed29b0b4SJens Axboe } 3573ed29b0b4SJens Axboe 3574ed29b0b4SJens Axboe static __cold int io_probe(struct io_ring_ctx *ctx, void __user *arg, 3575ed29b0b4SJens Axboe unsigned nr_args) 3576ed29b0b4SJens Axboe { 3577ed29b0b4SJens Axboe struct io_uring_probe *p; 3578ed29b0b4SJens Axboe size_t size; 3579ed29b0b4SJens Axboe int i, ret; 3580ed29b0b4SJens Axboe 3581ed29b0b4SJens Axboe size = struct_size(p, ops, nr_args); 3582ed29b0b4SJens Axboe if (size == SIZE_MAX) 3583ed29b0b4SJens Axboe return -EOVERFLOW; 3584ed29b0b4SJens Axboe p = kzalloc(size, GFP_KERNEL); 3585ed29b0b4SJens Axboe if (!p) 3586ed29b0b4SJens Axboe return -ENOMEM; 3587ed29b0b4SJens Axboe 3588ed29b0b4SJens Axboe ret = -EFAULT; 3589ed29b0b4SJens Axboe if (copy_from_user(p, arg, size)) 3590ed29b0b4SJens Axboe goto out; 3591ed29b0b4SJens Axboe ret = -EINVAL; 3592ed29b0b4SJens Axboe if (memchr_inv(p, 0, size)) 3593ed29b0b4SJens Axboe goto out; 3594ed29b0b4SJens Axboe 3595ed29b0b4SJens Axboe p->last_op = IORING_OP_LAST - 1; 3596ed29b0b4SJens Axboe if (nr_args > IORING_OP_LAST) 3597ed29b0b4SJens Axboe nr_args = IORING_OP_LAST; 3598ed29b0b4SJens Axboe 3599ed29b0b4SJens Axboe for (i = 0; i < nr_args; i++) { 3600ed29b0b4SJens Axboe p->ops[i].op = i; 3601ed29b0b4SJens Axboe if (!io_op_defs[i].not_supported) 3602ed29b0b4SJens Axboe p->ops[i].flags = IO_URING_OP_SUPPORTED; 3603ed29b0b4SJens Axboe } 3604ed29b0b4SJens Axboe p->ops_len = i; 3605ed29b0b4SJens Axboe 3606ed29b0b4SJens Axboe ret = 0; 3607ed29b0b4SJens Axboe if (copy_to_user(arg, p, size)) 3608ed29b0b4SJens Axboe ret = -EFAULT; 3609ed29b0b4SJens Axboe out: 3610ed29b0b4SJens Axboe kfree(p); 3611ed29b0b4SJens Axboe return ret; 3612ed29b0b4SJens Axboe } 3613ed29b0b4SJens Axboe 3614ed29b0b4SJens Axboe static int io_register_personality(struct io_ring_ctx *ctx) 3615ed29b0b4SJens Axboe { 3616ed29b0b4SJens Axboe const struct cred *creds; 3617ed29b0b4SJens Axboe u32 id; 3618ed29b0b4SJens Axboe int ret; 3619ed29b0b4SJens Axboe 3620ed29b0b4SJens Axboe creds = get_current_cred(); 3621ed29b0b4SJens Axboe 3622ed29b0b4SJens Axboe ret = xa_alloc_cyclic(&ctx->personalities, &id, (void *)creds, 3623ed29b0b4SJens Axboe XA_LIMIT(0, USHRT_MAX), &ctx->pers_next, GFP_KERNEL); 3624ed29b0b4SJens Axboe if (ret < 0) { 3625ed29b0b4SJens Axboe put_cred(creds); 3626ed29b0b4SJens Axboe return ret; 3627ed29b0b4SJens Axboe } 3628ed29b0b4SJens Axboe return id; 3629ed29b0b4SJens Axboe } 3630ed29b0b4SJens Axboe 3631ed29b0b4SJens Axboe static __cold int io_register_restrictions(struct io_ring_ctx *ctx, 3632ed29b0b4SJens Axboe void __user *arg, unsigned int nr_args) 3633ed29b0b4SJens Axboe { 3634ed29b0b4SJens Axboe struct io_uring_restriction *res; 3635ed29b0b4SJens Axboe size_t size; 3636ed29b0b4SJens Axboe int i, ret; 3637ed29b0b4SJens Axboe 3638ed29b0b4SJens Axboe /* Restrictions allowed only if rings started disabled */ 3639ed29b0b4SJens Axboe if (!(ctx->flags & IORING_SETUP_R_DISABLED)) 3640ed29b0b4SJens Axboe return -EBADFD; 3641ed29b0b4SJens Axboe 3642ed29b0b4SJens Axboe /* We allow only a single restrictions registration */ 3643ed29b0b4SJens Axboe if (ctx->restrictions.registered) 3644ed29b0b4SJens Axboe return -EBUSY; 3645ed29b0b4SJens Axboe 3646ed29b0b4SJens Axboe if (!arg || nr_args > IORING_MAX_RESTRICTIONS) 3647ed29b0b4SJens Axboe return -EINVAL; 3648ed29b0b4SJens Axboe 3649ed29b0b4SJens Axboe size = array_size(nr_args, sizeof(*res)); 3650ed29b0b4SJens Axboe if (size == SIZE_MAX) 3651ed29b0b4SJens Axboe return -EOVERFLOW; 3652ed29b0b4SJens Axboe 3653ed29b0b4SJens Axboe res = memdup_user(arg, size); 3654ed29b0b4SJens Axboe if (IS_ERR(res)) 3655ed29b0b4SJens Axboe return PTR_ERR(res); 3656ed29b0b4SJens Axboe 3657ed29b0b4SJens Axboe ret = 0; 3658ed29b0b4SJens Axboe 3659ed29b0b4SJens Axboe for (i = 0; i < nr_args; i++) { 3660ed29b0b4SJens Axboe switch (res[i].opcode) { 3661ed29b0b4SJens Axboe case IORING_RESTRICTION_REGISTER_OP: 3662ed29b0b4SJens Axboe if (res[i].register_op >= IORING_REGISTER_LAST) { 3663ed29b0b4SJens Axboe ret = -EINVAL; 3664ed29b0b4SJens Axboe goto out; 3665ed29b0b4SJens Axboe } 3666ed29b0b4SJens Axboe 3667ed29b0b4SJens Axboe __set_bit(res[i].register_op, 3668ed29b0b4SJens Axboe ctx->restrictions.register_op); 3669ed29b0b4SJens Axboe break; 3670ed29b0b4SJens Axboe case IORING_RESTRICTION_SQE_OP: 3671ed29b0b4SJens Axboe if (res[i].sqe_op >= IORING_OP_LAST) { 3672ed29b0b4SJens Axboe ret = -EINVAL; 3673ed29b0b4SJens Axboe goto out; 3674ed29b0b4SJens Axboe } 3675ed29b0b4SJens Axboe 3676ed29b0b4SJens Axboe __set_bit(res[i].sqe_op, ctx->restrictions.sqe_op); 3677ed29b0b4SJens Axboe break; 3678ed29b0b4SJens Axboe case IORING_RESTRICTION_SQE_FLAGS_ALLOWED: 3679ed29b0b4SJens Axboe ctx->restrictions.sqe_flags_allowed = res[i].sqe_flags; 3680ed29b0b4SJens Axboe break; 3681ed29b0b4SJens Axboe case IORING_RESTRICTION_SQE_FLAGS_REQUIRED: 3682ed29b0b4SJens Axboe ctx->restrictions.sqe_flags_required = res[i].sqe_flags; 3683ed29b0b4SJens Axboe break; 3684ed29b0b4SJens Axboe default: 3685ed29b0b4SJens Axboe ret = -EINVAL; 3686ed29b0b4SJens Axboe goto out; 3687ed29b0b4SJens Axboe } 3688ed29b0b4SJens Axboe } 3689ed29b0b4SJens Axboe 3690ed29b0b4SJens Axboe out: 3691ed29b0b4SJens Axboe /* Reset all restrictions if an error happened */ 3692ed29b0b4SJens Axboe if (ret != 0) 3693ed29b0b4SJens Axboe memset(&ctx->restrictions, 0, sizeof(ctx->restrictions)); 3694ed29b0b4SJens Axboe else 3695ed29b0b4SJens Axboe ctx->restrictions.registered = true; 3696ed29b0b4SJens Axboe 3697ed29b0b4SJens Axboe kfree(res); 3698ed29b0b4SJens Axboe return ret; 3699ed29b0b4SJens Axboe } 3700ed29b0b4SJens Axboe 3701ed29b0b4SJens Axboe static int io_register_enable_rings(struct io_ring_ctx *ctx) 3702ed29b0b4SJens Axboe { 3703ed29b0b4SJens Axboe if (!(ctx->flags & IORING_SETUP_R_DISABLED)) 3704ed29b0b4SJens Axboe return -EBADFD; 3705ed29b0b4SJens Axboe 37067cae596bSDylan Yudaken if (ctx->flags & IORING_SETUP_SINGLE_ISSUER && !ctx->submitter_task) 37077cae596bSDylan Yudaken ctx->submitter_task = get_task_struct(current); 37087cae596bSDylan Yudaken 3709ed29b0b4SJens Axboe if (ctx->restrictions.registered) 3710ed29b0b4SJens Axboe ctx->restricted = 1; 3711ed29b0b4SJens Axboe 3712ed29b0b4SJens Axboe ctx->flags &= ~IORING_SETUP_R_DISABLED; 3713ed29b0b4SJens Axboe if (ctx->sq_data && wq_has_sleeper(&ctx->sq_data->wait)) 3714ed29b0b4SJens Axboe wake_up(&ctx->sq_data->wait); 3715ed29b0b4SJens Axboe return 0; 3716ed29b0b4SJens Axboe } 3717ed29b0b4SJens Axboe 3718ed29b0b4SJens Axboe static __cold int io_register_iowq_aff(struct io_ring_ctx *ctx, 3719ed29b0b4SJens Axboe void __user *arg, unsigned len) 3720ed29b0b4SJens Axboe { 3721ed29b0b4SJens Axboe struct io_uring_task *tctx = current->io_uring; 3722ed29b0b4SJens Axboe cpumask_var_t new_mask; 3723ed29b0b4SJens Axboe int ret; 3724ed29b0b4SJens Axboe 3725ed29b0b4SJens Axboe if (!tctx || !tctx->io_wq) 3726ed29b0b4SJens Axboe return -EINVAL; 3727ed29b0b4SJens Axboe 3728ed29b0b4SJens Axboe if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) 3729ed29b0b4SJens Axboe return -ENOMEM; 3730ed29b0b4SJens Axboe 3731ed29b0b4SJens Axboe cpumask_clear(new_mask); 3732ed29b0b4SJens Axboe if (len > cpumask_size()) 3733ed29b0b4SJens Axboe len = cpumask_size(); 3734ed29b0b4SJens Axboe 3735ed29b0b4SJens Axboe if (in_compat_syscall()) { 3736ed29b0b4SJens Axboe ret = compat_get_bitmap(cpumask_bits(new_mask), 3737ed29b0b4SJens Axboe (const compat_ulong_t __user *)arg, 3738ed29b0b4SJens Axboe len * 8 /* CHAR_BIT */); 3739ed29b0b4SJens Axboe } else { 3740ed29b0b4SJens Axboe ret = copy_from_user(new_mask, arg, len); 3741ed29b0b4SJens Axboe } 3742ed29b0b4SJens Axboe 3743ed29b0b4SJens Axboe if (ret) { 3744ed29b0b4SJens Axboe free_cpumask_var(new_mask); 3745ed29b0b4SJens Axboe return -EFAULT; 3746ed29b0b4SJens Axboe } 3747ed29b0b4SJens Axboe 3748ed29b0b4SJens Axboe ret = io_wq_cpu_affinity(tctx->io_wq, new_mask); 3749ed29b0b4SJens Axboe free_cpumask_var(new_mask); 3750ed29b0b4SJens Axboe return ret; 3751ed29b0b4SJens Axboe } 3752ed29b0b4SJens Axboe 3753ed29b0b4SJens Axboe static __cold int io_unregister_iowq_aff(struct io_ring_ctx *ctx) 3754ed29b0b4SJens Axboe { 3755ed29b0b4SJens Axboe struct io_uring_task *tctx = current->io_uring; 3756ed29b0b4SJens Axboe 3757ed29b0b4SJens Axboe if (!tctx || !tctx->io_wq) 3758ed29b0b4SJens Axboe return -EINVAL; 3759ed29b0b4SJens Axboe 3760ed29b0b4SJens Axboe return io_wq_cpu_affinity(tctx->io_wq, NULL); 3761ed29b0b4SJens Axboe } 3762ed29b0b4SJens Axboe 3763ed29b0b4SJens Axboe static __cold int io_register_iowq_max_workers(struct io_ring_ctx *ctx, 3764ed29b0b4SJens Axboe void __user *arg) 3765ed29b0b4SJens Axboe __must_hold(&ctx->uring_lock) 3766ed29b0b4SJens Axboe { 3767ed29b0b4SJens Axboe struct io_tctx_node *node; 3768ed29b0b4SJens Axboe struct io_uring_task *tctx = NULL; 3769ed29b0b4SJens Axboe struct io_sq_data *sqd = NULL; 3770ed29b0b4SJens Axboe __u32 new_count[2]; 3771ed29b0b4SJens Axboe int i, ret; 3772ed29b0b4SJens Axboe 3773ed29b0b4SJens Axboe if (copy_from_user(new_count, arg, sizeof(new_count))) 3774ed29b0b4SJens Axboe return -EFAULT; 3775ed29b0b4SJens Axboe for (i = 0; i < ARRAY_SIZE(new_count); i++) 3776ed29b0b4SJens Axboe if (new_count[i] > INT_MAX) 3777ed29b0b4SJens Axboe return -EINVAL; 3778ed29b0b4SJens Axboe 3779ed29b0b4SJens Axboe if (ctx->flags & IORING_SETUP_SQPOLL) { 3780ed29b0b4SJens Axboe sqd = ctx->sq_data; 3781ed29b0b4SJens Axboe if (sqd) { 3782ed29b0b4SJens Axboe /* 3783ed29b0b4SJens Axboe * Observe the correct sqd->lock -> ctx->uring_lock 3784ed29b0b4SJens Axboe * ordering. Fine to drop uring_lock here, we hold 3785ed29b0b4SJens Axboe * a ref to the ctx. 3786ed29b0b4SJens Axboe */ 3787ed29b0b4SJens Axboe refcount_inc(&sqd->refs); 3788ed29b0b4SJens Axboe mutex_unlock(&ctx->uring_lock); 3789ed29b0b4SJens Axboe mutex_lock(&sqd->lock); 3790ed29b0b4SJens Axboe mutex_lock(&ctx->uring_lock); 3791ed29b0b4SJens Axboe if (sqd->thread) 3792ed29b0b4SJens Axboe tctx = sqd->thread->io_uring; 3793ed29b0b4SJens Axboe } 3794ed29b0b4SJens Axboe } else { 3795ed29b0b4SJens Axboe tctx = current->io_uring; 3796ed29b0b4SJens Axboe } 3797ed29b0b4SJens Axboe 3798ed29b0b4SJens Axboe BUILD_BUG_ON(sizeof(new_count) != sizeof(ctx->iowq_limits)); 3799ed29b0b4SJens Axboe 3800ed29b0b4SJens Axboe for (i = 0; i < ARRAY_SIZE(new_count); i++) 3801ed29b0b4SJens Axboe if (new_count[i]) 3802ed29b0b4SJens Axboe ctx->iowq_limits[i] = new_count[i]; 3803ed29b0b4SJens Axboe ctx->iowq_limits_set = true; 3804ed29b0b4SJens Axboe 3805ed29b0b4SJens Axboe if (tctx && tctx->io_wq) { 3806ed29b0b4SJens Axboe ret = io_wq_max_workers(tctx->io_wq, new_count); 3807ed29b0b4SJens Axboe if (ret) 3808ed29b0b4SJens Axboe goto err; 3809ed29b0b4SJens Axboe } else { 3810ed29b0b4SJens Axboe memset(new_count, 0, sizeof(new_count)); 3811ed29b0b4SJens Axboe } 3812ed29b0b4SJens Axboe 3813ed29b0b4SJens Axboe if (sqd) { 3814ed29b0b4SJens Axboe mutex_unlock(&sqd->lock); 3815ed29b0b4SJens Axboe io_put_sq_data(sqd); 3816ed29b0b4SJens Axboe } 3817ed29b0b4SJens Axboe 3818ed29b0b4SJens Axboe if (copy_to_user(arg, new_count, sizeof(new_count))) 3819ed29b0b4SJens Axboe return -EFAULT; 3820ed29b0b4SJens Axboe 3821ed29b0b4SJens Axboe /* that's it for SQPOLL, only the SQPOLL task creates requests */ 3822ed29b0b4SJens Axboe if (sqd) 3823ed29b0b4SJens Axboe return 0; 3824ed29b0b4SJens Axboe 3825ed29b0b4SJens Axboe /* now propagate the restriction to all registered users */ 3826ed29b0b4SJens Axboe list_for_each_entry(node, &ctx->tctx_list, ctx_node) { 3827ed29b0b4SJens Axboe struct io_uring_task *tctx = node->task->io_uring; 3828ed29b0b4SJens Axboe 3829ed29b0b4SJens Axboe if (WARN_ON_ONCE(!tctx->io_wq)) 3830ed29b0b4SJens Axboe continue; 3831ed29b0b4SJens Axboe 3832ed29b0b4SJens Axboe for (i = 0; i < ARRAY_SIZE(new_count); i++) 3833ed29b0b4SJens Axboe new_count[i] = ctx->iowq_limits[i]; 3834ed29b0b4SJens Axboe /* ignore errors, it always returns zero anyway */ 3835ed29b0b4SJens Axboe (void)io_wq_max_workers(tctx->io_wq, new_count); 3836ed29b0b4SJens Axboe } 3837ed29b0b4SJens Axboe return 0; 3838ed29b0b4SJens Axboe err: 3839ed29b0b4SJens Axboe if (sqd) { 3840ed29b0b4SJens Axboe mutex_unlock(&sqd->lock); 3841ed29b0b4SJens Axboe io_put_sq_data(sqd); 3842ed29b0b4SJens Axboe } 3843ed29b0b4SJens Axboe return ret; 3844ed29b0b4SJens Axboe } 3845ed29b0b4SJens Axboe 3846ed29b0b4SJens Axboe static int __io_uring_register(struct io_ring_ctx *ctx, unsigned opcode, 3847ed29b0b4SJens Axboe void __user *arg, unsigned nr_args) 3848ed29b0b4SJens Axboe __releases(ctx->uring_lock) 3849ed29b0b4SJens Axboe __acquires(ctx->uring_lock) 3850ed29b0b4SJens Axboe { 3851ed29b0b4SJens Axboe int ret; 3852ed29b0b4SJens Axboe 3853ed29b0b4SJens Axboe /* 3854fbb8bb02SPavel Begunkov * We don't quiesce the refs for register anymore and so it can't be 3855fbb8bb02SPavel Begunkov * dying as we're holding a file ref here. 3856ed29b0b4SJens Axboe */ 3857fbb8bb02SPavel Begunkov if (WARN_ON_ONCE(percpu_ref_is_dying(&ctx->refs))) 3858ed29b0b4SJens Axboe return -ENXIO; 3859ed29b0b4SJens Axboe 3860d7cce96cSPavel Begunkov if (ctx->submitter_task && ctx->submitter_task != current) 3861d7cce96cSPavel Begunkov return -EEXIST; 3862d7cce96cSPavel Begunkov 3863ed29b0b4SJens Axboe if (ctx->restricted) { 3864ed29b0b4SJens Axboe if (opcode >= IORING_REGISTER_LAST) 3865ed29b0b4SJens Axboe return -EINVAL; 3866ed29b0b4SJens Axboe opcode = array_index_nospec(opcode, IORING_REGISTER_LAST); 3867ed29b0b4SJens Axboe if (!test_bit(opcode, ctx->restrictions.register_op)) 3868ed29b0b4SJens Axboe return -EACCES; 3869ed29b0b4SJens Axboe } 3870ed29b0b4SJens Axboe 3871ed29b0b4SJens Axboe switch (opcode) { 3872ed29b0b4SJens Axboe case IORING_REGISTER_BUFFERS: 3873ed29b0b4SJens Axboe ret = -EFAULT; 3874ed29b0b4SJens Axboe if (!arg) 3875ed29b0b4SJens Axboe break; 3876ed29b0b4SJens Axboe ret = io_sqe_buffers_register(ctx, arg, nr_args, NULL); 3877ed29b0b4SJens Axboe break; 3878ed29b0b4SJens Axboe case IORING_UNREGISTER_BUFFERS: 3879ed29b0b4SJens Axboe ret = -EINVAL; 3880ed29b0b4SJens Axboe if (arg || nr_args) 3881ed29b0b4SJens Axboe break; 3882ed29b0b4SJens Axboe ret = io_sqe_buffers_unregister(ctx); 3883ed29b0b4SJens Axboe break; 3884ed29b0b4SJens Axboe case IORING_REGISTER_FILES: 3885ed29b0b4SJens Axboe ret = -EFAULT; 3886ed29b0b4SJens Axboe if (!arg) 3887ed29b0b4SJens Axboe break; 3888ed29b0b4SJens Axboe ret = io_sqe_files_register(ctx, arg, nr_args, NULL); 3889ed29b0b4SJens Axboe break; 3890ed29b0b4SJens Axboe case IORING_UNREGISTER_FILES: 3891ed29b0b4SJens Axboe ret = -EINVAL; 3892ed29b0b4SJens Axboe if (arg || nr_args) 3893ed29b0b4SJens Axboe break; 3894ed29b0b4SJens Axboe ret = io_sqe_files_unregister(ctx); 3895ed29b0b4SJens Axboe break; 3896ed29b0b4SJens Axboe case IORING_REGISTER_FILES_UPDATE: 3897ed29b0b4SJens Axboe ret = io_register_files_update(ctx, arg, nr_args); 3898ed29b0b4SJens Axboe break; 3899ed29b0b4SJens Axboe case IORING_REGISTER_EVENTFD: 3900ed29b0b4SJens Axboe ret = -EINVAL; 3901ed29b0b4SJens Axboe if (nr_args != 1) 3902ed29b0b4SJens Axboe break; 3903ed29b0b4SJens Axboe ret = io_eventfd_register(ctx, arg, 0); 3904ed29b0b4SJens Axboe break; 3905ed29b0b4SJens Axboe case IORING_REGISTER_EVENTFD_ASYNC: 3906ed29b0b4SJens Axboe ret = -EINVAL; 3907ed29b0b4SJens Axboe if (nr_args != 1) 3908ed29b0b4SJens Axboe break; 3909ed29b0b4SJens Axboe ret = io_eventfd_register(ctx, arg, 1); 3910ed29b0b4SJens Axboe break; 3911ed29b0b4SJens Axboe case IORING_UNREGISTER_EVENTFD: 3912ed29b0b4SJens Axboe ret = -EINVAL; 3913ed29b0b4SJens Axboe if (arg || nr_args) 3914ed29b0b4SJens Axboe break; 3915ed29b0b4SJens Axboe ret = io_eventfd_unregister(ctx); 3916ed29b0b4SJens Axboe break; 3917ed29b0b4SJens Axboe case IORING_REGISTER_PROBE: 3918ed29b0b4SJens Axboe ret = -EINVAL; 3919ed29b0b4SJens Axboe if (!arg || nr_args > 256) 3920ed29b0b4SJens Axboe break; 3921ed29b0b4SJens Axboe ret = io_probe(ctx, arg, nr_args); 3922ed29b0b4SJens Axboe break; 3923ed29b0b4SJens Axboe case IORING_REGISTER_PERSONALITY: 3924ed29b0b4SJens Axboe ret = -EINVAL; 3925ed29b0b4SJens Axboe if (arg || nr_args) 3926ed29b0b4SJens Axboe break; 3927ed29b0b4SJens Axboe ret = io_register_personality(ctx); 3928ed29b0b4SJens Axboe break; 3929ed29b0b4SJens Axboe case IORING_UNREGISTER_PERSONALITY: 3930ed29b0b4SJens Axboe ret = -EINVAL; 3931ed29b0b4SJens Axboe if (arg) 3932ed29b0b4SJens Axboe break; 3933ed29b0b4SJens Axboe ret = io_unregister_personality(ctx, nr_args); 3934ed29b0b4SJens Axboe break; 3935ed29b0b4SJens Axboe case IORING_REGISTER_ENABLE_RINGS: 3936ed29b0b4SJens Axboe ret = -EINVAL; 3937ed29b0b4SJens Axboe if (arg || nr_args) 3938ed29b0b4SJens Axboe break; 3939ed29b0b4SJens Axboe ret = io_register_enable_rings(ctx); 3940ed29b0b4SJens Axboe break; 3941ed29b0b4SJens Axboe case IORING_REGISTER_RESTRICTIONS: 3942ed29b0b4SJens Axboe ret = io_register_restrictions(ctx, arg, nr_args); 3943ed29b0b4SJens Axboe break; 3944ed29b0b4SJens Axboe case IORING_REGISTER_FILES2: 3945ed29b0b4SJens Axboe ret = io_register_rsrc(ctx, arg, nr_args, IORING_RSRC_FILE); 3946ed29b0b4SJens Axboe break; 3947ed29b0b4SJens Axboe case IORING_REGISTER_FILES_UPDATE2: 3948ed29b0b4SJens Axboe ret = io_register_rsrc_update(ctx, arg, nr_args, 3949ed29b0b4SJens Axboe IORING_RSRC_FILE); 3950ed29b0b4SJens Axboe break; 3951ed29b0b4SJens Axboe case IORING_REGISTER_BUFFERS2: 3952ed29b0b4SJens Axboe ret = io_register_rsrc(ctx, arg, nr_args, IORING_RSRC_BUFFER); 3953ed29b0b4SJens Axboe break; 3954ed29b0b4SJens Axboe case IORING_REGISTER_BUFFERS_UPDATE: 3955ed29b0b4SJens Axboe ret = io_register_rsrc_update(ctx, arg, nr_args, 3956ed29b0b4SJens Axboe IORING_RSRC_BUFFER); 3957ed29b0b4SJens Axboe break; 3958ed29b0b4SJens Axboe case IORING_REGISTER_IOWQ_AFF: 3959ed29b0b4SJens Axboe ret = -EINVAL; 3960ed29b0b4SJens Axboe if (!arg || !nr_args) 3961ed29b0b4SJens Axboe break; 3962ed29b0b4SJens Axboe ret = io_register_iowq_aff(ctx, arg, nr_args); 3963ed29b0b4SJens Axboe break; 3964ed29b0b4SJens Axboe case IORING_UNREGISTER_IOWQ_AFF: 3965ed29b0b4SJens Axboe ret = -EINVAL; 3966ed29b0b4SJens Axboe if (arg || nr_args) 3967ed29b0b4SJens Axboe break; 3968ed29b0b4SJens Axboe ret = io_unregister_iowq_aff(ctx); 3969ed29b0b4SJens Axboe break; 3970ed29b0b4SJens Axboe case IORING_REGISTER_IOWQ_MAX_WORKERS: 3971ed29b0b4SJens Axboe ret = -EINVAL; 3972ed29b0b4SJens Axboe if (!arg || nr_args != 2) 3973ed29b0b4SJens Axboe break; 3974ed29b0b4SJens Axboe ret = io_register_iowq_max_workers(ctx, arg); 3975ed29b0b4SJens Axboe break; 3976ed29b0b4SJens Axboe case IORING_REGISTER_RING_FDS: 3977ed29b0b4SJens Axboe ret = io_ringfd_register(ctx, arg, nr_args); 3978ed29b0b4SJens Axboe break; 3979ed29b0b4SJens Axboe case IORING_UNREGISTER_RING_FDS: 3980ed29b0b4SJens Axboe ret = io_ringfd_unregister(ctx, arg, nr_args); 3981ed29b0b4SJens Axboe break; 3982ed29b0b4SJens Axboe case IORING_REGISTER_PBUF_RING: 3983ed29b0b4SJens Axboe ret = -EINVAL; 3984ed29b0b4SJens Axboe if (!arg || nr_args != 1) 3985ed29b0b4SJens Axboe break; 3986ed29b0b4SJens Axboe ret = io_register_pbuf_ring(ctx, arg); 3987ed29b0b4SJens Axboe break; 3988ed29b0b4SJens Axboe case IORING_UNREGISTER_PBUF_RING: 3989ed29b0b4SJens Axboe ret = -EINVAL; 3990ed29b0b4SJens Axboe if (!arg || nr_args != 1) 3991ed29b0b4SJens Axboe break; 3992ed29b0b4SJens Axboe ret = io_unregister_pbuf_ring(ctx, arg); 3993ed29b0b4SJens Axboe break; 399478a861b9SJens Axboe case IORING_REGISTER_SYNC_CANCEL: 399578a861b9SJens Axboe ret = -EINVAL; 399678a861b9SJens Axboe if (!arg || nr_args != 1) 399778a861b9SJens Axboe break; 399878a861b9SJens Axboe ret = io_sync_cancel(ctx, arg); 399978a861b9SJens Axboe break; 40006e73dffbSPavel Begunkov case IORING_REGISTER_FILE_ALLOC_RANGE: 40016e73dffbSPavel Begunkov ret = -EINVAL; 40026e73dffbSPavel Begunkov if (!arg || nr_args) 40036e73dffbSPavel Begunkov break; 40046e73dffbSPavel Begunkov ret = io_register_file_alloc_range(ctx, arg); 40056e73dffbSPavel Begunkov break; 4006ed29b0b4SJens Axboe default: 4007ed29b0b4SJens Axboe ret = -EINVAL; 4008ed29b0b4SJens Axboe break; 4009ed29b0b4SJens Axboe } 4010ed29b0b4SJens Axboe 4011ed29b0b4SJens Axboe return ret; 4012ed29b0b4SJens Axboe } 4013ed29b0b4SJens Axboe 4014ed29b0b4SJens Axboe SYSCALL_DEFINE4(io_uring_register, unsigned int, fd, unsigned int, opcode, 4015ed29b0b4SJens Axboe void __user *, arg, unsigned int, nr_args) 4016ed29b0b4SJens Axboe { 4017ed29b0b4SJens Axboe struct io_ring_ctx *ctx; 4018ed29b0b4SJens Axboe long ret = -EBADF; 4019ed29b0b4SJens Axboe struct fd f; 4020ed29b0b4SJens Axboe 4021ed29b0b4SJens Axboe f = fdget(fd); 4022ed29b0b4SJens Axboe if (!f.file) 4023ed29b0b4SJens Axboe return -EBADF; 4024ed29b0b4SJens Axboe 4025ed29b0b4SJens Axboe ret = -EOPNOTSUPP; 4026e5550a14SJens Axboe if (!io_is_uring_fops(f.file)) 4027ed29b0b4SJens Axboe goto out_fput; 4028ed29b0b4SJens Axboe 4029ed29b0b4SJens Axboe ctx = f.file->private_data; 4030ed29b0b4SJens Axboe 4031ed29b0b4SJens Axboe mutex_lock(&ctx->uring_lock); 4032ed29b0b4SJens Axboe ret = __io_uring_register(ctx, opcode, arg, nr_args); 4033ed29b0b4SJens Axboe mutex_unlock(&ctx->uring_lock); 4034ed29b0b4SJens Axboe trace_io_uring_register(ctx, opcode, ctx->nr_user_files, ctx->nr_user_bufs, ret); 4035ed29b0b4SJens Axboe out_fput: 4036ed29b0b4SJens Axboe fdput(f); 4037ed29b0b4SJens Axboe return ret; 4038ed29b0b4SJens Axboe } 4039ed29b0b4SJens Axboe 4040ed29b0b4SJens Axboe static int __init io_uring_init(void) 4041ed29b0b4SJens Axboe { 40429c71d39aSStefan Metzmacher #define __BUILD_BUG_VERIFY_OFFSET_SIZE(stype, eoffset, esize, ename) do { \ 4043ed29b0b4SJens Axboe BUILD_BUG_ON(offsetof(stype, ename) != eoffset); \ 40449c71d39aSStefan Metzmacher BUILD_BUG_ON(sizeof_field(stype, ename) != esize); \ 4045ed29b0b4SJens Axboe } while (0) 4046ed29b0b4SJens Axboe 4047ed29b0b4SJens Axboe #define BUILD_BUG_SQE_ELEM(eoffset, etype, ename) \ 40489c71d39aSStefan Metzmacher __BUILD_BUG_VERIFY_OFFSET_SIZE(struct io_uring_sqe, eoffset, sizeof(etype), ename) 40499c71d39aSStefan Metzmacher #define BUILD_BUG_SQE_ELEM_SIZE(eoffset, esize, ename) \ 40509c71d39aSStefan Metzmacher __BUILD_BUG_VERIFY_OFFSET_SIZE(struct io_uring_sqe, eoffset, esize, ename) 4051ed29b0b4SJens Axboe BUILD_BUG_ON(sizeof(struct io_uring_sqe) != 64); 4052ed29b0b4SJens Axboe BUILD_BUG_SQE_ELEM(0, __u8, opcode); 4053ed29b0b4SJens Axboe BUILD_BUG_SQE_ELEM(1, __u8, flags); 4054ed29b0b4SJens Axboe BUILD_BUG_SQE_ELEM(2, __u16, ioprio); 4055ed29b0b4SJens Axboe BUILD_BUG_SQE_ELEM(4, __s32, fd); 4056ed29b0b4SJens Axboe BUILD_BUG_SQE_ELEM(8, __u64, off); 4057ed29b0b4SJens Axboe BUILD_BUG_SQE_ELEM(8, __u64, addr2); 40589c71d39aSStefan Metzmacher BUILD_BUG_SQE_ELEM(8, __u32, cmd_op); 40599c71d39aSStefan Metzmacher BUILD_BUG_SQE_ELEM(12, __u32, __pad1); 4060ed29b0b4SJens Axboe BUILD_BUG_SQE_ELEM(16, __u64, addr); 4061ed29b0b4SJens Axboe BUILD_BUG_SQE_ELEM(16, __u64, splice_off_in); 4062ed29b0b4SJens Axboe BUILD_BUG_SQE_ELEM(24, __u32, len); 4063ed29b0b4SJens Axboe BUILD_BUG_SQE_ELEM(28, __kernel_rwf_t, rw_flags); 4064ed29b0b4SJens Axboe BUILD_BUG_SQE_ELEM(28, /* compat */ int, rw_flags); 4065ed29b0b4SJens Axboe BUILD_BUG_SQE_ELEM(28, /* compat */ __u32, rw_flags); 4066ed29b0b4SJens Axboe BUILD_BUG_SQE_ELEM(28, __u32, fsync_flags); 4067ed29b0b4SJens Axboe BUILD_BUG_SQE_ELEM(28, /* compat */ __u16, poll_events); 4068ed29b0b4SJens Axboe BUILD_BUG_SQE_ELEM(28, __u32, poll32_events); 4069ed29b0b4SJens Axboe BUILD_BUG_SQE_ELEM(28, __u32, sync_range_flags); 4070ed29b0b4SJens Axboe BUILD_BUG_SQE_ELEM(28, __u32, msg_flags); 4071ed29b0b4SJens Axboe BUILD_BUG_SQE_ELEM(28, __u32, timeout_flags); 4072ed29b0b4SJens Axboe BUILD_BUG_SQE_ELEM(28, __u32, accept_flags); 4073ed29b0b4SJens Axboe BUILD_BUG_SQE_ELEM(28, __u32, cancel_flags); 4074ed29b0b4SJens Axboe BUILD_BUG_SQE_ELEM(28, __u32, open_flags); 4075ed29b0b4SJens Axboe BUILD_BUG_SQE_ELEM(28, __u32, statx_flags); 4076ed29b0b4SJens Axboe BUILD_BUG_SQE_ELEM(28, __u32, fadvise_advice); 4077ed29b0b4SJens Axboe BUILD_BUG_SQE_ELEM(28, __u32, splice_flags); 40789c71d39aSStefan Metzmacher BUILD_BUG_SQE_ELEM(28, __u32, rename_flags); 40799c71d39aSStefan Metzmacher BUILD_BUG_SQE_ELEM(28, __u32, unlink_flags); 40809c71d39aSStefan Metzmacher BUILD_BUG_SQE_ELEM(28, __u32, hardlink_flags); 40819c71d39aSStefan Metzmacher BUILD_BUG_SQE_ELEM(28, __u32, xattr_flags); 40829c71d39aSStefan Metzmacher BUILD_BUG_SQE_ELEM(28, __u32, msg_ring_flags); 4083ed29b0b4SJens Axboe BUILD_BUG_SQE_ELEM(32, __u64, user_data); 4084ed29b0b4SJens Axboe BUILD_BUG_SQE_ELEM(40, __u16, buf_index); 4085ed29b0b4SJens Axboe BUILD_BUG_SQE_ELEM(40, __u16, buf_group); 4086ed29b0b4SJens Axboe BUILD_BUG_SQE_ELEM(42, __u16, personality); 4087ed29b0b4SJens Axboe BUILD_BUG_SQE_ELEM(44, __s32, splice_fd_in); 4088ed29b0b4SJens Axboe BUILD_BUG_SQE_ELEM(44, __u32, file_index); 4089b48c312bSPavel Begunkov BUILD_BUG_SQE_ELEM(44, __u16, addr_len); 4090b48c312bSPavel Begunkov BUILD_BUG_SQE_ELEM(46, __u16, __pad3[0]); 4091ed29b0b4SJens Axboe BUILD_BUG_SQE_ELEM(48, __u64, addr3); 40929c71d39aSStefan Metzmacher BUILD_BUG_SQE_ELEM_SIZE(48, 0, cmd); 40939c71d39aSStefan Metzmacher BUILD_BUG_SQE_ELEM(56, __u64, __pad2); 4094ed29b0b4SJens Axboe 4095ed29b0b4SJens Axboe BUILD_BUG_ON(sizeof(struct io_uring_files_update) != 4096ed29b0b4SJens Axboe sizeof(struct io_uring_rsrc_update)); 4097ed29b0b4SJens Axboe BUILD_BUG_ON(sizeof(struct io_uring_rsrc_update) > 4098ed29b0b4SJens Axboe sizeof(struct io_uring_rsrc_update2)); 4099ed29b0b4SJens Axboe 4100ed29b0b4SJens Axboe /* ->buf_index is u16 */ 4101ed29b0b4SJens Axboe BUILD_BUG_ON(offsetof(struct io_uring_buf_ring, bufs) != 0); 4102ed29b0b4SJens Axboe BUILD_BUG_ON(offsetof(struct io_uring_buf, resv) != 4103ed29b0b4SJens Axboe offsetof(struct io_uring_buf_ring, tail)); 4104ed29b0b4SJens Axboe 4105ed29b0b4SJens Axboe /* should fit into one byte */ 4106ed29b0b4SJens Axboe BUILD_BUG_ON(SQE_VALID_FLAGS >= (1 << 8)); 4107ed29b0b4SJens Axboe BUILD_BUG_ON(SQE_COMMON_FLAGS >= (1 << 8)); 4108ed29b0b4SJens Axboe BUILD_BUG_ON((SQE_VALID_FLAGS | SQE_COMMON_FLAGS) != SQE_VALID_FLAGS); 4109ed29b0b4SJens Axboe 4110ed29b0b4SJens Axboe BUILD_BUG_ON(__REQ_F_LAST_BIT > 8 * sizeof(int)); 4111ed29b0b4SJens Axboe 4112ed29b0b4SJens Axboe BUILD_BUG_ON(sizeof(atomic_t) != sizeof(u32)); 4113ed29b0b4SJens Axboe 4114d9b57aa3SJens Axboe io_uring_optable_init(); 4115ed29b0b4SJens Axboe 4116ed29b0b4SJens Axboe req_cachep = KMEM_CACHE(io_kiocb, SLAB_HWCACHE_ALIGN | SLAB_PANIC | 4117ed29b0b4SJens Axboe SLAB_ACCOUNT); 4118ed29b0b4SJens Axboe return 0; 4119ed29b0b4SJens Axboe }; 4120ed29b0b4SJens Axboe __initcall(io_uring_init); 4121