157dacad5SJay Sternberg /* 257dacad5SJay Sternberg * NVM Express device driver 357dacad5SJay Sternberg * Copyright (c) 2011-2014, Intel Corporation. 457dacad5SJay Sternberg * 557dacad5SJay Sternberg * This program is free software; you can redistribute it and/or modify it 657dacad5SJay Sternberg * under the terms and conditions of the GNU General Public License, 757dacad5SJay Sternberg * version 2, as published by the Free Software Foundation. 857dacad5SJay Sternberg * 957dacad5SJay Sternberg * This program is distributed in the hope it will be useful, but WITHOUT 1057dacad5SJay Sternberg * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 1157dacad5SJay Sternberg * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for 1257dacad5SJay Sternberg * more details. 1357dacad5SJay Sternberg */ 1457dacad5SJay Sternberg 15a0a3408eSKeith Busch #include <linux/aer.h> 1657dacad5SJay Sternberg #include <linux/bitops.h> 1757dacad5SJay Sternberg #include <linux/blkdev.h> 1857dacad5SJay Sternberg #include <linux/blk-mq.h> 1957dacad5SJay Sternberg #include <linux/cpu.h> 2057dacad5SJay Sternberg #include <linux/delay.h> 2157dacad5SJay Sternberg #include <linux/errno.h> 2257dacad5SJay Sternberg #include <linux/fs.h> 2357dacad5SJay Sternberg #include <linux/genhd.h> 2457dacad5SJay Sternberg #include <linux/hdreg.h> 2557dacad5SJay Sternberg #include <linux/idr.h> 2657dacad5SJay Sternberg #include <linux/init.h> 2757dacad5SJay Sternberg #include <linux/interrupt.h> 2857dacad5SJay Sternberg #include <linux/io.h> 2957dacad5SJay Sternberg #include <linux/kdev_t.h> 3057dacad5SJay Sternberg #include <linux/kthread.h> 3157dacad5SJay Sternberg #include <linux/kernel.h> 3257dacad5SJay Sternberg #include <linux/mm.h> 3357dacad5SJay Sternberg #include <linux/module.h> 3457dacad5SJay Sternberg #include <linux/moduleparam.h> 3577bf25eaSKeith Busch #include <linux/mutex.h> 3657dacad5SJay Sternberg #include <linux/pci.h> 3757dacad5SJay Sternberg #include <linux/poison.h> 3857dacad5SJay Sternberg #include <linux/ptrace.h> 3957dacad5SJay Sternberg #include <linux/sched.h> 4057dacad5SJay Sternberg #include <linux/slab.h> 4157dacad5SJay Sternberg #include <linux/t10-pi.h> 4257dacad5SJay Sternberg #include <linux/types.h> 439cf5c095SLinus Torvalds #include <linux/io-64-nonatomic-lo-hi.h> 441d277a63SKeith Busch #include <asm/unaligned.h> 4557dacad5SJay Sternberg 4657dacad5SJay Sternberg #include "nvme.h" 4757dacad5SJay Sternberg 4857dacad5SJay Sternberg #define NVME_Q_DEPTH 1024 4957dacad5SJay Sternberg #define NVME_AQ_DEPTH 256 5057dacad5SJay Sternberg #define SQ_SIZE(depth) (depth * sizeof(struct nvme_command)) 5157dacad5SJay Sternberg #define CQ_SIZE(depth) (depth * sizeof(struct nvme_completion)) 5257dacad5SJay Sternberg 53adf68f21SChristoph Hellwig /* 54adf68f21SChristoph Hellwig * We handle AEN commands ourselves and don't even let the 55adf68f21SChristoph Hellwig * block layer know about them. 56adf68f21SChristoph Hellwig */ 57adf68f21SChristoph Hellwig #define NVME_NR_AEN_COMMANDS 1 58adf68f21SChristoph Hellwig #define NVME_AQ_BLKMQ_DEPTH (NVME_AQ_DEPTH - NVME_NR_AEN_COMMANDS) 59adf68f21SChristoph Hellwig 6021d34711SChristoph Hellwig unsigned char admin_timeout = 60; 6157dacad5SJay Sternberg module_param(admin_timeout, byte, 0644); 6257dacad5SJay Sternberg MODULE_PARM_DESC(admin_timeout, "timeout in seconds for admin commands"); 6357dacad5SJay Sternberg 6457dacad5SJay Sternberg unsigned char nvme_io_timeout = 30; 6557dacad5SJay Sternberg module_param_named(io_timeout, nvme_io_timeout, byte, 0644); 6657dacad5SJay Sternberg MODULE_PARM_DESC(io_timeout, "timeout in seconds for I/O"); 6757dacad5SJay Sternberg 685fd4ce1bSChristoph Hellwig unsigned char shutdown_timeout = 5; 6957dacad5SJay Sternberg module_param(shutdown_timeout, byte, 0644); 7057dacad5SJay Sternberg MODULE_PARM_DESC(shutdown_timeout, "timeout in seconds for controller shutdown"); 7157dacad5SJay Sternberg 7257dacad5SJay Sternberg static int use_threaded_interrupts; 7357dacad5SJay Sternberg module_param(use_threaded_interrupts, int, 0); 7457dacad5SJay Sternberg 7557dacad5SJay Sternberg static bool use_cmb_sqes = true; 7657dacad5SJay Sternberg module_param(use_cmb_sqes, bool, 0644); 7757dacad5SJay Sternberg MODULE_PARM_DESC(use_cmb_sqes, "use controller's memory buffer for I/O SQes"); 7857dacad5SJay Sternberg 7957dacad5SJay Sternberg static LIST_HEAD(dev_list); 8057dacad5SJay Sternberg static struct task_struct *nvme_thread; 8157dacad5SJay Sternberg static struct workqueue_struct *nvme_workq; 8257dacad5SJay Sternberg static wait_queue_head_t nvme_kthread_wait; 8357dacad5SJay Sternberg 841c63dc66SChristoph Hellwig struct nvme_dev; 851c63dc66SChristoph Hellwig struct nvme_queue; 8657dacad5SJay Sternberg 8757dacad5SJay Sternberg static int nvme_reset(struct nvme_dev *dev); 88a0fa9647SJens Axboe static void nvme_process_cq(struct nvme_queue *nvmeq); 895c8809e6SChristoph Hellwig static void nvme_remove_dead_ctrl(struct nvme_dev *dev); 90a5cdb68cSKeith Busch static void nvme_dev_disable(struct nvme_dev *dev, bool shutdown); 9157dacad5SJay Sternberg 9257dacad5SJay Sternberg /* 931c63dc66SChristoph Hellwig * Represents an NVM Express device. Each nvme_dev is a PCI function. 941c63dc66SChristoph Hellwig */ 951c63dc66SChristoph Hellwig struct nvme_dev { 961c63dc66SChristoph Hellwig struct list_head node; 971c63dc66SChristoph Hellwig struct nvme_queue **queues; 981c63dc66SChristoph Hellwig struct blk_mq_tag_set tagset; 991c63dc66SChristoph Hellwig struct blk_mq_tag_set admin_tagset; 1001c63dc66SChristoph Hellwig u32 __iomem *dbs; 1011c63dc66SChristoph Hellwig struct device *dev; 1021c63dc66SChristoph Hellwig struct dma_pool *prp_page_pool; 1031c63dc66SChristoph Hellwig struct dma_pool *prp_small_pool; 1041c63dc66SChristoph Hellwig unsigned queue_count; 1051c63dc66SChristoph Hellwig unsigned online_queues; 1061c63dc66SChristoph Hellwig unsigned max_qid; 1071c63dc66SChristoph Hellwig int q_depth; 1081c63dc66SChristoph Hellwig u32 db_stride; 1091c63dc66SChristoph Hellwig struct msix_entry *entry; 1101c63dc66SChristoph Hellwig void __iomem *bar; 1111c63dc66SChristoph Hellwig struct work_struct reset_work; 1121c63dc66SChristoph Hellwig struct work_struct scan_work; 1135c8809e6SChristoph Hellwig struct work_struct remove_work; 11477bf25eaSKeith Busch struct mutex shutdown_lock; 1151c63dc66SChristoph Hellwig bool subsystem; 1161c63dc66SChristoph Hellwig void __iomem *cmb; 1171c63dc66SChristoph Hellwig dma_addr_t cmb_dma_addr; 1181c63dc66SChristoph Hellwig u64 cmb_size; 1191c63dc66SChristoph Hellwig u32 cmbsz; 120fd634f41SChristoph Hellwig unsigned long flags; 121db3cbfffSKeith Busch 122fd634f41SChristoph Hellwig #define NVME_CTRL_RESETTING 0 1231c63dc66SChristoph Hellwig 1241c63dc66SChristoph Hellwig struct nvme_ctrl ctrl; 125db3cbfffSKeith Busch struct completion ioq_wait; 12657dacad5SJay Sternberg }; 12757dacad5SJay Sternberg 1281c63dc66SChristoph Hellwig static inline struct nvme_dev *to_nvme_dev(struct nvme_ctrl *ctrl) 1291c63dc66SChristoph Hellwig { 1301c63dc66SChristoph Hellwig return container_of(ctrl, struct nvme_dev, ctrl); 1311c63dc66SChristoph Hellwig } 1321c63dc66SChristoph Hellwig 13357dacad5SJay Sternberg /* 13457dacad5SJay Sternberg * An NVM Express queue. Each device has at least two (one for admin 13557dacad5SJay Sternberg * commands and one for I/O commands). 13657dacad5SJay Sternberg */ 13757dacad5SJay Sternberg struct nvme_queue { 13857dacad5SJay Sternberg struct device *q_dmadev; 13957dacad5SJay Sternberg struct nvme_dev *dev; 14057dacad5SJay Sternberg char irqname[24]; /* nvme4294967295-65535\0 */ 14157dacad5SJay Sternberg spinlock_t q_lock; 14257dacad5SJay Sternberg struct nvme_command *sq_cmds; 14357dacad5SJay Sternberg struct nvme_command __iomem *sq_cmds_io; 14457dacad5SJay Sternberg volatile struct nvme_completion *cqes; 14557dacad5SJay Sternberg struct blk_mq_tags **tags; 14657dacad5SJay Sternberg dma_addr_t sq_dma_addr; 14757dacad5SJay Sternberg dma_addr_t cq_dma_addr; 14857dacad5SJay Sternberg u32 __iomem *q_db; 14957dacad5SJay Sternberg u16 q_depth; 15057dacad5SJay Sternberg s16 cq_vector; 15157dacad5SJay Sternberg u16 sq_head; 15257dacad5SJay Sternberg u16 sq_tail; 15357dacad5SJay Sternberg u16 cq_head; 15457dacad5SJay Sternberg u16 qid; 15557dacad5SJay Sternberg u8 cq_phase; 15657dacad5SJay Sternberg u8 cqe_seen; 15757dacad5SJay Sternberg }; 15857dacad5SJay Sternberg 15957dacad5SJay Sternberg /* 16071bd150cSChristoph Hellwig * The nvme_iod describes the data in an I/O, including the list of PRP 16171bd150cSChristoph Hellwig * entries. You can't see it in this data structure because C doesn't let 162f4800d6dSChristoph Hellwig * me express that. Use nvme_init_iod to ensure there's enough space 16371bd150cSChristoph Hellwig * allocated to store the PRP list. 16471bd150cSChristoph Hellwig */ 16571bd150cSChristoph Hellwig struct nvme_iod { 166f4800d6dSChristoph Hellwig struct nvme_queue *nvmeq; 167f4800d6dSChristoph Hellwig int aborted; 16871bd150cSChristoph Hellwig int npages; /* In the PRP list. 0 means small pool in use */ 16971bd150cSChristoph Hellwig int nents; /* Used in scatterlist */ 17071bd150cSChristoph Hellwig int length; /* Of data, in bytes */ 17171bd150cSChristoph Hellwig dma_addr_t first_dma; 172bf684057SChristoph Hellwig struct scatterlist meta_sg; /* metadata requires single contiguous buffer */ 173f4800d6dSChristoph Hellwig struct scatterlist *sg; 174f4800d6dSChristoph Hellwig struct scatterlist inline_sg[0]; 17557dacad5SJay Sternberg }; 17657dacad5SJay Sternberg 17757dacad5SJay Sternberg /* 17857dacad5SJay Sternberg * Check we didin't inadvertently grow the command struct 17957dacad5SJay Sternberg */ 18057dacad5SJay Sternberg static inline void _nvme_check_size(void) 18157dacad5SJay Sternberg { 18257dacad5SJay Sternberg BUILD_BUG_ON(sizeof(struct nvme_rw_command) != 64); 18357dacad5SJay Sternberg BUILD_BUG_ON(sizeof(struct nvme_create_cq) != 64); 18457dacad5SJay Sternberg BUILD_BUG_ON(sizeof(struct nvme_create_sq) != 64); 18557dacad5SJay Sternberg BUILD_BUG_ON(sizeof(struct nvme_delete_queue) != 64); 18657dacad5SJay Sternberg BUILD_BUG_ON(sizeof(struct nvme_features) != 64); 18757dacad5SJay Sternberg BUILD_BUG_ON(sizeof(struct nvme_format_cmd) != 64); 18857dacad5SJay Sternberg BUILD_BUG_ON(sizeof(struct nvme_abort_cmd) != 64); 18957dacad5SJay Sternberg BUILD_BUG_ON(sizeof(struct nvme_command) != 64); 19057dacad5SJay Sternberg BUILD_BUG_ON(sizeof(struct nvme_id_ctrl) != 4096); 19157dacad5SJay Sternberg BUILD_BUG_ON(sizeof(struct nvme_id_ns) != 4096); 19257dacad5SJay Sternberg BUILD_BUG_ON(sizeof(struct nvme_lba_range_type) != 64); 19357dacad5SJay Sternberg BUILD_BUG_ON(sizeof(struct nvme_smart_log) != 512); 19457dacad5SJay Sternberg } 19557dacad5SJay Sternberg 19657dacad5SJay Sternberg /* 19757dacad5SJay Sternberg * Max size of iod being embedded in the request payload 19857dacad5SJay Sternberg */ 19957dacad5SJay Sternberg #define NVME_INT_PAGES 2 2005fd4ce1bSChristoph Hellwig #define NVME_INT_BYTES(dev) (NVME_INT_PAGES * (dev)->ctrl.page_size) 20157dacad5SJay Sternberg 20257dacad5SJay Sternberg /* 20357dacad5SJay Sternberg * Will slightly overestimate the number of pages needed. This is OK 20457dacad5SJay Sternberg * as it only leads to a small amount of wasted memory for the lifetime of 20557dacad5SJay Sternberg * the I/O. 20657dacad5SJay Sternberg */ 20757dacad5SJay Sternberg static int nvme_npages(unsigned size, struct nvme_dev *dev) 20857dacad5SJay Sternberg { 2095fd4ce1bSChristoph Hellwig unsigned nprps = DIV_ROUND_UP(size + dev->ctrl.page_size, 2105fd4ce1bSChristoph Hellwig dev->ctrl.page_size); 21157dacad5SJay Sternberg return DIV_ROUND_UP(8 * nprps, PAGE_SIZE - 8); 21257dacad5SJay Sternberg } 21357dacad5SJay Sternberg 214f4800d6dSChristoph Hellwig static unsigned int nvme_iod_alloc_size(struct nvme_dev *dev, 215f4800d6dSChristoph Hellwig unsigned int size, unsigned int nseg) 216f4800d6dSChristoph Hellwig { 217f4800d6dSChristoph Hellwig return sizeof(__le64 *) * nvme_npages(size, dev) + 218f4800d6dSChristoph Hellwig sizeof(struct scatterlist) * nseg; 219f4800d6dSChristoph Hellwig } 220f4800d6dSChristoph Hellwig 22157dacad5SJay Sternberg static unsigned int nvme_cmd_size(struct nvme_dev *dev) 22257dacad5SJay Sternberg { 223f4800d6dSChristoph Hellwig return sizeof(struct nvme_iod) + 224f4800d6dSChristoph Hellwig nvme_iod_alloc_size(dev, NVME_INT_BYTES(dev), NVME_INT_PAGES); 22557dacad5SJay Sternberg } 22657dacad5SJay Sternberg 22757dacad5SJay Sternberg static int nvme_admin_init_hctx(struct blk_mq_hw_ctx *hctx, void *data, 22857dacad5SJay Sternberg unsigned int hctx_idx) 22957dacad5SJay Sternberg { 23057dacad5SJay Sternberg struct nvme_dev *dev = data; 23157dacad5SJay Sternberg struct nvme_queue *nvmeq = dev->queues[0]; 23257dacad5SJay Sternberg 23357dacad5SJay Sternberg WARN_ON(hctx_idx != 0); 23457dacad5SJay Sternberg WARN_ON(dev->admin_tagset.tags[0] != hctx->tags); 23557dacad5SJay Sternberg WARN_ON(nvmeq->tags); 23657dacad5SJay Sternberg 23757dacad5SJay Sternberg hctx->driver_data = nvmeq; 23857dacad5SJay Sternberg nvmeq->tags = &dev->admin_tagset.tags[0]; 23957dacad5SJay Sternberg return 0; 24057dacad5SJay Sternberg } 24157dacad5SJay Sternberg 24257dacad5SJay Sternberg static void nvme_admin_exit_hctx(struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx) 24357dacad5SJay Sternberg { 24457dacad5SJay Sternberg struct nvme_queue *nvmeq = hctx->driver_data; 24557dacad5SJay Sternberg 24657dacad5SJay Sternberg nvmeq->tags = NULL; 24757dacad5SJay Sternberg } 24857dacad5SJay Sternberg 24957dacad5SJay Sternberg static int nvme_admin_init_request(void *data, struct request *req, 25057dacad5SJay Sternberg unsigned int hctx_idx, unsigned int rq_idx, 25157dacad5SJay Sternberg unsigned int numa_node) 25257dacad5SJay Sternberg { 25357dacad5SJay Sternberg struct nvme_dev *dev = data; 254f4800d6dSChristoph Hellwig struct nvme_iod *iod = blk_mq_rq_to_pdu(req); 25557dacad5SJay Sternberg struct nvme_queue *nvmeq = dev->queues[0]; 25657dacad5SJay Sternberg 25757dacad5SJay Sternberg BUG_ON(!nvmeq); 258f4800d6dSChristoph Hellwig iod->nvmeq = nvmeq; 25957dacad5SJay Sternberg return 0; 26057dacad5SJay Sternberg } 26157dacad5SJay Sternberg 26257dacad5SJay Sternberg static int nvme_init_hctx(struct blk_mq_hw_ctx *hctx, void *data, 26357dacad5SJay Sternberg unsigned int hctx_idx) 26457dacad5SJay Sternberg { 26557dacad5SJay Sternberg struct nvme_dev *dev = data; 26657dacad5SJay Sternberg struct nvme_queue *nvmeq = dev->queues[hctx_idx + 1]; 26757dacad5SJay Sternberg 26857dacad5SJay Sternberg if (!nvmeq->tags) 26957dacad5SJay Sternberg nvmeq->tags = &dev->tagset.tags[hctx_idx]; 27057dacad5SJay Sternberg 27157dacad5SJay Sternberg WARN_ON(dev->tagset.tags[hctx_idx] != hctx->tags); 27257dacad5SJay Sternberg hctx->driver_data = nvmeq; 27357dacad5SJay Sternberg return 0; 27457dacad5SJay Sternberg } 27557dacad5SJay Sternberg 27657dacad5SJay Sternberg static int nvme_init_request(void *data, struct request *req, 27757dacad5SJay Sternberg unsigned int hctx_idx, unsigned int rq_idx, 27857dacad5SJay Sternberg unsigned int numa_node) 27957dacad5SJay Sternberg { 28057dacad5SJay Sternberg struct nvme_dev *dev = data; 281f4800d6dSChristoph Hellwig struct nvme_iod *iod = blk_mq_rq_to_pdu(req); 28257dacad5SJay Sternberg struct nvme_queue *nvmeq = dev->queues[hctx_idx + 1]; 28357dacad5SJay Sternberg 28457dacad5SJay Sternberg BUG_ON(!nvmeq); 285f4800d6dSChristoph Hellwig iod->nvmeq = nvmeq; 28657dacad5SJay Sternberg return 0; 28757dacad5SJay Sternberg } 28857dacad5SJay Sternberg 289adf68f21SChristoph Hellwig static void nvme_complete_async_event(struct nvme_dev *dev, 29057dacad5SJay Sternberg struct nvme_completion *cqe) 29157dacad5SJay Sternberg { 292adf68f21SChristoph Hellwig u16 status = le16_to_cpu(cqe->status) >> 1; 293adf68f21SChristoph Hellwig u32 result = le32_to_cpu(cqe->result); 29457dacad5SJay Sternberg 29557dacad5SJay Sternberg if (status == NVME_SC_SUCCESS || status == NVME_SC_ABORT_REQ) 296adf68f21SChristoph Hellwig ++dev->ctrl.event_limit; 29757dacad5SJay Sternberg if (status != NVME_SC_SUCCESS) 29857dacad5SJay Sternberg return; 29957dacad5SJay Sternberg 30057dacad5SJay Sternberg switch (result & 0xff07) { 30157dacad5SJay Sternberg case NVME_AER_NOTICE_NS_CHANGED: 302adf68f21SChristoph Hellwig dev_info(dev->dev, "rescanning\n"); 303adf68f21SChristoph Hellwig queue_work(nvme_workq, &dev->scan_work); 30457dacad5SJay Sternberg default: 305adf68f21SChristoph Hellwig dev_warn(dev->dev, "async event result %08x\n", result); 30657dacad5SJay Sternberg } 30757dacad5SJay Sternberg } 30857dacad5SJay Sternberg 30957dacad5SJay Sternberg /** 310adf68f21SChristoph Hellwig * __nvme_submit_cmd() - Copy a command into a queue and ring the doorbell 31157dacad5SJay Sternberg * @nvmeq: The queue to use 31257dacad5SJay Sternberg * @cmd: The command to send 31357dacad5SJay Sternberg * 31457dacad5SJay Sternberg * Safe to use from interrupt context 31557dacad5SJay Sternberg */ 31657dacad5SJay Sternberg static void __nvme_submit_cmd(struct nvme_queue *nvmeq, 31757dacad5SJay Sternberg struct nvme_command *cmd) 31857dacad5SJay Sternberg { 31957dacad5SJay Sternberg u16 tail = nvmeq->sq_tail; 32057dacad5SJay Sternberg 32157dacad5SJay Sternberg if (nvmeq->sq_cmds_io) 32257dacad5SJay Sternberg memcpy_toio(&nvmeq->sq_cmds_io[tail], cmd, sizeof(*cmd)); 32357dacad5SJay Sternberg else 32457dacad5SJay Sternberg memcpy(&nvmeq->sq_cmds[tail], cmd, sizeof(*cmd)); 32557dacad5SJay Sternberg 32657dacad5SJay Sternberg if (++tail == nvmeq->q_depth) 32757dacad5SJay Sternberg tail = 0; 32857dacad5SJay Sternberg writel(tail, nvmeq->q_db); 32957dacad5SJay Sternberg nvmeq->sq_tail = tail; 33057dacad5SJay Sternberg } 33157dacad5SJay Sternberg 332f4800d6dSChristoph Hellwig static __le64 **iod_list(struct request *req) 33357dacad5SJay Sternberg { 334f4800d6dSChristoph Hellwig struct nvme_iod *iod = blk_mq_rq_to_pdu(req); 335f4800d6dSChristoph Hellwig return (__le64 **)(iod->sg + req->nr_phys_segments); 33657dacad5SJay Sternberg } 33757dacad5SJay Sternberg 338f4800d6dSChristoph Hellwig static int nvme_init_iod(struct request *rq, struct nvme_dev *dev) 33957dacad5SJay Sternberg { 340f4800d6dSChristoph Hellwig struct nvme_iod *iod = blk_mq_rq_to_pdu(rq); 341f4800d6dSChristoph Hellwig int nseg = rq->nr_phys_segments; 342f4800d6dSChristoph Hellwig unsigned size; 343f4800d6dSChristoph Hellwig 344f4800d6dSChristoph Hellwig if (rq->cmd_flags & REQ_DISCARD) 345f4800d6dSChristoph Hellwig size = sizeof(struct nvme_dsm_range); 346f4800d6dSChristoph Hellwig else 347f4800d6dSChristoph Hellwig size = blk_rq_bytes(rq); 348f4800d6dSChristoph Hellwig 349f4800d6dSChristoph Hellwig if (nseg > NVME_INT_PAGES || size > NVME_INT_BYTES(dev)) { 350f4800d6dSChristoph Hellwig iod->sg = kmalloc(nvme_iod_alloc_size(dev, size, nseg), GFP_ATOMIC); 351f4800d6dSChristoph Hellwig if (!iod->sg) 352f4800d6dSChristoph Hellwig return BLK_MQ_RQ_QUEUE_BUSY; 353f4800d6dSChristoph Hellwig } else { 354f4800d6dSChristoph Hellwig iod->sg = iod->inline_sg; 35557dacad5SJay Sternberg } 35657dacad5SJay Sternberg 357f4800d6dSChristoph Hellwig iod->aborted = 0; 35857dacad5SJay Sternberg iod->npages = -1; 35957dacad5SJay Sternberg iod->nents = 0; 360f4800d6dSChristoph Hellwig iod->length = size; 361f4800d6dSChristoph Hellwig return 0; 36257dacad5SJay Sternberg } 36357dacad5SJay Sternberg 364f4800d6dSChristoph Hellwig static void nvme_free_iod(struct nvme_dev *dev, struct request *req) 36557dacad5SJay Sternberg { 366f4800d6dSChristoph Hellwig struct nvme_iod *iod = blk_mq_rq_to_pdu(req); 3675fd4ce1bSChristoph Hellwig const int last_prp = dev->ctrl.page_size / 8 - 1; 36857dacad5SJay Sternberg int i; 369f4800d6dSChristoph Hellwig __le64 **list = iod_list(req); 37057dacad5SJay Sternberg dma_addr_t prp_dma = iod->first_dma; 37157dacad5SJay Sternberg 37257dacad5SJay Sternberg if (iod->npages == 0) 37357dacad5SJay Sternberg dma_pool_free(dev->prp_small_pool, list[0], prp_dma); 37457dacad5SJay Sternberg for (i = 0; i < iod->npages; i++) { 37557dacad5SJay Sternberg __le64 *prp_list = list[i]; 37657dacad5SJay Sternberg dma_addr_t next_prp_dma = le64_to_cpu(prp_list[last_prp]); 37757dacad5SJay Sternberg dma_pool_free(dev->prp_page_pool, prp_list, prp_dma); 37857dacad5SJay Sternberg prp_dma = next_prp_dma; 37957dacad5SJay Sternberg } 38057dacad5SJay Sternberg 381f4800d6dSChristoph Hellwig if (iod->sg != iod->inline_sg) 382f4800d6dSChristoph Hellwig kfree(iod->sg); 38357dacad5SJay Sternberg } 38457dacad5SJay Sternberg 38557dacad5SJay Sternberg #ifdef CONFIG_BLK_DEV_INTEGRITY 38657dacad5SJay Sternberg static void nvme_dif_prep(u32 p, u32 v, struct t10_pi_tuple *pi) 38757dacad5SJay Sternberg { 38857dacad5SJay Sternberg if (be32_to_cpu(pi->ref_tag) == v) 38957dacad5SJay Sternberg pi->ref_tag = cpu_to_be32(p); 39057dacad5SJay Sternberg } 39157dacad5SJay Sternberg 39257dacad5SJay Sternberg static void nvme_dif_complete(u32 p, u32 v, struct t10_pi_tuple *pi) 39357dacad5SJay Sternberg { 39457dacad5SJay Sternberg if (be32_to_cpu(pi->ref_tag) == p) 39557dacad5SJay Sternberg pi->ref_tag = cpu_to_be32(v); 39657dacad5SJay Sternberg } 39757dacad5SJay Sternberg 39857dacad5SJay Sternberg /** 39957dacad5SJay Sternberg * nvme_dif_remap - remaps ref tags to bip seed and physical lba 40057dacad5SJay Sternberg * 40157dacad5SJay Sternberg * The virtual start sector is the one that was originally submitted by the 40257dacad5SJay Sternberg * block layer. Due to partitioning, MD/DM cloning, etc. the actual physical 40357dacad5SJay Sternberg * start sector may be different. Remap protection information to match the 40457dacad5SJay Sternberg * physical LBA on writes, and back to the original seed on reads. 40557dacad5SJay Sternberg * 40657dacad5SJay Sternberg * Type 0 and 3 do not have a ref tag, so no remapping required. 40757dacad5SJay Sternberg */ 40857dacad5SJay Sternberg static void nvme_dif_remap(struct request *req, 40957dacad5SJay Sternberg void (*dif_swap)(u32 p, u32 v, struct t10_pi_tuple *pi)) 41057dacad5SJay Sternberg { 41157dacad5SJay Sternberg struct nvme_ns *ns = req->rq_disk->private_data; 41257dacad5SJay Sternberg struct bio_integrity_payload *bip; 41357dacad5SJay Sternberg struct t10_pi_tuple *pi; 41457dacad5SJay Sternberg void *p, *pmap; 41557dacad5SJay Sternberg u32 i, nlb, ts, phys, virt; 41657dacad5SJay Sternberg 41757dacad5SJay Sternberg if (!ns->pi_type || ns->pi_type == NVME_NS_DPS_PI_TYPE3) 41857dacad5SJay Sternberg return; 41957dacad5SJay Sternberg 42057dacad5SJay Sternberg bip = bio_integrity(req->bio); 42157dacad5SJay Sternberg if (!bip) 42257dacad5SJay Sternberg return; 42357dacad5SJay Sternberg 42457dacad5SJay Sternberg pmap = kmap_atomic(bip->bip_vec->bv_page) + bip->bip_vec->bv_offset; 42557dacad5SJay Sternberg 42657dacad5SJay Sternberg p = pmap; 42757dacad5SJay Sternberg virt = bip_get_seed(bip); 42857dacad5SJay Sternberg phys = nvme_block_nr(ns, blk_rq_pos(req)); 42957dacad5SJay Sternberg nlb = (blk_rq_bytes(req) >> ns->lba_shift); 430ac6fc48cSDan Williams ts = ns->disk->queue->integrity.tuple_size; 43157dacad5SJay Sternberg 43257dacad5SJay Sternberg for (i = 0; i < nlb; i++, virt++, phys++) { 43357dacad5SJay Sternberg pi = (struct t10_pi_tuple *)p; 43457dacad5SJay Sternberg dif_swap(phys, virt, pi); 43557dacad5SJay Sternberg p += ts; 43657dacad5SJay Sternberg } 43757dacad5SJay Sternberg kunmap_atomic(pmap); 43857dacad5SJay Sternberg } 43957dacad5SJay Sternberg #else /* CONFIG_BLK_DEV_INTEGRITY */ 44057dacad5SJay Sternberg static void nvme_dif_remap(struct request *req, 44157dacad5SJay Sternberg void (*dif_swap)(u32 p, u32 v, struct t10_pi_tuple *pi)) 44257dacad5SJay Sternberg { 44357dacad5SJay Sternberg } 44457dacad5SJay Sternberg static void nvme_dif_prep(u32 p, u32 v, struct t10_pi_tuple *pi) 44557dacad5SJay Sternberg { 44657dacad5SJay Sternberg } 44757dacad5SJay Sternberg static void nvme_dif_complete(u32 p, u32 v, struct t10_pi_tuple *pi) 44857dacad5SJay Sternberg { 44957dacad5SJay Sternberg } 45057dacad5SJay Sternberg #endif 45157dacad5SJay Sternberg 452f4800d6dSChristoph Hellwig static bool nvme_setup_prps(struct nvme_dev *dev, struct request *req, 45369d2b571SChristoph Hellwig int total_len) 45457dacad5SJay Sternberg { 455f4800d6dSChristoph Hellwig struct nvme_iod *iod = blk_mq_rq_to_pdu(req); 45657dacad5SJay Sternberg struct dma_pool *pool; 45757dacad5SJay Sternberg int length = total_len; 45857dacad5SJay Sternberg struct scatterlist *sg = iod->sg; 45957dacad5SJay Sternberg int dma_len = sg_dma_len(sg); 46057dacad5SJay Sternberg u64 dma_addr = sg_dma_address(sg); 4615fd4ce1bSChristoph Hellwig u32 page_size = dev->ctrl.page_size; 46257dacad5SJay Sternberg int offset = dma_addr & (page_size - 1); 46357dacad5SJay Sternberg __le64 *prp_list; 464f4800d6dSChristoph Hellwig __le64 **list = iod_list(req); 46557dacad5SJay Sternberg dma_addr_t prp_dma; 46657dacad5SJay Sternberg int nprps, i; 46757dacad5SJay Sternberg 46857dacad5SJay Sternberg length -= (page_size - offset); 46957dacad5SJay Sternberg if (length <= 0) 47069d2b571SChristoph Hellwig return true; 47157dacad5SJay Sternberg 47257dacad5SJay Sternberg dma_len -= (page_size - offset); 47357dacad5SJay Sternberg if (dma_len) { 47457dacad5SJay Sternberg dma_addr += (page_size - offset); 47557dacad5SJay Sternberg } else { 47657dacad5SJay Sternberg sg = sg_next(sg); 47757dacad5SJay Sternberg dma_addr = sg_dma_address(sg); 47857dacad5SJay Sternberg dma_len = sg_dma_len(sg); 47957dacad5SJay Sternberg } 48057dacad5SJay Sternberg 48157dacad5SJay Sternberg if (length <= page_size) { 48257dacad5SJay Sternberg iod->first_dma = dma_addr; 48369d2b571SChristoph Hellwig return true; 48457dacad5SJay Sternberg } 48557dacad5SJay Sternberg 48657dacad5SJay Sternberg nprps = DIV_ROUND_UP(length, page_size); 48757dacad5SJay Sternberg if (nprps <= (256 / 8)) { 48857dacad5SJay Sternberg pool = dev->prp_small_pool; 48957dacad5SJay Sternberg iod->npages = 0; 49057dacad5SJay Sternberg } else { 49157dacad5SJay Sternberg pool = dev->prp_page_pool; 49257dacad5SJay Sternberg iod->npages = 1; 49357dacad5SJay Sternberg } 49457dacad5SJay Sternberg 49569d2b571SChristoph Hellwig prp_list = dma_pool_alloc(pool, GFP_ATOMIC, &prp_dma); 49657dacad5SJay Sternberg if (!prp_list) { 49757dacad5SJay Sternberg iod->first_dma = dma_addr; 49857dacad5SJay Sternberg iod->npages = -1; 49969d2b571SChristoph Hellwig return false; 50057dacad5SJay Sternberg } 50157dacad5SJay Sternberg list[0] = prp_list; 50257dacad5SJay Sternberg iod->first_dma = prp_dma; 50357dacad5SJay Sternberg i = 0; 50457dacad5SJay Sternberg for (;;) { 50557dacad5SJay Sternberg if (i == page_size >> 3) { 50657dacad5SJay Sternberg __le64 *old_prp_list = prp_list; 50769d2b571SChristoph Hellwig prp_list = dma_pool_alloc(pool, GFP_ATOMIC, &prp_dma); 50857dacad5SJay Sternberg if (!prp_list) 50969d2b571SChristoph Hellwig return false; 51057dacad5SJay Sternberg list[iod->npages++] = prp_list; 51157dacad5SJay Sternberg prp_list[0] = old_prp_list[i - 1]; 51257dacad5SJay Sternberg old_prp_list[i - 1] = cpu_to_le64(prp_dma); 51357dacad5SJay Sternberg i = 1; 51457dacad5SJay Sternberg } 51557dacad5SJay Sternberg prp_list[i++] = cpu_to_le64(dma_addr); 51657dacad5SJay Sternberg dma_len -= page_size; 51757dacad5SJay Sternberg dma_addr += page_size; 51857dacad5SJay Sternberg length -= page_size; 51957dacad5SJay Sternberg if (length <= 0) 52057dacad5SJay Sternberg break; 52157dacad5SJay Sternberg if (dma_len > 0) 52257dacad5SJay Sternberg continue; 52357dacad5SJay Sternberg BUG_ON(dma_len < 0); 52457dacad5SJay Sternberg sg = sg_next(sg); 52557dacad5SJay Sternberg dma_addr = sg_dma_address(sg); 52657dacad5SJay Sternberg dma_len = sg_dma_len(sg); 52757dacad5SJay Sternberg } 52857dacad5SJay Sternberg 52969d2b571SChristoph Hellwig return true; 53057dacad5SJay Sternberg } 53157dacad5SJay Sternberg 532f4800d6dSChristoph Hellwig static int nvme_map_data(struct nvme_dev *dev, struct request *req, 533ba1ca37eSChristoph Hellwig struct nvme_command *cmnd) 53457dacad5SJay Sternberg { 535f4800d6dSChristoph Hellwig struct nvme_iod *iod = blk_mq_rq_to_pdu(req); 536ba1ca37eSChristoph Hellwig struct request_queue *q = req->q; 537ba1ca37eSChristoph Hellwig enum dma_data_direction dma_dir = rq_data_dir(req) ? 538ba1ca37eSChristoph Hellwig DMA_TO_DEVICE : DMA_FROM_DEVICE; 539ba1ca37eSChristoph Hellwig int ret = BLK_MQ_RQ_QUEUE_ERROR; 54057dacad5SJay Sternberg 541ba1ca37eSChristoph Hellwig sg_init_table(iod->sg, req->nr_phys_segments); 542ba1ca37eSChristoph Hellwig iod->nents = blk_rq_map_sg(q, req, iod->sg); 543ba1ca37eSChristoph Hellwig if (!iod->nents) 544ba1ca37eSChristoph Hellwig goto out; 545ba1ca37eSChristoph Hellwig 546ba1ca37eSChristoph Hellwig ret = BLK_MQ_RQ_QUEUE_BUSY; 547ba1ca37eSChristoph Hellwig if (!dma_map_sg(dev->dev, iod->sg, iod->nents, dma_dir)) 548ba1ca37eSChristoph Hellwig goto out; 549ba1ca37eSChristoph Hellwig 550f4800d6dSChristoph Hellwig if (!nvme_setup_prps(dev, req, blk_rq_bytes(req))) 551ba1ca37eSChristoph Hellwig goto out_unmap; 552ba1ca37eSChristoph Hellwig 553ba1ca37eSChristoph Hellwig ret = BLK_MQ_RQ_QUEUE_ERROR; 554ba1ca37eSChristoph Hellwig if (blk_integrity_rq(req)) { 555ba1ca37eSChristoph Hellwig if (blk_rq_count_integrity_sg(q, req->bio) != 1) 556ba1ca37eSChristoph Hellwig goto out_unmap; 557ba1ca37eSChristoph Hellwig 558bf684057SChristoph Hellwig sg_init_table(&iod->meta_sg, 1); 559bf684057SChristoph Hellwig if (blk_rq_map_integrity_sg(q, req->bio, &iod->meta_sg) != 1) 560ba1ca37eSChristoph Hellwig goto out_unmap; 561ba1ca37eSChristoph Hellwig 562ba1ca37eSChristoph Hellwig if (rq_data_dir(req)) 563ba1ca37eSChristoph Hellwig nvme_dif_remap(req, nvme_dif_prep); 564ba1ca37eSChristoph Hellwig 565bf684057SChristoph Hellwig if (!dma_map_sg(dev->dev, &iod->meta_sg, 1, dma_dir)) 566ba1ca37eSChristoph Hellwig goto out_unmap; 56757dacad5SJay Sternberg } 56857dacad5SJay Sternberg 569ba1ca37eSChristoph Hellwig cmnd->rw.prp1 = cpu_to_le64(sg_dma_address(iod->sg)); 570ba1ca37eSChristoph Hellwig cmnd->rw.prp2 = cpu_to_le64(iod->first_dma); 571ba1ca37eSChristoph Hellwig if (blk_integrity_rq(req)) 572bf684057SChristoph Hellwig cmnd->rw.metadata = cpu_to_le64(sg_dma_address(&iod->meta_sg)); 573ba1ca37eSChristoph Hellwig return BLK_MQ_RQ_QUEUE_OK; 574ba1ca37eSChristoph Hellwig 575ba1ca37eSChristoph Hellwig out_unmap: 576ba1ca37eSChristoph Hellwig dma_unmap_sg(dev->dev, iod->sg, iod->nents, dma_dir); 577ba1ca37eSChristoph Hellwig out: 578ba1ca37eSChristoph Hellwig return ret; 57957dacad5SJay Sternberg } 58057dacad5SJay Sternberg 581f4800d6dSChristoph Hellwig static void nvme_unmap_data(struct nvme_dev *dev, struct request *req) 582d4f6c3abSChristoph Hellwig { 583f4800d6dSChristoph Hellwig struct nvme_iod *iod = blk_mq_rq_to_pdu(req); 584d4f6c3abSChristoph Hellwig enum dma_data_direction dma_dir = rq_data_dir(req) ? 585d4f6c3abSChristoph Hellwig DMA_TO_DEVICE : DMA_FROM_DEVICE; 586d4f6c3abSChristoph Hellwig 587d4f6c3abSChristoph Hellwig if (iod->nents) { 588d4f6c3abSChristoph Hellwig dma_unmap_sg(dev->dev, iod->sg, iod->nents, dma_dir); 589d4f6c3abSChristoph Hellwig if (blk_integrity_rq(req)) { 590d4f6c3abSChristoph Hellwig if (!rq_data_dir(req)) 591d4f6c3abSChristoph Hellwig nvme_dif_remap(req, nvme_dif_complete); 592bf684057SChristoph Hellwig dma_unmap_sg(dev->dev, &iod->meta_sg, 1, dma_dir); 593d4f6c3abSChristoph Hellwig } 594d4f6c3abSChristoph Hellwig } 595d4f6c3abSChristoph Hellwig 596f4800d6dSChristoph Hellwig nvme_free_iod(dev, req); 59757dacad5SJay Sternberg } 59857dacad5SJay Sternberg 59957dacad5SJay Sternberg /* 60057dacad5SJay Sternberg * We reuse the small pool to allocate the 16-byte range here as it is not 60157dacad5SJay Sternberg * worth having a special pool for these or additional cases to handle freeing 60257dacad5SJay Sternberg * the iod. 60357dacad5SJay Sternberg */ 604ba1ca37eSChristoph Hellwig static int nvme_setup_discard(struct nvme_queue *nvmeq, struct nvme_ns *ns, 605f4800d6dSChristoph Hellwig struct request *req, struct nvme_command *cmnd) 60657dacad5SJay Sternberg { 607f4800d6dSChristoph Hellwig struct nvme_iod *iod = blk_mq_rq_to_pdu(req); 608ba1ca37eSChristoph Hellwig struct nvme_dsm_range *range; 609ba1ca37eSChristoph Hellwig 610ba1ca37eSChristoph Hellwig range = dma_pool_alloc(nvmeq->dev->prp_small_pool, GFP_ATOMIC, 611ba1ca37eSChristoph Hellwig &iod->first_dma); 612ba1ca37eSChristoph Hellwig if (!range) 613ba1ca37eSChristoph Hellwig return BLK_MQ_RQ_QUEUE_BUSY; 614f4800d6dSChristoph Hellwig iod_list(req)[0] = (__le64 *)range; 615ba1ca37eSChristoph Hellwig iod->npages = 0; 61657dacad5SJay Sternberg 61757dacad5SJay Sternberg range->cattr = cpu_to_le32(0); 61857dacad5SJay Sternberg range->nlb = cpu_to_le32(blk_rq_bytes(req) >> ns->lba_shift); 61957dacad5SJay Sternberg range->slba = cpu_to_le64(nvme_block_nr(ns, blk_rq_pos(req))); 62057dacad5SJay Sternberg 621ba1ca37eSChristoph Hellwig memset(cmnd, 0, sizeof(*cmnd)); 622ba1ca37eSChristoph Hellwig cmnd->dsm.opcode = nvme_cmd_dsm; 623ba1ca37eSChristoph Hellwig cmnd->dsm.nsid = cpu_to_le32(ns->ns_id); 624ba1ca37eSChristoph Hellwig cmnd->dsm.prp1 = cpu_to_le64(iod->first_dma); 625ba1ca37eSChristoph Hellwig cmnd->dsm.nr = 0; 626ba1ca37eSChristoph Hellwig cmnd->dsm.attributes = cpu_to_le32(NVME_DSMGMT_AD); 627ba1ca37eSChristoph Hellwig return BLK_MQ_RQ_QUEUE_OK; 62857dacad5SJay Sternberg } 62957dacad5SJay Sternberg 63057dacad5SJay Sternberg /* 63157dacad5SJay Sternberg * NOTE: ns is NULL when called on the admin queue. 63257dacad5SJay Sternberg */ 63357dacad5SJay Sternberg static int nvme_queue_rq(struct blk_mq_hw_ctx *hctx, 63457dacad5SJay Sternberg const struct blk_mq_queue_data *bd) 63557dacad5SJay Sternberg { 63657dacad5SJay Sternberg struct nvme_ns *ns = hctx->queue->queuedata; 63757dacad5SJay Sternberg struct nvme_queue *nvmeq = hctx->driver_data; 63857dacad5SJay Sternberg struct nvme_dev *dev = nvmeq->dev; 63957dacad5SJay Sternberg struct request *req = bd->rq; 640ba1ca37eSChristoph Hellwig struct nvme_command cmnd; 641ba1ca37eSChristoph Hellwig int ret = BLK_MQ_RQ_QUEUE_OK; 64257dacad5SJay Sternberg 64357dacad5SJay Sternberg /* 64457dacad5SJay Sternberg * If formated with metadata, require the block layer provide a buffer 64557dacad5SJay Sternberg * unless this namespace is formated such that the metadata can be 64657dacad5SJay Sternberg * stripped/generated by the controller with PRACT=1. 64757dacad5SJay Sternberg */ 64857dacad5SJay Sternberg if (ns && ns->ms && !blk_integrity_rq(req)) { 64957dacad5SJay Sternberg if (!(ns->pi_type && ns->ms == 8) && 65057dacad5SJay Sternberg req->cmd_type != REQ_TYPE_DRV_PRIV) { 651eee417b0SChristoph Hellwig blk_mq_end_request(req, -EFAULT); 65257dacad5SJay Sternberg return BLK_MQ_RQ_QUEUE_OK; 65357dacad5SJay Sternberg } 65457dacad5SJay Sternberg } 65557dacad5SJay Sternberg 656f4800d6dSChristoph Hellwig ret = nvme_init_iod(req, dev); 657f4800d6dSChristoph Hellwig if (ret) 658f4800d6dSChristoph Hellwig return ret; 65957dacad5SJay Sternberg 66057dacad5SJay Sternberg if (req->cmd_flags & REQ_DISCARD) { 661f4800d6dSChristoph Hellwig ret = nvme_setup_discard(nvmeq, ns, req, &cmnd); 662ba1ca37eSChristoph Hellwig } else { 66357dacad5SJay Sternberg if (req->cmd_type == REQ_TYPE_DRV_PRIV) 664ba1ca37eSChristoph Hellwig memcpy(&cmnd, req->cmd, sizeof(cmnd)); 66557dacad5SJay Sternberg else if (req->cmd_flags & REQ_FLUSH) 666ba1ca37eSChristoph Hellwig nvme_setup_flush(ns, &cmnd); 66757dacad5SJay Sternberg else 668ba1ca37eSChristoph Hellwig nvme_setup_rw(ns, req, &cmnd); 66957dacad5SJay Sternberg 670ba1ca37eSChristoph Hellwig if (req->nr_phys_segments) 671f4800d6dSChristoph Hellwig ret = nvme_map_data(dev, req, &cmnd); 672ba1ca37eSChristoph Hellwig } 673ba1ca37eSChristoph Hellwig 674ba1ca37eSChristoph Hellwig if (ret) 675ba1ca37eSChristoph Hellwig goto out; 676ba1ca37eSChristoph Hellwig 677ba1ca37eSChristoph Hellwig cmnd.common.command_id = req->tag; 678aae239e1SChristoph Hellwig blk_mq_start_request(req); 679ba1ca37eSChristoph Hellwig 680ba1ca37eSChristoph Hellwig spin_lock_irq(&nvmeq->q_lock); 681ba1ca37eSChristoph Hellwig __nvme_submit_cmd(nvmeq, &cmnd); 68257dacad5SJay Sternberg nvme_process_cq(nvmeq); 68357dacad5SJay Sternberg spin_unlock_irq(&nvmeq->q_lock); 68457dacad5SJay Sternberg return BLK_MQ_RQ_QUEUE_OK; 685ba1ca37eSChristoph Hellwig out: 686f4800d6dSChristoph Hellwig nvme_free_iod(dev, req); 687ba1ca37eSChristoph Hellwig return ret; 68857dacad5SJay Sternberg } 68957dacad5SJay Sternberg 690eee417b0SChristoph Hellwig static void nvme_complete_rq(struct request *req) 691eee417b0SChristoph Hellwig { 692f4800d6dSChristoph Hellwig struct nvme_iod *iod = blk_mq_rq_to_pdu(req); 693f4800d6dSChristoph Hellwig struct nvme_dev *dev = iod->nvmeq->dev; 694eee417b0SChristoph Hellwig int error = 0; 695eee417b0SChristoph Hellwig 696f4800d6dSChristoph Hellwig nvme_unmap_data(dev, req); 697eee417b0SChristoph Hellwig 698eee417b0SChristoph Hellwig if (unlikely(req->errors)) { 699eee417b0SChristoph Hellwig if (nvme_req_needs_retry(req, req->errors)) { 700eee417b0SChristoph Hellwig nvme_requeue_req(req); 701eee417b0SChristoph Hellwig return; 702eee417b0SChristoph Hellwig } 703eee417b0SChristoph Hellwig 704eee417b0SChristoph Hellwig if (req->cmd_type == REQ_TYPE_DRV_PRIV) 705eee417b0SChristoph Hellwig error = req->errors; 706eee417b0SChristoph Hellwig else 707eee417b0SChristoph Hellwig error = nvme_error_status(req->errors); 708eee417b0SChristoph Hellwig } 709eee417b0SChristoph Hellwig 710f4800d6dSChristoph Hellwig if (unlikely(iod->aborted)) { 711eee417b0SChristoph Hellwig dev_warn(dev->dev, 712eee417b0SChristoph Hellwig "completing aborted command with status: %04x\n", 713eee417b0SChristoph Hellwig req->errors); 714eee417b0SChristoph Hellwig } 715eee417b0SChristoph Hellwig 716eee417b0SChristoph Hellwig blk_mq_end_request(req, error); 71757dacad5SJay Sternberg } 71857dacad5SJay Sternberg 719a0fa9647SJens Axboe static void __nvme_process_cq(struct nvme_queue *nvmeq, unsigned int *tag) 72057dacad5SJay Sternberg { 72157dacad5SJay Sternberg u16 head, phase; 72257dacad5SJay Sternberg 72357dacad5SJay Sternberg head = nvmeq->cq_head; 72457dacad5SJay Sternberg phase = nvmeq->cq_phase; 72557dacad5SJay Sternberg 72657dacad5SJay Sternberg for (;;) { 72757dacad5SJay Sternberg struct nvme_completion cqe = nvmeq->cqes[head]; 728adf68f21SChristoph Hellwig u16 status = le16_to_cpu(cqe.status); 729eee417b0SChristoph Hellwig struct request *req; 730adf68f21SChristoph Hellwig 731adf68f21SChristoph Hellwig if ((status & 1) != phase) 73257dacad5SJay Sternberg break; 73357dacad5SJay Sternberg nvmeq->sq_head = le16_to_cpu(cqe.sq_head); 73457dacad5SJay Sternberg if (++head == nvmeq->q_depth) { 73557dacad5SJay Sternberg head = 0; 73657dacad5SJay Sternberg phase = !phase; 73757dacad5SJay Sternberg } 738adf68f21SChristoph Hellwig 739a0fa9647SJens Axboe if (tag && *tag == cqe.command_id) 740a0fa9647SJens Axboe *tag = -1; 741adf68f21SChristoph Hellwig 742aae239e1SChristoph Hellwig if (unlikely(cqe.command_id >= nvmeq->q_depth)) { 743aae239e1SChristoph Hellwig dev_warn(nvmeq->q_dmadev, 744aae239e1SChristoph Hellwig "invalid id %d completed on queue %d\n", 745aae239e1SChristoph Hellwig cqe.command_id, le16_to_cpu(cqe.sq_id)); 746aae239e1SChristoph Hellwig continue; 747aae239e1SChristoph Hellwig } 748aae239e1SChristoph Hellwig 749adf68f21SChristoph Hellwig /* 750adf68f21SChristoph Hellwig * AEN requests are special as they don't time out and can 751adf68f21SChristoph Hellwig * survive any kind of queue freeze and often don't respond to 752adf68f21SChristoph Hellwig * aborts. We don't even bother to allocate a struct request 753adf68f21SChristoph Hellwig * for them but rather special case them here. 754adf68f21SChristoph Hellwig */ 755adf68f21SChristoph Hellwig if (unlikely(nvmeq->qid == 0 && 756adf68f21SChristoph Hellwig cqe.command_id >= NVME_AQ_BLKMQ_DEPTH)) { 757adf68f21SChristoph Hellwig nvme_complete_async_event(nvmeq->dev, &cqe); 758adf68f21SChristoph Hellwig continue; 759adf68f21SChristoph Hellwig } 760adf68f21SChristoph Hellwig 761eee417b0SChristoph Hellwig req = blk_mq_tag_to_rq(*nvmeq->tags, cqe.command_id); 762eee417b0SChristoph Hellwig if (req->cmd_type == REQ_TYPE_DRV_PRIV) { 763eee417b0SChristoph Hellwig u32 result = le32_to_cpu(cqe.result); 764eee417b0SChristoph Hellwig req->special = (void *)(uintptr_t)result; 765eee417b0SChristoph Hellwig } 766eee417b0SChristoph Hellwig blk_mq_complete_request(req, status >> 1); 767eee417b0SChristoph Hellwig 76857dacad5SJay Sternberg } 76957dacad5SJay Sternberg 77057dacad5SJay Sternberg /* If the controller ignores the cq head doorbell and continuously 77157dacad5SJay Sternberg * writes to the queue, it is theoretically possible to wrap around 77257dacad5SJay Sternberg * the queue twice and mistakenly return IRQ_NONE. Linux only 77357dacad5SJay Sternberg * requires that 0.1% of your interrupts are handled, so this isn't 77457dacad5SJay Sternberg * a big problem. 77557dacad5SJay Sternberg */ 77657dacad5SJay Sternberg if (head == nvmeq->cq_head && phase == nvmeq->cq_phase) 777a0fa9647SJens Axboe return; 77857dacad5SJay Sternberg 779604e8c8dSKeith Busch if (likely(nvmeq->cq_vector >= 0)) 78057dacad5SJay Sternberg writel(head, nvmeq->q_db + nvmeq->dev->db_stride); 78157dacad5SJay Sternberg nvmeq->cq_head = head; 78257dacad5SJay Sternberg nvmeq->cq_phase = phase; 78357dacad5SJay Sternberg 78457dacad5SJay Sternberg nvmeq->cqe_seen = 1; 785a0fa9647SJens Axboe } 786a0fa9647SJens Axboe 787a0fa9647SJens Axboe static void nvme_process_cq(struct nvme_queue *nvmeq) 788a0fa9647SJens Axboe { 789a0fa9647SJens Axboe __nvme_process_cq(nvmeq, NULL); 79057dacad5SJay Sternberg } 79157dacad5SJay Sternberg 79257dacad5SJay Sternberg static irqreturn_t nvme_irq(int irq, void *data) 79357dacad5SJay Sternberg { 79457dacad5SJay Sternberg irqreturn_t result; 79557dacad5SJay Sternberg struct nvme_queue *nvmeq = data; 79657dacad5SJay Sternberg spin_lock(&nvmeq->q_lock); 79757dacad5SJay Sternberg nvme_process_cq(nvmeq); 79857dacad5SJay Sternberg result = nvmeq->cqe_seen ? IRQ_HANDLED : IRQ_NONE; 79957dacad5SJay Sternberg nvmeq->cqe_seen = 0; 80057dacad5SJay Sternberg spin_unlock(&nvmeq->q_lock); 80157dacad5SJay Sternberg return result; 80257dacad5SJay Sternberg } 80357dacad5SJay Sternberg 80457dacad5SJay Sternberg static irqreturn_t nvme_irq_check(int irq, void *data) 80557dacad5SJay Sternberg { 80657dacad5SJay Sternberg struct nvme_queue *nvmeq = data; 80757dacad5SJay Sternberg struct nvme_completion cqe = nvmeq->cqes[nvmeq->cq_head]; 80857dacad5SJay Sternberg if ((le16_to_cpu(cqe.status) & 1) != nvmeq->cq_phase) 80957dacad5SJay Sternberg return IRQ_NONE; 81057dacad5SJay Sternberg return IRQ_WAKE_THREAD; 81157dacad5SJay Sternberg } 81257dacad5SJay Sternberg 813a0fa9647SJens Axboe static int nvme_poll(struct blk_mq_hw_ctx *hctx, unsigned int tag) 814a0fa9647SJens Axboe { 815a0fa9647SJens Axboe struct nvme_queue *nvmeq = hctx->driver_data; 816a0fa9647SJens Axboe 817a0fa9647SJens Axboe if ((le16_to_cpu(nvmeq->cqes[nvmeq->cq_head].status) & 1) == 818a0fa9647SJens Axboe nvmeq->cq_phase) { 819a0fa9647SJens Axboe spin_lock_irq(&nvmeq->q_lock); 820a0fa9647SJens Axboe __nvme_process_cq(nvmeq, &tag); 821a0fa9647SJens Axboe spin_unlock_irq(&nvmeq->q_lock); 822a0fa9647SJens Axboe 823a0fa9647SJens Axboe if (tag == -1) 824a0fa9647SJens Axboe return 1; 825a0fa9647SJens Axboe } 826a0fa9647SJens Axboe 827a0fa9647SJens Axboe return 0; 828a0fa9647SJens Axboe } 829a0fa9647SJens Axboe 830adf68f21SChristoph Hellwig static void nvme_submit_async_event(struct nvme_dev *dev) 83157dacad5SJay Sternberg { 83257dacad5SJay Sternberg struct nvme_command c; 83357dacad5SJay Sternberg 83457dacad5SJay Sternberg memset(&c, 0, sizeof(c)); 83557dacad5SJay Sternberg c.common.opcode = nvme_admin_async_event; 836adf68f21SChristoph Hellwig c.common.command_id = NVME_AQ_BLKMQ_DEPTH + --dev->ctrl.event_limit; 83757dacad5SJay Sternberg 838adf68f21SChristoph Hellwig __nvme_submit_cmd(dev->queues[0], &c); 83957dacad5SJay Sternberg } 84057dacad5SJay Sternberg 84157dacad5SJay Sternberg static int adapter_delete_queue(struct nvme_dev *dev, u8 opcode, u16 id) 84257dacad5SJay Sternberg { 84357dacad5SJay Sternberg struct nvme_command c; 84457dacad5SJay Sternberg 84557dacad5SJay Sternberg memset(&c, 0, sizeof(c)); 84657dacad5SJay Sternberg c.delete_queue.opcode = opcode; 84757dacad5SJay Sternberg c.delete_queue.qid = cpu_to_le16(id); 84857dacad5SJay Sternberg 8491c63dc66SChristoph Hellwig return nvme_submit_sync_cmd(dev->ctrl.admin_q, &c, NULL, 0); 85057dacad5SJay Sternberg } 85157dacad5SJay Sternberg 85257dacad5SJay Sternberg static int adapter_alloc_cq(struct nvme_dev *dev, u16 qid, 85357dacad5SJay Sternberg struct nvme_queue *nvmeq) 85457dacad5SJay Sternberg { 85557dacad5SJay Sternberg struct nvme_command c; 85657dacad5SJay Sternberg int flags = NVME_QUEUE_PHYS_CONTIG | NVME_CQ_IRQ_ENABLED; 85757dacad5SJay Sternberg 85857dacad5SJay Sternberg /* 85957dacad5SJay Sternberg * Note: we (ab)use the fact the the prp fields survive if no data 86057dacad5SJay Sternberg * is attached to the request. 86157dacad5SJay Sternberg */ 86257dacad5SJay Sternberg memset(&c, 0, sizeof(c)); 86357dacad5SJay Sternberg c.create_cq.opcode = nvme_admin_create_cq; 86457dacad5SJay Sternberg c.create_cq.prp1 = cpu_to_le64(nvmeq->cq_dma_addr); 86557dacad5SJay Sternberg c.create_cq.cqid = cpu_to_le16(qid); 86657dacad5SJay Sternberg c.create_cq.qsize = cpu_to_le16(nvmeq->q_depth - 1); 86757dacad5SJay Sternberg c.create_cq.cq_flags = cpu_to_le16(flags); 86857dacad5SJay Sternberg c.create_cq.irq_vector = cpu_to_le16(nvmeq->cq_vector); 86957dacad5SJay Sternberg 8701c63dc66SChristoph Hellwig return nvme_submit_sync_cmd(dev->ctrl.admin_q, &c, NULL, 0); 87157dacad5SJay Sternberg } 87257dacad5SJay Sternberg 87357dacad5SJay Sternberg static int adapter_alloc_sq(struct nvme_dev *dev, u16 qid, 87457dacad5SJay Sternberg struct nvme_queue *nvmeq) 87557dacad5SJay Sternberg { 87657dacad5SJay Sternberg struct nvme_command c; 87757dacad5SJay Sternberg int flags = NVME_QUEUE_PHYS_CONTIG | NVME_SQ_PRIO_MEDIUM; 87857dacad5SJay Sternberg 87957dacad5SJay Sternberg /* 88057dacad5SJay Sternberg * Note: we (ab)use the fact the the prp fields survive if no data 88157dacad5SJay Sternberg * is attached to the request. 88257dacad5SJay Sternberg */ 88357dacad5SJay Sternberg memset(&c, 0, sizeof(c)); 88457dacad5SJay Sternberg c.create_sq.opcode = nvme_admin_create_sq; 88557dacad5SJay Sternberg c.create_sq.prp1 = cpu_to_le64(nvmeq->sq_dma_addr); 88657dacad5SJay Sternberg c.create_sq.sqid = cpu_to_le16(qid); 88757dacad5SJay Sternberg c.create_sq.qsize = cpu_to_le16(nvmeq->q_depth - 1); 88857dacad5SJay Sternberg c.create_sq.sq_flags = cpu_to_le16(flags); 88957dacad5SJay Sternberg c.create_sq.cqid = cpu_to_le16(qid); 89057dacad5SJay Sternberg 8911c63dc66SChristoph Hellwig return nvme_submit_sync_cmd(dev->ctrl.admin_q, &c, NULL, 0); 89257dacad5SJay Sternberg } 89357dacad5SJay Sternberg 89457dacad5SJay Sternberg static int adapter_delete_cq(struct nvme_dev *dev, u16 cqid) 89557dacad5SJay Sternberg { 89657dacad5SJay Sternberg return adapter_delete_queue(dev, nvme_admin_delete_cq, cqid); 89757dacad5SJay Sternberg } 89857dacad5SJay Sternberg 89957dacad5SJay Sternberg static int adapter_delete_sq(struct nvme_dev *dev, u16 sqid) 90057dacad5SJay Sternberg { 90157dacad5SJay Sternberg return adapter_delete_queue(dev, nvme_admin_delete_sq, sqid); 90257dacad5SJay Sternberg } 90357dacad5SJay Sternberg 904e7a2a87dSChristoph Hellwig static void abort_endio(struct request *req, int error) 90557dacad5SJay Sternberg { 906f4800d6dSChristoph Hellwig struct nvme_iod *iod = blk_mq_rq_to_pdu(req); 907f4800d6dSChristoph Hellwig struct nvme_queue *nvmeq = iod->nvmeq; 908e7a2a87dSChristoph Hellwig u32 result = (u32)(uintptr_t)req->special; 909e7a2a87dSChristoph Hellwig u16 status = req->errors; 91057dacad5SJay Sternberg 911e7a2a87dSChristoph Hellwig dev_warn(nvmeq->q_dmadev, "Abort status:%x result:%x", status, result); 912e7a2a87dSChristoph Hellwig atomic_inc(&nvmeq->dev->ctrl.abort_limit); 91357dacad5SJay Sternberg 914e7a2a87dSChristoph Hellwig blk_mq_free_request(req); 91557dacad5SJay Sternberg } 91657dacad5SJay Sternberg 91731c7c7d2SChristoph Hellwig static enum blk_eh_timer_return nvme_timeout(struct request *req, bool reserved) 91857dacad5SJay Sternberg { 919f4800d6dSChristoph Hellwig struct nvme_iod *iod = blk_mq_rq_to_pdu(req); 920f4800d6dSChristoph Hellwig struct nvme_queue *nvmeq = iod->nvmeq; 92157dacad5SJay Sternberg struct nvme_dev *dev = nvmeq->dev; 92257dacad5SJay Sternberg struct request *abort_req; 92357dacad5SJay Sternberg struct nvme_command cmd; 92457dacad5SJay Sternberg 92531c7c7d2SChristoph Hellwig /* 926fd634f41SChristoph Hellwig * Shutdown immediately if controller times out while starting. The 927fd634f41SChristoph Hellwig * reset work will see the pci device disabled when it gets the forced 928fd634f41SChristoph Hellwig * cancellation error. All outstanding requests are completed on 929fd634f41SChristoph Hellwig * shutdown, so we return BLK_EH_HANDLED. 930fd634f41SChristoph Hellwig */ 931fd634f41SChristoph Hellwig if (test_bit(NVME_CTRL_RESETTING, &dev->flags)) { 932fd634f41SChristoph Hellwig dev_warn(dev->dev, 933fd634f41SChristoph Hellwig "I/O %d QID %d timeout, disable controller\n", 934fd634f41SChristoph Hellwig req->tag, nvmeq->qid); 935a5cdb68cSKeith Busch nvme_dev_disable(dev, false); 936fd634f41SChristoph Hellwig req->errors = NVME_SC_CANCELLED; 937fd634f41SChristoph Hellwig return BLK_EH_HANDLED; 938fd634f41SChristoph Hellwig } 939fd634f41SChristoph Hellwig 940fd634f41SChristoph Hellwig /* 941e1569a16SKeith Busch * Shutdown the controller immediately and schedule a reset if the 942e1569a16SKeith Busch * command was already aborted once before and still hasn't been 943e1569a16SKeith Busch * returned to the driver, or if this is the admin queue. 94431c7c7d2SChristoph Hellwig */ 945f4800d6dSChristoph Hellwig if (!nvmeq->qid || iod->aborted) { 94657dacad5SJay Sternberg dev_warn(dev->dev, 94757dacad5SJay Sternberg "I/O %d QID %d timeout, reset controller\n", 94857dacad5SJay Sternberg req->tag, nvmeq->qid); 949a5cdb68cSKeith Busch nvme_dev_disable(dev, false); 950e1569a16SKeith Busch queue_work(nvme_workq, &dev->reset_work); 951e1569a16SKeith Busch 952e1569a16SKeith Busch /* 953e1569a16SKeith Busch * Mark the request as handled, since the inline shutdown 954e1569a16SKeith Busch * forces all outstanding requests to complete. 955e1569a16SKeith Busch */ 956e1569a16SKeith Busch req->errors = NVME_SC_CANCELLED; 957e1569a16SKeith Busch return BLK_EH_HANDLED; 95857dacad5SJay Sternberg } 95957dacad5SJay Sternberg 960f4800d6dSChristoph Hellwig iod->aborted = 1; 96157dacad5SJay Sternberg 962e7a2a87dSChristoph Hellwig if (atomic_dec_return(&dev->ctrl.abort_limit) < 0) { 963e7a2a87dSChristoph Hellwig atomic_inc(&dev->ctrl.abort_limit); 964e7a2a87dSChristoph Hellwig return BLK_EH_RESET_TIMER; 965e7a2a87dSChristoph Hellwig } 96657dacad5SJay Sternberg 96757dacad5SJay Sternberg memset(&cmd, 0, sizeof(cmd)); 96857dacad5SJay Sternberg cmd.abort.opcode = nvme_admin_abort_cmd; 96957dacad5SJay Sternberg cmd.abort.cid = req->tag; 97057dacad5SJay Sternberg cmd.abort.sqid = cpu_to_le16(nvmeq->qid); 97157dacad5SJay Sternberg 972e7a2a87dSChristoph Hellwig dev_warn(nvmeq->q_dmadev, "I/O %d QID %d timeout, aborting\n", 97357dacad5SJay Sternberg req->tag, nvmeq->qid); 974e7a2a87dSChristoph Hellwig 975e7a2a87dSChristoph Hellwig abort_req = nvme_alloc_request(dev->ctrl.admin_q, &cmd, 9766f3b0e8bSChristoph Hellwig BLK_MQ_REQ_NOWAIT); 9776bf25d16SChristoph Hellwig if (IS_ERR(abort_req)) { 9786bf25d16SChristoph Hellwig atomic_inc(&dev->ctrl.abort_limit); 97931c7c7d2SChristoph Hellwig return BLK_EH_RESET_TIMER; 98057dacad5SJay Sternberg } 98157dacad5SJay Sternberg 982e7a2a87dSChristoph Hellwig abort_req->timeout = ADMIN_TIMEOUT; 983e7a2a87dSChristoph Hellwig abort_req->end_io_data = NULL; 984e7a2a87dSChristoph Hellwig blk_execute_rq_nowait(abort_req->q, NULL, abort_req, 0, abort_endio); 98557dacad5SJay Sternberg 98657dacad5SJay Sternberg /* 98757dacad5SJay Sternberg * The aborted req will be completed on receiving the abort req. 98857dacad5SJay Sternberg * We enable the timer again. If hit twice, it'll cause a device reset, 98957dacad5SJay Sternberg * as the device then is in a faulty state. 99057dacad5SJay Sternberg */ 99157dacad5SJay Sternberg return BLK_EH_RESET_TIMER; 99257dacad5SJay Sternberg } 99357dacad5SJay Sternberg 99457dacad5SJay Sternberg static void nvme_cancel_queue_ios(struct request *req, void *data, bool reserved) 99557dacad5SJay Sternberg { 99657dacad5SJay Sternberg struct nvme_queue *nvmeq = data; 997aae239e1SChristoph Hellwig int status; 99857dacad5SJay Sternberg 99957dacad5SJay Sternberg if (!blk_mq_request_started(req)) 100057dacad5SJay Sternberg return; 100157dacad5SJay Sternberg 1002aae239e1SChristoph Hellwig dev_warn(nvmeq->q_dmadev, 1003aae239e1SChristoph Hellwig "Cancelling I/O %d QID %d\n", req->tag, nvmeq->qid); 100457dacad5SJay Sternberg 10051d49c38cSKeith Busch status = NVME_SC_ABORT_REQ; 100657dacad5SJay Sternberg if (blk_queue_dying(req->q)) 1007aae239e1SChristoph Hellwig status |= NVME_SC_DNR; 1008aae239e1SChristoph Hellwig blk_mq_complete_request(req, status); 100957dacad5SJay Sternberg } 101057dacad5SJay Sternberg 101157dacad5SJay Sternberg static void nvme_free_queue(struct nvme_queue *nvmeq) 101257dacad5SJay Sternberg { 101357dacad5SJay Sternberg dma_free_coherent(nvmeq->q_dmadev, CQ_SIZE(nvmeq->q_depth), 101457dacad5SJay Sternberg (void *)nvmeq->cqes, nvmeq->cq_dma_addr); 101557dacad5SJay Sternberg if (nvmeq->sq_cmds) 101657dacad5SJay Sternberg dma_free_coherent(nvmeq->q_dmadev, SQ_SIZE(nvmeq->q_depth), 101757dacad5SJay Sternberg nvmeq->sq_cmds, nvmeq->sq_dma_addr); 101857dacad5SJay Sternberg kfree(nvmeq); 101957dacad5SJay Sternberg } 102057dacad5SJay Sternberg 102157dacad5SJay Sternberg static void nvme_free_queues(struct nvme_dev *dev, int lowest) 102257dacad5SJay Sternberg { 102357dacad5SJay Sternberg int i; 102457dacad5SJay Sternberg 102557dacad5SJay Sternberg for (i = dev->queue_count - 1; i >= lowest; i--) { 102657dacad5SJay Sternberg struct nvme_queue *nvmeq = dev->queues[i]; 102757dacad5SJay Sternberg dev->queue_count--; 102857dacad5SJay Sternberg dev->queues[i] = NULL; 102957dacad5SJay Sternberg nvme_free_queue(nvmeq); 103057dacad5SJay Sternberg } 103157dacad5SJay Sternberg } 103257dacad5SJay Sternberg 103357dacad5SJay Sternberg /** 103457dacad5SJay Sternberg * nvme_suspend_queue - put queue into suspended state 103557dacad5SJay Sternberg * @nvmeq - queue to suspend 103657dacad5SJay Sternberg */ 103757dacad5SJay Sternberg static int nvme_suspend_queue(struct nvme_queue *nvmeq) 103857dacad5SJay Sternberg { 103957dacad5SJay Sternberg int vector; 104057dacad5SJay Sternberg 104157dacad5SJay Sternberg spin_lock_irq(&nvmeq->q_lock); 104257dacad5SJay Sternberg if (nvmeq->cq_vector == -1) { 104357dacad5SJay Sternberg spin_unlock_irq(&nvmeq->q_lock); 104457dacad5SJay Sternberg return 1; 104557dacad5SJay Sternberg } 104657dacad5SJay Sternberg vector = nvmeq->dev->entry[nvmeq->cq_vector].vector; 104757dacad5SJay Sternberg nvmeq->dev->online_queues--; 104857dacad5SJay Sternberg nvmeq->cq_vector = -1; 104957dacad5SJay Sternberg spin_unlock_irq(&nvmeq->q_lock); 105057dacad5SJay Sternberg 10511c63dc66SChristoph Hellwig if (!nvmeq->qid && nvmeq->dev->ctrl.admin_q) 105225646264SKeith Busch blk_mq_stop_hw_queues(nvmeq->dev->ctrl.admin_q); 105357dacad5SJay Sternberg 105457dacad5SJay Sternberg irq_set_affinity_hint(vector, NULL); 105557dacad5SJay Sternberg free_irq(vector, nvmeq); 105657dacad5SJay Sternberg 105757dacad5SJay Sternberg return 0; 105857dacad5SJay Sternberg } 105957dacad5SJay Sternberg 106057dacad5SJay Sternberg static void nvme_clear_queue(struct nvme_queue *nvmeq) 106157dacad5SJay Sternberg { 106257dacad5SJay Sternberg spin_lock_irq(&nvmeq->q_lock); 106357dacad5SJay Sternberg if (nvmeq->tags && *nvmeq->tags) 106457dacad5SJay Sternberg blk_mq_all_tag_busy_iter(*nvmeq->tags, nvme_cancel_queue_ios, nvmeq); 106557dacad5SJay Sternberg spin_unlock_irq(&nvmeq->q_lock); 106657dacad5SJay Sternberg } 106757dacad5SJay Sternberg 1068a5cdb68cSKeith Busch static void nvme_disable_admin_queue(struct nvme_dev *dev, bool shutdown) 106957dacad5SJay Sternberg { 1070a5cdb68cSKeith Busch struct nvme_queue *nvmeq = dev->queues[0]; 107157dacad5SJay Sternberg 107257dacad5SJay Sternberg if (!nvmeq) 107357dacad5SJay Sternberg return; 107457dacad5SJay Sternberg if (nvme_suspend_queue(nvmeq)) 107557dacad5SJay Sternberg return; 107657dacad5SJay Sternberg 1077a5cdb68cSKeith Busch if (shutdown) 1078a5cdb68cSKeith Busch nvme_shutdown_ctrl(&dev->ctrl); 1079a5cdb68cSKeith Busch else 1080a5cdb68cSKeith Busch nvme_disable_ctrl(&dev->ctrl, lo_hi_readq( 1081a5cdb68cSKeith Busch dev->bar + NVME_REG_CAP)); 108257dacad5SJay Sternberg 108357dacad5SJay Sternberg spin_lock_irq(&nvmeq->q_lock); 108457dacad5SJay Sternberg nvme_process_cq(nvmeq); 108557dacad5SJay Sternberg spin_unlock_irq(&nvmeq->q_lock); 108657dacad5SJay Sternberg } 108757dacad5SJay Sternberg 108857dacad5SJay Sternberg static int nvme_cmb_qdepth(struct nvme_dev *dev, int nr_io_queues, 108957dacad5SJay Sternberg int entry_size) 109057dacad5SJay Sternberg { 109157dacad5SJay Sternberg int q_depth = dev->q_depth; 10925fd4ce1bSChristoph Hellwig unsigned q_size_aligned = roundup(q_depth * entry_size, 10935fd4ce1bSChristoph Hellwig dev->ctrl.page_size); 109457dacad5SJay Sternberg 109557dacad5SJay Sternberg if (q_size_aligned * nr_io_queues > dev->cmb_size) { 109657dacad5SJay Sternberg u64 mem_per_q = div_u64(dev->cmb_size, nr_io_queues); 10975fd4ce1bSChristoph Hellwig mem_per_q = round_down(mem_per_q, dev->ctrl.page_size); 109857dacad5SJay Sternberg q_depth = div_u64(mem_per_q, entry_size); 109957dacad5SJay Sternberg 110057dacad5SJay Sternberg /* 110157dacad5SJay Sternberg * Ensure the reduced q_depth is above some threshold where it 110257dacad5SJay Sternberg * would be better to map queues in system memory with the 110357dacad5SJay Sternberg * original depth 110457dacad5SJay Sternberg */ 110557dacad5SJay Sternberg if (q_depth < 64) 110657dacad5SJay Sternberg return -ENOMEM; 110757dacad5SJay Sternberg } 110857dacad5SJay Sternberg 110957dacad5SJay Sternberg return q_depth; 111057dacad5SJay Sternberg } 111157dacad5SJay Sternberg 111257dacad5SJay Sternberg static int nvme_alloc_sq_cmds(struct nvme_dev *dev, struct nvme_queue *nvmeq, 111357dacad5SJay Sternberg int qid, int depth) 111457dacad5SJay Sternberg { 111557dacad5SJay Sternberg if (qid && dev->cmb && use_cmb_sqes && NVME_CMB_SQS(dev->cmbsz)) { 11165fd4ce1bSChristoph Hellwig unsigned offset = (qid - 1) * roundup(SQ_SIZE(depth), 11175fd4ce1bSChristoph Hellwig dev->ctrl.page_size); 111857dacad5SJay Sternberg nvmeq->sq_dma_addr = dev->cmb_dma_addr + offset; 111957dacad5SJay Sternberg nvmeq->sq_cmds_io = dev->cmb + offset; 112057dacad5SJay Sternberg } else { 112157dacad5SJay Sternberg nvmeq->sq_cmds = dma_alloc_coherent(dev->dev, SQ_SIZE(depth), 112257dacad5SJay Sternberg &nvmeq->sq_dma_addr, GFP_KERNEL); 112357dacad5SJay Sternberg if (!nvmeq->sq_cmds) 112457dacad5SJay Sternberg return -ENOMEM; 112557dacad5SJay Sternberg } 112657dacad5SJay Sternberg 112757dacad5SJay Sternberg return 0; 112857dacad5SJay Sternberg } 112957dacad5SJay Sternberg 113057dacad5SJay Sternberg static struct nvme_queue *nvme_alloc_queue(struct nvme_dev *dev, int qid, 113157dacad5SJay Sternberg int depth) 113257dacad5SJay Sternberg { 113357dacad5SJay Sternberg struct nvme_queue *nvmeq = kzalloc(sizeof(*nvmeq), GFP_KERNEL); 113457dacad5SJay Sternberg if (!nvmeq) 113557dacad5SJay Sternberg return NULL; 113657dacad5SJay Sternberg 113757dacad5SJay Sternberg nvmeq->cqes = dma_zalloc_coherent(dev->dev, CQ_SIZE(depth), 113857dacad5SJay Sternberg &nvmeq->cq_dma_addr, GFP_KERNEL); 113957dacad5SJay Sternberg if (!nvmeq->cqes) 114057dacad5SJay Sternberg goto free_nvmeq; 114157dacad5SJay Sternberg 114257dacad5SJay Sternberg if (nvme_alloc_sq_cmds(dev, nvmeq, qid, depth)) 114357dacad5SJay Sternberg goto free_cqdma; 114457dacad5SJay Sternberg 114557dacad5SJay Sternberg nvmeq->q_dmadev = dev->dev; 114657dacad5SJay Sternberg nvmeq->dev = dev; 114757dacad5SJay Sternberg snprintf(nvmeq->irqname, sizeof(nvmeq->irqname), "nvme%dq%d", 11481c63dc66SChristoph Hellwig dev->ctrl.instance, qid); 114957dacad5SJay Sternberg spin_lock_init(&nvmeq->q_lock); 115057dacad5SJay Sternberg nvmeq->cq_head = 0; 115157dacad5SJay Sternberg nvmeq->cq_phase = 1; 115257dacad5SJay Sternberg nvmeq->q_db = &dev->dbs[qid * 2 * dev->db_stride]; 115357dacad5SJay Sternberg nvmeq->q_depth = depth; 115457dacad5SJay Sternberg nvmeq->qid = qid; 115557dacad5SJay Sternberg nvmeq->cq_vector = -1; 115657dacad5SJay Sternberg dev->queues[qid] = nvmeq; 115757dacad5SJay Sternberg 115857dacad5SJay Sternberg /* make sure queue descriptor is set before queue count, for kthread */ 115957dacad5SJay Sternberg mb(); 116057dacad5SJay Sternberg dev->queue_count++; 116157dacad5SJay Sternberg 116257dacad5SJay Sternberg return nvmeq; 116357dacad5SJay Sternberg 116457dacad5SJay Sternberg free_cqdma: 116557dacad5SJay Sternberg dma_free_coherent(dev->dev, CQ_SIZE(depth), (void *)nvmeq->cqes, 116657dacad5SJay Sternberg nvmeq->cq_dma_addr); 116757dacad5SJay Sternberg free_nvmeq: 116857dacad5SJay Sternberg kfree(nvmeq); 116957dacad5SJay Sternberg return NULL; 117057dacad5SJay Sternberg } 117157dacad5SJay Sternberg 117257dacad5SJay Sternberg static int queue_request_irq(struct nvme_dev *dev, struct nvme_queue *nvmeq, 117357dacad5SJay Sternberg const char *name) 117457dacad5SJay Sternberg { 117557dacad5SJay Sternberg if (use_threaded_interrupts) 117657dacad5SJay Sternberg return request_threaded_irq(dev->entry[nvmeq->cq_vector].vector, 117757dacad5SJay Sternberg nvme_irq_check, nvme_irq, IRQF_SHARED, 117857dacad5SJay Sternberg name, nvmeq); 117957dacad5SJay Sternberg return request_irq(dev->entry[nvmeq->cq_vector].vector, nvme_irq, 118057dacad5SJay Sternberg IRQF_SHARED, name, nvmeq); 118157dacad5SJay Sternberg } 118257dacad5SJay Sternberg 118357dacad5SJay Sternberg static void nvme_init_queue(struct nvme_queue *nvmeq, u16 qid) 118457dacad5SJay Sternberg { 118557dacad5SJay Sternberg struct nvme_dev *dev = nvmeq->dev; 118657dacad5SJay Sternberg 118757dacad5SJay Sternberg spin_lock_irq(&nvmeq->q_lock); 118857dacad5SJay Sternberg nvmeq->sq_tail = 0; 118957dacad5SJay Sternberg nvmeq->cq_head = 0; 119057dacad5SJay Sternberg nvmeq->cq_phase = 1; 119157dacad5SJay Sternberg nvmeq->q_db = &dev->dbs[qid * 2 * dev->db_stride]; 119257dacad5SJay Sternberg memset((void *)nvmeq->cqes, 0, CQ_SIZE(nvmeq->q_depth)); 119357dacad5SJay Sternberg dev->online_queues++; 119457dacad5SJay Sternberg spin_unlock_irq(&nvmeq->q_lock); 119557dacad5SJay Sternberg } 119657dacad5SJay Sternberg 119757dacad5SJay Sternberg static int nvme_create_queue(struct nvme_queue *nvmeq, int qid) 119857dacad5SJay Sternberg { 119957dacad5SJay Sternberg struct nvme_dev *dev = nvmeq->dev; 120057dacad5SJay Sternberg int result; 120157dacad5SJay Sternberg 120257dacad5SJay Sternberg nvmeq->cq_vector = qid - 1; 120357dacad5SJay Sternberg result = adapter_alloc_cq(dev, qid, nvmeq); 120457dacad5SJay Sternberg if (result < 0) 120557dacad5SJay Sternberg return result; 120657dacad5SJay Sternberg 120757dacad5SJay Sternberg result = adapter_alloc_sq(dev, qid, nvmeq); 120857dacad5SJay Sternberg if (result < 0) 120957dacad5SJay Sternberg goto release_cq; 121057dacad5SJay Sternberg 121157dacad5SJay Sternberg result = queue_request_irq(dev, nvmeq, nvmeq->irqname); 121257dacad5SJay Sternberg if (result < 0) 121357dacad5SJay Sternberg goto release_sq; 121457dacad5SJay Sternberg 121557dacad5SJay Sternberg nvme_init_queue(nvmeq, qid); 121657dacad5SJay Sternberg return result; 121757dacad5SJay Sternberg 121857dacad5SJay Sternberg release_sq: 121957dacad5SJay Sternberg adapter_delete_sq(dev, qid); 122057dacad5SJay Sternberg release_cq: 122157dacad5SJay Sternberg adapter_delete_cq(dev, qid); 122257dacad5SJay Sternberg return result; 122357dacad5SJay Sternberg } 122457dacad5SJay Sternberg 122557dacad5SJay Sternberg static struct blk_mq_ops nvme_mq_admin_ops = { 122657dacad5SJay Sternberg .queue_rq = nvme_queue_rq, 1227eee417b0SChristoph Hellwig .complete = nvme_complete_rq, 122857dacad5SJay Sternberg .map_queue = blk_mq_map_queue, 122957dacad5SJay Sternberg .init_hctx = nvme_admin_init_hctx, 123057dacad5SJay Sternberg .exit_hctx = nvme_admin_exit_hctx, 123157dacad5SJay Sternberg .init_request = nvme_admin_init_request, 123257dacad5SJay Sternberg .timeout = nvme_timeout, 123357dacad5SJay Sternberg }; 123457dacad5SJay Sternberg 123557dacad5SJay Sternberg static struct blk_mq_ops nvme_mq_ops = { 123657dacad5SJay Sternberg .queue_rq = nvme_queue_rq, 1237eee417b0SChristoph Hellwig .complete = nvme_complete_rq, 123857dacad5SJay Sternberg .map_queue = blk_mq_map_queue, 123957dacad5SJay Sternberg .init_hctx = nvme_init_hctx, 124057dacad5SJay Sternberg .init_request = nvme_init_request, 124157dacad5SJay Sternberg .timeout = nvme_timeout, 1242a0fa9647SJens Axboe .poll = nvme_poll, 124357dacad5SJay Sternberg }; 124457dacad5SJay Sternberg 124557dacad5SJay Sternberg static void nvme_dev_remove_admin(struct nvme_dev *dev) 124657dacad5SJay Sternberg { 12471c63dc66SChristoph Hellwig if (dev->ctrl.admin_q && !blk_queue_dying(dev->ctrl.admin_q)) { 12481c63dc66SChristoph Hellwig blk_cleanup_queue(dev->ctrl.admin_q); 124957dacad5SJay Sternberg blk_mq_free_tag_set(&dev->admin_tagset); 125057dacad5SJay Sternberg } 125157dacad5SJay Sternberg } 125257dacad5SJay Sternberg 125357dacad5SJay Sternberg static int nvme_alloc_admin_tags(struct nvme_dev *dev) 125457dacad5SJay Sternberg { 12551c63dc66SChristoph Hellwig if (!dev->ctrl.admin_q) { 125657dacad5SJay Sternberg dev->admin_tagset.ops = &nvme_mq_admin_ops; 125757dacad5SJay Sternberg dev->admin_tagset.nr_hw_queues = 1; 1258e3e9d50cSKeith Busch 1259e3e9d50cSKeith Busch /* 1260e3e9d50cSKeith Busch * Subtract one to leave an empty queue entry for 'Full Queue' 1261e3e9d50cSKeith Busch * condition. See NVM-Express 1.2 specification, section 4.1.2. 1262e3e9d50cSKeith Busch */ 1263e3e9d50cSKeith Busch dev->admin_tagset.queue_depth = NVME_AQ_BLKMQ_DEPTH - 1; 126457dacad5SJay Sternberg dev->admin_tagset.timeout = ADMIN_TIMEOUT; 126557dacad5SJay Sternberg dev->admin_tagset.numa_node = dev_to_node(dev->dev); 126657dacad5SJay Sternberg dev->admin_tagset.cmd_size = nvme_cmd_size(dev); 126757dacad5SJay Sternberg dev->admin_tagset.driver_data = dev; 126857dacad5SJay Sternberg 126957dacad5SJay Sternberg if (blk_mq_alloc_tag_set(&dev->admin_tagset)) 127057dacad5SJay Sternberg return -ENOMEM; 127157dacad5SJay Sternberg 12721c63dc66SChristoph Hellwig dev->ctrl.admin_q = blk_mq_init_queue(&dev->admin_tagset); 12731c63dc66SChristoph Hellwig if (IS_ERR(dev->ctrl.admin_q)) { 127457dacad5SJay Sternberg blk_mq_free_tag_set(&dev->admin_tagset); 127557dacad5SJay Sternberg return -ENOMEM; 127657dacad5SJay Sternberg } 12771c63dc66SChristoph Hellwig if (!blk_get_queue(dev->ctrl.admin_q)) { 127857dacad5SJay Sternberg nvme_dev_remove_admin(dev); 12791c63dc66SChristoph Hellwig dev->ctrl.admin_q = NULL; 128057dacad5SJay Sternberg return -ENODEV; 128157dacad5SJay Sternberg } 128257dacad5SJay Sternberg } else 128325646264SKeith Busch blk_mq_start_stopped_hw_queues(dev->ctrl.admin_q, true); 128457dacad5SJay Sternberg 128557dacad5SJay Sternberg return 0; 128657dacad5SJay Sternberg } 128757dacad5SJay Sternberg 128857dacad5SJay Sternberg static int nvme_configure_admin_queue(struct nvme_dev *dev) 128957dacad5SJay Sternberg { 129057dacad5SJay Sternberg int result; 129157dacad5SJay Sternberg u32 aqa; 12927a67cbeaSChristoph Hellwig u64 cap = lo_hi_readq(dev->bar + NVME_REG_CAP); 129357dacad5SJay Sternberg struct nvme_queue *nvmeq; 129457dacad5SJay Sternberg 12957a67cbeaSChristoph Hellwig dev->subsystem = readl(dev->bar + NVME_REG_VS) >= NVME_VS(1, 1) ? 129657dacad5SJay Sternberg NVME_CAP_NSSRC(cap) : 0; 129757dacad5SJay Sternberg 12987a67cbeaSChristoph Hellwig if (dev->subsystem && 12997a67cbeaSChristoph Hellwig (readl(dev->bar + NVME_REG_CSTS) & NVME_CSTS_NSSRO)) 13007a67cbeaSChristoph Hellwig writel(NVME_CSTS_NSSRO, dev->bar + NVME_REG_CSTS); 130157dacad5SJay Sternberg 13025fd4ce1bSChristoph Hellwig result = nvme_disable_ctrl(&dev->ctrl, cap); 130357dacad5SJay Sternberg if (result < 0) 130457dacad5SJay Sternberg return result; 130557dacad5SJay Sternberg 130657dacad5SJay Sternberg nvmeq = dev->queues[0]; 130757dacad5SJay Sternberg if (!nvmeq) { 130857dacad5SJay Sternberg nvmeq = nvme_alloc_queue(dev, 0, NVME_AQ_DEPTH); 130957dacad5SJay Sternberg if (!nvmeq) 131057dacad5SJay Sternberg return -ENOMEM; 131157dacad5SJay Sternberg } 131257dacad5SJay Sternberg 131357dacad5SJay Sternberg aqa = nvmeq->q_depth - 1; 131457dacad5SJay Sternberg aqa |= aqa << 16; 131557dacad5SJay Sternberg 13167a67cbeaSChristoph Hellwig writel(aqa, dev->bar + NVME_REG_AQA); 13177a67cbeaSChristoph Hellwig lo_hi_writeq(nvmeq->sq_dma_addr, dev->bar + NVME_REG_ASQ); 13187a67cbeaSChristoph Hellwig lo_hi_writeq(nvmeq->cq_dma_addr, dev->bar + NVME_REG_ACQ); 131957dacad5SJay Sternberg 13205fd4ce1bSChristoph Hellwig result = nvme_enable_ctrl(&dev->ctrl, cap); 132157dacad5SJay Sternberg if (result) 132257dacad5SJay Sternberg goto free_nvmeq; 132357dacad5SJay Sternberg 132457dacad5SJay Sternberg nvmeq->cq_vector = 0; 132557dacad5SJay Sternberg result = queue_request_irq(dev, nvmeq, nvmeq->irqname); 132657dacad5SJay Sternberg if (result) { 132757dacad5SJay Sternberg nvmeq->cq_vector = -1; 132857dacad5SJay Sternberg goto free_nvmeq; 132957dacad5SJay Sternberg } 133057dacad5SJay Sternberg 133157dacad5SJay Sternberg return result; 133257dacad5SJay Sternberg 133357dacad5SJay Sternberg free_nvmeq: 133457dacad5SJay Sternberg nvme_free_queues(dev, 0); 133557dacad5SJay Sternberg return result; 133657dacad5SJay Sternberg } 133757dacad5SJay Sternberg 133857dacad5SJay Sternberg static int nvme_kthread(void *data) 133957dacad5SJay Sternberg { 134057dacad5SJay Sternberg struct nvme_dev *dev, *next; 134157dacad5SJay Sternberg 134257dacad5SJay Sternberg while (!kthread_should_stop()) { 134357dacad5SJay Sternberg set_current_state(TASK_INTERRUPTIBLE); 134457dacad5SJay Sternberg spin_lock(&dev_list_lock); 134557dacad5SJay Sternberg list_for_each_entry_safe(dev, next, &dev_list, node) { 134657dacad5SJay Sternberg int i; 13477a67cbeaSChristoph Hellwig u32 csts = readl(dev->bar + NVME_REG_CSTS); 134857dacad5SJay Sternberg 1349846cc05fSChristoph Hellwig /* 1350846cc05fSChristoph Hellwig * Skip controllers currently under reset. 1351846cc05fSChristoph Hellwig */ 1352846cc05fSChristoph Hellwig if (work_pending(&dev->reset_work) || work_busy(&dev->reset_work)) 1353846cc05fSChristoph Hellwig continue; 135457dacad5SJay Sternberg 135557dacad5SJay Sternberg if ((dev->subsystem && (csts & NVME_CSTS_NSSRO)) || 135657dacad5SJay Sternberg csts & NVME_CSTS_CFS) { 1357846cc05fSChristoph Hellwig if (queue_work(nvme_workq, &dev->reset_work)) { 135857dacad5SJay Sternberg dev_warn(dev->dev, 135957dacad5SJay Sternberg "Failed status: %x, reset controller\n", 13607a67cbeaSChristoph Hellwig readl(dev->bar + NVME_REG_CSTS)); 136157dacad5SJay Sternberg } 136257dacad5SJay Sternberg continue; 136357dacad5SJay Sternberg } 136457dacad5SJay Sternberg for (i = 0; i < dev->queue_count; i++) { 136557dacad5SJay Sternberg struct nvme_queue *nvmeq = dev->queues[i]; 136657dacad5SJay Sternberg if (!nvmeq) 136757dacad5SJay Sternberg continue; 136857dacad5SJay Sternberg spin_lock_irq(&nvmeq->q_lock); 136957dacad5SJay Sternberg nvme_process_cq(nvmeq); 137057dacad5SJay Sternberg 1371adf68f21SChristoph Hellwig while (i == 0 && dev->ctrl.event_limit > 0) 1372adf68f21SChristoph Hellwig nvme_submit_async_event(dev); 137357dacad5SJay Sternberg spin_unlock_irq(&nvmeq->q_lock); 137457dacad5SJay Sternberg } 137557dacad5SJay Sternberg } 137657dacad5SJay Sternberg spin_unlock(&dev_list_lock); 137757dacad5SJay Sternberg schedule_timeout(round_jiffies_relative(HZ)); 137857dacad5SJay Sternberg } 137957dacad5SJay Sternberg return 0; 138057dacad5SJay Sternberg } 138157dacad5SJay Sternberg 1382749941f2SChristoph Hellwig static int nvme_create_io_queues(struct nvme_dev *dev) 138357dacad5SJay Sternberg { 1384949928c1SKeith Busch unsigned i, max; 1385749941f2SChristoph Hellwig int ret = 0; 138657dacad5SJay Sternberg 1387749941f2SChristoph Hellwig for (i = dev->queue_count; i <= dev->max_qid; i++) { 1388749941f2SChristoph Hellwig if (!nvme_alloc_queue(dev, i, dev->q_depth)) { 1389749941f2SChristoph Hellwig ret = -ENOMEM; 139057dacad5SJay Sternberg break; 1391749941f2SChristoph Hellwig } 1392749941f2SChristoph Hellwig } 139357dacad5SJay Sternberg 1394949928c1SKeith Busch max = min(dev->max_qid, dev->queue_count - 1); 1395949928c1SKeith Busch for (i = dev->online_queues; i <= max; i++) { 1396749941f2SChristoph Hellwig ret = nvme_create_queue(dev->queues[i], i); 1397749941f2SChristoph Hellwig if (ret) { 139857dacad5SJay Sternberg nvme_free_queues(dev, i); 139957dacad5SJay Sternberg break; 140057dacad5SJay Sternberg } 140157dacad5SJay Sternberg } 140257dacad5SJay Sternberg 1403749941f2SChristoph Hellwig /* 1404749941f2SChristoph Hellwig * Ignore failing Create SQ/CQ commands, we can continue with less 1405749941f2SChristoph Hellwig * than the desired aount of queues, and even a controller without 1406749941f2SChristoph Hellwig * I/O queues an still be used to issue admin commands. This might 1407749941f2SChristoph Hellwig * be useful to upgrade a buggy firmware for example. 1408749941f2SChristoph Hellwig */ 1409749941f2SChristoph Hellwig return ret >= 0 ? 0 : ret; 141057dacad5SJay Sternberg } 141157dacad5SJay Sternberg 141257dacad5SJay Sternberg static void __iomem *nvme_map_cmb(struct nvme_dev *dev) 141357dacad5SJay Sternberg { 141457dacad5SJay Sternberg u64 szu, size, offset; 141557dacad5SJay Sternberg u32 cmbloc; 141657dacad5SJay Sternberg resource_size_t bar_size; 141757dacad5SJay Sternberg struct pci_dev *pdev = to_pci_dev(dev->dev); 141857dacad5SJay Sternberg void __iomem *cmb; 141957dacad5SJay Sternberg dma_addr_t dma_addr; 142057dacad5SJay Sternberg 142157dacad5SJay Sternberg if (!use_cmb_sqes) 142257dacad5SJay Sternberg return NULL; 142357dacad5SJay Sternberg 14247a67cbeaSChristoph Hellwig dev->cmbsz = readl(dev->bar + NVME_REG_CMBSZ); 142557dacad5SJay Sternberg if (!(NVME_CMB_SZ(dev->cmbsz))) 142657dacad5SJay Sternberg return NULL; 142757dacad5SJay Sternberg 14287a67cbeaSChristoph Hellwig cmbloc = readl(dev->bar + NVME_REG_CMBLOC); 142957dacad5SJay Sternberg 143057dacad5SJay Sternberg szu = (u64)1 << (12 + 4 * NVME_CMB_SZU(dev->cmbsz)); 143157dacad5SJay Sternberg size = szu * NVME_CMB_SZ(dev->cmbsz); 143257dacad5SJay Sternberg offset = szu * NVME_CMB_OFST(cmbloc); 143357dacad5SJay Sternberg bar_size = pci_resource_len(pdev, NVME_CMB_BIR(cmbloc)); 143457dacad5SJay Sternberg 143557dacad5SJay Sternberg if (offset > bar_size) 143657dacad5SJay Sternberg return NULL; 143757dacad5SJay Sternberg 143857dacad5SJay Sternberg /* 143957dacad5SJay Sternberg * Controllers may support a CMB size larger than their BAR, 144057dacad5SJay Sternberg * for example, due to being behind a bridge. Reduce the CMB to 144157dacad5SJay Sternberg * the reported size of the BAR 144257dacad5SJay Sternberg */ 144357dacad5SJay Sternberg if (size > bar_size - offset) 144457dacad5SJay Sternberg size = bar_size - offset; 144557dacad5SJay Sternberg 144657dacad5SJay Sternberg dma_addr = pci_resource_start(pdev, NVME_CMB_BIR(cmbloc)) + offset; 144757dacad5SJay Sternberg cmb = ioremap_wc(dma_addr, size); 144857dacad5SJay Sternberg if (!cmb) 144957dacad5SJay Sternberg return NULL; 145057dacad5SJay Sternberg 145157dacad5SJay Sternberg dev->cmb_dma_addr = dma_addr; 145257dacad5SJay Sternberg dev->cmb_size = size; 145357dacad5SJay Sternberg return cmb; 145457dacad5SJay Sternberg } 145557dacad5SJay Sternberg 145657dacad5SJay Sternberg static inline void nvme_release_cmb(struct nvme_dev *dev) 145757dacad5SJay Sternberg { 145857dacad5SJay Sternberg if (dev->cmb) { 145957dacad5SJay Sternberg iounmap(dev->cmb); 146057dacad5SJay Sternberg dev->cmb = NULL; 146157dacad5SJay Sternberg } 146257dacad5SJay Sternberg } 146357dacad5SJay Sternberg 146457dacad5SJay Sternberg static size_t db_bar_size(struct nvme_dev *dev, unsigned nr_io_queues) 146557dacad5SJay Sternberg { 146657dacad5SJay Sternberg return 4096 + ((nr_io_queues + 1) * 8 * dev->db_stride); 146757dacad5SJay Sternberg } 146857dacad5SJay Sternberg 146957dacad5SJay Sternberg static int nvme_setup_io_queues(struct nvme_dev *dev) 147057dacad5SJay Sternberg { 147157dacad5SJay Sternberg struct nvme_queue *adminq = dev->queues[0]; 147257dacad5SJay Sternberg struct pci_dev *pdev = to_pci_dev(dev->dev); 147357dacad5SJay Sternberg int result, i, vecs, nr_io_queues, size; 147457dacad5SJay Sternberg 147557dacad5SJay Sternberg nr_io_queues = num_possible_cpus(); 14769a0be7abSChristoph Hellwig result = nvme_set_queue_count(&dev->ctrl, &nr_io_queues); 14779a0be7abSChristoph Hellwig if (result < 0) 147857dacad5SJay Sternberg return result; 14799a0be7abSChristoph Hellwig 14809a0be7abSChristoph Hellwig /* 14819a0be7abSChristoph Hellwig * Degraded controllers might return an error when setting the queue 14829a0be7abSChristoph Hellwig * count. We still want to be able to bring them online and offer 14839a0be7abSChristoph Hellwig * access to the admin queue, as that might be only way to fix them up. 14849a0be7abSChristoph Hellwig */ 14859a0be7abSChristoph Hellwig if (result > 0) { 14869a0be7abSChristoph Hellwig dev_err(dev->dev, "Could not set queue count (%d)\n", result); 14879a0be7abSChristoph Hellwig nr_io_queues = 0; 14889a0be7abSChristoph Hellwig result = 0; 14899a0be7abSChristoph Hellwig } 149057dacad5SJay Sternberg 149157dacad5SJay Sternberg if (dev->cmb && NVME_CMB_SQS(dev->cmbsz)) { 149257dacad5SJay Sternberg result = nvme_cmb_qdepth(dev, nr_io_queues, 149357dacad5SJay Sternberg sizeof(struct nvme_command)); 149457dacad5SJay Sternberg if (result > 0) 149557dacad5SJay Sternberg dev->q_depth = result; 149657dacad5SJay Sternberg else 149757dacad5SJay Sternberg nvme_release_cmb(dev); 149857dacad5SJay Sternberg } 149957dacad5SJay Sternberg 150057dacad5SJay Sternberg size = db_bar_size(dev, nr_io_queues); 150157dacad5SJay Sternberg if (size > 8192) { 150257dacad5SJay Sternberg iounmap(dev->bar); 150357dacad5SJay Sternberg do { 150457dacad5SJay Sternberg dev->bar = ioremap(pci_resource_start(pdev, 0), size); 150557dacad5SJay Sternberg if (dev->bar) 150657dacad5SJay Sternberg break; 150757dacad5SJay Sternberg if (!--nr_io_queues) 150857dacad5SJay Sternberg return -ENOMEM; 150957dacad5SJay Sternberg size = db_bar_size(dev, nr_io_queues); 151057dacad5SJay Sternberg } while (1); 15117a67cbeaSChristoph Hellwig dev->dbs = dev->bar + 4096; 151257dacad5SJay Sternberg adminq->q_db = dev->dbs; 151357dacad5SJay Sternberg } 151457dacad5SJay Sternberg 151557dacad5SJay Sternberg /* Deregister the admin queue's interrupt */ 151657dacad5SJay Sternberg free_irq(dev->entry[0].vector, adminq); 151757dacad5SJay Sternberg 151857dacad5SJay Sternberg /* 151957dacad5SJay Sternberg * If we enable msix early due to not intx, disable it again before 152057dacad5SJay Sternberg * setting up the full range we need. 152157dacad5SJay Sternberg */ 152257dacad5SJay Sternberg if (!pdev->irq) 152357dacad5SJay Sternberg pci_disable_msix(pdev); 152457dacad5SJay Sternberg 152557dacad5SJay Sternberg for (i = 0; i < nr_io_queues; i++) 152657dacad5SJay Sternberg dev->entry[i].entry = i; 152757dacad5SJay Sternberg vecs = pci_enable_msix_range(pdev, dev->entry, 1, nr_io_queues); 152857dacad5SJay Sternberg if (vecs < 0) { 152957dacad5SJay Sternberg vecs = pci_enable_msi_range(pdev, 1, min(nr_io_queues, 32)); 153057dacad5SJay Sternberg if (vecs < 0) { 153157dacad5SJay Sternberg vecs = 1; 153257dacad5SJay Sternberg } else { 153357dacad5SJay Sternberg for (i = 0; i < vecs; i++) 153457dacad5SJay Sternberg dev->entry[i].vector = i + pdev->irq; 153557dacad5SJay Sternberg } 153657dacad5SJay Sternberg } 153757dacad5SJay Sternberg 153857dacad5SJay Sternberg /* 153957dacad5SJay Sternberg * Should investigate if there's a performance win from allocating 154057dacad5SJay Sternberg * more queues than interrupt vectors; it might allow the submission 154157dacad5SJay Sternberg * path to scale better, even if the receive path is limited by the 154257dacad5SJay Sternberg * number of interrupts. 154357dacad5SJay Sternberg */ 154457dacad5SJay Sternberg nr_io_queues = vecs; 154557dacad5SJay Sternberg dev->max_qid = nr_io_queues; 154657dacad5SJay Sternberg 154757dacad5SJay Sternberg result = queue_request_irq(dev, adminq, adminq->irqname); 154857dacad5SJay Sternberg if (result) { 154957dacad5SJay Sternberg adminq->cq_vector = -1; 155057dacad5SJay Sternberg goto free_queues; 155157dacad5SJay Sternberg } 1552749941f2SChristoph Hellwig return nvme_create_io_queues(dev); 155357dacad5SJay Sternberg 155457dacad5SJay Sternberg free_queues: 155557dacad5SJay Sternberg nvme_free_queues(dev, 1); 155657dacad5SJay Sternberg return result; 155757dacad5SJay Sternberg } 155857dacad5SJay Sternberg 155957dacad5SJay Sternberg static void nvme_set_irq_hints(struct nvme_dev *dev) 156057dacad5SJay Sternberg { 156157dacad5SJay Sternberg struct nvme_queue *nvmeq; 156257dacad5SJay Sternberg int i; 156357dacad5SJay Sternberg 156457dacad5SJay Sternberg for (i = 0; i < dev->online_queues; i++) { 156557dacad5SJay Sternberg nvmeq = dev->queues[i]; 156657dacad5SJay Sternberg 156757dacad5SJay Sternberg if (!nvmeq->tags || !(*nvmeq->tags)) 156857dacad5SJay Sternberg continue; 156957dacad5SJay Sternberg 157057dacad5SJay Sternberg irq_set_affinity_hint(dev->entry[nvmeq->cq_vector].vector, 157157dacad5SJay Sternberg blk_mq_tags_cpumask(*nvmeq->tags)); 157257dacad5SJay Sternberg } 157357dacad5SJay Sternberg } 157457dacad5SJay Sternberg 157557dacad5SJay Sternberg static void nvme_dev_scan(struct work_struct *work) 157657dacad5SJay Sternberg { 157757dacad5SJay Sternberg struct nvme_dev *dev = container_of(work, struct nvme_dev, scan_work); 157857dacad5SJay Sternberg 157957dacad5SJay Sternberg if (!dev->tagset.tags) 158057dacad5SJay Sternberg return; 15815bae7f73SChristoph Hellwig nvme_scan_namespaces(&dev->ctrl); 158257dacad5SJay Sternberg nvme_set_irq_hints(dev); 158357dacad5SJay Sternberg } 158457dacad5SJay Sternberg 1585db3cbfffSKeith Busch static void nvme_del_queue_end(struct request *req, int error) 1586db3cbfffSKeith Busch { 1587db3cbfffSKeith Busch struct nvme_queue *nvmeq = req->end_io_data; 1588db3cbfffSKeith Busch 1589db3cbfffSKeith Busch blk_mq_free_request(req); 1590db3cbfffSKeith Busch complete(&nvmeq->dev->ioq_wait); 1591db3cbfffSKeith Busch } 1592db3cbfffSKeith Busch 1593db3cbfffSKeith Busch static void nvme_del_cq_end(struct request *req, int error) 1594db3cbfffSKeith Busch { 1595db3cbfffSKeith Busch struct nvme_queue *nvmeq = req->end_io_data; 1596db3cbfffSKeith Busch 1597db3cbfffSKeith Busch if (!error) { 1598db3cbfffSKeith Busch unsigned long flags; 1599db3cbfffSKeith Busch 1600db3cbfffSKeith Busch spin_lock_irqsave(&nvmeq->q_lock, flags); 1601db3cbfffSKeith Busch nvme_process_cq(nvmeq); 1602db3cbfffSKeith Busch spin_unlock_irqrestore(&nvmeq->q_lock, flags); 1603db3cbfffSKeith Busch } 1604db3cbfffSKeith Busch 1605db3cbfffSKeith Busch nvme_del_queue_end(req, error); 1606db3cbfffSKeith Busch } 1607db3cbfffSKeith Busch 1608db3cbfffSKeith Busch static int nvme_delete_queue(struct nvme_queue *nvmeq, u8 opcode) 1609db3cbfffSKeith Busch { 1610db3cbfffSKeith Busch struct request_queue *q = nvmeq->dev->ctrl.admin_q; 1611db3cbfffSKeith Busch struct request *req; 1612db3cbfffSKeith Busch struct nvme_command cmd; 1613db3cbfffSKeith Busch 1614db3cbfffSKeith Busch memset(&cmd, 0, sizeof(cmd)); 1615db3cbfffSKeith Busch cmd.delete_queue.opcode = opcode; 1616db3cbfffSKeith Busch cmd.delete_queue.qid = cpu_to_le16(nvmeq->qid); 1617db3cbfffSKeith Busch 1618db3cbfffSKeith Busch req = nvme_alloc_request(q, &cmd, BLK_MQ_REQ_NOWAIT); 1619db3cbfffSKeith Busch if (IS_ERR(req)) 1620db3cbfffSKeith Busch return PTR_ERR(req); 1621db3cbfffSKeith Busch 1622db3cbfffSKeith Busch req->timeout = ADMIN_TIMEOUT; 1623db3cbfffSKeith Busch req->end_io_data = nvmeq; 1624db3cbfffSKeith Busch 1625db3cbfffSKeith Busch blk_execute_rq_nowait(q, NULL, req, false, 1626db3cbfffSKeith Busch opcode == nvme_admin_delete_cq ? 1627db3cbfffSKeith Busch nvme_del_cq_end : nvme_del_queue_end); 1628db3cbfffSKeith Busch return 0; 1629db3cbfffSKeith Busch } 1630db3cbfffSKeith Busch 1631db3cbfffSKeith Busch static void nvme_disable_io_queues(struct nvme_dev *dev) 1632db3cbfffSKeith Busch { 1633db3cbfffSKeith Busch int pass; 1634db3cbfffSKeith Busch unsigned long timeout; 1635db3cbfffSKeith Busch u8 opcode = nvme_admin_delete_sq; 1636db3cbfffSKeith Busch 1637db3cbfffSKeith Busch for (pass = 0; pass < 2; pass++) { 1638db3cbfffSKeith Busch int sent = 0, i = dev->queue_count - 1; 1639db3cbfffSKeith Busch 1640db3cbfffSKeith Busch reinit_completion(&dev->ioq_wait); 1641db3cbfffSKeith Busch retry: 1642db3cbfffSKeith Busch timeout = ADMIN_TIMEOUT; 1643db3cbfffSKeith Busch for (; i > 0; i--) { 1644db3cbfffSKeith Busch struct nvme_queue *nvmeq = dev->queues[i]; 1645db3cbfffSKeith Busch 1646db3cbfffSKeith Busch if (!pass) 1647db3cbfffSKeith Busch nvme_suspend_queue(nvmeq); 1648db3cbfffSKeith Busch if (nvme_delete_queue(nvmeq, opcode)) 1649db3cbfffSKeith Busch break; 1650db3cbfffSKeith Busch ++sent; 1651db3cbfffSKeith Busch } 1652db3cbfffSKeith Busch while (sent--) { 1653db3cbfffSKeith Busch timeout = wait_for_completion_io_timeout(&dev->ioq_wait, timeout); 1654db3cbfffSKeith Busch if (timeout == 0) 1655db3cbfffSKeith Busch return; 1656db3cbfffSKeith Busch if (i) 1657db3cbfffSKeith Busch goto retry; 1658db3cbfffSKeith Busch } 1659db3cbfffSKeith Busch opcode = nvme_admin_delete_cq; 1660db3cbfffSKeith Busch } 1661db3cbfffSKeith Busch } 1662db3cbfffSKeith Busch 166357dacad5SJay Sternberg /* 166457dacad5SJay Sternberg * Return: error value if an error occurred setting up the queues or calling 166557dacad5SJay Sternberg * Identify Device. 0 if these succeeded, even if adding some of the 166657dacad5SJay Sternberg * namespaces failed. At the moment, these failures are silent. TBD which 166757dacad5SJay Sternberg * failures should be reported. 166857dacad5SJay Sternberg */ 166957dacad5SJay Sternberg static int nvme_dev_add(struct nvme_dev *dev) 167057dacad5SJay Sternberg { 16715bae7f73SChristoph Hellwig if (!dev->ctrl.tagset) { 167257dacad5SJay Sternberg dev->tagset.ops = &nvme_mq_ops; 167357dacad5SJay Sternberg dev->tagset.nr_hw_queues = dev->online_queues - 1; 167457dacad5SJay Sternberg dev->tagset.timeout = NVME_IO_TIMEOUT; 167557dacad5SJay Sternberg dev->tagset.numa_node = dev_to_node(dev->dev); 167657dacad5SJay Sternberg dev->tagset.queue_depth = 167757dacad5SJay Sternberg min_t(int, dev->q_depth, BLK_MQ_MAX_DEPTH) - 1; 167857dacad5SJay Sternberg dev->tagset.cmd_size = nvme_cmd_size(dev); 167957dacad5SJay Sternberg dev->tagset.flags = BLK_MQ_F_SHOULD_MERGE; 168057dacad5SJay Sternberg dev->tagset.driver_data = dev; 168157dacad5SJay Sternberg 168257dacad5SJay Sternberg if (blk_mq_alloc_tag_set(&dev->tagset)) 168357dacad5SJay Sternberg return 0; 16845bae7f73SChristoph Hellwig dev->ctrl.tagset = &dev->tagset; 1685949928c1SKeith Busch } else { 1686949928c1SKeith Busch blk_mq_update_nr_hw_queues(&dev->tagset, dev->online_queues - 1); 1687949928c1SKeith Busch 1688949928c1SKeith Busch /* Free previously allocated queues that are no longer usable */ 1689949928c1SKeith Busch nvme_free_queues(dev, dev->online_queues); 169057dacad5SJay Sternberg } 1691949928c1SKeith Busch 169292f7a162SKeith Busch queue_work(nvme_workq, &dev->scan_work); 169357dacad5SJay Sternberg return 0; 169457dacad5SJay Sternberg } 169557dacad5SJay Sternberg 169657dacad5SJay Sternberg static int nvme_dev_map(struct nvme_dev *dev) 169757dacad5SJay Sternberg { 169857dacad5SJay Sternberg u64 cap; 169957dacad5SJay Sternberg int bars, result = -ENOMEM; 170057dacad5SJay Sternberg struct pci_dev *pdev = to_pci_dev(dev->dev); 170157dacad5SJay Sternberg 170257dacad5SJay Sternberg if (pci_enable_device_mem(pdev)) 170357dacad5SJay Sternberg return result; 170457dacad5SJay Sternberg 170557dacad5SJay Sternberg dev->entry[0].vector = pdev->irq; 170657dacad5SJay Sternberg pci_set_master(pdev); 170757dacad5SJay Sternberg bars = pci_select_bars(pdev, IORESOURCE_MEM); 170857dacad5SJay Sternberg if (!bars) 170957dacad5SJay Sternberg goto disable_pci; 171057dacad5SJay Sternberg 171157dacad5SJay Sternberg if (pci_request_selected_regions(pdev, bars, "nvme")) 171257dacad5SJay Sternberg goto disable_pci; 171357dacad5SJay Sternberg 171457dacad5SJay Sternberg if (dma_set_mask_and_coherent(dev->dev, DMA_BIT_MASK(64)) && 171557dacad5SJay Sternberg dma_set_mask_and_coherent(dev->dev, DMA_BIT_MASK(32))) 171657dacad5SJay Sternberg goto disable; 171757dacad5SJay Sternberg 171857dacad5SJay Sternberg dev->bar = ioremap(pci_resource_start(pdev, 0), 8192); 171957dacad5SJay Sternberg if (!dev->bar) 172057dacad5SJay Sternberg goto disable; 172157dacad5SJay Sternberg 17227a67cbeaSChristoph Hellwig if (readl(dev->bar + NVME_REG_CSTS) == -1) { 172357dacad5SJay Sternberg result = -ENODEV; 172457dacad5SJay Sternberg goto unmap; 172557dacad5SJay Sternberg } 172657dacad5SJay Sternberg 172757dacad5SJay Sternberg /* 172857dacad5SJay Sternberg * Some devices don't advertse INTx interrupts, pre-enable a single 172957dacad5SJay Sternberg * MSIX vec for setup. We'll adjust this later. 173057dacad5SJay Sternberg */ 173157dacad5SJay Sternberg if (!pdev->irq) { 173257dacad5SJay Sternberg result = pci_enable_msix(pdev, dev->entry, 1); 173357dacad5SJay Sternberg if (result < 0) 173457dacad5SJay Sternberg goto unmap; 173557dacad5SJay Sternberg } 173657dacad5SJay Sternberg 17377a67cbeaSChristoph Hellwig cap = lo_hi_readq(dev->bar + NVME_REG_CAP); 17387a67cbeaSChristoph Hellwig 173957dacad5SJay Sternberg dev->q_depth = min_t(int, NVME_CAP_MQES(cap) + 1, NVME_Q_DEPTH); 174057dacad5SJay Sternberg dev->db_stride = 1 << NVME_CAP_STRIDE(cap); 17417a67cbeaSChristoph Hellwig dev->dbs = dev->bar + 4096; 17421f390c1fSStephan Günther 17431f390c1fSStephan Günther /* 17441f390c1fSStephan Günther * Temporary fix for the Apple controller found in the MacBook8,1 and 17451f390c1fSStephan Günther * some MacBook7,1 to avoid controller resets and data loss. 17461f390c1fSStephan Günther */ 17471f390c1fSStephan Günther if (pdev->vendor == PCI_VENDOR_ID_APPLE && pdev->device == 0x2001) { 17481f390c1fSStephan Günther dev->q_depth = 2; 17491f390c1fSStephan Günther dev_warn(dev->dev, "detected Apple NVMe controller, set " 17501f390c1fSStephan Günther "queue depth=%u to work around controller resets\n", 17511f390c1fSStephan Günther dev->q_depth); 17521f390c1fSStephan Günther } 17531f390c1fSStephan Günther 17547a67cbeaSChristoph Hellwig if (readl(dev->bar + NVME_REG_VS) >= NVME_VS(1, 2)) 175557dacad5SJay Sternberg dev->cmb = nvme_map_cmb(dev); 175657dacad5SJay Sternberg 1757a0a3408eSKeith Busch pci_enable_pcie_error_reporting(pdev); 1758a0a3408eSKeith Busch pci_save_state(pdev); 175957dacad5SJay Sternberg return 0; 176057dacad5SJay Sternberg 176157dacad5SJay Sternberg unmap: 176257dacad5SJay Sternberg iounmap(dev->bar); 176357dacad5SJay Sternberg dev->bar = NULL; 176457dacad5SJay Sternberg disable: 176557dacad5SJay Sternberg pci_release_regions(pdev); 176657dacad5SJay Sternberg disable_pci: 176757dacad5SJay Sternberg pci_disable_device(pdev); 176857dacad5SJay Sternberg return result; 176957dacad5SJay Sternberg } 177057dacad5SJay Sternberg 177157dacad5SJay Sternberg static void nvme_dev_unmap(struct nvme_dev *dev) 177257dacad5SJay Sternberg { 177357dacad5SJay Sternberg struct pci_dev *pdev = to_pci_dev(dev->dev); 177457dacad5SJay Sternberg 177557dacad5SJay Sternberg if (pdev->msi_enabled) 177657dacad5SJay Sternberg pci_disable_msi(pdev); 177757dacad5SJay Sternberg else if (pdev->msix_enabled) 177857dacad5SJay Sternberg pci_disable_msix(pdev); 177957dacad5SJay Sternberg 178057dacad5SJay Sternberg if (dev->bar) { 178157dacad5SJay Sternberg iounmap(dev->bar); 178257dacad5SJay Sternberg dev->bar = NULL; 178357dacad5SJay Sternberg pci_release_regions(pdev); 178457dacad5SJay Sternberg } 178557dacad5SJay Sternberg 1786a0a3408eSKeith Busch if (pci_is_enabled(pdev)) { 1787a0a3408eSKeith Busch pci_disable_pcie_error_reporting(pdev); 178857dacad5SJay Sternberg pci_disable_device(pdev); 178957dacad5SJay Sternberg } 1790a0a3408eSKeith Busch } 179157dacad5SJay Sternberg 17927385014cSChristoph Hellwig static int nvme_dev_list_add(struct nvme_dev *dev) 179357dacad5SJay Sternberg { 17947385014cSChristoph Hellwig bool start_thread = false; 179557dacad5SJay Sternberg 17967385014cSChristoph Hellwig spin_lock(&dev_list_lock); 17977385014cSChristoph Hellwig if (list_empty(&dev_list) && IS_ERR_OR_NULL(nvme_thread)) { 17987385014cSChristoph Hellwig start_thread = true; 17997385014cSChristoph Hellwig nvme_thread = NULL; 180057dacad5SJay Sternberg } 18017385014cSChristoph Hellwig list_add(&dev->node, &dev_list); 18027385014cSChristoph Hellwig spin_unlock(&dev_list_lock); 180357dacad5SJay Sternberg 18047385014cSChristoph Hellwig if (start_thread) { 18057385014cSChristoph Hellwig nvme_thread = kthread_run(nvme_kthread, NULL, "nvme"); 18067385014cSChristoph Hellwig wake_up_all(&nvme_kthread_wait); 18077385014cSChristoph Hellwig } else 18087385014cSChristoph Hellwig wait_event_killable(nvme_kthread_wait, nvme_thread); 180957dacad5SJay Sternberg 18107385014cSChristoph Hellwig if (IS_ERR_OR_NULL(nvme_thread)) 18117385014cSChristoph Hellwig return nvme_thread ? PTR_ERR(nvme_thread) : -EINTR; 181257dacad5SJay Sternberg 18137385014cSChristoph Hellwig return 0; 181457dacad5SJay Sternberg } 181557dacad5SJay Sternberg 181657dacad5SJay Sternberg /* 181757dacad5SJay Sternberg * Remove the node from the device list and check 181857dacad5SJay Sternberg * for whether or not we need to stop the nvme_thread. 181957dacad5SJay Sternberg */ 182057dacad5SJay Sternberg static void nvme_dev_list_remove(struct nvme_dev *dev) 182157dacad5SJay Sternberg { 182257dacad5SJay Sternberg struct task_struct *tmp = NULL; 182357dacad5SJay Sternberg 182457dacad5SJay Sternberg spin_lock(&dev_list_lock); 182557dacad5SJay Sternberg list_del_init(&dev->node); 182657dacad5SJay Sternberg if (list_empty(&dev_list) && !IS_ERR_OR_NULL(nvme_thread)) { 182757dacad5SJay Sternberg tmp = nvme_thread; 182857dacad5SJay Sternberg nvme_thread = NULL; 182957dacad5SJay Sternberg } 183057dacad5SJay Sternberg spin_unlock(&dev_list_lock); 183157dacad5SJay Sternberg 183257dacad5SJay Sternberg if (tmp) 183357dacad5SJay Sternberg kthread_stop(tmp); 183457dacad5SJay Sternberg } 183557dacad5SJay Sternberg 1836a5cdb68cSKeith Busch static void nvme_dev_disable(struct nvme_dev *dev, bool shutdown) 183757dacad5SJay Sternberg { 183857dacad5SJay Sternberg int i; 183957dacad5SJay Sternberg u32 csts = -1; 184057dacad5SJay Sternberg 184157dacad5SJay Sternberg nvme_dev_list_remove(dev); 184257dacad5SJay Sternberg 184377bf25eaSKeith Busch mutex_lock(&dev->shutdown_lock); 184457dacad5SJay Sternberg if (dev->bar) { 184525646264SKeith Busch nvme_stop_queues(&dev->ctrl); 18467a67cbeaSChristoph Hellwig csts = readl(dev->bar + NVME_REG_CSTS); 184757dacad5SJay Sternberg } 184857dacad5SJay Sternberg if (csts & NVME_CSTS_CFS || !(csts & NVME_CSTS_RDY)) { 184957dacad5SJay Sternberg for (i = dev->queue_count - 1; i >= 0; i--) { 185057dacad5SJay Sternberg struct nvme_queue *nvmeq = dev->queues[i]; 185157dacad5SJay Sternberg nvme_suspend_queue(nvmeq); 185257dacad5SJay Sternberg } 185357dacad5SJay Sternberg } else { 185457dacad5SJay Sternberg nvme_disable_io_queues(dev); 1855a5cdb68cSKeith Busch nvme_disable_admin_queue(dev, shutdown); 185657dacad5SJay Sternberg } 185757dacad5SJay Sternberg nvme_dev_unmap(dev); 185857dacad5SJay Sternberg 185957dacad5SJay Sternberg for (i = dev->queue_count - 1; i >= 0; i--) 186057dacad5SJay Sternberg nvme_clear_queue(dev->queues[i]); 186177bf25eaSKeith Busch mutex_unlock(&dev->shutdown_lock); 186257dacad5SJay Sternberg } 186357dacad5SJay Sternberg 186457dacad5SJay Sternberg static int nvme_setup_prp_pools(struct nvme_dev *dev) 186557dacad5SJay Sternberg { 186657dacad5SJay Sternberg dev->prp_page_pool = dma_pool_create("prp list page", dev->dev, 186757dacad5SJay Sternberg PAGE_SIZE, PAGE_SIZE, 0); 186857dacad5SJay Sternberg if (!dev->prp_page_pool) 186957dacad5SJay Sternberg return -ENOMEM; 187057dacad5SJay Sternberg 187157dacad5SJay Sternberg /* Optimisation for I/Os between 4k and 128k */ 187257dacad5SJay Sternberg dev->prp_small_pool = dma_pool_create("prp list 256", dev->dev, 187357dacad5SJay Sternberg 256, 256, 0); 187457dacad5SJay Sternberg if (!dev->prp_small_pool) { 187557dacad5SJay Sternberg dma_pool_destroy(dev->prp_page_pool); 187657dacad5SJay Sternberg return -ENOMEM; 187757dacad5SJay Sternberg } 187857dacad5SJay Sternberg return 0; 187957dacad5SJay Sternberg } 188057dacad5SJay Sternberg 188157dacad5SJay Sternberg static void nvme_release_prp_pools(struct nvme_dev *dev) 188257dacad5SJay Sternberg { 188357dacad5SJay Sternberg dma_pool_destroy(dev->prp_page_pool); 188457dacad5SJay Sternberg dma_pool_destroy(dev->prp_small_pool); 188557dacad5SJay Sternberg } 188657dacad5SJay Sternberg 18871673f1f0SChristoph Hellwig static void nvme_pci_free_ctrl(struct nvme_ctrl *ctrl) 188857dacad5SJay Sternberg { 18891673f1f0SChristoph Hellwig struct nvme_dev *dev = to_nvme_dev(ctrl); 189057dacad5SJay Sternberg 189157dacad5SJay Sternberg put_device(dev->dev); 189257dacad5SJay Sternberg if (dev->tagset.tags) 189357dacad5SJay Sternberg blk_mq_free_tag_set(&dev->tagset); 18941c63dc66SChristoph Hellwig if (dev->ctrl.admin_q) 18951c63dc66SChristoph Hellwig blk_put_queue(dev->ctrl.admin_q); 189657dacad5SJay Sternberg kfree(dev->queues); 189757dacad5SJay Sternberg kfree(dev->entry); 189857dacad5SJay Sternberg kfree(dev); 189957dacad5SJay Sternberg } 190057dacad5SJay Sternberg 1901fd634f41SChristoph Hellwig static void nvme_reset_work(struct work_struct *work) 190257dacad5SJay Sternberg { 1903fd634f41SChristoph Hellwig struct nvme_dev *dev = container_of(work, struct nvme_dev, reset_work); 190457dacad5SJay Sternberg int result; 190557dacad5SJay Sternberg 1906fd634f41SChristoph Hellwig if (WARN_ON(test_bit(NVME_CTRL_RESETTING, &dev->flags))) 1907fd634f41SChristoph Hellwig goto out; 1908fd634f41SChristoph Hellwig 1909fd634f41SChristoph Hellwig /* 1910fd634f41SChristoph Hellwig * If we're called to reset a live controller first shut it down before 1911fd634f41SChristoph Hellwig * moving on. 1912fd634f41SChristoph Hellwig */ 1913fd634f41SChristoph Hellwig if (dev->bar) 1914a5cdb68cSKeith Busch nvme_dev_disable(dev, false); 1915fd634f41SChristoph Hellwig 1916fd634f41SChristoph Hellwig set_bit(NVME_CTRL_RESETTING, &dev->flags); 1917fd634f41SChristoph Hellwig 191857dacad5SJay Sternberg result = nvme_dev_map(dev); 191957dacad5SJay Sternberg if (result) 192057dacad5SJay Sternberg goto out; 192157dacad5SJay Sternberg 192257dacad5SJay Sternberg result = nvme_configure_admin_queue(dev); 192357dacad5SJay Sternberg if (result) 192457dacad5SJay Sternberg goto unmap; 192557dacad5SJay Sternberg 192657dacad5SJay Sternberg nvme_init_queue(dev->queues[0], 0); 192757dacad5SJay Sternberg result = nvme_alloc_admin_tags(dev); 192857dacad5SJay Sternberg if (result) 192957dacad5SJay Sternberg goto disable; 193057dacad5SJay Sternberg 1931ce4541f4SChristoph Hellwig result = nvme_init_identify(&dev->ctrl); 1932ce4541f4SChristoph Hellwig if (result) 1933ce4541f4SChristoph Hellwig goto free_tags; 1934ce4541f4SChristoph Hellwig 193557dacad5SJay Sternberg result = nvme_setup_io_queues(dev); 193657dacad5SJay Sternberg if (result) 193757dacad5SJay Sternberg goto free_tags; 193857dacad5SJay Sternberg 1939adf68f21SChristoph Hellwig dev->ctrl.event_limit = NVME_NR_AEN_COMMANDS; 194057dacad5SJay Sternberg 19417385014cSChristoph Hellwig result = nvme_dev_list_add(dev); 19427385014cSChristoph Hellwig if (result) 19437385014cSChristoph Hellwig goto remove; 194457dacad5SJay Sternberg 194557dacad5SJay Sternberg /* 194657dacad5SJay Sternberg * Keep the controller around but remove all namespaces if we don't have 194757dacad5SJay Sternberg * any working I/O queue. 194857dacad5SJay Sternberg */ 194957dacad5SJay Sternberg if (dev->online_queues < 2) { 195057dacad5SJay Sternberg dev_warn(dev->dev, "IO queues not created\n"); 19515bae7f73SChristoph Hellwig nvme_remove_namespaces(&dev->ctrl); 195257dacad5SJay Sternberg } else { 195325646264SKeith Busch nvme_start_queues(&dev->ctrl); 195457dacad5SJay Sternberg nvme_dev_add(dev); 195557dacad5SJay Sternberg } 195657dacad5SJay Sternberg 1957fd634f41SChristoph Hellwig clear_bit(NVME_CTRL_RESETTING, &dev->flags); 195857dacad5SJay Sternberg return; 195957dacad5SJay Sternberg 19607385014cSChristoph Hellwig remove: 19617385014cSChristoph Hellwig nvme_dev_list_remove(dev); 196257dacad5SJay Sternberg free_tags: 196357dacad5SJay Sternberg nvme_dev_remove_admin(dev); 19641c63dc66SChristoph Hellwig blk_put_queue(dev->ctrl.admin_q); 19651c63dc66SChristoph Hellwig dev->ctrl.admin_q = NULL; 196657dacad5SJay Sternberg dev->queues[0]->tags = NULL; 196757dacad5SJay Sternberg disable: 1968a5cdb68cSKeith Busch nvme_disable_admin_queue(dev, false); 196957dacad5SJay Sternberg unmap: 197057dacad5SJay Sternberg nvme_dev_unmap(dev); 197157dacad5SJay Sternberg out: 19725c8809e6SChristoph Hellwig nvme_remove_dead_ctrl(dev); 197357dacad5SJay Sternberg } 197457dacad5SJay Sternberg 19755c8809e6SChristoph Hellwig static void nvme_remove_dead_ctrl_work(struct work_struct *work) 197657dacad5SJay Sternberg { 19775c8809e6SChristoph Hellwig struct nvme_dev *dev = container_of(work, struct nvme_dev, remove_work); 197857dacad5SJay Sternberg struct pci_dev *pdev = to_pci_dev(dev->dev); 197957dacad5SJay Sternberg 198057dacad5SJay Sternberg if (pci_get_drvdata(pdev)) 198157dacad5SJay Sternberg pci_stop_and_remove_bus_device_locked(pdev); 19821673f1f0SChristoph Hellwig nvme_put_ctrl(&dev->ctrl); 198357dacad5SJay Sternberg } 198457dacad5SJay Sternberg 19855c8809e6SChristoph Hellwig static void nvme_remove_dead_ctrl(struct nvme_dev *dev) 198657dacad5SJay Sternberg { 19875c8809e6SChristoph Hellwig dev_warn(dev->dev, "Removing after probe failure\n"); 19881673f1f0SChristoph Hellwig kref_get(&dev->ctrl.kref); 19895c8809e6SChristoph Hellwig if (!schedule_work(&dev->remove_work)) 19901673f1f0SChristoph Hellwig nvme_put_ctrl(&dev->ctrl); 199157dacad5SJay Sternberg } 199257dacad5SJay Sternberg 199357dacad5SJay Sternberg static int nvme_reset(struct nvme_dev *dev) 199457dacad5SJay Sternberg { 19951c63dc66SChristoph Hellwig if (!dev->ctrl.admin_q || blk_queue_dying(dev->ctrl.admin_q)) 199657dacad5SJay Sternberg return -ENODEV; 199757dacad5SJay Sternberg 1998846cc05fSChristoph Hellwig if (!queue_work(nvme_workq, &dev->reset_work)) 1999846cc05fSChristoph Hellwig return -EBUSY; 200057dacad5SJay Sternberg 200157dacad5SJay Sternberg flush_work(&dev->reset_work); 200257dacad5SJay Sternberg return 0; 200357dacad5SJay Sternberg } 200457dacad5SJay Sternberg 20051c63dc66SChristoph Hellwig static int nvme_pci_reg_read32(struct nvme_ctrl *ctrl, u32 off, u32 *val) 200657dacad5SJay Sternberg { 20071c63dc66SChristoph Hellwig *val = readl(to_nvme_dev(ctrl)->bar + off); 20081c63dc66SChristoph Hellwig return 0; 200957dacad5SJay Sternberg } 20101c63dc66SChristoph Hellwig 20115fd4ce1bSChristoph Hellwig static int nvme_pci_reg_write32(struct nvme_ctrl *ctrl, u32 off, u32 val) 20125fd4ce1bSChristoph Hellwig { 20135fd4ce1bSChristoph Hellwig writel(val, to_nvme_dev(ctrl)->bar + off); 20145fd4ce1bSChristoph Hellwig return 0; 20155fd4ce1bSChristoph Hellwig } 20165fd4ce1bSChristoph Hellwig 20177fd8930fSChristoph Hellwig static int nvme_pci_reg_read64(struct nvme_ctrl *ctrl, u32 off, u64 *val) 20187fd8930fSChristoph Hellwig { 20197fd8930fSChristoph Hellwig *val = readq(to_nvme_dev(ctrl)->bar + off); 20207fd8930fSChristoph Hellwig return 0; 20217fd8930fSChristoph Hellwig } 20227fd8930fSChristoph Hellwig 20235bae7f73SChristoph Hellwig static bool nvme_pci_io_incapable(struct nvme_ctrl *ctrl) 20245bae7f73SChristoph Hellwig { 20255bae7f73SChristoph Hellwig struct nvme_dev *dev = to_nvme_dev(ctrl); 20265bae7f73SChristoph Hellwig 20275bae7f73SChristoph Hellwig return !dev->bar || dev->online_queues < 2; 20285bae7f73SChristoph Hellwig } 20295bae7f73SChristoph Hellwig 2030f3ca80fcSChristoph Hellwig static int nvme_pci_reset_ctrl(struct nvme_ctrl *ctrl) 2031f3ca80fcSChristoph Hellwig { 2032f3ca80fcSChristoph Hellwig return nvme_reset(to_nvme_dev(ctrl)); 2033f3ca80fcSChristoph Hellwig } 2034f3ca80fcSChristoph Hellwig 20351c63dc66SChristoph Hellwig static const struct nvme_ctrl_ops nvme_pci_ctrl_ops = { 20361c63dc66SChristoph Hellwig .reg_read32 = nvme_pci_reg_read32, 20375fd4ce1bSChristoph Hellwig .reg_write32 = nvme_pci_reg_write32, 20387fd8930fSChristoph Hellwig .reg_read64 = nvme_pci_reg_read64, 20395bae7f73SChristoph Hellwig .io_incapable = nvme_pci_io_incapable, 2040f3ca80fcSChristoph Hellwig .reset_ctrl = nvme_pci_reset_ctrl, 20411673f1f0SChristoph Hellwig .free_ctrl = nvme_pci_free_ctrl, 20421c63dc66SChristoph Hellwig }; 204357dacad5SJay Sternberg 204457dacad5SJay Sternberg static int nvme_probe(struct pci_dev *pdev, const struct pci_device_id *id) 204557dacad5SJay Sternberg { 204657dacad5SJay Sternberg int node, result = -ENOMEM; 204757dacad5SJay Sternberg struct nvme_dev *dev; 204857dacad5SJay Sternberg 204957dacad5SJay Sternberg node = dev_to_node(&pdev->dev); 205057dacad5SJay Sternberg if (node == NUMA_NO_NODE) 205157dacad5SJay Sternberg set_dev_node(&pdev->dev, 0); 205257dacad5SJay Sternberg 205357dacad5SJay Sternberg dev = kzalloc_node(sizeof(*dev), GFP_KERNEL, node); 205457dacad5SJay Sternberg if (!dev) 205557dacad5SJay Sternberg return -ENOMEM; 205657dacad5SJay Sternberg dev->entry = kzalloc_node(num_possible_cpus() * sizeof(*dev->entry), 205757dacad5SJay Sternberg GFP_KERNEL, node); 205857dacad5SJay Sternberg if (!dev->entry) 205957dacad5SJay Sternberg goto free; 206057dacad5SJay Sternberg dev->queues = kzalloc_node((num_possible_cpus() + 1) * sizeof(void *), 206157dacad5SJay Sternberg GFP_KERNEL, node); 206257dacad5SJay Sternberg if (!dev->queues) 206357dacad5SJay Sternberg goto free; 206457dacad5SJay Sternberg 206557dacad5SJay Sternberg dev->dev = get_device(&pdev->dev); 206657dacad5SJay Sternberg pci_set_drvdata(pdev, dev); 206757dacad5SJay Sternberg 206857dacad5SJay Sternberg INIT_LIST_HEAD(&dev->node); 206957dacad5SJay Sternberg INIT_WORK(&dev->scan_work, nvme_dev_scan); 2070f3ca80fcSChristoph Hellwig INIT_WORK(&dev->reset_work, nvme_reset_work); 20715c8809e6SChristoph Hellwig INIT_WORK(&dev->remove_work, nvme_remove_dead_ctrl_work); 207277bf25eaSKeith Busch mutex_init(&dev->shutdown_lock); 2073db3cbfffSKeith Busch init_completion(&dev->ioq_wait); 2074f3ca80fcSChristoph Hellwig 2075f3ca80fcSChristoph Hellwig result = nvme_setup_prp_pools(dev); 2076f3ca80fcSChristoph Hellwig if (result) 2077f3ca80fcSChristoph Hellwig goto put_pci; 2078f3ca80fcSChristoph Hellwig 2079f3ca80fcSChristoph Hellwig result = nvme_init_ctrl(&dev->ctrl, &pdev->dev, &nvme_pci_ctrl_ops, 2080f3ca80fcSChristoph Hellwig id->driver_data); 2081f3ca80fcSChristoph Hellwig if (result) 2082f3ca80fcSChristoph Hellwig goto release_pools; 2083f3ca80fcSChristoph Hellwig 208492f7a162SKeith Busch queue_work(nvme_workq, &dev->reset_work); 208557dacad5SJay Sternberg return 0; 208657dacad5SJay Sternberg 208757dacad5SJay Sternberg release_pools: 208857dacad5SJay Sternberg nvme_release_prp_pools(dev); 208957dacad5SJay Sternberg put_pci: 209057dacad5SJay Sternberg put_device(dev->dev); 209157dacad5SJay Sternberg free: 209257dacad5SJay Sternberg kfree(dev->queues); 209357dacad5SJay Sternberg kfree(dev->entry); 209457dacad5SJay Sternberg kfree(dev); 209557dacad5SJay Sternberg return result; 209657dacad5SJay Sternberg } 209757dacad5SJay Sternberg 209857dacad5SJay Sternberg static void nvme_reset_notify(struct pci_dev *pdev, bool prepare) 209957dacad5SJay Sternberg { 210057dacad5SJay Sternberg struct nvme_dev *dev = pci_get_drvdata(pdev); 210157dacad5SJay Sternberg 210257dacad5SJay Sternberg if (prepare) 2103a5cdb68cSKeith Busch nvme_dev_disable(dev, false); 210457dacad5SJay Sternberg else 210592f7a162SKeith Busch queue_work(nvme_workq, &dev->reset_work); 210657dacad5SJay Sternberg } 210757dacad5SJay Sternberg 210857dacad5SJay Sternberg static void nvme_shutdown(struct pci_dev *pdev) 210957dacad5SJay Sternberg { 211057dacad5SJay Sternberg struct nvme_dev *dev = pci_get_drvdata(pdev); 2111a5cdb68cSKeith Busch nvme_dev_disable(dev, true); 211257dacad5SJay Sternberg } 211357dacad5SJay Sternberg 211457dacad5SJay Sternberg static void nvme_remove(struct pci_dev *pdev) 211557dacad5SJay Sternberg { 211657dacad5SJay Sternberg struct nvme_dev *dev = pci_get_drvdata(pdev); 211757dacad5SJay Sternberg 211857dacad5SJay Sternberg spin_lock(&dev_list_lock); 211957dacad5SJay Sternberg list_del_init(&dev->node); 212057dacad5SJay Sternberg spin_unlock(&dev_list_lock); 212157dacad5SJay Sternberg 212257dacad5SJay Sternberg pci_set_drvdata(pdev, NULL); 212357dacad5SJay Sternberg flush_work(&dev->reset_work); 212457dacad5SJay Sternberg flush_work(&dev->scan_work); 21255bae7f73SChristoph Hellwig nvme_remove_namespaces(&dev->ctrl); 212653029b04SKeith Busch nvme_uninit_ctrl(&dev->ctrl); 2127a5cdb68cSKeith Busch nvme_dev_disable(dev, true); 212857dacad5SJay Sternberg nvme_dev_remove_admin(dev); 212957dacad5SJay Sternberg nvme_free_queues(dev, 0); 213057dacad5SJay Sternberg nvme_release_cmb(dev); 213157dacad5SJay Sternberg nvme_release_prp_pools(dev); 21321673f1f0SChristoph Hellwig nvme_put_ctrl(&dev->ctrl); 213357dacad5SJay Sternberg } 213457dacad5SJay Sternberg 213557dacad5SJay Sternberg #ifdef CONFIG_PM_SLEEP 213657dacad5SJay Sternberg static int nvme_suspend(struct device *dev) 213757dacad5SJay Sternberg { 213857dacad5SJay Sternberg struct pci_dev *pdev = to_pci_dev(dev); 213957dacad5SJay Sternberg struct nvme_dev *ndev = pci_get_drvdata(pdev); 214057dacad5SJay Sternberg 2141a5cdb68cSKeith Busch nvme_dev_disable(ndev, true); 214257dacad5SJay Sternberg return 0; 214357dacad5SJay Sternberg } 214457dacad5SJay Sternberg 214557dacad5SJay Sternberg static int nvme_resume(struct device *dev) 214657dacad5SJay Sternberg { 214757dacad5SJay Sternberg struct pci_dev *pdev = to_pci_dev(dev); 214857dacad5SJay Sternberg struct nvme_dev *ndev = pci_get_drvdata(pdev); 214957dacad5SJay Sternberg 215092f7a162SKeith Busch queue_work(nvme_workq, &ndev->reset_work); 215157dacad5SJay Sternberg return 0; 215257dacad5SJay Sternberg } 215357dacad5SJay Sternberg #endif 215457dacad5SJay Sternberg 215557dacad5SJay Sternberg static SIMPLE_DEV_PM_OPS(nvme_dev_pm_ops, nvme_suspend, nvme_resume); 215657dacad5SJay Sternberg 2157a0a3408eSKeith Busch static pci_ers_result_t nvme_error_detected(struct pci_dev *pdev, 2158a0a3408eSKeith Busch pci_channel_state_t state) 2159a0a3408eSKeith Busch { 2160a0a3408eSKeith Busch struct nvme_dev *dev = pci_get_drvdata(pdev); 2161a0a3408eSKeith Busch 2162a0a3408eSKeith Busch /* 2163a0a3408eSKeith Busch * A frozen channel requires a reset. When detected, this method will 2164a0a3408eSKeith Busch * shutdown the controller to quiesce. The controller will be restarted 2165a0a3408eSKeith Busch * after the slot reset through driver's slot_reset callback. 2166a0a3408eSKeith Busch */ 2167a0a3408eSKeith Busch dev_warn(&pdev->dev, "error detected: state:%d\n", state); 2168a0a3408eSKeith Busch switch (state) { 2169a0a3408eSKeith Busch case pci_channel_io_normal: 2170a0a3408eSKeith Busch return PCI_ERS_RESULT_CAN_RECOVER; 2171a0a3408eSKeith Busch case pci_channel_io_frozen: 2172a5cdb68cSKeith Busch nvme_dev_disable(dev, false); 2173a0a3408eSKeith Busch return PCI_ERS_RESULT_NEED_RESET; 2174a0a3408eSKeith Busch case pci_channel_io_perm_failure: 2175a0a3408eSKeith Busch return PCI_ERS_RESULT_DISCONNECT; 2176a0a3408eSKeith Busch } 2177a0a3408eSKeith Busch return PCI_ERS_RESULT_NEED_RESET; 2178a0a3408eSKeith Busch } 2179a0a3408eSKeith Busch 2180a0a3408eSKeith Busch static pci_ers_result_t nvme_slot_reset(struct pci_dev *pdev) 2181a0a3408eSKeith Busch { 2182a0a3408eSKeith Busch struct nvme_dev *dev = pci_get_drvdata(pdev); 2183a0a3408eSKeith Busch 2184a0a3408eSKeith Busch dev_info(&pdev->dev, "restart after slot reset\n"); 2185a0a3408eSKeith Busch pci_restore_state(pdev); 2186a0a3408eSKeith Busch queue_work(nvme_workq, &dev->reset_work); 2187a0a3408eSKeith Busch return PCI_ERS_RESULT_RECOVERED; 2188a0a3408eSKeith Busch } 2189a0a3408eSKeith Busch 2190a0a3408eSKeith Busch static void nvme_error_resume(struct pci_dev *pdev) 2191a0a3408eSKeith Busch { 2192a0a3408eSKeith Busch pci_cleanup_aer_uncorrect_error_status(pdev); 2193a0a3408eSKeith Busch } 2194a0a3408eSKeith Busch 219557dacad5SJay Sternberg static const struct pci_error_handlers nvme_err_handler = { 219657dacad5SJay Sternberg .error_detected = nvme_error_detected, 219757dacad5SJay Sternberg .slot_reset = nvme_slot_reset, 219857dacad5SJay Sternberg .resume = nvme_error_resume, 219957dacad5SJay Sternberg .reset_notify = nvme_reset_notify, 220057dacad5SJay Sternberg }; 220157dacad5SJay Sternberg 220257dacad5SJay Sternberg /* Move to pci_ids.h later */ 220357dacad5SJay Sternberg #define PCI_CLASS_STORAGE_EXPRESS 0x010802 220457dacad5SJay Sternberg 220557dacad5SJay Sternberg static const struct pci_device_id nvme_id_table[] = { 2206106198edSChristoph Hellwig { PCI_VDEVICE(INTEL, 0x0953), 2207106198edSChristoph Hellwig .driver_data = NVME_QUIRK_STRIPE_SIZE, }, 2208540c801cSKeith Busch { PCI_VDEVICE(INTEL, 0x5845), /* Qemu emulated controller */ 2209540c801cSKeith Busch .driver_data = NVME_QUIRK_IDENTIFY_CNS, }, 221057dacad5SJay Sternberg { PCI_DEVICE_CLASS(PCI_CLASS_STORAGE_EXPRESS, 0xffffff) }, 2211c74dc780SStephan Günther { PCI_DEVICE(PCI_VENDOR_ID_APPLE, 0x2001) }, 221257dacad5SJay Sternberg { 0, } 221357dacad5SJay Sternberg }; 221457dacad5SJay Sternberg MODULE_DEVICE_TABLE(pci, nvme_id_table); 221557dacad5SJay Sternberg 221657dacad5SJay Sternberg static struct pci_driver nvme_driver = { 221757dacad5SJay Sternberg .name = "nvme", 221857dacad5SJay Sternberg .id_table = nvme_id_table, 221957dacad5SJay Sternberg .probe = nvme_probe, 222057dacad5SJay Sternberg .remove = nvme_remove, 222157dacad5SJay Sternberg .shutdown = nvme_shutdown, 222257dacad5SJay Sternberg .driver = { 222357dacad5SJay Sternberg .pm = &nvme_dev_pm_ops, 222457dacad5SJay Sternberg }, 222557dacad5SJay Sternberg .err_handler = &nvme_err_handler, 222657dacad5SJay Sternberg }; 222757dacad5SJay Sternberg 222857dacad5SJay Sternberg static int __init nvme_init(void) 222957dacad5SJay Sternberg { 223057dacad5SJay Sternberg int result; 223157dacad5SJay Sternberg 223257dacad5SJay Sternberg init_waitqueue_head(&nvme_kthread_wait); 223357dacad5SJay Sternberg 223492f7a162SKeith Busch nvme_workq = alloc_workqueue("nvme", WQ_UNBOUND | WQ_MEM_RECLAIM, 0); 223557dacad5SJay Sternberg if (!nvme_workq) 223657dacad5SJay Sternberg return -ENOMEM; 223757dacad5SJay Sternberg 22385bae7f73SChristoph Hellwig result = nvme_core_init(); 223957dacad5SJay Sternberg if (result < 0) 224057dacad5SJay Sternberg goto kill_workq; 224157dacad5SJay Sternberg 224257dacad5SJay Sternberg result = pci_register_driver(&nvme_driver); 224357dacad5SJay Sternberg if (result) 2244f3ca80fcSChristoph Hellwig goto core_exit; 224557dacad5SJay Sternberg return 0; 224657dacad5SJay Sternberg 2247f3ca80fcSChristoph Hellwig core_exit: 22485bae7f73SChristoph Hellwig nvme_core_exit(); 224957dacad5SJay Sternberg kill_workq: 225057dacad5SJay Sternberg destroy_workqueue(nvme_workq); 225157dacad5SJay Sternberg return result; 225257dacad5SJay Sternberg } 225357dacad5SJay Sternberg 225457dacad5SJay Sternberg static void __exit nvme_exit(void) 225557dacad5SJay Sternberg { 225657dacad5SJay Sternberg pci_unregister_driver(&nvme_driver); 22575bae7f73SChristoph Hellwig nvme_core_exit(); 225857dacad5SJay Sternberg destroy_workqueue(nvme_workq); 225957dacad5SJay Sternberg BUG_ON(nvme_thread && !IS_ERR(nvme_thread)); 226057dacad5SJay Sternberg _nvme_check_size(); 226157dacad5SJay Sternberg } 226257dacad5SJay Sternberg 226357dacad5SJay Sternberg MODULE_AUTHOR("Matthew Wilcox <willy@linux.intel.com>"); 226457dacad5SJay Sternberg MODULE_LICENSE("GPL"); 226557dacad5SJay Sternberg MODULE_VERSION("1.0"); 226657dacad5SJay Sternberg module_init(nvme_init); 226757dacad5SJay Sternberg module_exit(nvme_exit); 2268