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); 89a5cdb68cSKeith Busch static void nvme_dev_disable(struct nvme_dev *dev, bool shutdown); 9057dacad5SJay Sternberg 9157dacad5SJay Sternberg /* 921c63dc66SChristoph Hellwig * Represents an NVM Express device. Each nvme_dev is a PCI function. 931c63dc66SChristoph Hellwig */ 941c63dc66SChristoph Hellwig struct nvme_dev { 951c63dc66SChristoph Hellwig struct list_head node; 961c63dc66SChristoph Hellwig struct nvme_queue **queues; 971c63dc66SChristoph Hellwig struct blk_mq_tag_set tagset; 981c63dc66SChristoph Hellwig struct blk_mq_tag_set admin_tagset; 991c63dc66SChristoph Hellwig u32 __iomem *dbs; 1001c63dc66SChristoph Hellwig struct device *dev; 1011c63dc66SChristoph Hellwig struct dma_pool *prp_page_pool; 1021c63dc66SChristoph Hellwig struct dma_pool *prp_small_pool; 1031c63dc66SChristoph Hellwig unsigned queue_count; 1041c63dc66SChristoph Hellwig unsigned online_queues; 1051c63dc66SChristoph Hellwig unsigned max_qid; 1061c63dc66SChristoph Hellwig int q_depth; 1071c63dc66SChristoph Hellwig u32 db_stride; 1081c63dc66SChristoph Hellwig struct msix_entry *entry; 1091c63dc66SChristoph Hellwig void __iomem *bar; 1101c63dc66SChristoph Hellwig struct work_struct reset_work; 1111c63dc66SChristoph Hellwig struct work_struct scan_work; 1125c8809e6SChristoph Hellwig struct work_struct remove_work; 11377bf25eaSKeith Busch struct mutex shutdown_lock; 1141c63dc66SChristoph Hellwig bool subsystem; 1151c63dc66SChristoph Hellwig void __iomem *cmb; 1161c63dc66SChristoph Hellwig dma_addr_t cmb_dma_addr; 1171c63dc66SChristoph Hellwig u64 cmb_size; 1181c63dc66SChristoph Hellwig u32 cmbsz; 119fd634f41SChristoph Hellwig unsigned long flags; 120db3cbfffSKeith Busch 121fd634f41SChristoph Hellwig #define NVME_CTRL_RESETTING 0 122646017a6SKeith Busch #define NVME_CTRL_REMOVING 1 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 289646017a6SKeith Busch static void nvme_queue_scan(struct nvme_dev *dev) 290646017a6SKeith Busch { 291646017a6SKeith Busch /* 292646017a6SKeith Busch * Do not queue new scan work when a controller is reset during 293646017a6SKeith Busch * removal. 294646017a6SKeith Busch */ 295646017a6SKeith Busch if (test_bit(NVME_CTRL_REMOVING, &dev->flags)) 296646017a6SKeith Busch return; 297646017a6SKeith Busch queue_work(nvme_workq, &dev->scan_work); 298646017a6SKeith Busch } 299646017a6SKeith Busch 300adf68f21SChristoph Hellwig static void nvme_complete_async_event(struct nvme_dev *dev, 30157dacad5SJay Sternberg struct nvme_completion *cqe) 30257dacad5SJay Sternberg { 303adf68f21SChristoph Hellwig u16 status = le16_to_cpu(cqe->status) >> 1; 304adf68f21SChristoph Hellwig u32 result = le32_to_cpu(cqe->result); 30557dacad5SJay Sternberg 30657dacad5SJay Sternberg if (status == NVME_SC_SUCCESS || status == NVME_SC_ABORT_REQ) 307adf68f21SChristoph Hellwig ++dev->ctrl.event_limit; 30857dacad5SJay Sternberg if (status != NVME_SC_SUCCESS) 30957dacad5SJay Sternberg return; 31057dacad5SJay Sternberg 31157dacad5SJay Sternberg switch (result & 0xff07) { 31257dacad5SJay Sternberg case NVME_AER_NOTICE_NS_CHANGED: 313adf68f21SChristoph Hellwig dev_info(dev->dev, "rescanning\n"); 314646017a6SKeith Busch nvme_queue_scan(dev); 31557dacad5SJay Sternberg default: 316adf68f21SChristoph Hellwig dev_warn(dev->dev, "async event result %08x\n", result); 31757dacad5SJay Sternberg } 31857dacad5SJay Sternberg } 31957dacad5SJay Sternberg 32057dacad5SJay Sternberg /** 321adf68f21SChristoph Hellwig * __nvme_submit_cmd() - Copy a command into a queue and ring the doorbell 32257dacad5SJay Sternberg * @nvmeq: The queue to use 32357dacad5SJay Sternberg * @cmd: The command to send 32457dacad5SJay Sternberg * 32557dacad5SJay Sternberg * Safe to use from interrupt context 32657dacad5SJay Sternberg */ 32757dacad5SJay Sternberg static void __nvme_submit_cmd(struct nvme_queue *nvmeq, 32857dacad5SJay Sternberg struct nvme_command *cmd) 32957dacad5SJay Sternberg { 33057dacad5SJay Sternberg u16 tail = nvmeq->sq_tail; 33157dacad5SJay Sternberg 33257dacad5SJay Sternberg if (nvmeq->sq_cmds_io) 33357dacad5SJay Sternberg memcpy_toio(&nvmeq->sq_cmds_io[tail], cmd, sizeof(*cmd)); 33457dacad5SJay Sternberg else 33557dacad5SJay Sternberg memcpy(&nvmeq->sq_cmds[tail], cmd, sizeof(*cmd)); 33657dacad5SJay Sternberg 33757dacad5SJay Sternberg if (++tail == nvmeq->q_depth) 33857dacad5SJay Sternberg tail = 0; 33957dacad5SJay Sternberg writel(tail, nvmeq->q_db); 34057dacad5SJay Sternberg nvmeq->sq_tail = tail; 34157dacad5SJay Sternberg } 34257dacad5SJay Sternberg 343f4800d6dSChristoph Hellwig static __le64 **iod_list(struct request *req) 34457dacad5SJay Sternberg { 345f4800d6dSChristoph Hellwig struct nvme_iod *iod = blk_mq_rq_to_pdu(req); 346f4800d6dSChristoph Hellwig return (__le64 **)(iod->sg + req->nr_phys_segments); 34757dacad5SJay Sternberg } 34857dacad5SJay Sternberg 349f4800d6dSChristoph Hellwig static int nvme_init_iod(struct request *rq, struct nvme_dev *dev) 35057dacad5SJay Sternberg { 351f4800d6dSChristoph Hellwig struct nvme_iod *iod = blk_mq_rq_to_pdu(rq); 352f4800d6dSChristoph Hellwig int nseg = rq->nr_phys_segments; 353f4800d6dSChristoph Hellwig unsigned size; 354f4800d6dSChristoph Hellwig 355f4800d6dSChristoph Hellwig if (rq->cmd_flags & REQ_DISCARD) 356f4800d6dSChristoph Hellwig size = sizeof(struct nvme_dsm_range); 357f4800d6dSChristoph Hellwig else 358f4800d6dSChristoph Hellwig size = blk_rq_bytes(rq); 359f4800d6dSChristoph Hellwig 360f4800d6dSChristoph Hellwig if (nseg > NVME_INT_PAGES || size > NVME_INT_BYTES(dev)) { 361f4800d6dSChristoph Hellwig iod->sg = kmalloc(nvme_iod_alloc_size(dev, size, nseg), GFP_ATOMIC); 362f4800d6dSChristoph Hellwig if (!iod->sg) 363f4800d6dSChristoph Hellwig return BLK_MQ_RQ_QUEUE_BUSY; 364f4800d6dSChristoph Hellwig } else { 365f4800d6dSChristoph Hellwig iod->sg = iod->inline_sg; 36657dacad5SJay Sternberg } 36757dacad5SJay Sternberg 368f4800d6dSChristoph Hellwig iod->aborted = 0; 36957dacad5SJay Sternberg iod->npages = -1; 37057dacad5SJay Sternberg iod->nents = 0; 371f4800d6dSChristoph Hellwig iod->length = size; 372f4800d6dSChristoph Hellwig return 0; 37357dacad5SJay Sternberg } 37457dacad5SJay Sternberg 375f4800d6dSChristoph Hellwig static void nvme_free_iod(struct nvme_dev *dev, struct request *req) 37657dacad5SJay Sternberg { 377f4800d6dSChristoph Hellwig struct nvme_iod *iod = blk_mq_rq_to_pdu(req); 3785fd4ce1bSChristoph Hellwig const int last_prp = dev->ctrl.page_size / 8 - 1; 37957dacad5SJay Sternberg int i; 380f4800d6dSChristoph Hellwig __le64 **list = iod_list(req); 38157dacad5SJay Sternberg dma_addr_t prp_dma = iod->first_dma; 38257dacad5SJay Sternberg 38357dacad5SJay Sternberg if (iod->npages == 0) 38457dacad5SJay Sternberg dma_pool_free(dev->prp_small_pool, list[0], prp_dma); 38557dacad5SJay Sternberg for (i = 0; i < iod->npages; i++) { 38657dacad5SJay Sternberg __le64 *prp_list = list[i]; 38757dacad5SJay Sternberg dma_addr_t next_prp_dma = le64_to_cpu(prp_list[last_prp]); 38857dacad5SJay Sternberg dma_pool_free(dev->prp_page_pool, prp_list, prp_dma); 38957dacad5SJay Sternberg prp_dma = next_prp_dma; 39057dacad5SJay Sternberg } 39157dacad5SJay Sternberg 392f4800d6dSChristoph Hellwig if (iod->sg != iod->inline_sg) 393f4800d6dSChristoph Hellwig kfree(iod->sg); 39457dacad5SJay Sternberg } 39557dacad5SJay Sternberg 39657dacad5SJay Sternberg #ifdef CONFIG_BLK_DEV_INTEGRITY 39757dacad5SJay Sternberg static void nvme_dif_prep(u32 p, u32 v, struct t10_pi_tuple *pi) 39857dacad5SJay Sternberg { 39957dacad5SJay Sternberg if (be32_to_cpu(pi->ref_tag) == v) 40057dacad5SJay Sternberg pi->ref_tag = cpu_to_be32(p); 40157dacad5SJay Sternberg } 40257dacad5SJay Sternberg 40357dacad5SJay Sternberg static void nvme_dif_complete(u32 p, u32 v, struct t10_pi_tuple *pi) 40457dacad5SJay Sternberg { 40557dacad5SJay Sternberg if (be32_to_cpu(pi->ref_tag) == p) 40657dacad5SJay Sternberg pi->ref_tag = cpu_to_be32(v); 40757dacad5SJay Sternberg } 40857dacad5SJay Sternberg 40957dacad5SJay Sternberg /** 41057dacad5SJay Sternberg * nvme_dif_remap - remaps ref tags to bip seed and physical lba 41157dacad5SJay Sternberg * 41257dacad5SJay Sternberg * The virtual start sector is the one that was originally submitted by the 41357dacad5SJay Sternberg * block layer. Due to partitioning, MD/DM cloning, etc. the actual physical 41457dacad5SJay Sternberg * start sector may be different. Remap protection information to match the 41557dacad5SJay Sternberg * physical LBA on writes, and back to the original seed on reads. 41657dacad5SJay Sternberg * 41757dacad5SJay Sternberg * Type 0 and 3 do not have a ref tag, so no remapping required. 41857dacad5SJay Sternberg */ 41957dacad5SJay Sternberg static void nvme_dif_remap(struct request *req, 42057dacad5SJay Sternberg void (*dif_swap)(u32 p, u32 v, struct t10_pi_tuple *pi)) 42157dacad5SJay Sternberg { 42257dacad5SJay Sternberg struct nvme_ns *ns = req->rq_disk->private_data; 42357dacad5SJay Sternberg struct bio_integrity_payload *bip; 42457dacad5SJay Sternberg struct t10_pi_tuple *pi; 42557dacad5SJay Sternberg void *p, *pmap; 42657dacad5SJay Sternberg u32 i, nlb, ts, phys, virt; 42757dacad5SJay Sternberg 42857dacad5SJay Sternberg if (!ns->pi_type || ns->pi_type == NVME_NS_DPS_PI_TYPE3) 42957dacad5SJay Sternberg return; 43057dacad5SJay Sternberg 43157dacad5SJay Sternberg bip = bio_integrity(req->bio); 43257dacad5SJay Sternberg if (!bip) 43357dacad5SJay Sternberg return; 43457dacad5SJay Sternberg 43557dacad5SJay Sternberg pmap = kmap_atomic(bip->bip_vec->bv_page) + bip->bip_vec->bv_offset; 43657dacad5SJay Sternberg 43757dacad5SJay Sternberg p = pmap; 43857dacad5SJay Sternberg virt = bip_get_seed(bip); 43957dacad5SJay Sternberg phys = nvme_block_nr(ns, blk_rq_pos(req)); 44057dacad5SJay Sternberg nlb = (blk_rq_bytes(req) >> ns->lba_shift); 441ac6fc48cSDan Williams ts = ns->disk->queue->integrity.tuple_size; 44257dacad5SJay Sternberg 44357dacad5SJay Sternberg for (i = 0; i < nlb; i++, virt++, phys++) { 44457dacad5SJay Sternberg pi = (struct t10_pi_tuple *)p; 44557dacad5SJay Sternberg dif_swap(phys, virt, pi); 44657dacad5SJay Sternberg p += ts; 44757dacad5SJay Sternberg } 44857dacad5SJay Sternberg kunmap_atomic(pmap); 44957dacad5SJay Sternberg } 45057dacad5SJay Sternberg #else /* CONFIG_BLK_DEV_INTEGRITY */ 45157dacad5SJay Sternberg static void nvme_dif_remap(struct request *req, 45257dacad5SJay Sternberg void (*dif_swap)(u32 p, u32 v, struct t10_pi_tuple *pi)) 45357dacad5SJay Sternberg { 45457dacad5SJay Sternberg } 45557dacad5SJay Sternberg static void nvme_dif_prep(u32 p, u32 v, struct t10_pi_tuple *pi) 45657dacad5SJay Sternberg { 45757dacad5SJay Sternberg } 45857dacad5SJay Sternberg static void nvme_dif_complete(u32 p, u32 v, struct t10_pi_tuple *pi) 45957dacad5SJay Sternberg { 46057dacad5SJay Sternberg } 46157dacad5SJay Sternberg #endif 46257dacad5SJay Sternberg 463f4800d6dSChristoph Hellwig static bool nvme_setup_prps(struct nvme_dev *dev, struct request *req, 46469d2b571SChristoph Hellwig int total_len) 46557dacad5SJay Sternberg { 466f4800d6dSChristoph Hellwig struct nvme_iod *iod = blk_mq_rq_to_pdu(req); 46757dacad5SJay Sternberg struct dma_pool *pool; 46857dacad5SJay Sternberg int length = total_len; 46957dacad5SJay Sternberg struct scatterlist *sg = iod->sg; 47057dacad5SJay Sternberg int dma_len = sg_dma_len(sg); 47157dacad5SJay Sternberg u64 dma_addr = sg_dma_address(sg); 4725fd4ce1bSChristoph Hellwig u32 page_size = dev->ctrl.page_size; 47357dacad5SJay Sternberg int offset = dma_addr & (page_size - 1); 47457dacad5SJay Sternberg __le64 *prp_list; 475f4800d6dSChristoph Hellwig __le64 **list = iod_list(req); 47657dacad5SJay Sternberg dma_addr_t prp_dma; 47757dacad5SJay Sternberg int nprps, i; 47857dacad5SJay Sternberg 47957dacad5SJay Sternberg length -= (page_size - offset); 48057dacad5SJay Sternberg if (length <= 0) 48169d2b571SChristoph Hellwig return true; 48257dacad5SJay Sternberg 48357dacad5SJay Sternberg dma_len -= (page_size - offset); 48457dacad5SJay Sternberg if (dma_len) { 48557dacad5SJay Sternberg dma_addr += (page_size - offset); 48657dacad5SJay Sternberg } else { 48757dacad5SJay Sternberg sg = sg_next(sg); 48857dacad5SJay Sternberg dma_addr = sg_dma_address(sg); 48957dacad5SJay Sternberg dma_len = sg_dma_len(sg); 49057dacad5SJay Sternberg } 49157dacad5SJay Sternberg 49257dacad5SJay Sternberg if (length <= page_size) { 49357dacad5SJay Sternberg iod->first_dma = dma_addr; 49469d2b571SChristoph Hellwig return true; 49557dacad5SJay Sternberg } 49657dacad5SJay Sternberg 49757dacad5SJay Sternberg nprps = DIV_ROUND_UP(length, page_size); 49857dacad5SJay Sternberg if (nprps <= (256 / 8)) { 49957dacad5SJay Sternberg pool = dev->prp_small_pool; 50057dacad5SJay Sternberg iod->npages = 0; 50157dacad5SJay Sternberg } else { 50257dacad5SJay Sternberg pool = dev->prp_page_pool; 50357dacad5SJay Sternberg iod->npages = 1; 50457dacad5SJay Sternberg } 50557dacad5SJay Sternberg 50669d2b571SChristoph Hellwig prp_list = dma_pool_alloc(pool, GFP_ATOMIC, &prp_dma); 50757dacad5SJay Sternberg if (!prp_list) { 50857dacad5SJay Sternberg iod->first_dma = dma_addr; 50957dacad5SJay Sternberg iod->npages = -1; 51069d2b571SChristoph Hellwig return false; 51157dacad5SJay Sternberg } 51257dacad5SJay Sternberg list[0] = prp_list; 51357dacad5SJay Sternberg iod->first_dma = prp_dma; 51457dacad5SJay Sternberg i = 0; 51557dacad5SJay Sternberg for (;;) { 51657dacad5SJay Sternberg if (i == page_size >> 3) { 51757dacad5SJay Sternberg __le64 *old_prp_list = prp_list; 51869d2b571SChristoph Hellwig prp_list = dma_pool_alloc(pool, GFP_ATOMIC, &prp_dma); 51957dacad5SJay Sternberg if (!prp_list) 52069d2b571SChristoph Hellwig return false; 52157dacad5SJay Sternberg list[iod->npages++] = prp_list; 52257dacad5SJay Sternberg prp_list[0] = old_prp_list[i - 1]; 52357dacad5SJay Sternberg old_prp_list[i - 1] = cpu_to_le64(prp_dma); 52457dacad5SJay Sternberg i = 1; 52557dacad5SJay Sternberg } 52657dacad5SJay Sternberg prp_list[i++] = cpu_to_le64(dma_addr); 52757dacad5SJay Sternberg dma_len -= page_size; 52857dacad5SJay Sternberg dma_addr += page_size; 52957dacad5SJay Sternberg length -= page_size; 53057dacad5SJay Sternberg if (length <= 0) 53157dacad5SJay Sternberg break; 53257dacad5SJay Sternberg if (dma_len > 0) 53357dacad5SJay Sternberg continue; 53457dacad5SJay Sternberg BUG_ON(dma_len < 0); 53557dacad5SJay Sternberg sg = sg_next(sg); 53657dacad5SJay Sternberg dma_addr = sg_dma_address(sg); 53757dacad5SJay Sternberg dma_len = sg_dma_len(sg); 53857dacad5SJay Sternberg } 53957dacad5SJay Sternberg 54069d2b571SChristoph Hellwig return true; 54157dacad5SJay Sternberg } 54257dacad5SJay Sternberg 543f4800d6dSChristoph Hellwig static int nvme_map_data(struct nvme_dev *dev, struct request *req, 544ba1ca37eSChristoph Hellwig struct nvme_command *cmnd) 54557dacad5SJay Sternberg { 546f4800d6dSChristoph Hellwig struct nvme_iod *iod = blk_mq_rq_to_pdu(req); 547ba1ca37eSChristoph Hellwig struct request_queue *q = req->q; 548ba1ca37eSChristoph Hellwig enum dma_data_direction dma_dir = rq_data_dir(req) ? 549ba1ca37eSChristoph Hellwig DMA_TO_DEVICE : DMA_FROM_DEVICE; 550ba1ca37eSChristoph Hellwig int ret = BLK_MQ_RQ_QUEUE_ERROR; 55157dacad5SJay Sternberg 552ba1ca37eSChristoph Hellwig sg_init_table(iod->sg, req->nr_phys_segments); 553ba1ca37eSChristoph Hellwig iod->nents = blk_rq_map_sg(q, req, iod->sg); 554ba1ca37eSChristoph Hellwig if (!iod->nents) 555ba1ca37eSChristoph Hellwig goto out; 556ba1ca37eSChristoph Hellwig 557ba1ca37eSChristoph Hellwig ret = BLK_MQ_RQ_QUEUE_BUSY; 558ba1ca37eSChristoph Hellwig if (!dma_map_sg(dev->dev, iod->sg, iod->nents, dma_dir)) 559ba1ca37eSChristoph Hellwig goto out; 560ba1ca37eSChristoph Hellwig 561f4800d6dSChristoph Hellwig if (!nvme_setup_prps(dev, req, blk_rq_bytes(req))) 562ba1ca37eSChristoph Hellwig goto out_unmap; 563ba1ca37eSChristoph Hellwig 564ba1ca37eSChristoph Hellwig ret = BLK_MQ_RQ_QUEUE_ERROR; 565ba1ca37eSChristoph Hellwig if (blk_integrity_rq(req)) { 566ba1ca37eSChristoph Hellwig if (blk_rq_count_integrity_sg(q, req->bio) != 1) 567ba1ca37eSChristoph Hellwig goto out_unmap; 568ba1ca37eSChristoph Hellwig 569bf684057SChristoph Hellwig sg_init_table(&iod->meta_sg, 1); 570bf684057SChristoph Hellwig if (blk_rq_map_integrity_sg(q, req->bio, &iod->meta_sg) != 1) 571ba1ca37eSChristoph Hellwig goto out_unmap; 572ba1ca37eSChristoph Hellwig 573ba1ca37eSChristoph Hellwig if (rq_data_dir(req)) 574ba1ca37eSChristoph Hellwig nvme_dif_remap(req, nvme_dif_prep); 575ba1ca37eSChristoph Hellwig 576bf684057SChristoph Hellwig if (!dma_map_sg(dev->dev, &iod->meta_sg, 1, dma_dir)) 577ba1ca37eSChristoph Hellwig goto out_unmap; 57857dacad5SJay Sternberg } 57957dacad5SJay Sternberg 580ba1ca37eSChristoph Hellwig cmnd->rw.prp1 = cpu_to_le64(sg_dma_address(iod->sg)); 581ba1ca37eSChristoph Hellwig cmnd->rw.prp2 = cpu_to_le64(iod->first_dma); 582ba1ca37eSChristoph Hellwig if (blk_integrity_rq(req)) 583bf684057SChristoph Hellwig cmnd->rw.metadata = cpu_to_le64(sg_dma_address(&iod->meta_sg)); 584ba1ca37eSChristoph Hellwig return BLK_MQ_RQ_QUEUE_OK; 585ba1ca37eSChristoph Hellwig 586ba1ca37eSChristoph Hellwig out_unmap: 587ba1ca37eSChristoph Hellwig dma_unmap_sg(dev->dev, iod->sg, iod->nents, dma_dir); 588ba1ca37eSChristoph Hellwig out: 589ba1ca37eSChristoph Hellwig return ret; 59057dacad5SJay Sternberg } 59157dacad5SJay Sternberg 592f4800d6dSChristoph Hellwig static void nvme_unmap_data(struct nvme_dev *dev, struct request *req) 593d4f6c3abSChristoph Hellwig { 594f4800d6dSChristoph Hellwig struct nvme_iod *iod = blk_mq_rq_to_pdu(req); 595d4f6c3abSChristoph Hellwig enum dma_data_direction dma_dir = rq_data_dir(req) ? 596d4f6c3abSChristoph Hellwig DMA_TO_DEVICE : DMA_FROM_DEVICE; 597d4f6c3abSChristoph Hellwig 598d4f6c3abSChristoph Hellwig if (iod->nents) { 599d4f6c3abSChristoph Hellwig dma_unmap_sg(dev->dev, iod->sg, iod->nents, dma_dir); 600d4f6c3abSChristoph Hellwig if (blk_integrity_rq(req)) { 601d4f6c3abSChristoph Hellwig if (!rq_data_dir(req)) 602d4f6c3abSChristoph Hellwig nvme_dif_remap(req, nvme_dif_complete); 603bf684057SChristoph Hellwig dma_unmap_sg(dev->dev, &iod->meta_sg, 1, dma_dir); 604d4f6c3abSChristoph Hellwig } 605d4f6c3abSChristoph Hellwig } 606d4f6c3abSChristoph Hellwig 607f4800d6dSChristoph Hellwig nvme_free_iod(dev, req); 60857dacad5SJay Sternberg } 60957dacad5SJay Sternberg 61057dacad5SJay Sternberg /* 61157dacad5SJay Sternberg * We reuse the small pool to allocate the 16-byte range here as it is not 61257dacad5SJay Sternberg * worth having a special pool for these or additional cases to handle freeing 61357dacad5SJay Sternberg * the iod. 61457dacad5SJay Sternberg */ 615ba1ca37eSChristoph Hellwig static int nvme_setup_discard(struct nvme_queue *nvmeq, struct nvme_ns *ns, 616f4800d6dSChristoph Hellwig struct request *req, struct nvme_command *cmnd) 61757dacad5SJay Sternberg { 618f4800d6dSChristoph Hellwig struct nvme_iod *iod = blk_mq_rq_to_pdu(req); 619ba1ca37eSChristoph Hellwig struct nvme_dsm_range *range; 620ba1ca37eSChristoph Hellwig 621ba1ca37eSChristoph Hellwig range = dma_pool_alloc(nvmeq->dev->prp_small_pool, GFP_ATOMIC, 622ba1ca37eSChristoph Hellwig &iod->first_dma); 623ba1ca37eSChristoph Hellwig if (!range) 624ba1ca37eSChristoph Hellwig return BLK_MQ_RQ_QUEUE_BUSY; 625f4800d6dSChristoph Hellwig iod_list(req)[0] = (__le64 *)range; 626ba1ca37eSChristoph Hellwig iod->npages = 0; 62757dacad5SJay Sternberg 62857dacad5SJay Sternberg range->cattr = cpu_to_le32(0); 62957dacad5SJay Sternberg range->nlb = cpu_to_le32(blk_rq_bytes(req) >> ns->lba_shift); 63057dacad5SJay Sternberg range->slba = cpu_to_le64(nvme_block_nr(ns, blk_rq_pos(req))); 63157dacad5SJay Sternberg 632ba1ca37eSChristoph Hellwig memset(cmnd, 0, sizeof(*cmnd)); 633ba1ca37eSChristoph Hellwig cmnd->dsm.opcode = nvme_cmd_dsm; 634ba1ca37eSChristoph Hellwig cmnd->dsm.nsid = cpu_to_le32(ns->ns_id); 635ba1ca37eSChristoph Hellwig cmnd->dsm.prp1 = cpu_to_le64(iod->first_dma); 636ba1ca37eSChristoph Hellwig cmnd->dsm.nr = 0; 637ba1ca37eSChristoph Hellwig cmnd->dsm.attributes = cpu_to_le32(NVME_DSMGMT_AD); 638ba1ca37eSChristoph Hellwig return BLK_MQ_RQ_QUEUE_OK; 63957dacad5SJay Sternberg } 64057dacad5SJay Sternberg 64157dacad5SJay Sternberg /* 64257dacad5SJay Sternberg * NOTE: ns is NULL when called on the admin queue. 64357dacad5SJay Sternberg */ 64457dacad5SJay Sternberg static int nvme_queue_rq(struct blk_mq_hw_ctx *hctx, 64557dacad5SJay Sternberg const struct blk_mq_queue_data *bd) 64657dacad5SJay Sternberg { 64757dacad5SJay Sternberg struct nvme_ns *ns = hctx->queue->queuedata; 64857dacad5SJay Sternberg struct nvme_queue *nvmeq = hctx->driver_data; 64957dacad5SJay Sternberg struct nvme_dev *dev = nvmeq->dev; 65057dacad5SJay Sternberg struct request *req = bd->rq; 651ba1ca37eSChristoph Hellwig struct nvme_command cmnd; 652ba1ca37eSChristoph Hellwig int ret = BLK_MQ_RQ_QUEUE_OK; 65357dacad5SJay Sternberg 65457dacad5SJay Sternberg /* 65557dacad5SJay Sternberg * If formated with metadata, require the block layer provide a buffer 65657dacad5SJay Sternberg * unless this namespace is formated such that the metadata can be 65757dacad5SJay Sternberg * stripped/generated by the controller with PRACT=1. 65857dacad5SJay Sternberg */ 65957dacad5SJay Sternberg if (ns && ns->ms && !blk_integrity_rq(req)) { 66057dacad5SJay Sternberg if (!(ns->pi_type && ns->ms == 8) && 66157dacad5SJay Sternberg req->cmd_type != REQ_TYPE_DRV_PRIV) { 662eee417b0SChristoph Hellwig blk_mq_end_request(req, -EFAULT); 66357dacad5SJay Sternberg return BLK_MQ_RQ_QUEUE_OK; 66457dacad5SJay Sternberg } 66557dacad5SJay Sternberg } 66657dacad5SJay Sternberg 667f4800d6dSChristoph Hellwig ret = nvme_init_iod(req, dev); 668f4800d6dSChristoph Hellwig if (ret) 669f4800d6dSChristoph Hellwig return ret; 67057dacad5SJay Sternberg 67157dacad5SJay Sternberg if (req->cmd_flags & REQ_DISCARD) { 672f4800d6dSChristoph Hellwig ret = nvme_setup_discard(nvmeq, ns, req, &cmnd); 673ba1ca37eSChristoph Hellwig } else { 67457dacad5SJay Sternberg if (req->cmd_type == REQ_TYPE_DRV_PRIV) 675ba1ca37eSChristoph Hellwig memcpy(&cmnd, req->cmd, sizeof(cmnd)); 67657dacad5SJay Sternberg else if (req->cmd_flags & REQ_FLUSH) 677ba1ca37eSChristoph Hellwig nvme_setup_flush(ns, &cmnd); 67857dacad5SJay Sternberg else 679ba1ca37eSChristoph Hellwig nvme_setup_rw(ns, req, &cmnd); 68057dacad5SJay Sternberg 681ba1ca37eSChristoph Hellwig if (req->nr_phys_segments) 682f4800d6dSChristoph Hellwig ret = nvme_map_data(dev, req, &cmnd); 683ba1ca37eSChristoph Hellwig } 684ba1ca37eSChristoph Hellwig 685ba1ca37eSChristoph Hellwig if (ret) 686ba1ca37eSChristoph Hellwig goto out; 687ba1ca37eSChristoph Hellwig 688ba1ca37eSChristoph Hellwig cmnd.common.command_id = req->tag; 689aae239e1SChristoph Hellwig blk_mq_start_request(req); 690ba1ca37eSChristoph Hellwig 691ba1ca37eSChristoph Hellwig spin_lock_irq(&nvmeq->q_lock); 692ae1fba20SKeith Busch if (unlikely(nvmeq->cq_vector < 0)) { 69369d9a99cSKeith Busch if (ns && !test_bit(NVME_NS_DEAD, &ns->flags)) 694ae1fba20SKeith Busch ret = BLK_MQ_RQ_QUEUE_BUSY; 69569d9a99cSKeith Busch else 69669d9a99cSKeith Busch ret = BLK_MQ_RQ_QUEUE_ERROR; 697ae1fba20SKeith Busch spin_unlock_irq(&nvmeq->q_lock); 698ae1fba20SKeith Busch goto out; 699ae1fba20SKeith Busch } 700ba1ca37eSChristoph Hellwig __nvme_submit_cmd(nvmeq, &cmnd); 70157dacad5SJay Sternberg nvme_process_cq(nvmeq); 70257dacad5SJay Sternberg spin_unlock_irq(&nvmeq->q_lock); 70357dacad5SJay Sternberg return BLK_MQ_RQ_QUEUE_OK; 704ba1ca37eSChristoph Hellwig out: 705f4800d6dSChristoph Hellwig nvme_free_iod(dev, req); 706ba1ca37eSChristoph Hellwig return ret; 70757dacad5SJay Sternberg } 70857dacad5SJay Sternberg 709eee417b0SChristoph Hellwig static void nvme_complete_rq(struct request *req) 710eee417b0SChristoph Hellwig { 711f4800d6dSChristoph Hellwig struct nvme_iod *iod = blk_mq_rq_to_pdu(req); 712f4800d6dSChristoph Hellwig struct nvme_dev *dev = iod->nvmeq->dev; 713eee417b0SChristoph Hellwig int error = 0; 714eee417b0SChristoph Hellwig 715f4800d6dSChristoph Hellwig nvme_unmap_data(dev, req); 716eee417b0SChristoph Hellwig 717eee417b0SChristoph Hellwig if (unlikely(req->errors)) { 718eee417b0SChristoph Hellwig if (nvme_req_needs_retry(req, req->errors)) { 719eee417b0SChristoph Hellwig nvme_requeue_req(req); 720eee417b0SChristoph Hellwig return; 721eee417b0SChristoph Hellwig } 722eee417b0SChristoph Hellwig 723eee417b0SChristoph Hellwig if (req->cmd_type == REQ_TYPE_DRV_PRIV) 724eee417b0SChristoph Hellwig error = req->errors; 725eee417b0SChristoph Hellwig else 726eee417b0SChristoph Hellwig error = nvme_error_status(req->errors); 727eee417b0SChristoph Hellwig } 728eee417b0SChristoph Hellwig 729f4800d6dSChristoph Hellwig if (unlikely(iod->aborted)) { 730eee417b0SChristoph Hellwig dev_warn(dev->dev, 731eee417b0SChristoph Hellwig "completing aborted command with status: %04x\n", 732eee417b0SChristoph Hellwig req->errors); 733eee417b0SChristoph Hellwig } 734eee417b0SChristoph Hellwig 735eee417b0SChristoph Hellwig blk_mq_end_request(req, error); 73657dacad5SJay Sternberg } 73757dacad5SJay Sternberg 738a0fa9647SJens Axboe static void __nvme_process_cq(struct nvme_queue *nvmeq, unsigned int *tag) 73957dacad5SJay Sternberg { 74057dacad5SJay Sternberg u16 head, phase; 74157dacad5SJay Sternberg 74257dacad5SJay Sternberg head = nvmeq->cq_head; 74357dacad5SJay Sternberg phase = nvmeq->cq_phase; 74457dacad5SJay Sternberg 74557dacad5SJay Sternberg for (;;) { 74657dacad5SJay Sternberg struct nvme_completion cqe = nvmeq->cqes[head]; 747adf68f21SChristoph Hellwig u16 status = le16_to_cpu(cqe.status); 748eee417b0SChristoph Hellwig struct request *req; 749adf68f21SChristoph Hellwig 750adf68f21SChristoph Hellwig if ((status & 1) != phase) 75157dacad5SJay Sternberg break; 75257dacad5SJay Sternberg nvmeq->sq_head = le16_to_cpu(cqe.sq_head); 75357dacad5SJay Sternberg if (++head == nvmeq->q_depth) { 75457dacad5SJay Sternberg head = 0; 75557dacad5SJay Sternberg phase = !phase; 75657dacad5SJay Sternberg } 757adf68f21SChristoph Hellwig 758a0fa9647SJens Axboe if (tag && *tag == cqe.command_id) 759a0fa9647SJens Axboe *tag = -1; 760adf68f21SChristoph Hellwig 761aae239e1SChristoph Hellwig if (unlikely(cqe.command_id >= nvmeq->q_depth)) { 762aae239e1SChristoph Hellwig dev_warn(nvmeq->q_dmadev, 763aae239e1SChristoph Hellwig "invalid id %d completed on queue %d\n", 764aae239e1SChristoph Hellwig cqe.command_id, le16_to_cpu(cqe.sq_id)); 765aae239e1SChristoph Hellwig continue; 766aae239e1SChristoph Hellwig } 767aae239e1SChristoph Hellwig 768adf68f21SChristoph Hellwig /* 769adf68f21SChristoph Hellwig * AEN requests are special as they don't time out and can 770adf68f21SChristoph Hellwig * survive any kind of queue freeze and often don't respond to 771adf68f21SChristoph Hellwig * aborts. We don't even bother to allocate a struct request 772adf68f21SChristoph Hellwig * for them but rather special case them here. 773adf68f21SChristoph Hellwig */ 774adf68f21SChristoph Hellwig if (unlikely(nvmeq->qid == 0 && 775adf68f21SChristoph Hellwig cqe.command_id >= NVME_AQ_BLKMQ_DEPTH)) { 776adf68f21SChristoph Hellwig nvme_complete_async_event(nvmeq->dev, &cqe); 777adf68f21SChristoph Hellwig continue; 778adf68f21SChristoph Hellwig } 779adf68f21SChristoph Hellwig 780eee417b0SChristoph Hellwig req = blk_mq_tag_to_rq(*nvmeq->tags, cqe.command_id); 781eee417b0SChristoph Hellwig if (req->cmd_type == REQ_TYPE_DRV_PRIV) { 782eee417b0SChristoph Hellwig u32 result = le32_to_cpu(cqe.result); 783eee417b0SChristoph Hellwig req->special = (void *)(uintptr_t)result; 784eee417b0SChristoph Hellwig } 785eee417b0SChristoph Hellwig blk_mq_complete_request(req, status >> 1); 786eee417b0SChristoph Hellwig 78757dacad5SJay Sternberg } 78857dacad5SJay Sternberg 78957dacad5SJay Sternberg /* If the controller ignores the cq head doorbell and continuously 79057dacad5SJay Sternberg * writes to the queue, it is theoretically possible to wrap around 79157dacad5SJay Sternberg * the queue twice and mistakenly return IRQ_NONE. Linux only 79257dacad5SJay Sternberg * requires that 0.1% of your interrupts are handled, so this isn't 79357dacad5SJay Sternberg * a big problem. 79457dacad5SJay Sternberg */ 79557dacad5SJay Sternberg if (head == nvmeq->cq_head && phase == nvmeq->cq_phase) 796a0fa9647SJens Axboe return; 79757dacad5SJay Sternberg 798604e8c8dSKeith Busch if (likely(nvmeq->cq_vector >= 0)) 79957dacad5SJay Sternberg writel(head, nvmeq->q_db + nvmeq->dev->db_stride); 80057dacad5SJay Sternberg nvmeq->cq_head = head; 80157dacad5SJay Sternberg nvmeq->cq_phase = phase; 80257dacad5SJay Sternberg 80357dacad5SJay Sternberg nvmeq->cqe_seen = 1; 804a0fa9647SJens Axboe } 805a0fa9647SJens Axboe 806a0fa9647SJens Axboe static void nvme_process_cq(struct nvme_queue *nvmeq) 807a0fa9647SJens Axboe { 808a0fa9647SJens Axboe __nvme_process_cq(nvmeq, NULL); 80957dacad5SJay Sternberg } 81057dacad5SJay Sternberg 81157dacad5SJay Sternberg static irqreturn_t nvme_irq(int irq, void *data) 81257dacad5SJay Sternberg { 81357dacad5SJay Sternberg irqreturn_t result; 81457dacad5SJay Sternberg struct nvme_queue *nvmeq = data; 81557dacad5SJay Sternberg spin_lock(&nvmeq->q_lock); 81657dacad5SJay Sternberg nvme_process_cq(nvmeq); 81757dacad5SJay Sternberg result = nvmeq->cqe_seen ? IRQ_HANDLED : IRQ_NONE; 81857dacad5SJay Sternberg nvmeq->cqe_seen = 0; 81957dacad5SJay Sternberg spin_unlock(&nvmeq->q_lock); 82057dacad5SJay Sternberg return result; 82157dacad5SJay Sternberg } 82257dacad5SJay Sternberg 82357dacad5SJay Sternberg static irqreturn_t nvme_irq_check(int irq, void *data) 82457dacad5SJay Sternberg { 82557dacad5SJay Sternberg struct nvme_queue *nvmeq = data; 82657dacad5SJay Sternberg struct nvme_completion cqe = nvmeq->cqes[nvmeq->cq_head]; 82757dacad5SJay Sternberg if ((le16_to_cpu(cqe.status) & 1) != nvmeq->cq_phase) 82857dacad5SJay Sternberg return IRQ_NONE; 82957dacad5SJay Sternberg return IRQ_WAKE_THREAD; 83057dacad5SJay Sternberg } 83157dacad5SJay Sternberg 832a0fa9647SJens Axboe static int nvme_poll(struct blk_mq_hw_ctx *hctx, unsigned int tag) 833a0fa9647SJens Axboe { 834a0fa9647SJens Axboe struct nvme_queue *nvmeq = hctx->driver_data; 835a0fa9647SJens Axboe 836a0fa9647SJens Axboe if ((le16_to_cpu(nvmeq->cqes[nvmeq->cq_head].status) & 1) == 837a0fa9647SJens Axboe nvmeq->cq_phase) { 838a0fa9647SJens Axboe spin_lock_irq(&nvmeq->q_lock); 839a0fa9647SJens Axboe __nvme_process_cq(nvmeq, &tag); 840a0fa9647SJens Axboe spin_unlock_irq(&nvmeq->q_lock); 841a0fa9647SJens Axboe 842a0fa9647SJens Axboe if (tag == -1) 843a0fa9647SJens Axboe return 1; 844a0fa9647SJens Axboe } 845a0fa9647SJens Axboe 846a0fa9647SJens Axboe return 0; 847a0fa9647SJens Axboe } 848a0fa9647SJens Axboe 849adf68f21SChristoph Hellwig static void nvme_submit_async_event(struct nvme_dev *dev) 85057dacad5SJay Sternberg { 85157dacad5SJay Sternberg struct nvme_command c; 85257dacad5SJay Sternberg 85357dacad5SJay Sternberg memset(&c, 0, sizeof(c)); 85457dacad5SJay Sternberg c.common.opcode = nvme_admin_async_event; 855adf68f21SChristoph Hellwig c.common.command_id = NVME_AQ_BLKMQ_DEPTH + --dev->ctrl.event_limit; 85657dacad5SJay Sternberg 857adf68f21SChristoph Hellwig __nvme_submit_cmd(dev->queues[0], &c); 85857dacad5SJay Sternberg } 85957dacad5SJay Sternberg 86057dacad5SJay Sternberg static int adapter_delete_queue(struct nvme_dev *dev, u8 opcode, u16 id) 86157dacad5SJay Sternberg { 86257dacad5SJay Sternberg struct nvme_command c; 86357dacad5SJay Sternberg 86457dacad5SJay Sternberg memset(&c, 0, sizeof(c)); 86557dacad5SJay Sternberg c.delete_queue.opcode = opcode; 86657dacad5SJay Sternberg c.delete_queue.qid = cpu_to_le16(id); 86757dacad5SJay Sternberg 8681c63dc66SChristoph Hellwig return nvme_submit_sync_cmd(dev->ctrl.admin_q, &c, NULL, 0); 86957dacad5SJay Sternberg } 87057dacad5SJay Sternberg 87157dacad5SJay Sternberg static int adapter_alloc_cq(struct nvme_dev *dev, u16 qid, 87257dacad5SJay Sternberg struct nvme_queue *nvmeq) 87357dacad5SJay Sternberg { 87457dacad5SJay Sternberg struct nvme_command c; 87557dacad5SJay Sternberg int flags = NVME_QUEUE_PHYS_CONTIG | NVME_CQ_IRQ_ENABLED; 87657dacad5SJay Sternberg 87757dacad5SJay Sternberg /* 87857dacad5SJay Sternberg * Note: we (ab)use the fact the the prp fields survive if no data 87957dacad5SJay Sternberg * is attached to the request. 88057dacad5SJay Sternberg */ 88157dacad5SJay Sternberg memset(&c, 0, sizeof(c)); 88257dacad5SJay Sternberg c.create_cq.opcode = nvme_admin_create_cq; 88357dacad5SJay Sternberg c.create_cq.prp1 = cpu_to_le64(nvmeq->cq_dma_addr); 88457dacad5SJay Sternberg c.create_cq.cqid = cpu_to_le16(qid); 88557dacad5SJay Sternberg c.create_cq.qsize = cpu_to_le16(nvmeq->q_depth - 1); 88657dacad5SJay Sternberg c.create_cq.cq_flags = cpu_to_le16(flags); 88757dacad5SJay Sternberg c.create_cq.irq_vector = cpu_to_le16(nvmeq->cq_vector); 88857dacad5SJay Sternberg 8891c63dc66SChristoph Hellwig return nvme_submit_sync_cmd(dev->ctrl.admin_q, &c, NULL, 0); 89057dacad5SJay Sternberg } 89157dacad5SJay Sternberg 89257dacad5SJay Sternberg static int adapter_alloc_sq(struct nvme_dev *dev, u16 qid, 89357dacad5SJay Sternberg struct nvme_queue *nvmeq) 89457dacad5SJay Sternberg { 89557dacad5SJay Sternberg struct nvme_command c; 89657dacad5SJay Sternberg int flags = NVME_QUEUE_PHYS_CONTIG | NVME_SQ_PRIO_MEDIUM; 89757dacad5SJay Sternberg 89857dacad5SJay Sternberg /* 89957dacad5SJay Sternberg * Note: we (ab)use the fact the the prp fields survive if no data 90057dacad5SJay Sternberg * is attached to the request. 90157dacad5SJay Sternberg */ 90257dacad5SJay Sternberg memset(&c, 0, sizeof(c)); 90357dacad5SJay Sternberg c.create_sq.opcode = nvme_admin_create_sq; 90457dacad5SJay Sternberg c.create_sq.prp1 = cpu_to_le64(nvmeq->sq_dma_addr); 90557dacad5SJay Sternberg c.create_sq.sqid = cpu_to_le16(qid); 90657dacad5SJay Sternberg c.create_sq.qsize = cpu_to_le16(nvmeq->q_depth - 1); 90757dacad5SJay Sternberg c.create_sq.sq_flags = cpu_to_le16(flags); 90857dacad5SJay Sternberg c.create_sq.cqid = cpu_to_le16(qid); 90957dacad5SJay Sternberg 9101c63dc66SChristoph Hellwig return nvme_submit_sync_cmd(dev->ctrl.admin_q, &c, NULL, 0); 91157dacad5SJay Sternberg } 91257dacad5SJay Sternberg 91357dacad5SJay Sternberg static int adapter_delete_cq(struct nvme_dev *dev, u16 cqid) 91457dacad5SJay Sternberg { 91557dacad5SJay Sternberg return adapter_delete_queue(dev, nvme_admin_delete_cq, cqid); 91657dacad5SJay Sternberg } 91757dacad5SJay Sternberg 91857dacad5SJay Sternberg static int adapter_delete_sq(struct nvme_dev *dev, u16 sqid) 91957dacad5SJay Sternberg { 92057dacad5SJay Sternberg return adapter_delete_queue(dev, nvme_admin_delete_sq, sqid); 92157dacad5SJay Sternberg } 92257dacad5SJay Sternberg 923e7a2a87dSChristoph Hellwig static void abort_endio(struct request *req, int error) 92457dacad5SJay Sternberg { 925f4800d6dSChristoph Hellwig struct nvme_iod *iod = blk_mq_rq_to_pdu(req); 926f4800d6dSChristoph Hellwig struct nvme_queue *nvmeq = iod->nvmeq; 927e7a2a87dSChristoph Hellwig u32 result = (u32)(uintptr_t)req->special; 928e7a2a87dSChristoph Hellwig u16 status = req->errors; 92957dacad5SJay Sternberg 930e7a2a87dSChristoph Hellwig dev_warn(nvmeq->q_dmadev, "Abort status:%x result:%x", status, result); 931e7a2a87dSChristoph Hellwig atomic_inc(&nvmeq->dev->ctrl.abort_limit); 93257dacad5SJay Sternberg 933e7a2a87dSChristoph Hellwig blk_mq_free_request(req); 93457dacad5SJay Sternberg } 93557dacad5SJay Sternberg 93631c7c7d2SChristoph Hellwig static enum blk_eh_timer_return nvme_timeout(struct request *req, bool reserved) 93757dacad5SJay Sternberg { 938f4800d6dSChristoph Hellwig struct nvme_iod *iod = blk_mq_rq_to_pdu(req); 939f4800d6dSChristoph Hellwig struct nvme_queue *nvmeq = iod->nvmeq; 94057dacad5SJay Sternberg struct nvme_dev *dev = nvmeq->dev; 94157dacad5SJay Sternberg struct request *abort_req; 94257dacad5SJay Sternberg struct nvme_command cmd; 94357dacad5SJay Sternberg 94431c7c7d2SChristoph Hellwig /* 945fd634f41SChristoph Hellwig * Shutdown immediately if controller times out while starting. The 946fd634f41SChristoph Hellwig * reset work will see the pci device disabled when it gets the forced 947fd634f41SChristoph Hellwig * cancellation error. All outstanding requests are completed on 948fd634f41SChristoph Hellwig * shutdown, so we return BLK_EH_HANDLED. 949fd634f41SChristoph Hellwig */ 950fd634f41SChristoph Hellwig if (test_bit(NVME_CTRL_RESETTING, &dev->flags)) { 951fd634f41SChristoph Hellwig dev_warn(dev->dev, 952fd634f41SChristoph Hellwig "I/O %d QID %d timeout, disable controller\n", 953fd634f41SChristoph Hellwig req->tag, nvmeq->qid); 954a5cdb68cSKeith Busch nvme_dev_disable(dev, false); 955fd634f41SChristoph Hellwig req->errors = NVME_SC_CANCELLED; 956fd634f41SChristoph Hellwig return BLK_EH_HANDLED; 957fd634f41SChristoph Hellwig } 958fd634f41SChristoph Hellwig 959fd634f41SChristoph Hellwig /* 960e1569a16SKeith Busch * Shutdown the controller immediately and schedule a reset if the 961e1569a16SKeith Busch * command was already aborted once before and still hasn't been 962e1569a16SKeith Busch * returned to the driver, or if this is the admin queue. 96331c7c7d2SChristoph Hellwig */ 964f4800d6dSChristoph Hellwig if (!nvmeq->qid || iod->aborted) { 96557dacad5SJay Sternberg dev_warn(dev->dev, 96657dacad5SJay Sternberg "I/O %d QID %d timeout, reset controller\n", 96757dacad5SJay Sternberg req->tag, nvmeq->qid); 968a5cdb68cSKeith Busch nvme_dev_disable(dev, false); 969e1569a16SKeith Busch queue_work(nvme_workq, &dev->reset_work); 970e1569a16SKeith Busch 971e1569a16SKeith Busch /* 972e1569a16SKeith Busch * Mark the request as handled, since the inline shutdown 973e1569a16SKeith Busch * forces all outstanding requests to complete. 974e1569a16SKeith Busch */ 975e1569a16SKeith Busch req->errors = NVME_SC_CANCELLED; 976e1569a16SKeith Busch return BLK_EH_HANDLED; 97757dacad5SJay Sternberg } 97857dacad5SJay Sternberg 979f4800d6dSChristoph Hellwig iod->aborted = 1; 98057dacad5SJay Sternberg 981e7a2a87dSChristoph Hellwig if (atomic_dec_return(&dev->ctrl.abort_limit) < 0) { 982e7a2a87dSChristoph Hellwig atomic_inc(&dev->ctrl.abort_limit); 983e7a2a87dSChristoph Hellwig return BLK_EH_RESET_TIMER; 984e7a2a87dSChristoph Hellwig } 98557dacad5SJay Sternberg 98657dacad5SJay Sternberg memset(&cmd, 0, sizeof(cmd)); 98757dacad5SJay Sternberg cmd.abort.opcode = nvme_admin_abort_cmd; 98857dacad5SJay Sternberg cmd.abort.cid = req->tag; 98957dacad5SJay Sternberg cmd.abort.sqid = cpu_to_le16(nvmeq->qid); 99057dacad5SJay Sternberg 991e7a2a87dSChristoph Hellwig dev_warn(nvmeq->q_dmadev, "I/O %d QID %d timeout, aborting\n", 99257dacad5SJay Sternberg req->tag, nvmeq->qid); 993e7a2a87dSChristoph Hellwig 994e7a2a87dSChristoph Hellwig abort_req = nvme_alloc_request(dev->ctrl.admin_q, &cmd, 9956f3b0e8bSChristoph Hellwig BLK_MQ_REQ_NOWAIT); 9966bf25d16SChristoph Hellwig if (IS_ERR(abort_req)) { 9976bf25d16SChristoph Hellwig atomic_inc(&dev->ctrl.abort_limit); 99831c7c7d2SChristoph Hellwig return BLK_EH_RESET_TIMER; 99957dacad5SJay Sternberg } 100057dacad5SJay Sternberg 1001e7a2a87dSChristoph Hellwig abort_req->timeout = ADMIN_TIMEOUT; 1002e7a2a87dSChristoph Hellwig abort_req->end_io_data = NULL; 1003e7a2a87dSChristoph Hellwig blk_execute_rq_nowait(abort_req->q, NULL, abort_req, 0, abort_endio); 100457dacad5SJay Sternberg 100557dacad5SJay Sternberg /* 100657dacad5SJay Sternberg * The aborted req will be completed on receiving the abort req. 100757dacad5SJay Sternberg * We enable the timer again. If hit twice, it'll cause a device reset, 100857dacad5SJay Sternberg * as the device then is in a faulty state. 100957dacad5SJay Sternberg */ 101057dacad5SJay Sternberg return BLK_EH_RESET_TIMER; 101157dacad5SJay Sternberg } 101257dacad5SJay Sternberg 101357dacad5SJay Sternberg static void nvme_cancel_queue_ios(struct request *req, void *data, bool reserved) 101457dacad5SJay Sternberg { 101557dacad5SJay Sternberg struct nvme_queue *nvmeq = data; 1016aae239e1SChristoph Hellwig int status; 101757dacad5SJay Sternberg 101857dacad5SJay Sternberg if (!blk_mq_request_started(req)) 101957dacad5SJay Sternberg return; 102057dacad5SJay Sternberg 1021f8e68a7cSKeith Busch dev_dbg_ratelimited(nvmeq->q_dmadev, 1022aae239e1SChristoph Hellwig "Cancelling I/O %d QID %d\n", req->tag, nvmeq->qid); 102357dacad5SJay Sternberg 10241d49c38cSKeith Busch status = NVME_SC_ABORT_REQ; 102557dacad5SJay Sternberg if (blk_queue_dying(req->q)) 1026aae239e1SChristoph Hellwig status |= NVME_SC_DNR; 1027aae239e1SChristoph Hellwig blk_mq_complete_request(req, status); 102857dacad5SJay Sternberg } 102957dacad5SJay Sternberg 103057dacad5SJay Sternberg static void nvme_free_queue(struct nvme_queue *nvmeq) 103157dacad5SJay Sternberg { 103257dacad5SJay Sternberg dma_free_coherent(nvmeq->q_dmadev, CQ_SIZE(nvmeq->q_depth), 103357dacad5SJay Sternberg (void *)nvmeq->cqes, nvmeq->cq_dma_addr); 103457dacad5SJay Sternberg if (nvmeq->sq_cmds) 103557dacad5SJay Sternberg dma_free_coherent(nvmeq->q_dmadev, SQ_SIZE(nvmeq->q_depth), 103657dacad5SJay Sternberg nvmeq->sq_cmds, nvmeq->sq_dma_addr); 103757dacad5SJay Sternberg kfree(nvmeq); 103857dacad5SJay Sternberg } 103957dacad5SJay Sternberg 104057dacad5SJay Sternberg static void nvme_free_queues(struct nvme_dev *dev, int lowest) 104157dacad5SJay Sternberg { 104257dacad5SJay Sternberg int i; 104357dacad5SJay Sternberg 104457dacad5SJay Sternberg for (i = dev->queue_count - 1; i >= lowest; i--) { 104557dacad5SJay Sternberg struct nvme_queue *nvmeq = dev->queues[i]; 104657dacad5SJay Sternberg dev->queue_count--; 104757dacad5SJay Sternberg dev->queues[i] = NULL; 104857dacad5SJay Sternberg nvme_free_queue(nvmeq); 104957dacad5SJay Sternberg } 105057dacad5SJay Sternberg } 105157dacad5SJay Sternberg 105257dacad5SJay Sternberg /** 105357dacad5SJay Sternberg * nvme_suspend_queue - put queue into suspended state 105457dacad5SJay Sternberg * @nvmeq - queue to suspend 105557dacad5SJay Sternberg */ 105657dacad5SJay Sternberg static int nvme_suspend_queue(struct nvme_queue *nvmeq) 105757dacad5SJay Sternberg { 105857dacad5SJay Sternberg int vector; 105957dacad5SJay Sternberg 106057dacad5SJay Sternberg spin_lock_irq(&nvmeq->q_lock); 106157dacad5SJay Sternberg if (nvmeq->cq_vector == -1) { 106257dacad5SJay Sternberg spin_unlock_irq(&nvmeq->q_lock); 106357dacad5SJay Sternberg return 1; 106457dacad5SJay Sternberg } 106557dacad5SJay Sternberg vector = nvmeq->dev->entry[nvmeq->cq_vector].vector; 106657dacad5SJay Sternberg nvmeq->dev->online_queues--; 106757dacad5SJay Sternberg nvmeq->cq_vector = -1; 106857dacad5SJay Sternberg spin_unlock_irq(&nvmeq->q_lock); 106957dacad5SJay Sternberg 10701c63dc66SChristoph Hellwig if (!nvmeq->qid && nvmeq->dev->ctrl.admin_q) 107125646264SKeith Busch blk_mq_stop_hw_queues(nvmeq->dev->ctrl.admin_q); 107257dacad5SJay Sternberg 107357dacad5SJay Sternberg irq_set_affinity_hint(vector, NULL); 107457dacad5SJay Sternberg free_irq(vector, nvmeq); 107557dacad5SJay Sternberg 107657dacad5SJay Sternberg return 0; 107757dacad5SJay Sternberg } 107857dacad5SJay Sternberg 107957dacad5SJay Sternberg static void nvme_clear_queue(struct nvme_queue *nvmeq) 108057dacad5SJay Sternberg { 108157dacad5SJay Sternberg spin_lock_irq(&nvmeq->q_lock); 108257dacad5SJay Sternberg if (nvmeq->tags && *nvmeq->tags) 108357dacad5SJay Sternberg blk_mq_all_tag_busy_iter(*nvmeq->tags, nvme_cancel_queue_ios, nvmeq); 108457dacad5SJay Sternberg spin_unlock_irq(&nvmeq->q_lock); 108557dacad5SJay Sternberg } 108657dacad5SJay Sternberg 1087a5cdb68cSKeith Busch static void nvme_disable_admin_queue(struct nvme_dev *dev, bool shutdown) 108857dacad5SJay Sternberg { 1089a5cdb68cSKeith Busch struct nvme_queue *nvmeq = dev->queues[0]; 109057dacad5SJay Sternberg 109157dacad5SJay Sternberg if (!nvmeq) 109257dacad5SJay Sternberg return; 109357dacad5SJay Sternberg if (nvme_suspend_queue(nvmeq)) 109457dacad5SJay Sternberg return; 109557dacad5SJay Sternberg 1096a5cdb68cSKeith Busch if (shutdown) 1097a5cdb68cSKeith Busch nvme_shutdown_ctrl(&dev->ctrl); 1098a5cdb68cSKeith Busch else 1099a5cdb68cSKeith Busch nvme_disable_ctrl(&dev->ctrl, lo_hi_readq( 1100a5cdb68cSKeith Busch dev->bar + NVME_REG_CAP)); 110157dacad5SJay Sternberg 110257dacad5SJay Sternberg spin_lock_irq(&nvmeq->q_lock); 110357dacad5SJay Sternberg nvme_process_cq(nvmeq); 110457dacad5SJay Sternberg spin_unlock_irq(&nvmeq->q_lock); 110557dacad5SJay Sternberg } 110657dacad5SJay Sternberg 110757dacad5SJay Sternberg static int nvme_cmb_qdepth(struct nvme_dev *dev, int nr_io_queues, 110857dacad5SJay Sternberg int entry_size) 110957dacad5SJay Sternberg { 111057dacad5SJay Sternberg int q_depth = dev->q_depth; 11115fd4ce1bSChristoph Hellwig unsigned q_size_aligned = roundup(q_depth * entry_size, 11125fd4ce1bSChristoph Hellwig dev->ctrl.page_size); 111357dacad5SJay Sternberg 111457dacad5SJay Sternberg if (q_size_aligned * nr_io_queues > dev->cmb_size) { 111557dacad5SJay Sternberg u64 mem_per_q = div_u64(dev->cmb_size, nr_io_queues); 11165fd4ce1bSChristoph Hellwig mem_per_q = round_down(mem_per_q, dev->ctrl.page_size); 111757dacad5SJay Sternberg q_depth = div_u64(mem_per_q, entry_size); 111857dacad5SJay Sternberg 111957dacad5SJay Sternberg /* 112057dacad5SJay Sternberg * Ensure the reduced q_depth is above some threshold where it 112157dacad5SJay Sternberg * would be better to map queues in system memory with the 112257dacad5SJay Sternberg * original depth 112357dacad5SJay Sternberg */ 112457dacad5SJay Sternberg if (q_depth < 64) 112557dacad5SJay Sternberg return -ENOMEM; 112657dacad5SJay Sternberg } 112757dacad5SJay Sternberg 112857dacad5SJay Sternberg return q_depth; 112957dacad5SJay Sternberg } 113057dacad5SJay Sternberg 113157dacad5SJay Sternberg static int nvme_alloc_sq_cmds(struct nvme_dev *dev, struct nvme_queue *nvmeq, 113257dacad5SJay Sternberg int qid, int depth) 113357dacad5SJay Sternberg { 113457dacad5SJay Sternberg if (qid && dev->cmb && use_cmb_sqes && NVME_CMB_SQS(dev->cmbsz)) { 11355fd4ce1bSChristoph Hellwig unsigned offset = (qid - 1) * roundup(SQ_SIZE(depth), 11365fd4ce1bSChristoph Hellwig dev->ctrl.page_size); 113757dacad5SJay Sternberg nvmeq->sq_dma_addr = dev->cmb_dma_addr + offset; 113857dacad5SJay Sternberg nvmeq->sq_cmds_io = dev->cmb + offset; 113957dacad5SJay Sternberg } else { 114057dacad5SJay Sternberg nvmeq->sq_cmds = dma_alloc_coherent(dev->dev, SQ_SIZE(depth), 114157dacad5SJay Sternberg &nvmeq->sq_dma_addr, GFP_KERNEL); 114257dacad5SJay Sternberg if (!nvmeq->sq_cmds) 114357dacad5SJay Sternberg return -ENOMEM; 114457dacad5SJay Sternberg } 114557dacad5SJay Sternberg 114657dacad5SJay Sternberg return 0; 114757dacad5SJay Sternberg } 114857dacad5SJay Sternberg 114957dacad5SJay Sternberg static struct nvme_queue *nvme_alloc_queue(struct nvme_dev *dev, int qid, 115057dacad5SJay Sternberg int depth) 115157dacad5SJay Sternberg { 115257dacad5SJay Sternberg struct nvme_queue *nvmeq = kzalloc(sizeof(*nvmeq), GFP_KERNEL); 115357dacad5SJay Sternberg if (!nvmeq) 115457dacad5SJay Sternberg return NULL; 115557dacad5SJay Sternberg 115657dacad5SJay Sternberg nvmeq->cqes = dma_zalloc_coherent(dev->dev, CQ_SIZE(depth), 115757dacad5SJay Sternberg &nvmeq->cq_dma_addr, GFP_KERNEL); 115857dacad5SJay Sternberg if (!nvmeq->cqes) 115957dacad5SJay Sternberg goto free_nvmeq; 116057dacad5SJay Sternberg 116157dacad5SJay Sternberg if (nvme_alloc_sq_cmds(dev, nvmeq, qid, depth)) 116257dacad5SJay Sternberg goto free_cqdma; 116357dacad5SJay Sternberg 116457dacad5SJay Sternberg nvmeq->q_dmadev = dev->dev; 116557dacad5SJay Sternberg nvmeq->dev = dev; 116657dacad5SJay Sternberg snprintf(nvmeq->irqname, sizeof(nvmeq->irqname), "nvme%dq%d", 11671c63dc66SChristoph Hellwig dev->ctrl.instance, qid); 116857dacad5SJay Sternberg spin_lock_init(&nvmeq->q_lock); 116957dacad5SJay Sternberg nvmeq->cq_head = 0; 117057dacad5SJay Sternberg nvmeq->cq_phase = 1; 117157dacad5SJay Sternberg nvmeq->q_db = &dev->dbs[qid * 2 * dev->db_stride]; 117257dacad5SJay Sternberg nvmeq->q_depth = depth; 117357dacad5SJay Sternberg nvmeq->qid = qid; 117457dacad5SJay Sternberg nvmeq->cq_vector = -1; 117557dacad5SJay Sternberg dev->queues[qid] = nvmeq; 117657dacad5SJay Sternberg 117757dacad5SJay Sternberg /* make sure queue descriptor is set before queue count, for kthread */ 117857dacad5SJay Sternberg mb(); 117957dacad5SJay Sternberg dev->queue_count++; 118057dacad5SJay Sternberg 118157dacad5SJay Sternberg return nvmeq; 118257dacad5SJay Sternberg 118357dacad5SJay Sternberg free_cqdma: 118457dacad5SJay Sternberg dma_free_coherent(dev->dev, CQ_SIZE(depth), (void *)nvmeq->cqes, 118557dacad5SJay Sternberg nvmeq->cq_dma_addr); 118657dacad5SJay Sternberg free_nvmeq: 118757dacad5SJay Sternberg kfree(nvmeq); 118857dacad5SJay Sternberg return NULL; 118957dacad5SJay Sternberg } 119057dacad5SJay Sternberg 119157dacad5SJay Sternberg static int queue_request_irq(struct nvme_dev *dev, struct nvme_queue *nvmeq, 119257dacad5SJay Sternberg const char *name) 119357dacad5SJay Sternberg { 119457dacad5SJay Sternberg if (use_threaded_interrupts) 119557dacad5SJay Sternberg return request_threaded_irq(dev->entry[nvmeq->cq_vector].vector, 119657dacad5SJay Sternberg nvme_irq_check, nvme_irq, IRQF_SHARED, 119757dacad5SJay Sternberg name, nvmeq); 119857dacad5SJay Sternberg return request_irq(dev->entry[nvmeq->cq_vector].vector, nvme_irq, 119957dacad5SJay Sternberg IRQF_SHARED, name, nvmeq); 120057dacad5SJay Sternberg } 120157dacad5SJay Sternberg 120257dacad5SJay Sternberg static void nvme_init_queue(struct nvme_queue *nvmeq, u16 qid) 120357dacad5SJay Sternberg { 120457dacad5SJay Sternberg struct nvme_dev *dev = nvmeq->dev; 120557dacad5SJay Sternberg 120657dacad5SJay Sternberg spin_lock_irq(&nvmeq->q_lock); 120757dacad5SJay Sternberg nvmeq->sq_tail = 0; 120857dacad5SJay Sternberg nvmeq->cq_head = 0; 120957dacad5SJay Sternberg nvmeq->cq_phase = 1; 121057dacad5SJay Sternberg nvmeq->q_db = &dev->dbs[qid * 2 * dev->db_stride]; 121157dacad5SJay Sternberg memset((void *)nvmeq->cqes, 0, CQ_SIZE(nvmeq->q_depth)); 121257dacad5SJay Sternberg dev->online_queues++; 121357dacad5SJay Sternberg spin_unlock_irq(&nvmeq->q_lock); 121457dacad5SJay Sternberg } 121557dacad5SJay Sternberg 121657dacad5SJay Sternberg static int nvme_create_queue(struct nvme_queue *nvmeq, int qid) 121757dacad5SJay Sternberg { 121857dacad5SJay Sternberg struct nvme_dev *dev = nvmeq->dev; 121957dacad5SJay Sternberg int result; 122057dacad5SJay Sternberg 122157dacad5SJay Sternberg nvmeq->cq_vector = qid - 1; 122257dacad5SJay Sternberg result = adapter_alloc_cq(dev, qid, nvmeq); 122357dacad5SJay Sternberg if (result < 0) 122457dacad5SJay Sternberg return result; 122557dacad5SJay Sternberg 122657dacad5SJay Sternberg result = adapter_alloc_sq(dev, qid, nvmeq); 122757dacad5SJay Sternberg if (result < 0) 122857dacad5SJay Sternberg goto release_cq; 122957dacad5SJay Sternberg 123057dacad5SJay Sternberg result = queue_request_irq(dev, nvmeq, nvmeq->irqname); 123157dacad5SJay Sternberg if (result < 0) 123257dacad5SJay Sternberg goto release_sq; 123357dacad5SJay Sternberg 123457dacad5SJay Sternberg nvme_init_queue(nvmeq, qid); 123557dacad5SJay Sternberg return result; 123657dacad5SJay Sternberg 123757dacad5SJay Sternberg release_sq: 123857dacad5SJay Sternberg adapter_delete_sq(dev, qid); 123957dacad5SJay Sternberg release_cq: 124057dacad5SJay Sternberg adapter_delete_cq(dev, qid); 124157dacad5SJay Sternberg return result; 124257dacad5SJay Sternberg } 124357dacad5SJay Sternberg 124457dacad5SJay Sternberg static struct blk_mq_ops nvme_mq_admin_ops = { 124557dacad5SJay Sternberg .queue_rq = nvme_queue_rq, 1246eee417b0SChristoph Hellwig .complete = nvme_complete_rq, 124757dacad5SJay Sternberg .map_queue = blk_mq_map_queue, 124857dacad5SJay Sternberg .init_hctx = nvme_admin_init_hctx, 124957dacad5SJay Sternberg .exit_hctx = nvme_admin_exit_hctx, 125057dacad5SJay Sternberg .init_request = nvme_admin_init_request, 125157dacad5SJay Sternberg .timeout = nvme_timeout, 125257dacad5SJay Sternberg }; 125357dacad5SJay Sternberg 125457dacad5SJay Sternberg static struct blk_mq_ops nvme_mq_ops = { 125557dacad5SJay Sternberg .queue_rq = nvme_queue_rq, 1256eee417b0SChristoph Hellwig .complete = nvme_complete_rq, 125757dacad5SJay Sternberg .map_queue = blk_mq_map_queue, 125857dacad5SJay Sternberg .init_hctx = nvme_init_hctx, 125957dacad5SJay Sternberg .init_request = nvme_init_request, 126057dacad5SJay Sternberg .timeout = nvme_timeout, 1261a0fa9647SJens Axboe .poll = nvme_poll, 126257dacad5SJay Sternberg }; 126357dacad5SJay Sternberg 126457dacad5SJay Sternberg static void nvme_dev_remove_admin(struct nvme_dev *dev) 126557dacad5SJay Sternberg { 12661c63dc66SChristoph Hellwig if (dev->ctrl.admin_q && !blk_queue_dying(dev->ctrl.admin_q)) { 126769d9a99cSKeith Busch /* 126869d9a99cSKeith Busch * If the controller was reset during removal, it's possible 126969d9a99cSKeith Busch * user requests may be waiting on a stopped queue. Start the 127069d9a99cSKeith Busch * queue to flush these to completion. 127169d9a99cSKeith Busch */ 127269d9a99cSKeith Busch blk_mq_start_stopped_hw_queues(dev->ctrl.admin_q, true); 12731c63dc66SChristoph Hellwig blk_cleanup_queue(dev->ctrl.admin_q); 127457dacad5SJay Sternberg blk_mq_free_tag_set(&dev->admin_tagset); 127557dacad5SJay Sternberg } 127657dacad5SJay Sternberg } 127757dacad5SJay Sternberg 127857dacad5SJay Sternberg static int nvme_alloc_admin_tags(struct nvme_dev *dev) 127957dacad5SJay Sternberg { 12801c63dc66SChristoph Hellwig if (!dev->ctrl.admin_q) { 128157dacad5SJay Sternberg dev->admin_tagset.ops = &nvme_mq_admin_ops; 128257dacad5SJay Sternberg dev->admin_tagset.nr_hw_queues = 1; 1283e3e9d50cSKeith Busch 1284e3e9d50cSKeith Busch /* 1285e3e9d50cSKeith Busch * Subtract one to leave an empty queue entry for 'Full Queue' 1286e3e9d50cSKeith Busch * condition. See NVM-Express 1.2 specification, section 4.1.2. 1287e3e9d50cSKeith Busch */ 1288e3e9d50cSKeith Busch dev->admin_tagset.queue_depth = NVME_AQ_BLKMQ_DEPTH - 1; 128957dacad5SJay Sternberg dev->admin_tagset.timeout = ADMIN_TIMEOUT; 129057dacad5SJay Sternberg dev->admin_tagset.numa_node = dev_to_node(dev->dev); 129157dacad5SJay Sternberg dev->admin_tagset.cmd_size = nvme_cmd_size(dev); 129257dacad5SJay Sternberg dev->admin_tagset.driver_data = dev; 129357dacad5SJay Sternberg 129457dacad5SJay Sternberg if (blk_mq_alloc_tag_set(&dev->admin_tagset)) 129557dacad5SJay Sternberg return -ENOMEM; 129657dacad5SJay Sternberg 12971c63dc66SChristoph Hellwig dev->ctrl.admin_q = blk_mq_init_queue(&dev->admin_tagset); 12981c63dc66SChristoph Hellwig if (IS_ERR(dev->ctrl.admin_q)) { 129957dacad5SJay Sternberg blk_mq_free_tag_set(&dev->admin_tagset); 130057dacad5SJay Sternberg return -ENOMEM; 130157dacad5SJay Sternberg } 13021c63dc66SChristoph Hellwig if (!blk_get_queue(dev->ctrl.admin_q)) { 130357dacad5SJay Sternberg nvme_dev_remove_admin(dev); 13041c63dc66SChristoph Hellwig dev->ctrl.admin_q = NULL; 130557dacad5SJay Sternberg return -ENODEV; 130657dacad5SJay Sternberg } 130757dacad5SJay Sternberg } else 130825646264SKeith Busch blk_mq_start_stopped_hw_queues(dev->ctrl.admin_q, true); 130957dacad5SJay Sternberg 131057dacad5SJay Sternberg return 0; 131157dacad5SJay Sternberg } 131257dacad5SJay Sternberg 131357dacad5SJay Sternberg static int nvme_configure_admin_queue(struct nvme_dev *dev) 131457dacad5SJay Sternberg { 131557dacad5SJay Sternberg int result; 131657dacad5SJay Sternberg u32 aqa; 13177a67cbeaSChristoph Hellwig u64 cap = lo_hi_readq(dev->bar + NVME_REG_CAP); 131857dacad5SJay Sternberg struct nvme_queue *nvmeq; 131957dacad5SJay Sternberg 13207a67cbeaSChristoph Hellwig dev->subsystem = readl(dev->bar + NVME_REG_VS) >= NVME_VS(1, 1) ? 132157dacad5SJay Sternberg NVME_CAP_NSSRC(cap) : 0; 132257dacad5SJay Sternberg 13237a67cbeaSChristoph Hellwig if (dev->subsystem && 13247a67cbeaSChristoph Hellwig (readl(dev->bar + NVME_REG_CSTS) & NVME_CSTS_NSSRO)) 13257a67cbeaSChristoph Hellwig writel(NVME_CSTS_NSSRO, dev->bar + NVME_REG_CSTS); 132657dacad5SJay Sternberg 13275fd4ce1bSChristoph Hellwig result = nvme_disable_ctrl(&dev->ctrl, cap); 132857dacad5SJay Sternberg if (result < 0) 132957dacad5SJay Sternberg return result; 133057dacad5SJay Sternberg 133157dacad5SJay Sternberg nvmeq = dev->queues[0]; 133257dacad5SJay Sternberg if (!nvmeq) { 133357dacad5SJay Sternberg nvmeq = nvme_alloc_queue(dev, 0, NVME_AQ_DEPTH); 133457dacad5SJay Sternberg if (!nvmeq) 133557dacad5SJay Sternberg return -ENOMEM; 133657dacad5SJay Sternberg } 133757dacad5SJay Sternberg 133857dacad5SJay Sternberg aqa = nvmeq->q_depth - 1; 133957dacad5SJay Sternberg aqa |= aqa << 16; 134057dacad5SJay Sternberg 13417a67cbeaSChristoph Hellwig writel(aqa, dev->bar + NVME_REG_AQA); 13427a67cbeaSChristoph Hellwig lo_hi_writeq(nvmeq->sq_dma_addr, dev->bar + NVME_REG_ASQ); 13437a67cbeaSChristoph Hellwig lo_hi_writeq(nvmeq->cq_dma_addr, dev->bar + NVME_REG_ACQ); 134457dacad5SJay Sternberg 13455fd4ce1bSChristoph Hellwig result = nvme_enable_ctrl(&dev->ctrl, cap); 134657dacad5SJay Sternberg if (result) 134757dacad5SJay Sternberg goto free_nvmeq; 134857dacad5SJay Sternberg 134957dacad5SJay Sternberg nvmeq->cq_vector = 0; 135057dacad5SJay Sternberg result = queue_request_irq(dev, nvmeq, nvmeq->irqname); 135157dacad5SJay Sternberg if (result) { 135257dacad5SJay Sternberg nvmeq->cq_vector = -1; 135357dacad5SJay Sternberg goto free_nvmeq; 135457dacad5SJay Sternberg } 135557dacad5SJay Sternberg 135657dacad5SJay Sternberg return result; 135757dacad5SJay Sternberg 135857dacad5SJay Sternberg free_nvmeq: 135957dacad5SJay Sternberg nvme_free_queues(dev, 0); 136057dacad5SJay Sternberg return result; 136157dacad5SJay Sternberg } 136257dacad5SJay Sternberg 136357dacad5SJay Sternberg static int nvme_kthread(void *data) 136457dacad5SJay Sternberg { 136557dacad5SJay Sternberg struct nvme_dev *dev, *next; 136657dacad5SJay Sternberg 136757dacad5SJay Sternberg while (!kthread_should_stop()) { 136857dacad5SJay Sternberg set_current_state(TASK_INTERRUPTIBLE); 136957dacad5SJay Sternberg spin_lock(&dev_list_lock); 137057dacad5SJay Sternberg list_for_each_entry_safe(dev, next, &dev_list, node) { 137157dacad5SJay Sternberg int i; 13727a67cbeaSChristoph Hellwig u32 csts = readl(dev->bar + NVME_REG_CSTS); 137357dacad5SJay Sternberg 1374846cc05fSChristoph Hellwig /* 1375846cc05fSChristoph Hellwig * Skip controllers currently under reset. 1376846cc05fSChristoph Hellwig */ 1377846cc05fSChristoph Hellwig if (work_pending(&dev->reset_work) || work_busy(&dev->reset_work)) 1378846cc05fSChristoph Hellwig continue; 137957dacad5SJay Sternberg 138057dacad5SJay Sternberg if ((dev->subsystem && (csts & NVME_CSTS_NSSRO)) || 138157dacad5SJay Sternberg csts & NVME_CSTS_CFS) { 1382846cc05fSChristoph Hellwig if (queue_work(nvme_workq, &dev->reset_work)) { 138357dacad5SJay Sternberg dev_warn(dev->dev, 138457dacad5SJay Sternberg "Failed status: %x, reset controller\n", 13857a67cbeaSChristoph Hellwig readl(dev->bar + NVME_REG_CSTS)); 138657dacad5SJay Sternberg } 138757dacad5SJay Sternberg continue; 138857dacad5SJay Sternberg } 138957dacad5SJay Sternberg for (i = 0; i < dev->queue_count; i++) { 139057dacad5SJay Sternberg struct nvme_queue *nvmeq = dev->queues[i]; 139157dacad5SJay Sternberg if (!nvmeq) 139257dacad5SJay Sternberg continue; 139357dacad5SJay Sternberg spin_lock_irq(&nvmeq->q_lock); 139457dacad5SJay Sternberg nvme_process_cq(nvmeq); 139557dacad5SJay Sternberg 1396adf68f21SChristoph Hellwig while (i == 0 && dev->ctrl.event_limit > 0) 1397adf68f21SChristoph Hellwig nvme_submit_async_event(dev); 139857dacad5SJay Sternberg spin_unlock_irq(&nvmeq->q_lock); 139957dacad5SJay Sternberg } 140057dacad5SJay Sternberg } 140157dacad5SJay Sternberg spin_unlock(&dev_list_lock); 140257dacad5SJay Sternberg schedule_timeout(round_jiffies_relative(HZ)); 140357dacad5SJay Sternberg } 140457dacad5SJay Sternberg return 0; 140557dacad5SJay Sternberg } 140657dacad5SJay Sternberg 1407749941f2SChristoph Hellwig static int nvme_create_io_queues(struct nvme_dev *dev) 140857dacad5SJay Sternberg { 140957dacad5SJay Sternberg unsigned i; 1410749941f2SChristoph Hellwig int ret = 0; 141157dacad5SJay Sternberg 1412749941f2SChristoph Hellwig for (i = dev->queue_count; i <= dev->max_qid; i++) { 1413749941f2SChristoph Hellwig if (!nvme_alloc_queue(dev, i, dev->q_depth)) { 1414749941f2SChristoph Hellwig ret = -ENOMEM; 141557dacad5SJay Sternberg break; 1416749941f2SChristoph Hellwig } 1417749941f2SChristoph Hellwig } 141857dacad5SJay Sternberg 1419749941f2SChristoph Hellwig for (i = dev->online_queues; i <= dev->queue_count - 1; i++) { 1420749941f2SChristoph Hellwig ret = nvme_create_queue(dev->queues[i], i); 1421749941f2SChristoph Hellwig if (ret) { 142257dacad5SJay Sternberg nvme_free_queues(dev, i); 142357dacad5SJay Sternberg break; 142457dacad5SJay Sternberg } 142557dacad5SJay Sternberg } 142657dacad5SJay Sternberg 1427749941f2SChristoph Hellwig /* 1428749941f2SChristoph Hellwig * Ignore failing Create SQ/CQ commands, we can continue with less 1429749941f2SChristoph Hellwig * than the desired aount of queues, and even a controller without 1430749941f2SChristoph Hellwig * I/O queues an still be used to issue admin commands. This might 1431749941f2SChristoph Hellwig * be useful to upgrade a buggy firmware for example. 1432749941f2SChristoph Hellwig */ 1433749941f2SChristoph Hellwig return ret >= 0 ? 0 : ret; 143457dacad5SJay Sternberg } 143557dacad5SJay Sternberg 143657dacad5SJay Sternberg static void __iomem *nvme_map_cmb(struct nvme_dev *dev) 143757dacad5SJay Sternberg { 143857dacad5SJay Sternberg u64 szu, size, offset; 143957dacad5SJay Sternberg u32 cmbloc; 144057dacad5SJay Sternberg resource_size_t bar_size; 144157dacad5SJay Sternberg struct pci_dev *pdev = to_pci_dev(dev->dev); 144257dacad5SJay Sternberg void __iomem *cmb; 144357dacad5SJay Sternberg dma_addr_t dma_addr; 144457dacad5SJay Sternberg 144557dacad5SJay Sternberg if (!use_cmb_sqes) 144657dacad5SJay Sternberg return NULL; 144757dacad5SJay Sternberg 14487a67cbeaSChristoph Hellwig dev->cmbsz = readl(dev->bar + NVME_REG_CMBSZ); 144957dacad5SJay Sternberg if (!(NVME_CMB_SZ(dev->cmbsz))) 145057dacad5SJay Sternberg return NULL; 145157dacad5SJay Sternberg 14527a67cbeaSChristoph Hellwig cmbloc = readl(dev->bar + NVME_REG_CMBLOC); 145357dacad5SJay Sternberg 145457dacad5SJay Sternberg szu = (u64)1 << (12 + 4 * NVME_CMB_SZU(dev->cmbsz)); 145557dacad5SJay Sternberg size = szu * NVME_CMB_SZ(dev->cmbsz); 145657dacad5SJay Sternberg offset = szu * NVME_CMB_OFST(cmbloc); 145757dacad5SJay Sternberg bar_size = pci_resource_len(pdev, NVME_CMB_BIR(cmbloc)); 145857dacad5SJay Sternberg 145957dacad5SJay Sternberg if (offset > bar_size) 146057dacad5SJay Sternberg return NULL; 146157dacad5SJay Sternberg 146257dacad5SJay Sternberg /* 146357dacad5SJay Sternberg * Controllers may support a CMB size larger than their BAR, 146457dacad5SJay Sternberg * for example, due to being behind a bridge. Reduce the CMB to 146557dacad5SJay Sternberg * the reported size of the BAR 146657dacad5SJay Sternberg */ 146757dacad5SJay Sternberg if (size > bar_size - offset) 146857dacad5SJay Sternberg size = bar_size - offset; 146957dacad5SJay Sternberg 147057dacad5SJay Sternberg dma_addr = pci_resource_start(pdev, NVME_CMB_BIR(cmbloc)) + offset; 147157dacad5SJay Sternberg cmb = ioremap_wc(dma_addr, size); 147257dacad5SJay Sternberg if (!cmb) 147357dacad5SJay Sternberg return NULL; 147457dacad5SJay Sternberg 147557dacad5SJay Sternberg dev->cmb_dma_addr = dma_addr; 147657dacad5SJay Sternberg dev->cmb_size = size; 147757dacad5SJay Sternberg return cmb; 147857dacad5SJay Sternberg } 147957dacad5SJay Sternberg 148057dacad5SJay Sternberg static inline void nvme_release_cmb(struct nvme_dev *dev) 148157dacad5SJay Sternberg { 148257dacad5SJay Sternberg if (dev->cmb) { 148357dacad5SJay Sternberg iounmap(dev->cmb); 148457dacad5SJay Sternberg dev->cmb = NULL; 148557dacad5SJay Sternberg } 148657dacad5SJay Sternberg } 148757dacad5SJay Sternberg 148857dacad5SJay Sternberg static size_t db_bar_size(struct nvme_dev *dev, unsigned nr_io_queues) 148957dacad5SJay Sternberg { 149057dacad5SJay Sternberg return 4096 + ((nr_io_queues + 1) * 8 * dev->db_stride); 149157dacad5SJay Sternberg } 149257dacad5SJay Sternberg 149357dacad5SJay Sternberg static int nvme_setup_io_queues(struct nvme_dev *dev) 149457dacad5SJay Sternberg { 149557dacad5SJay Sternberg struct nvme_queue *adminq = dev->queues[0]; 149657dacad5SJay Sternberg struct pci_dev *pdev = to_pci_dev(dev->dev); 149757dacad5SJay Sternberg int result, i, vecs, nr_io_queues, size; 149857dacad5SJay Sternberg 149957dacad5SJay Sternberg nr_io_queues = num_possible_cpus(); 15009a0be7abSChristoph Hellwig result = nvme_set_queue_count(&dev->ctrl, &nr_io_queues); 15019a0be7abSChristoph Hellwig if (result < 0) 150257dacad5SJay Sternberg return result; 15039a0be7abSChristoph Hellwig 15049a0be7abSChristoph Hellwig /* 15059a0be7abSChristoph Hellwig * Degraded controllers might return an error when setting the queue 15069a0be7abSChristoph Hellwig * count. We still want to be able to bring them online and offer 15079a0be7abSChristoph Hellwig * access to the admin queue, as that might be only way to fix them up. 15089a0be7abSChristoph Hellwig */ 15099a0be7abSChristoph Hellwig if (result > 0) { 15109a0be7abSChristoph Hellwig dev_err(dev->dev, "Could not set queue count (%d)\n", result); 15119a0be7abSChristoph Hellwig nr_io_queues = 0; 15129a0be7abSChristoph Hellwig result = 0; 15139a0be7abSChristoph Hellwig } 151457dacad5SJay Sternberg 151557dacad5SJay Sternberg if (dev->cmb && NVME_CMB_SQS(dev->cmbsz)) { 151657dacad5SJay Sternberg result = nvme_cmb_qdepth(dev, nr_io_queues, 151757dacad5SJay Sternberg sizeof(struct nvme_command)); 151857dacad5SJay Sternberg if (result > 0) 151957dacad5SJay Sternberg dev->q_depth = result; 152057dacad5SJay Sternberg else 152157dacad5SJay Sternberg nvme_release_cmb(dev); 152257dacad5SJay Sternberg } 152357dacad5SJay Sternberg 152457dacad5SJay Sternberg size = db_bar_size(dev, nr_io_queues); 152557dacad5SJay Sternberg if (size > 8192) { 152657dacad5SJay Sternberg iounmap(dev->bar); 152757dacad5SJay Sternberg do { 152857dacad5SJay Sternberg dev->bar = ioremap(pci_resource_start(pdev, 0), size); 152957dacad5SJay Sternberg if (dev->bar) 153057dacad5SJay Sternberg break; 153157dacad5SJay Sternberg if (!--nr_io_queues) 153257dacad5SJay Sternberg return -ENOMEM; 153357dacad5SJay Sternberg size = db_bar_size(dev, nr_io_queues); 153457dacad5SJay Sternberg } while (1); 15357a67cbeaSChristoph Hellwig dev->dbs = dev->bar + 4096; 153657dacad5SJay Sternberg adminq->q_db = dev->dbs; 153757dacad5SJay Sternberg } 153857dacad5SJay Sternberg 153957dacad5SJay Sternberg /* Deregister the admin queue's interrupt */ 154057dacad5SJay Sternberg free_irq(dev->entry[0].vector, adminq); 154157dacad5SJay Sternberg 154257dacad5SJay Sternberg /* 154357dacad5SJay Sternberg * If we enable msix early due to not intx, disable it again before 154457dacad5SJay Sternberg * setting up the full range we need. 154557dacad5SJay Sternberg */ 154657dacad5SJay Sternberg if (!pdev->irq) 154757dacad5SJay Sternberg pci_disable_msix(pdev); 154857dacad5SJay Sternberg 154957dacad5SJay Sternberg for (i = 0; i < nr_io_queues; i++) 155057dacad5SJay Sternberg dev->entry[i].entry = i; 155157dacad5SJay Sternberg vecs = pci_enable_msix_range(pdev, dev->entry, 1, nr_io_queues); 155257dacad5SJay Sternberg if (vecs < 0) { 155357dacad5SJay Sternberg vecs = pci_enable_msi_range(pdev, 1, min(nr_io_queues, 32)); 155457dacad5SJay Sternberg if (vecs < 0) { 155557dacad5SJay Sternberg vecs = 1; 155657dacad5SJay Sternberg } else { 155757dacad5SJay Sternberg for (i = 0; i < vecs; i++) 155857dacad5SJay Sternberg dev->entry[i].vector = i + pdev->irq; 155957dacad5SJay Sternberg } 156057dacad5SJay Sternberg } 156157dacad5SJay Sternberg 156257dacad5SJay Sternberg /* 156357dacad5SJay Sternberg * Should investigate if there's a performance win from allocating 156457dacad5SJay Sternberg * more queues than interrupt vectors; it might allow the submission 156557dacad5SJay Sternberg * path to scale better, even if the receive path is limited by the 156657dacad5SJay Sternberg * number of interrupts. 156757dacad5SJay Sternberg */ 156857dacad5SJay Sternberg nr_io_queues = vecs; 156957dacad5SJay Sternberg dev->max_qid = nr_io_queues; 157057dacad5SJay Sternberg 157157dacad5SJay Sternberg result = queue_request_irq(dev, adminq, adminq->irqname); 157257dacad5SJay Sternberg if (result) { 157357dacad5SJay Sternberg adminq->cq_vector = -1; 157457dacad5SJay Sternberg goto free_queues; 157557dacad5SJay Sternberg } 157657dacad5SJay Sternberg 157757dacad5SJay Sternberg /* Free previously allocated queues that are no longer usable */ 157857dacad5SJay Sternberg nvme_free_queues(dev, nr_io_queues + 1); 1579749941f2SChristoph Hellwig return nvme_create_io_queues(dev); 158057dacad5SJay Sternberg 158157dacad5SJay Sternberg free_queues: 158257dacad5SJay Sternberg nvme_free_queues(dev, 1); 158357dacad5SJay Sternberg return result; 158457dacad5SJay Sternberg } 158557dacad5SJay Sternberg 158657dacad5SJay Sternberg static void nvme_set_irq_hints(struct nvme_dev *dev) 158757dacad5SJay Sternberg { 158857dacad5SJay Sternberg struct nvme_queue *nvmeq; 158957dacad5SJay Sternberg int i; 159057dacad5SJay Sternberg 159157dacad5SJay Sternberg for (i = 0; i < dev->online_queues; i++) { 159257dacad5SJay Sternberg nvmeq = dev->queues[i]; 159357dacad5SJay Sternberg 159457dacad5SJay Sternberg if (!nvmeq->tags || !(*nvmeq->tags)) 159557dacad5SJay Sternberg continue; 159657dacad5SJay Sternberg 159757dacad5SJay Sternberg irq_set_affinity_hint(dev->entry[nvmeq->cq_vector].vector, 159857dacad5SJay Sternberg blk_mq_tags_cpumask(*nvmeq->tags)); 159957dacad5SJay Sternberg } 160057dacad5SJay Sternberg } 160157dacad5SJay Sternberg 160257dacad5SJay Sternberg static void nvme_dev_scan(struct work_struct *work) 160357dacad5SJay Sternberg { 160457dacad5SJay Sternberg struct nvme_dev *dev = container_of(work, struct nvme_dev, scan_work); 160557dacad5SJay Sternberg 160657dacad5SJay Sternberg if (!dev->tagset.tags) 160757dacad5SJay Sternberg return; 16085bae7f73SChristoph Hellwig nvme_scan_namespaces(&dev->ctrl); 160957dacad5SJay Sternberg nvme_set_irq_hints(dev); 161057dacad5SJay Sternberg } 161157dacad5SJay Sternberg 1612db3cbfffSKeith Busch static void nvme_del_queue_end(struct request *req, int error) 1613db3cbfffSKeith Busch { 1614db3cbfffSKeith Busch struct nvme_queue *nvmeq = req->end_io_data; 1615db3cbfffSKeith Busch 1616db3cbfffSKeith Busch blk_mq_free_request(req); 1617db3cbfffSKeith Busch complete(&nvmeq->dev->ioq_wait); 1618db3cbfffSKeith Busch } 1619db3cbfffSKeith Busch 1620db3cbfffSKeith Busch static void nvme_del_cq_end(struct request *req, int error) 1621db3cbfffSKeith Busch { 1622db3cbfffSKeith Busch struct nvme_queue *nvmeq = req->end_io_data; 1623db3cbfffSKeith Busch 1624db3cbfffSKeith Busch if (!error) { 1625db3cbfffSKeith Busch unsigned long flags; 1626db3cbfffSKeith Busch 1627db3cbfffSKeith Busch spin_lock_irqsave(&nvmeq->q_lock, flags); 1628db3cbfffSKeith Busch nvme_process_cq(nvmeq); 1629db3cbfffSKeith Busch spin_unlock_irqrestore(&nvmeq->q_lock, flags); 1630db3cbfffSKeith Busch } 1631db3cbfffSKeith Busch 1632db3cbfffSKeith Busch nvme_del_queue_end(req, error); 1633db3cbfffSKeith Busch } 1634db3cbfffSKeith Busch 1635db3cbfffSKeith Busch static int nvme_delete_queue(struct nvme_queue *nvmeq, u8 opcode) 1636db3cbfffSKeith Busch { 1637db3cbfffSKeith Busch struct request_queue *q = nvmeq->dev->ctrl.admin_q; 1638db3cbfffSKeith Busch struct request *req; 1639db3cbfffSKeith Busch struct nvme_command cmd; 1640db3cbfffSKeith Busch 1641db3cbfffSKeith Busch memset(&cmd, 0, sizeof(cmd)); 1642db3cbfffSKeith Busch cmd.delete_queue.opcode = opcode; 1643db3cbfffSKeith Busch cmd.delete_queue.qid = cpu_to_le16(nvmeq->qid); 1644db3cbfffSKeith Busch 1645db3cbfffSKeith Busch req = nvme_alloc_request(q, &cmd, BLK_MQ_REQ_NOWAIT); 1646db3cbfffSKeith Busch if (IS_ERR(req)) 1647db3cbfffSKeith Busch return PTR_ERR(req); 1648db3cbfffSKeith Busch 1649db3cbfffSKeith Busch req->timeout = ADMIN_TIMEOUT; 1650db3cbfffSKeith Busch req->end_io_data = nvmeq; 1651db3cbfffSKeith Busch 1652db3cbfffSKeith Busch blk_execute_rq_nowait(q, NULL, req, false, 1653db3cbfffSKeith Busch opcode == nvme_admin_delete_cq ? 1654db3cbfffSKeith Busch nvme_del_cq_end : nvme_del_queue_end); 1655db3cbfffSKeith Busch return 0; 1656db3cbfffSKeith Busch } 1657db3cbfffSKeith Busch 1658db3cbfffSKeith Busch static void nvme_disable_io_queues(struct nvme_dev *dev) 1659db3cbfffSKeith Busch { 1660db3cbfffSKeith Busch int pass; 1661db3cbfffSKeith Busch unsigned long timeout; 1662db3cbfffSKeith Busch u8 opcode = nvme_admin_delete_sq; 1663db3cbfffSKeith Busch 1664db3cbfffSKeith Busch for (pass = 0; pass < 2; pass++) { 1665db3cbfffSKeith Busch int sent = 0, i = dev->queue_count - 1; 1666db3cbfffSKeith Busch 1667db3cbfffSKeith Busch reinit_completion(&dev->ioq_wait); 1668db3cbfffSKeith Busch retry: 1669db3cbfffSKeith Busch timeout = ADMIN_TIMEOUT; 1670db3cbfffSKeith Busch for (; i > 0; i--) { 1671db3cbfffSKeith Busch struct nvme_queue *nvmeq = dev->queues[i]; 1672db3cbfffSKeith Busch 1673db3cbfffSKeith Busch if (!pass) 1674db3cbfffSKeith Busch nvme_suspend_queue(nvmeq); 1675db3cbfffSKeith Busch if (nvme_delete_queue(nvmeq, opcode)) 1676db3cbfffSKeith Busch break; 1677db3cbfffSKeith Busch ++sent; 1678db3cbfffSKeith Busch } 1679db3cbfffSKeith Busch while (sent--) { 1680db3cbfffSKeith Busch timeout = wait_for_completion_io_timeout(&dev->ioq_wait, timeout); 1681db3cbfffSKeith Busch if (timeout == 0) 1682db3cbfffSKeith Busch return; 1683db3cbfffSKeith Busch if (i) 1684db3cbfffSKeith Busch goto retry; 1685db3cbfffSKeith Busch } 1686db3cbfffSKeith Busch opcode = nvme_admin_delete_cq; 1687db3cbfffSKeith Busch } 1688db3cbfffSKeith Busch } 1689db3cbfffSKeith Busch 169057dacad5SJay Sternberg /* 169157dacad5SJay Sternberg * Return: error value if an error occurred setting up the queues or calling 169257dacad5SJay Sternberg * Identify Device. 0 if these succeeded, even if adding some of the 169357dacad5SJay Sternberg * namespaces failed. At the moment, these failures are silent. TBD which 169457dacad5SJay Sternberg * failures should be reported. 169557dacad5SJay Sternberg */ 169657dacad5SJay Sternberg static int nvme_dev_add(struct nvme_dev *dev) 169757dacad5SJay Sternberg { 16985bae7f73SChristoph Hellwig if (!dev->ctrl.tagset) { 169957dacad5SJay Sternberg dev->tagset.ops = &nvme_mq_ops; 170057dacad5SJay Sternberg dev->tagset.nr_hw_queues = dev->online_queues - 1; 170157dacad5SJay Sternberg dev->tagset.timeout = NVME_IO_TIMEOUT; 170257dacad5SJay Sternberg dev->tagset.numa_node = dev_to_node(dev->dev); 170357dacad5SJay Sternberg dev->tagset.queue_depth = 170457dacad5SJay Sternberg min_t(int, dev->q_depth, BLK_MQ_MAX_DEPTH) - 1; 170557dacad5SJay Sternberg dev->tagset.cmd_size = nvme_cmd_size(dev); 170657dacad5SJay Sternberg dev->tagset.flags = BLK_MQ_F_SHOULD_MERGE; 170757dacad5SJay Sternberg dev->tagset.driver_data = dev; 170857dacad5SJay Sternberg 170957dacad5SJay Sternberg if (blk_mq_alloc_tag_set(&dev->tagset)) 171057dacad5SJay Sternberg return 0; 17115bae7f73SChristoph Hellwig dev->ctrl.tagset = &dev->tagset; 171257dacad5SJay Sternberg } 1713646017a6SKeith Busch nvme_queue_scan(dev); 171457dacad5SJay Sternberg return 0; 171557dacad5SJay Sternberg } 171657dacad5SJay Sternberg 1717b00a726aSKeith Busch static int nvme_pci_enable(struct nvme_dev *dev) 171857dacad5SJay Sternberg { 171957dacad5SJay Sternberg u64 cap; 1720b00a726aSKeith Busch int result = -ENOMEM; 172157dacad5SJay Sternberg struct pci_dev *pdev = to_pci_dev(dev->dev); 172257dacad5SJay Sternberg 172357dacad5SJay Sternberg if (pci_enable_device_mem(pdev)) 172457dacad5SJay Sternberg return result; 172557dacad5SJay Sternberg 172657dacad5SJay Sternberg dev->entry[0].vector = pdev->irq; 172757dacad5SJay Sternberg pci_set_master(pdev); 172857dacad5SJay Sternberg 172957dacad5SJay Sternberg if (dma_set_mask_and_coherent(dev->dev, DMA_BIT_MASK(64)) && 173057dacad5SJay Sternberg dma_set_mask_and_coherent(dev->dev, DMA_BIT_MASK(32))) 173157dacad5SJay Sternberg goto disable; 173257dacad5SJay Sternberg 17337a67cbeaSChristoph Hellwig if (readl(dev->bar + NVME_REG_CSTS) == -1) { 173457dacad5SJay Sternberg result = -ENODEV; 1735b00a726aSKeith Busch goto disable; 173657dacad5SJay Sternberg } 173757dacad5SJay Sternberg 173857dacad5SJay Sternberg /* 173957dacad5SJay Sternberg * Some devices don't advertse INTx interrupts, pre-enable a single 174057dacad5SJay Sternberg * MSIX vec for setup. We'll adjust this later. 174157dacad5SJay Sternberg */ 174257dacad5SJay Sternberg if (!pdev->irq) { 174357dacad5SJay Sternberg result = pci_enable_msix(pdev, dev->entry, 1); 174457dacad5SJay Sternberg if (result < 0) 1745b00a726aSKeith Busch goto disable; 174657dacad5SJay Sternberg } 174757dacad5SJay Sternberg 17487a67cbeaSChristoph Hellwig cap = lo_hi_readq(dev->bar + NVME_REG_CAP); 17497a67cbeaSChristoph Hellwig 175057dacad5SJay Sternberg dev->q_depth = min_t(int, NVME_CAP_MQES(cap) + 1, NVME_Q_DEPTH); 175157dacad5SJay Sternberg dev->db_stride = 1 << NVME_CAP_STRIDE(cap); 17527a67cbeaSChristoph Hellwig dev->dbs = dev->bar + 4096; 17531f390c1fSStephan Günther 17541f390c1fSStephan Günther /* 17551f390c1fSStephan Günther * Temporary fix for the Apple controller found in the MacBook8,1 and 17561f390c1fSStephan Günther * some MacBook7,1 to avoid controller resets and data loss. 17571f390c1fSStephan Günther */ 17581f390c1fSStephan Günther if (pdev->vendor == PCI_VENDOR_ID_APPLE && pdev->device == 0x2001) { 17591f390c1fSStephan Günther dev->q_depth = 2; 17601f390c1fSStephan Günther dev_warn(dev->dev, "detected Apple NVMe controller, set " 17611f390c1fSStephan Günther "queue depth=%u to work around controller resets\n", 17621f390c1fSStephan Günther dev->q_depth); 17631f390c1fSStephan Günther } 17641f390c1fSStephan Günther 17657a67cbeaSChristoph Hellwig if (readl(dev->bar + NVME_REG_VS) >= NVME_VS(1, 2)) 176657dacad5SJay Sternberg dev->cmb = nvme_map_cmb(dev); 176757dacad5SJay Sternberg 1768a0a3408eSKeith Busch pci_enable_pcie_error_reporting(pdev); 1769a0a3408eSKeith Busch pci_save_state(pdev); 177057dacad5SJay Sternberg return 0; 177157dacad5SJay Sternberg 177257dacad5SJay Sternberg disable: 177357dacad5SJay Sternberg pci_disable_device(pdev); 177457dacad5SJay Sternberg return result; 177557dacad5SJay Sternberg } 177657dacad5SJay Sternberg 177757dacad5SJay Sternberg static void nvme_dev_unmap(struct nvme_dev *dev) 177857dacad5SJay Sternberg { 1779b00a726aSKeith Busch if (dev->bar) 1780b00a726aSKeith Busch iounmap(dev->bar); 1781b00a726aSKeith Busch pci_release_regions(to_pci_dev(dev->dev)); 1782b00a726aSKeith Busch } 1783b00a726aSKeith Busch 1784b00a726aSKeith Busch static void nvme_pci_disable(struct nvme_dev *dev) 1785b00a726aSKeith Busch { 178657dacad5SJay Sternberg struct pci_dev *pdev = to_pci_dev(dev->dev); 178757dacad5SJay Sternberg 178857dacad5SJay Sternberg if (pdev->msi_enabled) 178957dacad5SJay Sternberg pci_disable_msi(pdev); 179057dacad5SJay Sternberg else if (pdev->msix_enabled) 179157dacad5SJay Sternberg pci_disable_msix(pdev); 179257dacad5SJay Sternberg 1793a0a3408eSKeith Busch if (pci_is_enabled(pdev)) { 1794a0a3408eSKeith Busch pci_disable_pcie_error_reporting(pdev); 179557dacad5SJay Sternberg pci_disable_device(pdev); 179657dacad5SJay Sternberg } 1797a0a3408eSKeith Busch } 179857dacad5SJay Sternberg 17997385014cSChristoph Hellwig static int nvme_dev_list_add(struct nvme_dev *dev) 180057dacad5SJay Sternberg { 18017385014cSChristoph Hellwig bool start_thread = false; 180257dacad5SJay Sternberg 18037385014cSChristoph Hellwig spin_lock(&dev_list_lock); 18047385014cSChristoph Hellwig if (list_empty(&dev_list) && IS_ERR_OR_NULL(nvme_thread)) { 18057385014cSChristoph Hellwig start_thread = true; 18067385014cSChristoph Hellwig nvme_thread = NULL; 180757dacad5SJay Sternberg } 18087385014cSChristoph Hellwig list_add(&dev->node, &dev_list); 18097385014cSChristoph Hellwig spin_unlock(&dev_list_lock); 181057dacad5SJay Sternberg 18117385014cSChristoph Hellwig if (start_thread) { 18127385014cSChristoph Hellwig nvme_thread = kthread_run(nvme_kthread, NULL, "nvme"); 18137385014cSChristoph Hellwig wake_up_all(&nvme_kthread_wait); 18147385014cSChristoph Hellwig } else 18157385014cSChristoph Hellwig wait_event_killable(nvme_kthread_wait, nvme_thread); 181657dacad5SJay Sternberg 18177385014cSChristoph Hellwig if (IS_ERR_OR_NULL(nvme_thread)) 18187385014cSChristoph Hellwig return nvme_thread ? PTR_ERR(nvme_thread) : -EINTR; 181957dacad5SJay Sternberg 18207385014cSChristoph Hellwig return 0; 182157dacad5SJay Sternberg } 182257dacad5SJay Sternberg 182357dacad5SJay Sternberg /* 182457dacad5SJay Sternberg * Remove the node from the device list and check 182557dacad5SJay Sternberg * for whether or not we need to stop the nvme_thread. 182657dacad5SJay Sternberg */ 182757dacad5SJay Sternberg static void nvme_dev_list_remove(struct nvme_dev *dev) 182857dacad5SJay Sternberg { 182957dacad5SJay Sternberg struct task_struct *tmp = NULL; 183057dacad5SJay Sternberg 183157dacad5SJay Sternberg spin_lock(&dev_list_lock); 183257dacad5SJay Sternberg list_del_init(&dev->node); 183357dacad5SJay Sternberg if (list_empty(&dev_list) && !IS_ERR_OR_NULL(nvme_thread)) { 183457dacad5SJay Sternberg tmp = nvme_thread; 183557dacad5SJay Sternberg nvme_thread = NULL; 183657dacad5SJay Sternberg } 183757dacad5SJay Sternberg spin_unlock(&dev_list_lock); 183857dacad5SJay Sternberg 183957dacad5SJay Sternberg if (tmp) 184057dacad5SJay Sternberg kthread_stop(tmp); 184157dacad5SJay Sternberg } 184257dacad5SJay Sternberg 1843a5cdb68cSKeith Busch static void nvme_dev_disable(struct nvme_dev *dev, bool shutdown) 184457dacad5SJay Sternberg { 184557dacad5SJay Sternberg int i; 184657dacad5SJay Sternberg u32 csts = -1; 184757dacad5SJay Sternberg 184857dacad5SJay Sternberg nvme_dev_list_remove(dev); 184957dacad5SJay Sternberg 185077bf25eaSKeith Busch mutex_lock(&dev->shutdown_lock); 1851b00a726aSKeith Busch if (pci_is_enabled(to_pci_dev(dev->dev))) { 185225646264SKeith Busch nvme_stop_queues(&dev->ctrl); 18537a67cbeaSChristoph Hellwig csts = readl(dev->bar + NVME_REG_CSTS); 185457dacad5SJay Sternberg } 185557dacad5SJay Sternberg if (csts & NVME_CSTS_CFS || !(csts & NVME_CSTS_RDY)) { 185657dacad5SJay Sternberg for (i = dev->queue_count - 1; i >= 0; i--) { 185757dacad5SJay Sternberg struct nvme_queue *nvmeq = dev->queues[i]; 185857dacad5SJay Sternberg nvme_suspend_queue(nvmeq); 185957dacad5SJay Sternberg } 186057dacad5SJay Sternberg } else { 186157dacad5SJay Sternberg nvme_disable_io_queues(dev); 1862a5cdb68cSKeith Busch nvme_disable_admin_queue(dev, shutdown); 186357dacad5SJay Sternberg } 1864b00a726aSKeith Busch nvme_pci_disable(dev); 186557dacad5SJay Sternberg 186657dacad5SJay Sternberg for (i = dev->queue_count - 1; i >= 0; i--) 186757dacad5SJay Sternberg nvme_clear_queue(dev->queues[i]); 186877bf25eaSKeith Busch mutex_unlock(&dev->shutdown_lock); 186957dacad5SJay Sternberg } 187057dacad5SJay Sternberg 187157dacad5SJay Sternberg static int nvme_setup_prp_pools(struct nvme_dev *dev) 187257dacad5SJay Sternberg { 187357dacad5SJay Sternberg dev->prp_page_pool = dma_pool_create("prp list page", dev->dev, 187457dacad5SJay Sternberg PAGE_SIZE, PAGE_SIZE, 0); 187557dacad5SJay Sternberg if (!dev->prp_page_pool) 187657dacad5SJay Sternberg return -ENOMEM; 187757dacad5SJay Sternberg 187857dacad5SJay Sternberg /* Optimisation for I/Os between 4k and 128k */ 187957dacad5SJay Sternberg dev->prp_small_pool = dma_pool_create("prp list 256", dev->dev, 188057dacad5SJay Sternberg 256, 256, 0); 188157dacad5SJay Sternberg if (!dev->prp_small_pool) { 188257dacad5SJay Sternberg dma_pool_destroy(dev->prp_page_pool); 188357dacad5SJay Sternberg return -ENOMEM; 188457dacad5SJay Sternberg } 188557dacad5SJay Sternberg return 0; 188657dacad5SJay Sternberg } 188757dacad5SJay Sternberg 188857dacad5SJay Sternberg static void nvme_release_prp_pools(struct nvme_dev *dev) 188957dacad5SJay Sternberg { 189057dacad5SJay Sternberg dma_pool_destroy(dev->prp_page_pool); 189157dacad5SJay Sternberg dma_pool_destroy(dev->prp_small_pool); 189257dacad5SJay Sternberg } 189357dacad5SJay Sternberg 18941673f1f0SChristoph Hellwig static void nvme_pci_free_ctrl(struct nvme_ctrl *ctrl) 189557dacad5SJay Sternberg { 18961673f1f0SChristoph Hellwig struct nvme_dev *dev = to_nvme_dev(ctrl); 189757dacad5SJay Sternberg 189857dacad5SJay Sternberg put_device(dev->dev); 189957dacad5SJay Sternberg if (dev->tagset.tags) 190057dacad5SJay Sternberg blk_mq_free_tag_set(&dev->tagset); 19011c63dc66SChristoph Hellwig if (dev->ctrl.admin_q) 19021c63dc66SChristoph Hellwig blk_put_queue(dev->ctrl.admin_q); 190357dacad5SJay Sternberg kfree(dev->queues); 190457dacad5SJay Sternberg kfree(dev->entry); 190557dacad5SJay Sternberg kfree(dev); 190657dacad5SJay Sternberg } 190757dacad5SJay Sternberg 1908f58944e2SKeith Busch static void nvme_remove_dead_ctrl(struct nvme_dev *dev, int status) 1909f58944e2SKeith Busch { 1910f58944e2SKeith Busch dev_warn(dev->dev, "Removing after probe failure status: %d\n", status); 1911f58944e2SKeith Busch 1912f58944e2SKeith Busch kref_get(&dev->ctrl.kref); 191369d9a99cSKeith Busch nvme_dev_disable(dev, false); 1914f58944e2SKeith Busch if (!schedule_work(&dev->remove_work)) 1915f58944e2SKeith Busch nvme_put_ctrl(&dev->ctrl); 1916f58944e2SKeith Busch } 1917f58944e2SKeith Busch 1918fd634f41SChristoph Hellwig static void nvme_reset_work(struct work_struct *work) 191957dacad5SJay Sternberg { 1920fd634f41SChristoph Hellwig struct nvme_dev *dev = container_of(work, struct nvme_dev, reset_work); 1921f58944e2SKeith Busch int result = -ENODEV; 192257dacad5SJay Sternberg 1923fd634f41SChristoph Hellwig if (WARN_ON(test_bit(NVME_CTRL_RESETTING, &dev->flags))) 1924fd634f41SChristoph Hellwig goto out; 1925fd634f41SChristoph Hellwig 1926fd634f41SChristoph Hellwig /* 1927fd634f41SChristoph Hellwig * If we're called to reset a live controller first shut it down before 1928fd634f41SChristoph Hellwig * moving on. 1929fd634f41SChristoph Hellwig */ 1930b00a726aSKeith Busch if (dev->ctrl.ctrl_config & NVME_CC_ENABLE) 1931a5cdb68cSKeith Busch nvme_dev_disable(dev, false); 1932fd634f41SChristoph Hellwig 1933fd634f41SChristoph Hellwig set_bit(NVME_CTRL_RESETTING, &dev->flags); 1934fd634f41SChristoph Hellwig 1935b00a726aSKeith Busch result = nvme_pci_enable(dev); 193657dacad5SJay Sternberg if (result) 193757dacad5SJay Sternberg goto out; 193857dacad5SJay Sternberg 193957dacad5SJay Sternberg result = nvme_configure_admin_queue(dev); 194057dacad5SJay Sternberg if (result) 1941f58944e2SKeith Busch goto out; 194257dacad5SJay Sternberg 194357dacad5SJay Sternberg nvme_init_queue(dev->queues[0], 0); 194457dacad5SJay Sternberg result = nvme_alloc_admin_tags(dev); 194557dacad5SJay Sternberg if (result) 1946f58944e2SKeith Busch goto out; 194757dacad5SJay Sternberg 1948ce4541f4SChristoph Hellwig result = nvme_init_identify(&dev->ctrl); 1949ce4541f4SChristoph Hellwig if (result) 1950f58944e2SKeith Busch goto out; 1951ce4541f4SChristoph Hellwig 195257dacad5SJay Sternberg result = nvme_setup_io_queues(dev); 195357dacad5SJay Sternberg if (result) 1954f58944e2SKeith Busch goto out; 195557dacad5SJay Sternberg 1956adf68f21SChristoph Hellwig dev->ctrl.event_limit = NVME_NR_AEN_COMMANDS; 195757dacad5SJay Sternberg 19587385014cSChristoph Hellwig result = nvme_dev_list_add(dev); 19597385014cSChristoph Hellwig if (result) 1960f58944e2SKeith Busch goto out; 196157dacad5SJay Sternberg 196257dacad5SJay Sternberg /* 196357dacad5SJay Sternberg * Keep the controller around but remove all namespaces if we don't have 196457dacad5SJay Sternberg * any working I/O queue. 196557dacad5SJay Sternberg */ 196657dacad5SJay Sternberg if (dev->online_queues < 2) { 196757dacad5SJay Sternberg dev_warn(dev->dev, "IO queues not created\n"); 19685bae7f73SChristoph Hellwig nvme_remove_namespaces(&dev->ctrl); 196957dacad5SJay Sternberg } else { 197025646264SKeith Busch nvme_start_queues(&dev->ctrl); 197157dacad5SJay Sternberg nvme_dev_add(dev); 197257dacad5SJay Sternberg } 197357dacad5SJay Sternberg 1974fd634f41SChristoph Hellwig clear_bit(NVME_CTRL_RESETTING, &dev->flags); 197557dacad5SJay Sternberg return; 197657dacad5SJay Sternberg 197757dacad5SJay Sternberg out: 1978f58944e2SKeith Busch nvme_remove_dead_ctrl(dev, result); 197957dacad5SJay Sternberg } 198057dacad5SJay Sternberg 19815c8809e6SChristoph Hellwig static void nvme_remove_dead_ctrl_work(struct work_struct *work) 198257dacad5SJay Sternberg { 19835c8809e6SChristoph Hellwig struct nvme_dev *dev = container_of(work, struct nvme_dev, remove_work); 198457dacad5SJay Sternberg struct pci_dev *pdev = to_pci_dev(dev->dev); 198557dacad5SJay Sternberg 198669d9a99cSKeith Busch nvme_kill_queues(&dev->ctrl); 198757dacad5SJay Sternberg if (pci_get_drvdata(pdev)) 198857dacad5SJay Sternberg pci_stop_and_remove_bus_device_locked(pdev); 19891673f1f0SChristoph Hellwig nvme_put_ctrl(&dev->ctrl); 199057dacad5SJay Sternberg } 199157dacad5SJay Sternberg 199257dacad5SJay Sternberg static int nvme_reset(struct nvme_dev *dev) 199357dacad5SJay Sternberg { 19941c63dc66SChristoph Hellwig if (!dev->ctrl.admin_q || blk_queue_dying(dev->ctrl.admin_q)) 199557dacad5SJay Sternberg return -ENODEV; 199657dacad5SJay Sternberg 1997846cc05fSChristoph Hellwig if (!queue_work(nvme_workq, &dev->reset_work)) 1998846cc05fSChristoph Hellwig return -EBUSY; 199957dacad5SJay Sternberg 200057dacad5SJay Sternberg flush_work(&dev->reset_work); 200157dacad5SJay Sternberg return 0; 200257dacad5SJay Sternberg } 200357dacad5SJay Sternberg 20041c63dc66SChristoph Hellwig static int nvme_pci_reg_read32(struct nvme_ctrl *ctrl, u32 off, u32 *val) 200557dacad5SJay Sternberg { 20061c63dc66SChristoph Hellwig *val = readl(to_nvme_dev(ctrl)->bar + off); 20071c63dc66SChristoph Hellwig return 0; 200857dacad5SJay Sternberg } 20091c63dc66SChristoph Hellwig 20105fd4ce1bSChristoph Hellwig static int nvme_pci_reg_write32(struct nvme_ctrl *ctrl, u32 off, u32 val) 20115fd4ce1bSChristoph Hellwig { 20125fd4ce1bSChristoph Hellwig writel(val, to_nvme_dev(ctrl)->bar + off); 20135fd4ce1bSChristoph Hellwig return 0; 20145fd4ce1bSChristoph Hellwig } 20155fd4ce1bSChristoph Hellwig 20167fd8930fSChristoph Hellwig static int nvme_pci_reg_read64(struct nvme_ctrl *ctrl, u32 off, u64 *val) 20177fd8930fSChristoph Hellwig { 20187fd8930fSChristoph Hellwig *val = readq(to_nvme_dev(ctrl)->bar + off); 20197fd8930fSChristoph Hellwig return 0; 20207fd8930fSChristoph Hellwig } 20217fd8930fSChristoph Hellwig 20225bae7f73SChristoph Hellwig static bool nvme_pci_io_incapable(struct nvme_ctrl *ctrl) 20235bae7f73SChristoph Hellwig { 20245bae7f73SChristoph Hellwig struct nvme_dev *dev = to_nvme_dev(ctrl); 20255bae7f73SChristoph Hellwig 20265bae7f73SChristoph Hellwig return !dev->bar || dev->online_queues < 2; 20275bae7f73SChristoph Hellwig } 20285bae7f73SChristoph Hellwig 2029f3ca80fcSChristoph Hellwig static int nvme_pci_reset_ctrl(struct nvme_ctrl *ctrl) 2030f3ca80fcSChristoph Hellwig { 2031f3ca80fcSChristoph Hellwig return nvme_reset(to_nvme_dev(ctrl)); 2032f3ca80fcSChristoph Hellwig } 2033f3ca80fcSChristoph Hellwig 20341c63dc66SChristoph Hellwig static const struct nvme_ctrl_ops nvme_pci_ctrl_ops = { 20351c63dc66SChristoph Hellwig .reg_read32 = nvme_pci_reg_read32, 20365fd4ce1bSChristoph Hellwig .reg_write32 = nvme_pci_reg_write32, 20377fd8930fSChristoph Hellwig .reg_read64 = nvme_pci_reg_read64, 20385bae7f73SChristoph Hellwig .io_incapable = nvme_pci_io_incapable, 2039f3ca80fcSChristoph Hellwig .reset_ctrl = nvme_pci_reset_ctrl, 20401673f1f0SChristoph Hellwig .free_ctrl = nvme_pci_free_ctrl, 20411c63dc66SChristoph Hellwig }; 204257dacad5SJay Sternberg 2043b00a726aSKeith Busch static int nvme_dev_map(struct nvme_dev *dev) 2044b00a726aSKeith Busch { 2045b00a726aSKeith Busch int bars; 2046b00a726aSKeith Busch struct pci_dev *pdev = to_pci_dev(dev->dev); 2047b00a726aSKeith Busch 2048b00a726aSKeith Busch bars = pci_select_bars(pdev, IORESOURCE_MEM); 2049b00a726aSKeith Busch if (!bars) 2050b00a726aSKeith Busch return -ENODEV; 2051b00a726aSKeith Busch if (pci_request_selected_regions(pdev, bars, "nvme")) 2052b00a726aSKeith Busch return -ENODEV; 2053b00a726aSKeith Busch 2054b00a726aSKeith Busch dev->bar = ioremap(pci_resource_start(pdev, 0), 8192); 2055b00a726aSKeith Busch if (!dev->bar) 2056b00a726aSKeith Busch goto release; 2057b00a726aSKeith Busch 2058b00a726aSKeith Busch return 0; 2059b00a726aSKeith Busch release: 2060b00a726aSKeith Busch pci_release_regions(pdev); 2061b00a726aSKeith Busch return -ENODEV; 2062b00a726aSKeith Busch } 2063b00a726aSKeith Busch 206457dacad5SJay Sternberg static int nvme_probe(struct pci_dev *pdev, const struct pci_device_id *id) 206557dacad5SJay Sternberg { 206657dacad5SJay Sternberg int node, result = -ENOMEM; 206757dacad5SJay Sternberg struct nvme_dev *dev; 206857dacad5SJay Sternberg 206957dacad5SJay Sternberg node = dev_to_node(&pdev->dev); 207057dacad5SJay Sternberg if (node == NUMA_NO_NODE) 207157dacad5SJay Sternberg set_dev_node(&pdev->dev, 0); 207257dacad5SJay Sternberg 207357dacad5SJay Sternberg dev = kzalloc_node(sizeof(*dev), GFP_KERNEL, node); 207457dacad5SJay Sternberg if (!dev) 207557dacad5SJay Sternberg return -ENOMEM; 207657dacad5SJay Sternberg dev->entry = kzalloc_node(num_possible_cpus() * sizeof(*dev->entry), 207757dacad5SJay Sternberg GFP_KERNEL, node); 207857dacad5SJay Sternberg if (!dev->entry) 207957dacad5SJay Sternberg goto free; 208057dacad5SJay Sternberg dev->queues = kzalloc_node((num_possible_cpus() + 1) * sizeof(void *), 208157dacad5SJay Sternberg GFP_KERNEL, node); 208257dacad5SJay Sternberg if (!dev->queues) 208357dacad5SJay Sternberg goto free; 208457dacad5SJay Sternberg 208557dacad5SJay Sternberg dev->dev = get_device(&pdev->dev); 208657dacad5SJay Sternberg pci_set_drvdata(pdev, dev); 208757dacad5SJay Sternberg 2088b00a726aSKeith Busch result = nvme_dev_map(dev); 2089b00a726aSKeith Busch if (result) 2090b00a726aSKeith Busch goto free; 2091b00a726aSKeith Busch 209257dacad5SJay Sternberg INIT_LIST_HEAD(&dev->node); 209357dacad5SJay Sternberg INIT_WORK(&dev->scan_work, nvme_dev_scan); 2094f3ca80fcSChristoph Hellwig INIT_WORK(&dev->reset_work, nvme_reset_work); 20955c8809e6SChristoph Hellwig INIT_WORK(&dev->remove_work, nvme_remove_dead_ctrl_work); 209677bf25eaSKeith Busch mutex_init(&dev->shutdown_lock); 2097db3cbfffSKeith Busch init_completion(&dev->ioq_wait); 2098f3ca80fcSChristoph Hellwig 2099f3ca80fcSChristoph Hellwig result = nvme_setup_prp_pools(dev); 2100f3ca80fcSChristoph Hellwig if (result) 2101f3ca80fcSChristoph Hellwig goto put_pci; 2102f3ca80fcSChristoph Hellwig 2103f3ca80fcSChristoph Hellwig result = nvme_init_ctrl(&dev->ctrl, &pdev->dev, &nvme_pci_ctrl_ops, 2104f3ca80fcSChristoph Hellwig id->driver_data); 2105f3ca80fcSChristoph Hellwig if (result) 2106f3ca80fcSChristoph Hellwig goto release_pools; 2107f3ca80fcSChristoph Hellwig 210892f7a162SKeith Busch queue_work(nvme_workq, &dev->reset_work); 210957dacad5SJay Sternberg return 0; 211057dacad5SJay Sternberg 211157dacad5SJay Sternberg release_pools: 211257dacad5SJay Sternberg nvme_release_prp_pools(dev); 211357dacad5SJay Sternberg put_pci: 211457dacad5SJay Sternberg put_device(dev->dev); 2115b00a726aSKeith Busch nvme_dev_unmap(dev); 211657dacad5SJay Sternberg free: 211757dacad5SJay Sternberg kfree(dev->queues); 211857dacad5SJay Sternberg kfree(dev->entry); 211957dacad5SJay Sternberg kfree(dev); 212057dacad5SJay Sternberg return result; 212157dacad5SJay Sternberg } 212257dacad5SJay Sternberg 212357dacad5SJay Sternberg static void nvme_reset_notify(struct pci_dev *pdev, bool prepare) 212457dacad5SJay Sternberg { 212557dacad5SJay Sternberg struct nvme_dev *dev = pci_get_drvdata(pdev); 212657dacad5SJay Sternberg 212757dacad5SJay Sternberg if (prepare) 2128a5cdb68cSKeith Busch nvme_dev_disable(dev, false); 212957dacad5SJay Sternberg else 213092f7a162SKeith Busch queue_work(nvme_workq, &dev->reset_work); 213157dacad5SJay Sternberg } 213257dacad5SJay Sternberg 213357dacad5SJay Sternberg static void nvme_shutdown(struct pci_dev *pdev) 213457dacad5SJay Sternberg { 213557dacad5SJay Sternberg struct nvme_dev *dev = pci_get_drvdata(pdev); 2136a5cdb68cSKeith Busch nvme_dev_disable(dev, true); 213757dacad5SJay Sternberg } 213857dacad5SJay Sternberg 2139f58944e2SKeith Busch /* 2140f58944e2SKeith Busch * The driver's remove may be called on a device in a partially initialized 2141f58944e2SKeith Busch * state. This function must not have any dependencies on the device state in 2142f58944e2SKeith Busch * order to proceed. 2143f58944e2SKeith Busch */ 214457dacad5SJay Sternberg static void nvme_remove(struct pci_dev *pdev) 214557dacad5SJay Sternberg { 214657dacad5SJay Sternberg struct nvme_dev *dev = pci_get_drvdata(pdev); 214757dacad5SJay Sternberg 2148646017a6SKeith Busch set_bit(NVME_CTRL_REMOVING, &dev->flags); 214957dacad5SJay Sternberg pci_set_drvdata(pdev, NULL); 215057dacad5SJay Sternberg flush_work(&dev->scan_work); 21515bae7f73SChristoph Hellwig nvme_remove_namespaces(&dev->ctrl); 215253029b04SKeith Busch nvme_uninit_ctrl(&dev->ctrl); 2153a5cdb68cSKeith Busch nvme_dev_disable(dev, true); 2154ff23a2a1SKeith Busch flush_work(&dev->reset_work); 215557dacad5SJay Sternberg nvme_dev_remove_admin(dev); 215657dacad5SJay Sternberg nvme_free_queues(dev, 0); 215757dacad5SJay Sternberg nvme_release_cmb(dev); 215857dacad5SJay Sternberg nvme_release_prp_pools(dev); 2159b00a726aSKeith Busch nvme_dev_unmap(dev); 21601673f1f0SChristoph Hellwig nvme_put_ctrl(&dev->ctrl); 216157dacad5SJay Sternberg } 216257dacad5SJay Sternberg 216357dacad5SJay Sternberg #ifdef CONFIG_PM_SLEEP 216457dacad5SJay Sternberg static int nvme_suspend(struct device *dev) 216557dacad5SJay Sternberg { 216657dacad5SJay Sternberg struct pci_dev *pdev = to_pci_dev(dev); 216757dacad5SJay Sternberg struct nvme_dev *ndev = pci_get_drvdata(pdev); 216857dacad5SJay Sternberg 2169a5cdb68cSKeith Busch nvme_dev_disable(ndev, true); 217057dacad5SJay Sternberg return 0; 217157dacad5SJay Sternberg } 217257dacad5SJay Sternberg 217357dacad5SJay Sternberg static int nvme_resume(struct device *dev) 217457dacad5SJay Sternberg { 217557dacad5SJay Sternberg struct pci_dev *pdev = to_pci_dev(dev); 217657dacad5SJay Sternberg struct nvme_dev *ndev = pci_get_drvdata(pdev); 217757dacad5SJay Sternberg 217892f7a162SKeith Busch queue_work(nvme_workq, &ndev->reset_work); 217957dacad5SJay Sternberg return 0; 218057dacad5SJay Sternberg } 218157dacad5SJay Sternberg #endif 218257dacad5SJay Sternberg 218357dacad5SJay Sternberg static SIMPLE_DEV_PM_OPS(nvme_dev_pm_ops, nvme_suspend, nvme_resume); 218457dacad5SJay Sternberg 2185a0a3408eSKeith Busch static pci_ers_result_t nvme_error_detected(struct pci_dev *pdev, 2186a0a3408eSKeith Busch pci_channel_state_t state) 2187a0a3408eSKeith Busch { 2188a0a3408eSKeith Busch struct nvme_dev *dev = pci_get_drvdata(pdev); 2189a0a3408eSKeith Busch 2190a0a3408eSKeith Busch /* 2191a0a3408eSKeith Busch * A frozen channel requires a reset. When detected, this method will 2192a0a3408eSKeith Busch * shutdown the controller to quiesce. The controller will be restarted 2193a0a3408eSKeith Busch * after the slot reset through driver's slot_reset callback. 2194a0a3408eSKeith Busch */ 2195a0a3408eSKeith Busch dev_warn(&pdev->dev, "error detected: state:%d\n", state); 2196a0a3408eSKeith Busch switch (state) { 2197a0a3408eSKeith Busch case pci_channel_io_normal: 2198a0a3408eSKeith Busch return PCI_ERS_RESULT_CAN_RECOVER; 2199a0a3408eSKeith Busch case pci_channel_io_frozen: 2200a5cdb68cSKeith Busch nvme_dev_disable(dev, false); 2201a0a3408eSKeith Busch return PCI_ERS_RESULT_NEED_RESET; 2202a0a3408eSKeith Busch case pci_channel_io_perm_failure: 2203a0a3408eSKeith Busch return PCI_ERS_RESULT_DISCONNECT; 2204a0a3408eSKeith Busch } 2205a0a3408eSKeith Busch return PCI_ERS_RESULT_NEED_RESET; 2206a0a3408eSKeith Busch } 2207a0a3408eSKeith Busch 2208a0a3408eSKeith Busch static pci_ers_result_t nvme_slot_reset(struct pci_dev *pdev) 2209a0a3408eSKeith Busch { 2210a0a3408eSKeith Busch struct nvme_dev *dev = pci_get_drvdata(pdev); 2211a0a3408eSKeith Busch 2212a0a3408eSKeith Busch dev_info(&pdev->dev, "restart after slot reset\n"); 2213a0a3408eSKeith Busch pci_restore_state(pdev); 2214a0a3408eSKeith Busch queue_work(nvme_workq, &dev->reset_work); 2215a0a3408eSKeith Busch return PCI_ERS_RESULT_RECOVERED; 2216a0a3408eSKeith Busch } 2217a0a3408eSKeith Busch 2218a0a3408eSKeith Busch static void nvme_error_resume(struct pci_dev *pdev) 2219a0a3408eSKeith Busch { 2220a0a3408eSKeith Busch pci_cleanup_aer_uncorrect_error_status(pdev); 2221a0a3408eSKeith Busch } 2222a0a3408eSKeith Busch 222357dacad5SJay Sternberg static const struct pci_error_handlers nvme_err_handler = { 222457dacad5SJay Sternberg .error_detected = nvme_error_detected, 222557dacad5SJay Sternberg .slot_reset = nvme_slot_reset, 222657dacad5SJay Sternberg .resume = nvme_error_resume, 222757dacad5SJay Sternberg .reset_notify = nvme_reset_notify, 222857dacad5SJay Sternberg }; 222957dacad5SJay Sternberg 223057dacad5SJay Sternberg /* Move to pci_ids.h later */ 223157dacad5SJay Sternberg #define PCI_CLASS_STORAGE_EXPRESS 0x010802 223257dacad5SJay Sternberg 223357dacad5SJay Sternberg static const struct pci_device_id nvme_id_table[] = { 2234106198edSChristoph Hellwig { PCI_VDEVICE(INTEL, 0x0953), 2235106198edSChristoph Hellwig .driver_data = NVME_QUIRK_STRIPE_SIZE, }, 2236540c801cSKeith Busch { PCI_VDEVICE(INTEL, 0x5845), /* Qemu emulated controller */ 2237540c801cSKeith Busch .driver_data = NVME_QUIRK_IDENTIFY_CNS, }, 223857dacad5SJay Sternberg { PCI_DEVICE_CLASS(PCI_CLASS_STORAGE_EXPRESS, 0xffffff) }, 2239c74dc780SStephan Günther { PCI_DEVICE(PCI_VENDOR_ID_APPLE, 0x2001) }, 224057dacad5SJay Sternberg { 0, } 224157dacad5SJay Sternberg }; 224257dacad5SJay Sternberg MODULE_DEVICE_TABLE(pci, nvme_id_table); 224357dacad5SJay Sternberg 224457dacad5SJay Sternberg static struct pci_driver nvme_driver = { 224557dacad5SJay Sternberg .name = "nvme", 224657dacad5SJay Sternberg .id_table = nvme_id_table, 224757dacad5SJay Sternberg .probe = nvme_probe, 224857dacad5SJay Sternberg .remove = nvme_remove, 224957dacad5SJay Sternberg .shutdown = nvme_shutdown, 225057dacad5SJay Sternberg .driver = { 225157dacad5SJay Sternberg .pm = &nvme_dev_pm_ops, 225257dacad5SJay Sternberg }, 225357dacad5SJay Sternberg .err_handler = &nvme_err_handler, 225457dacad5SJay Sternberg }; 225557dacad5SJay Sternberg 225657dacad5SJay Sternberg static int __init nvme_init(void) 225757dacad5SJay Sternberg { 225857dacad5SJay Sternberg int result; 225957dacad5SJay Sternberg 226057dacad5SJay Sternberg init_waitqueue_head(&nvme_kthread_wait); 226157dacad5SJay Sternberg 226292f7a162SKeith Busch nvme_workq = alloc_workqueue("nvme", WQ_UNBOUND | WQ_MEM_RECLAIM, 0); 226357dacad5SJay Sternberg if (!nvme_workq) 226457dacad5SJay Sternberg return -ENOMEM; 226557dacad5SJay Sternberg 22665bae7f73SChristoph Hellwig result = nvme_core_init(); 226757dacad5SJay Sternberg if (result < 0) 226857dacad5SJay Sternberg goto kill_workq; 226957dacad5SJay Sternberg 227057dacad5SJay Sternberg result = pci_register_driver(&nvme_driver); 227157dacad5SJay Sternberg if (result) 2272f3ca80fcSChristoph Hellwig goto core_exit; 227357dacad5SJay Sternberg return 0; 227457dacad5SJay Sternberg 2275f3ca80fcSChristoph Hellwig core_exit: 22765bae7f73SChristoph Hellwig nvme_core_exit(); 227757dacad5SJay Sternberg kill_workq: 227857dacad5SJay Sternberg destroy_workqueue(nvme_workq); 227957dacad5SJay Sternberg return result; 228057dacad5SJay Sternberg } 228157dacad5SJay Sternberg 228257dacad5SJay Sternberg static void __exit nvme_exit(void) 228357dacad5SJay Sternberg { 228457dacad5SJay Sternberg pci_unregister_driver(&nvme_driver); 22855bae7f73SChristoph Hellwig nvme_core_exit(); 228657dacad5SJay Sternberg destroy_workqueue(nvme_workq); 228757dacad5SJay Sternberg BUG_ON(nvme_thread && !IS_ERR(nvme_thread)); 228857dacad5SJay Sternberg _nvme_check_size(); 228957dacad5SJay Sternberg } 229057dacad5SJay Sternberg 229157dacad5SJay Sternberg MODULE_AUTHOR("Matthew Wilcox <willy@linux.intel.com>"); 229257dacad5SJay Sternberg MODULE_LICENSE("GPL"); 229357dacad5SJay Sternberg MODULE_VERSION("1.0"); 229457dacad5SJay Sternberg module_init(nvme_init); 229557dacad5SJay Sternberg module_exit(nvme_exit); 2296