xref: /openbmc/linux/drivers/nvme/host/pci.c (revision bb0eb050)
1 /*
2  * NVM Express device driver
3  * Copyright (c) 2011-2014, Intel Corporation.
4  *
5  * This program is free software; you can redistribute it and/or modify it
6  * under the terms and conditions of the GNU General Public License,
7  * version 2, as published by the Free Software Foundation.
8  *
9  * This program is distributed in the hope it will be useful, but WITHOUT
10  * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11  * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License for
12  * more details.
13  */
14 
15 #include <linux/aer.h>
16 #include <linux/bitops.h>
17 #include <linux/blkdev.h>
18 #include <linux/blk-mq.h>
19 #include <linux/blk-mq-pci.h>
20 #include <linux/cpu.h>
21 #include <linux/delay.h>
22 #include <linux/dmi.h>
23 #include <linux/errno.h>
24 #include <linux/fs.h>
25 #include <linux/genhd.h>
26 #include <linux/hdreg.h>
27 #include <linux/idr.h>
28 #include <linux/init.h>
29 #include <linux/interrupt.h>
30 #include <linux/io.h>
31 #include <linux/kdev_t.h>
32 #include <linux/kernel.h>
33 #include <linux/mm.h>
34 #include <linux/module.h>
35 #include <linux/moduleparam.h>
36 #include <linux/mutex.h>
37 #include <linux/pci.h>
38 #include <linux/poison.h>
39 #include <linux/ptrace.h>
40 #include <linux/sched.h>
41 #include <linux/slab.h>
42 #include <linux/t10-pi.h>
43 #include <linux/timer.h>
44 #include <linux/types.h>
45 #include <linux/io-64-nonatomic-lo-hi.h>
46 #include <asm/unaligned.h>
47 #include <linux/sed-opal.h>
48 
49 #include "nvme.h"
50 
51 #define NVME_Q_DEPTH		1024
52 #define NVME_AQ_DEPTH		256
53 #define SQ_SIZE(depth)		(depth * sizeof(struct nvme_command))
54 #define CQ_SIZE(depth)		(depth * sizeof(struct nvme_completion))
55 
56 /*
57  * We handle AEN commands ourselves and don't even let the
58  * block layer know about them.
59  */
60 #define NVME_AQ_BLKMQ_DEPTH	(NVME_AQ_DEPTH - NVME_NR_AERS)
61 
62 static int use_threaded_interrupts;
63 module_param(use_threaded_interrupts, int, 0);
64 
65 static bool use_cmb_sqes = true;
66 module_param(use_cmb_sqes, bool, 0644);
67 MODULE_PARM_DESC(use_cmb_sqes, "use controller's memory buffer for I/O SQes");
68 
69 static struct workqueue_struct *nvme_workq;
70 
71 struct nvme_dev;
72 struct nvme_queue;
73 
74 static int nvme_reset(struct nvme_dev *dev);
75 static void nvme_process_cq(struct nvme_queue *nvmeq);
76 static void nvme_dev_disable(struct nvme_dev *dev, bool shutdown);
77 
78 /*
79  * Represents an NVM Express device.  Each nvme_dev is a PCI function.
80  */
81 struct nvme_dev {
82 	struct nvme_queue **queues;
83 	struct blk_mq_tag_set tagset;
84 	struct blk_mq_tag_set admin_tagset;
85 	u32 __iomem *dbs;
86 	struct device *dev;
87 	struct dma_pool *prp_page_pool;
88 	struct dma_pool *prp_small_pool;
89 	unsigned queue_count;
90 	unsigned online_queues;
91 	unsigned max_qid;
92 	int q_depth;
93 	u32 db_stride;
94 	void __iomem *bar;
95 	struct work_struct reset_work;
96 	struct work_struct remove_work;
97 	struct timer_list watchdog_timer;
98 	struct mutex shutdown_lock;
99 	bool subsystem;
100 	void __iomem *cmb;
101 	dma_addr_t cmb_dma_addr;
102 	u64 cmb_size;
103 	u32 cmbsz;
104 	u32 cmbloc;
105 	struct nvme_ctrl ctrl;
106 	struct completion ioq_wait;
107 	u32 *dbbuf_dbs;
108 	dma_addr_t dbbuf_dbs_dma_addr;
109 	u32 *dbbuf_eis;
110 	dma_addr_t dbbuf_eis_dma_addr;
111 };
112 
113 static inline unsigned int sq_idx(unsigned int qid, u32 stride)
114 {
115 	return qid * 2 * stride;
116 }
117 
118 static inline unsigned int cq_idx(unsigned int qid, u32 stride)
119 {
120 	return (qid * 2 + 1) * stride;
121 }
122 
123 static inline struct nvme_dev *to_nvme_dev(struct nvme_ctrl *ctrl)
124 {
125 	return container_of(ctrl, struct nvme_dev, ctrl);
126 }
127 
128 /*
129  * An NVM Express queue.  Each device has at least two (one for admin
130  * commands and one for I/O commands).
131  */
132 struct nvme_queue {
133 	struct device *q_dmadev;
134 	struct nvme_dev *dev;
135 	spinlock_t q_lock;
136 	struct nvme_command *sq_cmds;
137 	struct nvme_command __iomem *sq_cmds_io;
138 	volatile struct nvme_completion *cqes;
139 	struct blk_mq_tags **tags;
140 	dma_addr_t sq_dma_addr;
141 	dma_addr_t cq_dma_addr;
142 	u32 __iomem *q_db;
143 	u16 q_depth;
144 	s16 cq_vector;
145 	u16 sq_tail;
146 	u16 cq_head;
147 	u16 qid;
148 	u8 cq_phase;
149 	u8 cqe_seen;
150 	u32 *dbbuf_sq_db;
151 	u32 *dbbuf_cq_db;
152 	u32 *dbbuf_sq_ei;
153 	u32 *dbbuf_cq_ei;
154 };
155 
156 /*
157  * The nvme_iod describes the data in an I/O, including the list of PRP
158  * entries.  You can't see it in this data structure because C doesn't let
159  * me express that.  Use nvme_init_iod to ensure there's enough space
160  * allocated to store the PRP list.
161  */
162 struct nvme_iod {
163 	struct nvme_request req;
164 	struct nvme_queue *nvmeq;
165 	int aborted;
166 	int npages;		/* In the PRP list. 0 means small pool in use */
167 	int nents;		/* Used in scatterlist */
168 	int length;		/* Of data, in bytes */
169 	dma_addr_t first_dma;
170 	struct scatterlist meta_sg; /* metadata requires single contiguous buffer */
171 	struct scatterlist *sg;
172 	struct scatterlist inline_sg[0];
173 };
174 
175 /*
176  * Check we didin't inadvertently grow the command struct
177  */
178 static inline void _nvme_check_size(void)
179 {
180 	BUILD_BUG_ON(sizeof(struct nvme_rw_command) != 64);
181 	BUILD_BUG_ON(sizeof(struct nvme_create_cq) != 64);
182 	BUILD_BUG_ON(sizeof(struct nvme_create_sq) != 64);
183 	BUILD_BUG_ON(sizeof(struct nvme_delete_queue) != 64);
184 	BUILD_BUG_ON(sizeof(struct nvme_features) != 64);
185 	BUILD_BUG_ON(sizeof(struct nvme_format_cmd) != 64);
186 	BUILD_BUG_ON(sizeof(struct nvme_abort_cmd) != 64);
187 	BUILD_BUG_ON(sizeof(struct nvme_command) != 64);
188 	BUILD_BUG_ON(sizeof(struct nvme_id_ctrl) != 4096);
189 	BUILD_BUG_ON(sizeof(struct nvme_id_ns) != 4096);
190 	BUILD_BUG_ON(sizeof(struct nvme_lba_range_type) != 64);
191 	BUILD_BUG_ON(sizeof(struct nvme_smart_log) != 512);
192 	BUILD_BUG_ON(sizeof(struct nvme_dbbuf) != 64);
193 }
194 
195 static inline unsigned int nvme_dbbuf_size(u32 stride)
196 {
197 	return ((num_possible_cpus() + 1) * 8 * stride);
198 }
199 
200 static int nvme_dbbuf_dma_alloc(struct nvme_dev *dev)
201 {
202 	unsigned int mem_size = nvme_dbbuf_size(dev->db_stride);
203 
204 	if (dev->dbbuf_dbs)
205 		return 0;
206 
207 	dev->dbbuf_dbs = dma_alloc_coherent(dev->dev, mem_size,
208 					    &dev->dbbuf_dbs_dma_addr,
209 					    GFP_KERNEL);
210 	if (!dev->dbbuf_dbs)
211 		return -ENOMEM;
212 	dev->dbbuf_eis = dma_alloc_coherent(dev->dev, mem_size,
213 					    &dev->dbbuf_eis_dma_addr,
214 					    GFP_KERNEL);
215 	if (!dev->dbbuf_eis) {
216 		dma_free_coherent(dev->dev, mem_size,
217 				  dev->dbbuf_dbs, dev->dbbuf_dbs_dma_addr);
218 		dev->dbbuf_dbs = NULL;
219 		return -ENOMEM;
220 	}
221 
222 	return 0;
223 }
224 
225 static void nvme_dbbuf_dma_free(struct nvme_dev *dev)
226 {
227 	unsigned int mem_size = nvme_dbbuf_size(dev->db_stride);
228 
229 	if (dev->dbbuf_dbs) {
230 		dma_free_coherent(dev->dev, mem_size,
231 				  dev->dbbuf_dbs, dev->dbbuf_dbs_dma_addr);
232 		dev->dbbuf_dbs = NULL;
233 	}
234 	if (dev->dbbuf_eis) {
235 		dma_free_coherent(dev->dev, mem_size,
236 				  dev->dbbuf_eis, dev->dbbuf_eis_dma_addr);
237 		dev->dbbuf_eis = NULL;
238 	}
239 }
240 
241 static void nvme_dbbuf_init(struct nvme_dev *dev,
242 			    struct nvme_queue *nvmeq, int qid)
243 {
244 	if (!dev->dbbuf_dbs || !qid)
245 		return;
246 
247 	nvmeq->dbbuf_sq_db = &dev->dbbuf_dbs[sq_idx(qid, dev->db_stride)];
248 	nvmeq->dbbuf_cq_db = &dev->dbbuf_dbs[cq_idx(qid, dev->db_stride)];
249 	nvmeq->dbbuf_sq_ei = &dev->dbbuf_eis[sq_idx(qid, dev->db_stride)];
250 	nvmeq->dbbuf_cq_ei = &dev->dbbuf_eis[cq_idx(qid, dev->db_stride)];
251 }
252 
253 static void nvme_dbbuf_set(struct nvme_dev *dev)
254 {
255 	struct nvme_command c;
256 
257 	if (!dev->dbbuf_dbs)
258 		return;
259 
260 	memset(&c, 0, sizeof(c));
261 	c.dbbuf.opcode = nvme_admin_dbbuf;
262 	c.dbbuf.prp1 = cpu_to_le64(dev->dbbuf_dbs_dma_addr);
263 	c.dbbuf.prp2 = cpu_to_le64(dev->dbbuf_eis_dma_addr);
264 
265 	if (nvme_submit_sync_cmd(dev->ctrl.admin_q, &c, NULL, 0)) {
266 		dev_warn(dev->ctrl.device, "unable to set dbbuf\n");
267 		/* Free memory and continue on */
268 		nvme_dbbuf_dma_free(dev);
269 	}
270 }
271 
272 static inline int nvme_dbbuf_need_event(u16 event_idx, u16 new_idx, u16 old)
273 {
274 	return (u16)(new_idx - event_idx - 1) < (u16)(new_idx - old);
275 }
276 
277 /* Update dbbuf and return true if an MMIO is required */
278 static bool nvme_dbbuf_update_and_check_event(u16 value, u32 *dbbuf_db,
279 					      volatile u32 *dbbuf_ei)
280 {
281 	if (dbbuf_db) {
282 		u16 old_value;
283 
284 		/*
285 		 * Ensure that the queue is written before updating
286 		 * the doorbell in memory
287 		 */
288 		wmb();
289 
290 		old_value = *dbbuf_db;
291 		*dbbuf_db = value;
292 
293 		if (!nvme_dbbuf_need_event(*dbbuf_ei, value, old_value))
294 			return false;
295 	}
296 
297 	return true;
298 }
299 
300 /*
301  * Max size of iod being embedded in the request payload
302  */
303 #define NVME_INT_PAGES		2
304 #define NVME_INT_BYTES(dev)	(NVME_INT_PAGES * (dev)->ctrl.page_size)
305 
306 /*
307  * Will slightly overestimate the number of pages needed.  This is OK
308  * as it only leads to a small amount of wasted memory for the lifetime of
309  * the I/O.
310  */
311 static int nvme_npages(unsigned size, struct nvme_dev *dev)
312 {
313 	unsigned nprps = DIV_ROUND_UP(size + dev->ctrl.page_size,
314 				      dev->ctrl.page_size);
315 	return DIV_ROUND_UP(8 * nprps, PAGE_SIZE - 8);
316 }
317 
318 static unsigned int nvme_iod_alloc_size(struct nvme_dev *dev,
319 		unsigned int size, unsigned int nseg)
320 {
321 	return sizeof(__le64 *) * nvme_npages(size, dev) +
322 			sizeof(struct scatterlist) * nseg;
323 }
324 
325 static unsigned int nvme_cmd_size(struct nvme_dev *dev)
326 {
327 	return sizeof(struct nvme_iod) +
328 		nvme_iod_alloc_size(dev, NVME_INT_BYTES(dev), NVME_INT_PAGES);
329 }
330 
331 static int nvme_admin_init_hctx(struct blk_mq_hw_ctx *hctx, void *data,
332 				unsigned int hctx_idx)
333 {
334 	struct nvme_dev *dev = data;
335 	struct nvme_queue *nvmeq = dev->queues[0];
336 
337 	WARN_ON(hctx_idx != 0);
338 	WARN_ON(dev->admin_tagset.tags[0] != hctx->tags);
339 	WARN_ON(nvmeq->tags);
340 
341 	hctx->driver_data = nvmeq;
342 	nvmeq->tags = &dev->admin_tagset.tags[0];
343 	return 0;
344 }
345 
346 static void nvme_admin_exit_hctx(struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx)
347 {
348 	struct nvme_queue *nvmeq = hctx->driver_data;
349 
350 	nvmeq->tags = NULL;
351 }
352 
353 static int nvme_admin_init_request(struct blk_mq_tag_set *set,
354 		struct request *req, unsigned int hctx_idx,
355 		unsigned int numa_node)
356 {
357 	struct nvme_dev *dev = set->driver_data;
358 	struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
359 	struct nvme_queue *nvmeq = dev->queues[0];
360 
361 	BUG_ON(!nvmeq);
362 	iod->nvmeq = nvmeq;
363 	return 0;
364 }
365 
366 static int nvme_init_hctx(struct blk_mq_hw_ctx *hctx, void *data,
367 			  unsigned int hctx_idx)
368 {
369 	struct nvme_dev *dev = data;
370 	struct nvme_queue *nvmeq = dev->queues[hctx_idx + 1];
371 
372 	if (!nvmeq->tags)
373 		nvmeq->tags = &dev->tagset.tags[hctx_idx];
374 
375 	WARN_ON(dev->tagset.tags[hctx_idx] != hctx->tags);
376 	hctx->driver_data = nvmeq;
377 	return 0;
378 }
379 
380 static int nvme_init_request(struct blk_mq_tag_set *set, struct request *req,
381 		unsigned int hctx_idx, unsigned int numa_node)
382 {
383 	struct nvme_dev *dev = set->driver_data;
384 	struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
385 	struct nvme_queue *nvmeq = dev->queues[hctx_idx + 1];
386 
387 	BUG_ON(!nvmeq);
388 	iod->nvmeq = nvmeq;
389 	return 0;
390 }
391 
392 static int nvme_pci_map_queues(struct blk_mq_tag_set *set)
393 {
394 	struct nvme_dev *dev = set->driver_data;
395 
396 	return blk_mq_pci_map_queues(set, to_pci_dev(dev->dev));
397 }
398 
399 /**
400  * __nvme_submit_cmd() - Copy a command into a queue and ring the doorbell
401  * @nvmeq: The queue to use
402  * @cmd: The command to send
403  *
404  * Safe to use from interrupt context
405  */
406 static void __nvme_submit_cmd(struct nvme_queue *nvmeq,
407 						struct nvme_command *cmd)
408 {
409 	u16 tail = nvmeq->sq_tail;
410 
411 	if (nvmeq->sq_cmds_io)
412 		memcpy_toio(&nvmeq->sq_cmds_io[tail], cmd, sizeof(*cmd));
413 	else
414 		memcpy(&nvmeq->sq_cmds[tail], cmd, sizeof(*cmd));
415 
416 	if (++tail == nvmeq->q_depth)
417 		tail = 0;
418 	if (nvme_dbbuf_update_and_check_event(tail, nvmeq->dbbuf_sq_db,
419 					      nvmeq->dbbuf_sq_ei))
420 		writel(tail, nvmeq->q_db);
421 	nvmeq->sq_tail = tail;
422 }
423 
424 static __le64 **iod_list(struct request *req)
425 {
426 	struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
427 	return (__le64 **)(iod->sg + blk_rq_nr_phys_segments(req));
428 }
429 
430 static int nvme_init_iod(struct request *rq, struct nvme_dev *dev)
431 {
432 	struct nvme_iod *iod = blk_mq_rq_to_pdu(rq);
433 	int nseg = blk_rq_nr_phys_segments(rq);
434 	unsigned int size = blk_rq_payload_bytes(rq);
435 
436 	if (nseg > NVME_INT_PAGES || size > NVME_INT_BYTES(dev)) {
437 		iod->sg = kmalloc(nvme_iod_alloc_size(dev, size, nseg), GFP_ATOMIC);
438 		if (!iod->sg)
439 			return BLK_MQ_RQ_QUEUE_BUSY;
440 	} else {
441 		iod->sg = iod->inline_sg;
442 	}
443 
444 	iod->aborted = 0;
445 	iod->npages = -1;
446 	iod->nents = 0;
447 	iod->length = size;
448 
449 	return BLK_MQ_RQ_QUEUE_OK;
450 }
451 
452 static void nvme_free_iod(struct nvme_dev *dev, struct request *req)
453 {
454 	struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
455 	const int last_prp = dev->ctrl.page_size / 8 - 1;
456 	int i;
457 	__le64 **list = iod_list(req);
458 	dma_addr_t prp_dma = iod->first_dma;
459 
460 	if (iod->npages == 0)
461 		dma_pool_free(dev->prp_small_pool, list[0], prp_dma);
462 	for (i = 0; i < iod->npages; i++) {
463 		__le64 *prp_list = list[i];
464 		dma_addr_t next_prp_dma = le64_to_cpu(prp_list[last_prp]);
465 		dma_pool_free(dev->prp_page_pool, prp_list, prp_dma);
466 		prp_dma = next_prp_dma;
467 	}
468 
469 	if (iod->sg != iod->inline_sg)
470 		kfree(iod->sg);
471 }
472 
473 #ifdef CONFIG_BLK_DEV_INTEGRITY
474 static void nvme_dif_prep(u32 p, u32 v, struct t10_pi_tuple *pi)
475 {
476 	if (be32_to_cpu(pi->ref_tag) == v)
477 		pi->ref_tag = cpu_to_be32(p);
478 }
479 
480 static void nvme_dif_complete(u32 p, u32 v, struct t10_pi_tuple *pi)
481 {
482 	if (be32_to_cpu(pi->ref_tag) == p)
483 		pi->ref_tag = cpu_to_be32(v);
484 }
485 
486 /**
487  * nvme_dif_remap - remaps ref tags to bip seed and physical lba
488  *
489  * The virtual start sector is the one that was originally submitted by the
490  * block layer.	Due to partitioning, MD/DM cloning, etc. the actual physical
491  * start sector may be different. Remap protection information to match the
492  * physical LBA on writes, and back to the original seed on reads.
493  *
494  * Type 0 and 3 do not have a ref tag, so no remapping required.
495  */
496 static void nvme_dif_remap(struct request *req,
497 			void (*dif_swap)(u32 p, u32 v, struct t10_pi_tuple *pi))
498 {
499 	struct nvme_ns *ns = req->rq_disk->private_data;
500 	struct bio_integrity_payload *bip;
501 	struct t10_pi_tuple *pi;
502 	void *p, *pmap;
503 	u32 i, nlb, ts, phys, virt;
504 
505 	if (!ns->pi_type || ns->pi_type == NVME_NS_DPS_PI_TYPE3)
506 		return;
507 
508 	bip = bio_integrity(req->bio);
509 	if (!bip)
510 		return;
511 
512 	pmap = kmap_atomic(bip->bip_vec->bv_page) + bip->bip_vec->bv_offset;
513 
514 	p = pmap;
515 	virt = bip_get_seed(bip);
516 	phys = nvme_block_nr(ns, blk_rq_pos(req));
517 	nlb = (blk_rq_bytes(req) >> ns->lba_shift);
518 	ts = ns->disk->queue->integrity.tuple_size;
519 
520 	for (i = 0; i < nlb; i++, virt++, phys++) {
521 		pi = (struct t10_pi_tuple *)p;
522 		dif_swap(phys, virt, pi);
523 		p += ts;
524 	}
525 	kunmap_atomic(pmap);
526 }
527 #else /* CONFIG_BLK_DEV_INTEGRITY */
528 static void nvme_dif_remap(struct request *req,
529 			void (*dif_swap)(u32 p, u32 v, struct t10_pi_tuple *pi))
530 {
531 }
532 static void nvme_dif_prep(u32 p, u32 v, struct t10_pi_tuple *pi)
533 {
534 }
535 static void nvme_dif_complete(u32 p, u32 v, struct t10_pi_tuple *pi)
536 {
537 }
538 #endif
539 
540 static bool nvme_setup_prps(struct nvme_dev *dev, struct request *req)
541 {
542 	struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
543 	struct dma_pool *pool;
544 	int length = blk_rq_payload_bytes(req);
545 	struct scatterlist *sg = iod->sg;
546 	int dma_len = sg_dma_len(sg);
547 	u64 dma_addr = sg_dma_address(sg);
548 	u32 page_size = dev->ctrl.page_size;
549 	int offset = dma_addr & (page_size - 1);
550 	__le64 *prp_list;
551 	__le64 **list = iod_list(req);
552 	dma_addr_t prp_dma;
553 	int nprps, i;
554 
555 	length -= (page_size - offset);
556 	if (length <= 0)
557 		return true;
558 
559 	dma_len -= (page_size - offset);
560 	if (dma_len) {
561 		dma_addr += (page_size - offset);
562 	} else {
563 		sg = sg_next(sg);
564 		dma_addr = sg_dma_address(sg);
565 		dma_len = sg_dma_len(sg);
566 	}
567 
568 	if (length <= page_size) {
569 		iod->first_dma = dma_addr;
570 		return true;
571 	}
572 
573 	nprps = DIV_ROUND_UP(length, page_size);
574 	if (nprps <= (256 / 8)) {
575 		pool = dev->prp_small_pool;
576 		iod->npages = 0;
577 	} else {
578 		pool = dev->prp_page_pool;
579 		iod->npages = 1;
580 	}
581 
582 	prp_list = dma_pool_alloc(pool, GFP_ATOMIC, &prp_dma);
583 	if (!prp_list) {
584 		iod->first_dma = dma_addr;
585 		iod->npages = -1;
586 		return false;
587 	}
588 	list[0] = prp_list;
589 	iod->first_dma = prp_dma;
590 	i = 0;
591 	for (;;) {
592 		if (i == page_size >> 3) {
593 			__le64 *old_prp_list = prp_list;
594 			prp_list = dma_pool_alloc(pool, GFP_ATOMIC, &prp_dma);
595 			if (!prp_list)
596 				return false;
597 			list[iod->npages++] = prp_list;
598 			prp_list[0] = old_prp_list[i - 1];
599 			old_prp_list[i - 1] = cpu_to_le64(prp_dma);
600 			i = 1;
601 		}
602 		prp_list[i++] = cpu_to_le64(dma_addr);
603 		dma_len -= page_size;
604 		dma_addr += page_size;
605 		length -= page_size;
606 		if (length <= 0)
607 			break;
608 		if (dma_len > 0)
609 			continue;
610 		BUG_ON(dma_len < 0);
611 		sg = sg_next(sg);
612 		dma_addr = sg_dma_address(sg);
613 		dma_len = sg_dma_len(sg);
614 	}
615 
616 	return true;
617 }
618 
619 static int nvme_map_data(struct nvme_dev *dev, struct request *req,
620 		struct nvme_command *cmnd)
621 {
622 	struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
623 	struct request_queue *q = req->q;
624 	enum dma_data_direction dma_dir = rq_data_dir(req) ?
625 			DMA_TO_DEVICE : DMA_FROM_DEVICE;
626 	int ret = BLK_MQ_RQ_QUEUE_ERROR;
627 
628 	sg_init_table(iod->sg, blk_rq_nr_phys_segments(req));
629 	iod->nents = blk_rq_map_sg(q, req, iod->sg);
630 	if (!iod->nents)
631 		goto out;
632 
633 	ret = BLK_MQ_RQ_QUEUE_BUSY;
634 	if (!dma_map_sg_attrs(dev->dev, iod->sg, iod->nents, dma_dir,
635 				DMA_ATTR_NO_WARN))
636 		goto out;
637 
638 	if (!nvme_setup_prps(dev, req))
639 		goto out_unmap;
640 
641 	ret = BLK_MQ_RQ_QUEUE_ERROR;
642 	if (blk_integrity_rq(req)) {
643 		if (blk_rq_count_integrity_sg(q, req->bio) != 1)
644 			goto out_unmap;
645 
646 		sg_init_table(&iod->meta_sg, 1);
647 		if (blk_rq_map_integrity_sg(q, req->bio, &iod->meta_sg) != 1)
648 			goto out_unmap;
649 
650 		if (rq_data_dir(req))
651 			nvme_dif_remap(req, nvme_dif_prep);
652 
653 		if (!dma_map_sg(dev->dev, &iod->meta_sg, 1, dma_dir))
654 			goto out_unmap;
655 	}
656 
657 	cmnd->rw.dptr.prp1 = cpu_to_le64(sg_dma_address(iod->sg));
658 	cmnd->rw.dptr.prp2 = cpu_to_le64(iod->first_dma);
659 	if (blk_integrity_rq(req))
660 		cmnd->rw.metadata = cpu_to_le64(sg_dma_address(&iod->meta_sg));
661 	return BLK_MQ_RQ_QUEUE_OK;
662 
663 out_unmap:
664 	dma_unmap_sg(dev->dev, iod->sg, iod->nents, dma_dir);
665 out:
666 	return ret;
667 }
668 
669 static void nvme_unmap_data(struct nvme_dev *dev, struct request *req)
670 {
671 	struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
672 	enum dma_data_direction dma_dir = rq_data_dir(req) ?
673 			DMA_TO_DEVICE : DMA_FROM_DEVICE;
674 
675 	if (iod->nents) {
676 		dma_unmap_sg(dev->dev, iod->sg, iod->nents, dma_dir);
677 		if (blk_integrity_rq(req)) {
678 			if (!rq_data_dir(req))
679 				nvme_dif_remap(req, nvme_dif_complete);
680 			dma_unmap_sg(dev->dev, &iod->meta_sg, 1, dma_dir);
681 		}
682 	}
683 
684 	nvme_cleanup_cmd(req);
685 	nvme_free_iod(dev, req);
686 }
687 
688 /*
689  * NOTE: ns is NULL when called on the admin queue.
690  */
691 static int nvme_queue_rq(struct blk_mq_hw_ctx *hctx,
692 			 const struct blk_mq_queue_data *bd)
693 {
694 	struct nvme_ns *ns = hctx->queue->queuedata;
695 	struct nvme_queue *nvmeq = hctx->driver_data;
696 	struct nvme_dev *dev = nvmeq->dev;
697 	struct request *req = bd->rq;
698 	struct nvme_command cmnd;
699 	int ret = BLK_MQ_RQ_QUEUE_OK;
700 
701 	/*
702 	 * If formated with metadata, require the block layer provide a buffer
703 	 * unless this namespace is formated such that the metadata can be
704 	 * stripped/generated by the controller with PRACT=1.
705 	 */
706 	if (ns && ns->ms && !blk_integrity_rq(req)) {
707 		if (!(ns->pi_type && ns->ms == 8) &&
708 		    !blk_rq_is_passthrough(req)) {
709 			blk_mq_end_request(req, -EFAULT);
710 			return BLK_MQ_RQ_QUEUE_OK;
711 		}
712 	}
713 
714 	ret = nvme_setup_cmd(ns, req, &cmnd);
715 	if (ret != BLK_MQ_RQ_QUEUE_OK)
716 		return ret;
717 
718 	ret = nvme_init_iod(req, dev);
719 	if (ret != BLK_MQ_RQ_QUEUE_OK)
720 		goto out_free_cmd;
721 
722 	if (blk_rq_nr_phys_segments(req))
723 		ret = nvme_map_data(dev, req, &cmnd);
724 
725 	if (ret != BLK_MQ_RQ_QUEUE_OK)
726 		goto out_cleanup_iod;
727 
728 	blk_mq_start_request(req);
729 
730 	spin_lock_irq(&nvmeq->q_lock);
731 	if (unlikely(nvmeq->cq_vector < 0)) {
732 		ret = BLK_MQ_RQ_QUEUE_ERROR;
733 		spin_unlock_irq(&nvmeq->q_lock);
734 		goto out_cleanup_iod;
735 	}
736 	__nvme_submit_cmd(nvmeq, &cmnd);
737 	nvme_process_cq(nvmeq);
738 	spin_unlock_irq(&nvmeq->q_lock);
739 	return BLK_MQ_RQ_QUEUE_OK;
740 out_cleanup_iod:
741 	nvme_free_iod(dev, req);
742 out_free_cmd:
743 	nvme_cleanup_cmd(req);
744 	return ret;
745 }
746 
747 static void nvme_pci_complete_rq(struct request *req)
748 {
749 	struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
750 
751 	nvme_unmap_data(iod->nvmeq->dev, req);
752 	nvme_complete_rq(req);
753 }
754 
755 /* We read the CQE phase first to check if the rest of the entry is valid */
756 static inline bool nvme_cqe_valid(struct nvme_queue *nvmeq, u16 head,
757 		u16 phase)
758 {
759 	return (le16_to_cpu(nvmeq->cqes[head].status) & 1) == phase;
760 }
761 
762 static void __nvme_process_cq(struct nvme_queue *nvmeq, unsigned int *tag)
763 {
764 	u16 head, phase;
765 
766 	head = nvmeq->cq_head;
767 	phase = nvmeq->cq_phase;
768 
769 	while (nvme_cqe_valid(nvmeq, head, phase)) {
770 		struct nvme_completion cqe = nvmeq->cqes[head];
771 		struct request *req;
772 
773 		if (++head == nvmeq->q_depth) {
774 			head = 0;
775 			phase = !phase;
776 		}
777 
778 		if (tag && *tag == cqe.command_id)
779 			*tag = -1;
780 
781 		if (unlikely(cqe.command_id >= nvmeq->q_depth)) {
782 			dev_warn(nvmeq->dev->ctrl.device,
783 				"invalid id %d completed on queue %d\n",
784 				cqe.command_id, le16_to_cpu(cqe.sq_id));
785 			continue;
786 		}
787 
788 		/*
789 		 * AEN requests are special as they don't time out and can
790 		 * survive any kind of queue freeze and often don't respond to
791 		 * aborts.  We don't even bother to allocate a struct request
792 		 * for them but rather special case them here.
793 		 */
794 		if (unlikely(nvmeq->qid == 0 &&
795 				cqe.command_id >= NVME_AQ_BLKMQ_DEPTH)) {
796 			nvme_complete_async_event(&nvmeq->dev->ctrl,
797 					cqe.status, &cqe.result);
798 			continue;
799 		}
800 
801 		req = blk_mq_tag_to_rq(*nvmeq->tags, cqe.command_id);
802 		nvme_end_request(req, cqe.status, cqe.result);
803 	}
804 
805 	if (head == nvmeq->cq_head && phase == nvmeq->cq_phase)
806 		return;
807 
808 	if (likely(nvmeq->cq_vector >= 0))
809 		if (nvme_dbbuf_update_and_check_event(head, nvmeq->dbbuf_cq_db,
810 						      nvmeq->dbbuf_cq_ei))
811 			writel(head, nvmeq->q_db + nvmeq->dev->db_stride);
812 	nvmeq->cq_head = head;
813 	nvmeq->cq_phase = phase;
814 
815 	nvmeq->cqe_seen = 1;
816 }
817 
818 static void nvme_process_cq(struct nvme_queue *nvmeq)
819 {
820 	__nvme_process_cq(nvmeq, NULL);
821 }
822 
823 static irqreturn_t nvme_irq(int irq, void *data)
824 {
825 	irqreturn_t result;
826 	struct nvme_queue *nvmeq = data;
827 	spin_lock(&nvmeq->q_lock);
828 	nvme_process_cq(nvmeq);
829 	result = nvmeq->cqe_seen ? IRQ_HANDLED : IRQ_NONE;
830 	nvmeq->cqe_seen = 0;
831 	spin_unlock(&nvmeq->q_lock);
832 	return result;
833 }
834 
835 static irqreturn_t nvme_irq_check(int irq, void *data)
836 {
837 	struct nvme_queue *nvmeq = data;
838 	if (nvme_cqe_valid(nvmeq, nvmeq->cq_head, nvmeq->cq_phase))
839 		return IRQ_WAKE_THREAD;
840 	return IRQ_NONE;
841 }
842 
843 static int __nvme_poll(struct nvme_queue *nvmeq, unsigned int tag)
844 {
845 	if (nvme_cqe_valid(nvmeq, nvmeq->cq_head, nvmeq->cq_phase)) {
846 		spin_lock_irq(&nvmeq->q_lock);
847 		__nvme_process_cq(nvmeq, &tag);
848 		spin_unlock_irq(&nvmeq->q_lock);
849 
850 		if (tag == -1)
851 			return 1;
852 	}
853 
854 	return 0;
855 }
856 
857 static int nvme_poll(struct blk_mq_hw_ctx *hctx, unsigned int tag)
858 {
859 	struct nvme_queue *nvmeq = hctx->driver_data;
860 
861 	return __nvme_poll(nvmeq, tag);
862 }
863 
864 static void nvme_pci_submit_async_event(struct nvme_ctrl *ctrl, int aer_idx)
865 {
866 	struct nvme_dev *dev = to_nvme_dev(ctrl);
867 	struct nvme_queue *nvmeq = dev->queues[0];
868 	struct nvme_command c;
869 
870 	memset(&c, 0, sizeof(c));
871 	c.common.opcode = nvme_admin_async_event;
872 	c.common.command_id = NVME_AQ_BLKMQ_DEPTH + aer_idx;
873 
874 	spin_lock_irq(&nvmeq->q_lock);
875 	__nvme_submit_cmd(nvmeq, &c);
876 	spin_unlock_irq(&nvmeq->q_lock);
877 }
878 
879 static int adapter_delete_queue(struct nvme_dev *dev, u8 opcode, u16 id)
880 {
881 	struct nvme_command c;
882 
883 	memset(&c, 0, sizeof(c));
884 	c.delete_queue.opcode = opcode;
885 	c.delete_queue.qid = cpu_to_le16(id);
886 
887 	return nvme_submit_sync_cmd(dev->ctrl.admin_q, &c, NULL, 0);
888 }
889 
890 static int adapter_alloc_cq(struct nvme_dev *dev, u16 qid,
891 						struct nvme_queue *nvmeq)
892 {
893 	struct nvme_command c;
894 	int flags = NVME_QUEUE_PHYS_CONTIG | NVME_CQ_IRQ_ENABLED;
895 
896 	/*
897 	 * Note: we (ab)use the fact the the prp fields survive if no data
898 	 * is attached to the request.
899 	 */
900 	memset(&c, 0, sizeof(c));
901 	c.create_cq.opcode = nvme_admin_create_cq;
902 	c.create_cq.prp1 = cpu_to_le64(nvmeq->cq_dma_addr);
903 	c.create_cq.cqid = cpu_to_le16(qid);
904 	c.create_cq.qsize = cpu_to_le16(nvmeq->q_depth - 1);
905 	c.create_cq.cq_flags = cpu_to_le16(flags);
906 	c.create_cq.irq_vector = cpu_to_le16(nvmeq->cq_vector);
907 
908 	return nvme_submit_sync_cmd(dev->ctrl.admin_q, &c, NULL, 0);
909 }
910 
911 static int adapter_alloc_sq(struct nvme_dev *dev, u16 qid,
912 						struct nvme_queue *nvmeq)
913 {
914 	struct nvme_command c;
915 	int flags = NVME_QUEUE_PHYS_CONTIG;
916 
917 	/*
918 	 * Note: we (ab)use the fact the the prp fields survive if no data
919 	 * is attached to the request.
920 	 */
921 	memset(&c, 0, sizeof(c));
922 	c.create_sq.opcode = nvme_admin_create_sq;
923 	c.create_sq.prp1 = cpu_to_le64(nvmeq->sq_dma_addr);
924 	c.create_sq.sqid = cpu_to_le16(qid);
925 	c.create_sq.qsize = cpu_to_le16(nvmeq->q_depth - 1);
926 	c.create_sq.sq_flags = cpu_to_le16(flags);
927 	c.create_sq.cqid = cpu_to_le16(qid);
928 
929 	return nvme_submit_sync_cmd(dev->ctrl.admin_q, &c, NULL, 0);
930 }
931 
932 static int adapter_delete_cq(struct nvme_dev *dev, u16 cqid)
933 {
934 	return adapter_delete_queue(dev, nvme_admin_delete_cq, cqid);
935 }
936 
937 static int adapter_delete_sq(struct nvme_dev *dev, u16 sqid)
938 {
939 	return adapter_delete_queue(dev, nvme_admin_delete_sq, sqid);
940 }
941 
942 static void abort_endio(struct request *req, int error)
943 {
944 	struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
945 	struct nvme_queue *nvmeq = iod->nvmeq;
946 
947 	dev_warn(nvmeq->dev->ctrl.device,
948 		 "Abort status: 0x%x", nvme_req(req)->status);
949 	atomic_inc(&nvmeq->dev->ctrl.abort_limit);
950 	blk_mq_free_request(req);
951 }
952 
953 static enum blk_eh_timer_return nvme_timeout(struct request *req, bool reserved)
954 {
955 	struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
956 	struct nvme_queue *nvmeq = iod->nvmeq;
957 	struct nvme_dev *dev = nvmeq->dev;
958 	struct request *abort_req;
959 	struct nvme_command cmd;
960 
961 	/*
962 	 * Did we miss an interrupt?
963 	 */
964 	if (__nvme_poll(nvmeq, req->tag)) {
965 		dev_warn(dev->ctrl.device,
966 			 "I/O %d QID %d timeout, completion polled\n",
967 			 req->tag, nvmeq->qid);
968 		return BLK_EH_HANDLED;
969 	}
970 
971 	/*
972 	 * Shutdown immediately if controller times out while starting. The
973 	 * reset work will see the pci device disabled when it gets the forced
974 	 * cancellation error. All outstanding requests are completed on
975 	 * shutdown, so we return BLK_EH_HANDLED.
976 	 */
977 	if (dev->ctrl.state == NVME_CTRL_RESETTING) {
978 		dev_warn(dev->ctrl.device,
979 			 "I/O %d QID %d timeout, disable controller\n",
980 			 req->tag, nvmeq->qid);
981 		nvme_dev_disable(dev, false);
982 		nvme_req(req)->flags |= NVME_REQ_CANCELLED;
983 		return BLK_EH_HANDLED;
984 	}
985 
986 	/*
987  	 * Shutdown the controller immediately and schedule a reset if the
988  	 * command was already aborted once before and still hasn't been
989  	 * returned to the driver, or if this is the admin queue.
990 	 */
991 	if (!nvmeq->qid || iod->aborted) {
992 		dev_warn(dev->ctrl.device,
993 			 "I/O %d QID %d timeout, reset controller\n",
994 			 req->tag, nvmeq->qid);
995 		nvme_dev_disable(dev, false);
996 		nvme_reset(dev);
997 
998 		/*
999 		 * Mark the request as handled, since the inline shutdown
1000 		 * forces all outstanding requests to complete.
1001 		 */
1002 		nvme_req(req)->flags |= NVME_REQ_CANCELLED;
1003 		return BLK_EH_HANDLED;
1004 	}
1005 
1006 	if (atomic_dec_return(&dev->ctrl.abort_limit) < 0) {
1007 		atomic_inc(&dev->ctrl.abort_limit);
1008 		return BLK_EH_RESET_TIMER;
1009 	}
1010 	iod->aborted = 1;
1011 
1012 	memset(&cmd, 0, sizeof(cmd));
1013 	cmd.abort.opcode = nvme_admin_abort_cmd;
1014 	cmd.abort.cid = req->tag;
1015 	cmd.abort.sqid = cpu_to_le16(nvmeq->qid);
1016 
1017 	dev_warn(nvmeq->dev->ctrl.device,
1018 		"I/O %d QID %d timeout, aborting\n",
1019 		 req->tag, nvmeq->qid);
1020 
1021 	abort_req = nvme_alloc_request(dev->ctrl.admin_q, &cmd,
1022 			BLK_MQ_REQ_NOWAIT, NVME_QID_ANY);
1023 	if (IS_ERR(abort_req)) {
1024 		atomic_inc(&dev->ctrl.abort_limit);
1025 		return BLK_EH_RESET_TIMER;
1026 	}
1027 
1028 	abort_req->timeout = ADMIN_TIMEOUT;
1029 	abort_req->end_io_data = NULL;
1030 	blk_execute_rq_nowait(abort_req->q, NULL, abort_req, 0, abort_endio);
1031 
1032 	/*
1033 	 * The aborted req will be completed on receiving the abort req.
1034 	 * We enable the timer again. If hit twice, it'll cause a device reset,
1035 	 * as the device then is in a faulty state.
1036 	 */
1037 	return BLK_EH_RESET_TIMER;
1038 }
1039 
1040 static void nvme_free_queue(struct nvme_queue *nvmeq)
1041 {
1042 	dma_free_coherent(nvmeq->q_dmadev, CQ_SIZE(nvmeq->q_depth),
1043 				(void *)nvmeq->cqes, nvmeq->cq_dma_addr);
1044 	if (nvmeq->sq_cmds)
1045 		dma_free_coherent(nvmeq->q_dmadev, SQ_SIZE(nvmeq->q_depth),
1046 					nvmeq->sq_cmds, nvmeq->sq_dma_addr);
1047 	kfree(nvmeq);
1048 }
1049 
1050 static void nvme_free_queues(struct nvme_dev *dev, int lowest)
1051 {
1052 	int i;
1053 
1054 	for (i = dev->queue_count - 1; i >= lowest; i--) {
1055 		struct nvme_queue *nvmeq = dev->queues[i];
1056 		dev->queue_count--;
1057 		dev->queues[i] = NULL;
1058 		nvme_free_queue(nvmeq);
1059 	}
1060 }
1061 
1062 /**
1063  * nvme_suspend_queue - put queue into suspended state
1064  * @nvmeq - queue to suspend
1065  */
1066 static int nvme_suspend_queue(struct nvme_queue *nvmeq)
1067 {
1068 	int vector;
1069 
1070 	spin_lock_irq(&nvmeq->q_lock);
1071 	if (nvmeq->cq_vector == -1) {
1072 		spin_unlock_irq(&nvmeq->q_lock);
1073 		return 1;
1074 	}
1075 	vector = nvmeq->cq_vector;
1076 	nvmeq->dev->online_queues--;
1077 	nvmeq->cq_vector = -1;
1078 	spin_unlock_irq(&nvmeq->q_lock);
1079 
1080 	if (!nvmeq->qid && nvmeq->dev->ctrl.admin_q)
1081 		blk_mq_stop_hw_queues(nvmeq->dev->ctrl.admin_q);
1082 
1083 	pci_free_irq(to_pci_dev(nvmeq->dev->dev), vector, nvmeq);
1084 
1085 	return 0;
1086 }
1087 
1088 static void nvme_disable_admin_queue(struct nvme_dev *dev, bool shutdown)
1089 {
1090 	struct nvme_queue *nvmeq = dev->queues[0];
1091 
1092 	if (!nvmeq)
1093 		return;
1094 	if (nvme_suspend_queue(nvmeq))
1095 		return;
1096 
1097 	if (shutdown)
1098 		nvme_shutdown_ctrl(&dev->ctrl);
1099 	else
1100 		nvme_disable_ctrl(&dev->ctrl, lo_hi_readq(
1101 						dev->bar + NVME_REG_CAP));
1102 
1103 	spin_lock_irq(&nvmeq->q_lock);
1104 	nvme_process_cq(nvmeq);
1105 	spin_unlock_irq(&nvmeq->q_lock);
1106 }
1107 
1108 static int nvme_cmb_qdepth(struct nvme_dev *dev, int nr_io_queues,
1109 				int entry_size)
1110 {
1111 	int q_depth = dev->q_depth;
1112 	unsigned q_size_aligned = roundup(q_depth * entry_size,
1113 					  dev->ctrl.page_size);
1114 
1115 	if (q_size_aligned * nr_io_queues > dev->cmb_size) {
1116 		u64 mem_per_q = div_u64(dev->cmb_size, nr_io_queues);
1117 		mem_per_q = round_down(mem_per_q, dev->ctrl.page_size);
1118 		q_depth = div_u64(mem_per_q, entry_size);
1119 
1120 		/*
1121 		 * Ensure the reduced q_depth is above some threshold where it
1122 		 * would be better to map queues in system memory with the
1123 		 * original depth
1124 		 */
1125 		if (q_depth < 64)
1126 			return -ENOMEM;
1127 	}
1128 
1129 	return q_depth;
1130 }
1131 
1132 static int nvme_alloc_sq_cmds(struct nvme_dev *dev, struct nvme_queue *nvmeq,
1133 				int qid, int depth)
1134 {
1135 	if (qid && dev->cmb && use_cmb_sqes && NVME_CMB_SQS(dev->cmbsz)) {
1136 		unsigned offset = (qid - 1) * roundup(SQ_SIZE(depth),
1137 						      dev->ctrl.page_size);
1138 		nvmeq->sq_dma_addr = dev->cmb_dma_addr + offset;
1139 		nvmeq->sq_cmds_io = dev->cmb + offset;
1140 	} else {
1141 		nvmeq->sq_cmds = dma_alloc_coherent(dev->dev, SQ_SIZE(depth),
1142 					&nvmeq->sq_dma_addr, GFP_KERNEL);
1143 		if (!nvmeq->sq_cmds)
1144 			return -ENOMEM;
1145 	}
1146 
1147 	return 0;
1148 }
1149 
1150 static struct nvme_queue *nvme_alloc_queue(struct nvme_dev *dev, int qid,
1151 							int depth, int node)
1152 {
1153 	struct nvme_queue *nvmeq = kzalloc_node(sizeof(*nvmeq), GFP_KERNEL,
1154 							node);
1155 	if (!nvmeq)
1156 		return NULL;
1157 
1158 	nvmeq->cqes = dma_zalloc_coherent(dev->dev, CQ_SIZE(depth),
1159 					  &nvmeq->cq_dma_addr, GFP_KERNEL);
1160 	if (!nvmeq->cqes)
1161 		goto free_nvmeq;
1162 
1163 	if (nvme_alloc_sq_cmds(dev, nvmeq, qid, depth))
1164 		goto free_cqdma;
1165 
1166 	nvmeq->q_dmadev = dev->dev;
1167 	nvmeq->dev = dev;
1168 	spin_lock_init(&nvmeq->q_lock);
1169 	nvmeq->cq_head = 0;
1170 	nvmeq->cq_phase = 1;
1171 	nvmeq->q_db = &dev->dbs[qid * 2 * dev->db_stride];
1172 	nvmeq->q_depth = depth;
1173 	nvmeq->qid = qid;
1174 	nvmeq->cq_vector = -1;
1175 	dev->queues[qid] = nvmeq;
1176 	dev->queue_count++;
1177 
1178 	return nvmeq;
1179 
1180  free_cqdma:
1181 	dma_free_coherent(dev->dev, CQ_SIZE(depth), (void *)nvmeq->cqes,
1182 							nvmeq->cq_dma_addr);
1183  free_nvmeq:
1184 	kfree(nvmeq);
1185 	return NULL;
1186 }
1187 
1188 static int queue_request_irq(struct nvme_queue *nvmeq)
1189 {
1190 	struct pci_dev *pdev = to_pci_dev(nvmeq->dev->dev);
1191 	int nr = nvmeq->dev->ctrl.instance;
1192 
1193 	if (use_threaded_interrupts) {
1194 		return pci_request_irq(pdev, nvmeq->cq_vector, nvme_irq_check,
1195 				nvme_irq, nvmeq, "nvme%dq%d", nr, nvmeq->qid);
1196 	} else {
1197 		return pci_request_irq(pdev, nvmeq->cq_vector, nvme_irq,
1198 				NULL, nvmeq, "nvme%dq%d", nr, nvmeq->qid);
1199 	}
1200 }
1201 
1202 static void nvme_init_queue(struct nvme_queue *nvmeq, u16 qid)
1203 {
1204 	struct nvme_dev *dev = nvmeq->dev;
1205 
1206 	spin_lock_irq(&nvmeq->q_lock);
1207 	nvmeq->sq_tail = 0;
1208 	nvmeq->cq_head = 0;
1209 	nvmeq->cq_phase = 1;
1210 	nvmeq->q_db = &dev->dbs[qid * 2 * dev->db_stride];
1211 	memset((void *)nvmeq->cqes, 0, CQ_SIZE(nvmeq->q_depth));
1212 	nvme_dbbuf_init(dev, nvmeq, qid);
1213 	dev->online_queues++;
1214 	spin_unlock_irq(&nvmeq->q_lock);
1215 }
1216 
1217 static int nvme_create_queue(struct nvme_queue *nvmeq, int qid)
1218 {
1219 	struct nvme_dev *dev = nvmeq->dev;
1220 	int result;
1221 
1222 	nvmeq->cq_vector = qid - 1;
1223 	result = adapter_alloc_cq(dev, qid, nvmeq);
1224 	if (result < 0)
1225 		return result;
1226 
1227 	result = adapter_alloc_sq(dev, qid, nvmeq);
1228 	if (result < 0)
1229 		goto release_cq;
1230 
1231 	result = queue_request_irq(nvmeq);
1232 	if (result < 0)
1233 		goto release_sq;
1234 
1235 	nvme_init_queue(nvmeq, qid);
1236 	return result;
1237 
1238  release_sq:
1239 	adapter_delete_sq(dev, qid);
1240  release_cq:
1241 	adapter_delete_cq(dev, qid);
1242 	return result;
1243 }
1244 
1245 static const struct blk_mq_ops nvme_mq_admin_ops = {
1246 	.queue_rq	= nvme_queue_rq,
1247 	.complete	= nvme_pci_complete_rq,
1248 	.init_hctx	= nvme_admin_init_hctx,
1249 	.exit_hctx      = nvme_admin_exit_hctx,
1250 	.init_request	= nvme_admin_init_request,
1251 	.timeout	= nvme_timeout,
1252 };
1253 
1254 static const struct blk_mq_ops nvme_mq_ops = {
1255 	.queue_rq	= nvme_queue_rq,
1256 	.complete	= nvme_pci_complete_rq,
1257 	.init_hctx	= nvme_init_hctx,
1258 	.init_request	= nvme_init_request,
1259 	.map_queues	= nvme_pci_map_queues,
1260 	.timeout	= nvme_timeout,
1261 	.poll		= nvme_poll,
1262 };
1263 
1264 static void nvme_dev_remove_admin(struct nvme_dev *dev)
1265 {
1266 	if (dev->ctrl.admin_q && !blk_queue_dying(dev->ctrl.admin_q)) {
1267 		/*
1268 		 * If the controller was reset during removal, it's possible
1269 		 * user requests may be waiting on a stopped queue. Start the
1270 		 * queue to flush these to completion.
1271 		 */
1272 		blk_mq_start_stopped_hw_queues(dev->ctrl.admin_q, true);
1273 		blk_cleanup_queue(dev->ctrl.admin_q);
1274 		blk_mq_free_tag_set(&dev->admin_tagset);
1275 	}
1276 }
1277 
1278 static int nvme_alloc_admin_tags(struct nvme_dev *dev)
1279 {
1280 	if (!dev->ctrl.admin_q) {
1281 		dev->admin_tagset.ops = &nvme_mq_admin_ops;
1282 		dev->admin_tagset.nr_hw_queues = 1;
1283 
1284 		/*
1285 		 * Subtract one to leave an empty queue entry for 'Full Queue'
1286 		 * condition. See NVM-Express 1.2 specification, section 4.1.2.
1287 		 */
1288 		dev->admin_tagset.queue_depth = NVME_AQ_BLKMQ_DEPTH - 1;
1289 		dev->admin_tagset.timeout = ADMIN_TIMEOUT;
1290 		dev->admin_tagset.numa_node = dev_to_node(dev->dev);
1291 		dev->admin_tagset.cmd_size = nvme_cmd_size(dev);
1292 		dev->admin_tagset.flags = BLK_MQ_F_NO_SCHED;
1293 		dev->admin_tagset.driver_data = dev;
1294 
1295 		if (blk_mq_alloc_tag_set(&dev->admin_tagset))
1296 			return -ENOMEM;
1297 
1298 		dev->ctrl.admin_q = blk_mq_init_queue(&dev->admin_tagset);
1299 		if (IS_ERR(dev->ctrl.admin_q)) {
1300 			blk_mq_free_tag_set(&dev->admin_tagset);
1301 			return -ENOMEM;
1302 		}
1303 		if (!blk_get_queue(dev->ctrl.admin_q)) {
1304 			nvme_dev_remove_admin(dev);
1305 			dev->ctrl.admin_q = NULL;
1306 			return -ENODEV;
1307 		}
1308 	} else
1309 		blk_mq_start_stopped_hw_queues(dev->ctrl.admin_q, true);
1310 
1311 	return 0;
1312 }
1313 
1314 static int nvme_configure_admin_queue(struct nvme_dev *dev)
1315 {
1316 	int result;
1317 	u32 aqa;
1318 	u64 cap = lo_hi_readq(dev->bar + NVME_REG_CAP);
1319 	struct nvme_queue *nvmeq;
1320 
1321 	dev->subsystem = readl(dev->bar + NVME_REG_VS) >= NVME_VS(1, 1, 0) ?
1322 						NVME_CAP_NSSRC(cap) : 0;
1323 
1324 	if (dev->subsystem &&
1325 	    (readl(dev->bar + NVME_REG_CSTS) & NVME_CSTS_NSSRO))
1326 		writel(NVME_CSTS_NSSRO, dev->bar + NVME_REG_CSTS);
1327 
1328 	result = nvme_disable_ctrl(&dev->ctrl, cap);
1329 	if (result < 0)
1330 		return result;
1331 
1332 	nvmeq = dev->queues[0];
1333 	if (!nvmeq) {
1334 		nvmeq = nvme_alloc_queue(dev, 0, NVME_AQ_DEPTH,
1335 					dev_to_node(dev->dev));
1336 		if (!nvmeq)
1337 			return -ENOMEM;
1338 	}
1339 
1340 	aqa = nvmeq->q_depth - 1;
1341 	aqa |= aqa << 16;
1342 
1343 	writel(aqa, dev->bar + NVME_REG_AQA);
1344 	lo_hi_writeq(nvmeq->sq_dma_addr, dev->bar + NVME_REG_ASQ);
1345 	lo_hi_writeq(nvmeq->cq_dma_addr, dev->bar + NVME_REG_ACQ);
1346 
1347 	result = nvme_enable_ctrl(&dev->ctrl, cap);
1348 	if (result)
1349 		return result;
1350 
1351 	nvmeq->cq_vector = 0;
1352 	result = queue_request_irq(nvmeq);
1353 	if (result) {
1354 		nvmeq->cq_vector = -1;
1355 		return result;
1356 	}
1357 
1358 	return result;
1359 }
1360 
1361 static bool nvme_should_reset(struct nvme_dev *dev, u32 csts)
1362 {
1363 
1364 	/* If true, indicates loss of adapter communication, possibly by a
1365 	 * NVMe Subsystem reset.
1366 	 */
1367 	bool nssro = dev->subsystem && (csts & NVME_CSTS_NSSRO);
1368 
1369 	/* If there is a reset ongoing, we shouldn't reset again. */
1370 	if (work_busy(&dev->reset_work))
1371 		return false;
1372 
1373 	/* We shouldn't reset unless the controller is on fatal error state
1374 	 * _or_ if we lost the communication with it.
1375 	 */
1376 	if (!(csts & NVME_CSTS_CFS) && !nssro)
1377 		return false;
1378 
1379 	/* If PCI error recovery process is happening, we cannot reset or
1380 	 * the recovery mechanism will surely fail.
1381 	 */
1382 	if (pci_channel_offline(to_pci_dev(dev->dev)))
1383 		return false;
1384 
1385 	return true;
1386 }
1387 
1388 static void nvme_warn_reset(struct nvme_dev *dev, u32 csts)
1389 {
1390 	/* Read a config register to help see what died. */
1391 	u16 pci_status;
1392 	int result;
1393 
1394 	result = pci_read_config_word(to_pci_dev(dev->dev), PCI_STATUS,
1395 				      &pci_status);
1396 	if (result == PCIBIOS_SUCCESSFUL)
1397 		dev_warn(dev->ctrl.device,
1398 			 "controller is down; will reset: CSTS=0x%x, PCI_STATUS=0x%hx\n",
1399 			 csts, pci_status);
1400 	else
1401 		dev_warn(dev->ctrl.device,
1402 			 "controller is down; will reset: CSTS=0x%x, PCI_STATUS read failed (%d)\n",
1403 			 csts, result);
1404 }
1405 
1406 static void nvme_watchdog_timer(unsigned long data)
1407 {
1408 	struct nvme_dev *dev = (struct nvme_dev *)data;
1409 	u32 csts = readl(dev->bar + NVME_REG_CSTS);
1410 
1411 	/* Skip controllers under certain specific conditions. */
1412 	if (nvme_should_reset(dev, csts)) {
1413 		if (!nvme_reset(dev))
1414 			nvme_warn_reset(dev, csts);
1415 		return;
1416 	}
1417 
1418 	mod_timer(&dev->watchdog_timer, round_jiffies(jiffies + HZ));
1419 }
1420 
1421 static int nvme_create_io_queues(struct nvme_dev *dev)
1422 {
1423 	unsigned i, max;
1424 	int ret = 0;
1425 
1426 	for (i = dev->queue_count; i <= dev->max_qid; i++) {
1427 		/* vector == qid - 1, match nvme_create_queue */
1428 		if (!nvme_alloc_queue(dev, i, dev->q_depth,
1429 		     pci_irq_get_node(to_pci_dev(dev->dev), i - 1))) {
1430 			ret = -ENOMEM;
1431 			break;
1432 		}
1433 	}
1434 
1435 	max = min(dev->max_qid, dev->queue_count - 1);
1436 	for (i = dev->online_queues; i <= max; i++) {
1437 		ret = nvme_create_queue(dev->queues[i], i);
1438 		if (ret)
1439 			break;
1440 	}
1441 
1442 	/*
1443 	 * Ignore failing Create SQ/CQ commands, we can continue with less
1444 	 * than the desired aount of queues, and even a controller without
1445 	 * I/O queues an still be used to issue admin commands.  This might
1446 	 * be useful to upgrade a buggy firmware for example.
1447 	 */
1448 	return ret >= 0 ? 0 : ret;
1449 }
1450 
1451 static ssize_t nvme_cmb_show(struct device *dev,
1452 			     struct device_attribute *attr,
1453 			     char *buf)
1454 {
1455 	struct nvme_dev *ndev = to_nvme_dev(dev_get_drvdata(dev));
1456 
1457 	return scnprintf(buf, PAGE_SIZE, "cmbloc : x%08x\ncmbsz  : x%08x\n",
1458 		       ndev->cmbloc, ndev->cmbsz);
1459 }
1460 static DEVICE_ATTR(cmb, S_IRUGO, nvme_cmb_show, NULL);
1461 
1462 static void __iomem *nvme_map_cmb(struct nvme_dev *dev)
1463 {
1464 	u64 szu, size, offset;
1465 	resource_size_t bar_size;
1466 	struct pci_dev *pdev = to_pci_dev(dev->dev);
1467 	void __iomem *cmb;
1468 	dma_addr_t dma_addr;
1469 
1470 	dev->cmbsz = readl(dev->bar + NVME_REG_CMBSZ);
1471 	if (!(NVME_CMB_SZ(dev->cmbsz)))
1472 		return NULL;
1473 	dev->cmbloc = readl(dev->bar + NVME_REG_CMBLOC);
1474 
1475 	if (!use_cmb_sqes)
1476 		return NULL;
1477 
1478 	szu = (u64)1 << (12 + 4 * NVME_CMB_SZU(dev->cmbsz));
1479 	size = szu * NVME_CMB_SZ(dev->cmbsz);
1480 	offset = szu * NVME_CMB_OFST(dev->cmbloc);
1481 	bar_size = pci_resource_len(pdev, NVME_CMB_BIR(dev->cmbloc));
1482 
1483 	if (offset > bar_size)
1484 		return NULL;
1485 
1486 	/*
1487 	 * Controllers may support a CMB size larger than their BAR,
1488 	 * for example, due to being behind a bridge. Reduce the CMB to
1489 	 * the reported size of the BAR
1490 	 */
1491 	if (size > bar_size - offset)
1492 		size = bar_size - offset;
1493 
1494 	dma_addr = pci_resource_start(pdev, NVME_CMB_BIR(dev->cmbloc)) + offset;
1495 	cmb = ioremap_wc(dma_addr, size);
1496 	if (!cmb)
1497 		return NULL;
1498 
1499 	dev->cmb_dma_addr = dma_addr;
1500 	dev->cmb_size = size;
1501 	return cmb;
1502 }
1503 
1504 static inline void nvme_release_cmb(struct nvme_dev *dev)
1505 {
1506 	if (dev->cmb) {
1507 		iounmap(dev->cmb);
1508 		dev->cmb = NULL;
1509 		if (dev->cmbsz) {
1510 			sysfs_remove_file_from_group(&dev->ctrl.device->kobj,
1511 						     &dev_attr_cmb.attr, NULL);
1512 			dev->cmbsz = 0;
1513 		}
1514 	}
1515 }
1516 
1517 static size_t db_bar_size(struct nvme_dev *dev, unsigned nr_io_queues)
1518 {
1519 	return 4096 + ((nr_io_queues + 1) * 8 * dev->db_stride);
1520 }
1521 
1522 static int nvme_setup_io_queues(struct nvme_dev *dev)
1523 {
1524 	struct nvme_queue *adminq = dev->queues[0];
1525 	struct pci_dev *pdev = to_pci_dev(dev->dev);
1526 	int result, nr_io_queues, size;
1527 
1528 	nr_io_queues = num_online_cpus();
1529 	result = nvme_set_queue_count(&dev->ctrl, &nr_io_queues);
1530 	if (result < 0)
1531 		return result;
1532 
1533 	if (nr_io_queues == 0)
1534 		return 0;
1535 
1536 	if (dev->cmb && NVME_CMB_SQS(dev->cmbsz)) {
1537 		result = nvme_cmb_qdepth(dev, nr_io_queues,
1538 				sizeof(struct nvme_command));
1539 		if (result > 0)
1540 			dev->q_depth = result;
1541 		else
1542 			nvme_release_cmb(dev);
1543 	}
1544 
1545 	size = db_bar_size(dev, nr_io_queues);
1546 	if (size > 8192) {
1547 		iounmap(dev->bar);
1548 		do {
1549 			dev->bar = ioremap(pci_resource_start(pdev, 0), size);
1550 			if (dev->bar)
1551 				break;
1552 			if (!--nr_io_queues)
1553 				return -ENOMEM;
1554 			size = db_bar_size(dev, nr_io_queues);
1555 		} while (1);
1556 		dev->dbs = dev->bar + 4096;
1557 		adminq->q_db = dev->dbs;
1558 	}
1559 
1560 	/* Deregister the admin queue's interrupt */
1561 	pci_free_irq(pdev, 0, adminq);
1562 
1563 	/*
1564 	 * If we enable msix early due to not intx, disable it again before
1565 	 * setting up the full range we need.
1566 	 */
1567 	pci_free_irq_vectors(pdev);
1568 	nr_io_queues = pci_alloc_irq_vectors(pdev, 1, nr_io_queues,
1569 			PCI_IRQ_ALL_TYPES | PCI_IRQ_AFFINITY);
1570 	if (nr_io_queues <= 0)
1571 		return -EIO;
1572 	dev->max_qid = nr_io_queues;
1573 
1574 	/*
1575 	 * Should investigate if there's a performance win from allocating
1576 	 * more queues than interrupt vectors; it might allow the submission
1577 	 * path to scale better, even if the receive path is limited by the
1578 	 * number of interrupts.
1579 	 */
1580 
1581 	result = queue_request_irq(adminq);
1582 	if (result) {
1583 		adminq->cq_vector = -1;
1584 		return result;
1585 	}
1586 	return nvme_create_io_queues(dev);
1587 }
1588 
1589 static void nvme_del_queue_end(struct request *req, int error)
1590 {
1591 	struct nvme_queue *nvmeq = req->end_io_data;
1592 
1593 	blk_mq_free_request(req);
1594 	complete(&nvmeq->dev->ioq_wait);
1595 }
1596 
1597 static void nvme_del_cq_end(struct request *req, int error)
1598 {
1599 	struct nvme_queue *nvmeq = req->end_io_data;
1600 
1601 	if (!error) {
1602 		unsigned long flags;
1603 
1604 		/*
1605 		 * We might be called with the AQ q_lock held
1606 		 * and the I/O queue q_lock should always
1607 		 * nest inside the AQ one.
1608 		 */
1609 		spin_lock_irqsave_nested(&nvmeq->q_lock, flags,
1610 					SINGLE_DEPTH_NESTING);
1611 		nvme_process_cq(nvmeq);
1612 		spin_unlock_irqrestore(&nvmeq->q_lock, flags);
1613 	}
1614 
1615 	nvme_del_queue_end(req, error);
1616 }
1617 
1618 static int nvme_delete_queue(struct nvme_queue *nvmeq, u8 opcode)
1619 {
1620 	struct request_queue *q = nvmeq->dev->ctrl.admin_q;
1621 	struct request *req;
1622 	struct nvme_command cmd;
1623 
1624 	memset(&cmd, 0, sizeof(cmd));
1625 	cmd.delete_queue.opcode = opcode;
1626 	cmd.delete_queue.qid = cpu_to_le16(nvmeq->qid);
1627 
1628 	req = nvme_alloc_request(q, &cmd, BLK_MQ_REQ_NOWAIT, NVME_QID_ANY);
1629 	if (IS_ERR(req))
1630 		return PTR_ERR(req);
1631 
1632 	req->timeout = ADMIN_TIMEOUT;
1633 	req->end_io_data = nvmeq;
1634 
1635 	blk_execute_rq_nowait(q, NULL, req, false,
1636 			opcode == nvme_admin_delete_cq ?
1637 				nvme_del_cq_end : nvme_del_queue_end);
1638 	return 0;
1639 }
1640 
1641 static void nvme_disable_io_queues(struct nvme_dev *dev, int queues)
1642 {
1643 	int pass;
1644 	unsigned long timeout;
1645 	u8 opcode = nvme_admin_delete_sq;
1646 
1647 	for (pass = 0; pass < 2; pass++) {
1648 		int sent = 0, i = queues;
1649 
1650 		reinit_completion(&dev->ioq_wait);
1651  retry:
1652 		timeout = ADMIN_TIMEOUT;
1653 		for (; i > 0; i--, sent++)
1654 			if (nvme_delete_queue(dev->queues[i], opcode))
1655 				break;
1656 
1657 		while (sent--) {
1658 			timeout = wait_for_completion_io_timeout(&dev->ioq_wait, timeout);
1659 			if (timeout == 0)
1660 				return;
1661 			if (i)
1662 				goto retry;
1663 		}
1664 		opcode = nvme_admin_delete_cq;
1665 	}
1666 }
1667 
1668 /*
1669  * Return: error value if an error occurred setting up the queues or calling
1670  * Identify Device.  0 if these succeeded, even if adding some of the
1671  * namespaces failed.  At the moment, these failures are silent.  TBD which
1672  * failures should be reported.
1673  */
1674 static int nvme_dev_add(struct nvme_dev *dev)
1675 {
1676 	if (!dev->ctrl.tagset) {
1677 		dev->tagset.ops = &nvme_mq_ops;
1678 		dev->tagset.nr_hw_queues = dev->online_queues - 1;
1679 		dev->tagset.timeout = NVME_IO_TIMEOUT;
1680 		dev->tagset.numa_node = dev_to_node(dev->dev);
1681 		dev->tagset.queue_depth =
1682 				min_t(int, dev->q_depth, BLK_MQ_MAX_DEPTH) - 1;
1683 		dev->tagset.cmd_size = nvme_cmd_size(dev);
1684 		dev->tagset.flags = BLK_MQ_F_SHOULD_MERGE;
1685 		dev->tagset.driver_data = dev;
1686 
1687 		if (blk_mq_alloc_tag_set(&dev->tagset))
1688 			return 0;
1689 		dev->ctrl.tagset = &dev->tagset;
1690 
1691 		nvme_dbbuf_set(dev);
1692 	} else {
1693 		blk_mq_update_nr_hw_queues(&dev->tagset, dev->online_queues - 1);
1694 
1695 		/* Free previously allocated queues that are no longer usable */
1696 		nvme_free_queues(dev, dev->online_queues);
1697 	}
1698 
1699 	return 0;
1700 }
1701 
1702 static int nvme_pci_enable(struct nvme_dev *dev)
1703 {
1704 	u64 cap;
1705 	int result = -ENOMEM;
1706 	struct pci_dev *pdev = to_pci_dev(dev->dev);
1707 
1708 	if (pci_enable_device_mem(pdev))
1709 		return result;
1710 
1711 	pci_set_master(pdev);
1712 
1713 	if (dma_set_mask_and_coherent(dev->dev, DMA_BIT_MASK(64)) &&
1714 	    dma_set_mask_and_coherent(dev->dev, DMA_BIT_MASK(32)))
1715 		goto disable;
1716 
1717 	if (readl(dev->bar + NVME_REG_CSTS) == -1) {
1718 		result = -ENODEV;
1719 		goto disable;
1720 	}
1721 
1722 	/*
1723 	 * Some devices and/or platforms don't advertise or work with INTx
1724 	 * interrupts. Pre-enable a single MSIX or MSI vec for setup. We'll
1725 	 * adjust this later.
1726 	 */
1727 	result = pci_alloc_irq_vectors(pdev, 1, 1, PCI_IRQ_ALL_TYPES);
1728 	if (result < 0)
1729 		return result;
1730 
1731 	cap = lo_hi_readq(dev->bar + NVME_REG_CAP);
1732 
1733 	dev->q_depth = min_t(int, NVME_CAP_MQES(cap) + 1, NVME_Q_DEPTH);
1734 	dev->db_stride = 1 << NVME_CAP_STRIDE(cap);
1735 	dev->dbs = dev->bar + 4096;
1736 
1737 	/*
1738 	 * Temporary fix for the Apple controller found in the MacBook8,1 and
1739 	 * some MacBook7,1 to avoid controller resets and data loss.
1740 	 */
1741 	if (pdev->vendor == PCI_VENDOR_ID_APPLE && pdev->device == 0x2001) {
1742 		dev->q_depth = 2;
1743 		dev_warn(dev->ctrl.device, "detected Apple NVMe controller, "
1744 			"set queue depth=%u to work around controller resets\n",
1745 			dev->q_depth);
1746 	}
1747 
1748 	/*
1749 	 * CMBs can currently only exist on >=1.2 PCIe devices. We only
1750 	 * populate sysfs if a CMB is implemented. Note that we add the
1751 	 * CMB attribute to the nvme_ctrl kobj which removes the need to remove
1752 	 * it on exit. Since nvme_dev_attrs_group has no name we can pass
1753 	 * NULL as final argument to sysfs_add_file_to_group.
1754 	 */
1755 
1756 	if (readl(dev->bar + NVME_REG_VS) >= NVME_VS(1, 2, 0)) {
1757 		dev->cmb = nvme_map_cmb(dev);
1758 
1759 		if (dev->cmbsz) {
1760 			if (sysfs_add_file_to_group(&dev->ctrl.device->kobj,
1761 						    &dev_attr_cmb.attr, NULL))
1762 				dev_warn(dev->ctrl.device,
1763 					 "failed to add sysfs attribute for CMB\n");
1764 		}
1765 	}
1766 
1767 	pci_enable_pcie_error_reporting(pdev);
1768 	pci_save_state(pdev);
1769 	return 0;
1770 
1771  disable:
1772 	pci_disable_device(pdev);
1773 	return result;
1774 }
1775 
1776 static void nvme_dev_unmap(struct nvme_dev *dev)
1777 {
1778 	if (dev->bar)
1779 		iounmap(dev->bar);
1780 	pci_release_mem_regions(to_pci_dev(dev->dev));
1781 }
1782 
1783 static void nvme_pci_disable(struct nvme_dev *dev)
1784 {
1785 	struct pci_dev *pdev = to_pci_dev(dev->dev);
1786 
1787 	nvme_release_cmb(dev);
1788 	pci_free_irq_vectors(pdev);
1789 
1790 	if (pci_is_enabled(pdev)) {
1791 		pci_disable_pcie_error_reporting(pdev);
1792 		pci_disable_device(pdev);
1793 	}
1794 }
1795 
1796 static void nvme_dev_disable(struct nvme_dev *dev, bool shutdown)
1797 {
1798 	int i, queues;
1799 	bool dead = true;
1800 	struct pci_dev *pdev = to_pci_dev(dev->dev);
1801 
1802 	del_timer_sync(&dev->watchdog_timer);
1803 
1804 	mutex_lock(&dev->shutdown_lock);
1805 	if (pci_is_enabled(pdev)) {
1806 		u32 csts = readl(dev->bar + NVME_REG_CSTS);
1807 
1808 		if (dev->ctrl.state == NVME_CTRL_LIVE)
1809 			nvme_start_freeze(&dev->ctrl);
1810 		dead = !!((csts & NVME_CSTS_CFS) || !(csts & NVME_CSTS_RDY) ||
1811 			pdev->error_state  != pci_channel_io_normal);
1812 	}
1813 
1814 	/*
1815 	 * Give the controller a chance to complete all entered requests if
1816 	 * doing a safe shutdown.
1817 	 */
1818 	if (!dead && shutdown)
1819 		nvme_wait_freeze_timeout(&dev->ctrl, NVME_IO_TIMEOUT);
1820 	nvme_stop_queues(&dev->ctrl);
1821 
1822 	queues = dev->online_queues - 1;
1823 	for (i = dev->queue_count - 1; i > 0; i--)
1824 		nvme_suspend_queue(dev->queues[i]);
1825 
1826 	if (dead) {
1827 		/* A device might become IO incapable very soon during
1828 		 * probe, before the admin queue is configured. Thus,
1829 		 * queue_count can be 0 here.
1830 		 */
1831 		if (dev->queue_count)
1832 			nvme_suspend_queue(dev->queues[0]);
1833 	} else {
1834 		nvme_disable_io_queues(dev, queues);
1835 		nvme_disable_admin_queue(dev, shutdown);
1836 	}
1837 	nvme_pci_disable(dev);
1838 
1839 	blk_mq_tagset_busy_iter(&dev->tagset, nvme_cancel_request, &dev->ctrl);
1840 	blk_mq_tagset_busy_iter(&dev->admin_tagset, nvme_cancel_request, &dev->ctrl);
1841 
1842 	/*
1843 	 * The driver will not be starting up queues again if shutting down so
1844 	 * must flush all entered requests to their failed completion to avoid
1845 	 * deadlocking blk-mq hot-cpu notifier.
1846 	 */
1847 	if (shutdown)
1848 		nvme_start_queues(&dev->ctrl);
1849 	mutex_unlock(&dev->shutdown_lock);
1850 }
1851 
1852 static int nvme_setup_prp_pools(struct nvme_dev *dev)
1853 {
1854 	dev->prp_page_pool = dma_pool_create("prp list page", dev->dev,
1855 						PAGE_SIZE, PAGE_SIZE, 0);
1856 	if (!dev->prp_page_pool)
1857 		return -ENOMEM;
1858 
1859 	/* Optimisation for I/Os between 4k and 128k */
1860 	dev->prp_small_pool = dma_pool_create("prp list 256", dev->dev,
1861 						256, 256, 0);
1862 	if (!dev->prp_small_pool) {
1863 		dma_pool_destroy(dev->prp_page_pool);
1864 		return -ENOMEM;
1865 	}
1866 	return 0;
1867 }
1868 
1869 static void nvme_release_prp_pools(struct nvme_dev *dev)
1870 {
1871 	dma_pool_destroy(dev->prp_page_pool);
1872 	dma_pool_destroy(dev->prp_small_pool);
1873 }
1874 
1875 static void nvme_pci_free_ctrl(struct nvme_ctrl *ctrl)
1876 {
1877 	struct nvme_dev *dev = to_nvme_dev(ctrl);
1878 
1879 	nvme_dbbuf_dma_free(dev);
1880 	put_device(dev->dev);
1881 	if (dev->tagset.tags)
1882 		blk_mq_free_tag_set(&dev->tagset);
1883 	if (dev->ctrl.admin_q)
1884 		blk_put_queue(dev->ctrl.admin_q);
1885 	kfree(dev->queues);
1886 	free_opal_dev(dev->ctrl.opal_dev);
1887 	kfree(dev);
1888 }
1889 
1890 static void nvme_remove_dead_ctrl(struct nvme_dev *dev, int status)
1891 {
1892 	dev_warn(dev->ctrl.device, "Removing after probe failure status: %d\n", status);
1893 
1894 	kref_get(&dev->ctrl.kref);
1895 	nvme_dev_disable(dev, false);
1896 	if (!schedule_work(&dev->remove_work))
1897 		nvme_put_ctrl(&dev->ctrl);
1898 }
1899 
1900 static void nvme_reset_work(struct work_struct *work)
1901 {
1902 	struct nvme_dev *dev = container_of(work, struct nvme_dev, reset_work);
1903 	bool was_suspend = !!(dev->ctrl.ctrl_config & NVME_CC_SHN_NORMAL);
1904 	int result = -ENODEV;
1905 
1906 	if (WARN_ON(dev->ctrl.state == NVME_CTRL_RESETTING))
1907 		goto out;
1908 
1909 	/*
1910 	 * If we're called to reset a live controller first shut it down before
1911 	 * moving on.
1912 	 */
1913 	if (dev->ctrl.ctrl_config & NVME_CC_ENABLE)
1914 		nvme_dev_disable(dev, false);
1915 
1916 	if (!nvme_change_ctrl_state(&dev->ctrl, NVME_CTRL_RESETTING))
1917 		goto out;
1918 
1919 	result = nvme_pci_enable(dev);
1920 	if (result)
1921 		goto out;
1922 
1923 	result = nvme_configure_admin_queue(dev);
1924 	if (result)
1925 		goto out;
1926 
1927 	nvme_init_queue(dev->queues[0], 0);
1928 	result = nvme_alloc_admin_tags(dev);
1929 	if (result)
1930 		goto out;
1931 
1932 	result = nvme_init_identify(&dev->ctrl);
1933 	if (result)
1934 		goto out;
1935 
1936 	if (dev->ctrl.oacs & NVME_CTRL_OACS_SEC_SUPP) {
1937 		if (!dev->ctrl.opal_dev)
1938 			dev->ctrl.opal_dev =
1939 				init_opal_dev(&dev->ctrl, &nvme_sec_submit);
1940 		else if (was_suspend)
1941 			opal_unlock_from_suspend(dev->ctrl.opal_dev);
1942 	} else {
1943 		free_opal_dev(dev->ctrl.opal_dev);
1944 		dev->ctrl.opal_dev = NULL;
1945 	}
1946 
1947 	if (dev->ctrl.oacs & NVME_CTRL_OACS_DBBUF_SUPP) {
1948 		result = nvme_dbbuf_dma_alloc(dev);
1949 		if (result)
1950 			dev_warn(dev->dev,
1951 				 "unable to allocate dma for dbbuf\n");
1952 	}
1953 
1954 	result = nvme_setup_io_queues(dev);
1955 	if (result)
1956 		goto out;
1957 
1958 	/*
1959 	 * A controller that can not execute IO typically requires user
1960 	 * intervention to correct. For such degraded controllers, the driver
1961 	 * should not submit commands the user did not request, so skip
1962 	 * registering for asynchronous event notification on this condition.
1963 	 */
1964 	if (dev->online_queues > 1)
1965 		nvme_queue_async_events(&dev->ctrl);
1966 
1967 	mod_timer(&dev->watchdog_timer, round_jiffies(jiffies + HZ));
1968 
1969 	/*
1970 	 * Keep the controller around but remove all namespaces if we don't have
1971 	 * any working I/O queue.
1972 	 */
1973 	if (dev->online_queues < 2) {
1974 		dev_warn(dev->ctrl.device, "IO queues not created\n");
1975 		nvme_kill_queues(&dev->ctrl);
1976 		nvme_remove_namespaces(&dev->ctrl);
1977 	} else {
1978 		nvme_start_queues(&dev->ctrl);
1979 		nvme_wait_freeze(&dev->ctrl);
1980 		nvme_dev_add(dev);
1981 		nvme_unfreeze(&dev->ctrl);
1982 	}
1983 
1984 	if (!nvme_change_ctrl_state(&dev->ctrl, NVME_CTRL_LIVE)) {
1985 		dev_warn(dev->ctrl.device, "failed to mark controller live\n");
1986 		goto out;
1987 	}
1988 
1989 	if (dev->online_queues > 1)
1990 		nvme_queue_scan(&dev->ctrl);
1991 	return;
1992 
1993  out:
1994 	nvme_remove_dead_ctrl(dev, result);
1995 }
1996 
1997 static void nvme_remove_dead_ctrl_work(struct work_struct *work)
1998 {
1999 	struct nvme_dev *dev = container_of(work, struct nvme_dev, remove_work);
2000 	struct pci_dev *pdev = to_pci_dev(dev->dev);
2001 
2002 	nvme_kill_queues(&dev->ctrl);
2003 	if (pci_get_drvdata(pdev))
2004 		device_release_driver(&pdev->dev);
2005 	nvme_put_ctrl(&dev->ctrl);
2006 }
2007 
2008 static int nvme_reset(struct nvme_dev *dev)
2009 {
2010 	if (!dev->ctrl.admin_q || blk_queue_dying(dev->ctrl.admin_q))
2011 		return -ENODEV;
2012 	if (work_busy(&dev->reset_work))
2013 		return -ENODEV;
2014 	if (!queue_work(nvme_workq, &dev->reset_work))
2015 		return -EBUSY;
2016 	return 0;
2017 }
2018 
2019 static int nvme_pci_reg_read32(struct nvme_ctrl *ctrl, u32 off, u32 *val)
2020 {
2021 	*val = readl(to_nvme_dev(ctrl)->bar + off);
2022 	return 0;
2023 }
2024 
2025 static int nvme_pci_reg_write32(struct nvme_ctrl *ctrl, u32 off, u32 val)
2026 {
2027 	writel(val, to_nvme_dev(ctrl)->bar + off);
2028 	return 0;
2029 }
2030 
2031 static int nvme_pci_reg_read64(struct nvme_ctrl *ctrl, u32 off, u64 *val)
2032 {
2033 	*val = readq(to_nvme_dev(ctrl)->bar + off);
2034 	return 0;
2035 }
2036 
2037 static int nvme_pci_reset_ctrl(struct nvme_ctrl *ctrl)
2038 {
2039 	struct nvme_dev *dev = to_nvme_dev(ctrl);
2040 	int ret = nvme_reset(dev);
2041 
2042 	if (!ret)
2043 		flush_work(&dev->reset_work);
2044 	return ret;
2045 }
2046 
2047 static const struct nvme_ctrl_ops nvme_pci_ctrl_ops = {
2048 	.name			= "pcie",
2049 	.module			= THIS_MODULE,
2050 	.flags			= NVME_F_METADATA_SUPPORTED,
2051 	.reg_read32		= nvme_pci_reg_read32,
2052 	.reg_write32		= nvme_pci_reg_write32,
2053 	.reg_read64		= nvme_pci_reg_read64,
2054 	.reset_ctrl		= nvme_pci_reset_ctrl,
2055 	.free_ctrl		= nvme_pci_free_ctrl,
2056 	.submit_async_event	= nvme_pci_submit_async_event,
2057 };
2058 
2059 static int nvme_dev_map(struct nvme_dev *dev)
2060 {
2061 	struct pci_dev *pdev = to_pci_dev(dev->dev);
2062 
2063 	if (pci_request_mem_regions(pdev, "nvme"))
2064 		return -ENODEV;
2065 
2066 	dev->bar = ioremap(pci_resource_start(pdev, 0), 8192);
2067 	if (!dev->bar)
2068 		goto release;
2069 
2070 	return 0;
2071   release:
2072 	pci_release_mem_regions(pdev);
2073 	return -ENODEV;
2074 }
2075 
2076 static unsigned long check_dell_samsung_bug(struct pci_dev *pdev)
2077 {
2078 	if (pdev->vendor == 0x144d && pdev->device == 0xa802) {
2079 		/*
2080 		 * Several Samsung devices seem to drop off the PCIe bus
2081 		 * randomly when APST is on and uses the deepest sleep state.
2082 		 * This has been observed on a Samsung "SM951 NVMe SAMSUNG
2083 		 * 256GB", a "PM951 NVMe SAMSUNG 512GB", and a "Samsung SSD
2084 		 * 950 PRO 256GB", but it seems to be restricted to two Dell
2085 		 * laptops.
2086 		 */
2087 		if (dmi_match(DMI_SYS_VENDOR, "Dell Inc.") &&
2088 		    (dmi_match(DMI_PRODUCT_NAME, "XPS 15 9550") ||
2089 		     dmi_match(DMI_PRODUCT_NAME, "Precision 5510")))
2090 			return NVME_QUIRK_NO_DEEPEST_PS;
2091 	}
2092 
2093 	return 0;
2094 }
2095 
2096 static int nvme_probe(struct pci_dev *pdev, const struct pci_device_id *id)
2097 {
2098 	int node, result = -ENOMEM;
2099 	struct nvme_dev *dev;
2100 	unsigned long quirks = id->driver_data;
2101 
2102 	node = dev_to_node(&pdev->dev);
2103 	if (node == NUMA_NO_NODE)
2104 		set_dev_node(&pdev->dev, first_memory_node);
2105 
2106 	dev = kzalloc_node(sizeof(*dev), GFP_KERNEL, node);
2107 	if (!dev)
2108 		return -ENOMEM;
2109 	dev->queues = kzalloc_node((num_possible_cpus() + 1) * sizeof(void *),
2110 							GFP_KERNEL, node);
2111 	if (!dev->queues)
2112 		goto free;
2113 
2114 	dev->dev = get_device(&pdev->dev);
2115 	pci_set_drvdata(pdev, dev);
2116 
2117 	result = nvme_dev_map(dev);
2118 	if (result)
2119 		goto free;
2120 
2121 	INIT_WORK(&dev->reset_work, nvme_reset_work);
2122 	INIT_WORK(&dev->remove_work, nvme_remove_dead_ctrl_work);
2123 	setup_timer(&dev->watchdog_timer, nvme_watchdog_timer,
2124 		(unsigned long)dev);
2125 	mutex_init(&dev->shutdown_lock);
2126 	init_completion(&dev->ioq_wait);
2127 
2128 	result = nvme_setup_prp_pools(dev);
2129 	if (result)
2130 		goto put_pci;
2131 
2132 	quirks |= check_dell_samsung_bug(pdev);
2133 
2134 	result = nvme_init_ctrl(&dev->ctrl, &pdev->dev, &nvme_pci_ctrl_ops,
2135 			quirks);
2136 	if (result)
2137 		goto release_pools;
2138 
2139 	dev_info(dev->ctrl.device, "pci function %s\n", dev_name(&pdev->dev));
2140 
2141 	queue_work(nvme_workq, &dev->reset_work);
2142 	return 0;
2143 
2144  release_pools:
2145 	nvme_release_prp_pools(dev);
2146  put_pci:
2147 	put_device(dev->dev);
2148 	nvme_dev_unmap(dev);
2149  free:
2150 	kfree(dev->queues);
2151 	kfree(dev);
2152 	return result;
2153 }
2154 
2155 static void nvme_reset_notify(struct pci_dev *pdev, bool prepare)
2156 {
2157 	struct nvme_dev *dev = pci_get_drvdata(pdev);
2158 
2159 	if (prepare)
2160 		nvme_dev_disable(dev, false);
2161 	else
2162 		nvme_reset(dev);
2163 }
2164 
2165 static void nvme_shutdown(struct pci_dev *pdev)
2166 {
2167 	struct nvme_dev *dev = pci_get_drvdata(pdev);
2168 	nvme_dev_disable(dev, true);
2169 }
2170 
2171 /*
2172  * The driver's remove may be called on a device in a partially initialized
2173  * state. This function must not have any dependencies on the device state in
2174  * order to proceed.
2175  */
2176 static void nvme_remove(struct pci_dev *pdev)
2177 {
2178 	struct nvme_dev *dev = pci_get_drvdata(pdev);
2179 
2180 	nvme_change_ctrl_state(&dev->ctrl, NVME_CTRL_DELETING);
2181 
2182 	pci_set_drvdata(pdev, NULL);
2183 
2184 	if (!pci_device_is_present(pdev)) {
2185 		nvme_change_ctrl_state(&dev->ctrl, NVME_CTRL_DEAD);
2186 		nvme_dev_disable(dev, false);
2187 	}
2188 
2189 	flush_work(&dev->reset_work);
2190 	nvme_uninit_ctrl(&dev->ctrl);
2191 	nvme_dev_disable(dev, true);
2192 	nvme_dev_remove_admin(dev);
2193 	nvme_free_queues(dev, 0);
2194 	nvme_release_prp_pools(dev);
2195 	nvme_dev_unmap(dev);
2196 	nvme_put_ctrl(&dev->ctrl);
2197 }
2198 
2199 static int nvme_pci_sriov_configure(struct pci_dev *pdev, int numvfs)
2200 {
2201 	int ret = 0;
2202 
2203 	if (numvfs == 0) {
2204 		if (pci_vfs_assigned(pdev)) {
2205 			dev_warn(&pdev->dev,
2206 				"Cannot disable SR-IOV VFs while assigned\n");
2207 			return -EPERM;
2208 		}
2209 		pci_disable_sriov(pdev);
2210 		return 0;
2211 	}
2212 
2213 	ret = pci_enable_sriov(pdev, numvfs);
2214 	return ret ? ret : numvfs;
2215 }
2216 
2217 #ifdef CONFIG_PM_SLEEP
2218 static int nvme_suspend(struct device *dev)
2219 {
2220 	struct pci_dev *pdev = to_pci_dev(dev);
2221 	struct nvme_dev *ndev = pci_get_drvdata(pdev);
2222 
2223 	nvme_dev_disable(ndev, true);
2224 	return 0;
2225 }
2226 
2227 static int nvme_resume(struct device *dev)
2228 {
2229 	struct pci_dev *pdev = to_pci_dev(dev);
2230 	struct nvme_dev *ndev = pci_get_drvdata(pdev);
2231 
2232 	nvme_reset(ndev);
2233 	return 0;
2234 }
2235 #endif
2236 
2237 static SIMPLE_DEV_PM_OPS(nvme_dev_pm_ops, nvme_suspend, nvme_resume);
2238 
2239 static pci_ers_result_t nvme_error_detected(struct pci_dev *pdev,
2240 						pci_channel_state_t state)
2241 {
2242 	struct nvme_dev *dev = pci_get_drvdata(pdev);
2243 
2244 	/*
2245 	 * A frozen channel requires a reset. When detected, this method will
2246 	 * shutdown the controller to quiesce. The controller will be restarted
2247 	 * after the slot reset through driver's slot_reset callback.
2248 	 */
2249 	switch (state) {
2250 	case pci_channel_io_normal:
2251 		return PCI_ERS_RESULT_CAN_RECOVER;
2252 	case pci_channel_io_frozen:
2253 		dev_warn(dev->ctrl.device,
2254 			"frozen state error detected, reset controller\n");
2255 		nvme_dev_disable(dev, false);
2256 		return PCI_ERS_RESULT_NEED_RESET;
2257 	case pci_channel_io_perm_failure:
2258 		dev_warn(dev->ctrl.device,
2259 			"failure state error detected, request disconnect\n");
2260 		return PCI_ERS_RESULT_DISCONNECT;
2261 	}
2262 	return PCI_ERS_RESULT_NEED_RESET;
2263 }
2264 
2265 static pci_ers_result_t nvme_slot_reset(struct pci_dev *pdev)
2266 {
2267 	struct nvme_dev *dev = pci_get_drvdata(pdev);
2268 
2269 	dev_info(dev->ctrl.device, "restart after slot reset\n");
2270 	pci_restore_state(pdev);
2271 	nvme_reset(dev);
2272 	return PCI_ERS_RESULT_RECOVERED;
2273 }
2274 
2275 static void nvme_error_resume(struct pci_dev *pdev)
2276 {
2277 	pci_cleanup_aer_uncorrect_error_status(pdev);
2278 }
2279 
2280 static const struct pci_error_handlers nvme_err_handler = {
2281 	.error_detected	= nvme_error_detected,
2282 	.slot_reset	= nvme_slot_reset,
2283 	.resume		= nvme_error_resume,
2284 	.reset_notify	= nvme_reset_notify,
2285 };
2286 
2287 static const struct pci_device_id nvme_id_table[] = {
2288 	{ PCI_VDEVICE(INTEL, 0x0953),
2289 		.driver_data = NVME_QUIRK_STRIPE_SIZE |
2290 				NVME_QUIRK_DEALLOCATE_ZEROES, },
2291 	{ PCI_VDEVICE(INTEL, 0x0a53),
2292 		.driver_data = NVME_QUIRK_STRIPE_SIZE |
2293 				NVME_QUIRK_DEALLOCATE_ZEROES, },
2294 	{ PCI_VDEVICE(INTEL, 0x0a54),
2295 		.driver_data = NVME_QUIRK_STRIPE_SIZE |
2296 				NVME_QUIRK_DEALLOCATE_ZEROES, },
2297 	{ PCI_VDEVICE(INTEL, 0xf1a5),	/* Intel 600P/P3100 */
2298 		.driver_data = NVME_QUIRK_NO_DEEPEST_PS },
2299 	{ PCI_VDEVICE(INTEL, 0x5845),	/* Qemu emulated controller */
2300 		.driver_data = NVME_QUIRK_IDENTIFY_CNS, },
2301 	{ PCI_DEVICE(0x1c58, 0x0003),	/* HGST adapter */
2302 		.driver_data = NVME_QUIRK_DELAY_BEFORE_CHK_RDY, },
2303 	{ PCI_DEVICE(0x1c5f, 0x0540),	/* Memblaze Pblaze4 adapter */
2304 		.driver_data = NVME_QUIRK_DELAY_BEFORE_CHK_RDY, },
2305 	{ PCI_DEVICE_CLASS(PCI_CLASS_STORAGE_EXPRESS, 0xffffff) },
2306 	{ PCI_DEVICE(PCI_VENDOR_ID_APPLE, 0x2001) },
2307 	{ PCI_DEVICE(PCI_VENDOR_ID_APPLE, 0x2003) },
2308 	{ 0, }
2309 };
2310 MODULE_DEVICE_TABLE(pci, nvme_id_table);
2311 
2312 static struct pci_driver nvme_driver = {
2313 	.name		= "nvme",
2314 	.id_table	= nvme_id_table,
2315 	.probe		= nvme_probe,
2316 	.remove		= nvme_remove,
2317 	.shutdown	= nvme_shutdown,
2318 	.driver		= {
2319 		.pm	= &nvme_dev_pm_ops,
2320 	},
2321 	.sriov_configure = nvme_pci_sriov_configure,
2322 	.err_handler	= &nvme_err_handler,
2323 };
2324 
2325 static int __init nvme_init(void)
2326 {
2327 	int result;
2328 
2329 	nvme_workq = alloc_workqueue("nvme", WQ_UNBOUND | WQ_MEM_RECLAIM, 0);
2330 	if (!nvme_workq)
2331 		return -ENOMEM;
2332 
2333 	result = pci_register_driver(&nvme_driver);
2334 	if (result)
2335 		destroy_workqueue(nvme_workq);
2336 	return result;
2337 }
2338 
2339 static void __exit nvme_exit(void)
2340 {
2341 	pci_unregister_driver(&nvme_driver);
2342 	destroy_workqueue(nvme_workq);
2343 	_nvme_check_size();
2344 }
2345 
2346 MODULE_AUTHOR("Matthew Wilcox <willy@linux.intel.com>");
2347 MODULE_LICENSE("GPL");
2348 MODULE_VERSION("1.0");
2349 module_init(nvme_init);
2350 module_exit(nvme_exit);
2351