xref: /openbmc/linux/drivers/nvme/host/pci.c (revision aeb64ff3)
1 // SPDX-License-Identifier: GPL-2.0
2 /*
3  * NVM Express device driver
4  * Copyright (c) 2011-2014, Intel Corporation.
5  */
6 
7 #include <linux/aer.h>
8 #include <linux/async.h>
9 #include <linux/blkdev.h>
10 #include <linux/blk-mq.h>
11 #include <linux/blk-mq-pci.h>
12 #include <linux/dmi.h>
13 #include <linux/init.h>
14 #include <linux/interrupt.h>
15 #include <linux/io.h>
16 #include <linux/mm.h>
17 #include <linux/module.h>
18 #include <linux/mutex.h>
19 #include <linux/once.h>
20 #include <linux/pci.h>
21 #include <linux/suspend.h>
22 #include <linux/t10-pi.h>
23 #include <linux/types.h>
24 #include <linux/io-64-nonatomic-lo-hi.h>
25 #include <linux/sed-opal.h>
26 #include <linux/pci-p2pdma.h>
27 
28 #include "trace.h"
29 #include "nvme.h"
30 
31 #define SQ_SIZE(q)	((q)->q_depth << (q)->sqes)
32 #define CQ_SIZE(q)	((q)->q_depth * sizeof(struct nvme_completion))
33 
34 #define SGES_PER_PAGE	(PAGE_SIZE / sizeof(struct nvme_sgl_desc))
35 
36 /*
37  * These can be higher, but we need to ensure that any command doesn't
38  * require an sg allocation that needs more than a page of data.
39  */
40 #define NVME_MAX_KB_SZ	4096
41 #define NVME_MAX_SEGS	127
42 
43 static int use_threaded_interrupts;
44 module_param(use_threaded_interrupts, int, 0);
45 
46 static bool use_cmb_sqes = true;
47 module_param(use_cmb_sqes, bool, 0444);
48 MODULE_PARM_DESC(use_cmb_sqes, "use controller's memory buffer for I/O SQes");
49 
50 static unsigned int max_host_mem_size_mb = 128;
51 module_param(max_host_mem_size_mb, uint, 0444);
52 MODULE_PARM_DESC(max_host_mem_size_mb,
53 	"Maximum Host Memory Buffer (HMB) size per controller (in MiB)");
54 
55 static unsigned int sgl_threshold = SZ_32K;
56 module_param(sgl_threshold, uint, 0644);
57 MODULE_PARM_DESC(sgl_threshold,
58 		"Use SGLs when average request segment size is larger or equal to "
59 		"this size. Use 0 to disable SGLs.");
60 
61 static int io_queue_depth_set(const char *val, const struct kernel_param *kp);
62 static const struct kernel_param_ops io_queue_depth_ops = {
63 	.set = io_queue_depth_set,
64 	.get = param_get_int,
65 };
66 
67 static int io_queue_depth = 1024;
68 module_param_cb(io_queue_depth, &io_queue_depth_ops, &io_queue_depth, 0644);
69 MODULE_PARM_DESC(io_queue_depth, "set io queue depth, should >= 2");
70 
71 static int write_queues;
72 module_param(write_queues, int, 0644);
73 MODULE_PARM_DESC(write_queues,
74 	"Number of queues to use for writes. If not set, reads and writes "
75 	"will share a queue set.");
76 
77 static int poll_queues;
78 module_param(poll_queues, int, 0644);
79 MODULE_PARM_DESC(poll_queues, "Number of queues to use for polled IO.");
80 
81 struct nvme_dev;
82 struct nvme_queue;
83 
84 static void nvme_dev_disable(struct nvme_dev *dev, bool shutdown);
85 static bool __nvme_disable_io_queues(struct nvme_dev *dev, u8 opcode);
86 
87 /*
88  * Represents an NVM Express device.  Each nvme_dev is a PCI function.
89  */
90 struct nvme_dev {
91 	struct nvme_queue *queues;
92 	struct blk_mq_tag_set tagset;
93 	struct blk_mq_tag_set admin_tagset;
94 	u32 __iomem *dbs;
95 	struct device *dev;
96 	struct dma_pool *prp_page_pool;
97 	struct dma_pool *prp_small_pool;
98 	unsigned online_queues;
99 	unsigned max_qid;
100 	unsigned io_queues[HCTX_MAX_TYPES];
101 	unsigned int num_vecs;
102 	int q_depth;
103 	int io_sqes;
104 	u32 db_stride;
105 	void __iomem *bar;
106 	unsigned long bar_mapped_size;
107 	struct work_struct remove_work;
108 	struct mutex shutdown_lock;
109 	bool subsystem;
110 	u64 cmb_size;
111 	bool cmb_use_sqes;
112 	u32 cmbsz;
113 	u32 cmbloc;
114 	struct nvme_ctrl ctrl;
115 	u32 last_ps;
116 
117 	mempool_t *iod_mempool;
118 
119 	/* shadow doorbell buffer support: */
120 	u32 *dbbuf_dbs;
121 	dma_addr_t dbbuf_dbs_dma_addr;
122 	u32 *dbbuf_eis;
123 	dma_addr_t dbbuf_eis_dma_addr;
124 
125 	/* host memory buffer support: */
126 	u64 host_mem_size;
127 	u32 nr_host_mem_descs;
128 	dma_addr_t host_mem_descs_dma;
129 	struct nvme_host_mem_buf_desc *host_mem_descs;
130 	void **host_mem_desc_bufs;
131 };
132 
133 static int io_queue_depth_set(const char *val, const struct kernel_param *kp)
134 {
135 	int n = 0, ret;
136 
137 	ret = kstrtoint(val, 10, &n);
138 	if (ret != 0 || n < 2)
139 		return -EINVAL;
140 
141 	return param_set_int(val, kp);
142 }
143 
144 static inline unsigned int sq_idx(unsigned int qid, u32 stride)
145 {
146 	return qid * 2 * stride;
147 }
148 
149 static inline unsigned int cq_idx(unsigned int qid, u32 stride)
150 {
151 	return (qid * 2 + 1) * stride;
152 }
153 
154 static inline struct nvme_dev *to_nvme_dev(struct nvme_ctrl *ctrl)
155 {
156 	return container_of(ctrl, struct nvme_dev, ctrl);
157 }
158 
159 /*
160  * An NVM Express queue.  Each device has at least two (one for admin
161  * commands and one for I/O commands).
162  */
163 struct nvme_queue {
164 	struct nvme_dev *dev;
165 	spinlock_t sq_lock;
166 	void *sq_cmds;
167 	 /* only used for poll queues: */
168 	spinlock_t cq_poll_lock ____cacheline_aligned_in_smp;
169 	volatile struct nvme_completion *cqes;
170 	struct blk_mq_tags **tags;
171 	dma_addr_t sq_dma_addr;
172 	dma_addr_t cq_dma_addr;
173 	u32 __iomem *q_db;
174 	u16 q_depth;
175 	u16 cq_vector;
176 	u16 sq_tail;
177 	u16 last_sq_tail;
178 	u16 cq_head;
179 	u16 last_cq_head;
180 	u16 qid;
181 	u8 cq_phase;
182 	u8 sqes;
183 	unsigned long flags;
184 #define NVMEQ_ENABLED		0
185 #define NVMEQ_SQ_CMB		1
186 #define NVMEQ_DELETE_ERROR	2
187 #define NVMEQ_POLLED		3
188 	u32 *dbbuf_sq_db;
189 	u32 *dbbuf_cq_db;
190 	u32 *dbbuf_sq_ei;
191 	u32 *dbbuf_cq_ei;
192 	struct completion delete_done;
193 };
194 
195 /*
196  * The nvme_iod describes the data in an I/O.
197  *
198  * The sg pointer contains the list of PRP/SGL chunk allocations in addition
199  * to the actual struct scatterlist.
200  */
201 struct nvme_iod {
202 	struct nvme_request req;
203 	struct nvme_queue *nvmeq;
204 	bool use_sgl;
205 	int aborted;
206 	int npages;		/* In the PRP list. 0 means small pool in use */
207 	int nents;		/* Used in scatterlist */
208 	dma_addr_t first_dma;
209 	unsigned int dma_len;	/* length of single DMA segment mapping */
210 	dma_addr_t meta_dma;
211 	struct scatterlist *sg;
212 };
213 
214 static unsigned int max_io_queues(void)
215 {
216 	return num_possible_cpus() + write_queues + poll_queues;
217 }
218 
219 static unsigned int max_queue_count(void)
220 {
221 	/* IO queues + admin queue */
222 	return 1 + max_io_queues();
223 }
224 
225 static inline unsigned int nvme_dbbuf_size(u32 stride)
226 {
227 	return (max_queue_count() * 8 * stride);
228 }
229 
230 static int nvme_dbbuf_dma_alloc(struct nvme_dev *dev)
231 {
232 	unsigned int mem_size = nvme_dbbuf_size(dev->db_stride);
233 
234 	if (dev->dbbuf_dbs)
235 		return 0;
236 
237 	dev->dbbuf_dbs = dma_alloc_coherent(dev->dev, mem_size,
238 					    &dev->dbbuf_dbs_dma_addr,
239 					    GFP_KERNEL);
240 	if (!dev->dbbuf_dbs)
241 		return -ENOMEM;
242 	dev->dbbuf_eis = dma_alloc_coherent(dev->dev, mem_size,
243 					    &dev->dbbuf_eis_dma_addr,
244 					    GFP_KERNEL);
245 	if (!dev->dbbuf_eis) {
246 		dma_free_coherent(dev->dev, mem_size,
247 				  dev->dbbuf_dbs, dev->dbbuf_dbs_dma_addr);
248 		dev->dbbuf_dbs = NULL;
249 		return -ENOMEM;
250 	}
251 
252 	return 0;
253 }
254 
255 static void nvme_dbbuf_dma_free(struct nvme_dev *dev)
256 {
257 	unsigned int mem_size = nvme_dbbuf_size(dev->db_stride);
258 
259 	if (dev->dbbuf_dbs) {
260 		dma_free_coherent(dev->dev, mem_size,
261 				  dev->dbbuf_dbs, dev->dbbuf_dbs_dma_addr);
262 		dev->dbbuf_dbs = NULL;
263 	}
264 	if (dev->dbbuf_eis) {
265 		dma_free_coherent(dev->dev, mem_size,
266 				  dev->dbbuf_eis, dev->dbbuf_eis_dma_addr);
267 		dev->dbbuf_eis = NULL;
268 	}
269 }
270 
271 static void nvme_dbbuf_init(struct nvme_dev *dev,
272 			    struct nvme_queue *nvmeq, int qid)
273 {
274 	if (!dev->dbbuf_dbs || !qid)
275 		return;
276 
277 	nvmeq->dbbuf_sq_db = &dev->dbbuf_dbs[sq_idx(qid, dev->db_stride)];
278 	nvmeq->dbbuf_cq_db = &dev->dbbuf_dbs[cq_idx(qid, dev->db_stride)];
279 	nvmeq->dbbuf_sq_ei = &dev->dbbuf_eis[sq_idx(qid, dev->db_stride)];
280 	nvmeq->dbbuf_cq_ei = &dev->dbbuf_eis[cq_idx(qid, dev->db_stride)];
281 }
282 
283 static void nvme_dbbuf_set(struct nvme_dev *dev)
284 {
285 	struct nvme_command c;
286 
287 	if (!dev->dbbuf_dbs)
288 		return;
289 
290 	memset(&c, 0, sizeof(c));
291 	c.dbbuf.opcode = nvme_admin_dbbuf;
292 	c.dbbuf.prp1 = cpu_to_le64(dev->dbbuf_dbs_dma_addr);
293 	c.dbbuf.prp2 = cpu_to_le64(dev->dbbuf_eis_dma_addr);
294 
295 	if (nvme_submit_sync_cmd(dev->ctrl.admin_q, &c, NULL, 0)) {
296 		dev_warn(dev->ctrl.device, "unable to set dbbuf\n");
297 		/* Free memory and continue on */
298 		nvme_dbbuf_dma_free(dev);
299 	}
300 }
301 
302 static inline int nvme_dbbuf_need_event(u16 event_idx, u16 new_idx, u16 old)
303 {
304 	return (u16)(new_idx - event_idx - 1) < (u16)(new_idx - old);
305 }
306 
307 /* Update dbbuf and return true if an MMIO is required */
308 static bool nvme_dbbuf_update_and_check_event(u16 value, u32 *dbbuf_db,
309 					      volatile u32 *dbbuf_ei)
310 {
311 	if (dbbuf_db) {
312 		u16 old_value;
313 
314 		/*
315 		 * Ensure that the queue is written before updating
316 		 * the doorbell in memory
317 		 */
318 		wmb();
319 
320 		old_value = *dbbuf_db;
321 		*dbbuf_db = value;
322 
323 		/*
324 		 * Ensure that the doorbell is updated before reading the event
325 		 * index from memory.  The controller needs to provide similar
326 		 * ordering to ensure the envent index is updated before reading
327 		 * the doorbell.
328 		 */
329 		mb();
330 
331 		if (!nvme_dbbuf_need_event(*dbbuf_ei, value, old_value))
332 			return false;
333 	}
334 
335 	return true;
336 }
337 
338 /*
339  * Will slightly overestimate the number of pages needed.  This is OK
340  * as it only leads to a small amount of wasted memory for the lifetime of
341  * the I/O.
342  */
343 static int nvme_npages(unsigned size, struct nvme_dev *dev)
344 {
345 	unsigned nprps = DIV_ROUND_UP(size + dev->ctrl.page_size,
346 				      dev->ctrl.page_size);
347 	return DIV_ROUND_UP(8 * nprps, PAGE_SIZE - 8);
348 }
349 
350 /*
351  * Calculates the number of pages needed for the SGL segments. For example a 4k
352  * page can accommodate 256 SGL descriptors.
353  */
354 static int nvme_pci_npages_sgl(unsigned int num_seg)
355 {
356 	return DIV_ROUND_UP(num_seg * sizeof(struct nvme_sgl_desc), PAGE_SIZE);
357 }
358 
359 static unsigned int nvme_pci_iod_alloc_size(struct nvme_dev *dev,
360 		unsigned int size, unsigned int nseg, bool use_sgl)
361 {
362 	size_t alloc_size;
363 
364 	if (use_sgl)
365 		alloc_size = sizeof(__le64 *) * nvme_pci_npages_sgl(nseg);
366 	else
367 		alloc_size = sizeof(__le64 *) * nvme_npages(size, dev);
368 
369 	return alloc_size + sizeof(struct scatterlist) * nseg;
370 }
371 
372 static int nvme_admin_init_hctx(struct blk_mq_hw_ctx *hctx, void *data,
373 				unsigned int hctx_idx)
374 {
375 	struct nvme_dev *dev = data;
376 	struct nvme_queue *nvmeq = &dev->queues[0];
377 
378 	WARN_ON(hctx_idx != 0);
379 	WARN_ON(dev->admin_tagset.tags[0] != hctx->tags);
380 	WARN_ON(nvmeq->tags);
381 
382 	hctx->driver_data = nvmeq;
383 	nvmeq->tags = &dev->admin_tagset.tags[0];
384 	return 0;
385 }
386 
387 static void nvme_admin_exit_hctx(struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx)
388 {
389 	struct nvme_queue *nvmeq = hctx->driver_data;
390 
391 	nvmeq->tags = NULL;
392 }
393 
394 static int nvme_init_hctx(struct blk_mq_hw_ctx *hctx, void *data,
395 			  unsigned int hctx_idx)
396 {
397 	struct nvme_dev *dev = data;
398 	struct nvme_queue *nvmeq = &dev->queues[hctx_idx + 1];
399 
400 	if (!nvmeq->tags)
401 		nvmeq->tags = &dev->tagset.tags[hctx_idx];
402 
403 	WARN_ON(dev->tagset.tags[hctx_idx] != hctx->tags);
404 	hctx->driver_data = nvmeq;
405 	return 0;
406 }
407 
408 static int nvme_init_request(struct blk_mq_tag_set *set, struct request *req,
409 		unsigned int hctx_idx, unsigned int numa_node)
410 {
411 	struct nvme_dev *dev = set->driver_data;
412 	struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
413 	int queue_idx = (set == &dev->tagset) ? hctx_idx + 1 : 0;
414 	struct nvme_queue *nvmeq = &dev->queues[queue_idx];
415 
416 	BUG_ON(!nvmeq);
417 	iod->nvmeq = nvmeq;
418 
419 	nvme_req(req)->ctrl = &dev->ctrl;
420 	return 0;
421 }
422 
423 static int queue_irq_offset(struct nvme_dev *dev)
424 {
425 	/* if we have more than 1 vec, admin queue offsets us by 1 */
426 	if (dev->num_vecs > 1)
427 		return 1;
428 
429 	return 0;
430 }
431 
432 static int nvme_pci_map_queues(struct blk_mq_tag_set *set)
433 {
434 	struct nvme_dev *dev = set->driver_data;
435 	int i, qoff, offset;
436 
437 	offset = queue_irq_offset(dev);
438 	for (i = 0, qoff = 0; i < set->nr_maps; i++) {
439 		struct blk_mq_queue_map *map = &set->map[i];
440 
441 		map->nr_queues = dev->io_queues[i];
442 		if (!map->nr_queues) {
443 			BUG_ON(i == HCTX_TYPE_DEFAULT);
444 			continue;
445 		}
446 
447 		/*
448 		 * The poll queue(s) doesn't have an IRQ (and hence IRQ
449 		 * affinity), so use the regular blk-mq cpu mapping
450 		 */
451 		map->queue_offset = qoff;
452 		if (i != HCTX_TYPE_POLL && offset)
453 			blk_mq_pci_map_queues(map, to_pci_dev(dev->dev), offset);
454 		else
455 			blk_mq_map_queues(map);
456 		qoff += map->nr_queues;
457 		offset += map->nr_queues;
458 	}
459 
460 	return 0;
461 }
462 
463 /*
464  * Write sq tail if we are asked to, or if the next command would wrap.
465  */
466 static inline void nvme_write_sq_db(struct nvme_queue *nvmeq, bool write_sq)
467 {
468 	if (!write_sq) {
469 		u16 next_tail = nvmeq->sq_tail + 1;
470 
471 		if (next_tail == nvmeq->q_depth)
472 			next_tail = 0;
473 		if (next_tail != nvmeq->last_sq_tail)
474 			return;
475 	}
476 
477 	if (nvme_dbbuf_update_and_check_event(nvmeq->sq_tail,
478 			nvmeq->dbbuf_sq_db, nvmeq->dbbuf_sq_ei))
479 		writel(nvmeq->sq_tail, nvmeq->q_db);
480 	nvmeq->last_sq_tail = nvmeq->sq_tail;
481 }
482 
483 /**
484  * nvme_submit_cmd() - Copy a command into a queue and ring the doorbell
485  * @nvmeq: The queue to use
486  * @cmd: The command to send
487  * @write_sq: whether to write to the SQ doorbell
488  */
489 static void nvme_submit_cmd(struct nvme_queue *nvmeq, struct nvme_command *cmd,
490 			    bool write_sq)
491 {
492 	spin_lock(&nvmeq->sq_lock);
493 	memcpy(nvmeq->sq_cmds + (nvmeq->sq_tail << nvmeq->sqes),
494 	       cmd, sizeof(*cmd));
495 	if (++nvmeq->sq_tail == nvmeq->q_depth)
496 		nvmeq->sq_tail = 0;
497 	nvme_write_sq_db(nvmeq, write_sq);
498 	spin_unlock(&nvmeq->sq_lock);
499 }
500 
501 static void nvme_commit_rqs(struct blk_mq_hw_ctx *hctx)
502 {
503 	struct nvme_queue *nvmeq = hctx->driver_data;
504 
505 	spin_lock(&nvmeq->sq_lock);
506 	if (nvmeq->sq_tail != nvmeq->last_sq_tail)
507 		nvme_write_sq_db(nvmeq, true);
508 	spin_unlock(&nvmeq->sq_lock);
509 }
510 
511 static void **nvme_pci_iod_list(struct request *req)
512 {
513 	struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
514 	return (void **)(iod->sg + blk_rq_nr_phys_segments(req));
515 }
516 
517 static inline bool nvme_pci_use_sgls(struct nvme_dev *dev, struct request *req)
518 {
519 	struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
520 	int nseg = blk_rq_nr_phys_segments(req);
521 	unsigned int avg_seg_size;
522 
523 	if (nseg == 0)
524 		return false;
525 
526 	avg_seg_size = DIV_ROUND_UP(blk_rq_payload_bytes(req), nseg);
527 
528 	if (!(dev->ctrl.sgls & ((1 << 0) | (1 << 1))))
529 		return false;
530 	if (!iod->nvmeq->qid)
531 		return false;
532 	if (!sgl_threshold || avg_seg_size < sgl_threshold)
533 		return false;
534 	return true;
535 }
536 
537 static void nvme_unmap_data(struct nvme_dev *dev, struct request *req)
538 {
539 	struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
540 	const int last_prp = dev->ctrl.page_size / sizeof(__le64) - 1;
541 	dma_addr_t dma_addr = iod->first_dma, next_dma_addr;
542 	int i;
543 
544 	if (iod->dma_len) {
545 		dma_unmap_page(dev->dev, dma_addr, iod->dma_len,
546 			       rq_dma_dir(req));
547 		return;
548 	}
549 
550 	WARN_ON_ONCE(!iod->nents);
551 
552 	if (is_pci_p2pdma_page(sg_page(iod->sg)))
553 		pci_p2pdma_unmap_sg(dev->dev, iod->sg, iod->nents,
554 				    rq_dma_dir(req));
555 	else
556 		dma_unmap_sg(dev->dev, iod->sg, iod->nents, rq_dma_dir(req));
557 
558 
559 	if (iod->npages == 0)
560 		dma_pool_free(dev->prp_small_pool, nvme_pci_iod_list(req)[0],
561 			dma_addr);
562 
563 	for (i = 0; i < iod->npages; i++) {
564 		void *addr = nvme_pci_iod_list(req)[i];
565 
566 		if (iod->use_sgl) {
567 			struct nvme_sgl_desc *sg_list = addr;
568 
569 			next_dma_addr =
570 			    le64_to_cpu((sg_list[SGES_PER_PAGE - 1]).addr);
571 		} else {
572 			__le64 *prp_list = addr;
573 
574 			next_dma_addr = le64_to_cpu(prp_list[last_prp]);
575 		}
576 
577 		dma_pool_free(dev->prp_page_pool, addr, dma_addr);
578 		dma_addr = next_dma_addr;
579 	}
580 
581 	mempool_free(iod->sg, dev->iod_mempool);
582 }
583 
584 static void nvme_print_sgl(struct scatterlist *sgl, int nents)
585 {
586 	int i;
587 	struct scatterlist *sg;
588 
589 	for_each_sg(sgl, sg, nents, i) {
590 		dma_addr_t phys = sg_phys(sg);
591 		pr_warn("sg[%d] phys_addr:%pad offset:%d length:%d "
592 			"dma_address:%pad dma_length:%d\n",
593 			i, &phys, sg->offset, sg->length, &sg_dma_address(sg),
594 			sg_dma_len(sg));
595 	}
596 }
597 
598 static blk_status_t nvme_pci_setup_prps(struct nvme_dev *dev,
599 		struct request *req, struct nvme_rw_command *cmnd)
600 {
601 	struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
602 	struct dma_pool *pool;
603 	int length = blk_rq_payload_bytes(req);
604 	struct scatterlist *sg = iod->sg;
605 	int dma_len = sg_dma_len(sg);
606 	u64 dma_addr = sg_dma_address(sg);
607 	u32 page_size = dev->ctrl.page_size;
608 	int offset = dma_addr & (page_size - 1);
609 	__le64 *prp_list;
610 	void **list = nvme_pci_iod_list(req);
611 	dma_addr_t prp_dma;
612 	int nprps, i;
613 
614 	length -= (page_size - offset);
615 	if (length <= 0) {
616 		iod->first_dma = 0;
617 		goto done;
618 	}
619 
620 	dma_len -= (page_size - offset);
621 	if (dma_len) {
622 		dma_addr += (page_size - offset);
623 	} else {
624 		sg = sg_next(sg);
625 		dma_addr = sg_dma_address(sg);
626 		dma_len = sg_dma_len(sg);
627 	}
628 
629 	if (length <= page_size) {
630 		iod->first_dma = dma_addr;
631 		goto done;
632 	}
633 
634 	nprps = DIV_ROUND_UP(length, page_size);
635 	if (nprps <= (256 / 8)) {
636 		pool = dev->prp_small_pool;
637 		iod->npages = 0;
638 	} else {
639 		pool = dev->prp_page_pool;
640 		iod->npages = 1;
641 	}
642 
643 	prp_list = dma_pool_alloc(pool, GFP_ATOMIC, &prp_dma);
644 	if (!prp_list) {
645 		iod->first_dma = dma_addr;
646 		iod->npages = -1;
647 		return BLK_STS_RESOURCE;
648 	}
649 	list[0] = prp_list;
650 	iod->first_dma = prp_dma;
651 	i = 0;
652 	for (;;) {
653 		if (i == page_size >> 3) {
654 			__le64 *old_prp_list = prp_list;
655 			prp_list = dma_pool_alloc(pool, GFP_ATOMIC, &prp_dma);
656 			if (!prp_list)
657 				return BLK_STS_RESOURCE;
658 			list[iod->npages++] = prp_list;
659 			prp_list[0] = old_prp_list[i - 1];
660 			old_prp_list[i - 1] = cpu_to_le64(prp_dma);
661 			i = 1;
662 		}
663 		prp_list[i++] = cpu_to_le64(dma_addr);
664 		dma_len -= page_size;
665 		dma_addr += page_size;
666 		length -= page_size;
667 		if (length <= 0)
668 			break;
669 		if (dma_len > 0)
670 			continue;
671 		if (unlikely(dma_len < 0))
672 			goto bad_sgl;
673 		sg = sg_next(sg);
674 		dma_addr = sg_dma_address(sg);
675 		dma_len = sg_dma_len(sg);
676 	}
677 
678 done:
679 	cmnd->dptr.prp1 = cpu_to_le64(sg_dma_address(iod->sg));
680 	cmnd->dptr.prp2 = cpu_to_le64(iod->first_dma);
681 
682 	return BLK_STS_OK;
683 
684  bad_sgl:
685 	WARN(DO_ONCE(nvme_print_sgl, iod->sg, iod->nents),
686 			"Invalid SGL for payload:%d nents:%d\n",
687 			blk_rq_payload_bytes(req), iod->nents);
688 	return BLK_STS_IOERR;
689 }
690 
691 static void nvme_pci_sgl_set_data(struct nvme_sgl_desc *sge,
692 		struct scatterlist *sg)
693 {
694 	sge->addr = cpu_to_le64(sg_dma_address(sg));
695 	sge->length = cpu_to_le32(sg_dma_len(sg));
696 	sge->type = NVME_SGL_FMT_DATA_DESC << 4;
697 }
698 
699 static void nvme_pci_sgl_set_seg(struct nvme_sgl_desc *sge,
700 		dma_addr_t dma_addr, int entries)
701 {
702 	sge->addr = cpu_to_le64(dma_addr);
703 	if (entries < SGES_PER_PAGE) {
704 		sge->length = cpu_to_le32(entries * sizeof(*sge));
705 		sge->type = NVME_SGL_FMT_LAST_SEG_DESC << 4;
706 	} else {
707 		sge->length = cpu_to_le32(PAGE_SIZE);
708 		sge->type = NVME_SGL_FMT_SEG_DESC << 4;
709 	}
710 }
711 
712 static blk_status_t nvme_pci_setup_sgls(struct nvme_dev *dev,
713 		struct request *req, struct nvme_rw_command *cmd, int entries)
714 {
715 	struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
716 	struct dma_pool *pool;
717 	struct nvme_sgl_desc *sg_list;
718 	struct scatterlist *sg = iod->sg;
719 	dma_addr_t sgl_dma;
720 	int i = 0;
721 
722 	/* setting the transfer type as SGL */
723 	cmd->flags = NVME_CMD_SGL_METABUF;
724 
725 	if (entries == 1) {
726 		nvme_pci_sgl_set_data(&cmd->dptr.sgl, sg);
727 		return BLK_STS_OK;
728 	}
729 
730 	if (entries <= (256 / sizeof(struct nvme_sgl_desc))) {
731 		pool = dev->prp_small_pool;
732 		iod->npages = 0;
733 	} else {
734 		pool = dev->prp_page_pool;
735 		iod->npages = 1;
736 	}
737 
738 	sg_list = dma_pool_alloc(pool, GFP_ATOMIC, &sgl_dma);
739 	if (!sg_list) {
740 		iod->npages = -1;
741 		return BLK_STS_RESOURCE;
742 	}
743 
744 	nvme_pci_iod_list(req)[0] = sg_list;
745 	iod->first_dma = sgl_dma;
746 
747 	nvme_pci_sgl_set_seg(&cmd->dptr.sgl, sgl_dma, entries);
748 
749 	do {
750 		if (i == SGES_PER_PAGE) {
751 			struct nvme_sgl_desc *old_sg_desc = sg_list;
752 			struct nvme_sgl_desc *link = &old_sg_desc[i - 1];
753 
754 			sg_list = dma_pool_alloc(pool, GFP_ATOMIC, &sgl_dma);
755 			if (!sg_list)
756 				return BLK_STS_RESOURCE;
757 
758 			i = 0;
759 			nvme_pci_iod_list(req)[iod->npages++] = sg_list;
760 			sg_list[i++] = *link;
761 			nvme_pci_sgl_set_seg(link, sgl_dma, entries);
762 		}
763 
764 		nvme_pci_sgl_set_data(&sg_list[i++], sg);
765 		sg = sg_next(sg);
766 	} while (--entries > 0);
767 
768 	return BLK_STS_OK;
769 }
770 
771 static blk_status_t nvme_setup_prp_simple(struct nvme_dev *dev,
772 		struct request *req, struct nvme_rw_command *cmnd,
773 		struct bio_vec *bv)
774 {
775 	struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
776 	unsigned int offset = bv->bv_offset & (dev->ctrl.page_size - 1);
777 	unsigned int first_prp_len = dev->ctrl.page_size - offset;
778 
779 	iod->first_dma = dma_map_bvec(dev->dev, bv, rq_dma_dir(req), 0);
780 	if (dma_mapping_error(dev->dev, iod->first_dma))
781 		return BLK_STS_RESOURCE;
782 	iod->dma_len = bv->bv_len;
783 
784 	cmnd->dptr.prp1 = cpu_to_le64(iod->first_dma);
785 	if (bv->bv_len > first_prp_len)
786 		cmnd->dptr.prp2 = cpu_to_le64(iod->first_dma + first_prp_len);
787 	return 0;
788 }
789 
790 static blk_status_t nvme_setup_sgl_simple(struct nvme_dev *dev,
791 		struct request *req, struct nvme_rw_command *cmnd,
792 		struct bio_vec *bv)
793 {
794 	struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
795 
796 	iod->first_dma = dma_map_bvec(dev->dev, bv, rq_dma_dir(req), 0);
797 	if (dma_mapping_error(dev->dev, iod->first_dma))
798 		return BLK_STS_RESOURCE;
799 	iod->dma_len = bv->bv_len;
800 
801 	cmnd->flags = NVME_CMD_SGL_METABUF;
802 	cmnd->dptr.sgl.addr = cpu_to_le64(iod->first_dma);
803 	cmnd->dptr.sgl.length = cpu_to_le32(iod->dma_len);
804 	cmnd->dptr.sgl.type = NVME_SGL_FMT_DATA_DESC << 4;
805 	return 0;
806 }
807 
808 static blk_status_t nvme_map_data(struct nvme_dev *dev, struct request *req,
809 		struct nvme_command *cmnd)
810 {
811 	struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
812 	blk_status_t ret = BLK_STS_RESOURCE;
813 	int nr_mapped;
814 
815 	if (blk_rq_nr_phys_segments(req) == 1) {
816 		struct bio_vec bv = req_bvec(req);
817 
818 		if (!is_pci_p2pdma_page(bv.bv_page)) {
819 			if (bv.bv_offset + bv.bv_len <= dev->ctrl.page_size * 2)
820 				return nvme_setup_prp_simple(dev, req,
821 							     &cmnd->rw, &bv);
822 
823 			if (iod->nvmeq->qid &&
824 			    dev->ctrl.sgls & ((1 << 0) | (1 << 1)))
825 				return nvme_setup_sgl_simple(dev, req,
826 							     &cmnd->rw, &bv);
827 		}
828 	}
829 
830 	iod->dma_len = 0;
831 	iod->sg = mempool_alloc(dev->iod_mempool, GFP_ATOMIC);
832 	if (!iod->sg)
833 		return BLK_STS_RESOURCE;
834 	sg_init_table(iod->sg, blk_rq_nr_phys_segments(req));
835 	iod->nents = blk_rq_map_sg(req->q, req, iod->sg);
836 	if (!iod->nents)
837 		goto out;
838 
839 	if (is_pci_p2pdma_page(sg_page(iod->sg)))
840 		nr_mapped = pci_p2pdma_map_sg_attrs(dev->dev, iod->sg,
841 				iod->nents, rq_dma_dir(req), DMA_ATTR_NO_WARN);
842 	else
843 		nr_mapped = dma_map_sg_attrs(dev->dev, iod->sg, iod->nents,
844 					     rq_dma_dir(req), DMA_ATTR_NO_WARN);
845 	if (!nr_mapped)
846 		goto out;
847 
848 	iod->use_sgl = nvme_pci_use_sgls(dev, req);
849 	if (iod->use_sgl)
850 		ret = nvme_pci_setup_sgls(dev, req, &cmnd->rw, nr_mapped);
851 	else
852 		ret = nvme_pci_setup_prps(dev, req, &cmnd->rw);
853 out:
854 	if (ret != BLK_STS_OK)
855 		nvme_unmap_data(dev, req);
856 	return ret;
857 }
858 
859 static blk_status_t nvme_map_metadata(struct nvme_dev *dev, struct request *req,
860 		struct nvme_command *cmnd)
861 {
862 	struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
863 
864 	iod->meta_dma = dma_map_bvec(dev->dev, rq_integrity_vec(req),
865 			rq_dma_dir(req), 0);
866 	if (dma_mapping_error(dev->dev, iod->meta_dma))
867 		return BLK_STS_IOERR;
868 	cmnd->rw.metadata = cpu_to_le64(iod->meta_dma);
869 	return 0;
870 }
871 
872 /*
873  * NOTE: ns is NULL when called on the admin queue.
874  */
875 static blk_status_t nvme_queue_rq(struct blk_mq_hw_ctx *hctx,
876 			 const struct blk_mq_queue_data *bd)
877 {
878 	struct nvme_ns *ns = hctx->queue->queuedata;
879 	struct nvme_queue *nvmeq = hctx->driver_data;
880 	struct nvme_dev *dev = nvmeq->dev;
881 	struct request *req = bd->rq;
882 	struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
883 	struct nvme_command cmnd;
884 	blk_status_t ret;
885 
886 	iod->aborted = 0;
887 	iod->npages = -1;
888 	iod->nents = 0;
889 
890 	/*
891 	 * We should not need to do this, but we're still using this to
892 	 * ensure we can drain requests on a dying queue.
893 	 */
894 	if (unlikely(!test_bit(NVMEQ_ENABLED, &nvmeq->flags)))
895 		return BLK_STS_IOERR;
896 
897 	ret = nvme_setup_cmd(ns, req, &cmnd);
898 	if (ret)
899 		return ret;
900 
901 	if (blk_rq_nr_phys_segments(req)) {
902 		ret = nvme_map_data(dev, req, &cmnd);
903 		if (ret)
904 			goto out_free_cmd;
905 	}
906 
907 	if (blk_integrity_rq(req)) {
908 		ret = nvme_map_metadata(dev, req, &cmnd);
909 		if (ret)
910 			goto out_unmap_data;
911 	}
912 
913 	blk_mq_start_request(req);
914 	nvme_submit_cmd(nvmeq, &cmnd, bd->last);
915 	return BLK_STS_OK;
916 out_unmap_data:
917 	nvme_unmap_data(dev, req);
918 out_free_cmd:
919 	nvme_cleanup_cmd(req);
920 	return ret;
921 }
922 
923 static void nvme_pci_complete_rq(struct request *req)
924 {
925 	struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
926 	struct nvme_dev *dev = iod->nvmeq->dev;
927 
928 	if (blk_integrity_rq(req))
929 		dma_unmap_page(dev->dev, iod->meta_dma,
930 			       rq_integrity_vec(req)->bv_len, rq_data_dir(req));
931 	if (blk_rq_nr_phys_segments(req))
932 		nvme_unmap_data(dev, req);
933 	nvme_complete_rq(req);
934 }
935 
936 /* We read the CQE phase first to check if the rest of the entry is valid */
937 static inline bool nvme_cqe_pending(struct nvme_queue *nvmeq)
938 {
939 	return (le16_to_cpu(nvmeq->cqes[nvmeq->cq_head].status) & 1) ==
940 			nvmeq->cq_phase;
941 }
942 
943 static inline void nvme_ring_cq_doorbell(struct nvme_queue *nvmeq)
944 {
945 	u16 head = nvmeq->cq_head;
946 
947 	if (nvme_dbbuf_update_and_check_event(head, nvmeq->dbbuf_cq_db,
948 					      nvmeq->dbbuf_cq_ei))
949 		writel(head, nvmeq->q_db + nvmeq->dev->db_stride);
950 }
951 
952 static inline void nvme_handle_cqe(struct nvme_queue *nvmeq, u16 idx)
953 {
954 	volatile struct nvme_completion *cqe = &nvmeq->cqes[idx];
955 	struct request *req;
956 
957 	if (unlikely(cqe->command_id >= nvmeq->q_depth)) {
958 		dev_warn(nvmeq->dev->ctrl.device,
959 			"invalid id %d completed on queue %d\n",
960 			cqe->command_id, le16_to_cpu(cqe->sq_id));
961 		return;
962 	}
963 
964 	/*
965 	 * AEN requests are special as they don't time out and can
966 	 * survive any kind of queue freeze and often don't respond to
967 	 * aborts.  We don't even bother to allocate a struct request
968 	 * for them but rather special case them here.
969 	 */
970 	if (unlikely(nvme_is_aen_req(nvmeq->qid, cqe->command_id))) {
971 		nvme_complete_async_event(&nvmeq->dev->ctrl,
972 				cqe->status, &cqe->result);
973 		return;
974 	}
975 
976 	req = blk_mq_tag_to_rq(*nvmeq->tags, cqe->command_id);
977 	trace_nvme_sq(req, cqe->sq_head, nvmeq->sq_tail);
978 	nvme_end_request(req, cqe->status, cqe->result);
979 }
980 
981 static void nvme_complete_cqes(struct nvme_queue *nvmeq, u16 start, u16 end)
982 {
983 	while (start != end) {
984 		nvme_handle_cqe(nvmeq, start);
985 		if (++start == nvmeq->q_depth)
986 			start = 0;
987 	}
988 }
989 
990 static inline void nvme_update_cq_head(struct nvme_queue *nvmeq)
991 {
992 	if (nvmeq->cq_head == nvmeq->q_depth - 1) {
993 		nvmeq->cq_head = 0;
994 		nvmeq->cq_phase = !nvmeq->cq_phase;
995 	} else {
996 		nvmeq->cq_head++;
997 	}
998 }
999 
1000 static inline int nvme_process_cq(struct nvme_queue *nvmeq, u16 *start,
1001 				  u16 *end, unsigned int tag)
1002 {
1003 	int found = 0;
1004 
1005 	*start = nvmeq->cq_head;
1006 	while (nvme_cqe_pending(nvmeq)) {
1007 		if (tag == -1U || nvmeq->cqes[nvmeq->cq_head].command_id == tag)
1008 			found++;
1009 		nvme_update_cq_head(nvmeq);
1010 	}
1011 	*end = nvmeq->cq_head;
1012 
1013 	if (*start != *end)
1014 		nvme_ring_cq_doorbell(nvmeq);
1015 	return found;
1016 }
1017 
1018 static irqreturn_t nvme_irq(int irq, void *data)
1019 {
1020 	struct nvme_queue *nvmeq = data;
1021 	irqreturn_t ret = IRQ_NONE;
1022 	u16 start, end;
1023 
1024 	/*
1025 	 * The rmb/wmb pair ensures we see all updates from a previous run of
1026 	 * the irq handler, even if that was on another CPU.
1027 	 */
1028 	rmb();
1029 	if (nvmeq->cq_head != nvmeq->last_cq_head)
1030 		ret = IRQ_HANDLED;
1031 	nvme_process_cq(nvmeq, &start, &end, -1);
1032 	nvmeq->last_cq_head = nvmeq->cq_head;
1033 	wmb();
1034 
1035 	if (start != end) {
1036 		nvme_complete_cqes(nvmeq, start, end);
1037 		return IRQ_HANDLED;
1038 	}
1039 
1040 	return ret;
1041 }
1042 
1043 static irqreturn_t nvme_irq_check(int irq, void *data)
1044 {
1045 	struct nvme_queue *nvmeq = data;
1046 	if (nvme_cqe_pending(nvmeq))
1047 		return IRQ_WAKE_THREAD;
1048 	return IRQ_NONE;
1049 }
1050 
1051 /*
1052  * Poll for completions any queue, including those not dedicated to polling.
1053  * Can be called from any context.
1054  */
1055 static int nvme_poll_irqdisable(struct nvme_queue *nvmeq, unsigned int tag)
1056 {
1057 	struct pci_dev *pdev = to_pci_dev(nvmeq->dev->dev);
1058 	u16 start, end;
1059 	int found;
1060 
1061 	/*
1062 	 * For a poll queue we need to protect against the polling thread
1063 	 * using the CQ lock.  For normal interrupt driven threads we have
1064 	 * to disable the interrupt to avoid racing with it.
1065 	 */
1066 	if (test_bit(NVMEQ_POLLED, &nvmeq->flags)) {
1067 		spin_lock(&nvmeq->cq_poll_lock);
1068 		found = nvme_process_cq(nvmeq, &start, &end, tag);
1069 		spin_unlock(&nvmeq->cq_poll_lock);
1070 	} else {
1071 		disable_irq(pci_irq_vector(pdev, nvmeq->cq_vector));
1072 		found = nvme_process_cq(nvmeq, &start, &end, tag);
1073 		enable_irq(pci_irq_vector(pdev, nvmeq->cq_vector));
1074 	}
1075 
1076 	nvme_complete_cqes(nvmeq, start, end);
1077 	return found;
1078 }
1079 
1080 static int nvme_poll(struct blk_mq_hw_ctx *hctx)
1081 {
1082 	struct nvme_queue *nvmeq = hctx->driver_data;
1083 	u16 start, end;
1084 	bool found;
1085 
1086 	if (!nvme_cqe_pending(nvmeq))
1087 		return 0;
1088 
1089 	spin_lock(&nvmeq->cq_poll_lock);
1090 	found = nvme_process_cq(nvmeq, &start, &end, -1);
1091 	spin_unlock(&nvmeq->cq_poll_lock);
1092 
1093 	nvme_complete_cqes(nvmeq, start, end);
1094 	return found;
1095 }
1096 
1097 static void nvme_pci_submit_async_event(struct nvme_ctrl *ctrl)
1098 {
1099 	struct nvme_dev *dev = to_nvme_dev(ctrl);
1100 	struct nvme_queue *nvmeq = &dev->queues[0];
1101 	struct nvme_command c;
1102 
1103 	memset(&c, 0, sizeof(c));
1104 	c.common.opcode = nvme_admin_async_event;
1105 	c.common.command_id = NVME_AQ_BLK_MQ_DEPTH;
1106 	nvme_submit_cmd(nvmeq, &c, true);
1107 }
1108 
1109 static int adapter_delete_queue(struct nvme_dev *dev, u8 opcode, u16 id)
1110 {
1111 	struct nvme_command c;
1112 
1113 	memset(&c, 0, sizeof(c));
1114 	c.delete_queue.opcode = opcode;
1115 	c.delete_queue.qid = cpu_to_le16(id);
1116 
1117 	return nvme_submit_sync_cmd(dev->ctrl.admin_q, &c, NULL, 0);
1118 }
1119 
1120 static int adapter_alloc_cq(struct nvme_dev *dev, u16 qid,
1121 		struct nvme_queue *nvmeq, s16 vector)
1122 {
1123 	struct nvme_command c;
1124 	int flags = NVME_QUEUE_PHYS_CONTIG;
1125 
1126 	if (!test_bit(NVMEQ_POLLED, &nvmeq->flags))
1127 		flags |= NVME_CQ_IRQ_ENABLED;
1128 
1129 	/*
1130 	 * Note: we (ab)use the fact that the prp fields survive if no data
1131 	 * is attached to the request.
1132 	 */
1133 	memset(&c, 0, sizeof(c));
1134 	c.create_cq.opcode = nvme_admin_create_cq;
1135 	c.create_cq.prp1 = cpu_to_le64(nvmeq->cq_dma_addr);
1136 	c.create_cq.cqid = cpu_to_le16(qid);
1137 	c.create_cq.qsize = cpu_to_le16(nvmeq->q_depth - 1);
1138 	c.create_cq.cq_flags = cpu_to_le16(flags);
1139 	c.create_cq.irq_vector = cpu_to_le16(vector);
1140 
1141 	return nvme_submit_sync_cmd(dev->ctrl.admin_q, &c, NULL, 0);
1142 }
1143 
1144 static int adapter_alloc_sq(struct nvme_dev *dev, u16 qid,
1145 						struct nvme_queue *nvmeq)
1146 {
1147 	struct nvme_ctrl *ctrl = &dev->ctrl;
1148 	struct nvme_command c;
1149 	int flags = NVME_QUEUE_PHYS_CONTIG;
1150 
1151 	/*
1152 	 * Some drives have a bug that auto-enables WRRU if MEDIUM isn't
1153 	 * set. Since URGENT priority is zeroes, it makes all queues
1154 	 * URGENT.
1155 	 */
1156 	if (ctrl->quirks & NVME_QUIRK_MEDIUM_PRIO_SQ)
1157 		flags |= NVME_SQ_PRIO_MEDIUM;
1158 
1159 	/*
1160 	 * Note: we (ab)use the fact that the prp fields survive if no data
1161 	 * is attached to the request.
1162 	 */
1163 	memset(&c, 0, sizeof(c));
1164 	c.create_sq.opcode = nvme_admin_create_sq;
1165 	c.create_sq.prp1 = cpu_to_le64(nvmeq->sq_dma_addr);
1166 	c.create_sq.sqid = cpu_to_le16(qid);
1167 	c.create_sq.qsize = cpu_to_le16(nvmeq->q_depth - 1);
1168 	c.create_sq.sq_flags = cpu_to_le16(flags);
1169 	c.create_sq.cqid = cpu_to_le16(qid);
1170 
1171 	return nvme_submit_sync_cmd(dev->ctrl.admin_q, &c, NULL, 0);
1172 }
1173 
1174 static int adapter_delete_cq(struct nvme_dev *dev, u16 cqid)
1175 {
1176 	return adapter_delete_queue(dev, nvme_admin_delete_cq, cqid);
1177 }
1178 
1179 static int adapter_delete_sq(struct nvme_dev *dev, u16 sqid)
1180 {
1181 	return adapter_delete_queue(dev, nvme_admin_delete_sq, sqid);
1182 }
1183 
1184 static void abort_endio(struct request *req, blk_status_t error)
1185 {
1186 	struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
1187 	struct nvme_queue *nvmeq = iod->nvmeq;
1188 
1189 	dev_warn(nvmeq->dev->ctrl.device,
1190 		 "Abort status: 0x%x", nvme_req(req)->status);
1191 	atomic_inc(&nvmeq->dev->ctrl.abort_limit);
1192 	blk_mq_free_request(req);
1193 }
1194 
1195 static bool nvme_should_reset(struct nvme_dev *dev, u32 csts)
1196 {
1197 
1198 	/* If true, indicates loss of adapter communication, possibly by a
1199 	 * NVMe Subsystem reset.
1200 	 */
1201 	bool nssro = dev->subsystem && (csts & NVME_CSTS_NSSRO);
1202 
1203 	/* If there is a reset/reinit ongoing, we shouldn't reset again. */
1204 	switch (dev->ctrl.state) {
1205 	case NVME_CTRL_RESETTING:
1206 	case NVME_CTRL_CONNECTING:
1207 		return false;
1208 	default:
1209 		break;
1210 	}
1211 
1212 	/* We shouldn't reset unless the controller is on fatal error state
1213 	 * _or_ if we lost the communication with it.
1214 	 */
1215 	if (!(csts & NVME_CSTS_CFS) && !nssro)
1216 		return false;
1217 
1218 	return true;
1219 }
1220 
1221 static void nvme_warn_reset(struct nvme_dev *dev, u32 csts)
1222 {
1223 	/* Read a config register to help see what died. */
1224 	u16 pci_status;
1225 	int result;
1226 
1227 	result = pci_read_config_word(to_pci_dev(dev->dev), PCI_STATUS,
1228 				      &pci_status);
1229 	if (result == PCIBIOS_SUCCESSFUL)
1230 		dev_warn(dev->ctrl.device,
1231 			 "controller is down; will reset: CSTS=0x%x, PCI_STATUS=0x%hx\n",
1232 			 csts, pci_status);
1233 	else
1234 		dev_warn(dev->ctrl.device,
1235 			 "controller is down; will reset: CSTS=0x%x, PCI_STATUS read failed (%d)\n",
1236 			 csts, result);
1237 }
1238 
1239 static enum blk_eh_timer_return nvme_timeout(struct request *req, bool reserved)
1240 {
1241 	struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
1242 	struct nvme_queue *nvmeq = iod->nvmeq;
1243 	struct nvme_dev *dev = nvmeq->dev;
1244 	struct request *abort_req;
1245 	struct nvme_command cmd;
1246 	u32 csts = readl(dev->bar + NVME_REG_CSTS);
1247 
1248 	/* If PCI error recovery process is happening, we cannot reset or
1249 	 * the recovery mechanism will surely fail.
1250 	 */
1251 	mb();
1252 	if (pci_channel_offline(to_pci_dev(dev->dev)))
1253 		return BLK_EH_RESET_TIMER;
1254 
1255 	/*
1256 	 * Reset immediately if the controller is failed
1257 	 */
1258 	if (nvme_should_reset(dev, csts)) {
1259 		nvme_warn_reset(dev, csts);
1260 		nvme_dev_disable(dev, false);
1261 		nvme_reset_ctrl(&dev->ctrl);
1262 		return BLK_EH_DONE;
1263 	}
1264 
1265 	/*
1266 	 * Did we miss an interrupt?
1267 	 */
1268 	if (nvme_poll_irqdisable(nvmeq, req->tag)) {
1269 		dev_warn(dev->ctrl.device,
1270 			 "I/O %d QID %d timeout, completion polled\n",
1271 			 req->tag, nvmeq->qid);
1272 		return BLK_EH_DONE;
1273 	}
1274 
1275 	/*
1276 	 * Shutdown immediately if controller times out while starting. The
1277 	 * reset work will see the pci device disabled when it gets the forced
1278 	 * cancellation error. All outstanding requests are completed on
1279 	 * shutdown, so we return BLK_EH_DONE.
1280 	 */
1281 	switch (dev->ctrl.state) {
1282 	case NVME_CTRL_CONNECTING:
1283 		nvme_change_ctrl_state(&dev->ctrl, NVME_CTRL_DELETING);
1284 		/* fall through */
1285 	case NVME_CTRL_DELETING:
1286 		dev_warn_ratelimited(dev->ctrl.device,
1287 			 "I/O %d QID %d timeout, disable controller\n",
1288 			 req->tag, nvmeq->qid);
1289 		nvme_dev_disable(dev, true);
1290 		nvme_req(req)->flags |= NVME_REQ_CANCELLED;
1291 		return BLK_EH_DONE;
1292 	case NVME_CTRL_RESETTING:
1293 		return BLK_EH_RESET_TIMER;
1294 	default:
1295 		break;
1296 	}
1297 
1298 	/*
1299  	 * Shutdown the controller immediately and schedule a reset if the
1300  	 * command was already aborted once before and still hasn't been
1301  	 * returned to the driver, or if this is the admin queue.
1302 	 */
1303 	if (!nvmeq->qid || iod->aborted) {
1304 		dev_warn(dev->ctrl.device,
1305 			 "I/O %d QID %d timeout, reset controller\n",
1306 			 req->tag, nvmeq->qid);
1307 		nvme_dev_disable(dev, false);
1308 		nvme_reset_ctrl(&dev->ctrl);
1309 
1310 		nvme_req(req)->flags |= NVME_REQ_CANCELLED;
1311 		return BLK_EH_DONE;
1312 	}
1313 
1314 	if (atomic_dec_return(&dev->ctrl.abort_limit) < 0) {
1315 		atomic_inc(&dev->ctrl.abort_limit);
1316 		return BLK_EH_RESET_TIMER;
1317 	}
1318 	iod->aborted = 1;
1319 
1320 	memset(&cmd, 0, sizeof(cmd));
1321 	cmd.abort.opcode = nvme_admin_abort_cmd;
1322 	cmd.abort.cid = req->tag;
1323 	cmd.abort.sqid = cpu_to_le16(nvmeq->qid);
1324 
1325 	dev_warn(nvmeq->dev->ctrl.device,
1326 		"I/O %d QID %d timeout, aborting\n",
1327 		 req->tag, nvmeq->qid);
1328 
1329 	abort_req = nvme_alloc_request(dev->ctrl.admin_q, &cmd,
1330 			BLK_MQ_REQ_NOWAIT, NVME_QID_ANY);
1331 	if (IS_ERR(abort_req)) {
1332 		atomic_inc(&dev->ctrl.abort_limit);
1333 		return BLK_EH_RESET_TIMER;
1334 	}
1335 
1336 	abort_req->timeout = ADMIN_TIMEOUT;
1337 	abort_req->end_io_data = NULL;
1338 	blk_execute_rq_nowait(abort_req->q, NULL, abort_req, 0, abort_endio);
1339 
1340 	/*
1341 	 * The aborted req will be completed on receiving the abort req.
1342 	 * We enable the timer again. If hit twice, it'll cause a device reset,
1343 	 * as the device then is in a faulty state.
1344 	 */
1345 	return BLK_EH_RESET_TIMER;
1346 }
1347 
1348 static void nvme_free_queue(struct nvme_queue *nvmeq)
1349 {
1350 	dma_free_coherent(nvmeq->dev->dev, CQ_SIZE(nvmeq),
1351 				(void *)nvmeq->cqes, nvmeq->cq_dma_addr);
1352 	if (!nvmeq->sq_cmds)
1353 		return;
1354 
1355 	if (test_and_clear_bit(NVMEQ_SQ_CMB, &nvmeq->flags)) {
1356 		pci_free_p2pmem(to_pci_dev(nvmeq->dev->dev),
1357 				nvmeq->sq_cmds, SQ_SIZE(nvmeq));
1358 	} else {
1359 		dma_free_coherent(nvmeq->dev->dev, SQ_SIZE(nvmeq),
1360 				nvmeq->sq_cmds, nvmeq->sq_dma_addr);
1361 	}
1362 }
1363 
1364 static void nvme_free_queues(struct nvme_dev *dev, int lowest)
1365 {
1366 	int i;
1367 
1368 	for (i = dev->ctrl.queue_count - 1; i >= lowest; i--) {
1369 		dev->ctrl.queue_count--;
1370 		nvme_free_queue(&dev->queues[i]);
1371 	}
1372 }
1373 
1374 /**
1375  * nvme_suspend_queue - put queue into suspended state
1376  * @nvmeq: queue to suspend
1377  */
1378 static int nvme_suspend_queue(struct nvme_queue *nvmeq)
1379 {
1380 	if (!test_and_clear_bit(NVMEQ_ENABLED, &nvmeq->flags))
1381 		return 1;
1382 
1383 	/* ensure that nvme_queue_rq() sees NVMEQ_ENABLED cleared */
1384 	mb();
1385 
1386 	nvmeq->dev->online_queues--;
1387 	if (!nvmeq->qid && nvmeq->dev->ctrl.admin_q)
1388 		blk_mq_quiesce_queue(nvmeq->dev->ctrl.admin_q);
1389 	if (!test_and_clear_bit(NVMEQ_POLLED, &nvmeq->flags))
1390 		pci_free_irq(to_pci_dev(nvmeq->dev->dev), nvmeq->cq_vector, nvmeq);
1391 	return 0;
1392 }
1393 
1394 static void nvme_suspend_io_queues(struct nvme_dev *dev)
1395 {
1396 	int i;
1397 
1398 	for (i = dev->ctrl.queue_count - 1; i > 0; i--)
1399 		nvme_suspend_queue(&dev->queues[i]);
1400 }
1401 
1402 static void nvme_disable_admin_queue(struct nvme_dev *dev, bool shutdown)
1403 {
1404 	struct nvme_queue *nvmeq = &dev->queues[0];
1405 
1406 	if (shutdown)
1407 		nvme_shutdown_ctrl(&dev->ctrl);
1408 	else
1409 		nvme_disable_ctrl(&dev->ctrl);
1410 
1411 	nvme_poll_irqdisable(nvmeq, -1);
1412 }
1413 
1414 static int nvme_cmb_qdepth(struct nvme_dev *dev, int nr_io_queues,
1415 				int entry_size)
1416 {
1417 	int q_depth = dev->q_depth;
1418 	unsigned q_size_aligned = roundup(q_depth * entry_size,
1419 					  dev->ctrl.page_size);
1420 
1421 	if (q_size_aligned * nr_io_queues > dev->cmb_size) {
1422 		u64 mem_per_q = div_u64(dev->cmb_size, nr_io_queues);
1423 		mem_per_q = round_down(mem_per_q, dev->ctrl.page_size);
1424 		q_depth = div_u64(mem_per_q, entry_size);
1425 
1426 		/*
1427 		 * Ensure the reduced q_depth is above some threshold where it
1428 		 * would be better to map queues in system memory with the
1429 		 * original depth
1430 		 */
1431 		if (q_depth < 64)
1432 			return -ENOMEM;
1433 	}
1434 
1435 	return q_depth;
1436 }
1437 
1438 static int nvme_alloc_sq_cmds(struct nvme_dev *dev, struct nvme_queue *nvmeq,
1439 				int qid)
1440 {
1441 	struct pci_dev *pdev = to_pci_dev(dev->dev);
1442 
1443 	if (qid && dev->cmb_use_sqes && (dev->cmbsz & NVME_CMBSZ_SQS)) {
1444 		nvmeq->sq_cmds = pci_alloc_p2pmem(pdev, SQ_SIZE(nvmeq));
1445 		if (nvmeq->sq_cmds) {
1446 			nvmeq->sq_dma_addr = pci_p2pmem_virt_to_bus(pdev,
1447 							nvmeq->sq_cmds);
1448 			if (nvmeq->sq_dma_addr) {
1449 				set_bit(NVMEQ_SQ_CMB, &nvmeq->flags);
1450 				return 0;
1451 			}
1452 
1453 			pci_free_p2pmem(pdev, nvmeq->sq_cmds, SQ_SIZE(nvmeq));
1454 		}
1455 	}
1456 
1457 	nvmeq->sq_cmds = dma_alloc_coherent(dev->dev, SQ_SIZE(nvmeq),
1458 				&nvmeq->sq_dma_addr, GFP_KERNEL);
1459 	if (!nvmeq->sq_cmds)
1460 		return -ENOMEM;
1461 	return 0;
1462 }
1463 
1464 static int nvme_alloc_queue(struct nvme_dev *dev, int qid, int depth)
1465 {
1466 	struct nvme_queue *nvmeq = &dev->queues[qid];
1467 
1468 	if (dev->ctrl.queue_count > qid)
1469 		return 0;
1470 
1471 	nvmeq->sqes = qid ? dev->io_sqes : NVME_ADM_SQES;
1472 	nvmeq->q_depth = depth;
1473 	nvmeq->cqes = dma_alloc_coherent(dev->dev, CQ_SIZE(nvmeq),
1474 					 &nvmeq->cq_dma_addr, GFP_KERNEL);
1475 	if (!nvmeq->cqes)
1476 		goto free_nvmeq;
1477 
1478 	if (nvme_alloc_sq_cmds(dev, nvmeq, qid))
1479 		goto free_cqdma;
1480 
1481 	nvmeq->dev = dev;
1482 	spin_lock_init(&nvmeq->sq_lock);
1483 	spin_lock_init(&nvmeq->cq_poll_lock);
1484 	nvmeq->cq_head = 0;
1485 	nvmeq->cq_phase = 1;
1486 	nvmeq->q_db = &dev->dbs[qid * 2 * dev->db_stride];
1487 	nvmeq->qid = qid;
1488 	dev->ctrl.queue_count++;
1489 
1490 	return 0;
1491 
1492  free_cqdma:
1493 	dma_free_coherent(dev->dev, CQ_SIZE(nvmeq), (void *)nvmeq->cqes,
1494 			  nvmeq->cq_dma_addr);
1495  free_nvmeq:
1496 	return -ENOMEM;
1497 }
1498 
1499 static int queue_request_irq(struct nvme_queue *nvmeq)
1500 {
1501 	struct pci_dev *pdev = to_pci_dev(nvmeq->dev->dev);
1502 	int nr = nvmeq->dev->ctrl.instance;
1503 
1504 	if (use_threaded_interrupts) {
1505 		return pci_request_irq(pdev, nvmeq->cq_vector, nvme_irq_check,
1506 				nvme_irq, nvmeq, "nvme%dq%d", nr, nvmeq->qid);
1507 	} else {
1508 		return pci_request_irq(pdev, nvmeq->cq_vector, nvme_irq,
1509 				NULL, nvmeq, "nvme%dq%d", nr, nvmeq->qid);
1510 	}
1511 }
1512 
1513 static void nvme_init_queue(struct nvme_queue *nvmeq, u16 qid)
1514 {
1515 	struct nvme_dev *dev = nvmeq->dev;
1516 
1517 	nvmeq->sq_tail = 0;
1518 	nvmeq->last_sq_tail = 0;
1519 	nvmeq->cq_head = 0;
1520 	nvmeq->cq_phase = 1;
1521 	nvmeq->q_db = &dev->dbs[qid * 2 * dev->db_stride];
1522 	memset((void *)nvmeq->cqes, 0, CQ_SIZE(nvmeq));
1523 	nvme_dbbuf_init(dev, nvmeq, qid);
1524 	dev->online_queues++;
1525 	wmb(); /* ensure the first interrupt sees the initialization */
1526 }
1527 
1528 static int nvme_create_queue(struct nvme_queue *nvmeq, int qid, bool polled)
1529 {
1530 	struct nvme_dev *dev = nvmeq->dev;
1531 	int result;
1532 	u16 vector = 0;
1533 
1534 	clear_bit(NVMEQ_DELETE_ERROR, &nvmeq->flags);
1535 
1536 	/*
1537 	 * A queue's vector matches the queue identifier unless the controller
1538 	 * has only one vector available.
1539 	 */
1540 	if (!polled)
1541 		vector = dev->num_vecs == 1 ? 0 : qid;
1542 	else
1543 		set_bit(NVMEQ_POLLED, &nvmeq->flags);
1544 
1545 	result = adapter_alloc_cq(dev, qid, nvmeq, vector);
1546 	if (result)
1547 		return result;
1548 
1549 	result = adapter_alloc_sq(dev, qid, nvmeq);
1550 	if (result < 0)
1551 		return result;
1552 	else if (result)
1553 		goto release_cq;
1554 
1555 	nvmeq->cq_vector = vector;
1556 	nvme_init_queue(nvmeq, qid);
1557 
1558 	if (!polled) {
1559 		result = queue_request_irq(nvmeq);
1560 		if (result < 0)
1561 			goto release_sq;
1562 	}
1563 
1564 	set_bit(NVMEQ_ENABLED, &nvmeq->flags);
1565 	return result;
1566 
1567 release_sq:
1568 	dev->online_queues--;
1569 	adapter_delete_sq(dev, qid);
1570 release_cq:
1571 	adapter_delete_cq(dev, qid);
1572 	return result;
1573 }
1574 
1575 static const struct blk_mq_ops nvme_mq_admin_ops = {
1576 	.queue_rq	= nvme_queue_rq,
1577 	.complete	= nvme_pci_complete_rq,
1578 	.init_hctx	= nvme_admin_init_hctx,
1579 	.exit_hctx      = nvme_admin_exit_hctx,
1580 	.init_request	= nvme_init_request,
1581 	.timeout	= nvme_timeout,
1582 };
1583 
1584 static const struct blk_mq_ops nvme_mq_ops = {
1585 	.queue_rq	= nvme_queue_rq,
1586 	.complete	= nvme_pci_complete_rq,
1587 	.commit_rqs	= nvme_commit_rqs,
1588 	.init_hctx	= nvme_init_hctx,
1589 	.init_request	= nvme_init_request,
1590 	.map_queues	= nvme_pci_map_queues,
1591 	.timeout	= nvme_timeout,
1592 	.poll		= nvme_poll,
1593 };
1594 
1595 static void nvme_dev_remove_admin(struct nvme_dev *dev)
1596 {
1597 	if (dev->ctrl.admin_q && !blk_queue_dying(dev->ctrl.admin_q)) {
1598 		/*
1599 		 * If the controller was reset during removal, it's possible
1600 		 * user requests may be waiting on a stopped queue. Start the
1601 		 * queue to flush these to completion.
1602 		 */
1603 		blk_mq_unquiesce_queue(dev->ctrl.admin_q);
1604 		blk_cleanup_queue(dev->ctrl.admin_q);
1605 		blk_mq_free_tag_set(&dev->admin_tagset);
1606 	}
1607 }
1608 
1609 static int nvme_alloc_admin_tags(struct nvme_dev *dev)
1610 {
1611 	if (!dev->ctrl.admin_q) {
1612 		dev->admin_tagset.ops = &nvme_mq_admin_ops;
1613 		dev->admin_tagset.nr_hw_queues = 1;
1614 
1615 		dev->admin_tagset.queue_depth = NVME_AQ_MQ_TAG_DEPTH;
1616 		dev->admin_tagset.timeout = ADMIN_TIMEOUT;
1617 		dev->admin_tagset.numa_node = dev_to_node(dev->dev);
1618 		dev->admin_tagset.cmd_size = sizeof(struct nvme_iod);
1619 		dev->admin_tagset.flags = BLK_MQ_F_NO_SCHED;
1620 		dev->admin_tagset.driver_data = dev;
1621 
1622 		if (blk_mq_alloc_tag_set(&dev->admin_tagset))
1623 			return -ENOMEM;
1624 		dev->ctrl.admin_tagset = &dev->admin_tagset;
1625 
1626 		dev->ctrl.admin_q = blk_mq_init_queue(&dev->admin_tagset);
1627 		if (IS_ERR(dev->ctrl.admin_q)) {
1628 			blk_mq_free_tag_set(&dev->admin_tagset);
1629 			return -ENOMEM;
1630 		}
1631 		if (!blk_get_queue(dev->ctrl.admin_q)) {
1632 			nvme_dev_remove_admin(dev);
1633 			dev->ctrl.admin_q = NULL;
1634 			return -ENODEV;
1635 		}
1636 	} else
1637 		blk_mq_unquiesce_queue(dev->ctrl.admin_q);
1638 
1639 	return 0;
1640 }
1641 
1642 static unsigned long db_bar_size(struct nvme_dev *dev, unsigned nr_io_queues)
1643 {
1644 	return NVME_REG_DBS + ((nr_io_queues + 1) * 8 * dev->db_stride);
1645 }
1646 
1647 static int nvme_remap_bar(struct nvme_dev *dev, unsigned long size)
1648 {
1649 	struct pci_dev *pdev = to_pci_dev(dev->dev);
1650 
1651 	if (size <= dev->bar_mapped_size)
1652 		return 0;
1653 	if (size > pci_resource_len(pdev, 0))
1654 		return -ENOMEM;
1655 	if (dev->bar)
1656 		iounmap(dev->bar);
1657 	dev->bar = ioremap(pci_resource_start(pdev, 0), size);
1658 	if (!dev->bar) {
1659 		dev->bar_mapped_size = 0;
1660 		return -ENOMEM;
1661 	}
1662 	dev->bar_mapped_size = size;
1663 	dev->dbs = dev->bar + NVME_REG_DBS;
1664 
1665 	return 0;
1666 }
1667 
1668 static int nvme_pci_configure_admin_queue(struct nvme_dev *dev)
1669 {
1670 	int result;
1671 	u32 aqa;
1672 	struct nvme_queue *nvmeq;
1673 
1674 	result = nvme_remap_bar(dev, db_bar_size(dev, 0));
1675 	if (result < 0)
1676 		return result;
1677 
1678 	dev->subsystem = readl(dev->bar + NVME_REG_VS) >= NVME_VS(1, 1, 0) ?
1679 				NVME_CAP_NSSRC(dev->ctrl.cap) : 0;
1680 
1681 	if (dev->subsystem &&
1682 	    (readl(dev->bar + NVME_REG_CSTS) & NVME_CSTS_NSSRO))
1683 		writel(NVME_CSTS_NSSRO, dev->bar + NVME_REG_CSTS);
1684 
1685 	result = nvme_disable_ctrl(&dev->ctrl);
1686 	if (result < 0)
1687 		return result;
1688 
1689 	result = nvme_alloc_queue(dev, 0, NVME_AQ_DEPTH);
1690 	if (result)
1691 		return result;
1692 
1693 	nvmeq = &dev->queues[0];
1694 	aqa = nvmeq->q_depth - 1;
1695 	aqa |= aqa << 16;
1696 
1697 	writel(aqa, dev->bar + NVME_REG_AQA);
1698 	lo_hi_writeq(nvmeq->sq_dma_addr, dev->bar + NVME_REG_ASQ);
1699 	lo_hi_writeq(nvmeq->cq_dma_addr, dev->bar + NVME_REG_ACQ);
1700 
1701 	result = nvme_enable_ctrl(&dev->ctrl);
1702 	if (result)
1703 		return result;
1704 
1705 	nvmeq->cq_vector = 0;
1706 	nvme_init_queue(nvmeq, 0);
1707 	result = queue_request_irq(nvmeq);
1708 	if (result) {
1709 		dev->online_queues--;
1710 		return result;
1711 	}
1712 
1713 	set_bit(NVMEQ_ENABLED, &nvmeq->flags);
1714 	return result;
1715 }
1716 
1717 static int nvme_create_io_queues(struct nvme_dev *dev)
1718 {
1719 	unsigned i, max, rw_queues;
1720 	int ret = 0;
1721 
1722 	for (i = dev->ctrl.queue_count; i <= dev->max_qid; i++) {
1723 		if (nvme_alloc_queue(dev, i, dev->q_depth)) {
1724 			ret = -ENOMEM;
1725 			break;
1726 		}
1727 	}
1728 
1729 	max = min(dev->max_qid, dev->ctrl.queue_count - 1);
1730 	if (max != 1 && dev->io_queues[HCTX_TYPE_POLL]) {
1731 		rw_queues = dev->io_queues[HCTX_TYPE_DEFAULT] +
1732 				dev->io_queues[HCTX_TYPE_READ];
1733 	} else {
1734 		rw_queues = max;
1735 	}
1736 
1737 	for (i = dev->online_queues; i <= max; i++) {
1738 		bool polled = i > rw_queues;
1739 
1740 		ret = nvme_create_queue(&dev->queues[i], i, polled);
1741 		if (ret)
1742 			break;
1743 	}
1744 
1745 	/*
1746 	 * Ignore failing Create SQ/CQ commands, we can continue with less
1747 	 * than the desired amount of queues, and even a controller without
1748 	 * I/O queues can still be used to issue admin commands.  This might
1749 	 * be useful to upgrade a buggy firmware for example.
1750 	 */
1751 	return ret >= 0 ? 0 : ret;
1752 }
1753 
1754 static ssize_t nvme_cmb_show(struct device *dev,
1755 			     struct device_attribute *attr,
1756 			     char *buf)
1757 {
1758 	struct nvme_dev *ndev = to_nvme_dev(dev_get_drvdata(dev));
1759 
1760 	return scnprintf(buf, PAGE_SIZE, "cmbloc : x%08x\ncmbsz  : x%08x\n",
1761 		       ndev->cmbloc, ndev->cmbsz);
1762 }
1763 static DEVICE_ATTR(cmb, S_IRUGO, nvme_cmb_show, NULL);
1764 
1765 static u64 nvme_cmb_size_unit(struct nvme_dev *dev)
1766 {
1767 	u8 szu = (dev->cmbsz >> NVME_CMBSZ_SZU_SHIFT) & NVME_CMBSZ_SZU_MASK;
1768 
1769 	return 1ULL << (12 + 4 * szu);
1770 }
1771 
1772 static u32 nvme_cmb_size(struct nvme_dev *dev)
1773 {
1774 	return (dev->cmbsz >> NVME_CMBSZ_SZ_SHIFT) & NVME_CMBSZ_SZ_MASK;
1775 }
1776 
1777 static void nvme_map_cmb(struct nvme_dev *dev)
1778 {
1779 	u64 size, offset;
1780 	resource_size_t bar_size;
1781 	struct pci_dev *pdev = to_pci_dev(dev->dev);
1782 	int bar;
1783 
1784 	if (dev->cmb_size)
1785 		return;
1786 
1787 	dev->cmbsz = readl(dev->bar + NVME_REG_CMBSZ);
1788 	if (!dev->cmbsz)
1789 		return;
1790 	dev->cmbloc = readl(dev->bar + NVME_REG_CMBLOC);
1791 
1792 	size = nvme_cmb_size_unit(dev) * nvme_cmb_size(dev);
1793 	offset = nvme_cmb_size_unit(dev) * NVME_CMB_OFST(dev->cmbloc);
1794 	bar = NVME_CMB_BIR(dev->cmbloc);
1795 	bar_size = pci_resource_len(pdev, bar);
1796 
1797 	if (offset > bar_size)
1798 		return;
1799 
1800 	/*
1801 	 * Controllers may support a CMB size larger than their BAR,
1802 	 * for example, due to being behind a bridge. Reduce the CMB to
1803 	 * the reported size of the BAR
1804 	 */
1805 	if (size > bar_size - offset)
1806 		size = bar_size - offset;
1807 
1808 	if (pci_p2pdma_add_resource(pdev, bar, size, offset)) {
1809 		dev_warn(dev->ctrl.device,
1810 			 "failed to register the CMB\n");
1811 		return;
1812 	}
1813 
1814 	dev->cmb_size = size;
1815 	dev->cmb_use_sqes = use_cmb_sqes && (dev->cmbsz & NVME_CMBSZ_SQS);
1816 
1817 	if ((dev->cmbsz & (NVME_CMBSZ_WDS | NVME_CMBSZ_RDS)) ==
1818 			(NVME_CMBSZ_WDS | NVME_CMBSZ_RDS))
1819 		pci_p2pmem_publish(pdev, true);
1820 
1821 	if (sysfs_add_file_to_group(&dev->ctrl.device->kobj,
1822 				    &dev_attr_cmb.attr, NULL))
1823 		dev_warn(dev->ctrl.device,
1824 			 "failed to add sysfs attribute for CMB\n");
1825 }
1826 
1827 static inline void nvme_release_cmb(struct nvme_dev *dev)
1828 {
1829 	if (dev->cmb_size) {
1830 		sysfs_remove_file_from_group(&dev->ctrl.device->kobj,
1831 					     &dev_attr_cmb.attr, NULL);
1832 		dev->cmb_size = 0;
1833 	}
1834 }
1835 
1836 static int nvme_set_host_mem(struct nvme_dev *dev, u32 bits)
1837 {
1838 	u64 dma_addr = dev->host_mem_descs_dma;
1839 	struct nvme_command c;
1840 	int ret;
1841 
1842 	memset(&c, 0, sizeof(c));
1843 	c.features.opcode	= nvme_admin_set_features;
1844 	c.features.fid		= cpu_to_le32(NVME_FEAT_HOST_MEM_BUF);
1845 	c.features.dword11	= cpu_to_le32(bits);
1846 	c.features.dword12	= cpu_to_le32(dev->host_mem_size >>
1847 					      ilog2(dev->ctrl.page_size));
1848 	c.features.dword13	= cpu_to_le32(lower_32_bits(dma_addr));
1849 	c.features.dword14	= cpu_to_le32(upper_32_bits(dma_addr));
1850 	c.features.dword15	= cpu_to_le32(dev->nr_host_mem_descs);
1851 
1852 	ret = nvme_submit_sync_cmd(dev->ctrl.admin_q, &c, NULL, 0);
1853 	if (ret) {
1854 		dev_warn(dev->ctrl.device,
1855 			 "failed to set host mem (err %d, flags %#x).\n",
1856 			 ret, bits);
1857 	}
1858 	return ret;
1859 }
1860 
1861 static void nvme_free_host_mem(struct nvme_dev *dev)
1862 {
1863 	int i;
1864 
1865 	for (i = 0; i < dev->nr_host_mem_descs; i++) {
1866 		struct nvme_host_mem_buf_desc *desc = &dev->host_mem_descs[i];
1867 		size_t size = le32_to_cpu(desc->size) * dev->ctrl.page_size;
1868 
1869 		dma_free_attrs(dev->dev, size, dev->host_mem_desc_bufs[i],
1870 			       le64_to_cpu(desc->addr),
1871 			       DMA_ATTR_NO_KERNEL_MAPPING | DMA_ATTR_NO_WARN);
1872 	}
1873 
1874 	kfree(dev->host_mem_desc_bufs);
1875 	dev->host_mem_desc_bufs = NULL;
1876 	dma_free_coherent(dev->dev,
1877 			dev->nr_host_mem_descs * sizeof(*dev->host_mem_descs),
1878 			dev->host_mem_descs, dev->host_mem_descs_dma);
1879 	dev->host_mem_descs = NULL;
1880 	dev->nr_host_mem_descs = 0;
1881 }
1882 
1883 static int __nvme_alloc_host_mem(struct nvme_dev *dev, u64 preferred,
1884 		u32 chunk_size)
1885 {
1886 	struct nvme_host_mem_buf_desc *descs;
1887 	u32 max_entries, len;
1888 	dma_addr_t descs_dma;
1889 	int i = 0;
1890 	void **bufs;
1891 	u64 size, tmp;
1892 
1893 	tmp = (preferred + chunk_size - 1);
1894 	do_div(tmp, chunk_size);
1895 	max_entries = tmp;
1896 
1897 	if (dev->ctrl.hmmaxd && dev->ctrl.hmmaxd < max_entries)
1898 		max_entries = dev->ctrl.hmmaxd;
1899 
1900 	descs = dma_alloc_coherent(dev->dev, max_entries * sizeof(*descs),
1901 				   &descs_dma, GFP_KERNEL);
1902 	if (!descs)
1903 		goto out;
1904 
1905 	bufs = kcalloc(max_entries, sizeof(*bufs), GFP_KERNEL);
1906 	if (!bufs)
1907 		goto out_free_descs;
1908 
1909 	for (size = 0; size < preferred && i < max_entries; size += len) {
1910 		dma_addr_t dma_addr;
1911 
1912 		len = min_t(u64, chunk_size, preferred - size);
1913 		bufs[i] = dma_alloc_attrs(dev->dev, len, &dma_addr, GFP_KERNEL,
1914 				DMA_ATTR_NO_KERNEL_MAPPING | DMA_ATTR_NO_WARN);
1915 		if (!bufs[i])
1916 			break;
1917 
1918 		descs[i].addr = cpu_to_le64(dma_addr);
1919 		descs[i].size = cpu_to_le32(len / dev->ctrl.page_size);
1920 		i++;
1921 	}
1922 
1923 	if (!size)
1924 		goto out_free_bufs;
1925 
1926 	dev->nr_host_mem_descs = i;
1927 	dev->host_mem_size = size;
1928 	dev->host_mem_descs = descs;
1929 	dev->host_mem_descs_dma = descs_dma;
1930 	dev->host_mem_desc_bufs = bufs;
1931 	return 0;
1932 
1933 out_free_bufs:
1934 	while (--i >= 0) {
1935 		size_t size = le32_to_cpu(descs[i].size) * dev->ctrl.page_size;
1936 
1937 		dma_free_attrs(dev->dev, size, bufs[i],
1938 			       le64_to_cpu(descs[i].addr),
1939 			       DMA_ATTR_NO_KERNEL_MAPPING | DMA_ATTR_NO_WARN);
1940 	}
1941 
1942 	kfree(bufs);
1943 out_free_descs:
1944 	dma_free_coherent(dev->dev, max_entries * sizeof(*descs), descs,
1945 			descs_dma);
1946 out:
1947 	dev->host_mem_descs = NULL;
1948 	return -ENOMEM;
1949 }
1950 
1951 static int nvme_alloc_host_mem(struct nvme_dev *dev, u64 min, u64 preferred)
1952 {
1953 	u32 chunk_size;
1954 
1955 	/* start big and work our way down */
1956 	for (chunk_size = min_t(u64, preferred, PAGE_SIZE * MAX_ORDER_NR_PAGES);
1957 	     chunk_size >= max_t(u32, dev->ctrl.hmminds * 4096, PAGE_SIZE * 2);
1958 	     chunk_size /= 2) {
1959 		if (!__nvme_alloc_host_mem(dev, preferred, chunk_size)) {
1960 			if (!min || dev->host_mem_size >= min)
1961 				return 0;
1962 			nvme_free_host_mem(dev);
1963 		}
1964 	}
1965 
1966 	return -ENOMEM;
1967 }
1968 
1969 static int nvme_setup_host_mem(struct nvme_dev *dev)
1970 {
1971 	u64 max = (u64)max_host_mem_size_mb * SZ_1M;
1972 	u64 preferred = (u64)dev->ctrl.hmpre * 4096;
1973 	u64 min = (u64)dev->ctrl.hmmin * 4096;
1974 	u32 enable_bits = NVME_HOST_MEM_ENABLE;
1975 	int ret;
1976 
1977 	preferred = min(preferred, max);
1978 	if (min > max) {
1979 		dev_warn(dev->ctrl.device,
1980 			"min host memory (%lld MiB) above limit (%d MiB).\n",
1981 			min >> ilog2(SZ_1M), max_host_mem_size_mb);
1982 		nvme_free_host_mem(dev);
1983 		return 0;
1984 	}
1985 
1986 	/*
1987 	 * If we already have a buffer allocated check if we can reuse it.
1988 	 */
1989 	if (dev->host_mem_descs) {
1990 		if (dev->host_mem_size >= min)
1991 			enable_bits |= NVME_HOST_MEM_RETURN;
1992 		else
1993 			nvme_free_host_mem(dev);
1994 	}
1995 
1996 	if (!dev->host_mem_descs) {
1997 		if (nvme_alloc_host_mem(dev, min, preferred)) {
1998 			dev_warn(dev->ctrl.device,
1999 				"failed to allocate host memory buffer.\n");
2000 			return 0; /* controller must work without HMB */
2001 		}
2002 
2003 		dev_info(dev->ctrl.device,
2004 			"allocated %lld MiB host memory buffer.\n",
2005 			dev->host_mem_size >> ilog2(SZ_1M));
2006 	}
2007 
2008 	ret = nvme_set_host_mem(dev, enable_bits);
2009 	if (ret)
2010 		nvme_free_host_mem(dev);
2011 	return ret;
2012 }
2013 
2014 /*
2015  * nirqs is the number of interrupts available for write and read
2016  * queues. The core already reserved an interrupt for the admin queue.
2017  */
2018 static void nvme_calc_irq_sets(struct irq_affinity *affd, unsigned int nrirqs)
2019 {
2020 	struct nvme_dev *dev = affd->priv;
2021 	unsigned int nr_read_queues;
2022 
2023 	/*
2024 	 * If there is no interupt available for queues, ensure that
2025 	 * the default queue is set to 1. The affinity set size is
2026 	 * also set to one, but the irq core ignores it for this case.
2027 	 *
2028 	 * If only one interrupt is available or 'write_queue' == 0, combine
2029 	 * write and read queues.
2030 	 *
2031 	 * If 'write_queues' > 0, ensure it leaves room for at least one read
2032 	 * queue.
2033 	 */
2034 	if (!nrirqs) {
2035 		nrirqs = 1;
2036 		nr_read_queues = 0;
2037 	} else if (nrirqs == 1 || !write_queues) {
2038 		nr_read_queues = 0;
2039 	} else if (write_queues >= nrirqs) {
2040 		nr_read_queues = 1;
2041 	} else {
2042 		nr_read_queues = nrirqs - write_queues;
2043 	}
2044 
2045 	dev->io_queues[HCTX_TYPE_DEFAULT] = nrirqs - nr_read_queues;
2046 	affd->set_size[HCTX_TYPE_DEFAULT] = nrirqs - nr_read_queues;
2047 	dev->io_queues[HCTX_TYPE_READ] = nr_read_queues;
2048 	affd->set_size[HCTX_TYPE_READ] = nr_read_queues;
2049 	affd->nr_sets = nr_read_queues ? 2 : 1;
2050 }
2051 
2052 static int nvme_setup_irqs(struct nvme_dev *dev, unsigned int nr_io_queues)
2053 {
2054 	struct pci_dev *pdev = to_pci_dev(dev->dev);
2055 	struct irq_affinity affd = {
2056 		.pre_vectors	= 1,
2057 		.calc_sets	= nvme_calc_irq_sets,
2058 		.priv		= dev,
2059 	};
2060 	unsigned int irq_queues, this_p_queues;
2061 	unsigned int nr_cpus = num_possible_cpus();
2062 
2063 	/*
2064 	 * Poll queues don't need interrupts, but we need at least one IO
2065 	 * queue left over for non-polled IO.
2066 	 */
2067 	this_p_queues = poll_queues;
2068 	if (this_p_queues >= nr_io_queues) {
2069 		this_p_queues = nr_io_queues - 1;
2070 		irq_queues = 1;
2071 	} else {
2072 		if (nr_cpus < nr_io_queues - this_p_queues)
2073 			irq_queues = nr_cpus + 1;
2074 		else
2075 			irq_queues = nr_io_queues - this_p_queues + 1;
2076 	}
2077 	dev->io_queues[HCTX_TYPE_POLL] = this_p_queues;
2078 
2079 	/* Initialize for the single interrupt case */
2080 	dev->io_queues[HCTX_TYPE_DEFAULT] = 1;
2081 	dev->io_queues[HCTX_TYPE_READ] = 0;
2082 
2083 	/*
2084 	 * Some Apple controllers require all queues to use the
2085 	 * first vector.
2086 	 */
2087 	if (dev->ctrl.quirks & NVME_QUIRK_SINGLE_VECTOR)
2088 		irq_queues = 1;
2089 
2090 	return pci_alloc_irq_vectors_affinity(pdev, 1, irq_queues,
2091 			      PCI_IRQ_ALL_TYPES | PCI_IRQ_AFFINITY, &affd);
2092 }
2093 
2094 static void nvme_disable_io_queues(struct nvme_dev *dev)
2095 {
2096 	if (__nvme_disable_io_queues(dev, nvme_admin_delete_sq))
2097 		__nvme_disable_io_queues(dev, nvme_admin_delete_cq);
2098 }
2099 
2100 static int nvme_setup_io_queues(struct nvme_dev *dev)
2101 {
2102 	struct nvme_queue *adminq = &dev->queues[0];
2103 	struct pci_dev *pdev = to_pci_dev(dev->dev);
2104 	int result, nr_io_queues;
2105 	unsigned long size;
2106 
2107 	nr_io_queues = max_io_queues();
2108 
2109 	/*
2110 	 * If tags are shared with admin queue (Apple bug), then
2111 	 * make sure we only use one IO queue.
2112 	 */
2113 	if (dev->ctrl.quirks & NVME_QUIRK_SHARED_TAGS)
2114 		nr_io_queues = 1;
2115 
2116 	result = nvme_set_queue_count(&dev->ctrl, &nr_io_queues);
2117 	if (result < 0)
2118 		return result;
2119 
2120 	if (nr_io_queues == 0)
2121 		return 0;
2122 
2123 	clear_bit(NVMEQ_ENABLED, &adminq->flags);
2124 
2125 	if (dev->cmb_use_sqes) {
2126 		result = nvme_cmb_qdepth(dev, nr_io_queues,
2127 				sizeof(struct nvme_command));
2128 		if (result > 0)
2129 			dev->q_depth = result;
2130 		else
2131 			dev->cmb_use_sqes = false;
2132 	}
2133 
2134 	do {
2135 		size = db_bar_size(dev, nr_io_queues);
2136 		result = nvme_remap_bar(dev, size);
2137 		if (!result)
2138 			break;
2139 		if (!--nr_io_queues)
2140 			return -ENOMEM;
2141 	} while (1);
2142 	adminq->q_db = dev->dbs;
2143 
2144  retry:
2145 	/* Deregister the admin queue's interrupt */
2146 	pci_free_irq(pdev, 0, adminq);
2147 
2148 	/*
2149 	 * If we enable msix early due to not intx, disable it again before
2150 	 * setting up the full range we need.
2151 	 */
2152 	pci_free_irq_vectors(pdev);
2153 
2154 	result = nvme_setup_irqs(dev, nr_io_queues);
2155 	if (result <= 0)
2156 		return -EIO;
2157 
2158 	dev->num_vecs = result;
2159 	result = max(result - 1, 1);
2160 	dev->max_qid = result + dev->io_queues[HCTX_TYPE_POLL];
2161 
2162 	/*
2163 	 * Should investigate if there's a performance win from allocating
2164 	 * more queues than interrupt vectors; it might allow the submission
2165 	 * path to scale better, even if the receive path is limited by the
2166 	 * number of interrupts.
2167 	 */
2168 	result = queue_request_irq(adminq);
2169 	if (result)
2170 		return result;
2171 	set_bit(NVMEQ_ENABLED, &adminq->flags);
2172 
2173 	result = nvme_create_io_queues(dev);
2174 	if (result || dev->online_queues < 2)
2175 		return result;
2176 
2177 	if (dev->online_queues - 1 < dev->max_qid) {
2178 		nr_io_queues = dev->online_queues - 1;
2179 		nvme_disable_io_queues(dev);
2180 		nvme_suspend_io_queues(dev);
2181 		goto retry;
2182 	}
2183 	dev_info(dev->ctrl.device, "%d/%d/%d default/read/poll queues\n",
2184 					dev->io_queues[HCTX_TYPE_DEFAULT],
2185 					dev->io_queues[HCTX_TYPE_READ],
2186 					dev->io_queues[HCTX_TYPE_POLL]);
2187 	return 0;
2188 }
2189 
2190 static void nvme_del_queue_end(struct request *req, blk_status_t error)
2191 {
2192 	struct nvme_queue *nvmeq = req->end_io_data;
2193 
2194 	blk_mq_free_request(req);
2195 	complete(&nvmeq->delete_done);
2196 }
2197 
2198 static void nvme_del_cq_end(struct request *req, blk_status_t error)
2199 {
2200 	struct nvme_queue *nvmeq = req->end_io_data;
2201 
2202 	if (error)
2203 		set_bit(NVMEQ_DELETE_ERROR, &nvmeq->flags);
2204 
2205 	nvme_del_queue_end(req, error);
2206 }
2207 
2208 static int nvme_delete_queue(struct nvme_queue *nvmeq, u8 opcode)
2209 {
2210 	struct request_queue *q = nvmeq->dev->ctrl.admin_q;
2211 	struct request *req;
2212 	struct nvme_command cmd;
2213 
2214 	memset(&cmd, 0, sizeof(cmd));
2215 	cmd.delete_queue.opcode = opcode;
2216 	cmd.delete_queue.qid = cpu_to_le16(nvmeq->qid);
2217 
2218 	req = nvme_alloc_request(q, &cmd, BLK_MQ_REQ_NOWAIT, NVME_QID_ANY);
2219 	if (IS_ERR(req))
2220 		return PTR_ERR(req);
2221 
2222 	req->timeout = ADMIN_TIMEOUT;
2223 	req->end_io_data = nvmeq;
2224 
2225 	init_completion(&nvmeq->delete_done);
2226 	blk_execute_rq_nowait(q, NULL, req, false,
2227 			opcode == nvme_admin_delete_cq ?
2228 				nvme_del_cq_end : nvme_del_queue_end);
2229 	return 0;
2230 }
2231 
2232 static bool __nvme_disable_io_queues(struct nvme_dev *dev, u8 opcode)
2233 {
2234 	int nr_queues = dev->online_queues - 1, sent = 0;
2235 	unsigned long timeout;
2236 
2237  retry:
2238 	timeout = ADMIN_TIMEOUT;
2239 	while (nr_queues > 0) {
2240 		if (nvme_delete_queue(&dev->queues[nr_queues], opcode))
2241 			break;
2242 		nr_queues--;
2243 		sent++;
2244 	}
2245 	while (sent) {
2246 		struct nvme_queue *nvmeq = &dev->queues[nr_queues + sent];
2247 
2248 		timeout = wait_for_completion_io_timeout(&nvmeq->delete_done,
2249 				timeout);
2250 		if (timeout == 0)
2251 			return false;
2252 
2253 		/* handle any remaining CQEs */
2254 		if (opcode == nvme_admin_delete_cq &&
2255 		    !test_bit(NVMEQ_DELETE_ERROR, &nvmeq->flags))
2256 			nvme_poll_irqdisable(nvmeq, -1);
2257 
2258 		sent--;
2259 		if (nr_queues)
2260 			goto retry;
2261 	}
2262 	return true;
2263 }
2264 
2265 static void nvme_dev_add(struct nvme_dev *dev)
2266 {
2267 	int ret;
2268 
2269 	if (!dev->ctrl.tagset) {
2270 		dev->tagset.ops = &nvme_mq_ops;
2271 		dev->tagset.nr_hw_queues = dev->online_queues - 1;
2272 		dev->tagset.nr_maps = 2; /* default + read */
2273 		if (dev->io_queues[HCTX_TYPE_POLL])
2274 			dev->tagset.nr_maps++;
2275 		dev->tagset.timeout = NVME_IO_TIMEOUT;
2276 		dev->tagset.numa_node = dev_to_node(dev->dev);
2277 		dev->tagset.queue_depth =
2278 				min_t(int, dev->q_depth, BLK_MQ_MAX_DEPTH) - 1;
2279 		dev->tagset.cmd_size = sizeof(struct nvme_iod);
2280 		dev->tagset.flags = BLK_MQ_F_SHOULD_MERGE;
2281 		dev->tagset.driver_data = dev;
2282 
2283 		/*
2284 		 * Some Apple controllers requires tags to be unique
2285 		 * across admin and IO queue, so reserve the first 32
2286 		 * tags of the IO queue.
2287 		 */
2288 		if (dev->ctrl.quirks & NVME_QUIRK_SHARED_TAGS)
2289 			dev->tagset.reserved_tags = NVME_AQ_DEPTH;
2290 
2291 		ret = blk_mq_alloc_tag_set(&dev->tagset);
2292 		if (ret) {
2293 			dev_warn(dev->ctrl.device,
2294 				"IO queues tagset allocation failed %d\n", ret);
2295 			return;
2296 		}
2297 		dev->ctrl.tagset = &dev->tagset;
2298 	} else {
2299 		blk_mq_update_nr_hw_queues(&dev->tagset, dev->online_queues - 1);
2300 
2301 		/* Free previously allocated queues that are no longer usable */
2302 		nvme_free_queues(dev, dev->online_queues);
2303 	}
2304 
2305 	nvme_dbbuf_set(dev);
2306 }
2307 
2308 static int nvme_pci_enable(struct nvme_dev *dev)
2309 {
2310 	int result = -ENOMEM;
2311 	struct pci_dev *pdev = to_pci_dev(dev->dev);
2312 
2313 	if (pci_enable_device_mem(pdev))
2314 		return result;
2315 
2316 	pci_set_master(pdev);
2317 
2318 	if (dma_set_mask_and_coherent(dev->dev, DMA_BIT_MASK(64)))
2319 		goto disable;
2320 
2321 	if (readl(dev->bar + NVME_REG_CSTS) == -1) {
2322 		result = -ENODEV;
2323 		goto disable;
2324 	}
2325 
2326 	/*
2327 	 * Some devices and/or platforms don't advertise or work with INTx
2328 	 * interrupts. Pre-enable a single MSIX or MSI vec for setup. We'll
2329 	 * adjust this later.
2330 	 */
2331 	result = pci_alloc_irq_vectors(pdev, 1, 1, PCI_IRQ_ALL_TYPES);
2332 	if (result < 0)
2333 		return result;
2334 
2335 	dev->ctrl.cap = lo_hi_readq(dev->bar + NVME_REG_CAP);
2336 
2337 	dev->q_depth = min_t(int, NVME_CAP_MQES(dev->ctrl.cap) + 1,
2338 				io_queue_depth);
2339 	dev->ctrl.sqsize = dev->q_depth - 1; /* 0's based queue depth */
2340 	dev->db_stride = 1 << NVME_CAP_STRIDE(dev->ctrl.cap);
2341 	dev->dbs = dev->bar + 4096;
2342 
2343 	/*
2344 	 * Some Apple controllers require a non-standard SQE size.
2345 	 * Interestingly they also seem to ignore the CC:IOSQES register
2346 	 * so we don't bother updating it here.
2347 	 */
2348 	if (dev->ctrl.quirks & NVME_QUIRK_128_BYTES_SQES)
2349 		dev->io_sqes = 7;
2350 	else
2351 		dev->io_sqes = NVME_NVM_IOSQES;
2352 
2353 	/*
2354 	 * Temporary fix for the Apple controller found in the MacBook8,1 and
2355 	 * some MacBook7,1 to avoid controller resets and data loss.
2356 	 */
2357 	if (pdev->vendor == PCI_VENDOR_ID_APPLE && pdev->device == 0x2001) {
2358 		dev->q_depth = 2;
2359 		dev_warn(dev->ctrl.device, "detected Apple NVMe controller, "
2360 			"set queue depth=%u to work around controller resets\n",
2361 			dev->q_depth);
2362 	} else if (pdev->vendor == PCI_VENDOR_ID_SAMSUNG &&
2363 		   (pdev->device == 0xa821 || pdev->device == 0xa822) &&
2364 		   NVME_CAP_MQES(dev->ctrl.cap) == 0) {
2365 		dev->q_depth = 64;
2366 		dev_err(dev->ctrl.device, "detected PM1725 NVMe controller, "
2367                         "set queue depth=%u\n", dev->q_depth);
2368 	}
2369 
2370 	/*
2371 	 * Controllers with the shared tags quirk need the IO queue to be
2372 	 * big enough so that we get 32 tags for the admin queue
2373 	 */
2374 	if ((dev->ctrl.quirks & NVME_QUIRK_SHARED_TAGS) &&
2375 	    (dev->q_depth < (NVME_AQ_DEPTH + 2))) {
2376 		dev->q_depth = NVME_AQ_DEPTH + 2;
2377 		dev_warn(dev->ctrl.device, "IO queue depth clamped to %d\n",
2378 			 dev->q_depth);
2379 	}
2380 
2381 
2382 	nvme_map_cmb(dev);
2383 
2384 	pci_enable_pcie_error_reporting(pdev);
2385 	pci_save_state(pdev);
2386 	return 0;
2387 
2388  disable:
2389 	pci_disable_device(pdev);
2390 	return result;
2391 }
2392 
2393 static void nvme_dev_unmap(struct nvme_dev *dev)
2394 {
2395 	if (dev->bar)
2396 		iounmap(dev->bar);
2397 	pci_release_mem_regions(to_pci_dev(dev->dev));
2398 }
2399 
2400 static void nvme_pci_disable(struct nvme_dev *dev)
2401 {
2402 	struct pci_dev *pdev = to_pci_dev(dev->dev);
2403 
2404 	pci_free_irq_vectors(pdev);
2405 
2406 	if (pci_is_enabled(pdev)) {
2407 		pci_disable_pcie_error_reporting(pdev);
2408 		pci_disable_device(pdev);
2409 	}
2410 }
2411 
2412 static void nvme_dev_disable(struct nvme_dev *dev, bool shutdown)
2413 {
2414 	bool dead = true, freeze = false;
2415 	struct pci_dev *pdev = to_pci_dev(dev->dev);
2416 
2417 	mutex_lock(&dev->shutdown_lock);
2418 	if (pci_is_enabled(pdev)) {
2419 		u32 csts = readl(dev->bar + NVME_REG_CSTS);
2420 
2421 		if (dev->ctrl.state == NVME_CTRL_LIVE ||
2422 		    dev->ctrl.state == NVME_CTRL_RESETTING) {
2423 			freeze = true;
2424 			nvme_start_freeze(&dev->ctrl);
2425 		}
2426 		dead = !!((csts & NVME_CSTS_CFS) || !(csts & NVME_CSTS_RDY) ||
2427 			pdev->error_state  != pci_channel_io_normal);
2428 	}
2429 
2430 	/*
2431 	 * Give the controller a chance to complete all entered requests if
2432 	 * doing a safe shutdown.
2433 	 */
2434 	if (!dead && shutdown && freeze)
2435 		nvme_wait_freeze_timeout(&dev->ctrl, NVME_IO_TIMEOUT);
2436 
2437 	nvme_stop_queues(&dev->ctrl);
2438 
2439 	if (!dead && dev->ctrl.queue_count > 0) {
2440 		nvme_disable_io_queues(dev);
2441 		nvme_disable_admin_queue(dev, shutdown);
2442 	}
2443 	nvme_suspend_io_queues(dev);
2444 	nvme_suspend_queue(&dev->queues[0]);
2445 	nvme_pci_disable(dev);
2446 
2447 	blk_mq_tagset_busy_iter(&dev->tagset, nvme_cancel_request, &dev->ctrl);
2448 	blk_mq_tagset_busy_iter(&dev->admin_tagset, nvme_cancel_request, &dev->ctrl);
2449 	blk_mq_tagset_wait_completed_request(&dev->tagset);
2450 	blk_mq_tagset_wait_completed_request(&dev->admin_tagset);
2451 
2452 	/*
2453 	 * The driver will not be starting up queues again if shutting down so
2454 	 * must flush all entered requests to their failed completion to avoid
2455 	 * deadlocking blk-mq hot-cpu notifier.
2456 	 */
2457 	if (shutdown) {
2458 		nvme_start_queues(&dev->ctrl);
2459 		if (dev->ctrl.admin_q && !blk_queue_dying(dev->ctrl.admin_q))
2460 			blk_mq_unquiesce_queue(dev->ctrl.admin_q);
2461 	}
2462 	mutex_unlock(&dev->shutdown_lock);
2463 }
2464 
2465 static int nvme_disable_prepare_reset(struct nvme_dev *dev, bool shutdown)
2466 {
2467 	if (!nvme_wait_reset(&dev->ctrl))
2468 		return -EBUSY;
2469 	nvme_dev_disable(dev, shutdown);
2470 	return 0;
2471 }
2472 
2473 static int nvme_setup_prp_pools(struct nvme_dev *dev)
2474 {
2475 	dev->prp_page_pool = dma_pool_create("prp list page", dev->dev,
2476 						PAGE_SIZE, PAGE_SIZE, 0);
2477 	if (!dev->prp_page_pool)
2478 		return -ENOMEM;
2479 
2480 	/* Optimisation for I/Os between 4k and 128k */
2481 	dev->prp_small_pool = dma_pool_create("prp list 256", dev->dev,
2482 						256, 256, 0);
2483 	if (!dev->prp_small_pool) {
2484 		dma_pool_destroy(dev->prp_page_pool);
2485 		return -ENOMEM;
2486 	}
2487 	return 0;
2488 }
2489 
2490 static void nvme_release_prp_pools(struct nvme_dev *dev)
2491 {
2492 	dma_pool_destroy(dev->prp_page_pool);
2493 	dma_pool_destroy(dev->prp_small_pool);
2494 }
2495 
2496 static void nvme_free_tagset(struct nvme_dev *dev)
2497 {
2498 	if (dev->tagset.tags)
2499 		blk_mq_free_tag_set(&dev->tagset);
2500 	dev->ctrl.tagset = NULL;
2501 }
2502 
2503 static void nvme_pci_free_ctrl(struct nvme_ctrl *ctrl)
2504 {
2505 	struct nvme_dev *dev = to_nvme_dev(ctrl);
2506 
2507 	nvme_dbbuf_dma_free(dev);
2508 	put_device(dev->dev);
2509 	nvme_free_tagset(dev);
2510 	if (dev->ctrl.admin_q)
2511 		blk_put_queue(dev->ctrl.admin_q);
2512 	kfree(dev->queues);
2513 	free_opal_dev(dev->ctrl.opal_dev);
2514 	mempool_destroy(dev->iod_mempool);
2515 	kfree(dev);
2516 }
2517 
2518 static void nvme_remove_dead_ctrl(struct nvme_dev *dev)
2519 {
2520 	/*
2521 	 * Set state to deleting now to avoid blocking nvme_wait_reset(), which
2522 	 * may be holding this pci_dev's device lock.
2523 	 */
2524 	nvme_change_ctrl_state(&dev->ctrl, NVME_CTRL_DELETING);
2525 	nvme_get_ctrl(&dev->ctrl);
2526 	nvme_dev_disable(dev, false);
2527 	nvme_kill_queues(&dev->ctrl);
2528 	if (!queue_work(nvme_wq, &dev->remove_work))
2529 		nvme_put_ctrl(&dev->ctrl);
2530 }
2531 
2532 static void nvme_reset_work(struct work_struct *work)
2533 {
2534 	struct nvme_dev *dev =
2535 		container_of(work, struct nvme_dev, ctrl.reset_work);
2536 	bool was_suspend = !!(dev->ctrl.ctrl_config & NVME_CC_SHN_NORMAL);
2537 	int result;
2538 
2539 	if (WARN_ON(dev->ctrl.state != NVME_CTRL_RESETTING)) {
2540 		result = -ENODEV;
2541 		goto out;
2542 	}
2543 
2544 	/*
2545 	 * If we're called to reset a live controller first shut it down before
2546 	 * moving on.
2547 	 */
2548 	if (dev->ctrl.ctrl_config & NVME_CC_ENABLE)
2549 		nvme_dev_disable(dev, false);
2550 	nvme_sync_queues(&dev->ctrl);
2551 
2552 	mutex_lock(&dev->shutdown_lock);
2553 	result = nvme_pci_enable(dev);
2554 	if (result)
2555 		goto out_unlock;
2556 
2557 	result = nvme_pci_configure_admin_queue(dev);
2558 	if (result)
2559 		goto out_unlock;
2560 
2561 	result = nvme_alloc_admin_tags(dev);
2562 	if (result)
2563 		goto out_unlock;
2564 
2565 	/*
2566 	 * Limit the max command size to prevent iod->sg allocations going
2567 	 * over a single page.
2568 	 */
2569 	dev->ctrl.max_hw_sectors = min_t(u32,
2570 		NVME_MAX_KB_SZ << 1, dma_max_mapping_size(dev->dev) >> 9);
2571 	dev->ctrl.max_segments = NVME_MAX_SEGS;
2572 
2573 	/*
2574 	 * Don't limit the IOMMU merged segment size.
2575 	 */
2576 	dma_set_max_seg_size(dev->dev, 0xffffffff);
2577 
2578 	mutex_unlock(&dev->shutdown_lock);
2579 
2580 	/*
2581 	 * Introduce CONNECTING state from nvme-fc/rdma transports to mark the
2582 	 * initializing procedure here.
2583 	 */
2584 	if (!nvme_change_ctrl_state(&dev->ctrl, NVME_CTRL_CONNECTING)) {
2585 		dev_warn(dev->ctrl.device,
2586 			"failed to mark controller CONNECTING\n");
2587 		result = -EBUSY;
2588 		goto out;
2589 	}
2590 
2591 	result = nvme_init_identify(&dev->ctrl);
2592 	if (result)
2593 		goto out;
2594 
2595 	if (dev->ctrl.oacs & NVME_CTRL_OACS_SEC_SUPP) {
2596 		if (!dev->ctrl.opal_dev)
2597 			dev->ctrl.opal_dev =
2598 				init_opal_dev(&dev->ctrl, &nvme_sec_submit);
2599 		else if (was_suspend)
2600 			opal_unlock_from_suspend(dev->ctrl.opal_dev);
2601 	} else {
2602 		free_opal_dev(dev->ctrl.opal_dev);
2603 		dev->ctrl.opal_dev = NULL;
2604 	}
2605 
2606 	if (dev->ctrl.oacs & NVME_CTRL_OACS_DBBUF_SUPP) {
2607 		result = nvme_dbbuf_dma_alloc(dev);
2608 		if (result)
2609 			dev_warn(dev->dev,
2610 				 "unable to allocate dma for dbbuf\n");
2611 	}
2612 
2613 	if (dev->ctrl.hmpre) {
2614 		result = nvme_setup_host_mem(dev);
2615 		if (result < 0)
2616 			goto out;
2617 	}
2618 
2619 	result = nvme_setup_io_queues(dev);
2620 	if (result)
2621 		goto out;
2622 
2623 	/*
2624 	 * Keep the controller around but remove all namespaces if we don't have
2625 	 * any working I/O queue.
2626 	 */
2627 	if (dev->online_queues < 2) {
2628 		dev_warn(dev->ctrl.device, "IO queues not created\n");
2629 		nvme_kill_queues(&dev->ctrl);
2630 		nvme_remove_namespaces(&dev->ctrl);
2631 		nvme_free_tagset(dev);
2632 	} else {
2633 		nvme_start_queues(&dev->ctrl);
2634 		nvme_wait_freeze(&dev->ctrl);
2635 		nvme_dev_add(dev);
2636 		nvme_unfreeze(&dev->ctrl);
2637 	}
2638 
2639 	/*
2640 	 * If only admin queue live, keep it to do further investigation or
2641 	 * recovery.
2642 	 */
2643 	if (!nvme_change_ctrl_state(&dev->ctrl, NVME_CTRL_LIVE)) {
2644 		dev_warn(dev->ctrl.device,
2645 			"failed to mark controller live state\n");
2646 		result = -ENODEV;
2647 		goto out;
2648 	}
2649 
2650 	nvme_start_ctrl(&dev->ctrl);
2651 	return;
2652 
2653  out_unlock:
2654 	mutex_unlock(&dev->shutdown_lock);
2655  out:
2656 	if (result)
2657 		dev_warn(dev->ctrl.device,
2658 			 "Removing after probe failure status: %d\n", result);
2659 	nvme_remove_dead_ctrl(dev);
2660 }
2661 
2662 static void nvme_remove_dead_ctrl_work(struct work_struct *work)
2663 {
2664 	struct nvme_dev *dev = container_of(work, struct nvme_dev, remove_work);
2665 	struct pci_dev *pdev = to_pci_dev(dev->dev);
2666 
2667 	if (pci_get_drvdata(pdev))
2668 		device_release_driver(&pdev->dev);
2669 	nvme_put_ctrl(&dev->ctrl);
2670 }
2671 
2672 static int nvme_pci_reg_read32(struct nvme_ctrl *ctrl, u32 off, u32 *val)
2673 {
2674 	*val = readl(to_nvme_dev(ctrl)->bar + off);
2675 	return 0;
2676 }
2677 
2678 static int nvme_pci_reg_write32(struct nvme_ctrl *ctrl, u32 off, u32 val)
2679 {
2680 	writel(val, to_nvme_dev(ctrl)->bar + off);
2681 	return 0;
2682 }
2683 
2684 static int nvme_pci_reg_read64(struct nvme_ctrl *ctrl, u32 off, u64 *val)
2685 {
2686 	*val = lo_hi_readq(to_nvme_dev(ctrl)->bar + off);
2687 	return 0;
2688 }
2689 
2690 static int nvme_pci_get_address(struct nvme_ctrl *ctrl, char *buf, int size)
2691 {
2692 	struct pci_dev *pdev = to_pci_dev(to_nvme_dev(ctrl)->dev);
2693 
2694 	return snprintf(buf, size, "%s", dev_name(&pdev->dev));
2695 }
2696 
2697 static const struct nvme_ctrl_ops nvme_pci_ctrl_ops = {
2698 	.name			= "pcie",
2699 	.module			= THIS_MODULE,
2700 	.flags			= NVME_F_METADATA_SUPPORTED |
2701 				  NVME_F_PCI_P2PDMA,
2702 	.reg_read32		= nvme_pci_reg_read32,
2703 	.reg_write32		= nvme_pci_reg_write32,
2704 	.reg_read64		= nvme_pci_reg_read64,
2705 	.free_ctrl		= nvme_pci_free_ctrl,
2706 	.submit_async_event	= nvme_pci_submit_async_event,
2707 	.get_address		= nvme_pci_get_address,
2708 };
2709 
2710 static int nvme_dev_map(struct nvme_dev *dev)
2711 {
2712 	struct pci_dev *pdev = to_pci_dev(dev->dev);
2713 
2714 	if (pci_request_mem_regions(pdev, "nvme"))
2715 		return -ENODEV;
2716 
2717 	if (nvme_remap_bar(dev, NVME_REG_DBS + 4096))
2718 		goto release;
2719 
2720 	return 0;
2721   release:
2722 	pci_release_mem_regions(pdev);
2723 	return -ENODEV;
2724 }
2725 
2726 static unsigned long check_vendor_combination_bug(struct pci_dev *pdev)
2727 {
2728 	if (pdev->vendor == 0x144d && pdev->device == 0xa802) {
2729 		/*
2730 		 * Several Samsung devices seem to drop off the PCIe bus
2731 		 * randomly when APST is on and uses the deepest sleep state.
2732 		 * This has been observed on a Samsung "SM951 NVMe SAMSUNG
2733 		 * 256GB", a "PM951 NVMe SAMSUNG 512GB", and a "Samsung SSD
2734 		 * 950 PRO 256GB", but it seems to be restricted to two Dell
2735 		 * laptops.
2736 		 */
2737 		if (dmi_match(DMI_SYS_VENDOR, "Dell Inc.") &&
2738 		    (dmi_match(DMI_PRODUCT_NAME, "XPS 15 9550") ||
2739 		     dmi_match(DMI_PRODUCT_NAME, "Precision 5510")))
2740 			return NVME_QUIRK_NO_DEEPEST_PS;
2741 	} else if (pdev->vendor == 0x144d && pdev->device == 0xa804) {
2742 		/*
2743 		 * Samsung SSD 960 EVO drops off the PCIe bus after system
2744 		 * suspend on a Ryzen board, ASUS PRIME B350M-A, as well as
2745 		 * within few minutes after bootup on a Coffee Lake board -
2746 		 * ASUS PRIME Z370-A
2747 		 */
2748 		if (dmi_match(DMI_BOARD_VENDOR, "ASUSTeK COMPUTER INC.") &&
2749 		    (dmi_match(DMI_BOARD_NAME, "PRIME B350M-A") ||
2750 		     dmi_match(DMI_BOARD_NAME, "PRIME Z370-A")))
2751 			return NVME_QUIRK_NO_APST;
2752 	}
2753 
2754 	return 0;
2755 }
2756 
2757 static void nvme_async_probe(void *data, async_cookie_t cookie)
2758 {
2759 	struct nvme_dev *dev = data;
2760 
2761 	flush_work(&dev->ctrl.reset_work);
2762 	flush_work(&dev->ctrl.scan_work);
2763 	nvme_put_ctrl(&dev->ctrl);
2764 }
2765 
2766 static int nvme_probe(struct pci_dev *pdev, const struct pci_device_id *id)
2767 {
2768 	int node, result = -ENOMEM;
2769 	struct nvme_dev *dev;
2770 	unsigned long quirks = id->driver_data;
2771 	size_t alloc_size;
2772 
2773 	node = dev_to_node(&pdev->dev);
2774 	if (node == NUMA_NO_NODE)
2775 		set_dev_node(&pdev->dev, first_memory_node);
2776 
2777 	dev = kzalloc_node(sizeof(*dev), GFP_KERNEL, node);
2778 	if (!dev)
2779 		return -ENOMEM;
2780 
2781 	dev->queues = kcalloc_node(max_queue_count(), sizeof(struct nvme_queue),
2782 					GFP_KERNEL, node);
2783 	if (!dev->queues)
2784 		goto free;
2785 
2786 	dev->dev = get_device(&pdev->dev);
2787 	pci_set_drvdata(pdev, dev);
2788 
2789 	result = nvme_dev_map(dev);
2790 	if (result)
2791 		goto put_pci;
2792 
2793 	INIT_WORK(&dev->ctrl.reset_work, nvme_reset_work);
2794 	INIT_WORK(&dev->remove_work, nvme_remove_dead_ctrl_work);
2795 	mutex_init(&dev->shutdown_lock);
2796 
2797 	result = nvme_setup_prp_pools(dev);
2798 	if (result)
2799 		goto unmap;
2800 
2801 	quirks |= check_vendor_combination_bug(pdev);
2802 
2803 	/*
2804 	 * Double check that our mempool alloc size will cover the biggest
2805 	 * command we support.
2806 	 */
2807 	alloc_size = nvme_pci_iod_alloc_size(dev, NVME_MAX_KB_SZ,
2808 						NVME_MAX_SEGS, true);
2809 	WARN_ON_ONCE(alloc_size > PAGE_SIZE);
2810 
2811 	dev->iod_mempool = mempool_create_node(1, mempool_kmalloc,
2812 						mempool_kfree,
2813 						(void *) alloc_size,
2814 						GFP_KERNEL, node);
2815 	if (!dev->iod_mempool) {
2816 		result = -ENOMEM;
2817 		goto release_pools;
2818 	}
2819 
2820 	result = nvme_init_ctrl(&dev->ctrl, &pdev->dev, &nvme_pci_ctrl_ops,
2821 			quirks);
2822 	if (result)
2823 		goto release_mempool;
2824 
2825 	dev_info(dev->ctrl.device, "pci function %s\n", dev_name(&pdev->dev));
2826 
2827 	nvme_reset_ctrl(&dev->ctrl);
2828 	nvme_get_ctrl(&dev->ctrl);
2829 	async_schedule(nvme_async_probe, dev);
2830 
2831 	return 0;
2832 
2833  release_mempool:
2834 	mempool_destroy(dev->iod_mempool);
2835  release_pools:
2836 	nvme_release_prp_pools(dev);
2837  unmap:
2838 	nvme_dev_unmap(dev);
2839  put_pci:
2840 	put_device(dev->dev);
2841  free:
2842 	kfree(dev->queues);
2843 	kfree(dev);
2844 	return result;
2845 }
2846 
2847 static void nvme_reset_prepare(struct pci_dev *pdev)
2848 {
2849 	struct nvme_dev *dev = pci_get_drvdata(pdev);
2850 
2851 	/*
2852 	 * We don't need to check the return value from waiting for the reset
2853 	 * state as pci_dev device lock is held, making it impossible to race
2854 	 * with ->remove().
2855 	 */
2856 	nvme_disable_prepare_reset(dev, false);
2857 	nvme_sync_queues(&dev->ctrl);
2858 }
2859 
2860 static void nvme_reset_done(struct pci_dev *pdev)
2861 {
2862 	struct nvme_dev *dev = pci_get_drvdata(pdev);
2863 
2864 	if (!nvme_try_sched_reset(&dev->ctrl))
2865 		flush_work(&dev->ctrl.reset_work);
2866 }
2867 
2868 static void nvme_shutdown(struct pci_dev *pdev)
2869 {
2870 	struct nvme_dev *dev = pci_get_drvdata(pdev);
2871 	nvme_disable_prepare_reset(dev, true);
2872 }
2873 
2874 /*
2875  * The driver's remove may be called on a device in a partially initialized
2876  * state. This function must not have any dependencies on the device state in
2877  * order to proceed.
2878  */
2879 static void nvme_remove(struct pci_dev *pdev)
2880 {
2881 	struct nvme_dev *dev = pci_get_drvdata(pdev);
2882 
2883 	nvme_change_ctrl_state(&dev->ctrl, NVME_CTRL_DELETING);
2884 	pci_set_drvdata(pdev, NULL);
2885 
2886 	if (!pci_device_is_present(pdev)) {
2887 		nvme_change_ctrl_state(&dev->ctrl, NVME_CTRL_DEAD);
2888 		nvme_dev_disable(dev, true);
2889 		nvme_dev_remove_admin(dev);
2890 	}
2891 
2892 	flush_work(&dev->ctrl.reset_work);
2893 	nvme_stop_ctrl(&dev->ctrl);
2894 	nvme_remove_namespaces(&dev->ctrl);
2895 	nvme_dev_disable(dev, true);
2896 	nvme_release_cmb(dev);
2897 	nvme_free_host_mem(dev);
2898 	nvme_dev_remove_admin(dev);
2899 	nvme_free_queues(dev, 0);
2900 	nvme_uninit_ctrl(&dev->ctrl);
2901 	nvme_release_prp_pools(dev);
2902 	nvme_dev_unmap(dev);
2903 	nvme_put_ctrl(&dev->ctrl);
2904 }
2905 
2906 #ifdef CONFIG_PM_SLEEP
2907 static int nvme_get_power_state(struct nvme_ctrl *ctrl, u32 *ps)
2908 {
2909 	return nvme_get_features(ctrl, NVME_FEAT_POWER_MGMT, 0, NULL, 0, ps);
2910 }
2911 
2912 static int nvme_set_power_state(struct nvme_ctrl *ctrl, u32 ps)
2913 {
2914 	return nvme_set_features(ctrl, NVME_FEAT_POWER_MGMT, ps, NULL, 0, NULL);
2915 }
2916 
2917 static int nvme_resume(struct device *dev)
2918 {
2919 	struct nvme_dev *ndev = pci_get_drvdata(to_pci_dev(dev));
2920 	struct nvme_ctrl *ctrl = &ndev->ctrl;
2921 
2922 	if (ndev->last_ps == U32_MAX ||
2923 	    nvme_set_power_state(ctrl, ndev->last_ps) != 0)
2924 		return nvme_try_sched_reset(&ndev->ctrl);
2925 	return 0;
2926 }
2927 
2928 static int nvme_suspend(struct device *dev)
2929 {
2930 	struct pci_dev *pdev = to_pci_dev(dev);
2931 	struct nvme_dev *ndev = pci_get_drvdata(pdev);
2932 	struct nvme_ctrl *ctrl = &ndev->ctrl;
2933 	int ret = -EBUSY;
2934 
2935 	ndev->last_ps = U32_MAX;
2936 
2937 	/*
2938 	 * The platform does not remove power for a kernel managed suspend so
2939 	 * use host managed nvme power settings for lowest idle power if
2940 	 * possible. This should have quicker resume latency than a full device
2941 	 * shutdown.  But if the firmware is involved after the suspend or the
2942 	 * device does not support any non-default power states, shut down the
2943 	 * device fully.
2944 	 *
2945 	 * If ASPM is not enabled for the device, shut down the device and allow
2946 	 * the PCI bus layer to put it into D3 in order to take the PCIe link
2947 	 * down, so as to allow the platform to achieve its minimum low-power
2948 	 * state (which may not be possible if the link is up).
2949 	 */
2950 	if (pm_suspend_via_firmware() || !ctrl->npss ||
2951 	    !pcie_aspm_enabled(pdev) ||
2952 	    (ndev->ctrl.quirks & NVME_QUIRK_SIMPLE_SUSPEND))
2953 		return nvme_disable_prepare_reset(ndev, true);
2954 
2955 	nvme_start_freeze(ctrl);
2956 	nvme_wait_freeze(ctrl);
2957 	nvme_sync_queues(ctrl);
2958 
2959 	if (ctrl->state != NVME_CTRL_LIVE)
2960 		goto unfreeze;
2961 
2962 	ret = nvme_get_power_state(ctrl, &ndev->last_ps);
2963 	if (ret < 0)
2964 		goto unfreeze;
2965 
2966 	/*
2967 	 * A saved state prevents pci pm from generically controlling the
2968 	 * device's power. If we're using protocol specific settings, we don't
2969 	 * want pci interfering.
2970 	 */
2971 	pci_save_state(pdev);
2972 
2973 	ret = nvme_set_power_state(ctrl, ctrl->npss);
2974 	if (ret < 0)
2975 		goto unfreeze;
2976 
2977 	if (ret) {
2978 		/* discard the saved state */
2979 		pci_load_saved_state(pdev, NULL);
2980 
2981 		/*
2982 		 * Clearing npss forces a controller reset on resume. The
2983 		 * correct value will be rediscovered then.
2984 		 */
2985 		ret = nvme_disable_prepare_reset(ndev, true);
2986 		ctrl->npss = 0;
2987 	}
2988 unfreeze:
2989 	nvme_unfreeze(ctrl);
2990 	return ret;
2991 }
2992 
2993 static int nvme_simple_suspend(struct device *dev)
2994 {
2995 	struct nvme_dev *ndev = pci_get_drvdata(to_pci_dev(dev));
2996 	return nvme_disable_prepare_reset(ndev, true);
2997 }
2998 
2999 static int nvme_simple_resume(struct device *dev)
3000 {
3001 	struct pci_dev *pdev = to_pci_dev(dev);
3002 	struct nvme_dev *ndev = pci_get_drvdata(pdev);
3003 
3004 	return nvme_try_sched_reset(&ndev->ctrl);
3005 }
3006 
3007 static const struct dev_pm_ops nvme_dev_pm_ops = {
3008 	.suspend	= nvme_suspend,
3009 	.resume		= nvme_resume,
3010 	.freeze		= nvme_simple_suspend,
3011 	.thaw		= nvme_simple_resume,
3012 	.poweroff	= nvme_simple_suspend,
3013 	.restore	= nvme_simple_resume,
3014 };
3015 #endif /* CONFIG_PM_SLEEP */
3016 
3017 static pci_ers_result_t nvme_error_detected(struct pci_dev *pdev,
3018 						pci_channel_state_t state)
3019 {
3020 	struct nvme_dev *dev = pci_get_drvdata(pdev);
3021 
3022 	/*
3023 	 * A frozen channel requires a reset. When detected, this method will
3024 	 * shutdown the controller to quiesce. The controller will be restarted
3025 	 * after the slot reset through driver's slot_reset callback.
3026 	 */
3027 	switch (state) {
3028 	case pci_channel_io_normal:
3029 		return PCI_ERS_RESULT_CAN_RECOVER;
3030 	case pci_channel_io_frozen:
3031 		dev_warn(dev->ctrl.device,
3032 			"frozen state error detected, reset controller\n");
3033 		nvme_dev_disable(dev, false);
3034 		return PCI_ERS_RESULT_NEED_RESET;
3035 	case pci_channel_io_perm_failure:
3036 		dev_warn(dev->ctrl.device,
3037 			"failure state error detected, request disconnect\n");
3038 		return PCI_ERS_RESULT_DISCONNECT;
3039 	}
3040 	return PCI_ERS_RESULT_NEED_RESET;
3041 }
3042 
3043 static pci_ers_result_t nvme_slot_reset(struct pci_dev *pdev)
3044 {
3045 	struct nvme_dev *dev = pci_get_drvdata(pdev);
3046 
3047 	dev_info(dev->ctrl.device, "restart after slot reset\n");
3048 	pci_restore_state(pdev);
3049 	nvme_reset_ctrl(&dev->ctrl);
3050 	return PCI_ERS_RESULT_RECOVERED;
3051 }
3052 
3053 static void nvme_error_resume(struct pci_dev *pdev)
3054 {
3055 	struct nvme_dev *dev = pci_get_drvdata(pdev);
3056 
3057 	flush_work(&dev->ctrl.reset_work);
3058 }
3059 
3060 static const struct pci_error_handlers nvme_err_handler = {
3061 	.error_detected	= nvme_error_detected,
3062 	.slot_reset	= nvme_slot_reset,
3063 	.resume		= nvme_error_resume,
3064 	.reset_prepare	= nvme_reset_prepare,
3065 	.reset_done	= nvme_reset_done,
3066 };
3067 
3068 static const struct pci_device_id nvme_id_table[] = {
3069 	{ PCI_VDEVICE(INTEL, 0x0953),
3070 		.driver_data = NVME_QUIRK_STRIPE_SIZE |
3071 				NVME_QUIRK_DEALLOCATE_ZEROES, },
3072 	{ PCI_VDEVICE(INTEL, 0x0a53),
3073 		.driver_data = NVME_QUIRK_STRIPE_SIZE |
3074 				NVME_QUIRK_DEALLOCATE_ZEROES, },
3075 	{ PCI_VDEVICE(INTEL, 0x0a54),
3076 		.driver_data = NVME_QUIRK_STRIPE_SIZE |
3077 				NVME_QUIRK_DEALLOCATE_ZEROES, },
3078 	{ PCI_VDEVICE(INTEL, 0x0a55),
3079 		.driver_data = NVME_QUIRK_STRIPE_SIZE |
3080 				NVME_QUIRK_DEALLOCATE_ZEROES, },
3081 	{ PCI_VDEVICE(INTEL, 0xf1a5),	/* Intel 600P/P3100 */
3082 		.driver_data = NVME_QUIRK_NO_DEEPEST_PS |
3083 				NVME_QUIRK_MEDIUM_PRIO_SQ |
3084 				NVME_QUIRK_NO_TEMP_THRESH_CHANGE },
3085 	{ PCI_VDEVICE(INTEL, 0xf1a6),	/* Intel 760p/Pro 7600p */
3086 		.driver_data = NVME_QUIRK_IGNORE_DEV_SUBNQN, },
3087 	{ PCI_VDEVICE(INTEL, 0x5845),	/* Qemu emulated controller */
3088 		.driver_data = NVME_QUIRK_IDENTIFY_CNS |
3089 				NVME_QUIRK_DISABLE_WRITE_ZEROES, },
3090 	{ PCI_DEVICE(0x1bb1, 0x0100),   /* Seagate Nytro Flash Storage */
3091 		.driver_data = NVME_QUIRK_DELAY_BEFORE_CHK_RDY, },
3092 	{ PCI_DEVICE(0x1c58, 0x0003),	/* HGST adapter */
3093 		.driver_data = NVME_QUIRK_DELAY_BEFORE_CHK_RDY, },
3094 	{ PCI_DEVICE(0x1c58, 0x0023),	/* WDC SN200 adapter */
3095 		.driver_data = NVME_QUIRK_DELAY_BEFORE_CHK_RDY, },
3096 	{ PCI_DEVICE(0x1c5f, 0x0540),	/* Memblaze Pblaze4 adapter */
3097 		.driver_data = NVME_QUIRK_DELAY_BEFORE_CHK_RDY, },
3098 	{ PCI_DEVICE(0x144d, 0xa821),   /* Samsung PM1725 */
3099 		.driver_data = NVME_QUIRK_DELAY_BEFORE_CHK_RDY, },
3100 	{ PCI_DEVICE(0x144d, 0xa822),   /* Samsung PM1725a */
3101 		.driver_data = NVME_QUIRK_DELAY_BEFORE_CHK_RDY, },
3102 	{ PCI_DEVICE(0x1d1d, 0x1f1f),	/* LighNVM qemu device */
3103 		.driver_data = NVME_QUIRK_LIGHTNVM, },
3104 	{ PCI_DEVICE(0x1d1d, 0x2807),	/* CNEX WL */
3105 		.driver_data = NVME_QUIRK_LIGHTNVM, },
3106 	{ PCI_DEVICE(0x1d1d, 0x2601),	/* CNEX Granby */
3107 		.driver_data = NVME_QUIRK_LIGHTNVM, },
3108 	{ PCI_DEVICE(0x10ec, 0x5762),   /* ADATA SX6000LNP */
3109 		.driver_data = NVME_QUIRK_IGNORE_DEV_SUBNQN, },
3110 	{ PCI_DEVICE(0x1cc1, 0x8201),   /* ADATA SX8200PNP 512GB */
3111 		.driver_data = NVME_QUIRK_NO_DEEPEST_PS |
3112 				NVME_QUIRK_IGNORE_DEV_SUBNQN, },
3113 	{ PCI_DEVICE_CLASS(PCI_CLASS_STORAGE_EXPRESS, 0xffffff) },
3114 	{ PCI_DEVICE(PCI_VENDOR_ID_APPLE, 0x2001) },
3115 	{ PCI_DEVICE(PCI_VENDOR_ID_APPLE, 0x2003) },
3116 	{ PCI_DEVICE(PCI_VENDOR_ID_APPLE, 0x2005),
3117 		.driver_data = NVME_QUIRK_SINGLE_VECTOR |
3118 				NVME_QUIRK_128_BYTES_SQES |
3119 				NVME_QUIRK_SHARED_TAGS },
3120 	{ 0, }
3121 };
3122 MODULE_DEVICE_TABLE(pci, nvme_id_table);
3123 
3124 static struct pci_driver nvme_driver = {
3125 	.name		= "nvme",
3126 	.id_table	= nvme_id_table,
3127 	.probe		= nvme_probe,
3128 	.remove		= nvme_remove,
3129 	.shutdown	= nvme_shutdown,
3130 #ifdef CONFIG_PM_SLEEP
3131 	.driver		= {
3132 		.pm	= &nvme_dev_pm_ops,
3133 	},
3134 #endif
3135 	.sriov_configure = pci_sriov_configure_simple,
3136 	.err_handler	= &nvme_err_handler,
3137 };
3138 
3139 static int __init nvme_init(void)
3140 {
3141 	BUILD_BUG_ON(sizeof(struct nvme_create_cq) != 64);
3142 	BUILD_BUG_ON(sizeof(struct nvme_create_sq) != 64);
3143 	BUILD_BUG_ON(sizeof(struct nvme_delete_queue) != 64);
3144 	BUILD_BUG_ON(IRQ_AFFINITY_MAX_SETS < 2);
3145 	return pci_register_driver(&nvme_driver);
3146 }
3147 
3148 static void __exit nvme_exit(void)
3149 {
3150 	pci_unregister_driver(&nvme_driver);
3151 	flush_workqueue(nvme_wq);
3152 }
3153 
3154 MODULE_AUTHOR("Matthew Wilcox <willy@linux.intel.com>");
3155 MODULE_LICENSE("GPL");
3156 MODULE_VERSION("1.0");
3157 module_init(nvme_init);
3158 module_exit(nvme_exit);
3159