xref: /openbmc/linux/drivers/nvme/host/pci.c (revision 6355592e)
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 	/* P2PDMA requests do not need to be unmapped */
553 	if (!is_pci_p2pdma_page(sg_page(iod->sg)))
554 		dma_unmap_sg(dev->dev, iod->sg, iod->nents, rq_dma_dir(req));
555 
556 
557 	if (iod->npages == 0)
558 		dma_pool_free(dev->prp_small_pool, nvme_pci_iod_list(req)[0],
559 			dma_addr);
560 
561 	for (i = 0; i < iod->npages; i++) {
562 		void *addr = nvme_pci_iod_list(req)[i];
563 
564 		if (iod->use_sgl) {
565 			struct nvme_sgl_desc *sg_list = addr;
566 
567 			next_dma_addr =
568 			    le64_to_cpu((sg_list[SGES_PER_PAGE - 1]).addr);
569 		} else {
570 			__le64 *prp_list = addr;
571 
572 			next_dma_addr = le64_to_cpu(prp_list[last_prp]);
573 		}
574 
575 		dma_pool_free(dev->prp_page_pool, addr, dma_addr);
576 		dma_addr = next_dma_addr;
577 	}
578 
579 	mempool_free(iod->sg, dev->iod_mempool);
580 }
581 
582 static void nvme_print_sgl(struct scatterlist *sgl, int nents)
583 {
584 	int i;
585 	struct scatterlist *sg;
586 
587 	for_each_sg(sgl, sg, nents, i) {
588 		dma_addr_t phys = sg_phys(sg);
589 		pr_warn("sg[%d] phys_addr:%pad offset:%d length:%d "
590 			"dma_address:%pad dma_length:%d\n",
591 			i, &phys, sg->offset, sg->length, &sg_dma_address(sg),
592 			sg_dma_len(sg));
593 	}
594 }
595 
596 static blk_status_t nvme_pci_setup_prps(struct nvme_dev *dev,
597 		struct request *req, struct nvme_rw_command *cmnd)
598 {
599 	struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
600 	struct dma_pool *pool;
601 	int length = blk_rq_payload_bytes(req);
602 	struct scatterlist *sg = iod->sg;
603 	int dma_len = sg_dma_len(sg);
604 	u64 dma_addr = sg_dma_address(sg);
605 	u32 page_size = dev->ctrl.page_size;
606 	int offset = dma_addr & (page_size - 1);
607 	__le64 *prp_list;
608 	void **list = nvme_pci_iod_list(req);
609 	dma_addr_t prp_dma;
610 	int nprps, i;
611 
612 	length -= (page_size - offset);
613 	if (length <= 0) {
614 		iod->first_dma = 0;
615 		goto done;
616 	}
617 
618 	dma_len -= (page_size - offset);
619 	if (dma_len) {
620 		dma_addr += (page_size - offset);
621 	} else {
622 		sg = sg_next(sg);
623 		dma_addr = sg_dma_address(sg);
624 		dma_len = sg_dma_len(sg);
625 	}
626 
627 	if (length <= page_size) {
628 		iod->first_dma = dma_addr;
629 		goto done;
630 	}
631 
632 	nprps = DIV_ROUND_UP(length, page_size);
633 	if (nprps <= (256 / 8)) {
634 		pool = dev->prp_small_pool;
635 		iod->npages = 0;
636 	} else {
637 		pool = dev->prp_page_pool;
638 		iod->npages = 1;
639 	}
640 
641 	prp_list = dma_pool_alloc(pool, GFP_ATOMIC, &prp_dma);
642 	if (!prp_list) {
643 		iod->first_dma = dma_addr;
644 		iod->npages = -1;
645 		return BLK_STS_RESOURCE;
646 	}
647 	list[0] = prp_list;
648 	iod->first_dma = prp_dma;
649 	i = 0;
650 	for (;;) {
651 		if (i == page_size >> 3) {
652 			__le64 *old_prp_list = prp_list;
653 			prp_list = dma_pool_alloc(pool, GFP_ATOMIC, &prp_dma);
654 			if (!prp_list)
655 				return BLK_STS_RESOURCE;
656 			list[iod->npages++] = prp_list;
657 			prp_list[0] = old_prp_list[i - 1];
658 			old_prp_list[i - 1] = cpu_to_le64(prp_dma);
659 			i = 1;
660 		}
661 		prp_list[i++] = cpu_to_le64(dma_addr);
662 		dma_len -= page_size;
663 		dma_addr += page_size;
664 		length -= page_size;
665 		if (length <= 0)
666 			break;
667 		if (dma_len > 0)
668 			continue;
669 		if (unlikely(dma_len < 0))
670 			goto bad_sgl;
671 		sg = sg_next(sg);
672 		dma_addr = sg_dma_address(sg);
673 		dma_len = sg_dma_len(sg);
674 	}
675 
676 done:
677 	cmnd->dptr.prp1 = cpu_to_le64(sg_dma_address(iod->sg));
678 	cmnd->dptr.prp2 = cpu_to_le64(iod->first_dma);
679 
680 	return BLK_STS_OK;
681 
682  bad_sgl:
683 	WARN(DO_ONCE(nvme_print_sgl, iod->sg, iod->nents),
684 			"Invalid SGL for payload:%d nents:%d\n",
685 			blk_rq_payload_bytes(req), iod->nents);
686 	return BLK_STS_IOERR;
687 }
688 
689 static void nvme_pci_sgl_set_data(struct nvme_sgl_desc *sge,
690 		struct scatterlist *sg)
691 {
692 	sge->addr = cpu_to_le64(sg_dma_address(sg));
693 	sge->length = cpu_to_le32(sg_dma_len(sg));
694 	sge->type = NVME_SGL_FMT_DATA_DESC << 4;
695 }
696 
697 static void nvme_pci_sgl_set_seg(struct nvme_sgl_desc *sge,
698 		dma_addr_t dma_addr, int entries)
699 {
700 	sge->addr = cpu_to_le64(dma_addr);
701 	if (entries < SGES_PER_PAGE) {
702 		sge->length = cpu_to_le32(entries * sizeof(*sge));
703 		sge->type = NVME_SGL_FMT_LAST_SEG_DESC << 4;
704 	} else {
705 		sge->length = cpu_to_le32(PAGE_SIZE);
706 		sge->type = NVME_SGL_FMT_SEG_DESC << 4;
707 	}
708 }
709 
710 static blk_status_t nvme_pci_setup_sgls(struct nvme_dev *dev,
711 		struct request *req, struct nvme_rw_command *cmd, int entries)
712 {
713 	struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
714 	struct dma_pool *pool;
715 	struct nvme_sgl_desc *sg_list;
716 	struct scatterlist *sg = iod->sg;
717 	dma_addr_t sgl_dma;
718 	int i = 0;
719 
720 	/* setting the transfer type as SGL */
721 	cmd->flags = NVME_CMD_SGL_METABUF;
722 
723 	if (entries == 1) {
724 		nvme_pci_sgl_set_data(&cmd->dptr.sgl, sg);
725 		return BLK_STS_OK;
726 	}
727 
728 	if (entries <= (256 / sizeof(struct nvme_sgl_desc))) {
729 		pool = dev->prp_small_pool;
730 		iod->npages = 0;
731 	} else {
732 		pool = dev->prp_page_pool;
733 		iod->npages = 1;
734 	}
735 
736 	sg_list = dma_pool_alloc(pool, GFP_ATOMIC, &sgl_dma);
737 	if (!sg_list) {
738 		iod->npages = -1;
739 		return BLK_STS_RESOURCE;
740 	}
741 
742 	nvme_pci_iod_list(req)[0] = sg_list;
743 	iod->first_dma = sgl_dma;
744 
745 	nvme_pci_sgl_set_seg(&cmd->dptr.sgl, sgl_dma, entries);
746 
747 	do {
748 		if (i == SGES_PER_PAGE) {
749 			struct nvme_sgl_desc *old_sg_desc = sg_list;
750 			struct nvme_sgl_desc *link = &old_sg_desc[i - 1];
751 
752 			sg_list = dma_pool_alloc(pool, GFP_ATOMIC, &sgl_dma);
753 			if (!sg_list)
754 				return BLK_STS_RESOURCE;
755 
756 			i = 0;
757 			nvme_pci_iod_list(req)[iod->npages++] = sg_list;
758 			sg_list[i++] = *link;
759 			nvme_pci_sgl_set_seg(link, sgl_dma, entries);
760 		}
761 
762 		nvme_pci_sgl_set_data(&sg_list[i++], sg);
763 		sg = sg_next(sg);
764 	} while (--entries > 0);
765 
766 	return BLK_STS_OK;
767 }
768 
769 static blk_status_t nvme_setup_prp_simple(struct nvme_dev *dev,
770 		struct request *req, struct nvme_rw_command *cmnd,
771 		struct bio_vec *bv)
772 {
773 	struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
774 	unsigned int first_prp_len = dev->ctrl.page_size - bv->bv_offset;
775 
776 	iod->first_dma = dma_map_bvec(dev->dev, bv, rq_dma_dir(req), 0);
777 	if (dma_mapping_error(dev->dev, iod->first_dma))
778 		return BLK_STS_RESOURCE;
779 	iod->dma_len = bv->bv_len;
780 
781 	cmnd->dptr.prp1 = cpu_to_le64(iod->first_dma);
782 	if (bv->bv_len > first_prp_len)
783 		cmnd->dptr.prp2 = cpu_to_le64(iod->first_dma + first_prp_len);
784 	return 0;
785 }
786 
787 static blk_status_t nvme_setup_sgl_simple(struct nvme_dev *dev,
788 		struct request *req, struct nvme_rw_command *cmnd,
789 		struct bio_vec *bv)
790 {
791 	struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
792 
793 	iod->first_dma = dma_map_bvec(dev->dev, bv, rq_dma_dir(req), 0);
794 	if (dma_mapping_error(dev->dev, iod->first_dma))
795 		return BLK_STS_RESOURCE;
796 	iod->dma_len = bv->bv_len;
797 
798 	cmnd->flags = NVME_CMD_SGL_METABUF;
799 	cmnd->dptr.sgl.addr = cpu_to_le64(iod->first_dma);
800 	cmnd->dptr.sgl.length = cpu_to_le32(iod->dma_len);
801 	cmnd->dptr.sgl.type = NVME_SGL_FMT_DATA_DESC << 4;
802 	return 0;
803 }
804 
805 static blk_status_t nvme_map_data(struct nvme_dev *dev, struct request *req,
806 		struct nvme_command *cmnd)
807 {
808 	struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
809 	blk_status_t ret = BLK_STS_RESOURCE;
810 	int nr_mapped;
811 
812 	if (blk_rq_nr_phys_segments(req) == 1) {
813 		struct bio_vec bv = req_bvec(req);
814 
815 		if (!is_pci_p2pdma_page(bv.bv_page)) {
816 			if (bv.bv_offset + bv.bv_len <= dev->ctrl.page_size * 2)
817 				return nvme_setup_prp_simple(dev, req,
818 							     &cmnd->rw, &bv);
819 
820 			if (iod->nvmeq->qid &&
821 			    dev->ctrl.sgls & ((1 << 0) | (1 << 1)))
822 				return nvme_setup_sgl_simple(dev, req,
823 							     &cmnd->rw, &bv);
824 		}
825 	}
826 
827 	iod->dma_len = 0;
828 	iod->sg = mempool_alloc(dev->iod_mempool, GFP_ATOMIC);
829 	if (!iod->sg)
830 		return BLK_STS_RESOURCE;
831 	sg_init_table(iod->sg, blk_rq_nr_phys_segments(req));
832 	iod->nents = blk_rq_map_sg(req->q, req, iod->sg);
833 	if (!iod->nents)
834 		goto out;
835 
836 	if (is_pci_p2pdma_page(sg_page(iod->sg)))
837 		nr_mapped = pci_p2pdma_map_sg(dev->dev, iod->sg, iod->nents,
838 					      rq_dma_dir(req));
839 	else
840 		nr_mapped = dma_map_sg_attrs(dev->dev, iod->sg, iod->nents,
841 					     rq_dma_dir(req), DMA_ATTR_NO_WARN);
842 	if (!nr_mapped)
843 		goto out;
844 
845 	iod->use_sgl = nvme_pci_use_sgls(dev, req);
846 	if (iod->use_sgl)
847 		ret = nvme_pci_setup_sgls(dev, req, &cmnd->rw, nr_mapped);
848 	else
849 		ret = nvme_pci_setup_prps(dev, req, &cmnd->rw);
850 out:
851 	if (ret != BLK_STS_OK)
852 		nvme_unmap_data(dev, req);
853 	return ret;
854 }
855 
856 static blk_status_t nvme_map_metadata(struct nvme_dev *dev, struct request *req,
857 		struct nvme_command *cmnd)
858 {
859 	struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
860 
861 	iod->meta_dma = dma_map_bvec(dev->dev, rq_integrity_vec(req),
862 			rq_dma_dir(req), 0);
863 	if (dma_mapping_error(dev->dev, iod->meta_dma))
864 		return BLK_STS_IOERR;
865 	cmnd->rw.metadata = cpu_to_le64(iod->meta_dma);
866 	return 0;
867 }
868 
869 /*
870  * NOTE: ns is NULL when called on the admin queue.
871  */
872 static blk_status_t nvme_queue_rq(struct blk_mq_hw_ctx *hctx,
873 			 const struct blk_mq_queue_data *bd)
874 {
875 	struct nvme_ns *ns = hctx->queue->queuedata;
876 	struct nvme_queue *nvmeq = hctx->driver_data;
877 	struct nvme_dev *dev = nvmeq->dev;
878 	struct request *req = bd->rq;
879 	struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
880 	struct nvme_command cmnd;
881 	blk_status_t ret;
882 
883 	iod->aborted = 0;
884 	iod->npages = -1;
885 	iod->nents = 0;
886 
887 	/*
888 	 * We should not need to do this, but we're still using this to
889 	 * ensure we can drain requests on a dying queue.
890 	 */
891 	if (unlikely(!test_bit(NVMEQ_ENABLED, &nvmeq->flags)))
892 		return BLK_STS_IOERR;
893 
894 	ret = nvme_setup_cmd(ns, req, &cmnd);
895 	if (ret)
896 		return ret;
897 
898 	if (blk_rq_nr_phys_segments(req)) {
899 		ret = nvme_map_data(dev, req, &cmnd);
900 		if (ret)
901 			goto out_free_cmd;
902 	}
903 
904 	if (blk_integrity_rq(req)) {
905 		ret = nvme_map_metadata(dev, req, &cmnd);
906 		if (ret)
907 			goto out_unmap_data;
908 	}
909 
910 	blk_mq_start_request(req);
911 	nvme_submit_cmd(nvmeq, &cmnd, bd->last);
912 	return BLK_STS_OK;
913 out_unmap_data:
914 	nvme_unmap_data(dev, req);
915 out_free_cmd:
916 	nvme_cleanup_cmd(req);
917 	return ret;
918 }
919 
920 static void nvme_pci_complete_rq(struct request *req)
921 {
922 	struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
923 	struct nvme_dev *dev = iod->nvmeq->dev;
924 
925 	nvme_cleanup_cmd(req);
926 	if (blk_integrity_rq(req))
927 		dma_unmap_page(dev->dev, iod->meta_dma,
928 			       rq_integrity_vec(req)->bv_len, rq_data_dir(req));
929 	if (blk_rq_nr_phys_segments(req))
930 		nvme_unmap_data(dev, req);
931 	nvme_complete_rq(req);
932 }
933 
934 /* We read the CQE phase first to check if the rest of the entry is valid */
935 static inline bool nvme_cqe_pending(struct nvme_queue *nvmeq)
936 {
937 	return (le16_to_cpu(nvmeq->cqes[nvmeq->cq_head].status) & 1) ==
938 			nvmeq->cq_phase;
939 }
940 
941 static inline void nvme_ring_cq_doorbell(struct nvme_queue *nvmeq)
942 {
943 	u16 head = nvmeq->cq_head;
944 
945 	if (nvme_dbbuf_update_and_check_event(head, nvmeq->dbbuf_cq_db,
946 					      nvmeq->dbbuf_cq_ei))
947 		writel(head, nvmeq->q_db + nvmeq->dev->db_stride);
948 }
949 
950 static inline void nvme_handle_cqe(struct nvme_queue *nvmeq, u16 idx)
951 {
952 	volatile struct nvme_completion *cqe = &nvmeq->cqes[idx];
953 	struct request *req;
954 
955 	if (unlikely(cqe->command_id >= nvmeq->q_depth)) {
956 		dev_warn(nvmeq->dev->ctrl.device,
957 			"invalid id %d completed on queue %d\n",
958 			cqe->command_id, le16_to_cpu(cqe->sq_id));
959 		return;
960 	}
961 
962 	/*
963 	 * AEN requests are special as they don't time out and can
964 	 * survive any kind of queue freeze and often don't respond to
965 	 * aborts.  We don't even bother to allocate a struct request
966 	 * for them but rather special case them here.
967 	 */
968 	if (unlikely(nvmeq->qid == 0 &&
969 			cqe->command_id >= NVME_AQ_BLK_MQ_DEPTH)) {
970 		nvme_complete_async_event(&nvmeq->dev->ctrl,
971 				cqe->status, &cqe->result);
972 		return;
973 	}
974 
975 	req = blk_mq_tag_to_rq(*nvmeq->tags, cqe->command_id);
976 	trace_nvme_sq(req, cqe->sq_head, nvmeq->sq_tail);
977 	nvme_end_request(req, cqe->status, cqe->result);
978 }
979 
980 static void nvme_complete_cqes(struct nvme_queue *nvmeq, u16 start, u16 end)
981 {
982 	while (start != end) {
983 		nvme_handle_cqe(nvmeq, start);
984 		if (++start == nvmeq->q_depth)
985 			start = 0;
986 	}
987 }
988 
989 static inline void nvme_update_cq_head(struct nvme_queue *nvmeq)
990 {
991 	if (nvmeq->cq_head == nvmeq->q_depth - 1) {
992 		nvmeq->cq_head = 0;
993 		nvmeq->cq_phase = !nvmeq->cq_phase;
994 	} else {
995 		nvmeq->cq_head++;
996 	}
997 }
998 
999 static inline int nvme_process_cq(struct nvme_queue *nvmeq, u16 *start,
1000 				  u16 *end, unsigned int tag)
1001 {
1002 	int found = 0;
1003 
1004 	*start = nvmeq->cq_head;
1005 	while (nvme_cqe_pending(nvmeq)) {
1006 		if (tag == -1U || nvmeq->cqes[nvmeq->cq_head].command_id == tag)
1007 			found++;
1008 		nvme_update_cq_head(nvmeq);
1009 	}
1010 	*end = nvmeq->cq_head;
1011 
1012 	if (*start != *end)
1013 		nvme_ring_cq_doorbell(nvmeq);
1014 	return found;
1015 }
1016 
1017 static irqreturn_t nvme_irq(int irq, void *data)
1018 {
1019 	struct nvme_queue *nvmeq = data;
1020 	irqreturn_t ret = IRQ_NONE;
1021 	u16 start, end;
1022 
1023 	/*
1024 	 * The rmb/wmb pair ensures we see all updates from a previous run of
1025 	 * the irq handler, even if that was on another CPU.
1026 	 */
1027 	rmb();
1028 	if (nvmeq->cq_head != nvmeq->last_cq_head)
1029 		ret = IRQ_HANDLED;
1030 	nvme_process_cq(nvmeq, &start, &end, -1);
1031 	nvmeq->last_cq_head = nvmeq->cq_head;
1032 	wmb();
1033 
1034 	if (start != end) {
1035 		nvme_complete_cqes(nvmeq, start, end);
1036 		return IRQ_HANDLED;
1037 	}
1038 
1039 	return ret;
1040 }
1041 
1042 static irqreturn_t nvme_irq_check(int irq, void *data)
1043 {
1044 	struct nvme_queue *nvmeq = data;
1045 	if (nvme_cqe_pending(nvmeq))
1046 		return IRQ_WAKE_THREAD;
1047 	return IRQ_NONE;
1048 }
1049 
1050 /*
1051  * Poll for completions any queue, including those not dedicated to polling.
1052  * Can be called from any context.
1053  */
1054 static int nvme_poll_irqdisable(struct nvme_queue *nvmeq, unsigned int tag)
1055 {
1056 	struct pci_dev *pdev = to_pci_dev(nvmeq->dev->dev);
1057 	u16 start, end;
1058 	int found;
1059 
1060 	/*
1061 	 * For a poll queue we need to protect against the polling thread
1062 	 * using the CQ lock.  For normal interrupt driven threads we have
1063 	 * to disable the interrupt to avoid racing with it.
1064 	 */
1065 	if (test_bit(NVMEQ_POLLED, &nvmeq->flags)) {
1066 		spin_lock(&nvmeq->cq_poll_lock);
1067 		found = nvme_process_cq(nvmeq, &start, &end, tag);
1068 		spin_unlock(&nvmeq->cq_poll_lock);
1069 	} else {
1070 		disable_irq(pci_irq_vector(pdev, nvmeq->cq_vector));
1071 		found = nvme_process_cq(nvmeq, &start, &end, tag);
1072 		enable_irq(pci_irq_vector(pdev, nvmeq->cq_vector));
1073 	}
1074 
1075 	nvme_complete_cqes(nvmeq, start, end);
1076 	return found;
1077 }
1078 
1079 static int nvme_poll(struct blk_mq_hw_ctx *hctx)
1080 {
1081 	struct nvme_queue *nvmeq = hctx->driver_data;
1082 	u16 start, end;
1083 	bool found;
1084 
1085 	if (!nvme_cqe_pending(nvmeq))
1086 		return 0;
1087 
1088 	spin_lock(&nvmeq->cq_poll_lock);
1089 	found = nvme_process_cq(nvmeq, &start, &end, -1);
1090 	spin_unlock(&nvmeq->cq_poll_lock);
1091 
1092 	nvme_complete_cqes(nvmeq, start, end);
1093 	return found;
1094 }
1095 
1096 static void nvme_pci_submit_async_event(struct nvme_ctrl *ctrl)
1097 {
1098 	struct nvme_dev *dev = to_nvme_dev(ctrl);
1099 	struct nvme_queue *nvmeq = &dev->queues[0];
1100 	struct nvme_command c;
1101 
1102 	memset(&c, 0, sizeof(c));
1103 	c.common.opcode = nvme_admin_async_event;
1104 	c.common.command_id = NVME_AQ_BLK_MQ_DEPTH;
1105 	nvme_submit_cmd(nvmeq, &c, true);
1106 }
1107 
1108 static int adapter_delete_queue(struct nvme_dev *dev, u8 opcode, u16 id)
1109 {
1110 	struct nvme_command c;
1111 
1112 	memset(&c, 0, sizeof(c));
1113 	c.delete_queue.opcode = opcode;
1114 	c.delete_queue.qid = cpu_to_le16(id);
1115 
1116 	return nvme_submit_sync_cmd(dev->ctrl.admin_q, &c, NULL, 0);
1117 }
1118 
1119 static int adapter_alloc_cq(struct nvme_dev *dev, u16 qid,
1120 		struct nvme_queue *nvmeq, s16 vector)
1121 {
1122 	struct nvme_command c;
1123 	int flags = NVME_QUEUE_PHYS_CONTIG;
1124 
1125 	if (!test_bit(NVMEQ_POLLED, &nvmeq->flags))
1126 		flags |= NVME_CQ_IRQ_ENABLED;
1127 
1128 	/*
1129 	 * Note: we (ab)use the fact that the prp fields survive if no data
1130 	 * is attached to the request.
1131 	 */
1132 	memset(&c, 0, sizeof(c));
1133 	c.create_cq.opcode = nvme_admin_create_cq;
1134 	c.create_cq.prp1 = cpu_to_le64(nvmeq->cq_dma_addr);
1135 	c.create_cq.cqid = cpu_to_le16(qid);
1136 	c.create_cq.qsize = cpu_to_le16(nvmeq->q_depth - 1);
1137 	c.create_cq.cq_flags = cpu_to_le16(flags);
1138 	c.create_cq.irq_vector = cpu_to_le16(vector);
1139 
1140 	return nvme_submit_sync_cmd(dev->ctrl.admin_q, &c, NULL, 0);
1141 }
1142 
1143 static int adapter_alloc_sq(struct nvme_dev *dev, u16 qid,
1144 						struct nvme_queue *nvmeq)
1145 {
1146 	struct nvme_ctrl *ctrl = &dev->ctrl;
1147 	struct nvme_command c;
1148 	int flags = NVME_QUEUE_PHYS_CONTIG;
1149 
1150 	/*
1151 	 * Some drives have a bug that auto-enables WRRU if MEDIUM isn't
1152 	 * set. Since URGENT priority is zeroes, it makes all queues
1153 	 * URGENT.
1154 	 */
1155 	if (ctrl->quirks & NVME_QUIRK_MEDIUM_PRIO_SQ)
1156 		flags |= NVME_SQ_PRIO_MEDIUM;
1157 
1158 	/*
1159 	 * Note: we (ab)use the fact that the prp fields survive if no data
1160 	 * is attached to the request.
1161 	 */
1162 	memset(&c, 0, sizeof(c));
1163 	c.create_sq.opcode = nvme_admin_create_sq;
1164 	c.create_sq.prp1 = cpu_to_le64(nvmeq->sq_dma_addr);
1165 	c.create_sq.sqid = cpu_to_le16(qid);
1166 	c.create_sq.qsize = cpu_to_le16(nvmeq->q_depth - 1);
1167 	c.create_sq.sq_flags = cpu_to_le16(flags);
1168 	c.create_sq.cqid = cpu_to_le16(qid);
1169 
1170 	return nvme_submit_sync_cmd(dev->ctrl.admin_q, &c, NULL, 0);
1171 }
1172 
1173 static int adapter_delete_cq(struct nvme_dev *dev, u16 cqid)
1174 {
1175 	return adapter_delete_queue(dev, nvme_admin_delete_cq, cqid);
1176 }
1177 
1178 static int adapter_delete_sq(struct nvme_dev *dev, u16 sqid)
1179 {
1180 	return adapter_delete_queue(dev, nvme_admin_delete_sq, sqid);
1181 }
1182 
1183 static void abort_endio(struct request *req, blk_status_t error)
1184 {
1185 	struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
1186 	struct nvme_queue *nvmeq = iod->nvmeq;
1187 
1188 	dev_warn(nvmeq->dev->ctrl.device,
1189 		 "Abort status: 0x%x", nvme_req(req)->status);
1190 	atomic_inc(&nvmeq->dev->ctrl.abort_limit);
1191 	blk_mq_free_request(req);
1192 }
1193 
1194 static bool nvme_should_reset(struct nvme_dev *dev, u32 csts)
1195 {
1196 
1197 	/* If true, indicates loss of adapter communication, possibly by a
1198 	 * NVMe Subsystem reset.
1199 	 */
1200 	bool nssro = dev->subsystem && (csts & NVME_CSTS_NSSRO);
1201 
1202 	/* If there is a reset/reinit ongoing, we shouldn't reset again. */
1203 	switch (dev->ctrl.state) {
1204 	case NVME_CTRL_RESETTING:
1205 	case NVME_CTRL_CONNECTING:
1206 		return false;
1207 	default:
1208 		break;
1209 	}
1210 
1211 	/* We shouldn't reset unless the controller is on fatal error state
1212 	 * _or_ if we lost the communication with it.
1213 	 */
1214 	if (!(csts & NVME_CSTS_CFS) && !nssro)
1215 		return false;
1216 
1217 	return true;
1218 }
1219 
1220 static void nvme_warn_reset(struct nvme_dev *dev, u32 csts)
1221 {
1222 	/* Read a config register to help see what died. */
1223 	u16 pci_status;
1224 	int result;
1225 
1226 	result = pci_read_config_word(to_pci_dev(dev->dev), PCI_STATUS,
1227 				      &pci_status);
1228 	if (result == PCIBIOS_SUCCESSFUL)
1229 		dev_warn(dev->ctrl.device,
1230 			 "controller is down; will reset: CSTS=0x%x, PCI_STATUS=0x%hx\n",
1231 			 csts, pci_status);
1232 	else
1233 		dev_warn(dev->ctrl.device,
1234 			 "controller is down; will reset: CSTS=0x%x, PCI_STATUS read failed (%d)\n",
1235 			 csts, result);
1236 }
1237 
1238 static enum blk_eh_timer_return nvme_timeout(struct request *req, bool reserved)
1239 {
1240 	struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
1241 	struct nvme_queue *nvmeq = iod->nvmeq;
1242 	struct nvme_dev *dev = nvmeq->dev;
1243 	struct request *abort_req;
1244 	struct nvme_command cmd;
1245 	u32 csts = readl(dev->bar + NVME_REG_CSTS);
1246 
1247 	/* If PCI error recovery process is happening, we cannot reset or
1248 	 * the recovery mechanism will surely fail.
1249 	 */
1250 	mb();
1251 	if (pci_channel_offline(to_pci_dev(dev->dev)))
1252 		return BLK_EH_RESET_TIMER;
1253 
1254 	/*
1255 	 * Reset immediately if the controller is failed
1256 	 */
1257 	if (nvme_should_reset(dev, csts)) {
1258 		nvme_warn_reset(dev, csts);
1259 		nvme_dev_disable(dev, false);
1260 		nvme_reset_ctrl(&dev->ctrl);
1261 		return BLK_EH_DONE;
1262 	}
1263 
1264 	/*
1265 	 * Did we miss an interrupt?
1266 	 */
1267 	if (nvme_poll_irqdisable(nvmeq, req->tag)) {
1268 		dev_warn(dev->ctrl.device,
1269 			 "I/O %d QID %d timeout, completion polled\n",
1270 			 req->tag, nvmeq->qid);
1271 		return BLK_EH_DONE;
1272 	}
1273 
1274 	/*
1275 	 * Shutdown immediately if controller times out while starting. The
1276 	 * reset work will see the pci device disabled when it gets the forced
1277 	 * cancellation error. All outstanding requests are completed on
1278 	 * shutdown, so we return BLK_EH_DONE.
1279 	 */
1280 	switch (dev->ctrl.state) {
1281 	case NVME_CTRL_CONNECTING:
1282 		nvme_change_ctrl_state(&dev->ctrl, NVME_CTRL_DELETING);
1283 		/* fall through */
1284 	case NVME_CTRL_DELETING:
1285 		dev_warn_ratelimited(dev->ctrl.device,
1286 			 "I/O %d QID %d timeout, disable controller\n",
1287 			 req->tag, nvmeq->qid);
1288 		nvme_dev_disable(dev, true);
1289 		nvme_req(req)->flags |= NVME_REQ_CANCELLED;
1290 		return BLK_EH_DONE;
1291 	case NVME_CTRL_RESETTING:
1292 		return BLK_EH_RESET_TIMER;
1293 	default:
1294 		break;
1295 	}
1296 
1297 	/*
1298  	 * Shutdown the controller immediately and schedule a reset if the
1299  	 * command was already aborted once before and still hasn't been
1300  	 * returned to the driver, or if this is the admin queue.
1301 	 */
1302 	if (!nvmeq->qid || iod->aborted) {
1303 		dev_warn(dev->ctrl.device,
1304 			 "I/O %d QID %d timeout, reset controller\n",
1305 			 req->tag, nvmeq->qid);
1306 		nvme_dev_disable(dev, false);
1307 		nvme_reset_ctrl(&dev->ctrl);
1308 
1309 		nvme_req(req)->flags |= NVME_REQ_CANCELLED;
1310 		return BLK_EH_DONE;
1311 	}
1312 
1313 	if (atomic_dec_return(&dev->ctrl.abort_limit) < 0) {
1314 		atomic_inc(&dev->ctrl.abort_limit);
1315 		return BLK_EH_RESET_TIMER;
1316 	}
1317 	iod->aborted = 1;
1318 
1319 	memset(&cmd, 0, sizeof(cmd));
1320 	cmd.abort.opcode = nvme_admin_abort_cmd;
1321 	cmd.abort.cid = req->tag;
1322 	cmd.abort.sqid = cpu_to_le16(nvmeq->qid);
1323 
1324 	dev_warn(nvmeq->dev->ctrl.device,
1325 		"I/O %d QID %d timeout, aborting\n",
1326 		 req->tag, nvmeq->qid);
1327 
1328 	abort_req = nvme_alloc_request(dev->ctrl.admin_q, &cmd,
1329 			BLK_MQ_REQ_NOWAIT, NVME_QID_ANY);
1330 	if (IS_ERR(abort_req)) {
1331 		atomic_inc(&dev->ctrl.abort_limit);
1332 		return BLK_EH_RESET_TIMER;
1333 	}
1334 
1335 	abort_req->timeout = ADMIN_TIMEOUT;
1336 	abort_req->end_io_data = NULL;
1337 	blk_execute_rq_nowait(abort_req->q, NULL, abort_req, 0, abort_endio);
1338 
1339 	/*
1340 	 * The aborted req will be completed on receiving the abort req.
1341 	 * We enable the timer again. If hit twice, it'll cause a device reset,
1342 	 * as the device then is in a faulty state.
1343 	 */
1344 	return BLK_EH_RESET_TIMER;
1345 }
1346 
1347 static void nvme_free_queue(struct nvme_queue *nvmeq)
1348 {
1349 	dma_free_coherent(nvmeq->dev->dev, CQ_SIZE(nvmeq),
1350 				(void *)nvmeq->cqes, nvmeq->cq_dma_addr);
1351 	if (!nvmeq->sq_cmds)
1352 		return;
1353 
1354 	if (test_and_clear_bit(NVMEQ_SQ_CMB, &nvmeq->flags)) {
1355 		pci_free_p2pmem(to_pci_dev(nvmeq->dev->dev),
1356 				nvmeq->sq_cmds, SQ_SIZE(nvmeq));
1357 	} else {
1358 		dma_free_coherent(nvmeq->dev->dev, SQ_SIZE(nvmeq),
1359 				nvmeq->sq_cmds, nvmeq->sq_dma_addr);
1360 	}
1361 }
1362 
1363 static void nvme_free_queues(struct nvme_dev *dev, int lowest)
1364 {
1365 	int i;
1366 
1367 	for (i = dev->ctrl.queue_count - 1; i >= lowest; i--) {
1368 		dev->ctrl.queue_count--;
1369 		nvme_free_queue(&dev->queues[i]);
1370 	}
1371 }
1372 
1373 /**
1374  * nvme_suspend_queue - put queue into suspended state
1375  * @nvmeq: queue to suspend
1376  */
1377 static int nvme_suspend_queue(struct nvme_queue *nvmeq)
1378 {
1379 	if (!test_and_clear_bit(NVMEQ_ENABLED, &nvmeq->flags))
1380 		return 1;
1381 
1382 	/* ensure that nvme_queue_rq() sees NVMEQ_ENABLED cleared */
1383 	mb();
1384 
1385 	nvmeq->dev->online_queues--;
1386 	if (!nvmeq->qid && nvmeq->dev->ctrl.admin_q)
1387 		blk_mq_quiesce_queue(nvmeq->dev->ctrl.admin_q);
1388 	if (!test_and_clear_bit(NVMEQ_POLLED, &nvmeq->flags))
1389 		pci_free_irq(to_pci_dev(nvmeq->dev->dev), nvmeq->cq_vector, nvmeq);
1390 	return 0;
1391 }
1392 
1393 static void nvme_suspend_io_queues(struct nvme_dev *dev)
1394 {
1395 	int i;
1396 
1397 	for (i = dev->ctrl.queue_count - 1; i > 0; i--)
1398 		nvme_suspend_queue(&dev->queues[i]);
1399 }
1400 
1401 static void nvme_disable_admin_queue(struct nvme_dev *dev, bool shutdown)
1402 {
1403 	struct nvme_queue *nvmeq = &dev->queues[0];
1404 
1405 	if (shutdown)
1406 		nvme_shutdown_ctrl(&dev->ctrl);
1407 	else
1408 		nvme_disable_ctrl(&dev->ctrl);
1409 
1410 	nvme_poll_irqdisable(nvmeq, -1);
1411 }
1412 
1413 static int nvme_cmb_qdepth(struct nvme_dev *dev, int nr_io_queues,
1414 				int entry_size)
1415 {
1416 	int q_depth = dev->q_depth;
1417 	unsigned q_size_aligned = roundup(q_depth * entry_size,
1418 					  dev->ctrl.page_size);
1419 
1420 	if (q_size_aligned * nr_io_queues > dev->cmb_size) {
1421 		u64 mem_per_q = div_u64(dev->cmb_size, nr_io_queues);
1422 		mem_per_q = round_down(mem_per_q, dev->ctrl.page_size);
1423 		q_depth = div_u64(mem_per_q, entry_size);
1424 
1425 		/*
1426 		 * Ensure the reduced q_depth is above some threshold where it
1427 		 * would be better to map queues in system memory with the
1428 		 * original depth
1429 		 */
1430 		if (q_depth < 64)
1431 			return -ENOMEM;
1432 	}
1433 
1434 	return q_depth;
1435 }
1436 
1437 static int nvme_alloc_sq_cmds(struct nvme_dev *dev, struct nvme_queue *nvmeq,
1438 				int qid)
1439 {
1440 	struct pci_dev *pdev = to_pci_dev(dev->dev);
1441 
1442 	if (qid && dev->cmb_use_sqes && (dev->cmbsz & NVME_CMBSZ_SQS)) {
1443 		nvmeq->sq_cmds = pci_alloc_p2pmem(pdev, SQ_SIZE(nvmeq));
1444 		if (nvmeq->sq_cmds) {
1445 			nvmeq->sq_dma_addr = pci_p2pmem_virt_to_bus(pdev,
1446 							nvmeq->sq_cmds);
1447 			if (nvmeq->sq_dma_addr) {
1448 				set_bit(NVMEQ_SQ_CMB, &nvmeq->flags);
1449 				return 0;
1450 			}
1451 
1452 			pci_free_p2pmem(pdev, nvmeq->sq_cmds, SQ_SIZE(nvmeq));
1453 		}
1454 	}
1455 
1456 	nvmeq->sq_cmds = dma_alloc_coherent(dev->dev, SQ_SIZE(nvmeq),
1457 				&nvmeq->sq_dma_addr, GFP_KERNEL);
1458 	if (!nvmeq->sq_cmds)
1459 		return -ENOMEM;
1460 	return 0;
1461 }
1462 
1463 static int nvme_alloc_queue(struct nvme_dev *dev, int qid, int depth)
1464 {
1465 	struct nvme_queue *nvmeq = &dev->queues[qid];
1466 
1467 	if (dev->ctrl.queue_count > qid)
1468 		return 0;
1469 
1470 	nvmeq->sqes = qid ? dev->io_sqes : NVME_ADM_SQES;
1471 	nvmeq->q_depth = depth;
1472 	nvmeq->cqes = dma_alloc_coherent(dev->dev, CQ_SIZE(nvmeq),
1473 					 &nvmeq->cq_dma_addr, GFP_KERNEL);
1474 	if (!nvmeq->cqes)
1475 		goto free_nvmeq;
1476 
1477 	if (nvme_alloc_sq_cmds(dev, nvmeq, qid))
1478 		goto free_cqdma;
1479 
1480 	nvmeq->dev = dev;
1481 	spin_lock_init(&nvmeq->sq_lock);
1482 	spin_lock_init(&nvmeq->cq_poll_lock);
1483 	nvmeq->cq_head = 0;
1484 	nvmeq->cq_phase = 1;
1485 	nvmeq->q_db = &dev->dbs[qid * 2 * dev->db_stride];
1486 	nvmeq->qid = qid;
1487 	dev->ctrl.queue_count++;
1488 
1489 	return 0;
1490 
1491  free_cqdma:
1492 	dma_free_coherent(dev->dev, CQ_SIZE(nvmeq), (void *)nvmeq->cqes,
1493 			  nvmeq->cq_dma_addr);
1494  free_nvmeq:
1495 	return -ENOMEM;
1496 }
1497 
1498 static int queue_request_irq(struct nvme_queue *nvmeq)
1499 {
1500 	struct pci_dev *pdev = to_pci_dev(nvmeq->dev->dev);
1501 	int nr = nvmeq->dev->ctrl.instance;
1502 
1503 	if (use_threaded_interrupts) {
1504 		return pci_request_irq(pdev, nvmeq->cq_vector, nvme_irq_check,
1505 				nvme_irq, nvmeq, "nvme%dq%d", nr, nvmeq->qid);
1506 	} else {
1507 		return pci_request_irq(pdev, nvmeq->cq_vector, nvme_irq,
1508 				NULL, nvmeq, "nvme%dq%d", nr, nvmeq->qid);
1509 	}
1510 }
1511 
1512 static void nvme_init_queue(struct nvme_queue *nvmeq, u16 qid)
1513 {
1514 	struct nvme_dev *dev = nvmeq->dev;
1515 
1516 	nvmeq->sq_tail = 0;
1517 	nvmeq->last_sq_tail = 0;
1518 	nvmeq->cq_head = 0;
1519 	nvmeq->cq_phase = 1;
1520 	nvmeq->q_db = &dev->dbs[qid * 2 * dev->db_stride];
1521 	memset((void *)nvmeq->cqes, 0, CQ_SIZE(nvmeq));
1522 	nvme_dbbuf_init(dev, nvmeq, qid);
1523 	dev->online_queues++;
1524 	wmb(); /* ensure the first interrupt sees the initialization */
1525 }
1526 
1527 static int nvme_create_queue(struct nvme_queue *nvmeq, int qid, bool polled)
1528 {
1529 	struct nvme_dev *dev = nvmeq->dev;
1530 	int result;
1531 	u16 vector = 0;
1532 
1533 	clear_bit(NVMEQ_DELETE_ERROR, &nvmeq->flags);
1534 
1535 	/*
1536 	 * A queue's vector matches the queue identifier unless the controller
1537 	 * has only one vector available.
1538 	 */
1539 	if (!polled)
1540 		vector = dev->num_vecs == 1 ? 0 : qid;
1541 	else
1542 		set_bit(NVMEQ_POLLED, &nvmeq->flags);
1543 
1544 	result = adapter_alloc_cq(dev, qid, nvmeq, vector);
1545 	if (result)
1546 		return result;
1547 
1548 	result = adapter_alloc_sq(dev, qid, nvmeq);
1549 	if (result < 0)
1550 		return result;
1551 	else if (result)
1552 		goto release_cq;
1553 
1554 	nvmeq->cq_vector = vector;
1555 	nvme_init_queue(nvmeq, qid);
1556 
1557 	if (!polled) {
1558 		result = queue_request_irq(nvmeq);
1559 		if (result < 0)
1560 			goto release_sq;
1561 	}
1562 
1563 	set_bit(NVMEQ_ENABLED, &nvmeq->flags);
1564 	return result;
1565 
1566 release_sq:
1567 	dev->online_queues--;
1568 	adapter_delete_sq(dev, qid);
1569 release_cq:
1570 	adapter_delete_cq(dev, qid);
1571 	return result;
1572 }
1573 
1574 static const struct blk_mq_ops nvme_mq_admin_ops = {
1575 	.queue_rq	= nvme_queue_rq,
1576 	.complete	= nvme_pci_complete_rq,
1577 	.init_hctx	= nvme_admin_init_hctx,
1578 	.exit_hctx      = nvme_admin_exit_hctx,
1579 	.init_request	= nvme_init_request,
1580 	.timeout	= nvme_timeout,
1581 };
1582 
1583 static const struct blk_mq_ops nvme_mq_ops = {
1584 	.queue_rq	= nvme_queue_rq,
1585 	.complete	= nvme_pci_complete_rq,
1586 	.commit_rqs	= nvme_commit_rqs,
1587 	.init_hctx	= nvme_init_hctx,
1588 	.init_request	= nvme_init_request,
1589 	.map_queues	= nvme_pci_map_queues,
1590 	.timeout	= nvme_timeout,
1591 	.poll		= nvme_poll,
1592 };
1593 
1594 static void nvme_dev_remove_admin(struct nvme_dev *dev)
1595 {
1596 	if (dev->ctrl.admin_q && !blk_queue_dying(dev->ctrl.admin_q)) {
1597 		/*
1598 		 * If the controller was reset during removal, it's possible
1599 		 * user requests may be waiting on a stopped queue. Start the
1600 		 * queue to flush these to completion.
1601 		 */
1602 		blk_mq_unquiesce_queue(dev->ctrl.admin_q);
1603 		blk_cleanup_queue(dev->ctrl.admin_q);
1604 		blk_mq_free_tag_set(&dev->admin_tagset);
1605 	}
1606 }
1607 
1608 static int nvme_alloc_admin_tags(struct nvme_dev *dev)
1609 {
1610 	if (!dev->ctrl.admin_q) {
1611 		dev->admin_tagset.ops = &nvme_mq_admin_ops;
1612 		dev->admin_tagset.nr_hw_queues = 1;
1613 
1614 		dev->admin_tagset.queue_depth = NVME_AQ_MQ_TAG_DEPTH;
1615 		dev->admin_tagset.timeout = ADMIN_TIMEOUT;
1616 		dev->admin_tagset.numa_node = dev_to_node(dev->dev);
1617 		dev->admin_tagset.cmd_size = sizeof(struct nvme_iod);
1618 		dev->admin_tagset.flags = BLK_MQ_F_NO_SCHED;
1619 		dev->admin_tagset.driver_data = dev;
1620 
1621 		if (blk_mq_alloc_tag_set(&dev->admin_tagset))
1622 			return -ENOMEM;
1623 		dev->ctrl.admin_tagset = &dev->admin_tagset;
1624 
1625 		dev->ctrl.admin_q = blk_mq_init_queue(&dev->admin_tagset);
1626 		if (IS_ERR(dev->ctrl.admin_q)) {
1627 			blk_mq_free_tag_set(&dev->admin_tagset);
1628 			return -ENOMEM;
1629 		}
1630 		if (!blk_get_queue(dev->ctrl.admin_q)) {
1631 			nvme_dev_remove_admin(dev);
1632 			dev->ctrl.admin_q = NULL;
1633 			return -ENODEV;
1634 		}
1635 	} else
1636 		blk_mq_unquiesce_queue(dev->ctrl.admin_q);
1637 
1638 	return 0;
1639 }
1640 
1641 static unsigned long db_bar_size(struct nvme_dev *dev, unsigned nr_io_queues)
1642 {
1643 	return NVME_REG_DBS + ((nr_io_queues + 1) * 8 * dev->db_stride);
1644 }
1645 
1646 static int nvme_remap_bar(struct nvme_dev *dev, unsigned long size)
1647 {
1648 	struct pci_dev *pdev = to_pci_dev(dev->dev);
1649 
1650 	if (size <= dev->bar_mapped_size)
1651 		return 0;
1652 	if (size > pci_resource_len(pdev, 0))
1653 		return -ENOMEM;
1654 	if (dev->bar)
1655 		iounmap(dev->bar);
1656 	dev->bar = ioremap(pci_resource_start(pdev, 0), size);
1657 	if (!dev->bar) {
1658 		dev->bar_mapped_size = 0;
1659 		return -ENOMEM;
1660 	}
1661 	dev->bar_mapped_size = size;
1662 	dev->dbs = dev->bar + NVME_REG_DBS;
1663 
1664 	return 0;
1665 }
1666 
1667 static int nvme_pci_configure_admin_queue(struct nvme_dev *dev)
1668 {
1669 	int result;
1670 	u32 aqa;
1671 	struct nvme_queue *nvmeq;
1672 
1673 	result = nvme_remap_bar(dev, db_bar_size(dev, 0));
1674 	if (result < 0)
1675 		return result;
1676 
1677 	dev->subsystem = readl(dev->bar + NVME_REG_VS) >= NVME_VS(1, 1, 0) ?
1678 				NVME_CAP_NSSRC(dev->ctrl.cap) : 0;
1679 
1680 	if (dev->subsystem &&
1681 	    (readl(dev->bar + NVME_REG_CSTS) & NVME_CSTS_NSSRO))
1682 		writel(NVME_CSTS_NSSRO, dev->bar + NVME_REG_CSTS);
1683 
1684 	result = nvme_disable_ctrl(&dev->ctrl);
1685 	if (result < 0)
1686 		return result;
1687 
1688 	result = nvme_alloc_queue(dev, 0, NVME_AQ_DEPTH);
1689 	if (result)
1690 		return result;
1691 
1692 	nvmeq = &dev->queues[0];
1693 	aqa = nvmeq->q_depth - 1;
1694 	aqa |= aqa << 16;
1695 
1696 	writel(aqa, dev->bar + NVME_REG_AQA);
1697 	lo_hi_writeq(nvmeq->sq_dma_addr, dev->bar + NVME_REG_ASQ);
1698 	lo_hi_writeq(nvmeq->cq_dma_addr, dev->bar + NVME_REG_ACQ);
1699 
1700 	result = nvme_enable_ctrl(&dev->ctrl);
1701 	if (result)
1702 		return result;
1703 
1704 	nvmeq->cq_vector = 0;
1705 	nvme_init_queue(nvmeq, 0);
1706 	result = queue_request_irq(nvmeq);
1707 	if (result) {
1708 		dev->online_queues--;
1709 		return result;
1710 	}
1711 
1712 	set_bit(NVMEQ_ENABLED, &nvmeq->flags);
1713 	return result;
1714 }
1715 
1716 static int nvme_create_io_queues(struct nvme_dev *dev)
1717 {
1718 	unsigned i, max, rw_queues;
1719 	int ret = 0;
1720 
1721 	for (i = dev->ctrl.queue_count; i <= dev->max_qid; i++) {
1722 		if (nvme_alloc_queue(dev, i, dev->q_depth)) {
1723 			ret = -ENOMEM;
1724 			break;
1725 		}
1726 	}
1727 
1728 	max = min(dev->max_qid, dev->ctrl.queue_count - 1);
1729 	if (max != 1 && dev->io_queues[HCTX_TYPE_POLL]) {
1730 		rw_queues = dev->io_queues[HCTX_TYPE_DEFAULT] +
1731 				dev->io_queues[HCTX_TYPE_READ];
1732 	} else {
1733 		rw_queues = max;
1734 	}
1735 
1736 	for (i = dev->online_queues; i <= max; i++) {
1737 		bool polled = i > rw_queues;
1738 
1739 		ret = nvme_create_queue(&dev->queues[i], i, polled);
1740 		if (ret)
1741 			break;
1742 	}
1743 
1744 	/*
1745 	 * Ignore failing Create SQ/CQ commands, we can continue with less
1746 	 * than the desired amount of queues, and even a controller without
1747 	 * I/O queues can still be used to issue admin commands.  This might
1748 	 * be useful to upgrade a buggy firmware for example.
1749 	 */
1750 	return ret >= 0 ? 0 : ret;
1751 }
1752 
1753 static ssize_t nvme_cmb_show(struct device *dev,
1754 			     struct device_attribute *attr,
1755 			     char *buf)
1756 {
1757 	struct nvme_dev *ndev = to_nvme_dev(dev_get_drvdata(dev));
1758 
1759 	return scnprintf(buf, PAGE_SIZE, "cmbloc : x%08x\ncmbsz  : x%08x\n",
1760 		       ndev->cmbloc, ndev->cmbsz);
1761 }
1762 static DEVICE_ATTR(cmb, S_IRUGO, nvme_cmb_show, NULL);
1763 
1764 static u64 nvme_cmb_size_unit(struct nvme_dev *dev)
1765 {
1766 	u8 szu = (dev->cmbsz >> NVME_CMBSZ_SZU_SHIFT) & NVME_CMBSZ_SZU_MASK;
1767 
1768 	return 1ULL << (12 + 4 * szu);
1769 }
1770 
1771 static u32 nvme_cmb_size(struct nvme_dev *dev)
1772 {
1773 	return (dev->cmbsz >> NVME_CMBSZ_SZ_SHIFT) & NVME_CMBSZ_SZ_MASK;
1774 }
1775 
1776 static void nvme_map_cmb(struct nvme_dev *dev)
1777 {
1778 	u64 size, offset;
1779 	resource_size_t bar_size;
1780 	struct pci_dev *pdev = to_pci_dev(dev->dev);
1781 	int bar;
1782 
1783 	if (dev->cmb_size)
1784 		return;
1785 
1786 	dev->cmbsz = readl(dev->bar + NVME_REG_CMBSZ);
1787 	if (!dev->cmbsz)
1788 		return;
1789 	dev->cmbloc = readl(dev->bar + NVME_REG_CMBLOC);
1790 
1791 	size = nvme_cmb_size_unit(dev) * nvme_cmb_size(dev);
1792 	offset = nvme_cmb_size_unit(dev) * NVME_CMB_OFST(dev->cmbloc);
1793 	bar = NVME_CMB_BIR(dev->cmbloc);
1794 	bar_size = pci_resource_len(pdev, bar);
1795 
1796 	if (offset > bar_size)
1797 		return;
1798 
1799 	/*
1800 	 * Controllers may support a CMB size larger than their BAR,
1801 	 * for example, due to being behind a bridge. Reduce the CMB to
1802 	 * the reported size of the BAR
1803 	 */
1804 	if (size > bar_size - offset)
1805 		size = bar_size - offset;
1806 
1807 	if (pci_p2pdma_add_resource(pdev, bar, size, offset)) {
1808 		dev_warn(dev->ctrl.device,
1809 			 "failed to register the CMB\n");
1810 		return;
1811 	}
1812 
1813 	dev->cmb_size = size;
1814 	dev->cmb_use_sqes = use_cmb_sqes && (dev->cmbsz & NVME_CMBSZ_SQS);
1815 
1816 	if ((dev->cmbsz & (NVME_CMBSZ_WDS | NVME_CMBSZ_RDS)) ==
1817 			(NVME_CMBSZ_WDS | NVME_CMBSZ_RDS))
1818 		pci_p2pmem_publish(pdev, true);
1819 
1820 	if (sysfs_add_file_to_group(&dev->ctrl.device->kobj,
1821 				    &dev_attr_cmb.attr, NULL))
1822 		dev_warn(dev->ctrl.device,
1823 			 "failed to add sysfs attribute for CMB\n");
1824 }
1825 
1826 static inline void nvme_release_cmb(struct nvme_dev *dev)
1827 {
1828 	if (dev->cmb_size) {
1829 		sysfs_remove_file_from_group(&dev->ctrl.device->kobj,
1830 					     &dev_attr_cmb.attr, NULL);
1831 		dev->cmb_size = 0;
1832 	}
1833 }
1834 
1835 static int nvme_set_host_mem(struct nvme_dev *dev, u32 bits)
1836 {
1837 	u64 dma_addr = dev->host_mem_descs_dma;
1838 	struct nvme_command c;
1839 	int ret;
1840 
1841 	memset(&c, 0, sizeof(c));
1842 	c.features.opcode	= nvme_admin_set_features;
1843 	c.features.fid		= cpu_to_le32(NVME_FEAT_HOST_MEM_BUF);
1844 	c.features.dword11	= cpu_to_le32(bits);
1845 	c.features.dword12	= cpu_to_le32(dev->host_mem_size >>
1846 					      ilog2(dev->ctrl.page_size));
1847 	c.features.dword13	= cpu_to_le32(lower_32_bits(dma_addr));
1848 	c.features.dword14	= cpu_to_le32(upper_32_bits(dma_addr));
1849 	c.features.dword15	= cpu_to_le32(dev->nr_host_mem_descs);
1850 
1851 	ret = nvme_submit_sync_cmd(dev->ctrl.admin_q, &c, NULL, 0);
1852 	if (ret) {
1853 		dev_warn(dev->ctrl.device,
1854 			 "failed to set host mem (err %d, flags %#x).\n",
1855 			 ret, bits);
1856 	}
1857 	return ret;
1858 }
1859 
1860 static void nvme_free_host_mem(struct nvme_dev *dev)
1861 {
1862 	int i;
1863 
1864 	for (i = 0; i < dev->nr_host_mem_descs; i++) {
1865 		struct nvme_host_mem_buf_desc *desc = &dev->host_mem_descs[i];
1866 		size_t size = le32_to_cpu(desc->size) * dev->ctrl.page_size;
1867 
1868 		dma_free_attrs(dev->dev, size, dev->host_mem_desc_bufs[i],
1869 			       le64_to_cpu(desc->addr),
1870 			       DMA_ATTR_NO_KERNEL_MAPPING | DMA_ATTR_NO_WARN);
1871 	}
1872 
1873 	kfree(dev->host_mem_desc_bufs);
1874 	dev->host_mem_desc_bufs = NULL;
1875 	dma_free_coherent(dev->dev,
1876 			dev->nr_host_mem_descs * sizeof(*dev->host_mem_descs),
1877 			dev->host_mem_descs, dev->host_mem_descs_dma);
1878 	dev->host_mem_descs = NULL;
1879 	dev->nr_host_mem_descs = 0;
1880 }
1881 
1882 static int __nvme_alloc_host_mem(struct nvme_dev *dev, u64 preferred,
1883 		u32 chunk_size)
1884 {
1885 	struct nvme_host_mem_buf_desc *descs;
1886 	u32 max_entries, len;
1887 	dma_addr_t descs_dma;
1888 	int i = 0;
1889 	void **bufs;
1890 	u64 size, tmp;
1891 
1892 	tmp = (preferred + chunk_size - 1);
1893 	do_div(tmp, chunk_size);
1894 	max_entries = tmp;
1895 
1896 	if (dev->ctrl.hmmaxd && dev->ctrl.hmmaxd < max_entries)
1897 		max_entries = dev->ctrl.hmmaxd;
1898 
1899 	descs = dma_alloc_coherent(dev->dev, max_entries * sizeof(*descs),
1900 				   &descs_dma, GFP_KERNEL);
1901 	if (!descs)
1902 		goto out;
1903 
1904 	bufs = kcalloc(max_entries, sizeof(*bufs), GFP_KERNEL);
1905 	if (!bufs)
1906 		goto out_free_descs;
1907 
1908 	for (size = 0; size < preferred && i < max_entries; size += len) {
1909 		dma_addr_t dma_addr;
1910 
1911 		len = min_t(u64, chunk_size, preferred - size);
1912 		bufs[i] = dma_alloc_attrs(dev->dev, len, &dma_addr, GFP_KERNEL,
1913 				DMA_ATTR_NO_KERNEL_MAPPING | DMA_ATTR_NO_WARN);
1914 		if (!bufs[i])
1915 			break;
1916 
1917 		descs[i].addr = cpu_to_le64(dma_addr);
1918 		descs[i].size = cpu_to_le32(len / dev->ctrl.page_size);
1919 		i++;
1920 	}
1921 
1922 	if (!size)
1923 		goto out_free_bufs;
1924 
1925 	dev->nr_host_mem_descs = i;
1926 	dev->host_mem_size = size;
1927 	dev->host_mem_descs = descs;
1928 	dev->host_mem_descs_dma = descs_dma;
1929 	dev->host_mem_desc_bufs = bufs;
1930 	return 0;
1931 
1932 out_free_bufs:
1933 	while (--i >= 0) {
1934 		size_t size = le32_to_cpu(descs[i].size) * dev->ctrl.page_size;
1935 
1936 		dma_free_attrs(dev->dev, size, bufs[i],
1937 			       le64_to_cpu(descs[i].addr),
1938 			       DMA_ATTR_NO_KERNEL_MAPPING | DMA_ATTR_NO_WARN);
1939 	}
1940 
1941 	kfree(bufs);
1942 out_free_descs:
1943 	dma_free_coherent(dev->dev, max_entries * sizeof(*descs), descs,
1944 			descs_dma);
1945 out:
1946 	dev->host_mem_descs = NULL;
1947 	return -ENOMEM;
1948 }
1949 
1950 static int nvme_alloc_host_mem(struct nvme_dev *dev, u64 min, u64 preferred)
1951 {
1952 	u32 chunk_size;
1953 
1954 	/* start big and work our way down */
1955 	for (chunk_size = min_t(u64, preferred, PAGE_SIZE * MAX_ORDER_NR_PAGES);
1956 	     chunk_size >= max_t(u32, dev->ctrl.hmminds * 4096, PAGE_SIZE * 2);
1957 	     chunk_size /= 2) {
1958 		if (!__nvme_alloc_host_mem(dev, preferred, chunk_size)) {
1959 			if (!min || dev->host_mem_size >= min)
1960 				return 0;
1961 			nvme_free_host_mem(dev);
1962 		}
1963 	}
1964 
1965 	return -ENOMEM;
1966 }
1967 
1968 static int nvme_setup_host_mem(struct nvme_dev *dev)
1969 {
1970 	u64 max = (u64)max_host_mem_size_mb * SZ_1M;
1971 	u64 preferred = (u64)dev->ctrl.hmpre * 4096;
1972 	u64 min = (u64)dev->ctrl.hmmin * 4096;
1973 	u32 enable_bits = NVME_HOST_MEM_ENABLE;
1974 	int ret;
1975 
1976 	preferred = min(preferred, max);
1977 	if (min > max) {
1978 		dev_warn(dev->ctrl.device,
1979 			"min host memory (%lld MiB) above limit (%d MiB).\n",
1980 			min >> ilog2(SZ_1M), max_host_mem_size_mb);
1981 		nvme_free_host_mem(dev);
1982 		return 0;
1983 	}
1984 
1985 	/*
1986 	 * If we already have a buffer allocated check if we can reuse it.
1987 	 */
1988 	if (dev->host_mem_descs) {
1989 		if (dev->host_mem_size >= min)
1990 			enable_bits |= NVME_HOST_MEM_RETURN;
1991 		else
1992 			nvme_free_host_mem(dev);
1993 	}
1994 
1995 	if (!dev->host_mem_descs) {
1996 		if (nvme_alloc_host_mem(dev, min, preferred)) {
1997 			dev_warn(dev->ctrl.device,
1998 				"failed to allocate host memory buffer.\n");
1999 			return 0; /* controller must work without HMB */
2000 		}
2001 
2002 		dev_info(dev->ctrl.device,
2003 			"allocated %lld MiB host memory buffer.\n",
2004 			dev->host_mem_size >> ilog2(SZ_1M));
2005 	}
2006 
2007 	ret = nvme_set_host_mem(dev, enable_bits);
2008 	if (ret)
2009 		nvme_free_host_mem(dev);
2010 	return ret;
2011 }
2012 
2013 /*
2014  * nirqs is the number of interrupts available for write and read
2015  * queues. The core already reserved an interrupt for the admin queue.
2016  */
2017 static void nvme_calc_irq_sets(struct irq_affinity *affd, unsigned int nrirqs)
2018 {
2019 	struct nvme_dev *dev = affd->priv;
2020 	unsigned int nr_read_queues;
2021 
2022 	/*
2023 	 * If there is no interupt available for queues, ensure that
2024 	 * the default queue is set to 1. The affinity set size is
2025 	 * also set to one, but the irq core ignores it for this case.
2026 	 *
2027 	 * If only one interrupt is available or 'write_queue' == 0, combine
2028 	 * write and read queues.
2029 	 *
2030 	 * If 'write_queues' > 0, ensure it leaves room for at least one read
2031 	 * queue.
2032 	 */
2033 	if (!nrirqs) {
2034 		nrirqs = 1;
2035 		nr_read_queues = 0;
2036 	} else if (nrirqs == 1 || !write_queues) {
2037 		nr_read_queues = 0;
2038 	} else if (write_queues >= nrirqs) {
2039 		nr_read_queues = 1;
2040 	} else {
2041 		nr_read_queues = nrirqs - write_queues;
2042 	}
2043 
2044 	dev->io_queues[HCTX_TYPE_DEFAULT] = nrirqs - nr_read_queues;
2045 	affd->set_size[HCTX_TYPE_DEFAULT] = nrirqs - nr_read_queues;
2046 	dev->io_queues[HCTX_TYPE_READ] = nr_read_queues;
2047 	affd->set_size[HCTX_TYPE_READ] = nr_read_queues;
2048 	affd->nr_sets = nr_read_queues ? 2 : 1;
2049 }
2050 
2051 static int nvme_setup_irqs(struct nvme_dev *dev, unsigned int nr_io_queues)
2052 {
2053 	struct pci_dev *pdev = to_pci_dev(dev->dev);
2054 	struct irq_affinity affd = {
2055 		.pre_vectors	= 1,
2056 		.calc_sets	= nvme_calc_irq_sets,
2057 		.priv		= dev,
2058 	};
2059 	unsigned int irq_queues, this_p_queues;
2060 	unsigned int nr_cpus = num_possible_cpus();
2061 
2062 	/*
2063 	 * Poll queues don't need interrupts, but we need at least one IO
2064 	 * queue left over for non-polled IO.
2065 	 */
2066 	this_p_queues = poll_queues;
2067 	if (this_p_queues >= nr_io_queues) {
2068 		this_p_queues = nr_io_queues - 1;
2069 		irq_queues = 1;
2070 	} else {
2071 		if (nr_cpus < nr_io_queues - this_p_queues)
2072 			irq_queues = nr_cpus + 1;
2073 		else
2074 			irq_queues = nr_io_queues - this_p_queues + 1;
2075 	}
2076 	dev->io_queues[HCTX_TYPE_POLL] = this_p_queues;
2077 
2078 	/* Initialize for the single interrupt case */
2079 	dev->io_queues[HCTX_TYPE_DEFAULT] = 1;
2080 	dev->io_queues[HCTX_TYPE_READ] = 0;
2081 
2082 	/*
2083 	 * Some Apple controllers require all queues to use the
2084 	 * first vector.
2085 	 */
2086 	if (dev->ctrl.quirks & NVME_QUIRK_SINGLE_VECTOR)
2087 		irq_queues = 1;
2088 
2089 	return pci_alloc_irq_vectors_affinity(pdev, 1, irq_queues,
2090 			      PCI_IRQ_ALL_TYPES | PCI_IRQ_AFFINITY, &affd);
2091 }
2092 
2093 static void nvme_disable_io_queues(struct nvme_dev *dev)
2094 {
2095 	if (__nvme_disable_io_queues(dev, nvme_admin_delete_sq))
2096 		__nvme_disable_io_queues(dev, nvme_admin_delete_cq);
2097 }
2098 
2099 static int nvme_setup_io_queues(struct nvme_dev *dev)
2100 {
2101 	struct nvme_queue *adminq = &dev->queues[0];
2102 	struct pci_dev *pdev = to_pci_dev(dev->dev);
2103 	int result, nr_io_queues;
2104 	unsigned long size;
2105 
2106 	nr_io_queues = max_io_queues();
2107 
2108 	/*
2109 	 * If tags are shared with admin queue (Apple bug), then
2110 	 * make sure we only use one IO queue.
2111 	 */
2112 	if (dev->ctrl.quirks & NVME_QUIRK_SHARED_TAGS)
2113 		nr_io_queues = 1;
2114 
2115 	result = nvme_set_queue_count(&dev->ctrl, &nr_io_queues);
2116 	if (result < 0)
2117 		return result;
2118 
2119 	if (nr_io_queues == 0)
2120 		return 0;
2121 
2122 	clear_bit(NVMEQ_ENABLED, &adminq->flags);
2123 
2124 	if (dev->cmb_use_sqes) {
2125 		result = nvme_cmb_qdepth(dev, nr_io_queues,
2126 				sizeof(struct nvme_command));
2127 		if (result > 0)
2128 			dev->q_depth = result;
2129 		else
2130 			dev->cmb_use_sqes = false;
2131 	}
2132 
2133 	do {
2134 		size = db_bar_size(dev, nr_io_queues);
2135 		result = nvme_remap_bar(dev, size);
2136 		if (!result)
2137 			break;
2138 		if (!--nr_io_queues)
2139 			return -ENOMEM;
2140 	} while (1);
2141 	adminq->q_db = dev->dbs;
2142 
2143  retry:
2144 	/* Deregister the admin queue's interrupt */
2145 	pci_free_irq(pdev, 0, adminq);
2146 
2147 	/*
2148 	 * If we enable msix early due to not intx, disable it again before
2149 	 * setting up the full range we need.
2150 	 */
2151 	pci_free_irq_vectors(pdev);
2152 
2153 	result = nvme_setup_irqs(dev, nr_io_queues);
2154 	if (result <= 0)
2155 		return -EIO;
2156 
2157 	dev->num_vecs = result;
2158 	result = max(result - 1, 1);
2159 	dev->max_qid = result + dev->io_queues[HCTX_TYPE_POLL];
2160 
2161 	/*
2162 	 * Should investigate if there's a performance win from allocating
2163 	 * more queues than interrupt vectors; it might allow the submission
2164 	 * path to scale better, even if the receive path is limited by the
2165 	 * number of interrupts.
2166 	 */
2167 	result = queue_request_irq(adminq);
2168 	if (result)
2169 		return result;
2170 	set_bit(NVMEQ_ENABLED, &adminq->flags);
2171 
2172 	result = nvme_create_io_queues(dev);
2173 	if (result || dev->online_queues < 2)
2174 		return result;
2175 
2176 	if (dev->online_queues - 1 < dev->max_qid) {
2177 		nr_io_queues = dev->online_queues - 1;
2178 		nvme_disable_io_queues(dev);
2179 		nvme_suspend_io_queues(dev);
2180 		goto retry;
2181 	}
2182 	dev_info(dev->ctrl.device, "%d/%d/%d default/read/poll queues\n",
2183 					dev->io_queues[HCTX_TYPE_DEFAULT],
2184 					dev->io_queues[HCTX_TYPE_READ],
2185 					dev->io_queues[HCTX_TYPE_POLL]);
2186 	return 0;
2187 }
2188 
2189 static void nvme_del_queue_end(struct request *req, blk_status_t error)
2190 {
2191 	struct nvme_queue *nvmeq = req->end_io_data;
2192 
2193 	blk_mq_free_request(req);
2194 	complete(&nvmeq->delete_done);
2195 }
2196 
2197 static void nvme_del_cq_end(struct request *req, blk_status_t error)
2198 {
2199 	struct nvme_queue *nvmeq = req->end_io_data;
2200 
2201 	if (error)
2202 		set_bit(NVMEQ_DELETE_ERROR, &nvmeq->flags);
2203 
2204 	nvme_del_queue_end(req, error);
2205 }
2206 
2207 static int nvme_delete_queue(struct nvme_queue *nvmeq, u8 opcode)
2208 {
2209 	struct request_queue *q = nvmeq->dev->ctrl.admin_q;
2210 	struct request *req;
2211 	struct nvme_command cmd;
2212 
2213 	memset(&cmd, 0, sizeof(cmd));
2214 	cmd.delete_queue.opcode = opcode;
2215 	cmd.delete_queue.qid = cpu_to_le16(nvmeq->qid);
2216 
2217 	req = nvme_alloc_request(q, &cmd, BLK_MQ_REQ_NOWAIT, NVME_QID_ANY);
2218 	if (IS_ERR(req))
2219 		return PTR_ERR(req);
2220 
2221 	req->timeout = ADMIN_TIMEOUT;
2222 	req->end_io_data = nvmeq;
2223 
2224 	init_completion(&nvmeq->delete_done);
2225 	blk_execute_rq_nowait(q, NULL, req, false,
2226 			opcode == nvme_admin_delete_cq ?
2227 				nvme_del_cq_end : nvme_del_queue_end);
2228 	return 0;
2229 }
2230 
2231 static bool __nvme_disable_io_queues(struct nvme_dev *dev, u8 opcode)
2232 {
2233 	int nr_queues = dev->online_queues - 1, sent = 0;
2234 	unsigned long timeout;
2235 
2236  retry:
2237 	timeout = ADMIN_TIMEOUT;
2238 	while (nr_queues > 0) {
2239 		if (nvme_delete_queue(&dev->queues[nr_queues], opcode))
2240 			break;
2241 		nr_queues--;
2242 		sent++;
2243 	}
2244 	while (sent) {
2245 		struct nvme_queue *nvmeq = &dev->queues[nr_queues + sent];
2246 
2247 		timeout = wait_for_completion_io_timeout(&nvmeq->delete_done,
2248 				timeout);
2249 		if (timeout == 0)
2250 			return false;
2251 
2252 		/* handle any remaining CQEs */
2253 		if (opcode == nvme_admin_delete_cq &&
2254 		    !test_bit(NVMEQ_DELETE_ERROR, &nvmeq->flags))
2255 			nvme_poll_irqdisable(nvmeq, -1);
2256 
2257 		sent--;
2258 		if (nr_queues)
2259 			goto retry;
2260 	}
2261 	return true;
2262 }
2263 
2264 /*
2265  * return error value only when tagset allocation failed
2266  */
2267 static int nvme_dev_add(struct nvme_dev *dev)
2268 {
2269 	int ret;
2270 
2271 	if (!dev->ctrl.tagset) {
2272 		dev->tagset.ops = &nvme_mq_ops;
2273 		dev->tagset.nr_hw_queues = dev->online_queues - 1;
2274 		dev->tagset.nr_maps = 2; /* default + read */
2275 		if (dev->io_queues[HCTX_TYPE_POLL])
2276 			dev->tagset.nr_maps++;
2277 		dev->tagset.timeout = NVME_IO_TIMEOUT;
2278 		dev->tagset.numa_node = dev_to_node(dev->dev);
2279 		dev->tagset.queue_depth =
2280 				min_t(int, dev->q_depth, BLK_MQ_MAX_DEPTH) - 1;
2281 		dev->tagset.cmd_size = sizeof(struct nvme_iod);
2282 		dev->tagset.flags = BLK_MQ_F_SHOULD_MERGE;
2283 		dev->tagset.driver_data = dev;
2284 
2285 		/*
2286 		 * Some Apple controllers requires tags to be unique
2287 		 * across admin and IO queue, so reserve the first 32
2288 		 * tags of the IO queue.
2289 		 */
2290 		if (dev->ctrl.quirks & NVME_QUIRK_SHARED_TAGS)
2291 			dev->tagset.reserved_tags = NVME_AQ_DEPTH;
2292 
2293 		ret = blk_mq_alloc_tag_set(&dev->tagset);
2294 		if (ret) {
2295 			dev_warn(dev->ctrl.device,
2296 				"IO queues tagset allocation failed %d\n", ret);
2297 			return ret;
2298 		}
2299 		dev->ctrl.tagset = &dev->tagset;
2300 	} else {
2301 		blk_mq_update_nr_hw_queues(&dev->tagset, dev->online_queues - 1);
2302 
2303 		/* Free previously allocated queues that are no longer usable */
2304 		nvme_free_queues(dev, dev->online_queues);
2305 	}
2306 
2307 	nvme_dbbuf_set(dev);
2308 	return 0;
2309 }
2310 
2311 static int nvme_pci_enable(struct nvme_dev *dev)
2312 {
2313 	int result = -ENOMEM;
2314 	struct pci_dev *pdev = to_pci_dev(dev->dev);
2315 
2316 	if (pci_enable_device_mem(pdev))
2317 		return result;
2318 
2319 	pci_set_master(pdev);
2320 
2321 	if (dma_set_mask_and_coherent(dev->dev, DMA_BIT_MASK(64)))
2322 		goto disable;
2323 
2324 	if (readl(dev->bar + NVME_REG_CSTS) == -1) {
2325 		result = -ENODEV;
2326 		goto disable;
2327 	}
2328 
2329 	/*
2330 	 * Some devices and/or platforms don't advertise or work with INTx
2331 	 * interrupts. Pre-enable a single MSIX or MSI vec for setup. We'll
2332 	 * adjust this later.
2333 	 */
2334 	result = pci_alloc_irq_vectors(pdev, 1, 1, PCI_IRQ_ALL_TYPES);
2335 	if (result < 0)
2336 		return result;
2337 
2338 	dev->ctrl.cap = lo_hi_readq(dev->bar + NVME_REG_CAP);
2339 
2340 	dev->q_depth = min_t(int, NVME_CAP_MQES(dev->ctrl.cap) + 1,
2341 				io_queue_depth);
2342 	dev->ctrl.sqsize = dev->q_depth - 1; /* 0's based queue depth */
2343 	dev->db_stride = 1 << NVME_CAP_STRIDE(dev->ctrl.cap);
2344 	dev->dbs = dev->bar + 4096;
2345 
2346 	/*
2347 	 * Some Apple controllers require a non-standard SQE size.
2348 	 * Interestingly they also seem to ignore the CC:IOSQES register
2349 	 * so we don't bother updating it here.
2350 	 */
2351 	if (dev->ctrl.quirks & NVME_QUIRK_128_BYTES_SQES)
2352 		dev->io_sqes = 7;
2353 	else
2354 		dev->io_sqes = NVME_NVM_IOSQES;
2355 
2356 	/*
2357 	 * Temporary fix for the Apple controller found in the MacBook8,1 and
2358 	 * some MacBook7,1 to avoid controller resets and data loss.
2359 	 */
2360 	if (pdev->vendor == PCI_VENDOR_ID_APPLE && pdev->device == 0x2001) {
2361 		dev->q_depth = 2;
2362 		dev_warn(dev->ctrl.device, "detected Apple NVMe controller, "
2363 			"set queue depth=%u to work around controller resets\n",
2364 			dev->q_depth);
2365 	} else if (pdev->vendor == PCI_VENDOR_ID_SAMSUNG &&
2366 		   (pdev->device == 0xa821 || pdev->device == 0xa822) &&
2367 		   NVME_CAP_MQES(dev->ctrl.cap) == 0) {
2368 		dev->q_depth = 64;
2369 		dev_err(dev->ctrl.device, "detected PM1725 NVMe controller, "
2370                         "set queue depth=%u\n", dev->q_depth);
2371 	}
2372 
2373 	/*
2374 	 * Controllers with the shared tags quirk need the IO queue to be
2375 	 * big enough so that we get 32 tags for the admin queue
2376 	 */
2377 	if ((dev->ctrl.quirks & NVME_QUIRK_SHARED_TAGS) &&
2378 	    (dev->q_depth < (NVME_AQ_DEPTH + 2))) {
2379 		dev->q_depth = NVME_AQ_DEPTH + 2;
2380 		dev_warn(dev->ctrl.device, "IO queue depth clamped to %d\n",
2381 			 dev->q_depth);
2382 	}
2383 
2384 
2385 	nvme_map_cmb(dev);
2386 
2387 	pci_enable_pcie_error_reporting(pdev);
2388 	pci_save_state(pdev);
2389 	return 0;
2390 
2391  disable:
2392 	pci_disable_device(pdev);
2393 	return result;
2394 }
2395 
2396 static void nvme_dev_unmap(struct nvme_dev *dev)
2397 {
2398 	if (dev->bar)
2399 		iounmap(dev->bar);
2400 	pci_release_mem_regions(to_pci_dev(dev->dev));
2401 }
2402 
2403 static void nvme_pci_disable(struct nvme_dev *dev)
2404 {
2405 	struct pci_dev *pdev = to_pci_dev(dev->dev);
2406 
2407 	pci_free_irq_vectors(pdev);
2408 
2409 	if (pci_is_enabled(pdev)) {
2410 		pci_disable_pcie_error_reporting(pdev);
2411 		pci_disable_device(pdev);
2412 	}
2413 }
2414 
2415 static void nvme_dev_disable(struct nvme_dev *dev, bool shutdown)
2416 {
2417 	bool dead = true, freeze = false;
2418 	struct pci_dev *pdev = to_pci_dev(dev->dev);
2419 
2420 	mutex_lock(&dev->shutdown_lock);
2421 	if (pci_is_enabled(pdev)) {
2422 		u32 csts = readl(dev->bar + NVME_REG_CSTS);
2423 
2424 		if (dev->ctrl.state == NVME_CTRL_LIVE ||
2425 		    dev->ctrl.state == NVME_CTRL_RESETTING) {
2426 			freeze = true;
2427 			nvme_start_freeze(&dev->ctrl);
2428 		}
2429 		dead = !!((csts & NVME_CSTS_CFS) || !(csts & NVME_CSTS_RDY) ||
2430 			pdev->error_state  != pci_channel_io_normal);
2431 	}
2432 
2433 	/*
2434 	 * Give the controller a chance to complete all entered requests if
2435 	 * doing a safe shutdown.
2436 	 */
2437 	if (!dead && shutdown && freeze)
2438 		nvme_wait_freeze_timeout(&dev->ctrl, NVME_IO_TIMEOUT);
2439 
2440 	nvme_stop_queues(&dev->ctrl);
2441 
2442 	if (!dead && dev->ctrl.queue_count > 0) {
2443 		nvme_disable_io_queues(dev);
2444 		nvme_disable_admin_queue(dev, shutdown);
2445 	}
2446 	nvme_suspend_io_queues(dev);
2447 	nvme_suspend_queue(&dev->queues[0]);
2448 	nvme_pci_disable(dev);
2449 
2450 	blk_mq_tagset_busy_iter(&dev->tagset, nvme_cancel_request, &dev->ctrl);
2451 	blk_mq_tagset_busy_iter(&dev->admin_tagset, nvme_cancel_request, &dev->ctrl);
2452 	blk_mq_tagset_wait_completed_request(&dev->tagset);
2453 	blk_mq_tagset_wait_completed_request(&dev->admin_tagset);
2454 
2455 	/*
2456 	 * The driver will not be starting up queues again if shutting down so
2457 	 * must flush all entered requests to their failed completion to avoid
2458 	 * deadlocking blk-mq hot-cpu notifier.
2459 	 */
2460 	if (shutdown) {
2461 		nvme_start_queues(&dev->ctrl);
2462 		if (dev->ctrl.admin_q && !blk_queue_dying(dev->ctrl.admin_q))
2463 			blk_mq_unquiesce_queue(dev->ctrl.admin_q);
2464 	}
2465 	mutex_unlock(&dev->shutdown_lock);
2466 }
2467 
2468 static int nvme_setup_prp_pools(struct nvme_dev *dev)
2469 {
2470 	dev->prp_page_pool = dma_pool_create("prp list page", dev->dev,
2471 						PAGE_SIZE, PAGE_SIZE, 0);
2472 	if (!dev->prp_page_pool)
2473 		return -ENOMEM;
2474 
2475 	/* Optimisation for I/Os between 4k and 128k */
2476 	dev->prp_small_pool = dma_pool_create("prp list 256", dev->dev,
2477 						256, 256, 0);
2478 	if (!dev->prp_small_pool) {
2479 		dma_pool_destroy(dev->prp_page_pool);
2480 		return -ENOMEM;
2481 	}
2482 	return 0;
2483 }
2484 
2485 static void nvme_release_prp_pools(struct nvme_dev *dev)
2486 {
2487 	dma_pool_destroy(dev->prp_page_pool);
2488 	dma_pool_destroy(dev->prp_small_pool);
2489 }
2490 
2491 static void nvme_pci_free_ctrl(struct nvme_ctrl *ctrl)
2492 {
2493 	struct nvme_dev *dev = to_nvme_dev(ctrl);
2494 
2495 	nvme_dbbuf_dma_free(dev);
2496 	put_device(dev->dev);
2497 	if (dev->tagset.tags)
2498 		blk_mq_free_tag_set(&dev->tagset);
2499 	if (dev->ctrl.admin_q)
2500 		blk_put_queue(dev->ctrl.admin_q);
2501 	kfree(dev->queues);
2502 	free_opal_dev(dev->ctrl.opal_dev);
2503 	mempool_destroy(dev->iod_mempool);
2504 	kfree(dev);
2505 }
2506 
2507 static void nvme_remove_dead_ctrl(struct nvme_dev *dev)
2508 {
2509 	nvme_get_ctrl(&dev->ctrl);
2510 	nvme_dev_disable(dev, false);
2511 	nvme_kill_queues(&dev->ctrl);
2512 	if (!queue_work(nvme_wq, &dev->remove_work))
2513 		nvme_put_ctrl(&dev->ctrl);
2514 }
2515 
2516 static void nvme_reset_work(struct work_struct *work)
2517 {
2518 	struct nvme_dev *dev =
2519 		container_of(work, struct nvme_dev, ctrl.reset_work);
2520 	bool was_suspend = !!(dev->ctrl.ctrl_config & NVME_CC_SHN_NORMAL);
2521 	int result;
2522 	enum nvme_ctrl_state new_state = NVME_CTRL_LIVE;
2523 
2524 	if (WARN_ON(dev->ctrl.state != NVME_CTRL_RESETTING)) {
2525 		result = -ENODEV;
2526 		goto out;
2527 	}
2528 
2529 	/*
2530 	 * If we're called to reset a live controller first shut it down before
2531 	 * moving on.
2532 	 */
2533 	if (dev->ctrl.ctrl_config & NVME_CC_ENABLE)
2534 		nvme_dev_disable(dev, false);
2535 	nvme_sync_queues(&dev->ctrl);
2536 
2537 	mutex_lock(&dev->shutdown_lock);
2538 	result = nvme_pci_enable(dev);
2539 	if (result)
2540 		goto out_unlock;
2541 
2542 	result = nvme_pci_configure_admin_queue(dev);
2543 	if (result)
2544 		goto out_unlock;
2545 
2546 	result = nvme_alloc_admin_tags(dev);
2547 	if (result)
2548 		goto out_unlock;
2549 
2550 	/*
2551 	 * Limit the max command size to prevent iod->sg allocations going
2552 	 * over a single page.
2553 	 */
2554 	dev->ctrl.max_hw_sectors = min_t(u32,
2555 		NVME_MAX_KB_SZ << 1, dma_max_mapping_size(dev->dev) >> 9);
2556 	dev->ctrl.max_segments = NVME_MAX_SEGS;
2557 
2558 	/*
2559 	 * Don't limit the IOMMU merged segment size.
2560 	 */
2561 	dma_set_max_seg_size(dev->dev, 0xffffffff);
2562 
2563 	mutex_unlock(&dev->shutdown_lock);
2564 
2565 	/*
2566 	 * Introduce CONNECTING state from nvme-fc/rdma transports to mark the
2567 	 * initializing procedure here.
2568 	 */
2569 	if (!nvme_change_ctrl_state(&dev->ctrl, NVME_CTRL_CONNECTING)) {
2570 		dev_warn(dev->ctrl.device,
2571 			"failed to mark controller CONNECTING\n");
2572 		result = -EBUSY;
2573 		goto out;
2574 	}
2575 
2576 	result = nvme_init_identify(&dev->ctrl);
2577 	if (result)
2578 		goto out;
2579 
2580 	if (dev->ctrl.oacs & NVME_CTRL_OACS_SEC_SUPP) {
2581 		if (!dev->ctrl.opal_dev)
2582 			dev->ctrl.opal_dev =
2583 				init_opal_dev(&dev->ctrl, &nvme_sec_submit);
2584 		else if (was_suspend)
2585 			opal_unlock_from_suspend(dev->ctrl.opal_dev);
2586 	} else {
2587 		free_opal_dev(dev->ctrl.opal_dev);
2588 		dev->ctrl.opal_dev = NULL;
2589 	}
2590 
2591 	if (dev->ctrl.oacs & NVME_CTRL_OACS_DBBUF_SUPP) {
2592 		result = nvme_dbbuf_dma_alloc(dev);
2593 		if (result)
2594 			dev_warn(dev->dev,
2595 				 "unable to allocate dma for dbbuf\n");
2596 	}
2597 
2598 	if (dev->ctrl.hmpre) {
2599 		result = nvme_setup_host_mem(dev);
2600 		if (result < 0)
2601 			goto out;
2602 	}
2603 
2604 	result = nvme_setup_io_queues(dev);
2605 	if (result)
2606 		goto out;
2607 
2608 	/*
2609 	 * Keep the controller around but remove all namespaces if we don't have
2610 	 * any working I/O queue.
2611 	 */
2612 	if (dev->online_queues < 2) {
2613 		dev_warn(dev->ctrl.device, "IO queues not created\n");
2614 		nvme_kill_queues(&dev->ctrl);
2615 		nvme_remove_namespaces(&dev->ctrl);
2616 		new_state = NVME_CTRL_ADMIN_ONLY;
2617 	} else {
2618 		nvme_start_queues(&dev->ctrl);
2619 		nvme_wait_freeze(&dev->ctrl);
2620 		/* hit this only when allocate tagset fails */
2621 		if (nvme_dev_add(dev))
2622 			new_state = NVME_CTRL_ADMIN_ONLY;
2623 		nvme_unfreeze(&dev->ctrl);
2624 	}
2625 
2626 	/*
2627 	 * If only admin queue live, keep it to do further investigation or
2628 	 * recovery.
2629 	 */
2630 	if (!nvme_change_ctrl_state(&dev->ctrl, new_state)) {
2631 		dev_warn(dev->ctrl.device,
2632 			"failed to mark controller state %d\n", new_state);
2633 		result = -ENODEV;
2634 		goto out;
2635 	}
2636 
2637 	nvme_start_ctrl(&dev->ctrl);
2638 	return;
2639 
2640  out_unlock:
2641 	mutex_unlock(&dev->shutdown_lock);
2642  out:
2643 	if (result)
2644 		dev_warn(dev->ctrl.device,
2645 			 "Removing after probe failure status: %d\n", result);
2646 	nvme_remove_dead_ctrl(dev);
2647 }
2648 
2649 static void nvme_remove_dead_ctrl_work(struct work_struct *work)
2650 {
2651 	struct nvme_dev *dev = container_of(work, struct nvme_dev, remove_work);
2652 	struct pci_dev *pdev = to_pci_dev(dev->dev);
2653 
2654 	if (pci_get_drvdata(pdev))
2655 		device_release_driver(&pdev->dev);
2656 	nvme_put_ctrl(&dev->ctrl);
2657 }
2658 
2659 static int nvme_pci_reg_read32(struct nvme_ctrl *ctrl, u32 off, u32 *val)
2660 {
2661 	*val = readl(to_nvme_dev(ctrl)->bar + off);
2662 	return 0;
2663 }
2664 
2665 static int nvme_pci_reg_write32(struct nvme_ctrl *ctrl, u32 off, u32 val)
2666 {
2667 	writel(val, to_nvme_dev(ctrl)->bar + off);
2668 	return 0;
2669 }
2670 
2671 static int nvme_pci_reg_read64(struct nvme_ctrl *ctrl, u32 off, u64 *val)
2672 {
2673 	*val = readq(to_nvme_dev(ctrl)->bar + off);
2674 	return 0;
2675 }
2676 
2677 static int nvme_pci_get_address(struct nvme_ctrl *ctrl, char *buf, int size)
2678 {
2679 	struct pci_dev *pdev = to_pci_dev(to_nvme_dev(ctrl)->dev);
2680 
2681 	return snprintf(buf, size, "%s", dev_name(&pdev->dev));
2682 }
2683 
2684 static const struct nvme_ctrl_ops nvme_pci_ctrl_ops = {
2685 	.name			= "pcie",
2686 	.module			= THIS_MODULE,
2687 	.flags			= NVME_F_METADATA_SUPPORTED |
2688 				  NVME_F_PCI_P2PDMA,
2689 	.reg_read32		= nvme_pci_reg_read32,
2690 	.reg_write32		= nvme_pci_reg_write32,
2691 	.reg_read64		= nvme_pci_reg_read64,
2692 	.free_ctrl		= nvme_pci_free_ctrl,
2693 	.submit_async_event	= nvme_pci_submit_async_event,
2694 	.get_address		= nvme_pci_get_address,
2695 };
2696 
2697 static int nvme_dev_map(struct nvme_dev *dev)
2698 {
2699 	struct pci_dev *pdev = to_pci_dev(dev->dev);
2700 
2701 	if (pci_request_mem_regions(pdev, "nvme"))
2702 		return -ENODEV;
2703 
2704 	if (nvme_remap_bar(dev, NVME_REG_DBS + 4096))
2705 		goto release;
2706 
2707 	return 0;
2708   release:
2709 	pci_release_mem_regions(pdev);
2710 	return -ENODEV;
2711 }
2712 
2713 static unsigned long check_vendor_combination_bug(struct pci_dev *pdev)
2714 {
2715 	if (pdev->vendor == 0x144d && pdev->device == 0xa802) {
2716 		/*
2717 		 * Several Samsung devices seem to drop off the PCIe bus
2718 		 * randomly when APST is on and uses the deepest sleep state.
2719 		 * This has been observed on a Samsung "SM951 NVMe SAMSUNG
2720 		 * 256GB", a "PM951 NVMe SAMSUNG 512GB", and a "Samsung SSD
2721 		 * 950 PRO 256GB", but it seems to be restricted to two Dell
2722 		 * laptops.
2723 		 */
2724 		if (dmi_match(DMI_SYS_VENDOR, "Dell Inc.") &&
2725 		    (dmi_match(DMI_PRODUCT_NAME, "XPS 15 9550") ||
2726 		     dmi_match(DMI_PRODUCT_NAME, "Precision 5510")))
2727 			return NVME_QUIRK_NO_DEEPEST_PS;
2728 	} else if (pdev->vendor == 0x144d && pdev->device == 0xa804) {
2729 		/*
2730 		 * Samsung SSD 960 EVO drops off the PCIe bus after system
2731 		 * suspend on a Ryzen board, ASUS PRIME B350M-A, as well as
2732 		 * within few minutes after bootup on a Coffee Lake board -
2733 		 * ASUS PRIME Z370-A
2734 		 */
2735 		if (dmi_match(DMI_BOARD_VENDOR, "ASUSTeK COMPUTER INC.") &&
2736 		    (dmi_match(DMI_BOARD_NAME, "PRIME B350M-A") ||
2737 		     dmi_match(DMI_BOARD_NAME, "PRIME Z370-A")))
2738 			return NVME_QUIRK_NO_APST;
2739 	}
2740 
2741 	return 0;
2742 }
2743 
2744 static void nvme_async_probe(void *data, async_cookie_t cookie)
2745 {
2746 	struct nvme_dev *dev = data;
2747 
2748 	flush_work(&dev->ctrl.reset_work);
2749 	flush_work(&dev->ctrl.scan_work);
2750 	nvme_put_ctrl(&dev->ctrl);
2751 }
2752 
2753 static int nvme_probe(struct pci_dev *pdev, const struct pci_device_id *id)
2754 {
2755 	int node, result = -ENOMEM;
2756 	struct nvme_dev *dev;
2757 	unsigned long quirks = id->driver_data;
2758 	size_t alloc_size;
2759 
2760 	node = dev_to_node(&pdev->dev);
2761 	if (node == NUMA_NO_NODE)
2762 		set_dev_node(&pdev->dev, first_memory_node);
2763 
2764 	dev = kzalloc_node(sizeof(*dev), GFP_KERNEL, node);
2765 	if (!dev)
2766 		return -ENOMEM;
2767 
2768 	dev->queues = kcalloc_node(max_queue_count(), sizeof(struct nvme_queue),
2769 					GFP_KERNEL, node);
2770 	if (!dev->queues)
2771 		goto free;
2772 
2773 	dev->dev = get_device(&pdev->dev);
2774 	pci_set_drvdata(pdev, dev);
2775 
2776 	result = nvme_dev_map(dev);
2777 	if (result)
2778 		goto put_pci;
2779 
2780 	INIT_WORK(&dev->ctrl.reset_work, nvme_reset_work);
2781 	INIT_WORK(&dev->remove_work, nvme_remove_dead_ctrl_work);
2782 	mutex_init(&dev->shutdown_lock);
2783 
2784 	result = nvme_setup_prp_pools(dev);
2785 	if (result)
2786 		goto unmap;
2787 
2788 	quirks |= check_vendor_combination_bug(pdev);
2789 
2790 	/*
2791 	 * Double check that our mempool alloc size will cover the biggest
2792 	 * command we support.
2793 	 */
2794 	alloc_size = nvme_pci_iod_alloc_size(dev, NVME_MAX_KB_SZ,
2795 						NVME_MAX_SEGS, true);
2796 	WARN_ON_ONCE(alloc_size > PAGE_SIZE);
2797 
2798 	dev->iod_mempool = mempool_create_node(1, mempool_kmalloc,
2799 						mempool_kfree,
2800 						(void *) alloc_size,
2801 						GFP_KERNEL, node);
2802 	if (!dev->iod_mempool) {
2803 		result = -ENOMEM;
2804 		goto release_pools;
2805 	}
2806 
2807 	result = nvme_init_ctrl(&dev->ctrl, &pdev->dev, &nvme_pci_ctrl_ops,
2808 			quirks);
2809 	if (result)
2810 		goto release_mempool;
2811 
2812 	dev_info(dev->ctrl.device, "pci function %s\n", dev_name(&pdev->dev));
2813 
2814 	nvme_reset_ctrl(&dev->ctrl);
2815 	nvme_get_ctrl(&dev->ctrl);
2816 	async_schedule(nvme_async_probe, dev);
2817 
2818 	return 0;
2819 
2820  release_mempool:
2821 	mempool_destroy(dev->iod_mempool);
2822  release_pools:
2823 	nvme_release_prp_pools(dev);
2824  unmap:
2825 	nvme_dev_unmap(dev);
2826  put_pci:
2827 	put_device(dev->dev);
2828  free:
2829 	kfree(dev->queues);
2830 	kfree(dev);
2831 	return result;
2832 }
2833 
2834 static void nvme_reset_prepare(struct pci_dev *pdev)
2835 {
2836 	struct nvme_dev *dev = pci_get_drvdata(pdev);
2837 	nvme_dev_disable(dev, false);
2838 }
2839 
2840 static void nvme_reset_done(struct pci_dev *pdev)
2841 {
2842 	struct nvme_dev *dev = pci_get_drvdata(pdev);
2843 	nvme_reset_ctrl_sync(&dev->ctrl);
2844 }
2845 
2846 static void nvme_shutdown(struct pci_dev *pdev)
2847 {
2848 	struct nvme_dev *dev = pci_get_drvdata(pdev);
2849 	nvme_dev_disable(dev, true);
2850 }
2851 
2852 /*
2853  * The driver's remove may be called on a device in a partially initialized
2854  * state. This function must not have any dependencies on the device state in
2855  * order to proceed.
2856  */
2857 static void nvme_remove(struct pci_dev *pdev)
2858 {
2859 	struct nvme_dev *dev = pci_get_drvdata(pdev);
2860 
2861 	nvme_change_ctrl_state(&dev->ctrl, NVME_CTRL_DELETING);
2862 	pci_set_drvdata(pdev, NULL);
2863 
2864 	if (!pci_device_is_present(pdev)) {
2865 		nvme_change_ctrl_state(&dev->ctrl, NVME_CTRL_DEAD);
2866 		nvme_dev_disable(dev, true);
2867 		nvme_dev_remove_admin(dev);
2868 	}
2869 
2870 	flush_work(&dev->ctrl.reset_work);
2871 	nvme_stop_ctrl(&dev->ctrl);
2872 	nvme_remove_namespaces(&dev->ctrl);
2873 	nvme_dev_disable(dev, true);
2874 	nvme_release_cmb(dev);
2875 	nvme_free_host_mem(dev);
2876 	nvme_dev_remove_admin(dev);
2877 	nvme_free_queues(dev, 0);
2878 	nvme_uninit_ctrl(&dev->ctrl);
2879 	nvme_release_prp_pools(dev);
2880 	nvme_dev_unmap(dev);
2881 	nvme_put_ctrl(&dev->ctrl);
2882 }
2883 
2884 #ifdef CONFIG_PM_SLEEP
2885 static int nvme_get_power_state(struct nvme_ctrl *ctrl, u32 *ps)
2886 {
2887 	return nvme_get_features(ctrl, NVME_FEAT_POWER_MGMT, 0, NULL, 0, ps);
2888 }
2889 
2890 static int nvme_set_power_state(struct nvme_ctrl *ctrl, u32 ps)
2891 {
2892 	return nvme_set_features(ctrl, NVME_FEAT_POWER_MGMT, ps, NULL, 0, NULL);
2893 }
2894 
2895 static int nvme_resume(struct device *dev)
2896 {
2897 	struct nvme_dev *ndev = pci_get_drvdata(to_pci_dev(dev));
2898 	struct nvme_ctrl *ctrl = &ndev->ctrl;
2899 
2900 	if (ndev->last_ps == U32_MAX ||
2901 	    nvme_set_power_state(ctrl, ndev->last_ps) != 0)
2902 		nvme_reset_ctrl(ctrl);
2903 	return 0;
2904 }
2905 
2906 static int nvme_suspend(struct device *dev)
2907 {
2908 	struct pci_dev *pdev = to_pci_dev(dev);
2909 	struct nvme_dev *ndev = pci_get_drvdata(pdev);
2910 	struct nvme_ctrl *ctrl = &ndev->ctrl;
2911 	int ret = -EBUSY;
2912 
2913 	ndev->last_ps = U32_MAX;
2914 
2915 	/*
2916 	 * The platform does not remove power for a kernel managed suspend so
2917 	 * use host managed nvme power settings for lowest idle power if
2918 	 * possible. This should have quicker resume latency than a full device
2919 	 * shutdown.  But if the firmware is involved after the suspend or the
2920 	 * device does not support any non-default power states, shut down the
2921 	 * device fully.
2922 	 *
2923 	 * If ASPM is not enabled for the device, shut down the device and allow
2924 	 * the PCI bus layer to put it into D3 in order to take the PCIe link
2925 	 * down, so as to allow the platform to achieve its minimum low-power
2926 	 * state (which may not be possible if the link is up).
2927 	 */
2928 	if (pm_suspend_via_firmware() || !ctrl->npss ||
2929 	    !pcie_aspm_enabled(pdev) ||
2930 	    (ndev->ctrl.quirks & NVME_QUIRK_SIMPLE_SUSPEND)) {
2931 		nvme_dev_disable(ndev, true);
2932 		return 0;
2933 	}
2934 
2935 	nvme_start_freeze(ctrl);
2936 	nvme_wait_freeze(ctrl);
2937 	nvme_sync_queues(ctrl);
2938 
2939 	if (ctrl->state != NVME_CTRL_LIVE &&
2940 	    ctrl->state != NVME_CTRL_ADMIN_ONLY)
2941 		goto unfreeze;
2942 
2943 	ret = nvme_get_power_state(ctrl, &ndev->last_ps);
2944 	if (ret < 0)
2945 		goto unfreeze;
2946 
2947 	ret = nvme_set_power_state(ctrl, ctrl->npss);
2948 	if (ret < 0)
2949 		goto unfreeze;
2950 
2951 	if (ret) {
2952 		/*
2953 		 * Clearing npss forces a controller reset on resume. The
2954 		 * correct value will be resdicovered then.
2955 		 */
2956 		nvme_dev_disable(ndev, true);
2957 		ctrl->npss = 0;
2958 		ret = 0;
2959 		goto unfreeze;
2960 	}
2961 	/*
2962 	 * A saved state prevents pci pm from generically controlling the
2963 	 * device's power. If we're using protocol specific settings, we don't
2964 	 * want pci interfering.
2965 	 */
2966 	pci_save_state(pdev);
2967 unfreeze:
2968 	nvme_unfreeze(ctrl);
2969 	return ret;
2970 }
2971 
2972 static int nvme_simple_suspend(struct device *dev)
2973 {
2974 	struct nvme_dev *ndev = pci_get_drvdata(to_pci_dev(dev));
2975 
2976 	nvme_dev_disable(ndev, true);
2977 	return 0;
2978 }
2979 
2980 static int nvme_simple_resume(struct device *dev)
2981 {
2982 	struct pci_dev *pdev = to_pci_dev(dev);
2983 	struct nvme_dev *ndev = pci_get_drvdata(pdev);
2984 
2985 	nvme_reset_ctrl(&ndev->ctrl);
2986 	return 0;
2987 }
2988 
2989 static const struct dev_pm_ops nvme_dev_pm_ops = {
2990 	.suspend	= nvme_suspend,
2991 	.resume		= nvme_resume,
2992 	.freeze		= nvme_simple_suspend,
2993 	.thaw		= nvme_simple_resume,
2994 	.poweroff	= nvme_simple_suspend,
2995 	.restore	= nvme_simple_resume,
2996 };
2997 #endif /* CONFIG_PM_SLEEP */
2998 
2999 static pci_ers_result_t nvme_error_detected(struct pci_dev *pdev,
3000 						pci_channel_state_t state)
3001 {
3002 	struct nvme_dev *dev = pci_get_drvdata(pdev);
3003 
3004 	/*
3005 	 * A frozen channel requires a reset. When detected, this method will
3006 	 * shutdown the controller to quiesce. The controller will be restarted
3007 	 * after the slot reset through driver's slot_reset callback.
3008 	 */
3009 	switch (state) {
3010 	case pci_channel_io_normal:
3011 		return PCI_ERS_RESULT_CAN_RECOVER;
3012 	case pci_channel_io_frozen:
3013 		dev_warn(dev->ctrl.device,
3014 			"frozen state error detected, reset controller\n");
3015 		nvme_dev_disable(dev, false);
3016 		return PCI_ERS_RESULT_NEED_RESET;
3017 	case pci_channel_io_perm_failure:
3018 		dev_warn(dev->ctrl.device,
3019 			"failure state error detected, request disconnect\n");
3020 		return PCI_ERS_RESULT_DISCONNECT;
3021 	}
3022 	return PCI_ERS_RESULT_NEED_RESET;
3023 }
3024 
3025 static pci_ers_result_t nvme_slot_reset(struct pci_dev *pdev)
3026 {
3027 	struct nvme_dev *dev = pci_get_drvdata(pdev);
3028 
3029 	dev_info(dev->ctrl.device, "restart after slot reset\n");
3030 	pci_restore_state(pdev);
3031 	nvme_reset_ctrl(&dev->ctrl);
3032 	return PCI_ERS_RESULT_RECOVERED;
3033 }
3034 
3035 static void nvme_error_resume(struct pci_dev *pdev)
3036 {
3037 	struct nvme_dev *dev = pci_get_drvdata(pdev);
3038 
3039 	flush_work(&dev->ctrl.reset_work);
3040 }
3041 
3042 static const struct pci_error_handlers nvme_err_handler = {
3043 	.error_detected	= nvme_error_detected,
3044 	.slot_reset	= nvme_slot_reset,
3045 	.resume		= nvme_error_resume,
3046 	.reset_prepare	= nvme_reset_prepare,
3047 	.reset_done	= nvme_reset_done,
3048 };
3049 
3050 static const struct pci_device_id nvme_id_table[] = {
3051 	{ PCI_VDEVICE(INTEL, 0x0953),
3052 		.driver_data = NVME_QUIRK_STRIPE_SIZE |
3053 				NVME_QUIRK_DEALLOCATE_ZEROES, },
3054 	{ PCI_VDEVICE(INTEL, 0x0a53),
3055 		.driver_data = NVME_QUIRK_STRIPE_SIZE |
3056 				NVME_QUIRK_DEALLOCATE_ZEROES, },
3057 	{ PCI_VDEVICE(INTEL, 0x0a54),
3058 		.driver_data = NVME_QUIRK_STRIPE_SIZE |
3059 				NVME_QUIRK_DEALLOCATE_ZEROES, },
3060 	{ PCI_VDEVICE(INTEL, 0x0a55),
3061 		.driver_data = NVME_QUIRK_STRIPE_SIZE |
3062 				NVME_QUIRK_DEALLOCATE_ZEROES, },
3063 	{ PCI_VDEVICE(INTEL, 0xf1a5),	/* Intel 600P/P3100 */
3064 		.driver_data = NVME_QUIRK_NO_DEEPEST_PS |
3065 				NVME_QUIRK_MEDIUM_PRIO_SQ },
3066 	{ PCI_VDEVICE(INTEL, 0xf1a6),	/* Intel 760p/Pro 7600p */
3067 		.driver_data = NVME_QUIRK_IGNORE_DEV_SUBNQN, },
3068 	{ PCI_VDEVICE(INTEL, 0x5845),	/* Qemu emulated controller */
3069 		.driver_data = NVME_QUIRK_IDENTIFY_CNS |
3070 				NVME_QUIRK_DISABLE_WRITE_ZEROES, },
3071 	{ PCI_DEVICE(0x1bb1, 0x0100),   /* Seagate Nytro Flash Storage */
3072 		.driver_data = NVME_QUIRK_DELAY_BEFORE_CHK_RDY, },
3073 	{ PCI_DEVICE(0x1c58, 0x0003),	/* HGST adapter */
3074 		.driver_data = NVME_QUIRK_DELAY_BEFORE_CHK_RDY, },
3075 	{ PCI_DEVICE(0x1c58, 0x0023),	/* WDC SN200 adapter */
3076 		.driver_data = NVME_QUIRK_DELAY_BEFORE_CHK_RDY, },
3077 	{ PCI_DEVICE(0x1c5f, 0x0540),	/* Memblaze Pblaze4 adapter */
3078 		.driver_data = NVME_QUIRK_DELAY_BEFORE_CHK_RDY, },
3079 	{ PCI_DEVICE(0x144d, 0xa821),   /* Samsung PM1725 */
3080 		.driver_data = NVME_QUIRK_DELAY_BEFORE_CHK_RDY, },
3081 	{ PCI_DEVICE(0x144d, 0xa822),   /* Samsung PM1725a */
3082 		.driver_data = NVME_QUIRK_DELAY_BEFORE_CHK_RDY, },
3083 	{ PCI_DEVICE(0x1d1d, 0x1f1f),	/* LighNVM qemu device */
3084 		.driver_data = NVME_QUIRK_LIGHTNVM, },
3085 	{ PCI_DEVICE(0x1d1d, 0x2807),	/* CNEX WL */
3086 		.driver_data = NVME_QUIRK_LIGHTNVM, },
3087 	{ PCI_DEVICE(0x1d1d, 0x2601),	/* CNEX Granby */
3088 		.driver_data = NVME_QUIRK_LIGHTNVM, },
3089 	{ PCI_DEVICE(0x10ec, 0x5762),   /* ADATA SX6000LNP */
3090 		.driver_data = NVME_QUIRK_IGNORE_DEV_SUBNQN, },
3091 	{ PCI_DEVICE_CLASS(PCI_CLASS_STORAGE_EXPRESS, 0xffffff) },
3092 	{ PCI_DEVICE(PCI_VENDOR_ID_APPLE, 0x2001) },
3093 	{ PCI_DEVICE(PCI_VENDOR_ID_APPLE, 0x2003) },
3094 	{ PCI_DEVICE(PCI_VENDOR_ID_APPLE, 0x2005),
3095 		.driver_data = NVME_QUIRK_SINGLE_VECTOR |
3096 				NVME_QUIRK_128_BYTES_SQES |
3097 				NVME_QUIRK_SHARED_TAGS },
3098 	{ 0, }
3099 };
3100 MODULE_DEVICE_TABLE(pci, nvme_id_table);
3101 
3102 static struct pci_driver nvme_driver = {
3103 	.name		= "nvme",
3104 	.id_table	= nvme_id_table,
3105 	.probe		= nvme_probe,
3106 	.remove		= nvme_remove,
3107 	.shutdown	= nvme_shutdown,
3108 #ifdef CONFIG_PM_SLEEP
3109 	.driver		= {
3110 		.pm	= &nvme_dev_pm_ops,
3111 	},
3112 #endif
3113 	.sriov_configure = pci_sriov_configure_simple,
3114 	.err_handler	= &nvme_err_handler,
3115 };
3116 
3117 static int __init nvme_init(void)
3118 {
3119 	BUILD_BUG_ON(sizeof(struct nvme_create_cq) != 64);
3120 	BUILD_BUG_ON(sizeof(struct nvme_create_sq) != 64);
3121 	BUILD_BUG_ON(sizeof(struct nvme_delete_queue) != 64);
3122 	BUILD_BUG_ON(IRQ_AFFINITY_MAX_SETS < 2);
3123 	return pci_register_driver(&nvme_driver);
3124 }
3125 
3126 static void __exit nvme_exit(void)
3127 {
3128 	pci_unregister_driver(&nvme_driver);
3129 	flush_workqueue(nvme_wq);
3130 }
3131 
3132 MODULE_AUTHOR("Matthew Wilcox <willy@linux.intel.com>");
3133 MODULE_LICENSE("GPL");
3134 MODULE_VERSION("1.0");
3135 module_init(nvme_init);
3136 module_exit(nvme_exit);
3137