xref: /openbmc/linux/drivers/nvme/host/pci.c (revision 08283d30)
1 // SPDX-License-Identifier: GPL-2.0
2 /*
3  * NVM Express device driver
4  * Copyright (c) 2011-2014, Intel Corporation.
5  */
6 
7 #include <linux/aer.h>
8 #include <linux/async.h>
9 #include <linux/blkdev.h>
10 #include <linux/blk-mq.h>
11 #include <linux/blk-mq-pci.h>
12 #include <linux/dmi.h>
13 #include <linux/init.h>
14 #include <linux/interrupt.h>
15 #include <linux/io.h>
16 #include <linux/mm.h>
17 #include <linux/module.h>
18 #include <linux/mutex.h>
19 #include <linux/once.h>
20 #include <linux/pci.h>
21 #include <linux/suspend.h>
22 #include <linux/t10-pi.h>
23 #include <linux/types.h>
24 #include <linux/io-64-nonatomic-lo-hi.h>
25 #include <linux/sed-opal.h>
26 #include <linux/pci-p2pdma.h>
27 
28 #include "trace.h"
29 #include "nvme.h"
30 
31 #define SQ_SIZE(q)	((q)->q_depth << (q)->sqes)
32 #define CQ_SIZE(q)	((q)->q_depth * sizeof(struct nvme_completion))
33 
34 #define SGES_PER_PAGE	(PAGE_SIZE / sizeof(struct nvme_sgl_desc))
35 
36 /*
37  * These can be higher, but we need to ensure that any command doesn't
38  * require an sg allocation that needs more than a page of data.
39  */
40 #define NVME_MAX_KB_SZ	4096
41 #define NVME_MAX_SEGS	127
42 
43 static int use_threaded_interrupts;
44 module_param(use_threaded_interrupts, int, 0);
45 
46 static bool use_cmb_sqes = true;
47 module_param(use_cmb_sqes, bool, 0444);
48 MODULE_PARM_DESC(use_cmb_sqes, "use controller's memory buffer for I/O SQes");
49 
50 static unsigned int max_host_mem_size_mb = 128;
51 module_param(max_host_mem_size_mb, uint, 0444);
52 MODULE_PARM_DESC(max_host_mem_size_mb,
53 	"Maximum Host Memory Buffer (HMB) size per controller (in MiB)");
54 
55 static unsigned int sgl_threshold = SZ_32K;
56 module_param(sgl_threshold, uint, 0644);
57 MODULE_PARM_DESC(sgl_threshold,
58 		"Use SGLs when average request segment size is larger or equal to "
59 		"this size. Use 0 to disable SGLs.");
60 
61 static int io_queue_depth_set(const char *val, const struct kernel_param *kp);
62 static const struct kernel_param_ops io_queue_depth_ops = {
63 	.set = io_queue_depth_set,
64 	.get = param_get_int,
65 };
66 
67 static int io_queue_depth = 1024;
68 module_param_cb(io_queue_depth, &io_queue_depth_ops, &io_queue_depth, 0644);
69 MODULE_PARM_DESC(io_queue_depth, "set io queue depth, should >= 2");
70 
71 static int write_queues;
72 module_param(write_queues, int, 0644);
73 MODULE_PARM_DESC(write_queues,
74 	"Number of queues to use for writes. If not set, reads and writes "
75 	"will share a queue set.");
76 
77 static int poll_queues;
78 module_param(poll_queues, int, 0644);
79 MODULE_PARM_DESC(poll_queues, "Number of queues to use for polled IO.");
80 
81 struct nvme_dev;
82 struct nvme_queue;
83 
84 static void nvme_dev_disable(struct nvme_dev *dev, bool shutdown);
85 static bool __nvme_disable_io_queues(struct nvme_dev *dev, u8 opcode);
86 
87 /*
88  * Represents an NVM Express device.  Each nvme_dev is a PCI function.
89  */
90 struct nvme_dev {
91 	struct nvme_queue *queues;
92 	struct blk_mq_tag_set tagset;
93 	struct blk_mq_tag_set admin_tagset;
94 	u32 __iomem *dbs;
95 	struct device *dev;
96 	struct dma_pool *prp_page_pool;
97 	struct dma_pool *prp_small_pool;
98 	unsigned online_queues;
99 	unsigned max_qid;
100 	unsigned io_queues[HCTX_MAX_TYPES];
101 	unsigned int num_vecs;
102 	int q_depth;
103 	int io_sqes;
104 	u32 db_stride;
105 	void __iomem *bar;
106 	unsigned long bar_mapped_size;
107 	struct work_struct remove_work;
108 	struct mutex shutdown_lock;
109 	bool subsystem;
110 	u64 cmb_size;
111 	bool cmb_use_sqes;
112 	u32 cmbsz;
113 	u32 cmbloc;
114 	struct nvme_ctrl ctrl;
115 	u32 last_ps;
116 
117 	mempool_t *iod_mempool;
118 
119 	/* shadow doorbell buffer support: */
120 	u32 *dbbuf_dbs;
121 	dma_addr_t dbbuf_dbs_dma_addr;
122 	u32 *dbbuf_eis;
123 	dma_addr_t dbbuf_eis_dma_addr;
124 
125 	/* host memory buffer support: */
126 	u64 host_mem_size;
127 	u32 nr_host_mem_descs;
128 	dma_addr_t host_mem_descs_dma;
129 	struct nvme_host_mem_buf_desc *host_mem_descs;
130 	void **host_mem_desc_bufs;
131 };
132 
133 static int io_queue_depth_set(const char *val, const struct kernel_param *kp)
134 {
135 	int n = 0, ret;
136 
137 	ret = kstrtoint(val, 10, &n);
138 	if (ret != 0 || n < 2)
139 		return -EINVAL;
140 
141 	return param_set_int(val, kp);
142 }
143 
144 static inline unsigned int sq_idx(unsigned int qid, u32 stride)
145 {
146 	return qid * 2 * stride;
147 }
148 
149 static inline unsigned int cq_idx(unsigned int qid, u32 stride)
150 {
151 	return (qid * 2 + 1) * stride;
152 }
153 
154 static inline struct nvme_dev *to_nvme_dev(struct nvme_ctrl *ctrl)
155 {
156 	return container_of(ctrl, struct nvme_dev, ctrl);
157 }
158 
159 /*
160  * An NVM Express queue.  Each device has at least two (one for admin
161  * commands and one for I/O commands).
162  */
163 struct nvme_queue {
164 	struct nvme_dev *dev;
165 	spinlock_t sq_lock;
166 	void *sq_cmds;
167 	 /* only used for poll queues: */
168 	spinlock_t cq_poll_lock ____cacheline_aligned_in_smp;
169 	volatile struct nvme_completion *cqes;
170 	struct blk_mq_tags **tags;
171 	dma_addr_t sq_dma_addr;
172 	dma_addr_t cq_dma_addr;
173 	u32 __iomem *q_db;
174 	u16 q_depth;
175 	u16 cq_vector;
176 	u16 sq_tail;
177 	u16 last_sq_tail;
178 	u16 cq_head;
179 	u16 last_cq_head;
180 	u16 qid;
181 	u8 cq_phase;
182 	u8 sqes;
183 	unsigned long flags;
184 #define NVMEQ_ENABLED		0
185 #define NVMEQ_SQ_CMB		1
186 #define NVMEQ_DELETE_ERROR	2
187 #define NVMEQ_POLLED		3
188 	u32 *dbbuf_sq_db;
189 	u32 *dbbuf_cq_db;
190 	u32 *dbbuf_sq_ei;
191 	u32 *dbbuf_cq_ei;
192 	struct completion delete_done;
193 };
194 
195 /*
196  * The nvme_iod describes the data in an I/O.
197  *
198  * The sg pointer contains the list of PRP/SGL chunk allocations in addition
199  * to the actual struct scatterlist.
200  */
201 struct nvme_iod {
202 	struct nvme_request req;
203 	struct nvme_queue *nvmeq;
204 	bool use_sgl;
205 	int aborted;
206 	int npages;		/* In the PRP list. 0 means small pool in use */
207 	int nents;		/* Used in scatterlist */
208 	dma_addr_t first_dma;
209 	unsigned int dma_len;	/* length of single DMA segment mapping */
210 	dma_addr_t meta_dma;
211 	struct scatterlist *sg;
212 };
213 
214 static unsigned int max_io_queues(void)
215 {
216 	return num_possible_cpus() + write_queues + poll_queues;
217 }
218 
219 static unsigned int max_queue_count(void)
220 {
221 	/* IO queues + admin queue */
222 	return 1 + max_io_queues();
223 }
224 
225 static inline unsigned int nvme_dbbuf_size(u32 stride)
226 {
227 	return (max_queue_count() * 8 * stride);
228 }
229 
230 static int nvme_dbbuf_dma_alloc(struct nvme_dev *dev)
231 {
232 	unsigned int mem_size = nvme_dbbuf_size(dev->db_stride);
233 
234 	if (dev->dbbuf_dbs)
235 		return 0;
236 
237 	dev->dbbuf_dbs = dma_alloc_coherent(dev->dev, mem_size,
238 					    &dev->dbbuf_dbs_dma_addr,
239 					    GFP_KERNEL);
240 	if (!dev->dbbuf_dbs)
241 		return -ENOMEM;
242 	dev->dbbuf_eis = dma_alloc_coherent(dev->dev, mem_size,
243 					    &dev->dbbuf_eis_dma_addr,
244 					    GFP_KERNEL);
245 	if (!dev->dbbuf_eis) {
246 		dma_free_coherent(dev->dev, mem_size,
247 				  dev->dbbuf_dbs, dev->dbbuf_dbs_dma_addr);
248 		dev->dbbuf_dbs = NULL;
249 		return -ENOMEM;
250 	}
251 
252 	return 0;
253 }
254 
255 static void nvme_dbbuf_dma_free(struct nvme_dev *dev)
256 {
257 	unsigned int mem_size = nvme_dbbuf_size(dev->db_stride);
258 
259 	if (dev->dbbuf_dbs) {
260 		dma_free_coherent(dev->dev, mem_size,
261 				  dev->dbbuf_dbs, dev->dbbuf_dbs_dma_addr);
262 		dev->dbbuf_dbs = NULL;
263 	}
264 	if (dev->dbbuf_eis) {
265 		dma_free_coherent(dev->dev, mem_size,
266 				  dev->dbbuf_eis, dev->dbbuf_eis_dma_addr);
267 		dev->dbbuf_eis = NULL;
268 	}
269 }
270 
271 static void nvme_dbbuf_init(struct nvme_dev *dev,
272 			    struct nvme_queue *nvmeq, int qid)
273 {
274 	if (!dev->dbbuf_dbs || !qid)
275 		return;
276 
277 	nvmeq->dbbuf_sq_db = &dev->dbbuf_dbs[sq_idx(qid, dev->db_stride)];
278 	nvmeq->dbbuf_cq_db = &dev->dbbuf_dbs[cq_idx(qid, dev->db_stride)];
279 	nvmeq->dbbuf_sq_ei = &dev->dbbuf_eis[sq_idx(qid, dev->db_stride)];
280 	nvmeq->dbbuf_cq_ei = &dev->dbbuf_eis[cq_idx(qid, dev->db_stride)];
281 }
282 
283 static void nvme_dbbuf_set(struct nvme_dev *dev)
284 {
285 	struct nvme_command c;
286 
287 	if (!dev->dbbuf_dbs)
288 		return;
289 
290 	memset(&c, 0, sizeof(c));
291 	c.dbbuf.opcode = nvme_admin_dbbuf;
292 	c.dbbuf.prp1 = cpu_to_le64(dev->dbbuf_dbs_dma_addr);
293 	c.dbbuf.prp2 = cpu_to_le64(dev->dbbuf_eis_dma_addr);
294 
295 	if (nvme_submit_sync_cmd(dev->ctrl.admin_q, &c, NULL, 0)) {
296 		dev_warn(dev->ctrl.device, "unable to set dbbuf\n");
297 		/* Free memory and continue on */
298 		nvme_dbbuf_dma_free(dev);
299 	}
300 }
301 
302 static inline int nvme_dbbuf_need_event(u16 event_idx, u16 new_idx, u16 old)
303 {
304 	return (u16)(new_idx - event_idx - 1) < (u16)(new_idx - old);
305 }
306 
307 /* Update dbbuf and return true if an MMIO is required */
308 static bool nvme_dbbuf_update_and_check_event(u16 value, u32 *dbbuf_db,
309 					      volatile u32 *dbbuf_ei)
310 {
311 	if (dbbuf_db) {
312 		u16 old_value;
313 
314 		/*
315 		 * Ensure that the queue is written before updating
316 		 * the doorbell in memory
317 		 */
318 		wmb();
319 
320 		old_value = *dbbuf_db;
321 		*dbbuf_db = value;
322 
323 		/*
324 		 * Ensure that the doorbell is updated before reading the event
325 		 * index from memory.  The controller needs to provide similar
326 		 * ordering to ensure the envent index is updated before reading
327 		 * the doorbell.
328 		 */
329 		mb();
330 
331 		if (!nvme_dbbuf_need_event(*dbbuf_ei, value, old_value))
332 			return false;
333 	}
334 
335 	return true;
336 }
337 
338 /*
339  * Will slightly overestimate the number of pages needed.  This is OK
340  * as it only leads to a small amount of wasted memory for the lifetime of
341  * the I/O.
342  */
343 static int nvme_npages(unsigned size, struct nvme_dev *dev)
344 {
345 	unsigned nprps = DIV_ROUND_UP(size + dev->ctrl.page_size,
346 				      dev->ctrl.page_size);
347 	return DIV_ROUND_UP(8 * nprps, PAGE_SIZE - 8);
348 }
349 
350 /*
351  * Calculates the number of pages needed for the SGL segments. For example a 4k
352  * page can accommodate 256 SGL descriptors.
353  */
354 static int nvme_pci_npages_sgl(unsigned int num_seg)
355 {
356 	return DIV_ROUND_UP(num_seg * sizeof(struct nvme_sgl_desc), PAGE_SIZE);
357 }
358 
359 static unsigned int nvme_pci_iod_alloc_size(struct nvme_dev *dev,
360 		unsigned int size, unsigned int nseg, bool use_sgl)
361 {
362 	size_t alloc_size;
363 
364 	if (use_sgl)
365 		alloc_size = sizeof(__le64 *) * nvme_pci_npages_sgl(nseg);
366 	else
367 		alloc_size = sizeof(__le64 *) * nvme_npages(size, dev);
368 
369 	return alloc_size + sizeof(struct scatterlist) * nseg;
370 }
371 
372 static int nvme_admin_init_hctx(struct blk_mq_hw_ctx *hctx, void *data,
373 				unsigned int hctx_idx)
374 {
375 	struct nvme_dev *dev = data;
376 	struct nvme_queue *nvmeq = &dev->queues[0];
377 
378 	WARN_ON(hctx_idx != 0);
379 	WARN_ON(dev->admin_tagset.tags[0] != hctx->tags);
380 	WARN_ON(nvmeq->tags);
381 
382 	hctx->driver_data = nvmeq;
383 	nvmeq->tags = &dev->admin_tagset.tags[0];
384 	return 0;
385 }
386 
387 static void nvme_admin_exit_hctx(struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx)
388 {
389 	struct nvme_queue *nvmeq = hctx->driver_data;
390 
391 	nvmeq->tags = NULL;
392 }
393 
394 static int nvme_init_hctx(struct blk_mq_hw_ctx *hctx, void *data,
395 			  unsigned int hctx_idx)
396 {
397 	struct nvme_dev *dev = data;
398 	struct nvme_queue *nvmeq = &dev->queues[hctx_idx + 1];
399 
400 	if (!nvmeq->tags)
401 		nvmeq->tags = &dev->tagset.tags[hctx_idx];
402 
403 	WARN_ON(dev->tagset.tags[hctx_idx] != hctx->tags);
404 	hctx->driver_data = nvmeq;
405 	return 0;
406 }
407 
408 static int nvme_init_request(struct blk_mq_tag_set *set, struct request *req,
409 		unsigned int hctx_idx, unsigned int numa_node)
410 {
411 	struct nvme_dev *dev = set->driver_data;
412 	struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
413 	int queue_idx = (set == &dev->tagset) ? hctx_idx + 1 : 0;
414 	struct nvme_queue *nvmeq = &dev->queues[queue_idx];
415 
416 	BUG_ON(!nvmeq);
417 	iod->nvmeq = nvmeq;
418 
419 	nvme_req(req)->ctrl = &dev->ctrl;
420 	return 0;
421 }
422 
423 static int queue_irq_offset(struct nvme_dev *dev)
424 {
425 	/* if we have more than 1 vec, admin queue offsets us by 1 */
426 	if (dev->num_vecs > 1)
427 		return 1;
428 
429 	return 0;
430 }
431 
432 static int nvme_pci_map_queues(struct blk_mq_tag_set *set)
433 {
434 	struct nvme_dev *dev = set->driver_data;
435 	int i, qoff, offset;
436 
437 	offset = queue_irq_offset(dev);
438 	for (i = 0, qoff = 0; i < set->nr_maps; i++) {
439 		struct blk_mq_queue_map *map = &set->map[i];
440 
441 		map->nr_queues = dev->io_queues[i];
442 		if (!map->nr_queues) {
443 			BUG_ON(i == HCTX_TYPE_DEFAULT);
444 			continue;
445 		}
446 
447 		/*
448 		 * The poll queue(s) doesn't have an IRQ (and hence IRQ
449 		 * affinity), so use the regular blk-mq cpu mapping
450 		 */
451 		map->queue_offset = qoff;
452 		if (i != HCTX_TYPE_POLL && offset)
453 			blk_mq_pci_map_queues(map, to_pci_dev(dev->dev), offset);
454 		else
455 			blk_mq_map_queues(map);
456 		qoff += map->nr_queues;
457 		offset += map->nr_queues;
458 	}
459 
460 	return 0;
461 }
462 
463 /*
464  * Write sq tail if we are asked to, or if the next command would wrap.
465  */
466 static inline void nvme_write_sq_db(struct nvme_queue *nvmeq, bool write_sq)
467 {
468 	if (!write_sq) {
469 		u16 next_tail = nvmeq->sq_tail + 1;
470 
471 		if (next_tail == nvmeq->q_depth)
472 			next_tail = 0;
473 		if (next_tail != nvmeq->last_sq_tail)
474 			return;
475 	}
476 
477 	if (nvme_dbbuf_update_and_check_event(nvmeq->sq_tail,
478 			nvmeq->dbbuf_sq_db, nvmeq->dbbuf_sq_ei))
479 		writel(nvmeq->sq_tail, nvmeq->q_db);
480 	nvmeq->last_sq_tail = nvmeq->sq_tail;
481 }
482 
483 /**
484  * nvme_submit_cmd() - Copy a command into a queue and ring the doorbell
485  * @nvmeq: The queue to use
486  * @cmd: The command to send
487  * @write_sq: whether to write to the SQ doorbell
488  */
489 static void nvme_submit_cmd(struct nvme_queue *nvmeq, struct nvme_command *cmd,
490 			    bool write_sq)
491 {
492 	spin_lock(&nvmeq->sq_lock);
493 	memcpy(nvmeq->sq_cmds + (nvmeq->sq_tail << nvmeq->sqes),
494 	       cmd, sizeof(*cmd));
495 	if (++nvmeq->sq_tail == nvmeq->q_depth)
496 		nvmeq->sq_tail = 0;
497 	nvme_write_sq_db(nvmeq, write_sq);
498 	spin_unlock(&nvmeq->sq_lock);
499 }
500 
501 static void nvme_commit_rqs(struct blk_mq_hw_ctx *hctx)
502 {
503 	struct nvme_queue *nvmeq = hctx->driver_data;
504 
505 	spin_lock(&nvmeq->sq_lock);
506 	if (nvmeq->sq_tail != nvmeq->last_sq_tail)
507 		nvme_write_sq_db(nvmeq, true);
508 	spin_unlock(&nvmeq->sq_lock);
509 }
510 
511 static void **nvme_pci_iod_list(struct request *req)
512 {
513 	struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
514 	return (void **)(iod->sg + blk_rq_nr_phys_segments(req));
515 }
516 
517 static inline bool nvme_pci_use_sgls(struct nvme_dev *dev, struct request *req)
518 {
519 	struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
520 	int nseg = blk_rq_nr_phys_segments(req);
521 	unsigned int avg_seg_size;
522 
523 	if (nseg == 0)
524 		return false;
525 
526 	avg_seg_size = DIV_ROUND_UP(blk_rq_payload_bytes(req), nseg);
527 
528 	if (!(dev->ctrl.sgls & ((1 << 0) | (1 << 1))))
529 		return false;
530 	if (!iod->nvmeq->qid)
531 		return false;
532 	if (!sgl_threshold || avg_seg_size < sgl_threshold)
533 		return false;
534 	return true;
535 }
536 
537 static void nvme_unmap_data(struct nvme_dev *dev, struct request *req)
538 {
539 	struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
540 	const int last_prp = dev->ctrl.page_size / sizeof(__le64) - 1;
541 	dma_addr_t dma_addr = iod->first_dma, next_dma_addr;
542 	int i;
543 
544 	if (iod->dma_len) {
545 		dma_unmap_page(dev->dev, dma_addr, iod->dma_len,
546 			       rq_dma_dir(req));
547 		return;
548 	}
549 
550 	WARN_ON_ONCE(!iod->nents);
551 
552 	if (is_pci_p2pdma_page(sg_page(iod->sg)))
553 		pci_p2pdma_unmap_sg(dev->dev, iod->sg, iod->nents,
554 				    rq_dma_dir(req));
555 	else
556 		dma_unmap_sg(dev->dev, iod->sg, iod->nents, rq_dma_dir(req));
557 
558 
559 	if (iod->npages == 0)
560 		dma_pool_free(dev->prp_small_pool, nvme_pci_iod_list(req)[0],
561 			dma_addr);
562 
563 	for (i = 0; i < iod->npages; i++) {
564 		void *addr = nvme_pci_iod_list(req)[i];
565 
566 		if (iod->use_sgl) {
567 			struct nvme_sgl_desc *sg_list = addr;
568 
569 			next_dma_addr =
570 			    le64_to_cpu((sg_list[SGES_PER_PAGE - 1]).addr);
571 		} else {
572 			__le64 *prp_list = addr;
573 
574 			next_dma_addr = le64_to_cpu(prp_list[last_prp]);
575 		}
576 
577 		dma_pool_free(dev->prp_page_pool, addr, dma_addr);
578 		dma_addr = next_dma_addr;
579 	}
580 
581 	mempool_free(iod->sg, dev->iod_mempool);
582 }
583 
584 static void nvme_print_sgl(struct scatterlist *sgl, int nents)
585 {
586 	int i;
587 	struct scatterlist *sg;
588 
589 	for_each_sg(sgl, sg, nents, i) {
590 		dma_addr_t phys = sg_phys(sg);
591 		pr_warn("sg[%d] phys_addr:%pad offset:%d length:%d "
592 			"dma_address:%pad dma_length:%d\n",
593 			i, &phys, sg->offset, sg->length, &sg_dma_address(sg),
594 			sg_dma_len(sg));
595 	}
596 }
597 
598 static blk_status_t nvme_pci_setup_prps(struct nvme_dev *dev,
599 		struct request *req, struct nvme_rw_command *cmnd)
600 {
601 	struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
602 	struct dma_pool *pool;
603 	int length = blk_rq_payload_bytes(req);
604 	struct scatterlist *sg = iod->sg;
605 	int dma_len = sg_dma_len(sg);
606 	u64 dma_addr = sg_dma_address(sg);
607 	u32 page_size = dev->ctrl.page_size;
608 	int offset = dma_addr & (page_size - 1);
609 	__le64 *prp_list;
610 	void **list = nvme_pci_iod_list(req);
611 	dma_addr_t prp_dma;
612 	int nprps, i;
613 
614 	length -= (page_size - offset);
615 	if (length <= 0) {
616 		iod->first_dma = 0;
617 		goto done;
618 	}
619 
620 	dma_len -= (page_size - offset);
621 	if (dma_len) {
622 		dma_addr += (page_size - offset);
623 	} else {
624 		sg = sg_next(sg);
625 		dma_addr = sg_dma_address(sg);
626 		dma_len = sg_dma_len(sg);
627 	}
628 
629 	if (length <= page_size) {
630 		iod->first_dma = dma_addr;
631 		goto done;
632 	}
633 
634 	nprps = DIV_ROUND_UP(length, page_size);
635 	if (nprps <= (256 / 8)) {
636 		pool = dev->prp_small_pool;
637 		iod->npages = 0;
638 	} else {
639 		pool = dev->prp_page_pool;
640 		iod->npages = 1;
641 	}
642 
643 	prp_list = dma_pool_alloc(pool, GFP_ATOMIC, &prp_dma);
644 	if (!prp_list) {
645 		iod->first_dma = dma_addr;
646 		iod->npages = -1;
647 		return BLK_STS_RESOURCE;
648 	}
649 	list[0] = prp_list;
650 	iod->first_dma = prp_dma;
651 	i = 0;
652 	for (;;) {
653 		if (i == page_size >> 3) {
654 			__le64 *old_prp_list = prp_list;
655 			prp_list = dma_pool_alloc(pool, GFP_ATOMIC, &prp_dma);
656 			if (!prp_list)
657 				return BLK_STS_RESOURCE;
658 			list[iod->npages++] = prp_list;
659 			prp_list[0] = old_prp_list[i - 1];
660 			old_prp_list[i - 1] = cpu_to_le64(prp_dma);
661 			i = 1;
662 		}
663 		prp_list[i++] = cpu_to_le64(dma_addr);
664 		dma_len -= page_size;
665 		dma_addr += page_size;
666 		length -= page_size;
667 		if (length <= 0)
668 			break;
669 		if (dma_len > 0)
670 			continue;
671 		if (unlikely(dma_len < 0))
672 			goto bad_sgl;
673 		sg = sg_next(sg);
674 		dma_addr = sg_dma_address(sg);
675 		dma_len = sg_dma_len(sg);
676 	}
677 
678 done:
679 	cmnd->dptr.prp1 = cpu_to_le64(sg_dma_address(iod->sg));
680 	cmnd->dptr.prp2 = cpu_to_le64(iod->first_dma);
681 
682 	return BLK_STS_OK;
683 
684  bad_sgl:
685 	WARN(DO_ONCE(nvme_print_sgl, iod->sg, iod->nents),
686 			"Invalid SGL for payload:%d nents:%d\n",
687 			blk_rq_payload_bytes(req), iod->nents);
688 	return BLK_STS_IOERR;
689 }
690 
691 static void nvme_pci_sgl_set_data(struct nvme_sgl_desc *sge,
692 		struct scatterlist *sg)
693 {
694 	sge->addr = cpu_to_le64(sg_dma_address(sg));
695 	sge->length = cpu_to_le32(sg_dma_len(sg));
696 	sge->type = NVME_SGL_FMT_DATA_DESC << 4;
697 }
698 
699 static void nvme_pci_sgl_set_seg(struct nvme_sgl_desc *sge,
700 		dma_addr_t dma_addr, int entries)
701 {
702 	sge->addr = cpu_to_le64(dma_addr);
703 	if (entries < SGES_PER_PAGE) {
704 		sge->length = cpu_to_le32(entries * sizeof(*sge));
705 		sge->type = NVME_SGL_FMT_LAST_SEG_DESC << 4;
706 	} else {
707 		sge->length = cpu_to_le32(PAGE_SIZE);
708 		sge->type = NVME_SGL_FMT_SEG_DESC << 4;
709 	}
710 }
711 
712 static blk_status_t nvme_pci_setup_sgls(struct nvme_dev *dev,
713 		struct request *req, struct nvme_rw_command *cmd, int entries)
714 {
715 	struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
716 	struct dma_pool *pool;
717 	struct nvme_sgl_desc *sg_list;
718 	struct scatterlist *sg = iod->sg;
719 	dma_addr_t sgl_dma;
720 	int i = 0;
721 
722 	/* setting the transfer type as SGL */
723 	cmd->flags = NVME_CMD_SGL_METABUF;
724 
725 	if (entries == 1) {
726 		nvme_pci_sgl_set_data(&cmd->dptr.sgl, sg);
727 		return BLK_STS_OK;
728 	}
729 
730 	if (entries <= (256 / sizeof(struct nvme_sgl_desc))) {
731 		pool = dev->prp_small_pool;
732 		iod->npages = 0;
733 	} else {
734 		pool = dev->prp_page_pool;
735 		iod->npages = 1;
736 	}
737 
738 	sg_list = dma_pool_alloc(pool, GFP_ATOMIC, &sgl_dma);
739 	if (!sg_list) {
740 		iod->npages = -1;
741 		return BLK_STS_RESOURCE;
742 	}
743 
744 	nvme_pci_iod_list(req)[0] = sg_list;
745 	iod->first_dma = sgl_dma;
746 
747 	nvme_pci_sgl_set_seg(&cmd->dptr.sgl, sgl_dma, entries);
748 
749 	do {
750 		if (i == SGES_PER_PAGE) {
751 			struct nvme_sgl_desc *old_sg_desc = sg_list;
752 			struct nvme_sgl_desc *link = &old_sg_desc[i - 1];
753 
754 			sg_list = dma_pool_alloc(pool, GFP_ATOMIC, &sgl_dma);
755 			if (!sg_list)
756 				return BLK_STS_RESOURCE;
757 
758 			i = 0;
759 			nvme_pci_iod_list(req)[iod->npages++] = sg_list;
760 			sg_list[i++] = *link;
761 			nvme_pci_sgl_set_seg(link, sgl_dma, entries);
762 		}
763 
764 		nvme_pci_sgl_set_data(&sg_list[i++], sg);
765 		sg = sg_next(sg);
766 	} while (--entries > 0);
767 
768 	return BLK_STS_OK;
769 }
770 
771 static blk_status_t nvme_setup_prp_simple(struct nvme_dev *dev,
772 		struct request *req, struct nvme_rw_command *cmnd,
773 		struct bio_vec *bv)
774 {
775 	struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
776 	unsigned int offset = bv->bv_offset & (dev->ctrl.page_size - 1);
777 	unsigned int first_prp_len = dev->ctrl.page_size - offset;
778 
779 	iod->first_dma = dma_map_bvec(dev->dev, bv, rq_dma_dir(req), 0);
780 	if (dma_mapping_error(dev->dev, iod->first_dma))
781 		return BLK_STS_RESOURCE;
782 	iod->dma_len = bv->bv_len;
783 
784 	cmnd->dptr.prp1 = cpu_to_le64(iod->first_dma);
785 	if (bv->bv_len > first_prp_len)
786 		cmnd->dptr.prp2 = cpu_to_le64(iod->first_dma + first_prp_len);
787 	return 0;
788 }
789 
790 static blk_status_t nvme_setup_sgl_simple(struct nvme_dev *dev,
791 		struct request *req, struct nvme_rw_command *cmnd,
792 		struct bio_vec *bv)
793 {
794 	struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
795 
796 	iod->first_dma = dma_map_bvec(dev->dev, bv, rq_dma_dir(req), 0);
797 	if (dma_mapping_error(dev->dev, iod->first_dma))
798 		return BLK_STS_RESOURCE;
799 	iod->dma_len = bv->bv_len;
800 
801 	cmnd->flags = NVME_CMD_SGL_METABUF;
802 	cmnd->dptr.sgl.addr = cpu_to_le64(iod->first_dma);
803 	cmnd->dptr.sgl.length = cpu_to_le32(iod->dma_len);
804 	cmnd->dptr.sgl.type = NVME_SGL_FMT_DATA_DESC << 4;
805 	return 0;
806 }
807 
808 static blk_status_t nvme_map_data(struct nvme_dev *dev, struct request *req,
809 		struct nvme_command *cmnd)
810 {
811 	struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
812 	blk_status_t ret = BLK_STS_RESOURCE;
813 	int nr_mapped;
814 
815 	if (blk_rq_nr_phys_segments(req) == 1) {
816 		struct bio_vec bv = req_bvec(req);
817 
818 		if (!is_pci_p2pdma_page(bv.bv_page)) {
819 			if (bv.bv_offset + bv.bv_len <= dev->ctrl.page_size * 2)
820 				return nvme_setup_prp_simple(dev, req,
821 							     &cmnd->rw, &bv);
822 
823 			if (iod->nvmeq->qid &&
824 			    dev->ctrl.sgls & ((1 << 0) | (1 << 1)))
825 				return nvme_setup_sgl_simple(dev, req,
826 							     &cmnd->rw, &bv);
827 		}
828 	}
829 
830 	iod->dma_len = 0;
831 	iod->sg = mempool_alloc(dev->iod_mempool, GFP_ATOMIC);
832 	if (!iod->sg)
833 		return BLK_STS_RESOURCE;
834 	sg_init_table(iod->sg, blk_rq_nr_phys_segments(req));
835 	iod->nents = blk_rq_map_sg(req->q, req, iod->sg);
836 	if (!iod->nents)
837 		goto out;
838 
839 	if (is_pci_p2pdma_page(sg_page(iod->sg)))
840 		nr_mapped = pci_p2pdma_map_sg_attrs(dev->dev, iod->sg,
841 				iod->nents, rq_dma_dir(req), DMA_ATTR_NO_WARN);
842 	else
843 		nr_mapped = dma_map_sg_attrs(dev->dev, iod->sg, iod->nents,
844 					     rq_dma_dir(req), DMA_ATTR_NO_WARN);
845 	if (!nr_mapped)
846 		goto out;
847 
848 	iod->use_sgl = nvme_pci_use_sgls(dev, req);
849 	if (iod->use_sgl)
850 		ret = nvme_pci_setup_sgls(dev, req, &cmnd->rw, nr_mapped);
851 	else
852 		ret = nvme_pci_setup_prps(dev, req, &cmnd->rw);
853 out:
854 	if (ret != BLK_STS_OK)
855 		nvme_unmap_data(dev, req);
856 	return ret;
857 }
858 
859 static blk_status_t nvme_map_metadata(struct nvme_dev *dev, struct request *req,
860 		struct nvme_command *cmnd)
861 {
862 	struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
863 
864 	iod->meta_dma = dma_map_bvec(dev->dev, rq_integrity_vec(req),
865 			rq_dma_dir(req), 0);
866 	if (dma_mapping_error(dev->dev, iod->meta_dma))
867 		return BLK_STS_IOERR;
868 	cmnd->rw.metadata = cpu_to_le64(iod->meta_dma);
869 	return 0;
870 }
871 
872 /*
873  * NOTE: ns is NULL when called on the admin queue.
874  */
875 static blk_status_t nvme_queue_rq(struct blk_mq_hw_ctx *hctx,
876 			 const struct blk_mq_queue_data *bd)
877 {
878 	struct nvme_ns *ns = hctx->queue->queuedata;
879 	struct nvme_queue *nvmeq = hctx->driver_data;
880 	struct nvme_dev *dev = nvmeq->dev;
881 	struct request *req = bd->rq;
882 	struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
883 	struct nvme_command cmnd;
884 	blk_status_t ret;
885 
886 	iod->aborted = 0;
887 	iod->npages = -1;
888 	iod->nents = 0;
889 
890 	/*
891 	 * We should not need to do this, but we're still using this to
892 	 * ensure we can drain requests on a dying queue.
893 	 */
894 	if (unlikely(!test_bit(NVMEQ_ENABLED, &nvmeq->flags)))
895 		return BLK_STS_IOERR;
896 
897 	ret = nvme_setup_cmd(ns, req, &cmnd);
898 	if (ret)
899 		return ret;
900 
901 	if (blk_rq_nr_phys_segments(req)) {
902 		ret = nvme_map_data(dev, req, &cmnd);
903 		if (ret)
904 			goto out_free_cmd;
905 	}
906 
907 	if (blk_integrity_rq(req)) {
908 		ret = nvme_map_metadata(dev, req, &cmnd);
909 		if (ret)
910 			goto out_unmap_data;
911 	}
912 
913 	blk_mq_start_request(req);
914 	nvme_submit_cmd(nvmeq, &cmnd, bd->last);
915 	return BLK_STS_OK;
916 out_unmap_data:
917 	nvme_unmap_data(dev, req);
918 out_free_cmd:
919 	nvme_cleanup_cmd(req);
920 	return ret;
921 }
922 
923 static void nvme_pci_complete_rq(struct request *req)
924 {
925 	struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
926 	struct nvme_dev *dev = iod->nvmeq->dev;
927 
928 	nvme_cleanup_cmd(req);
929 	if (blk_integrity_rq(req))
930 		dma_unmap_page(dev->dev, iod->meta_dma,
931 			       rq_integrity_vec(req)->bv_len, rq_data_dir(req));
932 	if (blk_rq_nr_phys_segments(req))
933 		nvme_unmap_data(dev, req);
934 	nvme_complete_rq(req);
935 }
936 
937 /* We read the CQE phase first to check if the rest of the entry is valid */
938 static inline bool nvme_cqe_pending(struct nvme_queue *nvmeq)
939 {
940 	return (le16_to_cpu(nvmeq->cqes[nvmeq->cq_head].status) & 1) ==
941 			nvmeq->cq_phase;
942 }
943 
944 static inline void nvme_ring_cq_doorbell(struct nvme_queue *nvmeq)
945 {
946 	u16 head = nvmeq->cq_head;
947 
948 	if (nvme_dbbuf_update_and_check_event(head, nvmeq->dbbuf_cq_db,
949 					      nvmeq->dbbuf_cq_ei))
950 		writel(head, nvmeq->q_db + nvmeq->dev->db_stride);
951 }
952 
953 static inline void nvme_handle_cqe(struct nvme_queue *nvmeq, u16 idx)
954 {
955 	volatile struct nvme_completion *cqe = &nvmeq->cqes[idx];
956 	struct request *req;
957 
958 	if (unlikely(cqe->command_id >= nvmeq->q_depth)) {
959 		dev_warn(nvmeq->dev->ctrl.device,
960 			"invalid id %d completed on queue %d\n",
961 			cqe->command_id, le16_to_cpu(cqe->sq_id));
962 		return;
963 	}
964 
965 	/*
966 	 * AEN requests are special as they don't time out and can
967 	 * survive any kind of queue freeze and often don't respond to
968 	 * aborts.  We don't even bother to allocate a struct request
969 	 * for them but rather special case them here.
970 	 */
971 	if (unlikely(nvmeq->qid == 0 &&
972 			cqe->command_id >= NVME_AQ_BLK_MQ_DEPTH)) {
973 		nvme_complete_async_event(&nvmeq->dev->ctrl,
974 				cqe->status, &cqe->result);
975 		return;
976 	}
977 
978 	req = blk_mq_tag_to_rq(*nvmeq->tags, cqe->command_id);
979 	trace_nvme_sq(req, cqe->sq_head, nvmeq->sq_tail);
980 	nvme_end_request(req, cqe->status, cqe->result);
981 }
982 
983 static void nvme_complete_cqes(struct nvme_queue *nvmeq, u16 start, u16 end)
984 {
985 	while (start != end) {
986 		nvme_handle_cqe(nvmeq, start);
987 		if (++start == nvmeq->q_depth)
988 			start = 0;
989 	}
990 }
991 
992 static inline void nvme_update_cq_head(struct nvme_queue *nvmeq)
993 {
994 	if (nvmeq->cq_head == nvmeq->q_depth - 1) {
995 		nvmeq->cq_head = 0;
996 		nvmeq->cq_phase = !nvmeq->cq_phase;
997 	} else {
998 		nvmeq->cq_head++;
999 	}
1000 }
1001 
1002 static inline int nvme_process_cq(struct nvme_queue *nvmeq, u16 *start,
1003 				  u16 *end, unsigned int tag)
1004 {
1005 	int found = 0;
1006 
1007 	*start = nvmeq->cq_head;
1008 	while (nvme_cqe_pending(nvmeq)) {
1009 		if (tag == -1U || nvmeq->cqes[nvmeq->cq_head].command_id == tag)
1010 			found++;
1011 		nvme_update_cq_head(nvmeq);
1012 	}
1013 	*end = nvmeq->cq_head;
1014 
1015 	if (*start != *end)
1016 		nvme_ring_cq_doorbell(nvmeq);
1017 	return found;
1018 }
1019 
1020 static irqreturn_t nvme_irq(int irq, void *data)
1021 {
1022 	struct nvme_queue *nvmeq = data;
1023 	irqreturn_t ret = IRQ_NONE;
1024 	u16 start, end;
1025 
1026 	/*
1027 	 * The rmb/wmb pair ensures we see all updates from a previous run of
1028 	 * the irq handler, even if that was on another CPU.
1029 	 */
1030 	rmb();
1031 	if (nvmeq->cq_head != nvmeq->last_cq_head)
1032 		ret = IRQ_HANDLED;
1033 	nvme_process_cq(nvmeq, &start, &end, -1);
1034 	nvmeq->last_cq_head = nvmeq->cq_head;
1035 	wmb();
1036 
1037 	if (start != end) {
1038 		nvme_complete_cqes(nvmeq, start, end);
1039 		return IRQ_HANDLED;
1040 	}
1041 
1042 	return ret;
1043 }
1044 
1045 static irqreturn_t nvme_irq_check(int irq, void *data)
1046 {
1047 	struct nvme_queue *nvmeq = data;
1048 	if (nvme_cqe_pending(nvmeq))
1049 		return IRQ_WAKE_THREAD;
1050 	return IRQ_NONE;
1051 }
1052 
1053 /*
1054  * Poll for completions any queue, including those not dedicated to polling.
1055  * Can be called from any context.
1056  */
1057 static int nvme_poll_irqdisable(struct nvme_queue *nvmeq, unsigned int tag)
1058 {
1059 	struct pci_dev *pdev = to_pci_dev(nvmeq->dev->dev);
1060 	u16 start, end;
1061 	int found;
1062 
1063 	/*
1064 	 * For a poll queue we need to protect against the polling thread
1065 	 * using the CQ lock.  For normal interrupt driven threads we have
1066 	 * to disable the interrupt to avoid racing with it.
1067 	 */
1068 	if (test_bit(NVMEQ_POLLED, &nvmeq->flags)) {
1069 		spin_lock(&nvmeq->cq_poll_lock);
1070 		found = nvme_process_cq(nvmeq, &start, &end, tag);
1071 		spin_unlock(&nvmeq->cq_poll_lock);
1072 	} else {
1073 		disable_irq(pci_irq_vector(pdev, nvmeq->cq_vector));
1074 		found = nvme_process_cq(nvmeq, &start, &end, tag);
1075 		enable_irq(pci_irq_vector(pdev, nvmeq->cq_vector));
1076 	}
1077 
1078 	nvme_complete_cqes(nvmeq, start, end);
1079 	return found;
1080 }
1081 
1082 static int nvme_poll(struct blk_mq_hw_ctx *hctx)
1083 {
1084 	struct nvme_queue *nvmeq = hctx->driver_data;
1085 	u16 start, end;
1086 	bool found;
1087 
1088 	if (!nvme_cqe_pending(nvmeq))
1089 		return 0;
1090 
1091 	spin_lock(&nvmeq->cq_poll_lock);
1092 	found = nvme_process_cq(nvmeq, &start, &end, -1);
1093 	spin_unlock(&nvmeq->cq_poll_lock);
1094 
1095 	nvme_complete_cqes(nvmeq, start, end);
1096 	return found;
1097 }
1098 
1099 static void nvme_pci_submit_async_event(struct nvme_ctrl *ctrl)
1100 {
1101 	struct nvme_dev *dev = to_nvme_dev(ctrl);
1102 	struct nvme_queue *nvmeq = &dev->queues[0];
1103 	struct nvme_command c;
1104 
1105 	memset(&c, 0, sizeof(c));
1106 	c.common.opcode = nvme_admin_async_event;
1107 	c.common.command_id = NVME_AQ_BLK_MQ_DEPTH;
1108 	nvme_submit_cmd(nvmeq, &c, true);
1109 }
1110 
1111 static int adapter_delete_queue(struct nvme_dev *dev, u8 opcode, u16 id)
1112 {
1113 	struct nvme_command c;
1114 
1115 	memset(&c, 0, sizeof(c));
1116 	c.delete_queue.opcode = opcode;
1117 	c.delete_queue.qid = cpu_to_le16(id);
1118 
1119 	return nvme_submit_sync_cmd(dev->ctrl.admin_q, &c, NULL, 0);
1120 }
1121 
1122 static int adapter_alloc_cq(struct nvme_dev *dev, u16 qid,
1123 		struct nvme_queue *nvmeq, s16 vector)
1124 {
1125 	struct nvme_command c;
1126 	int flags = NVME_QUEUE_PHYS_CONTIG;
1127 
1128 	if (!test_bit(NVMEQ_POLLED, &nvmeq->flags))
1129 		flags |= NVME_CQ_IRQ_ENABLED;
1130 
1131 	/*
1132 	 * Note: we (ab)use the fact that the prp fields survive if no data
1133 	 * is attached to the request.
1134 	 */
1135 	memset(&c, 0, sizeof(c));
1136 	c.create_cq.opcode = nvme_admin_create_cq;
1137 	c.create_cq.prp1 = cpu_to_le64(nvmeq->cq_dma_addr);
1138 	c.create_cq.cqid = cpu_to_le16(qid);
1139 	c.create_cq.qsize = cpu_to_le16(nvmeq->q_depth - 1);
1140 	c.create_cq.cq_flags = cpu_to_le16(flags);
1141 	c.create_cq.irq_vector = cpu_to_le16(vector);
1142 
1143 	return nvme_submit_sync_cmd(dev->ctrl.admin_q, &c, NULL, 0);
1144 }
1145 
1146 static int adapter_alloc_sq(struct nvme_dev *dev, u16 qid,
1147 						struct nvme_queue *nvmeq)
1148 {
1149 	struct nvme_ctrl *ctrl = &dev->ctrl;
1150 	struct nvme_command c;
1151 	int flags = NVME_QUEUE_PHYS_CONTIG;
1152 
1153 	/*
1154 	 * Some drives have a bug that auto-enables WRRU if MEDIUM isn't
1155 	 * set. Since URGENT priority is zeroes, it makes all queues
1156 	 * URGENT.
1157 	 */
1158 	if (ctrl->quirks & NVME_QUIRK_MEDIUM_PRIO_SQ)
1159 		flags |= NVME_SQ_PRIO_MEDIUM;
1160 
1161 	/*
1162 	 * Note: we (ab)use the fact that the prp fields survive if no data
1163 	 * is attached to the request.
1164 	 */
1165 	memset(&c, 0, sizeof(c));
1166 	c.create_sq.opcode = nvme_admin_create_sq;
1167 	c.create_sq.prp1 = cpu_to_le64(nvmeq->sq_dma_addr);
1168 	c.create_sq.sqid = cpu_to_le16(qid);
1169 	c.create_sq.qsize = cpu_to_le16(nvmeq->q_depth - 1);
1170 	c.create_sq.sq_flags = cpu_to_le16(flags);
1171 	c.create_sq.cqid = cpu_to_le16(qid);
1172 
1173 	return nvme_submit_sync_cmd(dev->ctrl.admin_q, &c, NULL, 0);
1174 }
1175 
1176 static int adapter_delete_cq(struct nvme_dev *dev, u16 cqid)
1177 {
1178 	return adapter_delete_queue(dev, nvme_admin_delete_cq, cqid);
1179 }
1180 
1181 static int adapter_delete_sq(struct nvme_dev *dev, u16 sqid)
1182 {
1183 	return adapter_delete_queue(dev, nvme_admin_delete_sq, sqid);
1184 }
1185 
1186 static void abort_endio(struct request *req, blk_status_t error)
1187 {
1188 	struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
1189 	struct nvme_queue *nvmeq = iod->nvmeq;
1190 
1191 	dev_warn(nvmeq->dev->ctrl.device,
1192 		 "Abort status: 0x%x", nvme_req(req)->status);
1193 	atomic_inc(&nvmeq->dev->ctrl.abort_limit);
1194 	blk_mq_free_request(req);
1195 }
1196 
1197 static bool nvme_should_reset(struct nvme_dev *dev, u32 csts)
1198 {
1199 
1200 	/* If true, indicates loss of adapter communication, possibly by a
1201 	 * NVMe Subsystem reset.
1202 	 */
1203 	bool nssro = dev->subsystem && (csts & NVME_CSTS_NSSRO);
1204 
1205 	/* If there is a reset/reinit ongoing, we shouldn't reset again. */
1206 	switch (dev->ctrl.state) {
1207 	case NVME_CTRL_RESETTING:
1208 	case NVME_CTRL_CONNECTING:
1209 		return false;
1210 	default:
1211 		break;
1212 	}
1213 
1214 	/* We shouldn't reset unless the controller is on fatal error state
1215 	 * _or_ if we lost the communication with it.
1216 	 */
1217 	if (!(csts & NVME_CSTS_CFS) && !nssro)
1218 		return false;
1219 
1220 	return true;
1221 }
1222 
1223 static void nvme_warn_reset(struct nvme_dev *dev, u32 csts)
1224 {
1225 	/* Read a config register to help see what died. */
1226 	u16 pci_status;
1227 	int result;
1228 
1229 	result = pci_read_config_word(to_pci_dev(dev->dev), PCI_STATUS,
1230 				      &pci_status);
1231 	if (result == PCIBIOS_SUCCESSFUL)
1232 		dev_warn(dev->ctrl.device,
1233 			 "controller is down; will reset: CSTS=0x%x, PCI_STATUS=0x%hx\n",
1234 			 csts, pci_status);
1235 	else
1236 		dev_warn(dev->ctrl.device,
1237 			 "controller is down; will reset: CSTS=0x%x, PCI_STATUS read failed (%d)\n",
1238 			 csts, result);
1239 }
1240 
1241 static enum blk_eh_timer_return nvme_timeout(struct request *req, bool reserved)
1242 {
1243 	struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
1244 	struct nvme_queue *nvmeq = iod->nvmeq;
1245 	struct nvme_dev *dev = nvmeq->dev;
1246 	struct request *abort_req;
1247 	struct nvme_command cmd;
1248 	u32 csts = readl(dev->bar + NVME_REG_CSTS);
1249 
1250 	/* If PCI error recovery process is happening, we cannot reset or
1251 	 * the recovery mechanism will surely fail.
1252 	 */
1253 	mb();
1254 	if (pci_channel_offline(to_pci_dev(dev->dev)))
1255 		return BLK_EH_RESET_TIMER;
1256 
1257 	/*
1258 	 * Reset immediately if the controller is failed
1259 	 */
1260 	if (nvme_should_reset(dev, csts)) {
1261 		nvme_warn_reset(dev, csts);
1262 		nvme_dev_disable(dev, false);
1263 		nvme_reset_ctrl(&dev->ctrl);
1264 		return BLK_EH_DONE;
1265 	}
1266 
1267 	/*
1268 	 * Did we miss an interrupt?
1269 	 */
1270 	if (nvme_poll_irqdisable(nvmeq, req->tag)) {
1271 		dev_warn(dev->ctrl.device,
1272 			 "I/O %d QID %d timeout, completion polled\n",
1273 			 req->tag, nvmeq->qid);
1274 		return BLK_EH_DONE;
1275 	}
1276 
1277 	/*
1278 	 * Shutdown immediately if controller times out while starting. The
1279 	 * reset work will see the pci device disabled when it gets the forced
1280 	 * cancellation error. All outstanding requests are completed on
1281 	 * shutdown, so we return BLK_EH_DONE.
1282 	 */
1283 	switch (dev->ctrl.state) {
1284 	case NVME_CTRL_CONNECTING:
1285 		nvme_change_ctrl_state(&dev->ctrl, NVME_CTRL_DELETING);
1286 		/* fall through */
1287 	case NVME_CTRL_DELETING:
1288 		dev_warn_ratelimited(dev->ctrl.device,
1289 			 "I/O %d QID %d timeout, disable controller\n",
1290 			 req->tag, nvmeq->qid);
1291 		nvme_dev_disable(dev, true);
1292 		nvme_req(req)->flags |= NVME_REQ_CANCELLED;
1293 		return BLK_EH_DONE;
1294 	case NVME_CTRL_RESETTING:
1295 		return BLK_EH_RESET_TIMER;
1296 	default:
1297 		break;
1298 	}
1299 
1300 	/*
1301  	 * Shutdown the controller immediately and schedule a reset if the
1302  	 * command was already aborted once before and still hasn't been
1303  	 * returned to the driver, or if this is the admin queue.
1304 	 */
1305 	if (!nvmeq->qid || iod->aborted) {
1306 		dev_warn(dev->ctrl.device,
1307 			 "I/O %d QID %d timeout, reset controller\n",
1308 			 req->tag, nvmeq->qid);
1309 		nvme_dev_disable(dev, false);
1310 		nvme_reset_ctrl(&dev->ctrl);
1311 
1312 		nvme_req(req)->flags |= NVME_REQ_CANCELLED;
1313 		return BLK_EH_DONE;
1314 	}
1315 
1316 	if (atomic_dec_return(&dev->ctrl.abort_limit) < 0) {
1317 		atomic_inc(&dev->ctrl.abort_limit);
1318 		return BLK_EH_RESET_TIMER;
1319 	}
1320 	iod->aborted = 1;
1321 
1322 	memset(&cmd, 0, sizeof(cmd));
1323 	cmd.abort.opcode = nvme_admin_abort_cmd;
1324 	cmd.abort.cid = req->tag;
1325 	cmd.abort.sqid = cpu_to_le16(nvmeq->qid);
1326 
1327 	dev_warn(nvmeq->dev->ctrl.device,
1328 		"I/O %d QID %d timeout, aborting\n",
1329 		 req->tag, nvmeq->qid);
1330 
1331 	abort_req = nvme_alloc_request(dev->ctrl.admin_q, &cmd,
1332 			BLK_MQ_REQ_NOWAIT, NVME_QID_ANY);
1333 	if (IS_ERR(abort_req)) {
1334 		atomic_inc(&dev->ctrl.abort_limit);
1335 		return BLK_EH_RESET_TIMER;
1336 	}
1337 
1338 	abort_req->timeout = ADMIN_TIMEOUT;
1339 	abort_req->end_io_data = NULL;
1340 	blk_execute_rq_nowait(abort_req->q, NULL, abort_req, 0, abort_endio);
1341 
1342 	/*
1343 	 * The aborted req will be completed on receiving the abort req.
1344 	 * We enable the timer again. If hit twice, it'll cause a device reset,
1345 	 * as the device then is in a faulty state.
1346 	 */
1347 	return BLK_EH_RESET_TIMER;
1348 }
1349 
1350 static void nvme_free_queue(struct nvme_queue *nvmeq)
1351 {
1352 	dma_free_coherent(nvmeq->dev->dev, CQ_SIZE(nvmeq),
1353 				(void *)nvmeq->cqes, nvmeq->cq_dma_addr);
1354 	if (!nvmeq->sq_cmds)
1355 		return;
1356 
1357 	if (test_and_clear_bit(NVMEQ_SQ_CMB, &nvmeq->flags)) {
1358 		pci_free_p2pmem(to_pci_dev(nvmeq->dev->dev),
1359 				nvmeq->sq_cmds, SQ_SIZE(nvmeq));
1360 	} else {
1361 		dma_free_coherent(nvmeq->dev->dev, SQ_SIZE(nvmeq),
1362 				nvmeq->sq_cmds, nvmeq->sq_dma_addr);
1363 	}
1364 }
1365 
1366 static void nvme_free_queues(struct nvme_dev *dev, int lowest)
1367 {
1368 	int i;
1369 
1370 	for (i = dev->ctrl.queue_count - 1; i >= lowest; i--) {
1371 		dev->ctrl.queue_count--;
1372 		nvme_free_queue(&dev->queues[i]);
1373 	}
1374 }
1375 
1376 /**
1377  * nvme_suspend_queue - put queue into suspended state
1378  * @nvmeq: queue to suspend
1379  */
1380 static int nvme_suspend_queue(struct nvme_queue *nvmeq)
1381 {
1382 	if (!test_and_clear_bit(NVMEQ_ENABLED, &nvmeq->flags))
1383 		return 1;
1384 
1385 	/* ensure that nvme_queue_rq() sees NVMEQ_ENABLED cleared */
1386 	mb();
1387 
1388 	nvmeq->dev->online_queues--;
1389 	if (!nvmeq->qid && nvmeq->dev->ctrl.admin_q)
1390 		blk_mq_quiesce_queue(nvmeq->dev->ctrl.admin_q);
1391 	if (!test_and_clear_bit(NVMEQ_POLLED, &nvmeq->flags))
1392 		pci_free_irq(to_pci_dev(nvmeq->dev->dev), nvmeq->cq_vector, nvmeq);
1393 	return 0;
1394 }
1395 
1396 static void nvme_suspend_io_queues(struct nvme_dev *dev)
1397 {
1398 	int i;
1399 
1400 	for (i = dev->ctrl.queue_count - 1; i > 0; i--)
1401 		nvme_suspend_queue(&dev->queues[i]);
1402 }
1403 
1404 static void nvme_disable_admin_queue(struct nvme_dev *dev, bool shutdown)
1405 {
1406 	struct nvme_queue *nvmeq = &dev->queues[0];
1407 
1408 	if (shutdown)
1409 		nvme_shutdown_ctrl(&dev->ctrl);
1410 	else
1411 		nvme_disable_ctrl(&dev->ctrl);
1412 
1413 	nvme_poll_irqdisable(nvmeq, -1);
1414 }
1415 
1416 static int nvme_cmb_qdepth(struct nvme_dev *dev, int nr_io_queues,
1417 				int entry_size)
1418 {
1419 	int q_depth = dev->q_depth;
1420 	unsigned q_size_aligned = roundup(q_depth * entry_size,
1421 					  dev->ctrl.page_size);
1422 
1423 	if (q_size_aligned * nr_io_queues > dev->cmb_size) {
1424 		u64 mem_per_q = div_u64(dev->cmb_size, nr_io_queues);
1425 		mem_per_q = round_down(mem_per_q, dev->ctrl.page_size);
1426 		q_depth = div_u64(mem_per_q, entry_size);
1427 
1428 		/*
1429 		 * Ensure the reduced q_depth is above some threshold where it
1430 		 * would be better to map queues in system memory with the
1431 		 * original depth
1432 		 */
1433 		if (q_depth < 64)
1434 			return -ENOMEM;
1435 	}
1436 
1437 	return q_depth;
1438 }
1439 
1440 static int nvme_alloc_sq_cmds(struct nvme_dev *dev, struct nvme_queue *nvmeq,
1441 				int qid)
1442 {
1443 	struct pci_dev *pdev = to_pci_dev(dev->dev);
1444 
1445 	if (qid && dev->cmb_use_sqes && (dev->cmbsz & NVME_CMBSZ_SQS)) {
1446 		nvmeq->sq_cmds = pci_alloc_p2pmem(pdev, SQ_SIZE(nvmeq));
1447 		if (nvmeq->sq_cmds) {
1448 			nvmeq->sq_dma_addr = pci_p2pmem_virt_to_bus(pdev,
1449 							nvmeq->sq_cmds);
1450 			if (nvmeq->sq_dma_addr) {
1451 				set_bit(NVMEQ_SQ_CMB, &nvmeq->flags);
1452 				return 0;
1453 			}
1454 
1455 			pci_free_p2pmem(pdev, nvmeq->sq_cmds, SQ_SIZE(nvmeq));
1456 		}
1457 	}
1458 
1459 	nvmeq->sq_cmds = dma_alloc_coherent(dev->dev, SQ_SIZE(nvmeq),
1460 				&nvmeq->sq_dma_addr, GFP_KERNEL);
1461 	if (!nvmeq->sq_cmds)
1462 		return -ENOMEM;
1463 	return 0;
1464 }
1465 
1466 static int nvme_alloc_queue(struct nvme_dev *dev, int qid, int depth)
1467 {
1468 	struct nvme_queue *nvmeq = &dev->queues[qid];
1469 
1470 	if (dev->ctrl.queue_count > qid)
1471 		return 0;
1472 
1473 	nvmeq->sqes = qid ? dev->io_sqes : NVME_ADM_SQES;
1474 	nvmeq->q_depth = depth;
1475 	nvmeq->cqes = dma_alloc_coherent(dev->dev, CQ_SIZE(nvmeq),
1476 					 &nvmeq->cq_dma_addr, GFP_KERNEL);
1477 	if (!nvmeq->cqes)
1478 		goto free_nvmeq;
1479 
1480 	if (nvme_alloc_sq_cmds(dev, nvmeq, qid))
1481 		goto free_cqdma;
1482 
1483 	nvmeq->dev = dev;
1484 	spin_lock_init(&nvmeq->sq_lock);
1485 	spin_lock_init(&nvmeq->cq_poll_lock);
1486 	nvmeq->cq_head = 0;
1487 	nvmeq->cq_phase = 1;
1488 	nvmeq->q_db = &dev->dbs[qid * 2 * dev->db_stride];
1489 	nvmeq->qid = qid;
1490 	dev->ctrl.queue_count++;
1491 
1492 	return 0;
1493 
1494  free_cqdma:
1495 	dma_free_coherent(dev->dev, CQ_SIZE(nvmeq), (void *)nvmeq->cqes,
1496 			  nvmeq->cq_dma_addr);
1497  free_nvmeq:
1498 	return -ENOMEM;
1499 }
1500 
1501 static int queue_request_irq(struct nvme_queue *nvmeq)
1502 {
1503 	struct pci_dev *pdev = to_pci_dev(nvmeq->dev->dev);
1504 	int nr = nvmeq->dev->ctrl.instance;
1505 
1506 	if (use_threaded_interrupts) {
1507 		return pci_request_irq(pdev, nvmeq->cq_vector, nvme_irq_check,
1508 				nvme_irq, nvmeq, "nvme%dq%d", nr, nvmeq->qid);
1509 	} else {
1510 		return pci_request_irq(pdev, nvmeq->cq_vector, nvme_irq,
1511 				NULL, nvmeq, "nvme%dq%d", nr, nvmeq->qid);
1512 	}
1513 }
1514 
1515 static void nvme_init_queue(struct nvme_queue *nvmeq, u16 qid)
1516 {
1517 	struct nvme_dev *dev = nvmeq->dev;
1518 
1519 	nvmeq->sq_tail = 0;
1520 	nvmeq->last_sq_tail = 0;
1521 	nvmeq->cq_head = 0;
1522 	nvmeq->cq_phase = 1;
1523 	nvmeq->q_db = &dev->dbs[qid * 2 * dev->db_stride];
1524 	memset((void *)nvmeq->cqes, 0, CQ_SIZE(nvmeq));
1525 	nvme_dbbuf_init(dev, nvmeq, qid);
1526 	dev->online_queues++;
1527 	wmb(); /* ensure the first interrupt sees the initialization */
1528 }
1529 
1530 static int nvme_create_queue(struct nvme_queue *nvmeq, int qid, bool polled)
1531 {
1532 	struct nvme_dev *dev = nvmeq->dev;
1533 	int result;
1534 	u16 vector = 0;
1535 
1536 	clear_bit(NVMEQ_DELETE_ERROR, &nvmeq->flags);
1537 
1538 	/*
1539 	 * A queue's vector matches the queue identifier unless the controller
1540 	 * has only one vector available.
1541 	 */
1542 	if (!polled)
1543 		vector = dev->num_vecs == 1 ? 0 : qid;
1544 	else
1545 		set_bit(NVMEQ_POLLED, &nvmeq->flags);
1546 
1547 	result = adapter_alloc_cq(dev, qid, nvmeq, vector);
1548 	if (result)
1549 		return result;
1550 
1551 	result = adapter_alloc_sq(dev, qid, nvmeq);
1552 	if (result < 0)
1553 		return result;
1554 	else if (result)
1555 		goto release_cq;
1556 
1557 	nvmeq->cq_vector = vector;
1558 	nvme_init_queue(nvmeq, qid);
1559 
1560 	if (!polled) {
1561 		result = queue_request_irq(nvmeq);
1562 		if (result < 0)
1563 			goto release_sq;
1564 	}
1565 
1566 	set_bit(NVMEQ_ENABLED, &nvmeq->flags);
1567 	return result;
1568 
1569 release_sq:
1570 	dev->online_queues--;
1571 	adapter_delete_sq(dev, qid);
1572 release_cq:
1573 	adapter_delete_cq(dev, qid);
1574 	return result;
1575 }
1576 
1577 static const struct blk_mq_ops nvme_mq_admin_ops = {
1578 	.queue_rq	= nvme_queue_rq,
1579 	.complete	= nvme_pci_complete_rq,
1580 	.init_hctx	= nvme_admin_init_hctx,
1581 	.exit_hctx      = nvme_admin_exit_hctx,
1582 	.init_request	= nvme_init_request,
1583 	.timeout	= nvme_timeout,
1584 };
1585 
1586 static const struct blk_mq_ops nvme_mq_ops = {
1587 	.queue_rq	= nvme_queue_rq,
1588 	.complete	= nvme_pci_complete_rq,
1589 	.commit_rqs	= nvme_commit_rqs,
1590 	.init_hctx	= nvme_init_hctx,
1591 	.init_request	= nvme_init_request,
1592 	.map_queues	= nvme_pci_map_queues,
1593 	.timeout	= nvme_timeout,
1594 	.poll		= nvme_poll,
1595 };
1596 
1597 static void nvme_dev_remove_admin(struct nvme_dev *dev)
1598 {
1599 	if (dev->ctrl.admin_q && !blk_queue_dying(dev->ctrl.admin_q)) {
1600 		/*
1601 		 * If the controller was reset during removal, it's possible
1602 		 * user requests may be waiting on a stopped queue. Start the
1603 		 * queue to flush these to completion.
1604 		 */
1605 		blk_mq_unquiesce_queue(dev->ctrl.admin_q);
1606 		blk_cleanup_queue(dev->ctrl.admin_q);
1607 		blk_mq_free_tag_set(&dev->admin_tagset);
1608 	}
1609 }
1610 
1611 static int nvme_alloc_admin_tags(struct nvme_dev *dev)
1612 {
1613 	if (!dev->ctrl.admin_q) {
1614 		dev->admin_tagset.ops = &nvme_mq_admin_ops;
1615 		dev->admin_tagset.nr_hw_queues = 1;
1616 
1617 		dev->admin_tagset.queue_depth = NVME_AQ_MQ_TAG_DEPTH;
1618 		dev->admin_tagset.timeout = ADMIN_TIMEOUT;
1619 		dev->admin_tagset.numa_node = dev_to_node(dev->dev);
1620 		dev->admin_tagset.cmd_size = sizeof(struct nvme_iod);
1621 		dev->admin_tagset.flags = BLK_MQ_F_NO_SCHED;
1622 		dev->admin_tagset.driver_data = dev;
1623 
1624 		if (blk_mq_alloc_tag_set(&dev->admin_tagset))
1625 			return -ENOMEM;
1626 		dev->ctrl.admin_tagset = &dev->admin_tagset;
1627 
1628 		dev->ctrl.admin_q = blk_mq_init_queue(&dev->admin_tagset);
1629 		if (IS_ERR(dev->ctrl.admin_q)) {
1630 			blk_mq_free_tag_set(&dev->admin_tagset);
1631 			return -ENOMEM;
1632 		}
1633 		if (!blk_get_queue(dev->ctrl.admin_q)) {
1634 			nvme_dev_remove_admin(dev);
1635 			dev->ctrl.admin_q = NULL;
1636 			return -ENODEV;
1637 		}
1638 	} else
1639 		blk_mq_unquiesce_queue(dev->ctrl.admin_q);
1640 
1641 	return 0;
1642 }
1643 
1644 static unsigned long db_bar_size(struct nvme_dev *dev, unsigned nr_io_queues)
1645 {
1646 	return NVME_REG_DBS + ((nr_io_queues + 1) * 8 * dev->db_stride);
1647 }
1648 
1649 static int nvme_remap_bar(struct nvme_dev *dev, unsigned long size)
1650 {
1651 	struct pci_dev *pdev = to_pci_dev(dev->dev);
1652 
1653 	if (size <= dev->bar_mapped_size)
1654 		return 0;
1655 	if (size > pci_resource_len(pdev, 0))
1656 		return -ENOMEM;
1657 	if (dev->bar)
1658 		iounmap(dev->bar);
1659 	dev->bar = ioremap(pci_resource_start(pdev, 0), size);
1660 	if (!dev->bar) {
1661 		dev->bar_mapped_size = 0;
1662 		return -ENOMEM;
1663 	}
1664 	dev->bar_mapped_size = size;
1665 	dev->dbs = dev->bar + NVME_REG_DBS;
1666 
1667 	return 0;
1668 }
1669 
1670 static int nvme_pci_configure_admin_queue(struct nvme_dev *dev)
1671 {
1672 	int result;
1673 	u32 aqa;
1674 	struct nvme_queue *nvmeq;
1675 
1676 	result = nvme_remap_bar(dev, db_bar_size(dev, 0));
1677 	if (result < 0)
1678 		return result;
1679 
1680 	dev->subsystem = readl(dev->bar + NVME_REG_VS) >= NVME_VS(1, 1, 0) ?
1681 				NVME_CAP_NSSRC(dev->ctrl.cap) : 0;
1682 
1683 	if (dev->subsystem &&
1684 	    (readl(dev->bar + NVME_REG_CSTS) & NVME_CSTS_NSSRO))
1685 		writel(NVME_CSTS_NSSRO, dev->bar + NVME_REG_CSTS);
1686 
1687 	result = nvme_disable_ctrl(&dev->ctrl);
1688 	if (result < 0)
1689 		return result;
1690 
1691 	result = nvme_alloc_queue(dev, 0, NVME_AQ_DEPTH);
1692 	if (result)
1693 		return result;
1694 
1695 	nvmeq = &dev->queues[0];
1696 	aqa = nvmeq->q_depth - 1;
1697 	aqa |= aqa << 16;
1698 
1699 	writel(aqa, dev->bar + NVME_REG_AQA);
1700 	lo_hi_writeq(nvmeq->sq_dma_addr, dev->bar + NVME_REG_ASQ);
1701 	lo_hi_writeq(nvmeq->cq_dma_addr, dev->bar + NVME_REG_ACQ);
1702 
1703 	result = nvme_enable_ctrl(&dev->ctrl);
1704 	if (result)
1705 		return result;
1706 
1707 	nvmeq->cq_vector = 0;
1708 	nvme_init_queue(nvmeq, 0);
1709 	result = queue_request_irq(nvmeq);
1710 	if (result) {
1711 		dev->online_queues--;
1712 		return result;
1713 	}
1714 
1715 	set_bit(NVMEQ_ENABLED, &nvmeq->flags);
1716 	return result;
1717 }
1718 
1719 static int nvme_create_io_queues(struct nvme_dev *dev)
1720 {
1721 	unsigned i, max, rw_queues;
1722 	int ret = 0;
1723 
1724 	for (i = dev->ctrl.queue_count; i <= dev->max_qid; i++) {
1725 		if (nvme_alloc_queue(dev, i, dev->q_depth)) {
1726 			ret = -ENOMEM;
1727 			break;
1728 		}
1729 	}
1730 
1731 	max = min(dev->max_qid, dev->ctrl.queue_count - 1);
1732 	if (max != 1 && dev->io_queues[HCTX_TYPE_POLL]) {
1733 		rw_queues = dev->io_queues[HCTX_TYPE_DEFAULT] +
1734 				dev->io_queues[HCTX_TYPE_READ];
1735 	} else {
1736 		rw_queues = max;
1737 	}
1738 
1739 	for (i = dev->online_queues; i <= max; i++) {
1740 		bool polled = i > rw_queues;
1741 
1742 		ret = nvme_create_queue(&dev->queues[i], i, polled);
1743 		if (ret)
1744 			break;
1745 	}
1746 
1747 	/*
1748 	 * Ignore failing Create SQ/CQ commands, we can continue with less
1749 	 * than the desired amount of queues, and even a controller without
1750 	 * I/O queues can still be used to issue admin commands.  This might
1751 	 * be useful to upgrade a buggy firmware for example.
1752 	 */
1753 	return ret >= 0 ? 0 : ret;
1754 }
1755 
1756 static ssize_t nvme_cmb_show(struct device *dev,
1757 			     struct device_attribute *attr,
1758 			     char *buf)
1759 {
1760 	struct nvme_dev *ndev = to_nvme_dev(dev_get_drvdata(dev));
1761 
1762 	return scnprintf(buf, PAGE_SIZE, "cmbloc : x%08x\ncmbsz  : x%08x\n",
1763 		       ndev->cmbloc, ndev->cmbsz);
1764 }
1765 static DEVICE_ATTR(cmb, S_IRUGO, nvme_cmb_show, NULL);
1766 
1767 static u64 nvme_cmb_size_unit(struct nvme_dev *dev)
1768 {
1769 	u8 szu = (dev->cmbsz >> NVME_CMBSZ_SZU_SHIFT) & NVME_CMBSZ_SZU_MASK;
1770 
1771 	return 1ULL << (12 + 4 * szu);
1772 }
1773 
1774 static u32 nvme_cmb_size(struct nvme_dev *dev)
1775 {
1776 	return (dev->cmbsz >> NVME_CMBSZ_SZ_SHIFT) & NVME_CMBSZ_SZ_MASK;
1777 }
1778 
1779 static void nvme_map_cmb(struct nvme_dev *dev)
1780 {
1781 	u64 size, offset;
1782 	resource_size_t bar_size;
1783 	struct pci_dev *pdev = to_pci_dev(dev->dev);
1784 	int bar;
1785 
1786 	if (dev->cmb_size)
1787 		return;
1788 
1789 	dev->cmbsz = readl(dev->bar + NVME_REG_CMBSZ);
1790 	if (!dev->cmbsz)
1791 		return;
1792 	dev->cmbloc = readl(dev->bar + NVME_REG_CMBLOC);
1793 
1794 	size = nvme_cmb_size_unit(dev) * nvme_cmb_size(dev);
1795 	offset = nvme_cmb_size_unit(dev) * NVME_CMB_OFST(dev->cmbloc);
1796 	bar = NVME_CMB_BIR(dev->cmbloc);
1797 	bar_size = pci_resource_len(pdev, bar);
1798 
1799 	if (offset > bar_size)
1800 		return;
1801 
1802 	/*
1803 	 * Controllers may support a CMB size larger than their BAR,
1804 	 * for example, due to being behind a bridge. Reduce the CMB to
1805 	 * the reported size of the BAR
1806 	 */
1807 	if (size > bar_size - offset)
1808 		size = bar_size - offset;
1809 
1810 	if (pci_p2pdma_add_resource(pdev, bar, size, offset)) {
1811 		dev_warn(dev->ctrl.device,
1812 			 "failed to register the CMB\n");
1813 		return;
1814 	}
1815 
1816 	dev->cmb_size = size;
1817 	dev->cmb_use_sqes = use_cmb_sqes && (dev->cmbsz & NVME_CMBSZ_SQS);
1818 
1819 	if ((dev->cmbsz & (NVME_CMBSZ_WDS | NVME_CMBSZ_RDS)) ==
1820 			(NVME_CMBSZ_WDS | NVME_CMBSZ_RDS))
1821 		pci_p2pmem_publish(pdev, true);
1822 
1823 	if (sysfs_add_file_to_group(&dev->ctrl.device->kobj,
1824 				    &dev_attr_cmb.attr, NULL))
1825 		dev_warn(dev->ctrl.device,
1826 			 "failed to add sysfs attribute for CMB\n");
1827 }
1828 
1829 static inline void nvme_release_cmb(struct nvme_dev *dev)
1830 {
1831 	if (dev->cmb_size) {
1832 		sysfs_remove_file_from_group(&dev->ctrl.device->kobj,
1833 					     &dev_attr_cmb.attr, NULL);
1834 		dev->cmb_size = 0;
1835 	}
1836 }
1837 
1838 static int nvme_set_host_mem(struct nvme_dev *dev, u32 bits)
1839 {
1840 	u64 dma_addr = dev->host_mem_descs_dma;
1841 	struct nvme_command c;
1842 	int ret;
1843 
1844 	memset(&c, 0, sizeof(c));
1845 	c.features.opcode	= nvme_admin_set_features;
1846 	c.features.fid		= cpu_to_le32(NVME_FEAT_HOST_MEM_BUF);
1847 	c.features.dword11	= cpu_to_le32(bits);
1848 	c.features.dword12	= cpu_to_le32(dev->host_mem_size >>
1849 					      ilog2(dev->ctrl.page_size));
1850 	c.features.dword13	= cpu_to_le32(lower_32_bits(dma_addr));
1851 	c.features.dword14	= cpu_to_le32(upper_32_bits(dma_addr));
1852 	c.features.dword15	= cpu_to_le32(dev->nr_host_mem_descs);
1853 
1854 	ret = nvme_submit_sync_cmd(dev->ctrl.admin_q, &c, NULL, 0);
1855 	if (ret) {
1856 		dev_warn(dev->ctrl.device,
1857 			 "failed to set host mem (err %d, flags %#x).\n",
1858 			 ret, bits);
1859 	}
1860 	return ret;
1861 }
1862 
1863 static void nvme_free_host_mem(struct nvme_dev *dev)
1864 {
1865 	int i;
1866 
1867 	for (i = 0; i < dev->nr_host_mem_descs; i++) {
1868 		struct nvme_host_mem_buf_desc *desc = &dev->host_mem_descs[i];
1869 		size_t size = le32_to_cpu(desc->size) * dev->ctrl.page_size;
1870 
1871 		dma_free_attrs(dev->dev, size, dev->host_mem_desc_bufs[i],
1872 			       le64_to_cpu(desc->addr),
1873 			       DMA_ATTR_NO_KERNEL_MAPPING | DMA_ATTR_NO_WARN);
1874 	}
1875 
1876 	kfree(dev->host_mem_desc_bufs);
1877 	dev->host_mem_desc_bufs = NULL;
1878 	dma_free_coherent(dev->dev,
1879 			dev->nr_host_mem_descs * sizeof(*dev->host_mem_descs),
1880 			dev->host_mem_descs, dev->host_mem_descs_dma);
1881 	dev->host_mem_descs = NULL;
1882 	dev->nr_host_mem_descs = 0;
1883 }
1884 
1885 static int __nvme_alloc_host_mem(struct nvme_dev *dev, u64 preferred,
1886 		u32 chunk_size)
1887 {
1888 	struct nvme_host_mem_buf_desc *descs;
1889 	u32 max_entries, len;
1890 	dma_addr_t descs_dma;
1891 	int i = 0;
1892 	void **bufs;
1893 	u64 size, tmp;
1894 
1895 	tmp = (preferred + chunk_size - 1);
1896 	do_div(tmp, chunk_size);
1897 	max_entries = tmp;
1898 
1899 	if (dev->ctrl.hmmaxd && dev->ctrl.hmmaxd < max_entries)
1900 		max_entries = dev->ctrl.hmmaxd;
1901 
1902 	descs = dma_alloc_coherent(dev->dev, max_entries * sizeof(*descs),
1903 				   &descs_dma, GFP_KERNEL);
1904 	if (!descs)
1905 		goto out;
1906 
1907 	bufs = kcalloc(max_entries, sizeof(*bufs), GFP_KERNEL);
1908 	if (!bufs)
1909 		goto out_free_descs;
1910 
1911 	for (size = 0; size < preferred && i < max_entries; size += len) {
1912 		dma_addr_t dma_addr;
1913 
1914 		len = min_t(u64, chunk_size, preferred - size);
1915 		bufs[i] = dma_alloc_attrs(dev->dev, len, &dma_addr, GFP_KERNEL,
1916 				DMA_ATTR_NO_KERNEL_MAPPING | DMA_ATTR_NO_WARN);
1917 		if (!bufs[i])
1918 			break;
1919 
1920 		descs[i].addr = cpu_to_le64(dma_addr);
1921 		descs[i].size = cpu_to_le32(len / dev->ctrl.page_size);
1922 		i++;
1923 	}
1924 
1925 	if (!size)
1926 		goto out_free_bufs;
1927 
1928 	dev->nr_host_mem_descs = i;
1929 	dev->host_mem_size = size;
1930 	dev->host_mem_descs = descs;
1931 	dev->host_mem_descs_dma = descs_dma;
1932 	dev->host_mem_desc_bufs = bufs;
1933 	return 0;
1934 
1935 out_free_bufs:
1936 	while (--i >= 0) {
1937 		size_t size = le32_to_cpu(descs[i].size) * dev->ctrl.page_size;
1938 
1939 		dma_free_attrs(dev->dev, size, bufs[i],
1940 			       le64_to_cpu(descs[i].addr),
1941 			       DMA_ATTR_NO_KERNEL_MAPPING | DMA_ATTR_NO_WARN);
1942 	}
1943 
1944 	kfree(bufs);
1945 out_free_descs:
1946 	dma_free_coherent(dev->dev, max_entries * sizeof(*descs), descs,
1947 			descs_dma);
1948 out:
1949 	dev->host_mem_descs = NULL;
1950 	return -ENOMEM;
1951 }
1952 
1953 static int nvme_alloc_host_mem(struct nvme_dev *dev, u64 min, u64 preferred)
1954 {
1955 	u32 chunk_size;
1956 
1957 	/* start big and work our way down */
1958 	for (chunk_size = min_t(u64, preferred, PAGE_SIZE * MAX_ORDER_NR_PAGES);
1959 	     chunk_size >= max_t(u32, dev->ctrl.hmminds * 4096, PAGE_SIZE * 2);
1960 	     chunk_size /= 2) {
1961 		if (!__nvme_alloc_host_mem(dev, preferred, chunk_size)) {
1962 			if (!min || dev->host_mem_size >= min)
1963 				return 0;
1964 			nvme_free_host_mem(dev);
1965 		}
1966 	}
1967 
1968 	return -ENOMEM;
1969 }
1970 
1971 static int nvme_setup_host_mem(struct nvme_dev *dev)
1972 {
1973 	u64 max = (u64)max_host_mem_size_mb * SZ_1M;
1974 	u64 preferred = (u64)dev->ctrl.hmpre * 4096;
1975 	u64 min = (u64)dev->ctrl.hmmin * 4096;
1976 	u32 enable_bits = NVME_HOST_MEM_ENABLE;
1977 	int ret;
1978 
1979 	preferred = min(preferred, max);
1980 	if (min > max) {
1981 		dev_warn(dev->ctrl.device,
1982 			"min host memory (%lld MiB) above limit (%d MiB).\n",
1983 			min >> ilog2(SZ_1M), max_host_mem_size_mb);
1984 		nvme_free_host_mem(dev);
1985 		return 0;
1986 	}
1987 
1988 	/*
1989 	 * If we already have a buffer allocated check if we can reuse it.
1990 	 */
1991 	if (dev->host_mem_descs) {
1992 		if (dev->host_mem_size >= min)
1993 			enable_bits |= NVME_HOST_MEM_RETURN;
1994 		else
1995 			nvme_free_host_mem(dev);
1996 	}
1997 
1998 	if (!dev->host_mem_descs) {
1999 		if (nvme_alloc_host_mem(dev, min, preferred)) {
2000 			dev_warn(dev->ctrl.device,
2001 				"failed to allocate host memory buffer.\n");
2002 			return 0; /* controller must work without HMB */
2003 		}
2004 
2005 		dev_info(dev->ctrl.device,
2006 			"allocated %lld MiB host memory buffer.\n",
2007 			dev->host_mem_size >> ilog2(SZ_1M));
2008 	}
2009 
2010 	ret = nvme_set_host_mem(dev, enable_bits);
2011 	if (ret)
2012 		nvme_free_host_mem(dev);
2013 	return ret;
2014 }
2015 
2016 /*
2017  * nirqs is the number of interrupts available for write and read
2018  * queues. The core already reserved an interrupt for the admin queue.
2019  */
2020 static void nvme_calc_irq_sets(struct irq_affinity *affd, unsigned int nrirqs)
2021 {
2022 	struct nvme_dev *dev = affd->priv;
2023 	unsigned int nr_read_queues;
2024 
2025 	/*
2026 	 * If there is no interupt available for queues, ensure that
2027 	 * the default queue is set to 1. The affinity set size is
2028 	 * also set to one, but the irq core ignores it for this case.
2029 	 *
2030 	 * If only one interrupt is available or 'write_queue' == 0, combine
2031 	 * write and read queues.
2032 	 *
2033 	 * If 'write_queues' > 0, ensure it leaves room for at least one read
2034 	 * queue.
2035 	 */
2036 	if (!nrirqs) {
2037 		nrirqs = 1;
2038 		nr_read_queues = 0;
2039 	} else if (nrirqs == 1 || !write_queues) {
2040 		nr_read_queues = 0;
2041 	} else if (write_queues >= nrirqs) {
2042 		nr_read_queues = 1;
2043 	} else {
2044 		nr_read_queues = nrirqs - write_queues;
2045 	}
2046 
2047 	dev->io_queues[HCTX_TYPE_DEFAULT] = nrirqs - nr_read_queues;
2048 	affd->set_size[HCTX_TYPE_DEFAULT] = nrirqs - nr_read_queues;
2049 	dev->io_queues[HCTX_TYPE_READ] = nr_read_queues;
2050 	affd->set_size[HCTX_TYPE_READ] = nr_read_queues;
2051 	affd->nr_sets = nr_read_queues ? 2 : 1;
2052 }
2053 
2054 static int nvme_setup_irqs(struct nvme_dev *dev, unsigned int nr_io_queues)
2055 {
2056 	struct pci_dev *pdev = to_pci_dev(dev->dev);
2057 	struct irq_affinity affd = {
2058 		.pre_vectors	= 1,
2059 		.calc_sets	= nvme_calc_irq_sets,
2060 		.priv		= dev,
2061 	};
2062 	unsigned int irq_queues, this_p_queues;
2063 	unsigned int nr_cpus = num_possible_cpus();
2064 
2065 	/*
2066 	 * Poll queues don't need interrupts, but we need at least one IO
2067 	 * queue left over for non-polled IO.
2068 	 */
2069 	this_p_queues = poll_queues;
2070 	if (this_p_queues >= nr_io_queues) {
2071 		this_p_queues = nr_io_queues - 1;
2072 		irq_queues = 1;
2073 	} else {
2074 		if (nr_cpus < nr_io_queues - this_p_queues)
2075 			irq_queues = nr_cpus + 1;
2076 		else
2077 			irq_queues = nr_io_queues - this_p_queues + 1;
2078 	}
2079 	dev->io_queues[HCTX_TYPE_POLL] = this_p_queues;
2080 
2081 	/* Initialize for the single interrupt case */
2082 	dev->io_queues[HCTX_TYPE_DEFAULT] = 1;
2083 	dev->io_queues[HCTX_TYPE_READ] = 0;
2084 
2085 	/*
2086 	 * Some Apple controllers require all queues to use the
2087 	 * first vector.
2088 	 */
2089 	if (dev->ctrl.quirks & NVME_QUIRK_SINGLE_VECTOR)
2090 		irq_queues = 1;
2091 
2092 	return pci_alloc_irq_vectors_affinity(pdev, 1, irq_queues,
2093 			      PCI_IRQ_ALL_TYPES | PCI_IRQ_AFFINITY, &affd);
2094 }
2095 
2096 static void nvme_disable_io_queues(struct nvme_dev *dev)
2097 {
2098 	if (__nvme_disable_io_queues(dev, nvme_admin_delete_sq))
2099 		__nvme_disable_io_queues(dev, nvme_admin_delete_cq);
2100 }
2101 
2102 static int nvme_setup_io_queues(struct nvme_dev *dev)
2103 {
2104 	struct nvme_queue *adminq = &dev->queues[0];
2105 	struct pci_dev *pdev = to_pci_dev(dev->dev);
2106 	int result, nr_io_queues;
2107 	unsigned long size;
2108 
2109 	nr_io_queues = max_io_queues();
2110 
2111 	/*
2112 	 * If tags are shared with admin queue (Apple bug), then
2113 	 * make sure we only use one IO queue.
2114 	 */
2115 	if (dev->ctrl.quirks & NVME_QUIRK_SHARED_TAGS)
2116 		nr_io_queues = 1;
2117 
2118 	result = nvme_set_queue_count(&dev->ctrl, &nr_io_queues);
2119 	if (result < 0)
2120 		return result;
2121 
2122 	if (nr_io_queues == 0)
2123 		return 0;
2124 
2125 	clear_bit(NVMEQ_ENABLED, &adminq->flags);
2126 
2127 	if (dev->cmb_use_sqes) {
2128 		result = nvme_cmb_qdepth(dev, nr_io_queues,
2129 				sizeof(struct nvme_command));
2130 		if (result > 0)
2131 			dev->q_depth = result;
2132 		else
2133 			dev->cmb_use_sqes = false;
2134 	}
2135 
2136 	do {
2137 		size = db_bar_size(dev, nr_io_queues);
2138 		result = nvme_remap_bar(dev, size);
2139 		if (!result)
2140 			break;
2141 		if (!--nr_io_queues)
2142 			return -ENOMEM;
2143 	} while (1);
2144 	adminq->q_db = dev->dbs;
2145 
2146  retry:
2147 	/* Deregister the admin queue's interrupt */
2148 	pci_free_irq(pdev, 0, adminq);
2149 
2150 	/*
2151 	 * If we enable msix early due to not intx, disable it again before
2152 	 * setting up the full range we need.
2153 	 */
2154 	pci_free_irq_vectors(pdev);
2155 
2156 	result = nvme_setup_irqs(dev, nr_io_queues);
2157 	if (result <= 0)
2158 		return -EIO;
2159 
2160 	dev->num_vecs = result;
2161 	result = max(result - 1, 1);
2162 	dev->max_qid = result + dev->io_queues[HCTX_TYPE_POLL];
2163 
2164 	/*
2165 	 * Should investigate if there's a performance win from allocating
2166 	 * more queues than interrupt vectors; it might allow the submission
2167 	 * path to scale better, even if the receive path is limited by the
2168 	 * number of interrupts.
2169 	 */
2170 	result = queue_request_irq(adminq);
2171 	if (result)
2172 		return result;
2173 	set_bit(NVMEQ_ENABLED, &adminq->flags);
2174 
2175 	result = nvme_create_io_queues(dev);
2176 	if (result || dev->online_queues < 2)
2177 		return result;
2178 
2179 	if (dev->online_queues - 1 < dev->max_qid) {
2180 		nr_io_queues = dev->online_queues - 1;
2181 		nvme_disable_io_queues(dev);
2182 		nvme_suspend_io_queues(dev);
2183 		goto retry;
2184 	}
2185 	dev_info(dev->ctrl.device, "%d/%d/%d default/read/poll queues\n",
2186 					dev->io_queues[HCTX_TYPE_DEFAULT],
2187 					dev->io_queues[HCTX_TYPE_READ],
2188 					dev->io_queues[HCTX_TYPE_POLL]);
2189 	return 0;
2190 }
2191 
2192 static void nvme_del_queue_end(struct request *req, blk_status_t error)
2193 {
2194 	struct nvme_queue *nvmeq = req->end_io_data;
2195 
2196 	blk_mq_free_request(req);
2197 	complete(&nvmeq->delete_done);
2198 }
2199 
2200 static void nvme_del_cq_end(struct request *req, blk_status_t error)
2201 {
2202 	struct nvme_queue *nvmeq = req->end_io_data;
2203 
2204 	if (error)
2205 		set_bit(NVMEQ_DELETE_ERROR, &nvmeq->flags);
2206 
2207 	nvme_del_queue_end(req, error);
2208 }
2209 
2210 static int nvme_delete_queue(struct nvme_queue *nvmeq, u8 opcode)
2211 {
2212 	struct request_queue *q = nvmeq->dev->ctrl.admin_q;
2213 	struct request *req;
2214 	struct nvme_command cmd;
2215 
2216 	memset(&cmd, 0, sizeof(cmd));
2217 	cmd.delete_queue.opcode = opcode;
2218 	cmd.delete_queue.qid = cpu_to_le16(nvmeq->qid);
2219 
2220 	req = nvme_alloc_request(q, &cmd, BLK_MQ_REQ_NOWAIT, NVME_QID_ANY);
2221 	if (IS_ERR(req))
2222 		return PTR_ERR(req);
2223 
2224 	req->timeout = ADMIN_TIMEOUT;
2225 	req->end_io_data = nvmeq;
2226 
2227 	init_completion(&nvmeq->delete_done);
2228 	blk_execute_rq_nowait(q, NULL, req, false,
2229 			opcode == nvme_admin_delete_cq ?
2230 				nvme_del_cq_end : nvme_del_queue_end);
2231 	return 0;
2232 }
2233 
2234 static bool __nvme_disable_io_queues(struct nvme_dev *dev, u8 opcode)
2235 {
2236 	int nr_queues = dev->online_queues - 1, sent = 0;
2237 	unsigned long timeout;
2238 
2239  retry:
2240 	timeout = ADMIN_TIMEOUT;
2241 	while (nr_queues > 0) {
2242 		if (nvme_delete_queue(&dev->queues[nr_queues], opcode))
2243 			break;
2244 		nr_queues--;
2245 		sent++;
2246 	}
2247 	while (sent) {
2248 		struct nvme_queue *nvmeq = &dev->queues[nr_queues + sent];
2249 
2250 		timeout = wait_for_completion_io_timeout(&nvmeq->delete_done,
2251 				timeout);
2252 		if (timeout == 0)
2253 			return false;
2254 
2255 		/* handle any remaining CQEs */
2256 		if (opcode == nvme_admin_delete_cq &&
2257 		    !test_bit(NVMEQ_DELETE_ERROR, &nvmeq->flags))
2258 			nvme_poll_irqdisable(nvmeq, -1);
2259 
2260 		sent--;
2261 		if (nr_queues)
2262 			goto retry;
2263 	}
2264 	return true;
2265 }
2266 
2267 static void 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;
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 }
2309 
2310 static int nvme_pci_enable(struct nvme_dev *dev)
2311 {
2312 	int result = -ENOMEM;
2313 	struct pci_dev *pdev = to_pci_dev(dev->dev);
2314 
2315 	if (pci_enable_device_mem(pdev))
2316 		return result;
2317 
2318 	pci_set_master(pdev);
2319 
2320 	if (dma_set_mask_and_coherent(dev->dev, DMA_BIT_MASK(64)))
2321 		goto disable;
2322 
2323 	if (readl(dev->bar + NVME_REG_CSTS) == -1) {
2324 		result = -ENODEV;
2325 		goto disable;
2326 	}
2327 
2328 	/*
2329 	 * Some devices and/or platforms don't advertise or work with INTx
2330 	 * interrupts. Pre-enable a single MSIX or MSI vec for setup. We'll
2331 	 * adjust this later.
2332 	 */
2333 	result = pci_alloc_irq_vectors(pdev, 1, 1, PCI_IRQ_ALL_TYPES);
2334 	if (result < 0)
2335 		return result;
2336 
2337 	dev->ctrl.cap = lo_hi_readq(dev->bar + NVME_REG_CAP);
2338 
2339 	dev->q_depth = min_t(int, NVME_CAP_MQES(dev->ctrl.cap) + 1,
2340 				io_queue_depth);
2341 	dev->ctrl.sqsize = dev->q_depth - 1; /* 0's based queue depth */
2342 	dev->db_stride = 1 << NVME_CAP_STRIDE(dev->ctrl.cap);
2343 	dev->dbs = dev->bar + 4096;
2344 
2345 	/*
2346 	 * Some Apple controllers require a non-standard SQE size.
2347 	 * Interestingly they also seem to ignore the CC:IOSQES register
2348 	 * so we don't bother updating it here.
2349 	 */
2350 	if (dev->ctrl.quirks & NVME_QUIRK_128_BYTES_SQES)
2351 		dev->io_sqes = 7;
2352 	else
2353 		dev->io_sqes = NVME_NVM_IOSQES;
2354 
2355 	/*
2356 	 * Temporary fix for the Apple controller found in the MacBook8,1 and
2357 	 * some MacBook7,1 to avoid controller resets and data loss.
2358 	 */
2359 	if (pdev->vendor == PCI_VENDOR_ID_APPLE && pdev->device == 0x2001) {
2360 		dev->q_depth = 2;
2361 		dev_warn(dev->ctrl.device, "detected Apple NVMe controller, "
2362 			"set queue depth=%u to work around controller resets\n",
2363 			dev->q_depth);
2364 	} else if (pdev->vendor == PCI_VENDOR_ID_SAMSUNG &&
2365 		   (pdev->device == 0xa821 || pdev->device == 0xa822) &&
2366 		   NVME_CAP_MQES(dev->ctrl.cap) == 0) {
2367 		dev->q_depth = 64;
2368 		dev_err(dev->ctrl.device, "detected PM1725 NVMe controller, "
2369                         "set queue depth=%u\n", dev->q_depth);
2370 	}
2371 
2372 	/*
2373 	 * Controllers with the shared tags quirk need the IO queue to be
2374 	 * big enough so that we get 32 tags for the admin queue
2375 	 */
2376 	if ((dev->ctrl.quirks & NVME_QUIRK_SHARED_TAGS) &&
2377 	    (dev->q_depth < (NVME_AQ_DEPTH + 2))) {
2378 		dev->q_depth = NVME_AQ_DEPTH + 2;
2379 		dev_warn(dev->ctrl.device, "IO queue depth clamped to %d\n",
2380 			 dev->q_depth);
2381 	}
2382 
2383 
2384 	nvme_map_cmb(dev);
2385 
2386 	pci_enable_pcie_error_reporting(pdev);
2387 	pci_save_state(pdev);
2388 	return 0;
2389 
2390  disable:
2391 	pci_disable_device(pdev);
2392 	return result;
2393 }
2394 
2395 static void nvme_dev_unmap(struct nvme_dev *dev)
2396 {
2397 	if (dev->bar)
2398 		iounmap(dev->bar);
2399 	pci_release_mem_regions(to_pci_dev(dev->dev));
2400 }
2401 
2402 static void nvme_pci_disable(struct nvme_dev *dev)
2403 {
2404 	struct pci_dev *pdev = to_pci_dev(dev->dev);
2405 
2406 	pci_free_irq_vectors(pdev);
2407 
2408 	if (pci_is_enabled(pdev)) {
2409 		pci_disable_pcie_error_reporting(pdev);
2410 		pci_disable_device(pdev);
2411 	}
2412 }
2413 
2414 static void nvme_dev_disable(struct nvme_dev *dev, bool shutdown)
2415 {
2416 	bool dead = true, freeze = false;
2417 	struct pci_dev *pdev = to_pci_dev(dev->dev);
2418 
2419 	mutex_lock(&dev->shutdown_lock);
2420 	if (pci_is_enabled(pdev)) {
2421 		u32 csts = readl(dev->bar + NVME_REG_CSTS);
2422 
2423 		if (dev->ctrl.state == NVME_CTRL_LIVE ||
2424 		    dev->ctrl.state == NVME_CTRL_RESETTING) {
2425 			freeze = true;
2426 			nvme_start_freeze(&dev->ctrl);
2427 		}
2428 		dead = !!((csts & NVME_CSTS_CFS) || !(csts & NVME_CSTS_RDY) ||
2429 			pdev->error_state  != pci_channel_io_normal);
2430 	}
2431 
2432 	/*
2433 	 * Give the controller a chance to complete all entered requests if
2434 	 * doing a safe shutdown.
2435 	 */
2436 	if (!dead && shutdown && freeze)
2437 		nvme_wait_freeze_timeout(&dev->ctrl, NVME_IO_TIMEOUT);
2438 
2439 	nvme_stop_queues(&dev->ctrl);
2440 
2441 	if (!dead && dev->ctrl.queue_count > 0) {
2442 		nvme_disable_io_queues(dev);
2443 		nvme_disable_admin_queue(dev, shutdown);
2444 	}
2445 	nvme_suspend_io_queues(dev);
2446 	nvme_suspend_queue(&dev->queues[0]);
2447 	nvme_pci_disable(dev);
2448 
2449 	blk_mq_tagset_busy_iter(&dev->tagset, nvme_cancel_request, &dev->ctrl);
2450 	blk_mq_tagset_busy_iter(&dev->admin_tagset, nvme_cancel_request, &dev->ctrl);
2451 	blk_mq_tagset_wait_completed_request(&dev->tagset);
2452 	blk_mq_tagset_wait_completed_request(&dev->admin_tagset);
2453 
2454 	/*
2455 	 * The driver will not be starting up queues again if shutting down so
2456 	 * must flush all entered requests to their failed completion to avoid
2457 	 * deadlocking blk-mq hot-cpu notifier.
2458 	 */
2459 	if (shutdown) {
2460 		nvme_start_queues(&dev->ctrl);
2461 		if (dev->ctrl.admin_q && !blk_queue_dying(dev->ctrl.admin_q))
2462 			blk_mq_unquiesce_queue(dev->ctrl.admin_q);
2463 	}
2464 	mutex_unlock(&dev->shutdown_lock);
2465 }
2466 
2467 static int nvme_disable_prepare_reset(struct nvme_dev *dev, bool shutdown)
2468 {
2469 	if (!nvme_wait_reset(&dev->ctrl))
2470 		return -EBUSY;
2471 	nvme_dev_disable(dev, shutdown);
2472 	return 0;
2473 }
2474 
2475 static int nvme_setup_prp_pools(struct nvme_dev *dev)
2476 {
2477 	dev->prp_page_pool = dma_pool_create("prp list page", dev->dev,
2478 						PAGE_SIZE, PAGE_SIZE, 0);
2479 	if (!dev->prp_page_pool)
2480 		return -ENOMEM;
2481 
2482 	/* Optimisation for I/Os between 4k and 128k */
2483 	dev->prp_small_pool = dma_pool_create("prp list 256", dev->dev,
2484 						256, 256, 0);
2485 	if (!dev->prp_small_pool) {
2486 		dma_pool_destroy(dev->prp_page_pool);
2487 		return -ENOMEM;
2488 	}
2489 	return 0;
2490 }
2491 
2492 static void nvme_release_prp_pools(struct nvme_dev *dev)
2493 {
2494 	dma_pool_destroy(dev->prp_page_pool);
2495 	dma_pool_destroy(dev->prp_small_pool);
2496 }
2497 
2498 static void nvme_free_tagset(struct nvme_dev *dev)
2499 {
2500 	if (dev->tagset.tags)
2501 		blk_mq_free_tag_set(&dev->tagset);
2502 	dev->ctrl.tagset = NULL;
2503 }
2504 
2505 static void nvme_pci_free_ctrl(struct nvme_ctrl *ctrl)
2506 {
2507 	struct nvme_dev *dev = to_nvme_dev(ctrl);
2508 
2509 	nvme_dbbuf_dma_free(dev);
2510 	put_device(dev->dev);
2511 	nvme_free_tagset(dev);
2512 	if (dev->ctrl.admin_q)
2513 		blk_put_queue(dev->ctrl.admin_q);
2514 	kfree(dev->queues);
2515 	free_opal_dev(dev->ctrl.opal_dev);
2516 	mempool_destroy(dev->iod_mempool);
2517 	kfree(dev);
2518 }
2519 
2520 static void nvme_remove_dead_ctrl(struct nvme_dev *dev)
2521 {
2522 	/*
2523 	 * Set state to deleting now to avoid blocking nvme_wait_reset(), which
2524 	 * may be holding this pci_dev's device lock.
2525 	 */
2526 	nvme_change_ctrl_state(&dev->ctrl, NVME_CTRL_DELETING);
2527 	nvme_get_ctrl(&dev->ctrl);
2528 	nvme_dev_disable(dev, false);
2529 	nvme_kill_queues(&dev->ctrl);
2530 	if (!queue_work(nvme_wq, &dev->remove_work))
2531 		nvme_put_ctrl(&dev->ctrl);
2532 }
2533 
2534 static void nvme_reset_work(struct work_struct *work)
2535 {
2536 	struct nvme_dev *dev =
2537 		container_of(work, struct nvme_dev, ctrl.reset_work);
2538 	bool was_suspend = !!(dev->ctrl.ctrl_config & NVME_CC_SHN_NORMAL);
2539 	int result;
2540 
2541 	if (WARN_ON(dev->ctrl.state != NVME_CTRL_RESETTING)) {
2542 		result = -ENODEV;
2543 		goto out;
2544 	}
2545 
2546 	/*
2547 	 * If we're called to reset a live controller first shut it down before
2548 	 * moving on.
2549 	 */
2550 	if (dev->ctrl.ctrl_config & NVME_CC_ENABLE)
2551 		nvme_dev_disable(dev, false);
2552 	nvme_sync_queues(&dev->ctrl);
2553 
2554 	mutex_lock(&dev->shutdown_lock);
2555 	result = nvme_pci_enable(dev);
2556 	if (result)
2557 		goto out_unlock;
2558 
2559 	result = nvme_pci_configure_admin_queue(dev);
2560 	if (result)
2561 		goto out_unlock;
2562 
2563 	result = nvme_alloc_admin_tags(dev);
2564 	if (result)
2565 		goto out_unlock;
2566 
2567 	/*
2568 	 * Limit the max command size to prevent iod->sg allocations going
2569 	 * over a single page.
2570 	 */
2571 	dev->ctrl.max_hw_sectors = min_t(u32,
2572 		NVME_MAX_KB_SZ << 1, dma_max_mapping_size(dev->dev) >> 9);
2573 	dev->ctrl.max_segments = NVME_MAX_SEGS;
2574 
2575 	/*
2576 	 * Don't limit the IOMMU merged segment size.
2577 	 */
2578 	dma_set_max_seg_size(dev->dev, 0xffffffff);
2579 
2580 	mutex_unlock(&dev->shutdown_lock);
2581 
2582 	/*
2583 	 * Introduce CONNECTING state from nvme-fc/rdma transports to mark the
2584 	 * initializing procedure here.
2585 	 */
2586 	if (!nvme_change_ctrl_state(&dev->ctrl, NVME_CTRL_CONNECTING)) {
2587 		dev_warn(dev->ctrl.device,
2588 			"failed to mark controller CONNECTING\n");
2589 		result = -EBUSY;
2590 		goto out;
2591 	}
2592 
2593 	result = nvme_init_identify(&dev->ctrl);
2594 	if (result)
2595 		goto out;
2596 
2597 	if (dev->ctrl.oacs & NVME_CTRL_OACS_SEC_SUPP) {
2598 		if (!dev->ctrl.opal_dev)
2599 			dev->ctrl.opal_dev =
2600 				init_opal_dev(&dev->ctrl, &nvme_sec_submit);
2601 		else if (was_suspend)
2602 			opal_unlock_from_suspend(dev->ctrl.opal_dev);
2603 	} else {
2604 		free_opal_dev(dev->ctrl.opal_dev);
2605 		dev->ctrl.opal_dev = NULL;
2606 	}
2607 
2608 	if (dev->ctrl.oacs & NVME_CTRL_OACS_DBBUF_SUPP) {
2609 		result = nvme_dbbuf_dma_alloc(dev);
2610 		if (result)
2611 			dev_warn(dev->dev,
2612 				 "unable to allocate dma for dbbuf\n");
2613 	}
2614 
2615 	if (dev->ctrl.hmpre) {
2616 		result = nvme_setup_host_mem(dev);
2617 		if (result < 0)
2618 			goto out;
2619 	}
2620 
2621 	result = nvme_setup_io_queues(dev);
2622 	if (result)
2623 		goto out;
2624 
2625 	/*
2626 	 * Keep the controller around but remove all namespaces if we don't have
2627 	 * any working I/O queue.
2628 	 */
2629 	if (dev->online_queues < 2) {
2630 		dev_warn(dev->ctrl.device, "IO queues not created\n");
2631 		nvme_kill_queues(&dev->ctrl);
2632 		nvme_remove_namespaces(&dev->ctrl);
2633 		nvme_free_tagset(dev);
2634 	} else {
2635 		nvme_start_queues(&dev->ctrl);
2636 		nvme_wait_freeze(&dev->ctrl);
2637 		nvme_dev_add(dev);
2638 		nvme_unfreeze(&dev->ctrl);
2639 	}
2640 
2641 	/*
2642 	 * If only admin queue live, keep it to do further investigation or
2643 	 * recovery.
2644 	 */
2645 	if (!nvme_change_ctrl_state(&dev->ctrl, NVME_CTRL_LIVE)) {
2646 		dev_warn(dev->ctrl.device,
2647 			"failed to mark controller live state\n");
2648 		result = -ENODEV;
2649 		goto out;
2650 	}
2651 
2652 	nvme_start_ctrl(&dev->ctrl);
2653 	return;
2654 
2655  out_unlock:
2656 	mutex_unlock(&dev->shutdown_lock);
2657  out:
2658 	if (result)
2659 		dev_warn(dev->ctrl.device,
2660 			 "Removing after probe failure status: %d\n", result);
2661 	nvme_remove_dead_ctrl(dev);
2662 }
2663 
2664 static void nvme_remove_dead_ctrl_work(struct work_struct *work)
2665 {
2666 	struct nvme_dev *dev = container_of(work, struct nvme_dev, remove_work);
2667 	struct pci_dev *pdev = to_pci_dev(dev->dev);
2668 
2669 	if (pci_get_drvdata(pdev))
2670 		device_release_driver(&pdev->dev);
2671 	nvme_put_ctrl(&dev->ctrl);
2672 }
2673 
2674 static int nvme_pci_reg_read32(struct nvme_ctrl *ctrl, u32 off, u32 *val)
2675 {
2676 	*val = readl(to_nvme_dev(ctrl)->bar + off);
2677 	return 0;
2678 }
2679 
2680 static int nvme_pci_reg_write32(struct nvme_ctrl *ctrl, u32 off, u32 val)
2681 {
2682 	writel(val, to_nvme_dev(ctrl)->bar + off);
2683 	return 0;
2684 }
2685 
2686 static int nvme_pci_reg_read64(struct nvme_ctrl *ctrl, u32 off, u64 *val)
2687 {
2688 	*val = lo_hi_readq(to_nvme_dev(ctrl)->bar + off);
2689 	return 0;
2690 }
2691 
2692 static int nvme_pci_get_address(struct nvme_ctrl *ctrl, char *buf, int size)
2693 {
2694 	struct pci_dev *pdev = to_pci_dev(to_nvme_dev(ctrl)->dev);
2695 
2696 	return snprintf(buf, size, "%s", dev_name(&pdev->dev));
2697 }
2698 
2699 static const struct nvme_ctrl_ops nvme_pci_ctrl_ops = {
2700 	.name			= "pcie",
2701 	.module			= THIS_MODULE,
2702 	.flags			= NVME_F_METADATA_SUPPORTED |
2703 				  NVME_F_PCI_P2PDMA,
2704 	.reg_read32		= nvme_pci_reg_read32,
2705 	.reg_write32		= nvme_pci_reg_write32,
2706 	.reg_read64		= nvme_pci_reg_read64,
2707 	.free_ctrl		= nvme_pci_free_ctrl,
2708 	.submit_async_event	= nvme_pci_submit_async_event,
2709 	.get_address		= nvme_pci_get_address,
2710 };
2711 
2712 static int nvme_dev_map(struct nvme_dev *dev)
2713 {
2714 	struct pci_dev *pdev = to_pci_dev(dev->dev);
2715 
2716 	if (pci_request_mem_regions(pdev, "nvme"))
2717 		return -ENODEV;
2718 
2719 	if (nvme_remap_bar(dev, NVME_REG_DBS + 4096))
2720 		goto release;
2721 
2722 	return 0;
2723   release:
2724 	pci_release_mem_regions(pdev);
2725 	return -ENODEV;
2726 }
2727 
2728 static unsigned long check_vendor_combination_bug(struct pci_dev *pdev)
2729 {
2730 	if (pdev->vendor == 0x144d && pdev->device == 0xa802) {
2731 		/*
2732 		 * Several Samsung devices seem to drop off the PCIe bus
2733 		 * randomly when APST is on and uses the deepest sleep state.
2734 		 * This has been observed on a Samsung "SM951 NVMe SAMSUNG
2735 		 * 256GB", a "PM951 NVMe SAMSUNG 512GB", and a "Samsung SSD
2736 		 * 950 PRO 256GB", but it seems to be restricted to two Dell
2737 		 * laptops.
2738 		 */
2739 		if (dmi_match(DMI_SYS_VENDOR, "Dell Inc.") &&
2740 		    (dmi_match(DMI_PRODUCT_NAME, "XPS 15 9550") ||
2741 		     dmi_match(DMI_PRODUCT_NAME, "Precision 5510")))
2742 			return NVME_QUIRK_NO_DEEPEST_PS;
2743 	} else if (pdev->vendor == 0x144d && pdev->device == 0xa804) {
2744 		/*
2745 		 * Samsung SSD 960 EVO drops off the PCIe bus after system
2746 		 * suspend on a Ryzen board, ASUS PRIME B350M-A, as well as
2747 		 * within few minutes after bootup on a Coffee Lake board -
2748 		 * ASUS PRIME Z370-A
2749 		 */
2750 		if (dmi_match(DMI_BOARD_VENDOR, "ASUSTeK COMPUTER INC.") &&
2751 		    (dmi_match(DMI_BOARD_NAME, "PRIME B350M-A") ||
2752 		     dmi_match(DMI_BOARD_NAME, "PRIME Z370-A")))
2753 			return NVME_QUIRK_NO_APST;
2754 	}
2755 
2756 	return 0;
2757 }
2758 
2759 static void nvme_async_probe(void *data, async_cookie_t cookie)
2760 {
2761 	struct nvme_dev *dev = data;
2762 
2763 	flush_work(&dev->ctrl.reset_work);
2764 	flush_work(&dev->ctrl.scan_work);
2765 	nvme_put_ctrl(&dev->ctrl);
2766 }
2767 
2768 static int nvme_probe(struct pci_dev *pdev, const struct pci_device_id *id)
2769 {
2770 	int node, result = -ENOMEM;
2771 	struct nvme_dev *dev;
2772 	unsigned long quirks = id->driver_data;
2773 	size_t alloc_size;
2774 
2775 	node = dev_to_node(&pdev->dev);
2776 	if (node == NUMA_NO_NODE)
2777 		set_dev_node(&pdev->dev, first_memory_node);
2778 
2779 	dev = kzalloc_node(sizeof(*dev), GFP_KERNEL, node);
2780 	if (!dev)
2781 		return -ENOMEM;
2782 
2783 	dev->queues = kcalloc_node(max_queue_count(), sizeof(struct nvme_queue),
2784 					GFP_KERNEL, node);
2785 	if (!dev->queues)
2786 		goto free;
2787 
2788 	dev->dev = get_device(&pdev->dev);
2789 	pci_set_drvdata(pdev, dev);
2790 
2791 	result = nvme_dev_map(dev);
2792 	if (result)
2793 		goto put_pci;
2794 
2795 	INIT_WORK(&dev->ctrl.reset_work, nvme_reset_work);
2796 	INIT_WORK(&dev->remove_work, nvme_remove_dead_ctrl_work);
2797 	mutex_init(&dev->shutdown_lock);
2798 
2799 	result = nvme_setup_prp_pools(dev);
2800 	if (result)
2801 		goto unmap;
2802 
2803 	quirks |= check_vendor_combination_bug(pdev);
2804 
2805 	/*
2806 	 * Double check that our mempool alloc size will cover the biggest
2807 	 * command we support.
2808 	 */
2809 	alloc_size = nvme_pci_iod_alloc_size(dev, NVME_MAX_KB_SZ,
2810 						NVME_MAX_SEGS, true);
2811 	WARN_ON_ONCE(alloc_size > PAGE_SIZE);
2812 
2813 	dev->iod_mempool = mempool_create_node(1, mempool_kmalloc,
2814 						mempool_kfree,
2815 						(void *) alloc_size,
2816 						GFP_KERNEL, node);
2817 	if (!dev->iod_mempool) {
2818 		result = -ENOMEM;
2819 		goto release_pools;
2820 	}
2821 
2822 	result = nvme_init_ctrl(&dev->ctrl, &pdev->dev, &nvme_pci_ctrl_ops,
2823 			quirks);
2824 	if (result)
2825 		goto release_mempool;
2826 
2827 	dev_info(dev->ctrl.device, "pci function %s\n", dev_name(&pdev->dev));
2828 
2829 	nvme_reset_ctrl(&dev->ctrl);
2830 	nvme_get_ctrl(&dev->ctrl);
2831 	async_schedule(nvme_async_probe, dev);
2832 
2833 	return 0;
2834 
2835  release_mempool:
2836 	mempool_destroy(dev->iod_mempool);
2837  release_pools:
2838 	nvme_release_prp_pools(dev);
2839  unmap:
2840 	nvme_dev_unmap(dev);
2841  put_pci:
2842 	put_device(dev->dev);
2843  free:
2844 	kfree(dev->queues);
2845 	kfree(dev);
2846 	return result;
2847 }
2848 
2849 static void nvme_reset_prepare(struct pci_dev *pdev)
2850 {
2851 	struct nvme_dev *dev = pci_get_drvdata(pdev);
2852 
2853 	/*
2854 	 * We don't need to check the return value from waiting for the reset
2855 	 * state as pci_dev device lock is held, making it impossible to race
2856 	 * with ->remove().
2857 	 */
2858 	nvme_disable_prepare_reset(dev, false);
2859 	nvme_sync_queues(&dev->ctrl);
2860 }
2861 
2862 static void nvme_reset_done(struct pci_dev *pdev)
2863 {
2864 	struct nvme_dev *dev = pci_get_drvdata(pdev);
2865 
2866 	if (!nvme_try_sched_reset(&dev->ctrl))
2867 		flush_work(&dev->ctrl.reset_work);
2868 }
2869 
2870 static void nvme_shutdown(struct pci_dev *pdev)
2871 {
2872 	struct nvme_dev *dev = pci_get_drvdata(pdev);
2873 	nvme_disable_prepare_reset(dev, true);
2874 }
2875 
2876 /*
2877  * The driver's remove may be called on a device in a partially initialized
2878  * state. This function must not have any dependencies on the device state in
2879  * order to proceed.
2880  */
2881 static void nvme_remove(struct pci_dev *pdev)
2882 {
2883 	struct nvme_dev *dev = pci_get_drvdata(pdev);
2884 
2885 	nvme_change_ctrl_state(&dev->ctrl, NVME_CTRL_DELETING);
2886 	pci_set_drvdata(pdev, NULL);
2887 
2888 	if (!pci_device_is_present(pdev)) {
2889 		nvme_change_ctrl_state(&dev->ctrl, NVME_CTRL_DEAD);
2890 		nvme_dev_disable(dev, true);
2891 		nvme_dev_remove_admin(dev);
2892 	}
2893 
2894 	flush_work(&dev->ctrl.reset_work);
2895 	nvme_stop_ctrl(&dev->ctrl);
2896 	nvme_remove_namespaces(&dev->ctrl);
2897 	nvme_dev_disable(dev, true);
2898 	nvme_release_cmb(dev);
2899 	nvme_free_host_mem(dev);
2900 	nvme_dev_remove_admin(dev);
2901 	nvme_free_queues(dev, 0);
2902 	nvme_uninit_ctrl(&dev->ctrl);
2903 	nvme_release_prp_pools(dev);
2904 	nvme_dev_unmap(dev);
2905 	nvme_put_ctrl(&dev->ctrl);
2906 }
2907 
2908 #ifdef CONFIG_PM_SLEEP
2909 static int nvme_get_power_state(struct nvme_ctrl *ctrl, u32 *ps)
2910 {
2911 	return nvme_get_features(ctrl, NVME_FEAT_POWER_MGMT, 0, NULL, 0, ps);
2912 }
2913 
2914 static int nvme_set_power_state(struct nvme_ctrl *ctrl, u32 ps)
2915 {
2916 	return nvme_set_features(ctrl, NVME_FEAT_POWER_MGMT, ps, NULL, 0, NULL);
2917 }
2918 
2919 static int nvme_resume(struct device *dev)
2920 {
2921 	struct nvme_dev *ndev = pci_get_drvdata(to_pci_dev(dev));
2922 	struct nvme_ctrl *ctrl = &ndev->ctrl;
2923 
2924 	if (ndev->last_ps == U32_MAX ||
2925 	    nvme_set_power_state(ctrl, ndev->last_ps) != 0)
2926 		return nvme_try_sched_reset(&ndev->ctrl);
2927 	return 0;
2928 }
2929 
2930 static int nvme_suspend(struct device *dev)
2931 {
2932 	struct pci_dev *pdev = to_pci_dev(dev);
2933 	struct nvme_dev *ndev = pci_get_drvdata(pdev);
2934 	struct nvme_ctrl *ctrl = &ndev->ctrl;
2935 	int ret = -EBUSY;
2936 
2937 	ndev->last_ps = U32_MAX;
2938 
2939 	/*
2940 	 * The platform does not remove power for a kernel managed suspend so
2941 	 * use host managed nvme power settings for lowest idle power if
2942 	 * possible. This should have quicker resume latency than a full device
2943 	 * shutdown.  But if the firmware is involved after the suspend or the
2944 	 * device does not support any non-default power states, shut down the
2945 	 * device fully.
2946 	 *
2947 	 * If ASPM is not enabled for the device, shut down the device and allow
2948 	 * the PCI bus layer to put it into D3 in order to take the PCIe link
2949 	 * down, so as to allow the platform to achieve its minimum low-power
2950 	 * state (which may not be possible if the link is up).
2951 	 */
2952 	if (pm_suspend_via_firmware() || !ctrl->npss ||
2953 	    !pcie_aspm_enabled(pdev) ||
2954 	    (ndev->ctrl.quirks & NVME_QUIRK_SIMPLE_SUSPEND))
2955 		return nvme_disable_prepare_reset(ndev, true);
2956 
2957 	nvme_start_freeze(ctrl);
2958 	nvme_wait_freeze(ctrl);
2959 	nvme_sync_queues(ctrl);
2960 
2961 	if (ctrl->state != NVME_CTRL_LIVE)
2962 		goto unfreeze;
2963 
2964 	ret = nvme_get_power_state(ctrl, &ndev->last_ps);
2965 	if (ret < 0)
2966 		goto unfreeze;
2967 
2968 	/*
2969 	 * A saved state prevents pci pm from generically controlling the
2970 	 * device's power. If we're using protocol specific settings, we don't
2971 	 * want pci interfering.
2972 	 */
2973 	pci_save_state(pdev);
2974 
2975 	ret = nvme_set_power_state(ctrl, ctrl->npss);
2976 	if (ret < 0)
2977 		goto unfreeze;
2978 
2979 	if (ret) {
2980 		/* discard the saved state */
2981 		pci_load_saved_state(pdev, NULL);
2982 
2983 		/*
2984 		 * Clearing npss forces a controller reset on resume. The
2985 		 * correct value will be resdicovered then.
2986 		 */
2987 		ret = nvme_disable_prepare_reset(ndev, true);
2988 		ctrl->npss = 0;
2989 	}
2990 unfreeze:
2991 	nvme_unfreeze(ctrl);
2992 	return ret;
2993 }
2994 
2995 static int nvme_simple_suspend(struct device *dev)
2996 {
2997 	struct nvme_dev *ndev = pci_get_drvdata(to_pci_dev(dev));
2998 	return nvme_disable_prepare_reset(ndev, true);
2999 }
3000 
3001 static int nvme_simple_resume(struct device *dev)
3002 {
3003 	struct pci_dev *pdev = to_pci_dev(dev);
3004 	struct nvme_dev *ndev = pci_get_drvdata(pdev);
3005 
3006 	return nvme_try_sched_reset(&ndev->ctrl);
3007 }
3008 
3009 static const struct dev_pm_ops nvme_dev_pm_ops = {
3010 	.suspend	= nvme_suspend,
3011 	.resume		= nvme_resume,
3012 	.freeze		= nvme_simple_suspend,
3013 	.thaw		= nvme_simple_resume,
3014 	.poweroff	= nvme_simple_suspend,
3015 	.restore	= nvme_simple_resume,
3016 };
3017 #endif /* CONFIG_PM_SLEEP */
3018 
3019 static pci_ers_result_t nvme_error_detected(struct pci_dev *pdev,
3020 						pci_channel_state_t state)
3021 {
3022 	struct nvme_dev *dev = pci_get_drvdata(pdev);
3023 
3024 	/*
3025 	 * A frozen channel requires a reset. When detected, this method will
3026 	 * shutdown the controller to quiesce. The controller will be restarted
3027 	 * after the slot reset through driver's slot_reset callback.
3028 	 */
3029 	switch (state) {
3030 	case pci_channel_io_normal:
3031 		return PCI_ERS_RESULT_CAN_RECOVER;
3032 	case pci_channel_io_frozen:
3033 		dev_warn(dev->ctrl.device,
3034 			"frozen state error detected, reset controller\n");
3035 		nvme_dev_disable(dev, false);
3036 		return PCI_ERS_RESULT_NEED_RESET;
3037 	case pci_channel_io_perm_failure:
3038 		dev_warn(dev->ctrl.device,
3039 			"failure state error detected, request disconnect\n");
3040 		return PCI_ERS_RESULT_DISCONNECT;
3041 	}
3042 	return PCI_ERS_RESULT_NEED_RESET;
3043 }
3044 
3045 static pci_ers_result_t nvme_slot_reset(struct pci_dev *pdev)
3046 {
3047 	struct nvme_dev *dev = pci_get_drvdata(pdev);
3048 
3049 	dev_info(dev->ctrl.device, "restart after slot reset\n");
3050 	pci_restore_state(pdev);
3051 	nvme_reset_ctrl(&dev->ctrl);
3052 	return PCI_ERS_RESULT_RECOVERED;
3053 }
3054 
3055 static void nvme_error_resume(struct pci_dev *pdev)
3056 {
3057 	struct nvme_dev *dev = pci_get_drvdata(pdev);
3058 
3059 	flush_work(&dev->ctrl.reset_work);
3060 }
3061 
3062 static const struct pci_error_handlers nvme_err_handler = {
3063 	.error_detected	= nvme_error_detected,
3064 	.slot_reset	= nvme_slot_reset,
3065 	.resume		= nvme_error_resume,
3066 	.reset_prepare	= nvme_reset_prepare,
3067 	.reset_done	= nvme_reset_done,
3068 };
3069 
3070 static const struct pci_device_id nvme_id_table[] = {
3071 	{ PCI_VDEVICE(INTEL, 0x0953),
3072 		.driver_data = NVME_QUIRK_STRIPE_SIZE |
3073 				NVME_QUIRK_DEALLOCATE_ZEROES, },
3074 	{ PCI_VDEVICE(INTEL, 0x0a53),
3075 		.driver_data = NVME_QUIRK_STRIPE_SIZE |
3076 				NVME_QUIRK_DEALLOCATE_ZEROES, },
3077 	{ PCI_VDEVICE(INTEL, 0x0a54),
3078 		.driver_data = NVME_QUIRK_STRIPE_SIZE |
3079 				NVME_QUIRK_DEALLOCATE_ZEROES, },
3080 	{ PCI_VDEVICE(INTEL, 0x0a55),
3081 		.driver_data = NVME_QUIRK_STRIPE_SIZE |
3082 				NVME_QUIRK_DEALLOCATE_ZEROES, },
3083 	{ PCI_VDEVICE(INTEL, 0xf1a5),	/* Intel 600P/P3100 */
3084 		.driver_data = NVME_QUIRK_NO_DEEPEST_PS |
3085 				NVME_QUIRK_MEDIUM_PRIO_SQ },
3086 	{ PCI_VDEVICE(INTEL, 0xf1a6),	/* Intel 760p/Pro 7600p */
3087 		.driver_data = NVME_QUIRK_IGNORE_DEV_SUBNQN, },
3088 	{ PCI_VDEVICE(INTEL, 0x5845),	/* Qemu emulated controller */
3089 		.driver_data = NVME_QUIRK_IDENTIFY_CNS |
3090 				NVME_QUIRK_DISABLE_WRITE_ZEROES, },
3091 	{ PCI_DEVICE(0x1bb1, 0x0100),   /* Seagate Nytro Flash Storage */
3092 		.driver_data = NVME_QUIRK_DELAY_BEFORE_CHK_RDY, },
3093 	{ PCI_DEVICE(0x1c58, 0x0003),	/* HGST adapter */
3094 		.driver_data = NVME_QUIRK_DELAY_BEFORE_CHK_RDY, },
3095 	{ PCI_DEVICE(0x1c58, 0x0023),	/* WDC SN200 adapter */
3096 		.driver_data = NVME_QUIRK_DELAY_BEFORE_CHK_RDY, },
3097 	{ PCI_DEVICE(0x1c5f, 0x0540),	/* Memblaze Pblaze4 adapter */
3098 		.driver_data = NVME_QUIRK_DELAY_BEFORE_CHK_RDY, },
3099 	{ PCI_DEVICE(0x144d, 0xa821),   /* Samsung PM1725 */
3100 		.driver_data = NVME_QUIRK_DELAY_BEFORE_CHK_RDY, },
3101 	{ PCI_DEVICE(0x144d, 0xa822),   /* Samsung PM1725a */
3102 		.driver_data = NVME_QUIRK_DELAY_BEFORE_CHK_RDY, },
3103 	{ PCI_DEVICE(0x1d1d, 0x1f1f),	/* LighNVM qemu device */
3104 		.driver_data = NVME_QUIRK_LIGHTNVM, },
3105 	{ PCI_DEVICE(0x1d1d, 0x2807),	/* CNEX WL */
3106 		.driver_data = NVME_QUIRK_LIGHTNVM, },
3107 	{ PCI_DEVICE(0x1d1d, 0x2601),	/* CNEX Granby */
3108 		.driver_data = NVME_QUIRK_LIGHTNVM, },
3109 	{ PCI_DEVICE(0x10ec, 0x5762),   /* ADATA SX6000LNP */
3110 		.driver_data = NVME_QUIRK_IGNORE_DEV_SUBNQN, },
3111 	{ PCI_DEVICE(0x1cc1, 0x8201),   /* ADATA SX8200PNP 512GB */
3112 		.driver_data = NVME_QUIRK_NO_DEEPEST_PS |
3113 				NVME_QUIRK_IGNORE_DEV_SUBNQN, },
3114 	{ PCI_DEVICE_CLASS(PCI_CLASS_STORAGE_EXPRESS, 0xffffff) },
3115 	{ PCI_DEVICE(PCI_VENDOR_ID_APPLE, 0x2001) },
3116 	{ PCI_DEVICE(PCI_VENDOR_ID_APPLE, 0x2003) },
3117 	{ PCI_DEVICE(PCI_VENDOR_ID_APPLE, 0x2005),
3118 		.driver_data = NVME_QUIRK_SINGLE_VECTOR |
3119 				NVME_QUIRK_128_BYTES_SQES |
3120 				NVME_QUIRK_SHARED_TAGS },
3121 	{ 0, }
3122 };
3123 MODULE_DEVICE_TABLE(pci, nvme_id_table);
3124 
3125 static struct pci_driver nvme_driver = {
3126 	.name		= "nvme",
3127 	.id_table	= nvme_id_table,
3128 	.probe		= nvme_probe,
3129 	.remove		= nvme_remove,
3130 	.shutdown	= nvme_shutdown,
3131 #ifdef CONFIG_PM_SLEEP
3132 	.driver		= {
3133 		.pm	= &nvme_dev_pm_ops,
3134 	},
3135 #endif
3136 	.sriov_configure = pci_sriov_configure_simple,
3137 	.err_handler	= &nvme_err_handler,
3138 };
3139 
3140 static int __init nvme_init(void)
3141 {
3142 	BUILD_BUG_ON(sizeof(struct nvme_create_cq) != 64);
3143 	BUILD_BUG_ON(sizeof(struct nvme_create_sq) != 64);
3144 	BUILD_BUG_ON(sizeof(struct nvme_delete_queue) != 64);
3145 	BUILD_BUG_ON(IRQ_AFFINITY_MAX_SETS < 2);
3146 	return pci_register_driver(&nvme_driver);
3147 }
3148 
3149 static void __exit nvme_exit(void)
3150 {
3151 	pci_unregister_driver(&nvme_driver);
3152 	flush_workqueue(nvme_wq);
3153 }
3154 
3155 MODULE_AUTHOR("Matthew Wilcox <willy@linux.intel.com>");
3156 MODULE_LICENSE("GPL");
3157 MODULE_VERSION("1.0");
3158 module_init(nvme_init);
3159 module_exit(nvme_exit);
3160