xref: /openbmc/linux/drivers/nvme/host/pci.c (revision 09b35b41)
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 first_prp_len = dev->ctrl.page_size - bv->bv_offset;
777 
778 	iod->first_dma = dma_map_bvec(dev->dev, bv, rq_dma_dir(req), 0);
779 	if (dma_mapping_error(dev->dev, iod->first_dma))
780 		return BLK_STS_RESOURCE;
781 	iod->dma_len = bv->bv_len;
782 
783 	cmnd->dptr.prp1 = cpu_to_le64(iod->first_dma);
784 	if (bv->bv_len > first_prp_len)
785 		cmnd->dptr.prp2 = cpu_to_le64(iod->first_dma + first_prp_len);
786 	return 0;
787 }
788 
789 static blk_status_t nvme_setup_sgl_simple(struct nvme_dev *dev,
790 		struct request *req, struct nvme_rw_command *cmnd,
791 		struct bio_vec *bv)
792 {
793 	struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
794 
795 	iod->first_dma = dma_map_bvec(dev->dev, bv, rq_dma_dir(req), 0);
796 	if (dma_mapping_error(dev->dev, iod->first_dma))
797 		return BLK_STS_RESOURCE;
798 	iod->dma_len = bv->bv_len;
799 
800 	cmnd->flags = NVME_CMD_SGL_METABUF;
801 	cmnd->dptr.sgl.addr = cpu_to_le64(iod->first_dma);
802 	cmnd->dptr.sgl.length = cpu_to_le32(iod->dma_len);
803 	cmnd->dptr.sgl.type = NVME_SGL_FMT_DATA_DESC << 4;
804 	return 0;
805 }
806 
807 static blk_status_t nvme_map_data(struct nvme_dev *dev, struct request *req,
808 		struct nvme_command *cmnd)
809 {
810 	struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
811 	blk_status_t ret = BLK_STS_RESOURCE;
812 	int nr_mapped;
813 
814 	if (blk_rq_nr_phys_segments(req) == 1) {
815 		struct bio_vec bv = req_bvec(req);
816 
817 		if (!is_pci_p2pdma_page(bv.bv_page)) {
818 			if (bv.bv_offset + bv.bv_len <= dev->ctrl.page_size * 2)
819 				return nvme_setup_prp_simple(dev, req,
820 							     &cmnd->rw, &bv);
821 
822 			if (iod->nvmeq->qid &&
823 			    dev->ctrl.sgls & ((1 << 0) | (1 << 1)))
824 				return nvme_setup_sgl_simple(dev, req,
825 							     &cmnd->rw, &bv);
826 		}
827 	}
828 
829 	iod->dma_len = 0;
830 	iod->sg = mempool_alloc(dev->iod_mempool, GFP_ATOMIC);
831 	if (!iod->sg)
832 		return BLK_STS_RESOURCE;
833 	sg_init_table(iod->sg, blk_rq_nr_phys_segments(req));
834 	iod->nents = blk_rq_map_sg(req->q, req, iod->sg);
835 	if (!iod->nents)
836 		goto out;
837 
838 	if (is_pci_p2pdma_page(sg_page(iod->sg)))
839 		nr_mapped = pci_p2pdma_map_sg_attrs(dev->dev, iod->sg,
840 				iod->nents, rq_dma_dir(req), DMA_ATTR_NO_WARN);
841 	else
842 		nr_mapped = dma_map_sg_attrs(dev->dev, iod->sg, iod->nents,
843 					     rq_dma_dir(req), DMA_ATTR_NO_WARN);
844 	if (!nr_mapped)
845 		goto out;
846 
847 	iod->use_sgl = nvme_pci_use_sgls(dev, req);
848 	if (iod->use_sgl)
849 		ret = nvme_pci_setup_sgls(dev, req, &cmnd->rw, nr_mapped);
850 	else
851 		ret = nvme_pci_setup_prps(dev, req, &cmnd->rw);
852 out:
853 	if (ret != BLK_STS_OK)
854 		nvme_unmap_data(dev, req);
855 	return ret;
856 }
857 
858 static blk_status_t nvme_map_metadata(struct nvme_dev *dev, struct request *req,
859 		struct nvme_command *cmnd)
860 {
861 	struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
862 
863 	iod->meta_dma = dma_map_bvec(dev->dev, rq_integrity_vec(req),
864 			rq_dma_dir(req), 0);
865 	if (dma_mapping_error(dev->dev, iod->meta_dma))
866 		return BLK_STS_IOERR;
867 	cmnd->rw.metadata = cpu_to_le64(iod->meta_dma);
868 	return 0;
869 }
870 
871 /*
872  * NOTE: ns is NULL when called on the admin queue.
873  */
874 static blk_status_t nvme_queue_rq(struct blk_mq_hw_ctx *hctx,
875 			 const struct blk_mq_queue_data *bd)
876 {
877 	struct nvme_ns *ns = hctx->queue->queuedata;
878 	struct nvme_queue *nvmeq = hctx->driver_data;
879 	struct nvme_dev *dev = nvmeq->dev;
880 	struct request *req = bd->rq;
881 	struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
882 	struct nvme_command cmnd;
883 	blk_status_t ret;
884 
885 	iod->aborted = 0;
886 	iod->npages = -1;
887 	iod->nents = 0;
888 
889 	/*
890 	 * We should not need to do this, but we're still using this to
891 	 * ensure we can drain requests on a dying queue.
892 	 */
893 	if (unlikely(!test_bit(NVMEQ_ENABLED, &nvmeq->flags)))
894 		return BLK_STS_IOERR;
895 
896 	ret = nvme_setup_cmd(ns, req, &cmnd);
897 	if (ret)
898 		return ret;
899 
900 	if (blk_rq_nr_phys_segments(req)) {
901 		ret = nvme_map_data(dev, req, &cmnd);
902 		if (ret)
903 			goto out_free_cmd;
904 	}
905 
906 	if (blk_integrity_rq(req)) {
907 		ret = nvme_map_metadata(dev, req, &cmnd);
908 		if (ret)
909 			goto out_unmap_data;
910 	}
911 
912 	blk_mq_start_request(req);
913 	nvme_submit_cmd(nvmeq, &cmnd, bd->last);
914 	return BLK_STS_OK;
915 out_unmap_data:
916 	nvme_unmap_data(dev, req);
917 out_free_cmd:
918 	nvme_cleanup_cmd(req);
919 	return ret;
920 }
921 
922 static void nvme_pci_complete_rq(struct request *req)
923 {
924 	struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
925 	struct nvme_dev *dev = iod->nvmeq->dev;
926 
927 	nvme_cleanup_cmd(req);
928 	if (blk_integrity_rq(req))
929 		dma_unmap_page(dev->dev, iod->meta_dma,
930 			       rq_integrity_vec(req)->bv_len, rq_data_dir(req));
931 	if (blk_rq_nr_phys_segments(req))
932 		nvme_unmap_data(dev, req);
933 	nvme_complete_rq(req);
934 }
935 
936 /* We read the CQE phase first to check if the rest of the entry is valid */
937 static inline bool nvme_cqe_pending(struct nvme_queue *nvmeq)
938 {
939 	return (le16_to_cpu(nvmeq->cqes[nvmeq->cq_head].status) & 1) ==
940 			nvmeq->cq_phase;
941 }
942 
943 static inline void nvme_ring_cq_doorbell(struct nvme_queue *nvmeq)
944 {
945 	u16 head = nvmeq->cq_head;
946 
947 	if (nvme_dbbuf_update_and_check_event(head, nvmeq->dbbuf_cq_db,
948 					      nvmeq->dbbuf_cq_ei))
949 		writel(head, nvmeq->q_db + nvmeq->dev->db_stride);
950 }
951 
952 static inline void nvme_handle_cqe(struct nvme_queue *nvmeq, u16 idx)
953 {
954 	volatile struct nvme_completion *cqe = &nvmeq->cqes[idx];
955 	struct request *req;
956 
957 	if (unlikely(cqe->command_id >= nvmeq->q_depth)) {
958 		dev_warn(nvmeq->dev->ctrl.device,
959 			"invalid id %d completed on queue %d\n",
960 			cqe->command_id, le16_to_cpu(cqe->sq_id));
961 		return;
962 	}
963 
964 	/*
965 	 * AEN requests are special as they don't time out and can
966 	 * survive any kind of queue freeze and often don't respond to
967 	 * aborts.  We don't even bother to allocate a struct request
968 	 * for them but rather special case them here.
969 	 */
970 	if (unlikely(nvmeq->qid == 0 &&
971 			cqe->command_id >= NVME_AQ_BLK_MQ_DEPTH)) {
972 		nvme_complete_async_event(&nvmeq->dev->ctrl,
973 				cqe->status, &cqe->result);
974 		return;
975 	}
976 
977 	req = blk_mq_tag_to_rq(*nvmeq->tags, cqe->command_id);
978 	trace_nvme_sq(req, cqe->sq_head, nvmeq->sq_tail);
979 	nvme_end_request(req, cqe->status, cqe->result);
980 }
981 
982 static void nvme_complete_cqes(struct nvme_queue *nvmeq, u16 start, u16 end)
983 {
984 	while (start != end) {
985 		nvme_handle_cqe(nvmeq, start);
986 		if (++start == nvmeq->q_depth)
987 			start = 0;
988 	}
989 }
990 
991 static inline void nvme_update_cq_head(struct nvme_queue *nvmeq)
992 {
993 	if (nvmeq->cq_head == nvmeq->q_depth - 1) {
994 		nvmeq->cq_head = 0;
995 		nvmeq->cq_phase = !nvmeq->cq_phase;
996 	} else {
997 		nvmeq->cq_head++;
998 	}
999 }
1000 
1001 static inline int nvme_process_cq(struct nvme_queue *nvmeq, u16 *start,
1002 				  u16 *end, unsigned int tag)
1003 {
1004 	int found = 0;
1005 
1006 	*start = nvmeq->cq_head;
1007 	while (nvme_cqe_pending(nvmeq)) {
1008 		if (tag == -1U || nvmeq->cqes[nvmeq->cq_head].command_id == tag)
1009 			found++;
1010 		nvme_update_cq_head(nvmeq);
1011 	}
1012 	*end = nvmeq->cq_head;
1013 
1014 	if (*start != *end)
1015 		nvme_ring_cq_doorbell(nvmeq);
1016 	return found;
1017 }
1018 
1019 static irqreturn_t nvme_irq(int irq, void *data)
1020 {
1021 	struct nvme_queue *nvmeq = data;
1022 	irqreturn_t ret = IRQ_NONE;
1023 	u16 start, end;
1024 
1025 	/*
1026 	 * The rmb/wmb pair ensures we see all updates from a previous run of
1027 	 * the irq handler, even if that was on another CPU.
1028 	 */
1029 	rmb();
1030 	if (nvmeq->cq_head != nvmeq->last_cq_head)
1031 		ret = IRQ_HANDLED;
1032 	nvme_process_cq(nvmeq, &start, &end, -1);
1033 	nvmeq->last_cq_head = nvmeq->cq_head;
1034 	wmb();
1035 
1036 	if (start != end) {
1037 		nvme_complete_cqes(nvmeq, start, end);
1038 		return IRQ_HANDLED;
1039 	}
1040 
1041 	return ret;
1042 }
1043 
1044 static irqreturn_t nvme_irq_check(int irq, void *data)
1045 {
1046 	struct nvme_queue *nvmeq = data;
1047 	if (nvme_cqe_pending(nvmeq))
1048 		return IRQ_WAKE_THREAD;
1049 	return IRQ_NONE;
1050 }
1051 
1052 /*
1053  * Poll for completions any queue, including those not dedicated to polling.
1054  * Can be called from any context.
1055  */
1056 static int nvme_poll_irqdisable(struct nvme_queue *nvmeq, unsigned int tag)
1057 {
1058 	struct pci_dev *pdev = to_pci_dev(nvmeq->dev->dev);
1059 	u16 start, end;
1060 	int found;
1061 
1062 	/*
1063 	 * For a poll queue we need to protect against the polling thread
1064 	 * using the CQ lock.  For normal interrupt driven threads we have
1065 	 * to disable the interrupt to avoid racing with it.
1066 	 */
1067 	if (test_bit(NVMEQ_POLLED, &nvmeq->flags)) {
1068 		spin_lock(&nvmeq->cq_poll_lock);
1069 		found = nvme_process_cq(nvmeq, &start, &end, tag);
1070 		spin_unlock(&nvmeq->cq_poll_lock);
1071 	} else {
1072 		disable_irq(pci_irq_vector(pdev, nvmeq->cq_vector));
1073 		found = nvme_process_cq(nvmeq, &start, &end, tag);
1074 		enable_irq(pci_irq_vector(pdev, nvmeq->cq_vector));
1075 	}
1076 
1077 	nvme_complete_cqes(nvmeq, start, end);
1078 	return found;
1079 }
1080 
1081 static int nvme_poll(struct blk_mq_hw_ctx *hctx)
1082 {
1083 	struct nvme_queue *nvmeq = hctx->driver_data;
1084 	u16 start, end;
1085 	bool found;
1086 
1087 	if (!nvme_cqe_pending(nvmeq))
1088 		return 0;
1089 
1090 	spin_lock(&nvmeq->cq_poll_lock);
1091 	found = nvme_process_cq(nvmeq, &start, &end, -1);
1092 	spin_unlock(&nvmeq->cq_poll_lock);
1093 
1094 	nvme_complete_cqes(nvmeq, start, end);
1095 	return found;
1096 }
1097 
1098 static void nvme_pci_submit_async_event(struct nvme_ctrl *ctrl)
1099 {
1100 	struct nvme_dev *dev = to_nvme_dev(ctrl);
1101 	struct nvme_queue *nvmeq = &dev->queues[0];
1102 	struct nvme_command c;
1103 
1104 	memset(&c, 0, sizeof(c));
1105 	c.common.opcode = nvme_admin_async_event;
1106 	c.common.command_id = NVME_AQ_BLK_MQ_DEPTH;
1107 	nvme_submit_cmd(nvmeq, &c, true);
1108 }
1109 
1110 static int adapter_delete_queue(struct nvme_dev *dev, u8 opcode, u16 id)
1111 {
1112 	struct nvme_command c;
1113 
1114 	memset(&c, 0, sizeof(c));
1115 	c.delete_queue.opcode = opcode;
1116 	c.delete_queue.qid = cpu_to_le16(id);
1117 
1118 	return nvme_submit_sync_cmd(dev->ctrl.admin_q, &c, NULL, 0);
1119 }
1120 
1121 static int adapter_alloc_cq(struct nvme_dev *dev, u16 qid,
1122 		struct nvme_queue *nvmeq, s16 vector)
1123 {
1124 	struct nvme_command c;
1125 	int flags = NVME_QUEUE_PHYS_CONTIG;
1126 
1127 	if (!test_bit(NVMEQ_POLLED, &nvmeq->flags))
1128 		flags |= NVME_CQ_IRQ_ENABLED;
1129 
1130 	/*
1131 	 * Note: we (ab)use the fact that the prp fields survive if no data
1132 	 * is attached to the request.
1133 	 */
1134 	memset(&c, 0, sizeof(c));
1135 	c.create_cq.opcode = nvme_admin_create_cq;
1136 	c.create_cq.prp1 = cpu_to_le64(nvmeq->cq_dma_addr);
1137 	c.create_cq.cqid = cpu_to_le16(qid);
1138 	c.create_cq.qsize = cpu_to_le16(nvmeq->q_depth - 1);
1139 	c.create_cq.cq_flags = cpu_to_le16(flags);
1140 	c.create_cq.irq_vector = cpu_to_le16(vector);
1141 
1142 	return nvme_submit_sync_cmd(dev->ctrl.admin_q, &c, NULL, 0);
1143 }
1144 
1145 static int adapter_alloc_sq(struct nvme_dev *dev, u16 qid,
1146 						struct nvme_queue *nvmeq)
1147 {
1148 	struct nvme_ctrl *ctrl = &dev->ctrl;
1149 	struct nvme_command c;
1150 	int flags = NVME_QUEUE_PHYS_CONTIG;
1151 
1152 	/*
1153 	 * Some drives have a bug that auto-enables WRRU if MEDIUM isn't
1154 	 * set. Since URGENT priority is zeroes, it makes all queues
1155 	 * URGENT.
1156 	 */
1157 	if (ctrl->quirks & NVME_QUIRK_MEDIUM_PRIO_SQ)
1158 		flags |= NVME_SQ_PRIO_MEDIUM;
1159 
1160 	/*
1161 	 * Note: we (ab)use the fact that the prp fields survive if no data
1162 	 * is attached to the request.
1163 	 */
1164 	memset(&c, 0, sizeof(c));
1165 	c.create_sq.opcode = nvme_admin_create_sq;
1166 	c.create_sq.prp1 = cpu_to_le64(nvmeq->sq_dma_addr);
1167 	c.create_sq.sqid = cpu_to_le16(qid);
1168 	c.create_sq.qsize = cpu_to_le16(nvmeq->q_depth - 1);
1169 	c.create_sq.sq_flags = cpu_to_le16(flags);
1170 	c.create_sq.cqid = cpu_to_le16(qid);
1171 
1172 	return nvme_submit_sync_cmd(dev->ctrl.admin_q, &c, NULL, 0);
1173 }
1174 
1175 static int adapter_delete_cq(struct nvme_dev *dev, u16 cqid)
1176 {
1177 	return adapter_delete_queue(dev, nvme_admin_delete_cq, cqid);
1178 }
1179 
1180 static int adapter_delete_sq(struct nvme_dev *dev, u16 sqid)
1181 {
1182 	return adapter_delete_queue(dev, nvme_admin_delete_sq, sqid);
1183 }
1184 
1185 static void abort_endio(struct request *req, blk_status_t error)
1186 {
1187 	struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
1188 	struct nvme_queue *nvmeq = iod->nvmeq;
1189 
1190 	dev_warn(nvmeq->dev->ctrl.device,
1191 		 "Abort status: 0x%x", nvme_req(req)->status);
1192 	atomic_inc(&nvmeq->dev->ctrl.abort_limit);
1193 	blk_mq_free_request(req);
1194 }
1195 
1196 static bool nvme_should_reset(struct nvme_dev *dev, u32 csts)
1197 {
1198 
1199 	/* If true, indicates loss of adapter communication, possibly by a
1200 	 * NVMe Subsystem reset.
1201 	 */
1202 	bool nssro = dev->subsystem && (csts & NVME_CSTS_NSSRO);
1203 
1204 	/* If there is a reset/reinit ongoing, we shouldn't reset again. */
1205 	switch (dev->ctrl.state) {
1206 	case NVME_CTRL_RESETTING:
1207 	case NVME_CTRL_CONNECTING:
1208 		return false;
1209 	default:
1210 		break;
1211 	}
1212 
1213 	/* We shouldn't reset unless the controller is on fatal error state
1214 	 * _or_ if we lost the communication with it.
1215 	 */
1216 	if (!(csts & NVME_CSTS_CFS) && !nssro)
1217 		return false;
1218 
1219 	return true;
1220 }
1221 
1222 static void nvme_warn_reset(struct nvme_dev *dev, u32 csts)
1223 {
1224 	/* Read a config register to help see what died. */
1225 	u16 pci_status;
1226 	int result;
1227 
1228 	result = pci_read_config_word(to_pci_dev(dev->dev), PCI_STATUS,
1229 				      &pci_status);
1230 	if (result == PCIBIOS_SUCCESSFUL)
1231 		dev_warn(dev->ctrl.device,
1232 			 "controller is down; will reset: CSTS=0x%x, PCI_STATUS=0x%hx\n",
1233 			 csts, pci_status);
1234 	else
1235 		dev_warn(dev->ctrl.device,
1236 			 "controller is down; will reset: CSTS=0x%x, PCI_STATUS read failed (%d)\n",
1237 			 csts, result);
1238 }
1239 
1240 static enum blk_eh_timer_return nvme_timeout(struct request *req, bool reserved)
1241 {
1242 	struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
1243 	struct nvme_queue *nvmeq = iod->nvmeq;
1244 	struct nvme_dev *dev = nvmeq->dev;
1245 	struct request *abort_req;
1246 	struct nvme_command cmd;
1247 	u32 csts = readl(dev->bar + NVME_REG_CSTS);
1248 
1249 	/* If PCI error recovery process is happening, we cannot reset or
1250 	 * the recovery mechanism will surely fail.
1251 	 */
1252 	mb();
1253 	if (pci_channel_offline(to_pci_dev(dev->dev)))
1254 		return BLK_EH_RESET_TIMER;
1255 
1256 	/*
1257 	 * Reset immediately if the controller is failed
1258 	 */
1259 	if (nvme_should_reset(dev, csts)) {
1260 		nvme_warn_reset(dev, csts);
1261 		nvme_dev_disable(dev, false);
1262 		nvme_reset_ctrl(&dev->ctrl);
1263 		return BLK_EH_DONE;
1264 	}
1265 
1266 	/*
1267 	 * Did we miss an interrupt?
1268 	 */
1269 	if (nvme_poll_irqdisable(nvmeq, req->tag)) {
1270 		dev_warn(dev->ctrl.device,
1271 			 "I/O %d QID %d timeout, completion polled\n",
1272 			 req->tag, nvmeq->qid);
1273 		return BLK_EH_DONE;
1274 	}
1275 
1276 	/*
1277 	 * Shutdown immediately if controller times out while starting. The
1278 	 * reset work will see the pci device disabled when it gets the forced
1279 	 * cancellation error. All outstanding requests are completed on
1280 	 * shutdown, so we return BLK_EH_DONE.
1281 	 */
1282 	switch (dev->ctrl.state) {
1283 	case NVME_CTRL_CONNECTING:
1284 		nvme_change_ctrl_state(&dev->ctrl, NVME_CTRL_DELETING);
1285 		/* fall through */
1286 	case NVME_CTRL_DELETING:
1287 		dev_warn_ratelimited(dev->ctrl.device,
1288 			 "I/O %d QID %d timeout, disable controller\n",
1289 			 req->tag, nvmeq->qid);
1290 		nvme_dev_disable(dev, true);
1291 		nvme_req(req)->flags |= NVME_REQ_CANCELLED;
1292 		return BLK_EH_DONE;
1293 	case NVME_CTRL_RESETTING:
1294 		return BLK_EH_RESET_TIMER;
1295 	default:
1296 		break;
1297 	}
1298 
1299 	/*
1300  	 * Shutdown the controller immediately and schedule a reset if the
1301  	 * command was already aborted once before and still hasn't been
1302  	 * returned to the driver, or if this is the admin queue.
1303 	 */
1304 	if (!nvmeq->qid || iod->aborted) {
1305 		dev_warn(dev->ctrl.device,
1306 			 "I/O %d QID %d timeout, reset controller\n",
1307 			 req->tag, nvmeq->qid);
1308 		nvme_dev_disable(dev, false);
1309 		nvme_reset_ctrl(&dev->ctrl);
1310 
1311 		nvme_req(req)->flags |= NVME_REQ_CANCELLED;
1312 		return BLK_EH_DONE;
1313 	}
1314 
1315 	if (atomic_dec_return(&dev->ctrl.abort_limit) < 0) {
1316 		atomic_inc(&dev->ctrl.abort_limit);
1317 		return BLK_EH_RESET_TIMER;
1318 	}
1319 	iod->aborted = 1;
1320 
1321 	memset(&cmd, 0, sizeof(cmd));
1322 	cmd.abort.opcode = nvme_admin_abort_cmd;
1323 	cmd.abort.cid = req->tag;
1324 	cmd.abort.sqid = cpu_to_le16(nvmeq->qid);
1325 
1326 	dev_warn(nvmeq->dev->ctrl.device,
1327 		"I/O %d QID %d timeout, aborting\n",
1328 		 req->tag, nvmeq->qid);
1329 
1330 	abort_req = nvme_alloc_request(dev->ctrl.admin_q, &cmd,
1331 			BLK_MQ_REQ_NOWAIT, NVME_QID_ANY);
1332 	if (IS_ERR(abort_req)) {
1333 		atomic_inc(&dev->ctrl.abort_limit);
1334 		return BLK_EH_RESET_TIMER;
1335 	}
1336 
1337 	abort_req->timeout = ADMIN_TIMEOUT;
1338 	abort_req->end_io_data = NULL;
1339 	blk_execute_rq_nowait(abort_req->q, NULL, abort_req, 0, abort_endio);
1340 
1341 	/*
1342 	 * The aborted req will be completed on receiving the abort req.
1343 	 * We enable the timer again. If hit twice, it'll cause a device reset,
1344 	 * as the device then is in a faulty state.
1345 	 */
1346 	return BLK_EH_RESET_TIMER;
1347 }
1348 
1349 static void nvme_free_queue(struct nvme_queue *nvmeq)
1350 {
1351 	dma_free_coherent(nvmeq->dev->dev, CQ_SIZE(nvmeq),
1352 				(void *)nvmeq->cqes, nvmeq->cq_dma_addr);
1353 	if (!nvmeq->sq_cmds)
1354 		return;
1355 
1356 	if (test_and_clear_bit(NVMEQ_SQ_CMB, &nvmeq->flags)) {
1357 		pci_free_p2pmem(to_pci_dev(nvmeq->dev->dev),
1358 				nvmeq->sq_cmds, SQ_SIZE(nvmeq));
1359 	} else {
1360 		dma_free_coherent(nvmeq->dev->dev, SQ_SIZE(nvmeq),
1361 				nvmeq->sq_cmds, nvmeq->sq_dma_addr);
1362 	}
1363 }
1364 
1365 static void nvme_free_queues(struct nvme_dev *dev, int lowest)
1366 {
1367 	int i;
1368 
1369 	for (i = dev->ctrl.queue_count - 1; i >= lowest; i--) {
1370 		dev->ctrl.queue_count--;
1371 		nvme_free_queue(&dev->queues[i]);
1372 	}
1373 }
1374 
1375 /**
1376  * nvme_suspend_queue - put queue into suspended state
1377  * @nvmeq: queue to suspend
1378  */
1379 static int nvme_suspend_queue(struct nvme_queue *nvmeq)
1380 {
1381 	if (!test_and_clear_bit(NVMEQ_ENABLED, &nvmeq->flags))
1382 		return 1;
1383 
1384 	/* ensure that nvme_queue_rq() sees NVMEQ_ENABLED cleared */
1385 	mb();
1386 
1387 	nvmeq->dev->online_queues--;
1388 	if (!nvmeq->qid && nvmeq->dev->ctrl.admin_q)
1389 		blk_mq_quiesce_queue(nvmeq->dev->ctrl.admin_q);
1390 	if (!test_and_clear_bit(NVMEQ_POLLED, &nvmeq->flags))
1391 		pci_free_irq(to_pci_dev(nvmeq->dev->dev), nvmeq->cq_vector, nvmeq);
1392 	return 0;
1393 }
1394 
1395 static void nvme_suspend_io_queues(struct nvme_dev *dev)
1396 {
1397 	int i;
1398 
1399 	for (i = dev->ctrl.queue_count - 1; i > 0; i--)
1400 		nvme_suspend_queue(&dev->queues[i]);
1401 }
1402 
1403 static void nvme_disable_admin_queue(struct nvme_dev *dev, bool shutdown)
1404 {
1405 	struct nvme_queue *nvmeq = &dev->queues[0];
1406 
1407 	if (shutdown)
1408 		nvme_shutdown_ctrl(&dev->ctrl);
1409 	else
1410 		nvme_disable_ctrl(&dev->ctrl);
1411 
1412 	nvme_poll_irqdisable(nvmeq, -1);
1413 }
1414 
1415 static int nvme_cmb_qdepth(struct nvme_dev *dev, int nr_io_queues,
1416 				int entry_size)
1417 {
1418 	int q_depth = dev->q_depth;
1419 	unsigned q_size_aligned = roundup(q_depth * entry_size,
1420 					  dev->ctrl.page_size);
1421 
1422 	if (q_size_aligned * nr_io_queues > dev->cmb_size) {
1423 		u64 mem_per_q = div_u64(dev->cmb_size, nr_io_queues);
1424 		mem_per_q = round_down(mem_per_q, dev->ctrl.page_size);
1425 		q_depth = div_u64(mem_per_q, entry_size);
1426 
1427 		/*
1428 		 * Ensure the reduced q_depth is above some threshold where it
1429 		 * would be better to map queues in system memory with the
1430 		 * original depth
1431 		 */
1432 		if (q_depth < 64)
1433 			return -ENOMEM;
1434 	}
1435 
1436 	return q_depth;
1437 }
1438 
1439 static int nvme_alloc_sq_cmds(struct nvme_dev *dev, struct nvme_queue *nvmeq,
1440 				int qid)
1441 {
1442 	struct pci_dev *pdev = to_pci_dev(dev->dev);
1443 
1444 	if (qid && dev->cmb_use_sqes && (dev->cmbsz & NVME_CMBSZ_SQS)) {
1445 		nvmeq->sq_cmds = pci_alloc_p2pmem(pdev, SQ_SIZE(nvmeq));
1446 		if (nvmeq->sq_cmds) {
1447 			nvmeq->sq_dma_addr = pci_p2pmem_virt_to_bus(pdev,
1448 							nvmeq->sq_cmds);
1449 			if (nvmeq->sq_dma_addr) {
1450 				set_bit(NVMEQ_SQ_CMB, &nvmeq->flags);
1451 				return 0;
1452 			}
1453 
1454 			pci_free_p2pmem(pdev, nvmeq->sq_cmds, SQ_SIZE(nvmeq));
1455 		}
1456 	}
1457 
1458 	nvmeq->sq_cmds = dma_alloc_coherent(dev->dev, SQ_SIZE(nvmeq),
1459 				&nvmeq->sq_dma_addr, GFP_KERNEL);
1460 	if (!nvmeq->sq_cmds)
1461 		return -ENOMEM;
1462 	return 0;
1463 }
1464 
1465 static int nvme_alloc_queue(struct nvme_dev *dev, int qid, int depth)
1466 {
1467 	struct nvme_queue *nvmeq = &dev->queues[qid];
1468 
1469 	if (dev->ctrl.queue_count > qid)
1470 		return 0;
1471 
1472 	nvmeq->sqes = qid ? dev->io_sqes : NVME_ADM_SQES;
1473 	nvmeq->q_depth = depth;
1474 	nvmeq->cqes = dma_alloc_coherent(dev->dev, CQ_SIZE(nvmeq),
1475 					 &nvmeq->cq_dma_addr, GFP_KERNEL);
1476 	if (!nvmeq->cqes)
1477 		goto free_nvmeq;
1478 
1479 	if (nvme_alloc_sq_cmds(dev, nvmeq, qid))
1480 		goto free_cqdma;
1481 
1482 	nvmeq->dev = dev;
1483 	spin_lock_init(&nvmeq->sq_lock);
1484 	spin_lock_init(&nvmeq->cq_poll_lock);
1485 	nvmeq->cq_head = 0;
1486 	nvmeq->cq_phase = 1;
1487 	nvmeq->q_db = &dev->dbs[qid * 2 * dev->db_stride];
1488 	nvmeq->qid = qid;
1489 	dev->ctrl.queue_count++;
1490 
1491 	return 0;
1492 
1493  free_cqdma:
1494 	dma_free_coherent(dev->dev, CQ_SIZE(nvmeq), (void *)nvmeq->cqes,
1495 			  nvmeq->cq_dma_addr);
1496  free_nvmeq:
1497 	return -ENOMEM;
1498 }
1499 
1500 static int queue_request_irq(struct nvme_queue *nvmeq)
1501 {
1502 	struct pci_dev *pdev = to_pci_dev(nvmeq->dev->dev);
1503 	int nr = nvmeq->dev->ctrl.instance;
1504 
1505 	if (use_threaded_interrupts) {
1506 		return pci_request_irq(pdev, nvmeq->cq_vector, nvme_irq_check,
1507 				nvme_irq, nvmeq, "nvme%dq%d", nr, nvmeq->qid);
1508 	} else {
1509 		return pci_request_irq(pdev, nvmeq->cq_vector, nvme_irq,
1510 				NULL, nvmeq, "nvme%dq%d", nr, nvmeq->qid);
1511 	}
1512 }
1513 
1514 static void nvme_init_queue(struct nvme_queue *nvmeq, u16 qid)
1515 {
1516 	struct nvme_dev *dev = nvmeq->dev;
1517 
1518 	nvmeq->sq_tail = 0;
1519 	nvmeq->last_sq_tail = 0;
1520 	nvmeq->cq_head = 0;
1521 	nvmeq->cq_phase = 1;
1522 	nvmeq->q_db = &dev->dbs[qid * 2 * dev->db_stride];
1523 	memset((void *)nvmeq->cqes, 0, CQ_SIZE(nvmeq));
1524 	nvme_dbbuf_init(dev, nvmeq, qid);
1525 	dev->online_queues++;
1526 	wmb(); /* ensure the first interrupt sees the initialization */
1527 }
1528 
1529 static int nvme_create_queue(struct nvme_queue *nvmeq, int qid, bool polled)
1530 {
1531 	struct nvme_dev *dev = nvmeq->dev;
1532 	int result;
1533 	u16 vector = 0;
1534 
1535 	clear_bit(NVMEQ_DELETE_ERROR, &nvmeq->flags);
1536 
1537 	/*
1538 	 * A queue's vector matches the queue identifier unless the controller
1539 	 * has only one vector available.
1540 	 */
1541 	if (!polled)
1542 		vector = dev->num_vecs == 1 ? 0 : qid;
1543 	else
1544 		set_bit(NVMEQ_POLLED, &nvmeq->flags);
1545 
1546 	result = adapter_alloc_cq(dev, qid, nvmeq, vector);
1547 	if (result)
1548 		return result;
1549 
1550 	result = adapter_alloc_sq(dev, qid, nvmeq);
1551 	if (result < 0)
1552 		return result;
1553 	else if (result)
1554 		goto release_cq;
1555 
1556 	nvmeq->cq_vector = vector;
1557 	nvme_init_queue(nvmeq, qid);
1558 
1559 	if (!polled) {
1560 		result = queue_request_irq(nvmeq);
1561 		if (result < 0)
1562 			goto release_sq;
1563 	}
1564 
1565 	set_bit(NVMEQ_ENABLED, &nvmeq->flags);
1566 	return result;
1567 
1568 release_sq:
1569 	dev->online_queues--;
1570 	adapter_delete_sq(dev, qid);
1571 release_cq:
1572 	adapter_delete_cq(dev, qid);
1573 	return result;
1574 }
1575 
1576 static const struct blk_mq_ops nvme_mq_admin_ops = {
1577 	.queue_rq	= nvme_queue_rq,
1578 	.complete	= nvme_pci_complete_rq,
1579 	.init_hctx	= nvme_admin_init_hctx,
1580 	.exit_hctx      = nvme_admin_exit_hctx,
1581 	.init_request	= nvme_init_request,
1582 	.timeout	= nvme_timeout,
1583 };
1584 
1585 static const struct blk_mq_ops nvme_mq_ops = {
1586 	.queue_rq	= nvme_queue_rq,
1587 	.complete	= nvme_pci_complete_rq,
1588 	.commit_rqs	= nvme_commit_rqs,
1589 	.init_hctx	= nvme_init_hctx,
1590 	.init_request	= nvme_init_request,
1591 	.map_queues	= nvme_pci_map_queues,
1592 	.timeout	= nvme_timeout,
1593 	.poll		= nvme_poll,
1594 };
1595 
1596 static void nvme_dev_remove_admin(struct nvme_dev *dev)
1597 {
1598 	if (dev->ctrl.admin_q && !blk_queue_dying(dev->ctrl.admin_q)) {
1599 		/*
1600 		 * If the controller was reset during removal, it's possible
1601 		 * user requests may be waiting on a stopped queue. Start the
1602 		 * queue to flush these to completion.
1603 		 */
1604 		blk_mq_unquiesce_queue(dev->ctrl.admin_q);
1605 		blk_cleanup_queue(dev->ctrl.admin_q);
1606 		blk_mq_free_tag_set(&dev->admin_tagset);
1607 	}
1608 }
1609 
1610 static int nvme_alloc_admin_tags(struct nvme_dev *dev)
1611 {
1612 	if (!dev->ctrl.admin_q) {
1613 		dev->admin_tagset.ops = &nvme_mq_admin_ops;
1614 		dev->admin_tagset.nr_hw_queues = 1;
1615 
1616 		dev->admin_tagset.queue_depth = NVME_AQ_MQ_TAG_DEPTH;
1617 		dev->admin_tagset.timeout = ADMIN_TIMEOUT;
1618 		dev->admin_tagset.numa_node = dev_to_node(dev->dev);
1619 		dev->admin_tagset.cmd_size = sizeof(struct nvme_iod);
1620 		dev->admin_tagset.flags = BLK_MQ_F_NO_SCHED;
1621 		dev->admin_tagset.driver_data = dev;
1622 
1623 		if (blk_mq_alloc_tag_set(&dev->admin_tagset))
1624 			return -ENOMEM;
1625 		dev->ctrl.admin_tagset = &dev->admin_tagset;
1626 
1627 		dev->ctrl.admin_q = blk_mq_init_queue(&dev->admin_tagset);
1628 		if (IS_ERR(dev->ctrl.admin_q)) {
1629 			blk_mq_free_tag_set(&dev->admin_tagset);
1630 			return -ENOMEM;
1631 		}
1632 		if (!blk_get_queue(dev->ctrl.admin_q)) {
1633 			nvme_dev_remove_admin(dev);
1634 			dev->ctrl.admin_q = NULL;
1635 			return -ENODEV;
1636 		}
1637 	} else
1638 		blk_mq_unquiesce_queue(dev->ctrl.admin_q);
1639 
1640 	return 0;
1641 }
1642 
1643 static unsigned long db_bar_size(struct nvme_dev *dev, unsigned nr_io_queues)
1644 {
1645 	return NVME_REG_DBS + ((nr_io_queues + 1) * 8 * dev->db_stride);
1646 }
1647 
1648 static int nvme_remap_bar(struct nvme_dev *dev, unsigned long size)
1649 {
1650 	struct pci_dev *pdev = to_pci_dev(dev->dev);
1651 
1652 	if (size <= dev->bar_mapped_size)
1653 		return 0;
1654 	if (size > pci_resource_len(pdev, 0))
1655 		return -ENOMEM;
1656 	if (dev->bar)
1657 		iounmap(dev->bar);
1658 	dev->bar = ioremap(pci_resource_start(pdev, 0), size);
1659 	if (!dev->bar) {
1660 		dev->bar_mapped_size = 0;
1661 		return -ENOMEM;
1662 	}
1663 	dev->bar_mapped_size = size;
1664 	dev->dbs = dev->bar + NVME_REG_DBS;
1665 
1666 	return 0;
1667 }
1668 
1669 static int nvme_pci_configure_admin_queue(struct nvme_dev *dev)
1670 {
1671 	int result;
1672 	u32 aqa;
1673 	struct nvme_queue *nvmeq;
1674 
1675 	result = nvme_remap_bar(dev, db_bar_size(dev, 0));
1676 	if (result < 0)
1677 		return result;
1678 
1679 	dev->subsystem = readl(dev->bar + NVME_REG_VS) >= NVME_VS(1, 1, 0) ?
1680 				NVME_CAP_NSSRC(dev->ctrl.cap) : 0;
1681 
1682 	if (dev->subsystem &&
1683 	    (readl(dev->bar + NVME_REG_CSTS) & NVME_CSTS_NSSRO))
1684 		writel(NVME_CSTS_NSSRO, dev->bar + NVME_REG_CSTS);
1685 
1686 	result = nvme_disable_ctrl(&dev->ctrl);
1687 	if (result < 0)
1688 		return result;
1689 
1690 	result = nvme_alloc_queue(dev, 0, NVME_AQ_DEPTH);
1691 	if (result)
1692 		return result;
1693 
1694 	nvmeq = &dev->queues[0];
1695 	aqa = nvmeq->q_depth - 1;
1696 	aqa |= aqa << 16;
1697 
1698 	writel(aqa, dev->bar + NVME_REG_AQA);
1699 	lo_hi_writeq(nvmeq->sq_dma_addr, dev->bar + NVME_REG_ASQ);
1700 	lo_hi_writeq(nvmeq->cq_dma_addr, dev->bar + NVME_REG_ACQ);
1701 
1702 	result = nvme_enable_ctrl(&dev->ctrl);
1703 	if (result)
1704 		return result;
1705 
1706 	nvmeq->cq_vector = 0;
1707 	nvme_init_queue(nvmeq, 0);
1708 	result = queue_request_irq(nvmeq);
1709 	if (result) {
1710 		dev->online_queues--;
1711 		return result;
1712 	}
1713 
1714 	set_bit(NVMEQ_ENABLED, &nvmeq->flags);
1715 	return result;
1716 }
1717 
1718 static int nvme_create_io_queues(struct nvme_dev *dev)
1719 {
1720 	unsigned i, max, rw_queues;
1721 	int ret = 0;
1722 
1723 	for (i = dev->ctrl.queue_count; i <= dev->max_qid; i++) {
1724 		if (nvme_alloc_queue(dev, i, dev->q_depth)) {
1725 			ret = -ENOMEM;
1726 			break;
1727 		}
1728 	}
1729 
1730 	max = min(dev->max_qid, dev->ctrl.queue_count - 1);
1731 	if (max != 1 && dev->io_queues[HCTX_TYPE_POLL]) {
1732 		rw_queues = dev->io_queues[HCTX_TYPE_DEFAULT] +
1733 				dev->io_queues[HCTX_TYPE_READ];
1734 	} else {
1735 		rw_queues = max;
1736 	}
1737 
1738 	for (i = dev->online_queues; i <= max; i++) {
1739 		bool polled = i > rw_queues;
1740 
1741 		ret = nvme_create_queue(&dev->queues[i], i, polled);
1742 		if (ret)
1743 			break;
1744 	}
1745 
1746 	/*
1747 	 * Ignore failing Create SQ/CQ commands, we can continue with less
1748 	 * than the desired amount of queues, and even a controller without
1749 	 * I/O queues can still be used to issue admin commands.  This might
1750 	 * be useful to upgrade a buggy firmware for example.
1751 	 */
1752 	return ret >= 0 ? 0 : ret;
1753 }
1754 
1755 static ssize_t nvme_cmb_show(struct device *dev,
1756 			     struct device_attribute *attr,
1757 			     char *buf)
1758 {
1759 	struct nvme_dev *ndev = to_nvme_dev(dev_get_drvdata(dev));
1760 
1761 	return scnprintf(buf, PAGE_SIZE, "cmbloc : x%08x\ncmbsz  : x%08x\n",
1762 		       ndev->cmbloc, ndev->cmbsz);
1763 }
1764 static DEVICE_ATTR(cmb, S_IRUGO, nvme_cmb_show, NULL);
1765 
1766 static u64 nvme_cmb_size_unit(struct nvme_dev *dev)
1767 {
1768 	u8 szu = (dev->cmbsz >> NVME_CMBSZ_SZU_SHIFT) & NVME_CMBSZ_SZU_MASK;
1769 
1770 	return 1ULL << (12 + 4 * szu);
1771 }
1772 
1773 static u32 nvme_cmb_size(struct nvme_dev *dev)
1774 {
1775 	return (dev->cmbsz >> NVME_CMBSZ_SZ_SHIFT) & NVME_CMBSZ_SZ_MASK;
1776 }
1777 
1778 static void nvme_map_cmb(struct nvme_dev *dev)
1779 {
1780 	u64 size, offset;
1781 	resource_size_t bar_size;
1782 	struct pci_dev *pdev = to_pci_dev(dev->dev);
1783 	int bar;
1784 
1785 	if (dev->cmb_size)
1786 		return;
1787 
1788 	dev->cmbsz = readl(dev->bar + NVME_REG_CMBSZ);
1789 	if (!dev->cmbsz)
1790 		return;
1791 	dev->cmbloc = readl(dev->bar + NVME_REG_CMBLOC);
1792 
1793 	size = nvme_cmb_size_unit(dev) * nvme_cmb_size(dev);
1794 	offset = nvme_cmb_size_unit(dev) * NVME_CMB_OFST(dev->cmbloc);
1795 	bar = NVME_CMB_BIR(dev->cmbloc);
1796 	bar_size = pci_resource_len(pdev, bar);
1797 
1798 	if (offset > bar_size)
1799 		return;
1800 
1801 	/*
1802 	 * Controllers may support a CMB size larger than their BAR,
1803 	 * for example, due to being behind a bridge. Reduce the CMB to
1804 	 * the reported size of the BAR
1805 	 */
1806 	if (size > bar_size - offset)
1807 		size = bar_size - offset;
1808 
1809 	if (pci_p2pdma_add_resource(pdev, bar, size, offset)) {
1810 		dev_warn(dev->ctrl.device,
1811 			 "failed to register the CMB\n");
1812 		return;
1813 	}
1814 
1815 	dev->cmb_size = size;
1816 	dev->cmb_use_sqes = use_cmb_sqes && (dev->cmbsz & NVME_CMBSZ_SQS);
1817 
1818 	if ((dev->cmbsz & (NVME_CMBSZ_WDS | NVME_CMBSZ_RDS)) ==
1819 			(NVME_CMBSZ_WDS | NVME_CMBSZ_RDS))
1820 		pci_p2pmem_publish(pdev, true);
1821 
1822 	if (sysfs_add_file_to_group(&dev->ctrl.device->kobj,
1823 				    &dev_attr_cmb.attr, NULL))
1824 		dev_warn(dev->ctrl.device,
1825 			 "failed to add sysfs attribute for CMB\n");
1826 }
1827 
1828 static inline void nvme_release_cmb(struct nvme_dev *dev)
1829 {
1830 	if (dev->cmb_size) {
1831 		sysfs_remove_file_from_group(&dev->ctrl.device->kobj,
1832 					     &dev_attr_cmb.attr, NULL);
1833 		dev->cmb_size = 0;
1834 	}
1835 }
1836 
1837 static int nvme_set_host_mem(struct nvme_dev *dev, u32 bits)
1838 {
1839 	u64 dma_addr = dev->host_mem_descs_dma;
1840 	struct nvme_command c;
1841 	int ret;
1842 
1843 	memset(&c, 0, sizeof(c));
1844 	c.features.opcode	= nvme_admin_set_features;
1845 	c.features.fid		= cpu_to_le32(NVME_FEAT_HOST_MEM_BUF);
1846 	c.features.dword11	= cpu_to_le32(bits);
1847 	c.features.dword12	= cpu_to_le32(dev->host_mem_size >>
1848 					      ilog2(dev->ctrl.page_size));
1849 	c.features.dword13	= cpu_to_le32(lower_32_bits(dma_addr));
1850 	c.features.dword14	= cpu_to_le32(upper_32_bits(dma_addr));
1851 	c.features.dword15	= cpu_to_le32(dev->nr_host_mem_descs);
1852 
1853 	ret = nvme_submit_sync_cmd(dev->ctrl.admin_q, &c, NULL, 0);
1854 	if (ret) {
1855 		dev_warn(dev->ctrl.device,
1856 			 "failed to set host mem (err %d, flags %#x).\n",
1857 			 ret, bits);
1858 	}
1859 	return ret;
1860 }
1861 
1862 static void nvme_free_host_mem(struct nvme_dev *dev)
1863 {
1864 	int i;
1865 
1866 	for (i = 0; i < dev->nr_host_mem_descs; i++) {
1867 		struct nvme_host_mem_buf_desc *desc = &dev->host_mem_descs[i];
1868 		size_t size = le32_to_cpu(desc->size) * dev->ctrl.page_size;
1869 
1870 		dma_free_attrs(dev->dev, size, dev->host_mem_desc_bufs[i],
1871 			       le64_to_cpu(desc->addr),
1872 			       DMA_ATTR_NO_KERNEL_MAPPING | DMA_ATTR_NO_WARN);
1873 	}
1874 
1875 	kfree(dev->host_mem_desc_bufs);
1876 	dev->host_mem_desc_bufs = NULL;
1877 	dma_free_coherent(dev->dev,
1878 			dev->nr_host_mem_descs * sizeof(*dev->host_mem_descs),
1879 			dev->host_mem_descs, dev->host_mem_descs_dma);
1880 	dev->host_mem_descs = NULL;
1881 	dev->nr_host_mem_descs = 0;
1882 }
1883 
1884 static int __nvme_alloc_host_mem(struct nvme_dev *dev, u64 preferred,
1885 		u32 chunk_size)
1886 {
1887 	struct nvme_host_mem_buf_desc *descs;
1888 	u32 max_entries, len;
1889 	dma_addr_t descs_dma;
1890 	int i = 0;
1891 	void **bufs;
1892 	u64 size, tmp;
1893 
1894 	tmp = (preferred + chunk_size - 1);
1895 	do_div(tmp, chunk_size);
1896 	max_entries = tmp;
1897 
1898 	if (dev->ctrl.hmmaxd && dev->ctrl.hmmaxd < max_entries)
1899 		max_entries = dev->ctrl.hmmaxd;
1900 
1901 	descs = dma_alloc_coherent(dev->dev, max_entries * sizeof(*descs),
1902 				   &descs_dma, GFP_KERNEL);
1903 	if (!descs)
1904 		goto out;
1905 
1906 	bufs = kcalloc(max_entries, sizeof(*bufs), GFP_KERNEL);
1907 	if (!bufs)
1908 		goto out_free_descs;
1909 
1910 	for (size = 0; size < preferred && i < max_entries; size += len) {
1911 		dma_addr_t dma_addr;
1912 
1913 		len = min_t(u64, chunk_size, preferred - size);
1914 		bufs[i] = dma_alloc_attrs(dev->dev, len, &dma_addr, GFP_KERNEL,
1915 				DMA_ATTR_NO_KERNEL_MAPPING | DMA_ATTR_NO_WARN);
1916 		if (!bufs[i])
1917 			break;
1918 
1919 		descs[i].addr = cpu_to_le64(dma_addr);
1920 		descs[i].size = cpu_to_le32(len / dev->ctrl.page_size);
1921 		i++;
1922 	}
1923 
1924 	if (!size)
1925 		goto out_free_bufs;
1926 
1927 	dev->nr_host_mem_descs = i;
1928 	dev->host_mem_size = size;
1929 	dev->host_mem_descs = descs;
1930 	dev->host_mem_descs_dma = descs_dma;
1931 	dev->host_mem_desc_bufs = bufs;
1932 	return 0;
1933 
1934 out_free_bufs:
1935 	while (--i >= 0) {
1936 		size_t size = le32_to_cpu(descs[i].size) * dev->ctrl.page_size;
1937 
1938 		dma_free_attrs(dev->dev, size, bufs[i],
1939 			       le64_to_cpu(descs[i].addr),
1940 			       DMA_ATTR_NO_KERNEL_MAPPING | DMA_ATTR_NO_WARN);
1941 	}
1942 
1943 	kfree(bufs);
1944 out_free_descs:
1945 	dma_free_coherent(dev->dev, max_entries * sizeof(*descs), descs,
1946 			descs_dma);
1947 out:
1948 	dev->host_mem_descs = NULL;
1949 	return -ENOMEM;
1950 }
1951 
1952 static int nvme_alloc_host_mem(struct nvme_dev *dev, u64 min, u64 preferred)
1953 {
1954 	u32 chunk_size;
1955 
1956 	/* start big and work our way down */
1957 	for (chunk_size = min_t(u64, preferred, PAGE_SIZE * MAX_ORDER_NR_PAGES);
1958 	     chunk_size >= max_t(u32, dev->ctrl.hmminds * 4096, PAGE_SIZE * 2);
1959 	     chunk_size /= 2) {
1960 		if (!__nvme_alloc_host_mem(dev, preferred, chunk_size)) {
1961 			if (!min || dev->host_mem_size >= min)
1962 				return 0;
1963 			nvme_free_host_mem(dev);
1964 		}
1965 	}
1966 
1967 	return -ENOMEM;
1968 }
1969 
1970 static int nvme_setup_host_mem(struct nvme_dev *dev)
1971 {
1972 	u64 max = (u64)max_host_mem_size_mb * SZ_1M;
1973 	u64 preferred = (u64)dev->ctrl.hmpre * 4096;
1974 	u64 min = (u64)dev->ctrl.hmmin * 4096;
1975 	u32 enable_bits = NVME_HOST_MEM_ENABLE;
1976 	int ret;
1977 
1978 	preferred = min(preferred, max);
1979 	if (min > max) {
1980 		dev_warn(dev->ctrl.device,
1981 			"min host memory (%lld MiB) above limit (%d MiB).\n",
1982 			min >> ilog2(SZ_1M), max_host_mem_size_mb);
1983 		nvme_free_host_mem(dev);
1984 		return 0;
1985 	}
1986 
1987 	/*
1988 	 * If we already have a buffer allocated check if we can reuse it.
1989 	 */
1990 	if (dev->host_mem_descs) {
1991 		if (dev->host_mem_size >= min)
1992 			enable_bits |= NVME_HOST_MEM_RETURN;
1993 		else
1994 			nvme_free_host_mem(dev);
1995 	}
1996 
1997 	if (!dev->host_mem_descs) {
1998 		if (nvme_alloc_host_mem(dev, min, preferred)) {
1999 			dev_warn(dev->ctrl.device,
2000 				"failed to allocate host memory buffer.\n");
2001 			return 0; /* controller must work without HMB */
2002 		}
2003 
2004 		dev_info(dev->ctrl.device,
2005 			"allocated %lld MiB host memory buffer.\n",
2006 			dev->host_mem_size >> ilog2(SZ_1M));
2007 	}
2008 
2009 	ret = nvme_set_host_mem(dev, enable_bits);
2010 	if (ret)
2011 		nvme_free_host_mem(dev);
2012 	return ret;
2013 }
2014 
2015 /*
2016  * nirqs is the number of interrupts available for write and read
2017  * queues. The core already reserved an interrupt for the admin queue.
2018  */
2019 static void nvme_calc_irq_sets(struct irq_affinity *affd, unsigned int nrirqs)
2020 {
2021 	struct nvme_dev *dev = affd->priv;
2022 	unsigned int nr_read_queues;
2023 
2024 	/*
2025 	 * If there is no interupt available for queues, ensure that
2026 	 * the default queue is set to 1. The affinity set size is
2027 	 * also set to one, but the irq core ignores it for this case.
2028 	 *
2029 	 * If only one interrupt is available or 'write_queue' == 0, combine
2030 	 * write and read queues.
2031 	 *
2032 	 * If 'write_queues' > 0, ensure it leaves room for at least one read
2033 	 * queue.
2034 	 */
2035 	if (!nrirqs) {
2036 		nrirqs = 1;
2037 		nr_read_queues = 0;
2038 	} else if (nrirqs == 1 || !write_queues) {
2039 		nr_read_queues = 0;
2040 	} else if (write_queues >= nrirqs) {
2041 		nr_read_queues = 1;
2042 	} else {
2043 		nr_read_queues = nrirqs - write_queues;
2044 	}
2045 
2046 	dev->io_queues[HCTX_TYPE_DEFAULT] = nrirqs - nr_read_queues;
2047 	affd->set_size[HCTX_TYPE_DEFAULT] = nrirqs - nr_read_queues;
2048 	dev->io_queues[HCTX_TYPE_READ] = nr_read_queues;
2049 	affd->set_size[HCTX_TYPE_READ] = nr_read_queues;
2050 	affd->nr_sets = nr_read_queues ? 2 : 1;
2051 }
2052 
2053 static int nvme_setup_irqs(struct nvme_dev *dev, unsigned int nr_io_queues)
2054 {
2055 	struct pci_dev *pdev = to_pci_dev(dev->dev);
2056 	struct irq_affinity affd = {
2057 		.pre_vectors	= 1,
2058 		.calc_sets	= nvme_calc_irq_sets,
2059 		.priv		= dev,
2060 	};
2061 	unsigned int irq_queues, this_p_queues;
2062 	unsigned int nr_cpus = num_possible_cpus();
2063 
2064 	/*
2065 	 * Poll queues don't need interrupts, but we need at least one IO
2066 	 * queue left over for non-polled IO.
2067 	 */
2068 	this_p_queues = poll_queues;
2069 	if (this_p_queues >= nr_io_queues) {
2070 		this_p_queues = nr_io_queues - 1;
2071 		irq_queues = 1;
2072 	} else {
2073 		if (nr_cpus < nr_io_queues - this_p_queues)
2074 			irq_queues = nr_cpus + 1;
2075 		else
2076 			irq_queues = nr_io_queues - this_p_queues + 1;
2077 	}
2078 	dev->io_queues[HCTX_TYPE_POLL] = this_p_queues;
2079 
2080 	/* Initialize for the single interrupt case */
2081 	dev->io_queues[HCTX_TYPE_DEFAULT] = 1;
2082 	dev->io_queues[HCTX_TYPE_READ] = 0;
2083 
2084 	/*
2085 	 * Some Apple controllers require all queues to use the
2086 	 * first vector.
2087 	 */
2088 	if (dev->ctrl.quirks & NVME_QUIRK_SINGLE_VECTOR)
2089 		irq_queues = 1;
2090 
2091 	return pci_alloc_irq_vectors_affinity(pdev, 1, irq_queues,
2092 			      PCI_IRQ_ALL_TYPES | PCI_IRQ_AFFINITY, &affd);
2093 }
2094 
2095 static void nvme_disable_io_queues(struct nvme_dev *dev)
2096 {
2097 	if (__nvme_disable_io_queues(dev, nvme_admin_delete_sq))
2098 		__nvme_disable_io_queues(dev, nvme_admin_delete_cq);
2099 }
2100 
2101 static int nvme_setup_io_queues(struct nvme_dev *dev)
2102 {
2103 	struct nvme_queue *adminq = &dev->queues[0];
2104 	struct pci_dev *pdev = to_pci_dev(dev->dev);
2105 	int result, nr_io_queues;
2106 	unsigned long size;
2107 
2108 	nr_io_queues = max_io_queues();
2109 
2110 	/*
2111 	 * If tags are shared with admin queue (Apple bug), then
2112 	 * make sure we only use one IO queue.
2113 	 */
2114 	if (dev->ctrl.quirks & NVME_QUIRK_SHARED_TAGS)
2115 		nr_io_queues = 1;
2116 
2117 	result = nvme_set_queue_count(&dev->ctrl, &nr_io_queues);
2118 	if (result < 0)
2119 		return result;
2120 
2121 	if (nr_io_queues == 0)
2122 		return 0;
2123 
2124 	clear_bit(NVMEQ_ENABLED, &adminq->flags);
2125 
2126 	if (dev->cmb_use_sqes) {
2127 		result = nvme_cmb_qdepth(dev, nr_io_queues,
2128 				sizeof(struct nvme_command));
2129 		if (result > 0)
2130 			dev->q_depth = result;
2131 		else
2132 			dev->cmb_use_sqes = false;
2133 	}
2134 
2135 	do {
2136 		size = db_bar_size(dev, nr_io_queues);
2137 		result = nvme_remap_bar(dev, size);
2138 		if (!result)
2139 			break;
2140 		if (!--nr_io_queues)
2141 			return -ENOMEM;
2142 	} while (1);
2143 	adminq->q_db = dev->dbs;
2144 
2145  retry:
2146 	/* Deregister the admin queue's interrupt */
2147 	pci_free_irq(pdev, 0, adminq);
2148 
2149 	/*
2150 	 * If we enable msix early due to not intx, disable it again before
2151 	 * setting up the full range we need.
2152 	 */
2153 	pci_free_irq_vectors(pdev);
2154 
2155 	result = nvme_setup_irqs(dev, nr_io_queues);
2156 	if (result <= 0)
2157 		return -EIO;
2158 
2159 	dev->num_vecs = result;
2160 	result = max(result - 1, 1);
2161 	dev->max_qid = result + dev->io_queues[HCTX_TYPE_POLL];
2162 
2163 	/*
2164 	 * Should investigate if there's a performance win from allocating
2165 	 * more queues than interrupt vectors; it might allow the submission
2166 	 * path to scale better, even if the receive path is limited by the
2167 	 * number of interrupts.
2168 	 */
2169 	result = queue_request_irq(adminq);
2170 	if (result)
2171 		return result;
2172 	set_bit(NVMEQ_ENABLED, &adminq->flags);
2173 
2174 	result = nvme_create_io_queues(dev);
2175 	if (result || dev->online_queues < 2)
2176 		return result;
2177 
2178 	if (dev->online_queues - 1 < dev->max_qid) {
2179 		nr_io_queues = dev->online_queues - 1;
2180 		nvme_disable_io_queues(dev);
2181 		nvme_suspend_io_queues(dev);
2182 		goto retry;
2183 	}
2184 	dev_info(dev->ctrl.device, "%d/%d/%d default/read/poll queues\n",
2185 					dev->io_queues[HCTX_TYPE_DEFAULT],
2186 					dev->io_queues[HCTX_TYPE_READ],
2187 					dev->io_queues[HCTX_TYPE_POLL]);
2188 	return 0;
2189 }
2190 
2191 static void nvme_del_queue_end(struct request *req, blk_status_t error)
2192 {
2193 	struct nvme_queue *nvmeq = req->end_io_data;
2194 
2195 	blk_mq_free_request(req);
2196 	complete(&nvmeq->delete_done);
2197 }
2198 
2199 static void nvme_del_cq_end(struct request *req, blk_status_t error)
2200 {
2201 	struct nvme_queue *nvmeq = req->end_io_data;
2202 
2203 	if (error)
2204 		set_bit(NVMEQ_DELETE_ERROR, &nvmeq->flags);
2205 
2206 	nvme_del_queue_end(req, error);
2207 }
2208 
2209 static int nvme_delete_queue(struct nvme_queue *nvmeq, u8 opcode)
2210 {
2211 	struct request_queue *q = nvmeq->dev->ctrl.admin_q;
2212 	struct request *req;
2213 	struct nvme_command cmd;
2214 
2215 	memset(&cmd, 0, sizeof(cmd));
2216 	cmd.delete_queue.opcode = opcode;
2217 	cmd.delete_queue.qid = cpu_to_le16(nvmeq->qid);
2218 
2219 	req = nvme_alloc_request(q, &cmd, BLK_MQ_REQ_NOWAIT, NVME_QID_ANY);
2220 	if (IS_ERR(req))
2221 		return PTR_ERR(req);
2222 
2223 	req->timeout = ADMIN_TIMEOUT;
2224 	req->end_io_data = nvmeq;
2225 
2226 	init_completion(&nvmeq->delete_done);
2227 	blk_execute_rq_nowait(q, NULL, req, false,
2228 			opcode == nvme_admin_delete_cq ?
2229 				nvme_del_cq_end : nvme_del_queue_end);
2230 	return 0;
2231 }
2232 
2233 static bool __nvme_disable_io_queues(struct nvme_dev *dev, u8 opcode)
2234 {
2235 	int nr_queues = dev->online_queues - 1, sent = 0;
2236 	unsigned long timeout;
2237 
2238  retry:
2239 	timeout = ADMIN_TIMEOUT;
2240 	while (nr_queues > 0) {
2241 		if (nvme_delete_queue(&dev->queues[nr_queues], opcode))
2242 			break;
2243 		nr_queues--;
2244 		sent++;
2245 	}
2246 	while (sent) {
2247 		struct nvme_queue *nvmeq = &dev->queues[nr_queues + sent];
2248 
2249 		timeout = wait_for_completion_io_timeout(&nvmeq->delete_done,
2250 				timeout);
2251 		if (timeout == 0)
2252 			return false;
2253 
2254 		/* handle any remaining CQEs */
2255 		if (opcode == nvme_admin_delete_cq &&
2256 		    !test_bit(NVMEQ_DELETE_ERROR, &nvmeq->flags))
2257 			nvme_poll_irqdisable(nvmeq, -1);
2258 
2259 		sent--;
2260 		if (nr_queues)
2261 			goto retry;
2262 	}
2263 	return true;
2264 }
2265 
2266 /*
2267  * return error value only when tagset allocation failed
2268  */
2269 static int nvme_dev_add(struct nvme_dev *dev)
2270 {
2271 	int ret;
2272 
2273 	if (!dev->ctrl.tagset) {
2274 		dev->tagset.ops = &nvme_mq_ops;
2275 		dev->tagset.nr_hw_queues = dev->online_queues - 1;
2276 		dev->tagset.nr_maps = 2; /* default + read */
2277 		if (dev->io_queues[HCTX_TYPE_POLL])
2278 			dev->tagset.nr_maps++;
2279 		dev->tagset.timeout = NVME_IO_TIMEOUT;
2280 		dev->tagset.numa_node = dev_to_node(dev->dev);
2281 		dev->tagset.queue_depth =
2282 				min_t(int, dev->q_depth, BLK_MQ_MAX_DEPTH) - 1;
2283 		dev->tagset.cmd_size = sizeof(struct nvme_iod);
2284 		dev->tagset.flags = BLK_MQ_F_SHOULD_MERGE;
2285 		dev->tagset.driver_data = dev;
2286 
2287 		/*
2288 		 * Some Apple controllers requires tags to be unique
2289 		 * across admin and IO queue, so reserve the first 32
2290 		 * tags of the IO queue.
2291 		 */
2292 		if (dev->ctrl.quirks & NVME_QUIRK_SHARED_TAGS)
2293 			dev->tagset.reserved_tags = NVME_AQ_DEPTH;
2294 
2295 		ret = blk_mq_alloc_tag_set(&dev->tagset);
2296 		if (ret) {
2297 			dev_warn(dev->ctrl.device,
2298 				"IO queues tagset allocation failed %d\n", ret);
2299 			return ret;
2300 		}
2301 		dev->ctrl.tagset = &dev->tagset;
2302 	} else {
2303 		blk_mq_update_nr_hw_queues(&dev->tagset, dev->online_queues - 1);
2304 
2305 		/* Free previously allocated queues that are no longer usable */
2306 		nvme_free_queues(dev, dev->online_queues);
2307 	}
2308 
2309 	nvme_dbbuf_set(dev);
2310 	return 0;
2311 }
2312 
2313 static int nvme_pci_enable(struct nvme_dev *dev)
2314 {
2315 	int result = -ENOMEM;
2316 	struct pci_dev *pdev = to_pci_dev(dev->dev);
2317 
2318 	if (pci_enable_device_mem(pdev))
2319 		return result;
2320 
2321 	pci_set_master(pdev);
2322 
2323 	if (dma_set_mask_and_coherent(dev->dev, DMA_BIT_MASK(64)))
2324 		goto disable;
2325 
2326 	if (readl(dev->bar + NVME_REG_CSTS) == -1) {
2327 		result = -ENODEV;
2328 		goto disable;
2329 	}
2330 
2331 	/*
2332 	 * Some devices and/or platforms don't advertise or work with INTx
2333 	 * interrupts. Pre-enable a single MSIX or MSI vec for setup. We'll
2334 	 * adjust this later.
2335 	 */
2336 	result = pci_alloc_irq_vectors(pdev, 1, 1, PCI_IRQ_ALL_TYPES);
2337 	if (result < 0)
2338 		return result;
2339 
2340 	dev->ctrl.cap = lo_hi_readq(dev->bar + NVME_REG_CAP);
2341 
2342 	dev->q_depth = min_t(int, NVME_CAP_MQES(dev->ctrl.cap) + 1,
2343 				io_queue_depth);
2344 	dev->ctrl.sqsize = dev->q_depth - 1; /* 0's based queue depth */
2345 	dev->db_stride = 1 << NVME_CAP_STRIDE(dev->ctrl.cap);
2346 	dev->dbs = dev->bar + 4096;
2347 
2348 	/*
2349 	 * Some Apple controllers require a non-standard SQE size.
2350 	 * Interestingly they also seem to ignore the CC:IOSQES register
2351 	 * so we don't bother updating it here.
2352 	 */
2353 	if (dev->ctrl.quirks & NVME_QUIRK_128_BYTES_SQES)
2354 		dev->io_sqes = 7;
2355 	else
2356 		dev->io_sqes = NVME_NVM_IOSQES;
2357 
2358 	/*
2359 	 * Temporary fix for the Apple controller found in the MacBook8,1 and
2360 	 * some MacBook7,1 to avoid controller resets and data loss.
2361 	 */
2362 	if (pdev->vendor == PCI_VENDOR_ID_APPLE && pdev->device == 0x2001) {
2363 		dev->q_depth = 2;
2364 		dev_warn(dev->ctrl.device, "detected Apple NVMe controller, "
2365 			"set queue depth=%u to work around controller resets\n",
2366 			dev->q_depth);
2367 	} else if (pdev->vendor == PCI_VENDOR_ID_SAMSUNG &&
2368 		   (pdev->device == 0xa821 || pdev->device == 0xa822) &&
2369 		   NVME_CAP_MQES(dev->ctrl.cap) == 0) {
2370 		dev->q_depth = 64;
2371 		dev_err(dev->ctrl.device, "detected PM1725 NVMe controller, "
2372                         "set queue depth=%u\n", dev->q_depth);
2373 	}
2374 
2375 	/*
2376 	 * Controllers with the shared tags quirk need the IO queue to be
2377 	 * big enough so that we get 32 tags for the admin queue
2378 	 */
2379 	if ((dev->ctrl.quirks & NVME_QUIRK_SHARED_TAGS) &&
2380 	    (dev->q_depth < (NVME_AQ_DEPTH + 2))) {
2381 		dev->q_depth = NVME_AQ_DEPTH + 2;
2382 		dev_warn(dev->ctrl.device, "IO queue depth clamped to %d\n",
2383 			 dev->q_depth);
2384 	}
2385 
2386 
2387 	nvme_map_cmb(dev);
2388 
2389 	pci_enable_pcie_error_reporting(pdev);
2390 	pci_save_state(pdev);
2391 	return 0;
2392 
2393  disable:
2394 	pci_disable_device(pdev);
2395 	return result;
2396 }
2397 
2398 static void nvme_dev_unmap(struct nvme_dev *dev)
2399 {
2400 	if (dev->bar)
2401 		iounmap(dev->bar);
2402 	pci_release_mem_regions(to_pci_dev(dev->dev));
2403 }
2404 
2405 static void nvme_pci_disable(struct nvme_dev *dev)
2406 {
2407 	struct pci_dev *pdev = to_pci_dev(dev->dev);
2408 
2409 	pci_free_irq_vectors(pdev);
2410 
2411 	if (pci_is_enabled(pdev)) {
2412 		pci_disable_pcie_error_reporting(pdev);
2413 		pci_disable_device(pdev);
2414 	}
2415 }
2416 
2417 static void nvme_dev_disable(struct nvme_dev *dev, bool shutdown)
2418 {
2419 	bool dead = true, freeze = false;
2420 	struct pci_dev *pdev = to_pci_dev(dev->dev);
2421 
2422 	mutex_lock(&dev->shutdown_lock);
2423 	if (pci_is_enabled(pdev)) {
2424 		u32 csts = readl(dev->bar + NVME_REG_CSTS);
2425 
2426 		if (dev->ctrl.state == NVME_CTRL_LIVE ||
2427 		    dev->ctrl.state == NVME_CTRL_RESETTING) {
2428 			freeze = true;
2429 			nvme_start_freeze(&dev->ctrl);
2430 		}
2431 		dead = !!((csts & NVME_CSTS_CFS) || !(csts & NVME_CSTS_RDY) ||
2432 			pdev->error_state  != pci_channel_io_normal);
2433 	}
2434 
2435 	/*
2436 	 * Give the controller a chance to complete all entered requests if
2437 	 * doing a safe shutdown.
2438 	 */
2439 	if (!dead && shutdown && freeze)
2440 		nvme_wait_freeze_timeout(&dev->ctrl, NVME_IO_TIMEOUT);
2441 
2442 	nvme_stop_queues(&dev->ctrl);
2443 
2444 	if (!dead && dev->ctrl.queue_count > 0) {
2445 		nvme_disable_io_queues(dev);
2446 		nvme_disable_admin_queue(dev, shutdown);
2447 	}
2448 	nvme_suspend_io_queues(dev);
2449 	nvme_suspend_queue(&dev->queues[0]);
2450 	nvme_pci_disable(dev);
2451 
2452 	blk_mq_tagset_busy_iter(&dev->tagset, nvme_cancel_request, &dev->ctrl);
2453 	blk_mq_tagset_busy_iter(&dev->admin_tagset, nvme_cancel_request, &dev->ctrl);
2454 	blk_mq_tagset_wait_completed_request(&dev->tagset);
2455 	blk_mq_tagset_wait_completed_request(&dev->admin_tagset);
2456 
2457 	/*
2458 	 * The driver will not be starting up queues again if shutting down so
2459 	 * must flush all entered requests to their failed completion to avoid
2460 	 * deadlocking blk-mq hot-cpu notifier.
2461 	 */
2462 	if (shutdown) {
2463 		nvme_start_queues(&dev->ctrl);
2464 		if (dev->ctrl.admin_q && !blk_queue_dying(dev->ctrl.admin_q))
2465 			blk_mq_unquiesce_queue(dev->ctrl.admin_q);
2466 	}
2467 	mutex_unlock(&dev->shutdown_lock);
2468 }
2469 
2470 static int nvme_setup_prp_pools(struct nvme_dev *dev)
2471 {
2472 	dev->prp_page_pool = dma_pool_create("prp list page", dev->dev,
2473 						PAGE_SIZE, PAGE_SIZE, 0);
2474 	if (!dev->prp_page_pool)
2475 		return -ENOMEM;
2476 
2477 	/* Optimisation for I/Os between 4k and 128k */
2478 	dev->prp_small_pool = dma_pool_create("prp list 256", dev->dev,
2479 						256, 256, 0);
2480 	if (!dev->prp_small_pool) {
2481 		dma_pool_destroy(dev->prp_page_pool);
2482 		return -ENOMEM;
2483 	}
2484 	return 0;
2485 }
2486 
2487 static void nvme_release_prp_pools(struct nvme_dev *dev)
2488 {
2489 	dma_pool_destroy(dev->prp_page_pool);
2490 	dma_pool_destroy(dev->prp_small_pool);
2491 }
2492 
2493 static void nvme_pci_free_ctrl(struct nvme_ctrl *ctrl)
2494 {
2495 	struct nvme_dev *dev = to_nvme_dev(ctrl);
2496 
2497 	nvme_dbbuf_dma_free(dev);
2498 	put_device(dev->dev);
2499 	if (dev->tagset.tags)
2500 		blk_mq_free_tag_set(&dev->tagset);
2501 	if (dev->ctrl.admin_q)
2502 		blk_put_queue(dev->ctrl.admin_q);
2503 	kfree(dev->queues);
2504 	free_opal_dev(dev->ctrl.opal_dev);
2505 	mempool_destroy(dev->iod_mempool);
2506 	kfree(dev);
2507 }
2508 
2509 static void nvme_remove_dead_ctrl(struct nvme_dev *dev)
2510 {
2511 	nvme_get_ctrl(&dev->ctrl);
2512 	nvme_dev_disable(dev, false);
2513 	nvme_kill_queues(&dev->ctrl);
2514 	if (!queue_work(nvme_wq, &dev->remove_work))
2515 		nvme_put_ctrl(&dev->ctrl);
2516 }
2517 
2518 static void nvme_reset_work(struct work_struct *work)
2519 {
2520 	struct nvme_dev *dev =
2521 		container_of(work, struct nvme_dev, ctrl.reset_work);
2522 	bool was_suspend = !!(dev->ctrl.ctrl_config & NVME_CC_SHN_NORMAL);
2523 	int result;
2524 	enum nvme_ctrl_state new_state = NVME_CTRL_LIVE;
2525 
2526 	if (WARN_ON(dev->ctrl.state != NVME_CTRL_RESETTING)) {
2527 		result = -ENODEV;
2528 		goto out;
2529 	}
2530 
2531 	/*
2532 	 * If we're called to reset a live controller first shut it down before
2533 	 * moving on.
2534 	 */
2535 	if (dev->ctrl.ctrl_config & NVME_CC_ENABLE)
2536 		nvme_dev_disable(dev, false);
2537 	nvme_sync_queues(&dev->ctrl);
2538 
2539 	mutex_lock(&dev->shutdown_lock);
2540 	result = nvme_pci_enable(dev);
2541 	if (result)
2542 		goto out_unlock;
2543 
2544 	result = nvme_pci_configure_admin_queue(dev);
2545 	if (result)
2546 		goto out_unlock;
2547 
2548 	result = nvme_alloc_admin_tags(dev);
2549 	if (result)
2550 		goto out_unlock;
2551 
2552 	/*
2553 	 * Limit the max command size to prevent iod->sg allocations going
2554 	 * over a single page.
2555 	 */
2556 	dev->ctrl.max_hw_sectors = min_t(u32,
2557 		NVME_MAX_KB_SZ << 1, dma_max_mapping_size(dev->dev) >> 9);
2558 	dev->ctrl.max_segments = NVME_MAX_SEGS;
2559 
2560 	/*
2561 	 * Don't limit the IOMMU merged segment size.
2562 	 */
2563 	dma_set_max_seg_size(dev->dev, 0xffffffff);
2564 
2565 	mutex_unlock(&dev->shutdown_lock);
2566 
2567 	/*
2568 	 * Introduce CONNECTING state from nvme-fc/rdma transports to mark the
2569 	 * initializing procedure here.
2570 	 */
2571 	if (!nvme_change_ctrl_state(&dev->ctrl, NVME_CTRL_CONNECTING)) {
2572 		dev_warn(dev->ctrl.device,
2573 			"failed to mark controller CONNECTING\n");
2574 		result = -EBUSY;
2575 		goto out;
2576 	}
2577 
2578 	result = nvme_init_identify(&dev->ctrl);
2579 	if (result)
2580 		goto out;
2581 
2582 	if (dev->ctrl.oacs & NVME_CTRL_OACS_SEC_SUPP) {
2583 		if (!dev->ctrl.opal_dev)
2584 			dev->ctrl.opal_dev =
2585 				init_opal_dev(&dev->ctrl, &nvme_sec_submit);
2586 		else if (was_suspend)
2587 			opal_unlock_from_suspend(dev->ctrl.opal_dev);
2588 	} else {
2589 		free_opal_dev(dev->ctrl.opal_dev);
2590 		dev->ctrl.opal_dev = NULL;
2591 	}
2592 
2593 	if (dev->ctrl.oacs & NVME_CTRL_OACS_DBBUF_SUPP) {
2594 		result = nvme_dbbuf_dma_alloc(dev);
2595 		if (result)
2596 			dev_warn(dev->dev,
2597 				 "unable to allocate dma for dbbuf\n");
2598 	}
2599 
2600 	if (dev->ctrl.hmpre) {
2601 		result = nvme_setup_host_mem(dev);
2602 		if (result < 0)
2603 			goto out;
2604 	}
2605 
2606 	result = nvme_setup_io_queues(dev);
2607 	if (result)
2608 		goto out;
2609 
2610 	/*
2611 	 * Keep the controller around but remove all namespaces if we don't have
2612 	 * any working I/O queue.
2613 	 */
2614 	if (dev->online_queues < 2) {
2615 		dev_warn(dev->ctrl.device, "IO queues not created\n");
2616 		nvme_kill_queues(&dev->ctrl);
2617 		nvme_remove_namespaces(&dev->ctrl);
2618 		new_state = NVME_CTRL_ADMIN_ONLY;
2619 	} else {
2620 		nvme_start_queues(&dev->ctrl);
2621 		nvme_wait_freeze(&dev->ctrl);
2622 		/* hit this only when allocate tagset fails */
2623 		if (nvme_dev_add(dev))
2624 			new_state = NVME_CTRL_ADMIN_ONLY;
2625 		nvme_unfreeze(&dev->ctrl);
2626 	}
2627 
2628 	/*
2629 	 * If only admin queue live, keep it to do further investigation or
2630 	 * recovery.
2631 	 */
2632 	if (!nvme_change_ctrl_state(&dev->ctrl, new_state)) {
2633 		dev_warn(dev->ctrl.device,
2634 			"failed to mark controller state %d\n", new_state);
2635 		result = -ENODEV;
2636 		goto out;
2637 	}
2638 
2639 	nvme_start_ctrl(&dev->ctrl);
2640 	return;
2641 
2642  out_unlock:
2643 	mutex_unlock(&dev->shutdown_lock);
2644  out:
2645 	if (result)
2646 		dev_warn(dev->ctrl.device,
2647 			 "Removing after probe failure status: %d\n", result);
2648 	nvme_remove_dead_ctrl(dev);
2649 }
2650 
2651 static void nvme_remove_dead_ctrl_work(struct work_struct *work)
2652 {
2653 	struct nvme_dev *dev = container_of(work, struct nvme_dev, remove_work);
2654 	struct pci_dev *pdev = to_pci_dev(dev->dev);
2655 
2656 	if (pci_get_drvdata(pdev))
2657 		device_release_driver(&pdev->dev);
2658 	nvme_put_ctrl(&dev->ctrl);
2659 }
2660 
2661 static int nvme_pci_reg_read32(struct nvme_ctrl *ctrl, u32 off, u32 *val)
2662 {
2663 	*val = readl(to_nvme_dev(ctrl)->bar + off);
2664 	return 0;
2665 }
2666 
2667 static int nvme_pci_reg_write32(struct nvme_ctrl *ctrl, u32 off, u32 val)
2668 {
2669 	writel(val, to_nvme_dev(ctrl)->bar + off);
2670 	return 0;
2671 }
2672 
2673 static int nvme_pci_reg_read64(struct nvme_ctrl *ctrl, u32 off, u64 *val)
2674 {
2675 	*val = readq(to_nvme_dev(ctrl)->bar + off);
2676 	return 0;
2677 }
2678 
2679 static int nvme_pci_get_address(struct nvme_ctrl *ctrl, char *buf, int size)
2680 {
2681 	struct pci_dev *pdev = to_pci_dev(to_nvme_dev(ctrl)->dev);
2682 
2683 	return snprintf(buf, size, "%s", dev_name(&pdev->dev));
2684 }
2685 
2686 static const struct nvme_ctrl_ops nvme_pci_ctrl_ops = {
2687 	.name			= "pcie",
2688 	.module			= THIS_MODULE,
2689 	.flags			= NVME_F_METADATA_SUPPORTED |
2690 				  NVME_F_PCI_P2PDMA,
2691 	.reg_read32		= nvme_pci_reg_read32,
2692 	.reg_write32		= nvme_pci_reg_write32,
2693 	.reg_read64		= nvme_pci_reg_read64,
2694 	.free_ctrl		= nvme_pci_free_ctrl,
2695 	.submit_async_event	= nvme_pci_submit_async_event,
2696 	.get_address		= nvme_pci_get_address,
2697 };
2698 
2699 static int nvme_dev_map(struct nvme_dev *dev)
2700 {
2701 	struct pci_dev *pdev = to_pci_dev(dev->dev);
2702 
2703 	if (pci_request_mem_regions(pdev, "nvme"))
2704 		return -ENODEV;
2705 
2706 	if (nvme_remap_bar(dev, NVME_REG_DBS + 4096))
2707 		goto release;
2708 
2709 	return 0;
2710   release:
2711 	pci_release_mem_regions(pdev);
2712 	return -ENODEV;
2713 }
2714 
2715 static unsigned long check_vendor_combination_bug(struct pci_dev *pdev)
2716 {
2717 	if (pdev->vendor == 0x144d && pdev->device == 0xa802) {
2718 		/*
2719 		 * Several Samsung devices seem to drop off the PCIe bus
2720 		 * randomly when APST is on and uses the deepest sleep state.
2721 		 * This has been observed on a Samsung "SM951 NVMe SAMSUNG
2722 		 * 256GB", a "PM951 NVMe SAMSUNG 512GB", and a "Samsung SSD
2723 		 * 950 PRO 256GB", but it seems to be restricted to two Dell
2724 		 * laptops.
2725 		 */
2726 		if (dmi_match(DMI_SYS_VENDOR, "Dell Inc.") &&
2727 		    (dmi_match(DMI_PRODUCT_NAME, "XPS 15 9550") ||
2728 		     dmi_match(DMI_PRODUCT_NAME, "Precision 5510")))
2729 			return NVME_QUIRK_NO_DEEPEST_PS;
2730 	} else if (pdev->vendor == 0x144d && pdev->device == 0xa804) {
2731 		/*
2732 		 * Samsung SSD 960 EVO drops off the PCIe bus after system
2733 		 * suspend on a Ryzen board, ASUS PRIME B350M-A, as well as
2734 		 * within few minutes after bootup on a Coffee Lake board -
2735 		 * ASUS PRIME Z370-A
2736 		 */
2737 		if (dmi_match(DMI_BOARD_VENDOR, "ASUSTeK COMPUTER INC.") &&
2738 		    (dmi_match(DMI_BOARD_NAME, "PRIME B350M-A") ||
2739 		     dmi_match(DMI_BOARD_NAME, "PRIME Z370-A")))
2740 			return NVME_QUIRK_NO_APST;
2741 	}
2742 
2743 	return 0;
2744 }
2745 
2746 static void nvme_async_probe(void *data, async_cookie_t cookie)
2747 {
2748 	struct nvme_dev *dev = data;
2749 
2750 	flush_work(&dev->ctrl.reset_work);
2751 	flush_work(&dev->ctrl.scan_work);
2752 	nvme_put_ctrl(&dev->ctrl);
2753 }
2754 
2755 static int nvme_probe(struct pci_dev *pdev, const struct pci_device_id *id)
2756 {
2757 	int node, result = -ENOMEM;
2758 	struct nvme_dev *dev;
2759 	unsigned long quirks = id->driver_data;
2760 	size_t alloc_size;
2761 
2762 	node = dev_to_node(&pdev->dev);
2763 	if (node == NUMA_NO_NODE)
2764 		set_dev_node(&pdev->dev, first_memory_node);
2765 
2766 	dev = kzalloc_node(sizeof(*dev), GFP_KERNEL, node);
2767 	if (!dev)
2768 		return -ENOMEM;
2769 
2770 	dev->queues = kcalloc_node(max_queue_count(), sizeof(struct nvme_queue),
2771 					GFP_KERNEL, node);
2772 	if (!dev->queues)
2773 		goto free;
2774 
2775 	dev->dev = get_device(&pdev->dev);
2776 	pci_set_drvdata(pdev, dev);
2777 
2778 	result = nvme_dev_map(dev);
2779 	if (result)
2780 		goto put_pci;
2781 
2782 	INIT_WORK(&dev->ctrl.reset_work, nvme_reset_work);
2783 	INIT_WORK(&dev->remove_work, nvme_remove_dead_ctrl_work);
2784 	mutex_init(&dev->shutdown_lock);
2785 
2786 	result = nvme_setup_prp_pools(dev);
2787 	if (result)
2788 		goto unmap;
2789 
2790 	quirks |= check_vendor_combination_bug(pdev);
2791 
2792 	/*
2793 	 * Double check that our mempool alloc size will cover the biggest
2794 	 * command we support.
2795 	 */
2796 	alloc_size = nvme_pci_iod_alloc_size(dev, NVME_MAX_KB_SZ,
2797 						NVME_MAX_SEGS, true);
2798 	WARN_ON_ONCE(alloc_size > PAGE_SIZE);
2799 
2800 	dev->iod_mempool = mempool_create_node(1, mempool_kmalloc,
2801 						mempool_kfree,
2802 						(void *) alloc_size,
2803 						GFP_KERNEL, node);
2804 	if (!dev->iod_mempool) {
2805 		result = -ENOMEM;
2806 		goto release_pools;
2807 	}
2808 
2809 	result = nvme_init_ctrl(&dev->ctrl, &pdev->dev, &nvme_pci_ctrl_ops,
2810 			quirks);
2811 	if (result)
2812 		goto release_mempool;
2813 
2814 	dev_info(dev->ctrl.device, "pci function %s\n", dev_name(&pdev->dev));
2815 
2816 	nvme_reset_ctrl(&dev->ctrl);
2817 	nvme_get_ctrl(&dev->ctrl);
2818 	async_schedule(nvme_async_probe, dev);
2819 
2820 	return 0;
2821 
2822  release_mempool:
2823 	mempool_destroy(dev->iod_mempool);
2824  release_pools:
2825 	nvme_release_prp_pools(dev);
2826  unmap:
2827 	nvme_dev_unmap(dev);
2828  put_pci:
2829 	put_device(dev->dev);
2830  free:
2831 	kfree(dev->queues);
2832 	kfree(dev);
2833 	return result;
2834 }
2835 
2836 static void nvme_reset_prepare(struct pci_dev *pdev)
2837 {
2838 	struct nvme_dev *dev = pci_get_drvdata(pdev);
2839 	nvme_dev_disable(dev, false);
2840 }
2841 
2842 static void nvme_reset_done(struct pci_dev *pdev)
2843 {
2844 	struct nvme_dev *dev = pci_get_drvdata(pdev);
2845 	nvme_reset_ctrl_sync(&dev->ctrl);
2846 }
2847 
2848 static void nvme_shutdown(struct pci_dev *pdev)
2849 {
2850 	struct nvme_dev *dev = pci_get_drvdata(pdev);
2851 	nvme_dev_disable(dev, true);
2852 }
2853 
2854 /*
2855  * The driver's remove may be called on a device in a partially initialized
2856  * state. This function must not have any dependencies on the device state in
2857  * order to proceed.
2858  */
2859 static void nvme_remove(struct pci_dev *pdev)
2860 {
2861 	struct nvme_dev *dev = pci_get_drvdata(pdev);
2862 
2863 	nvme_change_ctrl_state(&dev->ctrl, NVME_CTRL_DELETING);
2864 	pci_set_drvdata(pdev, NULL);
2865 
2866 	if (!pci_device_is_present(pdev)) {
2867 		nvme_change_ctrl_state(&dev->ctrl, NVME_CTRL_DEAD);
2868 		nvme_dev_disable(dev, true);
2869 		nvme_dev_remove_admin(dev);
2870 	}
2871 
2872 	flush_work(&dev->ctrl.reset_work);
2873 	nvme_stop_ctrl(&dev->ctrl);
2874 	nvme_remove_namespaces(&dev->ctrl);
2875 	nvme_dev_disable(dev, true);
2876 	nvme_release_cmb(dev);
2877 	nvme_free_host_mem(dev);
2878 	nvme_dev_remove_admin(dev);
2879 	nvme_free_queues(dev, 0);
2880 	nvme_uninit_ctrl(&dev->ctrl);
2881 	nvme_release_prp_pools(dev);
2882 	nvme_dev_unmap(dev);
2883 	nvme_put_ctrl(&dev->ctrl);
2884 }
2885 
2886 #ifdef CONFIG_PM_SLEEP
2887 static int nvme_get_power_state(struct nvme_ctrl *ctrl, u32 *ps)
2888 {
2889 	return nvme_get_features(ctrl, NVME_FEAT_POWER_MGMT, 0, NULL, 0, ps);
2890 }
2891 
2892 static int nvme_set_power_state(struct nvme_ctrl *ctrl, u32 ps)
2893 {
2894 	return nvme_set_features(ctrl, NVME_FEAT_POWER_MGMT, ps, NULL, 0, NULL);
2895 }
2896 
2897 static int nvme_resume(struct device *dev)
2898 {
2899 	struct nvme_dev *ndev = pci_get_drvdata(to_pci_dev(dev));
2900 	struct nvme_ctrl *ctrl = &ndev->ctrl;
2901 
2902 	if (ndev->last_ps == U32_MAX ||
2903 	    nvme_set_power_state(ctrl, ndev->last_ps) != 0)
2904 		nvme_reset_ctrl(ctrl);
2905 	return 0;
2906 }
2907 
2908 static int nvme_suspend(struct device *dev)
2909 {
2910 	struct pci_dev *pdev = to_pci_dev(dev);
2911 	struct nvme_dev *ndev = pci_get_drvdata(pdev);
2912 	struct nvme_ctrl *ctrl = &ndev->ctrl;
2913 	int ret = -EBUSY;
2914 
2915 	ndev->last_ps = U32_MAX;
2916 
2917 	/*
2918 	 * The platform does not remove power for a kernel managed suspend so
2919 	 * use host managed nvme power settings for lowest idle power if
2920 	 * possible. This should have quicker resume latency than a full device
2921 	 * shutdown.  But if the firmware is involved after the suspend or the
2922 	 * device does not support any non-default power states, shut down the
2923 	 * device fully.
2924 	 *
2925 	 * If ASPM is not enabled for the device, shut down the device and allow
2926 	 * the PCI bus layer to put it into D3 in order to take the PCIe link
2927 	 * down, so as to allow the platform to achieve its minimum low-power
2928 	 * state (which may not be possible if the link is up).
2929 	 */
2930 	if (pm_suspend_via_firmware() || !ctrl->npss ||
2931 	    !pcie_aspm_enabled(pdev) ||
2932 	    (ndev->ctrl.quirks & NVME_QUIRK_SIMPLE_SUSPEND)) {
2933 		nvme_dev_disable(ndev, true);
2934 		return 0;
2935 	}
2936 
2937 	nvme_start_freeze(ctrl);
2938 	nvme_wait_freeze(ctrl);
2939 	nvme_sync_queues(ctrl);
2940 
2941 	if (ctrl->state != NVME_CTRL_LIVE &&
2942 	    ctrl->state != NVME_CTRL_ADMIN_ONLY)
2943 		goto unfreeze;
2944 
2945 	ret = nvme_get_power_state(ctrl, &ndev->last_ps);
2946 	if (ret < 0)
2947 		goto unfreeze;
2948 
2949 	ret = nvme_set_power_state(ctrl, ctrl->npss);
2950 	if (ret < 0)
2951 		goto unfreeze;
2952 
2953 	if (ret) {
2954 		/*
2955 		 * Clearing npss forces a controller reset on resume. The
2956 		 * correct value will be resdicovered then.
2957 		 */
2958 		nvme_dev_disable(ndev, true);
2959 		ctrl->npss = 0;
2960 		ret = 0;
2961 		goto unfreeze;
2962 	}
2963 	/*
2964 	 * A saved state prevents pci pm from generically controlling the
2965 	 * device's power. If we're using protocol specific settings, we don't
2966 	 * want pci interfering.
2967 	 */
2968 	pci_save_state(pdev);
2969 unfreeze:
2970 	nvme_unfreeze(ctrl);
2971 	return ret;
2972 }
2973 
2974 static int nvme_simple_suspend(struct device *dev)
2975 {
2976 	struct nvme_dev *ndev = pci_get_drvdata(to_pci_dev(dev));
2977 
2978 	nvme_dev_disable(ndev, true);
2979 	return 0;
2980 }
2981 
2982 static int nvme_simple_resume(struct device *dev)
2983 {
2984 	struct pci_dev *pdev = to_pci_dev(dev);
2985 	struct nvme_dev *ndev = pci_get_drvdata(pdev);
2986 
2987 	nvme_reset_ctrl(&ndev->ctrl);
2988 	return 0;
2989 }
2990 
2991 static const struct dev_pm_ops nvme_dev_pm_ops = {
2992 	.suspend	= nvme_suspend,
2993 	.resume		= nvme_resume,
2994 	.freeze		= nvme_simple_suspend,
2995 	.thaw		= nvme_simple_resume,
2996 	.poweroff	= nvme_simple_suspend,
2997 	.restore	= nvme_simple_resume,
2998 };
2999 #endif /* CONFIG_PM_SLEEP */
3000 
3001 static pci_ers_result_t nvme_error_detected(struct pci_dev *pdev,
3002 						pci_channel_state_t state)
3003 {
3004 	struct nvme_dev *dev = pci_get_drvdata(pdev);
3005 
3006 	/*
3007 	 * A frozen channel requires a reset. When detected, this method will
3008 	 * shutdown the controller to quiesce. The controller will be restarted
3009 	 * after the slot reset through driver's slot_reset callback.
3010 	 */
3011 	switch (state) {
3012 	case pci_channel_io_normal:
3013 		return PCI_ERS_RESULT_CAN_RECOVER;
3014 	case pci_channel_io_frozen:
3015 		dev_warn(dev->ctrl.device,
3016 			"frozen state error detected, reset controller\n");
3017 		nvme_dev_disable(dev, false);
3018 		return PCI_ERS_RESULT_NEED_RESET;
3019 	case pci_channel_io_perm_failure:
3020 		dev_warn(dev->ctrl.device,
3021 			"failure state error detected, request disconnect\n");
3022 		return PCI_ERS_RESULT_DISCONNECT;
3023 	}
3024 	return PCI_ERS_RESULT_NEED_RESET;
3025 }
3026 
3027 static pci_ers_result_t nvme_slot_reset(struct pci_dev *pdev)
3028 {
3029 	struct nvme_dev *dev = pci_get_drvdata(pdev);
3030 
3031 	dev_info(dev->ctrl.device, "restart after slot reset\n");
3032 	pci_restore_state(pdev);
3033 	nvme_reset_ctrl(&dev->ctrl);
3034 	return PCI_ERS_RESULT_RECOVERED;
3035 }
3036 
3037 static void nvme_error_resume(struct pci_dev *pdev)
3038 {
3039 	struct nvme_dev *dev = pci_get_drvdata(pdev);
3040 
3041 	flush_work(&dev->ctrl.reset_work);
3042 }
3043 
3044 static const struct pci_error_handlers nvme_err_handler = {
3045 	.error_detected	= nvme_error_detected,
3046 	.slot_reset	= nvme_slot_reset,
3047 	.resume		= nvme_error_resume,
3048 	.reset_prepare	= nvme_reset_prepare,
3049 	.reset_done	= nvme_reset_done,
3050 };
3051 
3052 static const struct pci_device_id nvme_id_table[] = {
3053 	{ PCI_VDEVICE(INTEL, 0x0953),
3054 		.driver_data = NVME_QUIRK_STRIPE_SIZE |
3055 				NVME_QUIRK_DEALLOCATE_ZEROES, },
3056 	{ PCI_VDEVICE(INTEL, 0x0a53),
3057 		.driver_data = NVME_QUIRK_STRIPE_SIZE |
3058 				NVME_QUIRK_DEALLOCATE_ZEROES, },
3059 	{ PCI_VDEVICE(INTEL, 0x0a54),
3060 		.driver_data = NVME_QUIRK_STRIPE_SIZE |
3061 				NVME_QUIRK_DEALLOCATE_ZEROES, },
3062 	{ PCI_VDEVICE(INTEL, 0x0a55),
3063 		.driver_data = NVME_QUIRK_STRIPE_SIZE |
3064 				NVME_QUIRK_DEALLOCATE_ZEROES, },
3065 	{ PCI_VDEVICE(INTEL, 0xf1a5),	/* Intel 600P/P3100 */
3066 		.driver_data = NVME_QUIRK_NO_DEEPEST_PS |
3067 				NVME_QUIRK_MEDIUM_PRIO_SQ },
3068 	{ PCI_VDEVICE(INTEL, 0xf1a6),	/* Intel 760p/Pro 7600p */
3069 		.driver_data = NVME_QUIRK_IGNORE_DEV_SUBNQN, },
3070 	{ PCI_VDEVICE(INTEL, 0x5845),	/* Qemu emulated controller */
3071 		.driver_data = NVME_QUIRK_IDENTIFY_CNS |
3072 				NVME_QUIRK_DISABLE_WRITE_ZEROES, },
3073 	{ PCI_DEVICE(0x1bb1, 0x0100),   /* Seagate Nytro Flash Storage */
3074 		.driver_data = NVME_QUIRK_DELAY_BEFORE_CHK_RDY, },
3075 	{ PCI_DEVICE(0x1c58, 0x0003),	/* HGST adapter */
3076 		.driver_data = NVME_QUIRK_DELAY_BEFORE_CHK_RDY, },
3077 	{ PCI_DEVICE(0x1c58, 0x0023),	/* WDC SN200 adapter */
3078 		.driver_data = NVME_QUIRK_DELAY_BEFORE_CHK_RDY, },
3079 	{ PCI_DEVICE(0x1c5f, 0x0540),	/* Memblaze Pblaze4 adapter */
3080 		.driver_data = NVME_QUIRK_DELAY_BEFORE_CHK_RDY, },
3081 	{ PCI_DEVICE(0x144d, 0xa821),   /* Samsung PM1725 */
3082 		.driver_data = NVME_QUIRK_DELAY_BEFORE_CHK_RDY, },
3083 	{ PCI_DEVICE(0x144d, 0xa822),   /* Samsung PM1725a */
3084 		.driver_data = NVME_QUIRK_DELAY_BEFORE_CHK_RDY, },
3085 	{ PCI_DEVICE(0x1d1d, 0x1f1f),	/* LighNVM qemu device */
3086 		.driver_data = NVME_QUIRK_LIGHTNVM, },
3087 	{ PCI_DEVICE(0x1d1d, 0x2807),	/* CNEX WL */
3088 		.driver_data = NVME_QUIRK_LIGHTNVM, },
3089 	{ PCI_DEVICE(0x1d1d, 0x2601),	/* CNEX Granby */
3090 		.driver_data = NVME_QUIRK_LIGHTNVM, },
3091 	{ PCI_DEVICE(0x10ec, 0x5762),   /* ADATA SX6000LNP */
3092 		.driver_data = NVME_QUIRK_IGNORE_DEV_SUBNQN, },
3093 	{ PCI_DEVICE_CLASS(PCI_CLASS_STORAGE_EXPRESS, 0xffffff) },
3094 	{ PCI_DEVICE(PCI_VENDOR_ID_APPLE, 0x2001) },
3095 	{ PCI_DEVICE(PCI_VENDOR_ID_APPLE, 0x2003) },
3096 	{ PCI_DEVICE(PCI_VENDOR_ID_APPLE, 0x2005),
3097 		.driver_data = NVME_QUIRK_SINGLE_VECTOR |
3098 				NVME_QUIRK_128_BYTES_SQES |
3099 				NVME_QUIRK_SHARED_TAGS },
3100 	{ 0, }
3101 };
3102 MODULE_DEVICE_TABLE(pci, nvme_id_table);
3103 
3104 static struct pci_driver nvme_driver = {
3105 	.name		= "nvme",
3106 	.id_table	= nvme_id_table,
3107 	.probe		= nvme_probe,
3108 	.remove		= nvme_remove,
3109 	.shutdown	= nvme_shutdown,
3110 #ifdef CONFIG_PM_SLEEP
3111 	.driver		= {
3112 		.pm	= &nvme_dev_pm_ops,
3113 	},
3114 #endif
3115 	.sriov_configure = pci_sriov_configure_simple,
3116 	.err_handler	= &nvme_err_handler,
3117 };
3118 
3119 static int __init nvme_init(void)
3120 {
3121 	BUILD_BUG_ON(sizeof(struct nvme_create_cq) != 64);
3122 	BUILD_BUG_ON(sizeof(struct nvme_create_sq) != 64);
3123 	BUILD_BUG_ON(sizeof(struct nvme_delete_queue) != 64);
3124 	BUILD_BUG_ON(IRQ_AFFINITY_MAX_SETS < 2);
3125 	return pci_register_driver(&nvme_driver);
3126 }
3127 
3128 static void __exit nvme_exit(void)
3129 {
3130 	pci_unregister_driver(&nvme_driver);
3131 	flush_workqueue(nvme_wq);
3132 }
3133 
3134 MODULE_AUTHOR("Matthew Wilcox <willy@linux.intel.com>");
3135 MODULE_LICENSE("GPL");
3136 MODULE_VERSION("1.0");
3137 module_init(nvme_init);
3138 module_exit(nvme_exit);
3139