xref: /openbmc/linux/drivers/nvme/host/pci.c (revision 3ce7547e)
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/acpi.h>
8 #include <linux/aer.h>
9 #include <linux/async.h>
10 #include <linux/blkdev.h>
11 #include <linux/blk-mq.h>
12 #include <linux/blk-mq-pci.h>
13 #include <linux/blk-integrity.h>
14 #include <linux/dmi.h>
15 #include <linux/init.h>
16 #include <linux/interrupt.h>
17 #include <linux/io.h>
18 #include <linux/memremap.h>
19 #include <linux/mm.h>
20 #include <linux/module.h>
21 #include <linux/mutex.h>
22 #include <linux/once.h>
23 #include <linux/pci.h>
24 #include <linux/suspend.h>
25 #include <linux/t10-pi.h>
26 #include <linux/types.h>
27 #include <linux/io-64-nonatomic-lo-hi.h>
28 #include <linux/io-64-nonatomic-hi-lo.h>
29 #include <linux/sed-opal.h>
30 #include <linux/pci-p2pdma.h>
31 
32 #include "trace.h"
33 #include "nvme.h"
34 
35 #define SQ_SIZE(q)	((q)->q_depth << (q)->sqes)
36 #define CQ_SIZE(q)	((q)->q_depth * sizeof(struct nvme_completion))
37 
38 #define SGES_PER_PAGE	(PAGE_SIZE / sizeof(struct nvme_sgl_desc))
39 
40 /*
41  * These can be higher, but we need to ensure that any command doesn't
42  * require an sg allocation that needs more than a page of data.
43  */
44 #define NVME_MAX_KB_SZ	4096
45 #define NVME_MAX_SEGS	127
46 
47 static int use_threaded_interrupts;
48 module_param(use_threaded_interrupts, int, 0444);
49 
50 static bool use_cmb_sqes = true;
51 module_param(use_cmb_sqes, bool, 0444);
52 MODULE_PARM_DESC(use_cmb_sqes, "use controller's memory buffer for I/O SQes");
53 
54 static unsigned int max_host_mem_size_mb = 128;
55 module_param(max_host_mem_size_mb, uint, 0444);
56 MODULE_PARM_DESC(max_host_mem_size_mb,
57 	"Maximum Host Memory Buffer (HMB) size per controller (in MiB)");
58 
59 static unsigned int sgl_threshold = SZ_32K;
60 module_param(sgl_threshold, uint, 0644);
61 MODULE_PARM_DESC(sgl_threshold,
62 		"Use SGLs when average request segment size is larger or equal to "
63 		"this size. Use 0 to disable SGLs.");
64 
65 #define NVME_PCI_MIN_QUEUE_SIZE 2
66 #define NVME_PCI_MAX_QUEUE_SIZE 4095
67 static int io_queue_depth_set(const char *val, const struct kernel_param *kp);
68 static const struct kernel_param_ops io_queue_depth_ops = {
69 	.set = io_queue_depth_set,
70 	.get = param_get_uint,
71 };
72 
73 static unsigned int io_queue_depth = 1024;
74 module_param_cb(io_queue_depth, &io_queue_depth_ops, &io_queue_depth, 0644);
75 MODULE_PARM_DESC(io_queue_depth, "set io queue depth, should >= 2 and < 4096");
76 
77 static int io_queue_count_set(const char *val, const struct kernel_param *kp)
78 {
79 	unsigned int n;
80 	int ret;
81 
82 	ret = kstrtouint(val, 10, &n);
83 	if (ret != 0 || n > num_possible_cpus())
84 		return -EINVAL;
85 	return param_set_uint(val, kp);
86 }
87 
88 static const struct kernel_param_ops io_queue_count_ops = {
89 	.set = io_queue_count_set,
90 	.get = param_get_uint,
91 };
92 
93 static unsigned int write_queues;
94 module_param_cb(write_queues, &io_queue_count_ops, &write_queues, 0644);
95 MODULE_PARM_DESC(write_queues,
96 	"Number of queues to use for writes. If not set, reads and writes "
97 	"will share a queue set.");
98 
99 static unsigned int poll_queues;
100 module_param_cb(poll_queues, &io_queue_count_ops, &poll_queues, 0644);
101 MODULE_PARM_DESC(poll_queues, "Number of queues to use for polled IO.");
102 
103 static bool noacpi;
104 module_param(noacpi, bool, 0444);
105 MODULE_PARM_DESC(noacpi, "disable acpi bios quirks");
106 
107 struct nvme_dev;
108 struct nvme_queue;
109 
110 static void nvme_dev_disable(struct nvme_dev *dev, bool shutdown);
111 static bool __nvme_disable_io_queues(struct nvme_dev *dev, u8 opcode);
112 
113 /*
114  * Represents an NVM Express device.  Each nvme_dev is a PCI function.
115  */
116 struct nvme_dev {
117 	struct nvme_queue *queues;
118 	struct blk_mq_tag_set tagset;
119 	struct blk_mq_tag_set admin_tagset;
120 	u32 __iomem *dbs;
121 	struct device *dev;
122 	struct dma_pool *prp_page_pool;
123 	struct dma_pool *prp_small_pool;
124 	unsigned online_queues;
125 	unsigned max_qid;
126 	unsigned io_queues[HCTX_MAX_TYPES];
127 	unsigned int num_vecs;
128 	u32 q_depth;
129 	int io_sqes;
130 	u32 db_stride;
131 	void __iomem *bar;
132 	unsigned long bar_mapped_size;
133 	struct work_struct remove_work;
134 	struct mutex shutdown_lock;
135 	bool subsystem;
136 	u64 cmb_size;
137 	bool cmb_use_sqes;
138 	u32 cmbsz;
139 	u32 cmbloc;
140 	struct nvme_ctrl ctrl;
141 	u32 last_ps;
142 	bool hmb;
143 
144 	mempool_t *iod_mempool;
145 
146 	/* shadow doorbell buffer support: */
147 	u32 *dbbuf_dbs;
148 	dma_addr_t dbbuf_dbs_dma_addr;
149 	u32 *dbbuf_eis;
150 	dma_addr_t dbbuf_eis_dma_addr;
151 
152 	/* host memory buffer support: */
153 	u64 host_mem_size;
154 	u32 nr_host_mem_descs;
155 	dma_addr_t host_mem_descs_dma;
156 	struct nvme_host_mem_buf_desc *host_mem_descs;
157 	void **host_mem_desc_bufs;
158 	unsigned int nr_allocated_queues;
159 	unsigned int nr_write_queues;
160 	unsigned int nr_poll_queues;
161 
162 	bool attrs_added;
163 };
164 
165 static int io_queue_depth_set(const char *val, const struct kernel_param *kp)
166 {
167 	return param_set_uint_minmax(val, kp, NVME_PCI_MIN_QUEUE_SIZE,
168 			NVME_PCI_MAX_QUEUE_SIZE);
169 }
170 
171 static inline unsigned int sq_idx(unsigned int qid, u32 stride)
172 {
173 	return qid * 2 * stride;
174 }
175 
176 static inline unsigned int cq_idx(unsigned int qid, u32 stride)
177 {
178 	return (qid * 2 + 1) * stride;
179 }
180 
181 static inline struct nvme_dev *to_nvme_dev(struct nvme_ctrl *ctrl)
182 {
183 	return container_of(ctrl, struct nvme_dev, ctrl);
184 }
185 
186 /*
187  * An NVM Express queue.  Each device has at least two (one for admin
188  * commands and one for I/O commands).
189  */
190 struct nvme_queue {
191 	struct nvme_dev *dev;
192 	spinlock_t sq_lock;
193 	void *sq_cmds;
194 	 /* only used for poll queues: */
195 	spinlock_t cq_poll_lock ____cacheline_aligned_in_smp;
196 	struct nvme_completion *cqes;
197 	dma_addr_t sq_dma_addr;
198 	dma_addr_t cq_dma_addr;
199 	u32 __iomem *q_db;
200 	u32 q_depth;
201 	u16 cq_vector;
202 	u16 sq_tail;
203 	u16 last_sq_tail;
204 	u16 cq_head;
205 	u16 qid;
206 	u8 cq_phase;
207 	u8 sqes;
208 	unsigned long flags;
209 #define NVMEQ_ENABLED		0
210 #define NVMEQ_SQ_CMB		1
211 #define NVMEQ_DELETE_ERROR	2
212 #define NVMEQ_POLLED		3
213 	u32 *dbbuf_sq_db;
214 	u32 *dbbuf_cq_db;
215 	u32 *dbbuf_sq_ei;
216 	u32 *dbbuf_cq_ei;
217 	struct completion delete_done;
218 };
219 
220 /*
221  * The nvme_iod describes the data in an I/O.
222  *
223  * The sg pointer contains the list of PRP/SGL chunk allocations in addition
224  * to the actual struct scatterlist.
225  */
226 struct nvme_iod {
227 	struct nvme_request req;
228 	struct nvme_command cmd;
229 	struct nvme_queue *nvmeq;
230 	bool use_sgl;
231 	int aborted;
232 	int npages;		/* In the PRP list. 0 means small pool in use */
233 	int nents;		/* Used in scatterlist */
234 	dma_addr_t first_dma;
235 	unsigned int dma_len;	/* length of single DMA segment mapping */
236 	dma_addr_t meta_dma;
237 	struct scatterlist *sg;
238 };
239 
240 static inline unsigned int nvme_dbbuf_size(struct nvme_dev *dev)
241 {
242 	return dev->nr_allocated_queues * 8 * dev->db_stride;
243 }
244 
245 static int nvme_dbbuf_dma_alloc(struct nvme_dev *dev)
246 {
247 	unsigned int mem_size = nvme_dbbuf_size(dev);
248 
249 	if (dev->dbbuf_dbs) {
250 		/*
251 		 * Clear the dbbuf memory so the driver doesn't observe stale
252 		 * values from the previous instantiation.
253 		 */
254 		memset(dev->dbbuf_dbs, 0, mem_size);
255 		memset(dev->dbbuf_eis, 0, mem_size);
256 		return 0;
257 	}
258 
259 	dev->dbbuf_dbs = dma_alloc_coherent(dev->dev, mem_size,
260 					    &dev->dbbuf_dbs_dma_addr,
261 					    GFP_KERNEL);
262 	if (!dev->dbbuf_dbs)
263 		return -ENOMEM;
264 	dev->dbbuf_eis = dma_alloc_coherent(dev->dev, mem_size,
265 					    &dev->dbbuf_eis_dma_addr,
266 					    GFP_KERNEL);
267 	if (!dev->dbbuf_eis) {
268 		dma_free_coherent(dev->dev, mem_size,
269 				  dev->dbbuf_dbs, dev->dbbuf_dbs_dma_addr);
270 		dev->dbbuf_dbs = NULL;
271 		return -ENOMEM;
272 	}
273 
274 	return 0;
275 }
276 
277 static void nvme_dbbuf_dma_free(struct nvme_dev *dev)
278 {
279 	unsigned int mem_size = nvme_dbbuf_size(dev);
280 
281 	if (dev->dbbuf_dbs) {
282 		dma_free_coherent(dev->dev, mem_size,
283 				  dev->dbbuf_dbs, dev->dbbuf_dbs_dma_addr);
284 		dev->dbbuf_dbs = NULL;
285 	}
286 	if (dev->dbbuf_eis) {
287 		dma_free_coherent(dev->dev, mem_size,
288 				  dev->dbbuf_eis, dev->dbbuf_eis_dma_addr);
289 		dev->dbbuf_eis = NULL;
290 	}
291 }
292 
293 static void nvme_dbbuf_init(struct nvme_dev *dev,
294 			    struct nvme_queue *nvmeq, int qid)
295 {
296 	if (!dev->dbbuf_dbs || !qid)
297 		return;
298 
299 	nvmeq->dbbuf_sq_db = &dev->dbbuf_dbs[sq_idx(qid, dev->db_stride)];
300 	nvmeq->dbbuf_cq_db = &dev->dbbuf_dbs[cq_idx(qid, dev->db_stride)];
301 	nvmeq->dbbuf_sq_ei = &dev->dbbuf_eis[sq_idx(qid, dev->db_stride)];
302 	nvmeq->dbbuf_cq_ei = &dev->dbbuf_eis[cq_idx(qid, dev->db_stride)];
303 }
304 
305 static void nvme_dbbuf_free(struct nvme_queue *nvmeq)
306 {
307 	if (!nvmeq->qid)
308 		return;
309 
310 	nvmeq->dbbuf_sq_db = NULL;
311 	nvmeq->dbbuf_cq_db = NULL;
312 	nvmeq->dbbuf_sq_ei = NULL;
313 	nvmeq->dbbuf_cq_ei = NULL;
314 }
315 
316 static void nvme_dbbuf_set(struct nvme_dev *dev)
317 {
318 	struct nvme_command c = { };
319 	unsigned int i;
320 
321 	if (!dev->dbbuf_dbs)
322 		return;
323 
324 	c.dbbuf.opcode = nvme_admin_dbbuf;
325 	c.dbbuf.prp1 = cpu_to_le64(dev->dbbuf_dbs_dma_addr);
326 	c.dbbuf.prp2 = cpu_to_le64(dev->dbbuf_eis_dma_addr);
327 
328 	if (nvme_submit_sync_cmd(dev->ctrl.admin_q, &c, NULL, 0)) {
329 		dev_warn(dev->ctrl.device, "unable to set dbbuf\n");
330 		/* Free memory and continue on */
331 		nvme_dbbuf_dma_free(dev);
332 
333 		for (i = 1; i <= dev->online_queues; i++)
334 			nvme_dbbuf_free(&dev->queues[i]);
335 	}
336 }
337 
338 static inline int nvme_dbbuf_need_event(u16 event_idx, u16 new_idx, u16 old)
339 {
340 	return (u16)(new_idx - event_idx - 1) < (u16)(new_idx - old);
341 }
342 
343 /* Update dbbuf and return true if an MMIO is required */
344 static bool nvme_dbbuf_update_and_check_event(u16 value, u32 *dbbuf_db,
345 					      volatile u32 *dbbuf_ei)
346 {
347 	if (dbbuf_db) {
348 		u16 old_value;
349 
350 		/*
351 		 * Ensure that the queue is written before updating
352 		 * the doorbell in memory
353 		 */
354 		wmb();
355 
356 		old_value = *dbbuf_db;
357 		*dbbuf_db = value;
358 
359 		/*
360 		 * Ensure that the doorbell is updated before reading the event
361 		 * index from memory.  The controller needs to provide similar
362 		 * ordering to ensure the envent index is updated before reading
363 		 * the doorbell.
364 		 */
365 		mb();
366 
367 		if (!nvme_dbbuf_need_event(*dbbuf_ei, value, old_value))
368 			return false;
369 	}
370 
371 	return true;
372 }
373 
374 /*
375  * Will slightly overestimate the number of pages needed.  This is OK
376  * as it only leads to a small amount of wasted memory for the lifetime of
377  * the I/O.
378  */
379 static int nvme_pci_npages_prp(void)
380 {
381 	unsigned nprps = DIV_ROUND_UP(NVME_MAX_KB_SZ + NVME_CTRL_PAGE_SIZE,
382 				      NVME_CTRL_PAGE_SIZE);
383 	return DIV_ROUND_UP(8 * nprps, PAGE_SIZE - 8);
384 }
385 
386 /*
387  * Calculates the number of pages needed for the SGL segments. For example a 4k
388  * page can accommodate 256 SGL descriptors.
389  */
390 static int nvme_pci_npages_sgl(void)
391 {
392 	return DIV_ROUND_UP(NVME_MAX_SEGS * sizeof(struct nvme_sgl_desc),
393 			PAGE_SIZE);
394 }
395 
396 static size_t nvme_pci_iod_alloc_size(void)
397 {
398 	size_t npages = max(nvme_pci_npages_prp(), nvme_pci_npages_sgl());
399 
400 	return sizeof(__le64 *) * npages +
401 		sizeof(struct scatterlist) * NVME_MAX_SEGS;
402 }
403 
404 static int nvme_admin_init_hctx(struct blk_mq_hw_ctx *hctx, void *data,
405 				unsigned int hctx_idx)
406 {
407 	struct nvme_dev *dev = data;
408 	struct nvme_queue *nvmeq = &dev->queues[0];
409 
410 	WARN_ON(hctx_idx != 0);
411 	WARN_ON(dev->admin_tagset.tags[0] != hctx->tags);
412 
413 	hctx->driver_data = nvmeq;
414 	return 0;
415 }
416 
417 static int nvme_init_hctx(struct blk_mq_hw_ctx *hctx, void *data,
418 			  unsigned int hctx_idx)
419 {
420 	struct nvme_dev *dev = data;
421 	struct nvme_queue *nvmeq = &dev->queues[hctx_idx + 1];
422 
423 	WARN_ON(dev->tagset.tags[hctx_idx] != hctx->tags);
424 	hctx->driver_data = nvmeq;
425 	return 0;
426 }
427 
428 static int nvme_pci_init_request(struct blk_mq_tag_set *set,
429 		struct request *req, unsigned int hctx_idx,
430 		unsigned int numa_node)
431 {
432 	struct nvme_dev *dev = set->driver_data;
433 	struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
434 	int queue_idx = (set == &dev->tagset) ? hctx_idx + 1 : 0;
435 	struct nvme_queue *nvmeq = &dev->queues[queue_idx];
436 
437 	BUG_ON(!nvmeq);
438 	iod->nvmeq = nvmeq;
439 
440 	nvme_req(req)->ctrl = &dev->ctrl;
441 	nvme_req(req)->cmd = &iod->cmd;
442 	return 0;
443 }
444 
445 static int queue_irq_offset(struct nvme_dev *dev)
446 {
447 	/* if we have more than 1 vec, admin queue offsets us by 1 */
448 	if (dev->num_vecs > 1)
449 		return 1;
450 
451 	return 0;
452 }
453 
454 static int nvme_pci_map_queues(struct blk_mq_tag_set *set)
455 {
456 	struct nvme_dev *dev = set->driver_data;
457 	int i, qoff, offset;
458 
459 	offset = queue_irq_offset(dev);
460 	for (i = 0, qoff = 0; i < set->nr_maps; i++) {
461 		struct blk_mq_queue_map *map = &set->map[i];
462 
463 		map->nr_queues = dev->io_queues[i];
464 		if (!map->nr_queues) {
465 			BUG_ON(i == HCTX_TYPE_DEFAULT);
466 			continue;
467 		}
468 
469 		/*
470 		 * The poll queue(s) doesn't have an IRQ (and hence IRQ
471 		 * affinity), so use the regular blk-mq cpu mapping
472 		 */
473 		map->queue_offset = qoff;
474 		if (i != HCTX_TYPE_POLL && offset)
475 			blk_mq_pci_map_queues(map, to_pci_dev(dev->dev), offset);
476 		else
477 			blk_mq_map_queues(map);
478 		qoff += map->nr_queues;
479 		offset += map->nr_queues;
480 	}
481 
482 	return 0;
483 }
484 
485 /*
486  * Write sq tail if we are asked to, or if the next command would wrap.
487  */
488 static inline void nvme_write_sq_db(struct nvme_queue *nvmeq, bool write_sq)
489 {
490 	if (!write_sq) {
491 		u16 next_tail = nvmeq->sq_tail + 1;
492 
493 		if (next_tail == nvmeq->q_depth)
494 			next_tail = 0;
495 		if (next_tail != nvmeq->last_sq_tail)
496 			return;
497 	}
498 
499 	if (nvme_dbbuf_update_and_check_event(nvmeq->sq_tail,
500 			nvmeq->dbbuf_sq_db, nvmeq->dbbuf_sq_ei))
501 		writel(nvmeq->sq_tail, nvmeq->q_db);
502 	nvmeq->last_sq_tail = nvmeq->sq_tail;
503 }
504 
505 static inline void nvme_sq_copy_cmd(struct nvme_queue *nvmeq,
506 				    struct nvme_command *cmd)
507 {
508 	memcpy(nvmeq->sq_cmds + (nvmeq->sq_tail << nvmeq->sqes),
509 		absolute_pointer(cmd), sizeof(*cmd));
510 	if (++nvmeq->sq_tail == nvmeq->q_depth)
511 		nvmeq->sq_tail = 0;
512 }
513 
514 static void nvme_commit_rqs(struct blk_mq_hw_ctx *hctx)
515 {
516 	struct nvme_queue *nvmeq = hctx->driver_data;
517 
518 	spin_lock(&nvmeq->sq_lock);
519 	if (nvmeq->sq_tail != nvmeq->last_sq_tail)
520 		nvme_write_sq_db(nvmeq, true);
521 	spin_unlock(&nvmeq->sq_lock);
522 }
523 
524 static void **nvme_pci_iod_list(struct request *req)
525 {
526 	struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
527 	return (void **)(iod->sg + blk_rq_nr_phys_segments(req));
528 }
529 
530 static inline bool nvme_pci_use_sgls(struct nvme_dev *dev, struct request *req)
531 {
532 	struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
533 	int nseg = blk_rq_nr_phys_segments(req);
534 	unsigned int avg_seg_size;
535 
536 	avg_seg_size = DIV_ROUND_UP(blk_rq_payload_bytes(req), nseg);
537 
538 	if (!nvme_ctrl_sgl_supported(&dev->ctrl))
539 		return false;
540 	if (!iod->nvmeq->qid)
541 		return false;
542 	if (!sgl_threshold || avg_seg_size < sgl_threshold)
543 		return false;
544 	return true;
545 }
546 
547 static void nvme_free_prps(struct nvme_dev *dev, struct request *req)
548 {
549 	const int last_prp = NVME_CTRL_PAGE_SIZE / sizeof(__le64) - 1;
550 	struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
551 	dma_addr_t dma_addr = iod->first_dma;
552 	int i;
553 
554 	for (i = 0; i < iod->npages; i++) {
555 		__le64 *prp_list = nvme_pci_iod_list(req)[i];
556 		dma_addr_t next_dma_addr = le64_to_cpu(prp_list[last_prp]);
557 
558 		dma_pool_free(dev->prp_page_pool, prp_list, dma_addr);
559 		dma_addr = next_dma_addr;
560 	}
561 }
562 
563 static void nvme_free_sgls(struct nvme_dev *dev, struct request *req)
564 {
565 	const int last_sg = SGES_PER_PAGE - 1;
566 	struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
567 	dma_addr_t dma_addr = iod->first_dma;
568 	int i;
569 
570 	for (i = 0; i < iod->npages; i++) {
571 		struct nvme_sgl_desc *sg_list = nvme_pci_iod_list(req)[i];
572 		dma_addr_t next_dma_addr = le64_to_cpu((sg_list[last_sg]).addr);
573 
574 		dma_pool_free(dev->prp_page_pool, sg_list, dma_addr);
575 		dma_addr = next_dma_addr;
576 	}
577 }
578 
579 static void nvme_unmap_sg(struct nvme_dev *dev, struct request *req)
580 {
581 	struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
582 
583 	if (is_pci_p2pdma_page(sg_page(iod->sg)))
584 		pci_p2pdma_unmap_sg(dev->dev, iod->sg, iod->nents,
585 				    rq_dma_dir(req));
586 	else
587 		dma_unmap_sg(dev->dev, iod->sg, iod->nents, rq_dma_dir(req));
588 }
589 
590 static void nvme_unmap_data(struct nvme_dev *dev, struct request *req)
591 {
592 	struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
593 
594 	if (iod->dma_len) {
595 		dma_unmap_page(dev->dev, iod->first_dma, iod->dma_len,
596 			       rq_dma_dir(req));
597 		return;
598 	}
599 
600 	WARN_ON_ONCE(!iod->nents);
601 
602 	nvme_unmap_sg(dev, req);
603 	if (iod->npages == 0)
604 		dma_pool_free(dev->prp_small_pool, nvme_pci_iod_list(req)[0],
605 			      iod->first_dma);
606 	else if (iod->use_sgl)
607 		nvme_free_sgls(dev, req);
608 	else
609 		nvme_free_prps(dev, req);
610 	mempool_free(iod->sg, dev->iod_mempool);
611 }
612 
613 static void nvme_print_sgl(struct scatterlist *sgl, int nents)
614 {
615 	int i;
616 	struct scatterlist *sg;
617 
618 	for_each_sg(sgl, sg, nents, i) {
619 		dma_addr_t phys = sg_phys(sg);
620 		pr_warn("sg[%d] phys_addr:%pad offset:%d length:%d "
621 			"dma_address:%pad dma_length:%d\n",
622 			i, &phys, sg->offset, sg->length, &sg_dma_address(sg),
623 			sg_dma_len(sg));
624 	}
625 }
626 
627 static blk_status_t nvme_pci_setup_prps(struct nvme_dev *dev,
628 		struct request *req, struct nvme_rw_command *cmnd)
629 {
630 	struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
631 	struct dma_pool *pool;
632 	int length = blk_rq_payload_bytes(req);
633 	struct scatterlist *sg = iod->sg;
634 	int dma_len = sg_dma_len(sg);
635 	u64 dma_addr = sg_dma_address(sg);
636 	int offset = dma_addr & (NVME_CTRL_PAGE_SIZE - 1);
637 	__le64 *prp_list;
638 	void **list = nvme_pci_iod_list(req);
639 	dma_addr_t prp_dma;
640 	int nprps, i;
641 
642 	length -= (NVME_CTRL_PAGE_SIZE - offset);
643 	if (length <= 0) {
644 		iod->first_dma = 0;
645 		goto done;
646 	}
647 
648 	dma_len -= (NVME_CTRL_PAGE_SIZE - offset);
649 	if (dma_len) {
650 		dma_addr += (NVME_CTRL_PAGE_SIZE - offset);
651 	} else {
652 		sg = sg_next(sg);
653 		dma_addr = sg_dma_address(sg);
654 		dma_len = sg_dma_len(sg);
655 	}
656 
657 	if (length <= NVME_CTRL_PAGE_SIZE) {
658 		iod->first_dma = dma_addr;
659 		goto done;
660 	}
661 
662 	nprps = DIV_ROUND_UP(length, NVME_CTRL_PAGE_SIZE);
663 	if (nprps <= (256 / 8)) {
664 		pool = dev->prp_small_pool;
665 		iod->npages = 0;
666 	} else {
667 		pool = dev->prp_page_pool;
668 		iod->npages = 1;
669 	}
670 
671 	prp_list = dma_pool_alloc(pool, GFP_ATOMIC, &prp_dma);
672 	if (!prp_list) {
673 		iod->first_dma = dma_addr;
674 		iod->npages = -1;
675 		return BLK_STS_RESOURCE;
676 	}
677 	list[0] = prp_list;
678 	iod->first_dma = prp_dma;
679 	i = 0;
680 	for (;;) {
681 		if (i == NVME_CTRL_PAGE_SIZE >> 3) {
682 			__le64 *old_prp_list = prp_list;
683 			prp_list = dma_pool_alloc(pool, GFP_ATOMIC, &prp_dma);
684 			if (!prp_list)
685 				goto free_prps;
686 			list[iod->npages++] = prp_list;
687 			prp_list[0] = old_prp_list[i - 1];
688 			old_prp_list[i - 1] = cpu_to_le64(prp_dma);
689 			i = 1;
690 		}
691 		prp_list[i++] = cpu_to_le64(dma_addr);
692 		dma_len -= NVME_CTRL_PAGE_SIZE;
693 		dma_addr += NVME_CTRL_PAGE_SIZE;
694 		length -= NVME_CTRL_PAGE_SIZE;
695 		if (length <= 0)
696 			break;
697 		if (dma_len > 0)
698 			continue;
699 		if (unlikely(dma_len < 0))
700 			goto bad_sgl;
701 		sg = sg_next(sg);
702 		dma_addr = sg_dma_address(sg);
703 		dma_len = sg_dma_len(sg);
704 	}
705 done:
706 	cmnd->dptr.prp1 = cpu_to_le64(sg_dma_address(iod->sg));
707 	cmnd->dptr.prp2 = cpu_to_le64(iod->first_dma);
708 	return BLK_STS_OK;
709 free_prps:
710 	nvme_free_prps(dev, req);
711 	return BLK_STS_RESOURCE;
712 bad_sgl:
713 	WARN(DO_ONCE(nvme_print_sgl, iod->sg, iod->nents),
714 			"Invalid SGL for payload:%d nents:%d\n",
715 			blk_rq_payload_bytes(req), iod->nents);
716 	return BLK_STS_IOERR;
717 }
718 
719 static void nvme_pci_sgl_set_data(struct nvme_sgl_desc *sge,
720 		struct scatterlist *sg)
721 {
722 	sge->addr = cpu_to_le64(sg_dma_address(sg));
723 	sge->length = cpu_to_le32(sg_dma_len(sg));
724 	sge->type = NVME_SGL_FMT_DATA_DESC << 4;
725 }
726 
727 static void nvme_pci_sgl_set_seg(struct nvme_sgl_desc *sge,
728 		dma_addr_t dma_addr, int entries)
729 {
730 	sge->addr = cpu_to_le64(dma_addr);
731 	if (entries < SGES_PER_PAGE) {
732 		sge->length = cpu_to_le32(entries * sizeof(*sge));
733 		sge->type = NVME_SGL_FMT_LAST_SEG_DESC << 4;
734 	} else {
735 		sge->length = cpu_to_le32(PAGE_SIZE);
736 		sge->type = NVME_SGL_FMT_SEG_DESC << 4;
737 	}
738 }
739 
740 static blk_status_t nvme_pci_setup_sgls(struct nvme_dev *dev,
741 		struct request *req, struct nvme_rw_command *cmd, int entries)
742 {
743 	struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
744 	struct dma_pool *pool;
745 	struct nvme_sgl_desc *sg_list;
746 	struct scatterlist *sg = iod->sg;
747 	dma_addr_t sgl_dma;
748 	int i = 0;
749 
750 	/* setting the transfer type as SGL */
751 	cmd->flags = NVME_CMD_SGL_METABUF;
752 
753 	if (entries == 1) {
754 		nvme_pci_sgl_set_data(&cmd->dptr.sgl, sg);
755 		return BLK_STS_OK;
756 	}
757 
758 	if (entries <= (256 / sizeof(struct nvme_sgl_desc))) {
759 		pool = dev->prp_small_pool;
760 		iod->npages = 0;
761 	} else {
762 		pool = dev->prp_page_pool;
763 		iod->npages = 1;
764 	}
765 
766 	sg_list = dma_pool_alloc(pool, GFP_ATOMIC, &sgl_dma);
767 	if (!sg_list) {
768 		iod->npages = -1;
769 		return BLK_STS_RESOURCE;
770 	}
771 
772 	nvme_pci_iod_list(req)[0] = sg_list;
773 	iod->first_dma = sgl_dma;
774 
775 	nvme_pci_sgl_set_seg(&cmd->dptr.sgl, sgl_dma, entries);
776 
777 	do {
778 		if (i == SGES_PER_PAGE) {
779 			struct nvme_sgl_desc *old_sg_desc = sg_list;
780 			struct nvme_sgl_desc *link = &old_sg_desc[i - 1];
781 
782 			sg_list = dma_pool_alloc(pool, GFP_ATOMIC, &sgl_dma);
783 			if (!sg_list)
784 				goto free_sgls;
785 
786 			i = 0;
787 			nvme_pci_iod_list(req)[iod->npages++] = sg_list;
788 			sg_list[i++] = *link;
789 			nvme_pci_sgl_set_seg(link, sgl_dma, entries);
790 		}
791 
792 		nvme_pci_sgl_set_data(&sg_list[i++], sg);
793 		sg = sg_next(sg);
794 	} while (--entries > 0);
795 
796 	return BLK_STS_OK;
797 free_sgls:
798 	nvme_free_sgls(dev, req);
799 	return BLK_STS_RESOURCE;
800 }
801 
802 static blk_status_t nvme_setup_prp_simple(struct nvme_dev *dev,
803 		struct request *req, struct nvme_rw_command *cmnd,
804 		struct bio_vec *bv)
805 {
806 	struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
807 	unsigned int offset = bv->bv_offset & (NVME_CTRL_PAGE_SIZE - 1);
808 	unsigned int first_prp_len = NVME_CTRL_PAGE_SIZE - offset;
809 
810 	iod->first_dma = dma_map_bvec(dev->dev, bv, rq_dma_dir(req), 0);
811 	if (dma_mapping_error(dev->dev, iod->first_dma))
812 		return BLK_STS_RESOURCE;
813 	iod->dma_len = bv->bv_len;
814 
815 	cmnd->dptr.prp1 = cpu_to_le64(iod->first_dma);
816 	if (bv->bv_len > first_prp_len)
817 		cmnd->dptr.prp2 = cpu_to_le64(iod->first_dma + first_prp_len);
818 	return BLK_STS_OK;
819 }
820 
821 static blk_status_t nvme_setup_sgl_simple(struct nvme_dev *dev,
822 		struct request *req, struct nvme_rw_command *cmnd,
823 		struct bio_vec *bv)
824 {
825 	struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
826 
827 	iod->first_dma = dma_map_bvec(dev->dev, bv, rq_dma_dir(req), 0);
828 	if (dma_mapping_error(dev->dev, iod->first_dma))
829 		return BLK_STS_RESOURCE;
830 	iod->dma_len = bv->bv_len;
831 
832 	cmnd->flags = NVME_CMD_SGL_METABUF;
833 	cmnd->dptr.sgl.addr = cpu_to_le64(iod->first_dma);
834 	cmnd->dptr.sgl.length = cpu_to_le32(iod->dma_len);
835 	cmnd->dptr.sgl.type = NVME_SGL_FMT_DATA_DESC << 4;
836 	return BLK_STS_OK;
837 }
838 
839 static blk_status_t nvme_map_data(struct nvme_dev *dev, struct request *req,
840 		struct nvme_command *cmnd)
841 {
842 	struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
843 	blk_status_t ret = BLK_STS_RESOURCE;
844 	int nr_mapped;
845 
846 	if (blk_rq_nr_phys_segments(req) == 1) {
847 		struct bio_vec bv = req_bvec(req);
848 
849 		if (!is_pci_p2pdma_page(bv.bv_page)) {
850 			if (bv.bv_offset + bv.bv_len <= NVME_CTRL_PAGE_SIZE * 2)
851 				return nvme_setup_prp_simple(dev, req,
852 							     &cmnd->rw, &bv);
853 
854 			if (iod->nvmeq->qid && sgl_threshold &&
855 			    nvme_ctrl_sgl_supported(&dev->ctrl))
856 				return nvme_setup_sgl_simple(dev, req,
857 							     &cmnd->rw, &bv);
858 		}
859 	}
860 
861 	iod->dma_len = 0;
862 	iod->sg = mempool_alloc(dev->iod_mempool, GFP_ATOMIC);
863 	if (!iod->sg)
864 		return BLK_STS_RESOURCE;
865 	sg_init_table(iod->sg, blk_rq_nr_phys_segments(req));
866 	iod->nents = blk_rq_map_sg(req->q, req, iod->sg);
867 	if (!iod->nents)
868 		goto out_free_sg;
869 
870 	if (is_pci_p2pdma_page(sg_page(iod->sg)))
871 		nr_mapped = pci_p2pdma_map_sg_attrs(dev->dev, iod->sg,
872 				iod->nents, rq_dma_dir(req), DMA_ATTR_NO_WARN);
873 	else
874 		nr_mapped = dma_map_sg_attrs(dev->dev, iod->sg, iod->nents,
875 					     rq_dma_dir(req), DMA_ATTR_NO_WARN);
876 	if (!nr_mapped)
877 		goto out_free_sg;
878 
879 	iod->use_sgl = nvme_pci_use_sgls(dev, req);
880 	if (iod->use_sgl)
881 		ret = nvme_pci_setup_sgls(dev, req, &cmnd->rw, nr_mapped);
882 	else
883 		ret = nvme_pci_setup_prps(dev, req, &cmnd->rw);
884 	if (ret != BLK_STS_OK)
885 		goto out_unmap_sg;
886 	return BLK_STS_OK;
887 
888 out_unmap_sg:
889 	nvme_unmap_sg(dev, req);
890 out_free_sg:
891 	mempool_free(iod->sg, dev->iod_mempool);
892 	return ret;
893 }
894 
895 static blk_status_t nvme_map_metadata(struct nvme_dev *dev, struct request *req,
896 		struct nvme_command *cmnd)
897 {
898 	struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
899 
900 	iod->meta_dma = dma_map_bvec(dev->dev, rq_integrity_vec(req),
901 			rq_dma_dir(req), 0);
902 	if (dma_mapping_error(dev->dev, iod->meta_dma))
903 		return BLK_STS_IOERR;
904 	cmnd->rw.metadata = cpu_to_le64(iod->meta_dma);
905 	return BLK_STS_OK;
906 }
907 
908 static blk_status_t nvme_prep_rq(struct nvme_dev *dev, struct request *req)
909 {
910 	struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
911 	blk_status_t ret;
912 
913 	iod->aborted = 0;
914 	iod->npages = -1;
915 	iod->nents = 0;
916 
917 	ret = nvme_setup_cmd(req->q->queuedata, req);
918 	if (ret)
919 		return ret;
920 
921 	if (blk_rq_nr_phys_segments(req)) {
922 		ret = nvme_map_data(dev, req, &iod->cmd);
923 		if (ret)
924 			goto out_free_cmd;
925 	}
926 
927 	if (blk_integrity_rq(req)) {
928 		ret = nvme_map_metadata(dev, req, &iod->cmd);
929 		if (ret)
930 			goto out_unmap_data;
931 	}
932 
933 	blk_mq_start_request(req);
934 	return BLK_STS_OK;
935 out_unmap_data:
936 	nvme_unmap_data(dev, req);
937 out_free_cmd:
938 	nvme_cleanup_cmd(req);
939 	return ret;
940 }
941 
942 /*
943  * NOTE: ns is NULL when called on the admin queue.
944  */
945 static blk_status_t nvme_queue_rq(struct blk_mq_hw_ctx *hctx,
946 			 const struct blk_mq_queue_data *bd)
947 {
948 	struct nvme_queue *nvmeq = hctx->driver_data;
949 	struct nvme_dev *dev = nvmeq->dev;
950 	struct request *req = bd->rq;
951 	struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
952 	blk_status_t ret;
953 
954 	/*
955 	 * We should not need to do this, but we're still using this to
956 	 * ensure we can drain requests on a dying queue.
957 	 */
958 	if (unlikely(!test_bit(NVMEQ_ENABLED, &nvmeq->flags)))
959 		return BLK_STS_IOERR;
960 
961 	if (unlikely(!nvme_check_ready(&dev->ctrl, req, true)))
962 		return nvme_fail_nonready_command(&dev->ctrl, req);
963 
964 	ret = nvme_prep_rq(dev, req);
965 	if (unlikely(ret))
966 		return ret;
967 	spin_lock(&nvmeq->sq_lock);
968 	nvme_sq_copy_cmd(nvmeq, &iod->cmd);
969 	nvme_write_sq_db(nvmeq, bd->last);
970 	spin_unlock(&nvmeq->sq_lock);
971 	return BLK_STS_OK;
972 }
973 
974 static void nvme_submit_cmds(struct nvme_queue *nvmeq, struct request **rqlist)
975 {
976 	spin_lock(&nvmeq->sq_lock);
977 	while (!rq_list_empty(*rqlist)) {
978 		struct request *req = rq_list_pop(rqlist);
979 		struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
980 
981 		nvme_sq_copy_cmd(nvmeq, &iod->cmd);
982 	}
983 	nvme_write_sq_db(nvmeq, true);
984 	spin_unlock(&nvmeq->sq_lock);
985 }
986 
987 static bool nvme_prep_rq_batch(struct nvme_queue *nvmeq, struct request *req)
988 {
989 	/*
990 	 * We should not need to do this, but we're still using this to
991 	 * ensure we can drain requests on a dying queue.
992 	 */
993 	if (unlikely(!test_bit(NVMEQ_ENABLED, &nvmeq->flags)))
994 		return false;
995 	if (unlikely(!nvme_check_ready(&nvmeq->dev->ctrl, req, true)))
996 		return false;
997 
998 	req->mq_hctx->tags->rqs[req->tag] = req;
999 	return nvme_prep_rq(nvmeq->dev, req) == BLK_STS_OK;
1000 }
1001 
1002 static void nvme_queue_rqs(struct request **rqlist)
1003 {
1004 	struct request *req, *next, *prev = NULL;
1005 	struct request *requeue_list = NULL;
1006 
1007 	rq_list_for_each_safe(rqlist, req, next) {
1008 		struct nvme_queue *nvmeq = req->mq_hctx->driver_data;
1009 
1010 		if (!nvme_prep_rq_batch(nvmeq, req)) {
1011 			/* detach 'req' and add to remainder list */
1012 			rq_list_move(rqlist, &requeue_list, req, prev);
1013 
1014 			req = prev;
1015 			if (!req)
1016 				continue;
1017 		}
1018 
1019 		if (!next || req->mq_hctx != next->mq_hctx) {
1020 			/* detach rest of list, and submit */
1021 			req->rq_next = NULL;
1022 			nvme_submit_cmds(nvmeq, rqlist);
1023 			*rqlist = next;
1024 			prev = NULL;
1025 		} else
1026 			prev = req;
1027 	}
1028 
1029 	*rqlist = requeue_list;
1030 }
1031 
1032 static __always_inline void nvme_pci_unmap_rq(struct request *req)
1033 {
1034 	struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
1035 	struct nvme_dev *dev = iod->nvmeq->dev;
1036 
1037 	if (blk_integrity_rq(req))
1038 		dma_unmap_page(dev->dev, iod->meta_dma,
1039 			       rq_integrity_vec(req)->bv_len, rq_data_dir(req));
1040 	if (blk_rq_nr_phys_segments(req))
1041 		nvme_unmap_data(dev, req);
1042 }
1043 
1044 static void nvme_pci_complete_rq(struct request *req)
1045 {
1046 	nvme_pci_unmap_rq(req);
1047 	nvme_complete_rq(req);
1048 }
1049 
1050 static void nvme_pci_complete_batch(struct io_comp_batch *iob)
1051 {
1052 	nvme_complete_batch(iob, nvme_pci_unmap_rq);
1053 }
1054 
1055 /* We read the CQE phase first to check if the rest of the entry is valid */
1056 static inline bool nvme_cqe_pending(struct nvme_queue *nvmeq)
1057 {
1058 	struct nvme_completion *hcqe = &nvmeq->cqes[nvmeq->cq_head];
1059 
1060 	return (le16_to_cpu(READ_ONCE(hcqe->status)) & 1) == nvmeq->cq_phase;
1061 }
1062 
1063 static inline void nvme_ring_cq_doorbell(struct nvme_queue *nvmeq)
1064 {
1065 	u16 head = nvmeq->cq_head;
1066 
1067 	if (nvme_dbbuf_update_and_check_event(head, nvmeq->dbbuf_cq_db,
1068 					      nvmeq->dbbuf_cq_ei))
1069 		writel(head, nvmeq->q_db + nvmeq->dev->db_stride);
1070 }
1071 
1072 static inline struct blk_mq_tags *nvme_queue_tagset(struct nvme_queue *nvmeq)
1073 {
1074 	if (!nvmeq->qid)
1075 		return nvmeq->dev->admin_tagset.tags[0];
1076 	return nvmeq->dev->tagset.tags[nvmeq->qid - 1];
1077 }
1078 
1079 static inline void nvme_handle_cqe(struct nvme_queue *nvmeq,
1080 				   struct io_comp_batch *iob, u16 idx)
1081 {
1082 	struct nvme_completion *cqe = &nvmeq->cqes[idx];
1083 	__u16 command_id = READ_ONCE(cqe->command_id);
1084 	struct request *req;
1085 
1086 	/*
1087 	 * AEN requests are special as they don't time out and can
1088 	 * survive any kind of queue freeze and often don't respond to
1089 	 * aborts.  We don't even bother to allocate a struct request
1090 	 * for them but rather special case them here.
1091 	 */
1092 	if (unlikely(nvme_is_aen_req(nvmeq->qid, command_id))) {
1093 		nvme_complete_async_event(&nvmeq->dev->ctrl,
1094 				cqe->status, &cqe->result);
1095 		return;
1096 	}
1097 
1098 	req = nvme_find_rq(nvme_queue_tagset(nvmeq), command_id);
1099 	if (unlikely(!req)) {
1100 		dev_warn(nvmeq->dev->ctrl.device,
1101 			"invalid id %d completed on queue %d\n",
1102 			command_id, le16_to_cpu(cqe->sq_id));
1103 		return;
1104 	}
1105 
1106 	trace_nvme_sq(req, cqe->sq_head, nvmeq->sq_tail);
1107 	if (!nvme_try_complete_req(req, cqe->status, cqe->result) &&
1108 	    !blk_mq_add_to_batch(req, iob, nvme_req(req)->status,
1109 					nvme_pci_complete_batch))
1110 		nvme_pci_complete_rq(req);
1111 }
1112 
1113 static inline void nvme_update_cq_head(struct nvme_queue *nvmeq)
1114 {
1115 	u32 tmp = nvmeq->cq_head + 1;
1116 
1117 	if (tmp == nvmeq->q_depth) {
1118 		nvmeq->cq_head = 0;
1119 		nvmeq->cq_phase ^= 1;
1120 	} else {
1121 		nvmeq->cq_head = tmp;
1122 	}
1123 }
1124 
1125 static inline int nvme_poll_cq(struct nvme_queue *nvmeq,
1126 			       struct io_comp_batch *iob)
1127 {
1128 	int found = 0;
1129 
1130 	while (nvme_cqe_pending(nvmeq)) {
1131 		found++;
1132 		/*
1133 		 * load-load control dependency between phase and the rest of
1134 		 * the cqe requires a full read memory barrier
1135 		 */
1136 		dma_rmb();
1137 		nvme_handle_cqe(nvmeq, iob, nvmeq->cq_head);
1138 		nvme_update_cq_head(nvmeq);
1139 	}
1140 
1141 	if (found)
1142 		nvme_ring_cq_doorbell(nvmeq);
1143 	return found;
1144 }
1145 
1146 static irqreturn_t nvme_irq(int irq, void *data)
1147 {
1148 	struct nvme_queue *nvmeq = data;
1149 	DEFINE_IO_COMP_BATCH(iob);
1150 
1151 	if (nvme_poll_cq(nvmeq, &iob)) {
1152 		if (!rq_list_empty(iob.req_list))
1153 			nvme_pci_complete_batch(&iob);
1154 		return IRQ_HANDLED;
1155 	}
1156 	return IRQ_NONE;
1157 }
1158 
1159 static irqreturn_t nvme_irq_check(int irq, void *data)
1160 {
1161 	struct nvme_queue *nvmeq = data;
1162 
1163 	if (nvme_cqe_pending(nvmeq))
1164 		return IRQ_WAKE_THREAD;
1165 	return IRQ_NONE;
1166 }
1167 
1168 /*
1169  * Poll for completions for any interrupt driven queue
1170  * Can be called from any context.
1171  */
1172 static void nvme_poll_irqdisable(struct nvme_queue *nvmeq)
1173 {
1174 	struct pci_dev *pdev = to_pci_dev(nvmeq->dev->dev);
1175 
1176 	WARN_ON_ONCE(test_bit(NVMEQ_POLLED, &nvmeq->flags));
1177 
1178 	disable_irq(pci_irq_vector(pdev, nvmeq->cq_vector));
1179 	nvme_poll_cq(nvmeq, NULL);
1180 	enable_irq(pci_irq_vector(pdev, nvmeq->cq_vector));
1181 }
1182 
1183 static int nvme_poll(struct blk_mq_hw_ctx *hctx, struct io_comp_batch *iob)
1184 {
1185 	struct nvme_queue *nvmeq = hctx->driver_data;
1186 	bool found;
1187 
1188 	if (!nvme_cqe_pending(nvmeq))
1189 		return 0;
1190 
1191 	spin_lock(&nvmeq->cq_poll_lock);
1192 	found = nvme_poll_cq(nvmeq, iob);
1193 	spin_unlock(&nvmeq->cq_poll_lock);
1194 
1195 	return found;
1196 }
1197 
1198 static void nvme_pci_submit_async_event(struct nvme_ctrl *ctrl)
1199 {
1200 	struct nvme_dev *dev = to_nvme_dev(ctrl);
1201 	struct nvme_queue *nvmeq = &dev->queues[0];
1202 	struct nvme_command c = { };
1203 
1204 	c.common.opcode = nvme_admin_async_event;
1205 	c.common.command_id = NVME_AQ_BLK_MQ_DEPTH;
1206 
1207 	spin_lock(&nvmeq->sq_lock);
1208 	nvme_sq_copy_cmd(nvmeq, &c);
1209 	nvme_write_sq_db(nvmeq, true);
1210 	spin_unlock(&nvmeq->sq_lock);
1211 }
1212 
1213 static int adapter_delete_queue(struct nvme_dev *dev, u8 opcode, u16 id)
1214 {
1215 	struct nvme_command c = { };
1216 
1217 	c.delete_queue.opcode = opcode;
1218 	c.delete_queue.qid = cpu_to_le16(id);
1219 
1220 	return nvme_submit_sync_cmd(dev->ctrl.admin_q, &c, NULL, 0);
1221 }
1222 
1223 static int adapter_alloc_cq(struct nvme_dev *dev, u16 qid,
1224 		struct nvme_queue *nvmeq, s16 vector)
1225 {
1226 	struct nvme_command c = { };
1227 	int flags = NVME_QUEUE_PHYS_CONTIG;
1228 
1229 	if (!test_bit(NVMEQ_POLLED, &nvmeq->flags))
1230 		flags |= NVME_CQ_IRQ_ENABLED;
1231 
1232 	/*
1233 	 * Note: we (ab)use the fact that the prp fields survive if no data
1234 	 * is attached to the request.
1235 	 */
1236 	c.create_cq.opcode = nvme_admin_create_cq;
1237 	c.create_cq.prp1 = cpu_to_le64(nvmeq->cq_dma_addr);
1238 	c.create_cq.cqid = cpu_to_le16(qid);
1239 	c.create_cq.qsize = cpu_to_le16(nvmeq->q_depth - 1);
1240 	c.create_cq.cq_flags = cpu_to_le16(flags);
1241 	c.create_cq.irq_vector = cpu_to_le16(vector);
1242 
1243 	return nvme_submit_sync_cmd(dev->ctrl.admin_q, &c, NULL, 0);
1244 }
1245 
1246 static int adapter_alloc_sq(struct nvme_dev *dev, u16 qid,
1247 						struct nvme_queue *nvmeq)
1248 {
1249 	struct nvme_ctrl *ctrl = &dev->ctrl;
1250 	struct nvme_command c = { };
1251 	int flags = NVME_QUEUE_PHYS_CONTIG;
1252 
1253 	/*
1254 	 * Some drives have a bug that auto-enables WRRU if MEDIUM isn't
1255 	 * set. Since URGENT priority is zeroes, it makes all queues
1256 	 * URGENT.
1257 	 */
1258 	if (ctrl->quirks & NVME_QUIRK_MEDIUM_PRIO_SQ)
1259 		flags |= NVME_SQ_PRIO_MEDIUM;
1260 
1261 	/*
1262 	 * Note: we (ab)use the fact that the prp fields survive if no data
1263 	 * is attached to the request.
1264 	 */
1265 	c.create_sq.opcode = nvme_admin_create_sq;
1266 	c.create_sq.prp1 = cpu_to_le64(nvmeq->sq_dma_addr);
1267 	c.create_sq.sqid = cpu_to_le16(qid);
1268 	c.create_sq.qsize = cpu_to_le16(nvmeq->q_depth - 1);
1269 	c.create_sq.sq_flags = cpu_to_le16(flags);
1270 	c.create_sq.cqid = cpu_to_le16(qid);
1271 
1272 	return nvme_submit_sync_cmd(dev->ctrl.admin_q, &c, NULL, 0);
1273 }
1274 
1275 static int adapter_delete_cq(struct nvme_dev *dev, u16 cqid)
1276 {
1277 	return adapter_delete_queue(dev, nvme_admin_delete_cq, cqid);
1278 }
1279 
1280 static int adapter_delete_sq(struct nvme_dev *dev, u16 sqid)
1281 {
1282 	return adapter_delete_queue(dev, nvme_admin_delete_sq, sqid);
1283 }
1284 
1285 static void abort_endio(struct request *req, blk_status_t error)
1286 {
1287 	struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
1288 	struct nvme_queue *nvmeq = iod->nvmeq;
1289 
1290 	dev_warn(nvmeq->dev->ctrl.device,
1291 		 "Abort status: 0x%x", nvme_req(req)->status);
1292 	atomic_inc(&nvmeq->dev->ctrl.abort_limit);
1293 	blk_mq_free_request(req);
1294 }
1295 
1296 static bool nvme_should_reset(struct nvme_dev *dev, u32 csts)
1297 {
1298 	/* If true, indicates loss of adapter communication, possibly by a
1299 	 * NVMe Subsystem reset.
1300 	 */
1301 	bool nssro = dev->subsystem && (csts & NVME_CSTS_NSSRO);
1302 
1303 	/* If there is a reset/reinit ongoing, we shouldn't reset again. */
1304 	switch (dev->ctrl.state) {
1305 	case NVME_CTRL_RESETTING:
1306 	case NVME_CTRL_CONNECTING:
1307 		return false;
1308 	default:
1309 		break;
1310 	}
1311 
1312 	/* We shouldn't reset unless the controller is on fatal error state
1313 	 * _or_ if we lost the communication with it.
1314 	 */
1315 	if (!(csts & NVME_CSTS_CFS) && !nssro)
1316 		return false;
1317 
1318 	return true;
1319 }
1320 
1321 static void nvme_warn_reset(struct nvme_dev *dev, u32 csts)
1322 {
1323 	/* Read a config register to help see what died. */
1324 	u16 pci_status;
1325 	int result;
1326 
1327 	result = pci_read_config_word(to_pci_dev(dev->dev), PCI_STATUS,
1328 				      &pci_status);
1329 	if (result == PCIBIOS_SUCCESSFUL)
1330 		dev_warn(dev->ctrl.device,
1331 			 "controller is down; will reset: CSTS=0x%x, PCI_STATUS=0x%hx\n",
1332 			 csts, pci_status);
1333 	else
1334 		dev_warn(dev->ctrl.device,
1335 			 "controller is down; will reset: CSTS=0x%x, PCI_STATUS read failed (%d)\n",
1336 			 csts, result);
1337 
1338 	if (csts != ~0)
1339 		return;
1340 
1341 	dev_warn(dev->ctrl.device,
1342 		 "Does your device have a faulty power saving mode enabled?\n");
1343 	dev_warn(dev->ctrl.device,
1344 		 "Try \"nvme_core.default_ps_max_latency_us=0 pcie_aspm=off\" and report a bug\n");
1345 }
1346 
1347 static enum blk_eh_timer_return nvme_timeout(struct request *req, bool reserved)
1348 {
1349 	struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
1350 	struct nvme_queue *nvmeq = iod->nvmeq;
1351 	struct nvme_dev *dev = nvmeq->dev;
1352 	struct request *abort_req;
1353 	struct nvme_command cmd = { };
1354 	u32 csts = readl(dev->bar + NVME_REG_CSTS);
1355 
1356 	/* If PCI error recovery process is happening, we cannot reset or
1357 	 * the recovery mechanism will surely fail.
1358 	 */
1359 	mb();
1360 	if (pci_channel_offline(to_pci_dev(dev->dev)))
1361 		return BLK_EH_RESET_TIMER;
1362 
1363 	/*
1364 	 * Reset immediately if the controller is failed
1365 	 */
1366 	if (nvme_should_reset(dev, csts)) {
1367 		nvme_warn_reset(dev, csts);
1368 		nvme_dev_disable(dev, false);
1369 		nvme_reset_ctrl(&dev->ctrl);
1370 		return BLK_EH_DONE;
1371 	}
1372 
1373 	/*
1374 	 * Did we miss an interrupt?
1375 	 */
1376 	if (test_bit(NVMEQ_POLLED, &nvmeq->flags))
1377 		nvme_poll(req->mq_hctx, NULL);
1378 	else
1379 		nvme_poll_irqdisable(nvmeq);
1380 
1381 	if (blk_mq_request_completed(req)) {
1382 		dev_warn(dev->ctrl.device,
1383 			 "I/O %d QID %d timeout, completion polled\n",
1384 			 req->tag, nvmeq->qid);
1385 		return BLK_EH_DONE;
1386 	}
1387 
1388 	/*
1389 	 * Shutdown immediately if controller times out while starting. The
1390 	 * reset work will see the pci device disabled when it gets the forced
1391 	 * cancellation error. All outstanding requests are completed on
1392 	 * shutdown, so we return BLK_EH_DONE.
1393 	 */
1394 	switch (dev->ctrl.state) {
1395 	case NVME_CTRL_CONNECTING:
1396 		nvme_change_ctrl_state(&dev->ctrl, NVME_CTRL_DELETING);
1397 		fallthrough;
1398 	case NVME_CTRL_DELETING:
1399 		dev_warn_ratelimited(dev->ctrl.device,
1400 			 "I/O %d QID %d timeout, disable controller\n",
1401 			 req->tag, nvmeq->qid);
1402 		nvme_req(req)->flags |= NVME_REQ_CANCELLED;
1403 		nvme_dev_disable(dev, true);
1404 		return BLK_EH_DONE;
1405 	case NVME_CTRL_RESETTING:
1406 		return BLK_EH_RESET_TIMER;
1407 	default:
1408 		break;
1409 	}
1410 
1411 	/*
1412 	 * Shutdown the controller immediately and schedule a reset if the
1413 	 * command was already aborted once before and still hasn't been
1414 	 * returned to the driver, or if this is the admin queue.
1415 	 */
1416 	if (!nvmeq->qid || iod->aborted) {
1417 		dev_warn(dev->ctrl.device,
1418 			 "I/O %d QID %d timeout, reset controller\n",
1419 			 req->tag, nvmeq->qid);
1420 		nvme_req(req)->flags |= NVME_REQ_CANCELLED;
1421 		nvme_dev_disable(dev, false);
1422 		nvme_reset_ctrl(&dev->ctrl);
1423 
1424 		return BLK_EH_DONE;
1425 	}
1426 
1427 	if (atomic_dec_return(&dev->ctrl.abort_limit) < 0) {
1428 		atomic_inc(&dev->ctrl.abort_limit);
1429 		return BLK_EH_RESET_TIMER;
1430 	}
1431 	iod->aborted = 1;
1432 
1433 	cmd.abort.opcode = nvme_admin_abort_cmd;
1434 	cmd.abort.cid = nvme_cid(req);
1435 	cmd.abort.sqid = cpu_to_le16(nvmeq->qid);
1436 
1437 	dev_warn(nvmeq->dev->ctrl.device,
1438 		"I/O %d QID %d timeout, aborting\n",
1439 		 req->tag, nvmeq->qid);
1440 
1441 	abort_req = blk_mq_alloc_request(dev->ctrl.admin_q, nvme_req_op(&cmd),
1442 					 BLK_MQ_REQ_NOWAIT);
1443 	if (IS_ERR(abort_req)) {
1444 		atomic_inc(&dev->ctrl.abort_limit);
1445 		return BLK_EH_RESET_TIMER;
1446 	}
1447 	nvme_init_request(abort_req, &cmd);
1448 
1449 	abort_req->end_io = abort_endio;
1450 	abort_req->end_io_data = NULL;
1451 	abort_req->rq_flags |= RQF_QUIET;
1452 	blk_execute_rq_nowait(abort_req, false);
1453 
1454 	/*
1455 	 * The aborted req will be completed on receiving the abort req.
1456 	 * We enable the timer again. If hit twice, it'll cause a device reset,
1457 	 * as the device then is in a faulty state.
1458 	 */
1459 	return BLK_EH_RESET_TIMER;
1460 }
1461 
1462 static void nvme_free_queue(struct nvme_queue *nvmeq)
1463 {
1464 	dma_free_coherent(nvmeq->dev->dev, CQ_SIZE(nvmeq),
1465 				(void *)nvmeq->cqes, nvmeq->cq_dma_addr);
1466 	if (!nvmeq->sq_cmds)
1467 		return;
1468 
1469 	if (test_and_clear_bit(NVMEQ_SQ_CMB, &nvmeq->flags)) {
1470 		pci_free_p2pmem(to_pci_dev(nvmeq->dev->dev),
1471 				nvmeq->sq_cmds, SQ_SIZE(nvmeq));
1472 	} else {
1473 		dma_free_coherent(nvmeq->dev->dev, SQ_SIZE(nvmeq),
1474 				nvmeq->sq_cmds, nvmeq->sq_dma_addr);
1475 	}
1476 }
1477 
1478 static void nvme_free_queues(struct nvme_dev *dev, int lowest)
1479 {
1480 	int i;
1481 
1482 	for (i = dev->ctrl.queue_count - 1; i >= lowest; i--) {
1483 		dev->ctrl.queue_count--;
1484 		nvme_free_queue(&dev->queues[i]);
1485 	}
1486 }
1487 
1488 /**
1489  * nvme_suspend_queue - put queue into suspended state
1490  * @nvmeq: queue to suspend
1491  */
1492 static int nvme_suspend_queue(struct nvme_queue *nvmeq)
1493 {
1494 	if (!test_and_clear_bit(NVMEQ_ENABLED, &nvmeq->flags))
1495 		return 1;
1496 
1497 	/* ensure that nvme_queue_rq() sees NVMEQ_ENABLED cleared */
1498 	mb();
1499 
1500 	nvmeq->dev->online_queues--;
1501 	if (!nvmeq->qid && nvmeq->dev->ctrl.admin_q)
1502 		nvme_stop_admin_queue(&nvmeq->dev->ctrl);
1503 	if (!test_and_clear_bit(NVMEQ_POLLED, &nvmeq->flags))
1504 		pci_free_irq(to_pci_dev(nvmeq->dev->dev), nvmeq->cq_vector, nvmeq);
1505 	return 0;
1506 }
1507 
1508 static void nvme_suspend_io_queues(struct nvme_dev *dev)
1509 {
1510 	int i;
1511 
1512 	for (i = dev->ctrl.queue_count - 1; i > 0; i--)
1513 		nvme_suspend_queue(&dev->queues[i]);
1514 }
1515 
1516 static void nvme_disable_admin_queue(struct nvme_dev *dev, bool shutdown)
1517 {
1518 	struct nvme_queue *nvmeq = &dev->queues[0];
1519 
1520 	if (shutdown)
1521 		nvme_shutdown_ctrl(&dev->ctrl);
1522 	else
1523 		nvme_disable_ctrl(&dev->ctrl);
1524 
1525 	nvme_poll_irqdisable(nvmeq);
1526 }
1527 
1528 /*
1529  * Called only on a device that has been disabled and after all other threads
1530  * that can check this device's completion queues have synced, except
1531  * nvme_poll(). This is the last chance for the driver to see a natural
1532  * completion before nvme_cancel_request() terminates all incomplete requests.
1533  */
1534 static void nvme_reap_pending_cqes(struct nvme_dev *dev)
1535 {
1536 	int i;
1537 
1538 	for (i = dev->ctrl.queue_count - 1; i > 0; i--) {
1539 		spin_lock(&dev->queues[i].cq_poll_lock);
1540 		nvme_poll_cq(&dev->queues[i], NULL);
1541 		spin_unlock(&dev->queues[i].cq_poll_lock);
1542 	}
1543 }
1544 
1545 static int nvme_cmb_qdepth(struct nvme_dev *dev, int nr_io_queues,
1546 				int entry_size)
1547 {
1548 	int q_depth = dev->q_depth;
1549 	unsigned q_size_aligned = roundup(q_depth * entry_size,
1550 					  NVME_CTRL_PAGE_SIZE);
1551 
1552 	if (q_size_aligned * nr_io_queues > dev->cmb_size) {
1553 		u64 mem_per_q = div_u64(dev->cmb_size, nr_io_queues);
1554 
1555 		mem_per_q = round_down(mem_per_q, NVME_CTRL_PAGE_SIZE);
1556 		q_depth = div_u64(mem_per_q, entry_size);
1557 
1558 		/*
1559 		 * Ensure the reduced q_depth is above some threshold where it
1560 		 * would be better to map queues in system memory with the
1561 		 * original depth
1562 		 */
1563 		if (q_depth < 64)
1564 			return -ENOMEM;
1565 	}
1566 
1567 	return q_depth;
1568 }
1569 
1570 static int nvme_alloc_sq_cmds(struct nvme_dev *dev, struct nvme_queue *nvmeq,
1571 				int qid)
1572 {
1573 	struct pci_dev *pdev = to_pci_dev(dev->dev);
1574 
1575 	if (qid && dev->cmb_use_sqes && (dev->cmbsz & NVME_CMBSZ_SQS)) {
1576 		nvmeq->sq_cmds = pci_alloc_p2pmem(pdev, SQ_SIZE(nvmeq));
1577 		if (nvmeq->sq_cmds) {
1578 			nvmeq->sq_dma_addr = pci_p2pmem_virt_to_bus(pdev,
1579 							nvmeq->sq_cmds);
1580 			if (nvmeq->sq_dma_addr) {
1581 				set_bit(NVMEQ_SQ_CMB, &nvmeq->flags);
1582 				return 0;
1583 			}
1584 
1585 			pci_free_p2pmem(pdev, nvmeq->sq_cmds, SQ_SIZE(nvmeq));
1586 		}
1587 	}
1588 
1589 	nvmeq->sq_cmds = dma_alloc_coherent(dev->dev, SQ_SIZE(nvmeq),
1590 				&nvmeq->sq_dma_addr, GFP_KERNEL);
1591 	if (!nvmeq->sq_cmds)
1592 		return -ENOMEM;
1593 	return 0;
1594 }
1595 
1596 static int nvme_alloc_queue(struct nvme_dev *dev, int qid, int depth)
1597 {
1598 	struct nvme_queue *nvmeq = &dev->queues[qid];
1599 
1600 	if (dev->ctrl.queue_count > qid)
1601 		return 0;
1602 
1603 	nvmeq->sqes = qid ? dev->io_sqes : NVME_ADM_SQES;
1604 	nvmeq->q_depth = depth;
1605 	nvmeq->cqes = dma_alloc_coherent(dev->dev, CQ_SIZE(nvmeq),
1606 					 &nvmeq->cq_dma_addr, GFP_KERNEL);
1607 	if (!nvmeq->cqes)
1608 		goto free_nvmeq;
1609 
1610 	if (nvme_alloc_sq_cmds(dev, nvmeq, qid))
1611 		goto free_cqdma;
1612 
1613 	nvmeq->dev = dev;
1614 	spin_lock_init(&nvmeq->sq_lock);
1615 	spin_lock_init(&nvmeq->cq_poll_lock);
1616 	nvmeq->cq_head = 0;
1617 	nvmeq->cq_phase = 1;
1618 	nvmeq->q_db = &dev->dbs[qid * 2 * dev->db_stride];
1619 	nvmeq->qid = qid;
1620 	dev->ctrl.queue_count++;
1621 
1622 	return 0;
1623 
1624  free_cqdma:
1625 	dma_free_coherent(dev->dev, CQ_SIZE(nvmeq), (void *)nvmeq->cqes,
1626 			  nvmeq->cq_dma_addr);
1627  free_nvmeq:
1628 	return -ENOMEM;
1629 }
1630 
1631 static int queue_request_irq(struct nvme_queue *nvmeq)
1632 {
1633 	struct pci_dev *pdev = to_pci_dev(nvmeq->dev->dev);
1634 	int nr = nvmeq->dev->ctrl.instance;
1635 
1636 	if (use_threaded_interrupts) {
1637 		return pci_request_irq(pdev, nvmeq->cq_vector, nvme_irq_check,
1638 				nvme_irq, nvmeq, "nvme%dq%d", nr, nvmeq->qid);
1639 	} else {
1640 		return pci_request_irq(pdev, nvmeq->cq_vector, nvme_irq,
1641 				NULL, nvmeq, "nvme%dq%d", nr, nvmeq->qid);
1642 	}
1643 }
1644 
1645 static void nvme_init_queue(struct nvme_queue *nvmeq, u16 qid)
1646 {
1647 	struct nvme_dev *dev = nvmeq->dev;
1648 
1649 	nvmeq->sq_tail = 0;
1650 	nvmeq->last_sq_tail = 0;
1651 	nvmeq->cq_head = 0;
1652 	nvmeq->cq_phase = 1;
1653 	nvmeq->q_db = &dev->dbs[qid * 2 * dev->db_stride];
1654 	memset((void *)nvmeq->cqes, 0, CQ_SIZE(nvmeq));
1655 	nvme_dbbuf_init(dev, nvmeq, qid);
1656 	dev->online_queues++;
1657 	wmb(); /* ensure the first interrupt sees the initialization */
1658 }
1659 
1660 /*
1661  * Try getting shutdown_lock while setting up IO queues.
1662  */
1663 static int nvme_setup_io_queues_trylock(struct nvme_dev *dev)
1664 {
1665 	/*
1666 	 * Give up if the lock is being held by nvme_dev_disable.
1667 	 */
1668 	if (!mutex_trylock(&dev->shutdown_lock))
1669 		return -ENODEV;
1670 
1671 	/*
1672 	 * Controller is in wrong state, fail early.
1673 	 */
1674 	if (dev->ctrl.state != NVME_CTRL_CONNECTING) {
1675 		mutex_unlock(&dev->shutdown_lock);
1676 		return -ENODEV;
1677 	}
1678 
1679 	return 0;
1680 }
1681 
1682 static int nvme_create_queue(struct nvme_queue *nvmeq, int qid, bool polled)
1683 {
1684 	struct nvme_dev *dev = nvmeq->dev;
1685 	int result;
1686 	u16 vector = 0;
1687 
1688 	clear_bit(NVMEQ_DELETE_ERROR, &nvmeq->flags);
1689 
1690 	/*
1691 	 * A queue's vector matches the queue identifier unless the controller
1692 	 * has only one vector available.
1693 	 */
1694 	if (!polled)
1695 		vector = dev->num_vecs == 1 ? 0 : qid;
1696 	else
1697 		set_bit(NVMEQ_POLLED, &nvmeq->flags);
1698 
1699 	result = adapter_alloc_cq(dev, qid, nvmeq, vector);
1700 	if (result)
1701 		return result;
1702 
1703 	result = adapter_alloc_sq(dev, qid, nvmeq);
1704 	if (result < 0)
1705 		return result;
1706 	if (result)
1707 		goto release_cq;
1708 
1709 	nvmeq->cq_vector = vector;
1710 
1711 	result = nvme_setup_io_queues_trylock(dev);
1712 	if (result)
1713 		return result;
1714 	nvme_init_queue(nvmeq, qid);
1715 	if (!polled) {
1716 		result = queue_request_irq(nvmeq);
1717 		if (result < 0)
1718 			goto release_sq;
1719 	}
1720 
1721 	set_bit(NVMEQ_ENABLED, &nvmeq->flags);
1722 	mutex_unlock(&dev->shutdown_lock);
1723 	return result;
1724 
1725 release_sq:
1726 	dev->online_queues--;
1727 	mutex_unlock(&dev->shutdown_lock);
1728 	adapter_delete_sq(dev, qid);
1729 release_cq:
1730 	adapter_delete_cq(dev, qid);
1731 	return result;
1732 }
1733 
1734 static const struct blk_mq_ops nvme_mq_admin_ops = {
1735 	.queue_rq	= nvme_queue_rq,
1736 	.complete	= nvme_pci_complete_rq,
1737 	.init_hctx	= nvme_admin_init_hctx,
1738 	.init_request	= nvme_pci_init_request,
1739 	.timeout	= nvme_timeout,
1740 };
1741 
1742 static const struct blk_mq_ops nvme_mq_ops = {
1743 	.queue_rq	= nvme_queue_rq,
1744 	.queue_rqs	= nvme_queue_rqs,
1745 	.complete	= nvme_pci_complete_rq,
1746 	.commit_rqs	= nvme_commit_rqs,
1747 	.init_hctx	= nvme_init_hctx,
1748 	.init_request	= nvme_pci_init_request,
1749 	.map_queues	= nvme_pci_map_queues,
1750 	.timeout	= nvme_timeout,
1751 	.poll		= nvme_poll,
1752 };
1753 
1754 static void nvme_dev_remove_admin(struct nvme_dev *dev)
1755 {
1756 	if (dev->ctrl.admin_q && !blk_queue_dying(dev->ctrl.admin_q)) {
1757 		/*
1758 		 * If the controller was reset during removal, it's possible
1759 		 * user requests may be waiting on a stopped queue. Start the
1760 		 * queue to flush these to completion.
1761 		 */
1762 		nvme_start_admin_queue(&dev->ctrl);
1763 		blk_cleanup_queue(dev->ctrl.admin_q);
1764 		blk_mq_free_tag_set(&dev->admin_tagset);
1765 	}
1766 }
1767 
1768 static int nvme_alloc_admin_tags(struct nvme_dev *dev)
1769 {
1770 	if (!dev->ctrl.admin_q) {
1771 		dev->admin_tagset.ops = &nvme_mq_admin_ops;
1772 		dev->admin_tagset.nr_hw_queues = 1;
1773 
1774 		dev->admin_tagset.queue_depth = NVME_AQ_MQ_TAG_DEPTH;
1775 		dev->admin_tagset.timeout = NVME_ADMIN_TIMEOUT;
1776 		dev->admin_tagset.numa_node = dev->ctrl.numa_node;
1777 		dev->admin_tagset.cmd_size = sizeof(struct nvme_iod);
1778 		dev->admin_tagset.flags = BLK_MQ_F_NO_SCHED;
1779 		dev->admin_tagset.driver_data = dev;
1780 
1781 		if (blk_mq_alloc_tag_set(&dev->admin_tagset))
1782 			return -ENOMEM;
1783 		dev->ctrl.admin_tagset = &dev->admin_tagset;
1784 
1785 		dev->ctrl.admin_q = blk_mq_init_queue(&dev->admin_tagset);
1786 		if (IS_ERR(dev->ctrl.admin_q)) {
1787 			blk_mq_free_tag_set(&dev->admin_tagset);
1788 			dev->ctrl.admin_q = NULL;
1789 			return -ENOMEM;
1790 		}
1791 		if (!blk_get_queue(dev->ctrl.admin_q)) {
1792 			nvme_dev_remove_admin(dev);
1793 			dev->ctrl.admin_q = NULL;
1794 			return -ENODEV;
1795 		}
1796 	} else
1797 		nvme_start_admin_queue(&dev->ctrl);
1798 
1799 	return 0;
1800 }
1801 
1802 static unsigned long db_bar_size(struct nvme_dev *dev, unsigned nr_io_queues)
1803 {
1804 	return NVME_REG_DBS + ((nr_io_queues + 1) * 8 * dev->db_stride);
1805 }
1806 
1807 static int nvme_remap_bar(struct nvme_dev *dev, unsigned long size)
1808 {
1809 	struct pci_dev *pdev = to_pci_dev(dev->dev);
1810 
1811 	if (size <= dev->bar_mapped_size)
1812 		return 0;
1813 	if (size > pci_resource_len(pdev, 0))
1814 		return -ENOMEM;
1815 	if (dev->bar)
1816 		iounmap(dev->bar);
1817 	dev->bar = ioremap(pci_resource_start(pdev, 0), size);
1818 	if (!dev->bar) {
1819 		dev->bar_mapped_size = 0;
1820 		return -ENOMEM;
1821 	}
1822 	dev->bar_mapped_size = size;
1823 	dev->dbs = dev->bar + NVME_REG_DBS;
1824 
1825 	return 0;
1826 }
1827 
1828 static int nvme_pci_configure_admin_queue(struct nvme_dev *dev)
1829 {
1830 	int result;
1831 	u32 aqa;
1832 	struct nvme_queue *nvmeq;
1833 
1834 	result = nvme_remap_bar(dev, db_bar_size(dev, 0));
1835 	if (result < 0)
1836 		return result;
1837 
1838 	dev->subsystem = readl(dev->bar + NVME_REG_VS) >= NVME_VS(1, 1, 0) ?
1839 				NVME_CAP_NSSRC(dev->ctrl.cap) : 0;
1840 
1841 	if (dev->subsystem &&
1842 	    (readl(dev->bar + NVME_REG_CSTS) & NVME_CSTS_NSSRO))
1843 		writel(NVME_CSTS_NSSRO, dev->bar + NVME_REG_CSTS);
1844 
1845 	result = nvme_disable_ctrl(&dev->ctrl);
1846 	if (result < 0)
1847 		return result;
1848 
1849 	result = nvme_alloc_queue(dev, 0, NVME_AQ_DEPTH);
1850 	if (result)
1851 		return result;
1852 
1853 	dev->ctrl.numa_node = dev_to_node(dev->dev);
1854 
1855 	nvmeq = &dev->queues[0];
1856 	aqa = nvmeq->q_depth - 1;
1857 	aqa |= aqa << 16;
1858 
1859 	writel(aqa, dev->bar + NVME_REG_AQA);
1860 	lo_hi_writeq(nvmeq->sq_dma_addr, dev->bar + NVME_REG_ASQ);
1861 	lo_hi_writeq(nvmeq->cq_dma_addr, dev->bar + NVME_REG_ACQ);
1862 
1863 	result = nvme_enable_ctrl(&dev->ctrl);
1864 	if (result)
1865 		return result;
1866 
1867 	nvmeq->cq_vector = 0;
1868 	nvme_init_queue(nvmeq, 0);
1869 	result = queue_request_irq(nvmeq);
1870 	if (result) {
1871 		dev->online_queues--;
1872 		return result;
1873 	}
1874 
1875 	set_bit(NVMEQ_ENABLED, &nvmeq->flags);
1876 	return result;
1877 }
1878 
1879 static int nvme_create_io_queues(struct nvme_dev *dev)
1880 {
1881 	unsigned i, max, rw_queues;
1882 	int ret = 0;
1883 
1884 	for (i = dev->ctrl.queue_count; i <= dev->max_qid; i++) {
1885 		if (nvme_alloc_queue(dev, i, dev->q_depth)) {
1886 			ret = -ENOMEM;
1887 			break;
1888 		}
1889 	}
1890 
1891 	max = min(dev->max_qid, dev->ctrl.queue_count - 1);
1892 	if (max != 1 && dev->io_queues[HCTX_TYPE_POLL]) {
1893 		rw_queues = dev->io_queues[HCTX_TYPE_DEFAULT] +
1894 				dev->io_queues[HCTX_TYPE_READ];
1895 	} else {
1896 		rw_queues = max;
1897 	}
1898 
1899 	for (i = dev->online_queues; i <= max; i++) {
1900 		bool polled = i > rw_queues;
1901 
1902 		ret = nvme_create_queue(&dev->queues[i], i, polled);
1903 		if (ret)
1904 			break;
1905 	}
1906 
1907 	/*
1908 	 * Ignore failing Create SQ/CQ commands, we can continue with less
1909 	 * than the desired amount of queues, and even a controller without
1910 	 * I/O queues can still be used to issue admin commands.  This might
1911 	 * be useful to upgrade a buggy firmware for example.
1912 	 */
1913 	return ret >= 0 ? 0 : ret;
1914 }
1915 
1916 static u64 nvme_cmb_size_unit(struct nvme_dev *dev)
1917 {
1918 	u8 szu = (dev->cmbsz >> NVME_CMBSZ_SZU_SHIFT) & NVME_CMBSZ_SZU_MASK;
1919 
1920 	return 1ULL << (12 + 4 * szu);
1921 }
1922 
1923 static u32 nvme_cmb_size(struct nvme_dev *dev)
1924 {
1925 	return (dev->cmbsz >> NVME_CMBSZ_SZ_SHIFT) & NVME_CMBSZ_SZ_MASK;
1926 }
1927 
1928 static void nvme_map_cmb(struct nvme_dev *dev)
1929 {
1930 	u64 size, offset;
1931 	resource_size_t bar_size;
1932 	struct pci_dev *pdev = to_pci_dev(dev->dev);
1933 	int bar;
1934 
1935 	if (dev->cmb_size)
1936 		return;
1937 
1938 	if (NVME_CAP_CMBS(dev->ctrl.cap))
1939 		writel(NVME_CMBMSC_CRE, dev->bar + NVME_REG_CMBMSC);
1940 
1941 	dev->cmbsz = readl(dev->bar + NVME_REG_CMBSZ);
1942 	if (!dev->cmbsz)
1943 		return;
1944 	dev->cmbloc = readl(dev->bar + NVME_REG_CMBLOC);
1945 
1946 	size = nvme_cmb_size_unit(dev) * nvme_cmb_size(dev);
1947 	offset = nvme_cmb_size_unit(dev) * NVME_CMB_OFST(dev->cmbloc);
1948 	bar = NVME_CMB_BIR(dev->cmbloc);
1949 	bar_size = pci_resource_len(pdev, bar);
1950 
1951 	if (offset > bar_size)
1952 		return;
1953 
1954 	/*
1955 	 * Tell the controller about the host side address mapping the CMB,
1956 	 * and enable CMB decoding for the NVMe 1.4+ scheme:
1957 	 */
1958 	if (NVME_CAP_CMBS(dev->ctrl.cap)) {
1959 		hi_lo_writeq(NVME_CMBMSC_CRE | NVME_CMBMSC_CMSE |
1960 			     (pci_bus_address(pdev, bar) + offset),
1961 			     dev->bar + NVME_REG_CMBMSC);
1962 	}
1963 
1964 	/*
1965 	 * Controllers may support a CMB size larger than their BAR,
1966 	 * for example, due to being behind a bridge. Reduce the CMB to
1967 	 * the reported size of the BAR
1968 	 */
1969 	if (size > bar_size - offset)
1970 		size = bar_size - offset;
1971 
1972 	if (pci_p2pdma_add_resource(pdev, bar, size, offset)) {
1973 		dev_warn(dev->ctrl.device,
1974 			 "failed to register the CMB\n");
1975 		return;
1976 	}
1977 
1978 	dev->cmb_size = size;
1979 	dev->cmb_use_sqes = use_cmb_sqes && (dev->cmbsz & NVME_CMBSZ_SQS);
1980 
1981 	if ((dev->cmbsz & (NVME_CMBSZ_WDS | NVME_CMBSZ_RDS)) ==
1982 			(NVME_CMBSZ_WDS | NVME_CMBSZ_RDS))
1983 		pci_p2pmem_publish(pdev, true);
1984 }
1985 
1986 static int nvme_set_host_mem(struct nvme_dev *dev, u32 bits)
1987 {
1988 	u32 host_mem_size = dev->host_mem_size >> NVME_CTRL_PAGE_SHIFT;
1989 	u64 dma_addr = dev->host_mem_descs_dma;
1990 	struct nvme_command c = { };
1991 	int ret;
1992 
1993 	c.features.opcode	= nvme_admin_set_features;
1994 	c.features.fid		= cpu_to_le32(NVME_FEAT_HOST_MEM_BUF);
1995 	c.features.dword11	= cpu_to_le32(bits);
1996 	c.features.dword12	= cpu_to_le32(host_mem_size);
1997 	c.features.dword13	= cpu_to_le32(lower_32_bits(dma_addr));
1998 	c.features.dword14	= cpu_to_le32(upper_32_bits(dma_addr));
1999 	c.features.dword15	= cpu_to_le32(dev->nr_host_mem_descs);
2000 
2001 	ret = nvme_submit_sync_cmd(dev->ctrl.admin_q, &c, NULL, 0);
2002 	if (ret) {
2003 		dev_warn(dev->ctrl.device,
2004 			 "failed to set host mem (err %d, flags %#x).\n",
2005 			 ret, bits);
2006 	} else
2007 		dev->hmb = bits & NVME_HOST_MEM_ENABLE;
2008 
2009 	return ret;
2010 }
2011 
2012 static void nvme_free_host_mem(struct nvme_dev *dev)
2013 {
2014 	int i;
2015 
2016 	for (i = 0; i < dev->nr_host_mem_descs; i++) {
2017 		struct nvme_host_mem_buf_desc *desc = &dev->host_mem_descs[i];
2018 		size_t size = le32_to_cpu(desc->size) * NVME_CTRL_PAGE_SIZE;
2019 
2020 		dma_free_attrs(dev->dev, size, dev->host_mem_desc_bufs[i],
2021 			       le64_to_cpu(desc->addr),
2022 			       DMA_ATTR_NO_KERNEL_MAPPING | DMA_ATTR_NO_WARN);
2023 	}
2024 
2025 	kfree(dev->host_mem_desc_bufs);
2026 	dev->host_mem_desc_bufs = NULL;
2027 	dma_free_coherent(dev->dev,
2028 			dev->nr_host_mem_descs * sizeof(*dev->host_mem_descs),
2029 			dev->host_mem_descs, dev->host_mem_descs_dma);
2030 	dev->host_mem_descs = NULL;
2031 	dev->nr_host_mem_descs = 0;
2032 }
2033 
2034 static int __nvme_alloc_host_mem(struct nvme_dev *dev, u64 preferred,
2035 		u32 chunk_size)
2036 {
2037 	struct nvme_host_mem_buf_desc *descs;
2038 	u32 max_entries, len;
2039 	dma_addr_t descs_dma;
2040 	int i = 0;
2041 	void **bufs;
2042 	u64 size, tmp;
2043 
2044 	tmp = (preferred + chunk_size - 1);
2045 	do_div(tmp, chunk_size);
2046 	max_entries = tmp;
2047 
2048 	if (dev->ctrl.hmmaxd && dev->ctrl.hmmaxd < max_entries)
2049 		max_entries = dev->ctrl.hmmaxd;
2050 
2051 	descs = dma_alloc_coherent(dev->dev, max_entries * sizeof(*descs),
2052 				   &descs_dma, GFP_KERNEL);
2053 	if (!descs)
2054 		goto out;
2055 
2056 	bufs = kcalloc(max_entries, sizeof(*bufs), GFP_KERNEL);
2057 	if (!bufs)
2058 		goto out_free_descs;
2059 
2060 	for (size = 0; size < preferred && i < max_entries; size += len) {
2061 		dma_addr_t dma_addr;
2062 
2063 		len = min_t(u64, chunk_size, preferred - size);
2064 		bufs[i] = dma_alloc_attrs(dev->dev, len, &dma_addr, GFP_KERNEL,
2065 				DMA_ATTR_NO_KERNEL_MAPPING | DMA_ATTR_NO_WARN);
2066 		if (!bufs[i])
2067 			break;
2068 
2069 		descs[i].addr = cpu_to_le64(dma_addr);
2070 		descs[i].size = cpu_to_le32(len / NVME_CTRL_PAGE_SIZE);
2071 		i++;
2072 	}
2073 
2074 	if (!size)
2075 		goto out_free_bufs;
2076 
2077 	dev->nr_host_mem_descs = i;
2078 	dev->host_mem_size = size;
2079 	dev->host_mem_descs = descs;
2080 	dev->host_mem_descs_dma = descs_dma;
2081 	dev->host_mem_desc_bufs = bufs;
2082 	return 0;
2083 
2084 out_free_bufs:
2085 	while (--i >= 0) {
2086 		size_t size = le32_to_cpu(descs[i].size) * NVME_CTRL_PAGE_SIZE;
2087 
2088 		dma_free_attrs(dev->dev, size, bufs[i],
2089 			       le64_to_cpu(descs[i].addr),
2090 			       DMA_ATTR_NO_KERNEL_MAPPING | DMA_ATTR_NO_WARN);
2091 	}
2092 
2093 	kfree(bufs);
2094 out_free_descs:
2095 	dma_free_coherent(dev->dev, max_entries * sizeof(*descs), descs,
2096 			descs_dma);
2097 out:
2098 	dev->host_mem_descs = NULL;
2099 	return -ENOMEM;
2100 }
2101 
2102 static int nvme_alloc_host_mem(struct nvme_dev *dev, u64 min, u64 preferred)
2103 {
2104 	u64 min_chunk = min_t(u64, preferred, PAGE_SIZE * MAX_ORDER_NR_PAGES);
2105 	u64 hmminds = max_t(u32, dev->ctrl.hmminds * 4096, PAGE_SIZE * 2);
2106 	u64 chunk_size;
2107 
2108 	/* start big and work our way down */
2109 	for (chunk_size = min_chunk; chunk_size >= hmminds; chunk_size /= 2) {
2110 		if (!__nvme_alloc_host_mem(dev, preferred, chunk_size)) {
2111 			if (!min || dev->host_mem_size >= min)
2112 				return 0;
2113 			nvme_free_host_mem(dev);
2114 		}
2115 	}
2116 
2117 	return -ENOMEM;
2118 }
2119 
2120 static int nvme_setup_host_mem(struct nvme_dev *dev)
2121 {
2122 	u64 max = (u64)max_host_mem_size_mb * SZ_1M;
2123 	u64 preferred = (u64)dev->ctrl.hmpre * 4096;
2124 	u64 min = (u64)dev->ctrl.hmmin * 4096;
2125 	u32 enable_bits = NVME_HOST_MEM_ENABLE;
2126 	int ret;
2127 
2128 	preferred = min(preferred, max);
2129 	if (min > max) {
2130 		dev_warn(dev->ctrl.device,
2131 			"min host memory (%lld MiB) above limit (%d MiB).\n",
2132 			min >> ilog2(SZ_1M), max_host_mem_size_mb);
2133 		nvme_free_host_mem(dev);
2134 		return 0;
2135 	}
2136 
2137 	/*
2138 	 * If we already have a buffer allocated check if we can reuse it.
2139 	 */
2140 	if (dev->host_mem_descs) {
2141 		if (dev->host_mem_size >= min)
2142 			enable_bits |= NVME_HOST_MEM_RETURN;
2143 		else
2144 			nvme_free_host_mem(dev);
2145 	}
2146 
2147 	if (!dev->host_mem_descs) {
2148 		if (nvme_alloc_host_mem(dev, min, preferred)) {
2149 			dev_warn(dev->ctrl.device,
2150 				"failed to allocate host memory buffer.\n");
2151 			return 0; /* controller must work without HMB */
2152 		}
2153 
2154 		dev_info(dev->ctrl.device,
2155 			"allocated %lld MiB host memory buffer.\n",
2156 			dev->host_mem_size >> ilog2(SZ_1M));
2157 	}
2158 
2159 	ret = nvme_set_host_mem(dev, enable_bits);
2160 	if (ret)
2161 		nvme_free_host_mem(dev);
2162 	return ret;
2163 }
2164 
2165 static ssize_t cmb_show(struct device *dev, struct device_attribute *attr,
2166 		char *buf)
2167 {
2168 	struct nvme_dev *ndev = to_nvme_dev(dev_get_drvdata(dev));
2169 
2170 	return sysfs_emit(buf, "cmbloc : x%08x\ncmbsz  : x%08x\n",
2171 		       ndev->cmbloc, ndev->cmbsz);
2172 }
2173 static DEVICE_ATTR_RO(cmb);
2174 
2175 static ssize_t cmbloc_show(struct device *dev, struct device_attribute *attr,
2176 		char *buf)
2177 {
2178 	struct nvme_dev *ndev = to_nvme_dev(dev_get_drvdata(dev));
2179 
2180 	return sysfs_emit(buf, "%u\n", ndev->cmbloc);
2181 }
2182 static DEVICE_ATTR_RO(cmbloc);
2183 
2184 static ssize_t cmbsz_show(struct device *dev, struct device_attribute *attr,
2185 		char *buf)
2186 {
2187 	struct nvme_dev *ndev = to_nvme_dev(dev_get_drvdata(dev));
2188 
2189 	return sysfs_emit(buf, "%u\n", ndev->cmbsz);
2190 }
2191 static DEVICE_ATTR_RO(cmbsz);
2192 
2193 static ssize_t hmb_show(struct device *dev, struct device_attribute *attr,
2194 			char *buf)
2195 {
2196 	struct nvme_dev *ndev = to_nvme_dev(dev_get_drvdata(dev));
2197 
2198 	return sysfs_emit(buf, "%d\n", ndev->hmb);
2199 }
2200 
2201 static ssize_t hmb_store(struct device *dev, struct device_attribute *attr,
2202 			 const char *buf, size_t count)
2203 {
2204 	struct nvme_dev *ndev = to_nvme_dev(dev_get_drvdata(dev));
2205 	bool new;
2206 	int ret;
2207 
2208 	if (strtobool(buf, &new) < 0)
2209 		return -EINVAL;
2210 
2211 	if (new == ndev->hmb)
2212 		return count;
2213 
2214 	if (new) {
2215 		ret = nvme_setup_host_mem(ndev);
2216 	} else {
2217 		ret = nvme_set_host_mem(ndev, 0);
2218 		if (!ret)
2219 			nvme_free_host_mem(ndev);
2220 	}
2221 
2222 	if (ret < 0)
2223 		return ret;
2224 
2225 	return count;
2226 }
2227 static DEVICE_ATTR_RW(hmb);
2228 
2229 static umode_t nvme_pci_attrs_are_visible(struct kobject *kobj,
2230 		struct attribute *a, int n)
2231 {
2232 	struct nvme_ctrl *ctrl =
2233 		dev_get_drvdata(container_of(kobj, struct device, kobj));
2234 	struct nvme_dev *dev = to_nvme_dev(ctrl);
2235 
2236 	if (a == &dev_attr_cmb.attr ||
2237 	    a == &dev_attr_cmbloc.attr ||
2238 	    a == &dev_attr_cmbsz.attr) {
2239 	    	if (!dev->cmbsz)
2240 			return 0;
2241 	}
2242 	if (a == &dev_attr_hmb.attr && !ctrl->hmpre)
2243 		return 0;
2244 
2245 	return a->mode;
2246 }
2247 
2248 static struct attribute *nvme_pci_attrs[] = {
2249 	&dev_attr_cmb.attr,
2250 	&dev_attr_cmbloc.attr,
2251 	&dev_attr_cmbsz.attr,
2252 	&dev_attr_hmb.attr,
2253 	NULL,
2254 };
2255 
2256 static const struct attribute_group nvme_pci_attr_group = {
2257 	.attrs		= nvme_pci_attrs,
2258 	.is_visible	= nvme_pci_attrs_are_visible,
2259 };
2260 
2261 /*
2262  * nirqs is the number of interrupts available for write and read
2263  * queues. The core already reserved an interrupt for the admin queue.
2264  */
2265 static void nvme_calc_irq_sets(struct irq_affinity *affd, unsigned int nrirqs)
2266 {
2267 	struct nvme_dev *dev = affd->priv;
2268 	unsigned int nr_read_queues, nr_write_queues = dev->nr_write_queues;
2269 
2270 	/*
2271 	 * If there is no interrupt available for queues, ensure that
2272 	 * the default queue is set to 1. The affinity set size is
2273 	 * also set to one, but the irq core ignores it for this case.
2274 	 *
2275 	 * If only one interrupt is available or 'write_queue' == 0, combine
2276 	 * write and read queues.
2277 	 *
2278 	 * If 'write_queues' > 0, ensure it leaves room for at least one read
2279 	 * queue.
2280 	 */
2281 	if (!nrirqs) {
2282 		nrirqs = 1;
2283 		nr_read_queues = 0;
2284 	} else if (nrirqs == 1 || !nr_write_queues) {
2285 		nr_read_queues = 0;
2286 	} else if (nr_write_queues >= nrirqs) {
2287 		nr_read_queues = 1;
2288 	} else {
2289 		nr_read_queues = nrirqs - nr_write_queues;
2290 	}
2291 
2292 	dev->io_queues[HCTX_TYPE_DEFAULT] = nrirqs - nr_read_queues;
2293 	affd->set_size[HCTX_TYPE_DEFAULT] = nrirqs - nr_read_queues;
2294 	dev->io_queues[HCTX_TYPE_READ] = nr_read_queues;
2295 	affd->set_size[HCTX_TYPE_READ] = nr_read_queues;
2296 	affd->nr_sets = nr_read_queues ? 2 : 1;
2297 }
2298 
2299 static int nvme_setup_irqs(struct nvme_dev *dev, unsigned int nr_io_queues)
2300 {
2301 	struct pci_dev *pdev = to_pci_dev(dev->dev);
2302 	struct irq_affinity affd = {
2303 		.pre_vectors	= 1,
2304 		.calc_sets	= nvme_calc_irq_sets,
2305 		.priv		= dev,
2306 	};
2307 	unsigned int irq_queues, poll_queues;
2308 
2309 	/*
2310 	 * Poll queues don't need interrupts, but we need at least one I/O queue
2311 	 * left over for non-polled I/O.
2312 	 */
2313 	poll_queues = min(dev->nr_poll_queues, nr_io_queues - 1);
2314 	dev->io_queues[HCTX_TYPE_POLL] = poll_queues;
2315 
2316 	/*
2317 	 * Initialize for the single interrupt case, will be updated in
2318 	 * nvme_calc_irq_sets().
2319 	 */
2320 	dev->io_queues[HCTX_TYPE_DEFAULT] = 1;
2321 	dev->io_queues[HCTX_TYPE_READ] = 0;
2322 
2323 	/*
2324 	 * We need interrupts for the admin queue and each non-polled I/O queue,
2325 	 * but some Apple controllers require all queues to use the first
2326 	 * vector.
2327 	 */
2328 	irq_queues = 1;
2329 	if (!(dev->ctrl.quirks & NVME_QUIRK_SINGLE_VECTOR))
2330 		irq_queues += (nr_io_queues - poll_queues);
2331 	return pci_alloc_irq_vectors_affinity(pdev, 1, irq_queues,
2332 			      PCI_IRQ_ALL_TYPES | PCI_IRQ_AFFINITY, &affd);
2333 }
2334 
2335 static void nvme_disable_io_queues(struct nvme_dev *dev)
2336 {
2337 	if (__nvme_disable_io_queues(dev, nvme_admin_delete_sq))
2338 		__nvme_disable_io_queues(dev, nvme_admin_delete_cq);
2339 }
2340 
2341 static unsigned int nvme_max_io_queues(struct nvme_dev *dev)
2342 {
2343 	/*
2344 	 * If tags are shared with admin queue (Apple bug), then
2345 	 * make sure we only use one IO queue.
2346 	 */
2347 	if (dev->ctrl.quirks & NVME_QUIRK_SHARED_TAGS)
2348 		return 1;
2349 	return num_possible_cpus() + dev->nr_write_queues + dev->nr_poll_queues;
2350 }
2351 
2352 static int nvme_setup_io_queues(struct nvme_dev *dev)
2353 {
2354 	struct nvme_queue *adminq = &dev->queues[0];
2355 	struct pci_dev *pdev = to_pci_dev(dev->dev);
2356 	unsigned int nr_io_queues;
2357 	unsigned long size;
2358 	int result;
2359 
2360 	/*
2361 	 * Sample the module parameters once at reset time so that we have
2362 	 * stable values to work with.
2363 	 */
2364 	dev->nr_write_queues = write_queues;
2365 	dev->nr_poll_queues = poll_queues;
2366 
2367 	nr_io_queues = dev->nr_allocated_queues - 1;
2368 	result = nvme_set_queue_count(&dev->ctrl, &nr_io_queues);
2369 	if (result < 0)
2370 		return result;
2371 
2372 	if (nr_io_queues == 0)
2373 		return 0;
2374 
2375 	/*
2376 	 * Free IRQ resources as soon as NVMEQ_ENABLED bit transitions
2377 	 * from set to unset. If there is a window to it is truely freed,
2378 	 * pci_free_irq_vectors() jumping into this window will crash.
2379 	 * And take lock to avoid racing with pci_free_irq_vectors() in
2380 	 * nvme_dev_disable() path.
2381 	 */
2382 	result = nvme_setup_io_queues_trylock(dev);
2383 	if (result)
2384 		return result;
2385 	if (test_and_clear_bit(NVMEQ_ENABLED, &adminq->flags))
2386 		pci_free_irq(pdev, 0, adminq);
2387 
2388 	if (dev->cmb_use_sqes) {
2389 		result = nvme_cmb_qdepth(dev, nr_io_queues,
2390 				sizeof(struct nvme_command));
2391 		if (result > 0)
2392 			dev->q_depth = result;
2393 		else
2394 			dev->cmb_use_sqes = false;
2395 	}
2396 
2397 	do {
2398 		size = db_bar_size(dev, nr_io_queues);
2399 		result = nvme_remap_bar(dev, size);
2400 		if (!result)
2401 			break;
2402 		if (!--nr_io_queues) {
2403 			result = -ENOMEM;
2404 			goto out_unlock;
2405 		}
2406 	} while (1);
2407 	adminq->q_db = dev->dbs;
2408 
2409  retry:
2410 	/* Deregister the admin queue's interrupt */
2411 	if (test_and_clear_bit(NVMEQ_ENABLED, &adminq->flags))
2412 		pci_free_irq(pdev, 0, adminq);
2413 
2414 	/*
2415 	 * If we enable msix early due to not intx, disable it again before
2416 	 * setting up the full range we need.
2417 	 */
2418 	pci_free_irq_vectors(pdev);
2419 
2420 	result = nvme_setup_irqs(dev, nr_io_queues);
2421 	if (result <= 0) {
2422 		result = -EIO;
2423 		goto out_unlock;
2424 	}
2425 
2426 	dev->num_vecs = result;
2427 	result = max(result - 1, 1);
2428 	dev->max_qid = result + dev->io_queues[HCTX_TYPE_POLL];
2429 
2430 	/*
2431 	 * Should investigate if there's a performance win from allocating
2432 	 * more queues than interrupt vectors; it might allow the submission
2433 	 * path to scale better, even if the receive path is limited by the
2434 	 * number of interrupts.
2435 	 */
2436 	result = queue_request_irq(adminq);
2437 	if (result)
2438 		goto out_unlock;
2439 	set_bit(NVMEQ_ENABLED, &adminq->flags);
2440 	mutex_unlock(&dev->shutdown_lock);
2441 
2442 	result = nvme_create_io_queues(dev);
2443 	if (result || dev->online_queues < 2)
2444 		return result;
2445 
2446 	if (dev->online_queues - 1 < dev->max_qid) {
2447 		nr_io_queues = dev->online_queues - 1;
2448 		nvme_disable_io_queues(dev);
2449 		result = nvme_setup_io_queues_trylock(dev);
2450 		if (result)
2451 			return result;
2452 		nvme_suspend_io_queues(dev);
2453 		goto retry;
2454 	}
2455 	dev_info(dev->ctrl.device, "%d/%d/%d default/read/poll queues\n",
2456 					dev->io_queues[HCTX_TYPE_DEFAULT],
2457 					dev->io_queues[HCTX_TYPE_READ],
2458 					dev->io_queues[HCTX_TYPE_POLL]);
2459 	return 0;
2460 out_unlock:
2461 	mutex_unlock(&dev->shutdown_lock);
2462 	return result;
2463 }
2464 
2465 static void nvme_del_queue_end(struct request *req, blk_status_t error)
2466 {
2467 	struct nvme_queue *nvmeq = req->end_io_data;
2468 
2469 	blk_mq_free_request(req);
2470 	complete(&nvmeq->delete_done);
2471 }
2472 
2473 static void nvme_del_cq_end(struct request *req, blk_status_t error)
2474 {
2475 	struct nvme_queue *nvmeq = req->end_io_data;
2476 
2477 	if (error)
2478 		set_bit(NVMEQ_DELETE_ERROR, &nvmeq->flags);
2479 
2480 	nvme_del_queue_end(req, error);
2481 }
2482 
2483 static int nvme_delete_queue(struct nvme_queue *nvmeq, u8 opcode)
2484 {
2485 	struct request_queue *q = nvmeq->dev->ctrl.admin_q;
2486 	struct request *req;
2487 	struct nvme_command cmd = { };
2488 
2489 	cmd.delete_queue.opcode = opcode;
2490 	cmd.delete_queue.qid = cpu_to_le16(nvmeq->qid);
2491 
2492 	req = blk_mq_alloc_request(q, nvme_req_op(&cmd), BLK_MQ_REQ_NOWAIT);
2493 	if (IS_ERR(req))
2494 		return PTR_ERR(req);
2495 	nvme_init_request(req, &cmd);
2496 
2497 	if (opcode == nvme_admin_delete_cq)
2498 		req->end_io = nvme_del_cq_end;
2499 	else
2500 		req->end_io = nvme_del_queue_end;
2501 	req->end_io_data = nvmeq;
2502 
2503 	init_completion(&nvmeq->delete_done);
2504 	req->rq_flags |= RQF_QUIET;
2505 	blk_execute_rq_nowait(req, false);
2506 	return 0;
2507 }
2508 
2509 static bool __nvme_disable_io_queues(struct nvme_dev *dev, u8 opcode)
2510 {
2511 	int nr_queues = dev->online_queues - 1, sent = 0;
2512 	unsigned long timeout;
2513 
2514  retry:
2515 	timeout = NVME_ADMIN_TIMEOUT;
2516 	while (nr_queues > 0) {
2517 		if (nvme_delete_queue(&dev->queues[nr_queues], opcode))
2518 			break;
2519 		nr_queues--;
2520 		sent++;
2521 	}
2522 	while (sent) {
2523 		struct nvme_queue *nvmeq = &dev->queues[nr_queues + sent];
2524 
2525 		timeout = wait_for_completion_io_timeout(&nvmeq->delete_done,
2526 				timeout);
2527 		if (timeout == 0)
2528 			return false;
2529 
2530 		sent--;
2531 		if (nr_queues)
2532 			goto retry;
2533 	}
2534 	return true;
2535 }
2536 
2537 static void nvme_dev_add(struct nvme_dev *dev)
2538 {
2539 	int ret;
2540 
2541 	if (!dev->ctrl.tagset) {
2542 		dev->tagset.ops = &nvme_mq_ops;
2543 		dev->tagset.nr_hw_queues = dev->online_queues - 1;
2544 		dev->tagset.nr_maps = 2; /* default + read */
2545 		if (dev->io_queues[HCTX_TYPE_POLL])
2546 			dev->tagset.nr_maps++;
2547 		dev->tagset.timeout = NVME_IO_TIMEOUT;
2548 		dev->tagset.numa_node = dev->ctrl.numa_node;
2549 		dev->tagset.queue_depth = min_t(unsigned int, dev->q_depth,
2550 						BLK_MQ_MAX_DEPTH) - 1;
2551 		dev->tagset.cmd_size = sizeof(struct nvme_iod);
2552 		dev->tagset.flags = BLK_MQ_F_SHOULD_MERGE;
2553 		dev->tagset.driver_data = dev;
2554 
2555 		/*
2556 		 * Some Apple controllers requires tags to be unique
2557 		 * across admin and IO queue, so reserve the first 32
2558 		 * tags of the IO queue.
2559 		 */
2560 		if (dev->ctrl.quirks & NVME_QUIRK_SHARED_TAGS)
2561 			dev->tagset.reserved_tags = NVME_AQ_DEPTH;
2562 
2563 		ret = blk_mq_alloc_tag_set(&dev->tagset);
2564 		if (ret) {
2565 			dev_warn(dev->ctrl.device,
2566 				"IO queues tagset allocation failed %d\n", ret);
2567 			return;
2568 		}
2569 		dev->ctrl.tagset = &dev->tagset;
2570 	} else {
2571 		blk_mq_update_nr_hw_queues(&dev->tagset, dev->online_queues - 1);
2572 
2573 		/* Free previously allocated queues that are no longer usable */
2574 		nvme_free_queues(dev, dev->online_queues);
2575 	}
2576 
2577 	nvme_dbbuf_set(dev);
2578 }
2579 
2580 static int nvme_pci_enable(struct nvme_dev *dev)
2581 {
2582 	int result = -ENOMEM;
2583 	struct pci_dev *pdev = to_pci_dev(dev->dev);
2584 	int dma_address_bits = 64;
2585 
2586 	if (pci_enable_device_mem(pdev))
2587 		return result;
2588 
2589 	pci_set_master(pdev);
2590 
2591 	if (dev->ctrl.quirks & NVME_QUIRK_DMA_ADDRESS_BITS_48)
2592 		dma_address_bits = 48;
2593 	if (dma_set_mask_and_coherent(dev->dev, DMA_BIT_MASK(dma_address_bits)))
2594 		goto disable;
2595 
2596 	if (readl(dev->bar + NVME_REG_CSTS) == -1) {
2597 		result = -ENODEV;
2598 		goto disable;
2599 	}
2600 
2601 	/*
2602 	 * Some devices and/or platforms don't advertise or work with INTx
2603 	 * interrupts. Pre-enable a single MSIX or MSI vec for setup. We'll
2604 	 * adjust this later.
2605 	 */
2606 	result = pci_alloc_irq_vectors(pdev, 1, 1, PCI_IRQ_ALL_TYPES);
2607 	if (result < 0)
2608 		return result;
2609 
2610 	dev->ctrl.cap = lo_hi_readq(dev->bar + NVME_REG_CAP);
2611 
2612 	dev->q_depth = min_t(u32, NVME_CAP_MQES(dev->ctrl.cap) + 1,
2613 				io_queue_depth);
2614 	dev->ctrl.sqsize = dev->q_depth - 1; /* 0's based queue depth */
2615 	dev->db_stride = 1 << NVME_CAP_STRIDE(dev->ctrl.cap);
2616 	dev->dbs = dev->bar + 4096;
2617 
2618 	/*
2619 	 * Some Apple controllers require a non-standard SQE size.
2620 	 * Interestingly they also seem to ignore the CC:IOSQES register
2621 	 * so we don't bother updating it here.
2622 	 */
2623 	if (dev->ctrl.quirks & NVME_QUIRK_128_BYTES_SQES)
2624 		dev->io_sqes = 7;
2625 	else
2626 		dev->io_sqes = NVME_NVM_IOSQES;
2627 
2628 	/*
2629 	 * Temporary fix for the Apple controller found in the MacBook8,1 and
2630 	 * some MacBook7,1 to avoid controller resets and data loss.
2631 	 */
2632 	if (pdev->vendor == PCI_VENDOR_ID_APPLE && pdev->device == 0x2001) {
2633 		dev->q_depth = 2;
2634 		dev_warn(dev->ctrl.device, "detected Apple NVMe controller, "
2635 			"set queue depth=%u to work around controller resets\n",
2636 			dev->q_depth);
2637 	} else if (pdev->vendor == PCI_VENDOR_ID_SAMSUNG &&
2638 		   (pdev->device == 0xa821 || pdev->device == 0xa822) &&
2639 		   NVME_CAP_MQES(dev->ctrl.cap) == 0) {
2640 		dev->q_depth = 64;
2641 		dev_err(dev->ctrl.device, "detected PM1725 NVMe controller, "
2642                         "set queue depth=%u\n", dev->q_depth);
2643 	}
2644 
2645 	/*
2646 	 * Controllers with the shared tags quirk need the IO queue to be
2647 	 * big enough so that we get 32 tags for the admin queue
2648 	 */
2649 	if ((dev->ctrl.quirks & NVME_QUIRK_SHARED_TAGS) &&
2650 	    (dev->q_depth < (NVME_AQ_DEPTH + 2))) {
2651 		dev->q_depth = NVME_AQ_DEPTH + 2;
2652 		dev_warn(dev->ctrl.device, "IO queue depth clamped to %d\n",
2653 			 dev->q_depth);
2654 	}
2655 
2656 
2657 	nvme_map_cmb(dev);
2658 
2659 	pci_enable_pcie_error_reporting(pdev);
2660 	pci_save_state(pdev);
2661 	return 0;
2662 
2663  disable:
2664 	pci_disable_device(pdev);
2665 	return result;
2666 }
2667 
2668 static void nvme_dev_unmap(struct nvme_dev *dev)
2669 {
2670 	if (dev->bar)
2671 		iounmap(dev->bar);
2672 	pci_release_mem_regions(to_pci_dev(dev->dev));
2673 }
2674 
2675 static void nvme_pci_disable(struct nvme_dev *dev)
2676 {
2677 	struct pci_dev *pdev = to_pci_dev(dev->dev);
2678 
2679 	pci_free_irq_vectors(pdev);
2680 
2681 	if (pci_is_enabled(pdev)) {
2682 		pci_disable_pcie_error_reporting(pdev);
2683 		pci_disable_device(pdev);
2684 	}
2685 }
2686 
2687 static void nvme_dev_disable(struct nvme_dev *dev, bool shutdown)
2688 {
2689 	bool dead = true, freeze = false;
2690 	struct pci_dev *pdev = to_pci_dev(dev->dev);
2691 
2692 	mutex_lock(&dev->shutdown_lock);
2693 	if (pci_device_is_present(pdev) && pci_is_enabled(pdev)) {
2694 		u32 csts = readl(dev->bar + NVME_REG_CSTS);
2695 
2696 		if (dev->ctrl.state == NVME_CTRL_LIVE ||
2697 		    dev->ctrl.state == NVME_CTRL_RESETTING) {
2698 			freeze = true;
2699 			nvme_start_freeze(&dev->ctrl);
2700 		}
2701 		dead = !!((csts & NVME_CSTS_CFS) || !(csts & NVME_CSTS_RDY) ||
2702 			pdev->error_state  != pci_channel_io_normal);
2703 	}
2704 
2705 	/*
2706 	 * Give the controller a chance to complete all entered requests if
2707 	 * doing a safe shutdown.
2708 	 */
2709 	if (!dead && shutdown && freeze)
2710 		nvme_wait_freeze_timeout(&dev->ctrl, NVME_IO_TIMEOUT);
2711 
2712 	nvme_stop_queues(&dev->ctrl);
2713 
2714 	if (!dead && dev->ctrl.queue_count > 0) {
2715 		nvme_disable_io_queues(dev);
2716 		nvme_disable_admin_queue(dev, shutdown);
2717 	}
2718 	nvme_suspend_io_queues(dev);
2719 	nvme_suspend_queue(&dev->queues[0]);
2720 	nvme_pci_disable(dev);
2721 	nvme_reap_pending_cqes(dev);
2722 
2723 	blk_mq_tagset_busy_iter(&dev->tagset, nvme_cancel_request, &dev->ctrl);
2724 	blk_mq_tagset_busy_iter(&dev->admin_tagset, nvme_cancel_request, &dev->ctrl);
2725 	blk_mq_tagset_wait_completed_request(&dev->tagset);
2726 	blk_mq_tagset_wait_completed_request(&dev->admin_tagset);
2727 
2728 	/*
2729 	 * The driver will not be starting up queues again if shutting down so
2730 	 * must flush all entered requests to their failed completion to avoid
2731 	 * deadlocking blk-mq hot-cpu notifier.
2732 	 */
2733 	if (shutdown) {
2734 		nvme_start_queues(&dev->ctrl);
2735 		if (dev->ctrl.admin_q && !blk_queue_dying(dev->ctrl.admin_q))
2736 			nvme_start_admin_queue(&dev->ctrl);
2737 	}
2738 	mutex_unlock(&dev->shutdown_lock);
2739 }
2740 
2741 static int nvme_disable_prepare_reset(struct nvme_dev *dev, bool shutdown)
2742 {
2743 	if (!nvme_wait_reset(&dev->ctrl))
2744 		return -EBUSY;
2745 	nvme_dev_disable(dev, shutdown);
2746 	return 0;
2747 }
2748 
2749 static int nvme_setup_prp_pools(struct nvme_dev *dev)
2750 {
2751 	dev->prp_page_pool = dma_pool_create("prp list page", dev->dev,
2752 						NVME_CTRL_PAGE_SIZE,
2753 						NVME_CTRL_PAGE_SIZE, 0);
2754 	if (!dev->prp_page_pool)
2755 		return -ENOMEM;
2756 
2757 	/* Optimisation for I/Os between 4k and 128k */
2758 	dev->prp_small_pool = dma_pool_create("prp list 256", dev->dev,
2759 						256, 256, 0);
2760 	if (!dev->prp_small_pool) {
2761 		dma_pool_destroy(dev->prp_page_pool);
2762 		return -ENOMEM;
2763 	}
2764 	return 0;
2765 }
2766 
2767 static void nvme_release_prp_pools(struct nvme_dev *dev)
2768 {
2769 	dma_pool_destroy(dev->prp_page_pool);
2770 	dma_pool_destroy(dev->prp_small_pool);
2771 }
2772 
2773 static void nvme_free_tagset(struct nvme_dev *dev)
2774 {
2775 	if (dev->tagset.tags)
2776 		blk_mq_free_tag_set(&dev->tagset);
2777 	dev->ctrl.tagset = NULL;
2778 }
2779 
2780 static void nvme_pci_free_ctrl(struct nvme_ctrl *ctrl)
2781 {
2782 	struct nvme_dev *dev = to_nvme_dev(ctrl);
2783 
2784 	nvme_dbbuf_dma_free(dev);
2785 	nvme_free_tagset(dev);
2786 	if (dev->ctrl.admin_q)
2787 		blk_put_queue(dev->ctrl.admin_q);
2788 	free_opal_dev(dev->ctrl.opal_dev);
2789 	mempool_destroy(dev->iod_mempool);
2790 	put_device(dev->dev);
2791 	kfree(dev->queues);
2792 	kfree(dev);
2793 }
2794 
2795 static void nvme_remove_dead_ctrl(struct nvme_dev *dev)
2796 {
2797 	/*
2798 	 * Set state to deleting now to avoid blocking nvme_wait_reset(), which
2799 	 * may be holding this pci_dev's device lock.
2800 	 */
2801 	nvme_change_ctrl_state(&dev->ctrl, NVME_CTRL_DELETING);
2802 	nvme_get_ctrl(&dev->ctrl);
2803 	nvme_dev_disable(dev, false);
2804 	nvme_kill_queues(&dev->ctrl);
2805 	if (!queue_work(nvme_wq, &dev->remove_work))
2806 		nvme_put_ctrl(&dev->ctrl);
2807 }
2808 
2809 static void nvme_reset_work(struct work_struct *work)
2810 {
2811 	struct nvme_dev *dev =
2812 		container_of(work, struct nvme_dev, ctrl.reset_work);
2813 	bool was_suspend = !!(dev->ctrl.ctrl_config & NVME_CC_SHN_NORMAL);
2814 	int result;
2815 
2816 	if (dev->ctrl.state != NVME_CTRL_RESETTING) {
2817 		dev_warn(dev->ctrl.device, "ctrl state %d is not RESETTING\n",
2818 			 dev->ctrl.state);
2819 		result = -ENODEV;
2820 		goto out;
2821 	}
2822 
2823 	/*
2824 	 * If we're called to reset a live controller first shut it down before
2825 	 * moving on.
2826 	 */
2827 	if (dev->ctrl.ctrl_config & NVME_CC_ENABLE)
2828 		nvme_dev_disable(dev, false);
2829 	nvme_sync_queues(&dev->ctrl);
2830 
2831 	mutex_lock(&dev->shutdown_lock);
2832 	result = nvme_pci_enable(dev);
2833 	if (result)
2834 		goto out_unlock;
2835 
2836 	result = nvme_pci_configure_admin_queue(dev);
2837 	if (result)
2838 		goto out_unlock;
2839 
2840 	result = nvme_alloc_admin_tags(dev);
2841 	if (result)
2842 		goto out_unlock;
2843 
2844 	/*
2845 	 * Limit the max command size to prevent iod->sg allocations going
2846 	 * over a single page.
2847 	 */
2848 	dev->ctrl.max_hw_sectors = min_t(u32,
2849 		NVME_MAX_KB_SZ << 1, dma_max_mapping_size(dev->dev) >> 9);
2850 	dev->ctrl.max_segments = NVME_MAX_SEGS;
2851 
2852 	/*
2853 	 * Don't limit the IOMMU merged segment size.
2854 	 */
2855 	dma_set_max_seg_size(dev->dev, 0xffffffff);
2856 	dma_set_min_align_mask(dev->dev, NVME_CTRL_PAGE_SIZE - 1);
2857 
2858 	mutex_unlock(&dev->shutdown_lock);
2859 
2860 	/*
2861 	 * Introduce CONNECTING state from nvme-fc/rdma transports to mark the
2862 	 * initializing procedure here.
2863 	 */
2864 	if (!nvme_change_ctrl_state(&dev->ctrl, NVME_CTRL_CONNECTING)) {
2865 		dev_warn(dev->ctrl.device,
2866 			"failed to mark controller CONNECTING\n");
2867 		result = -EBUSY;
2868 		goto out;
2869 	}
2870 
2871 	/*
2872 	 * We do not support an SGL for metadata (yet), so we are limited to a
2873 	 * single integrity segment for the separate metadata pointer.
2874 	 */
2875 	dev->ctrl.max_integrity_segments = 1;
2876 
2877 	result = nvme_init_ctrl_finish(&dev->ctrl);
2878 	if (result)
2879 		goto out;
2880 
2881 	if (dev->ctrl.oacs & NVME_CTRL_OACS_SEC_SUPP) {
2882 		if (!dev->ctrl.opal_dev)
2883 			dev->ctrl.opal_dev =
2884 				init_opal_dev(&dev->ctrl, &nvme_sec_submit);
2885 		else if (was_suspend)
2886 			opal_unlock_from_suspend(dev->ctrl.opal_dev);
2887 	} else {
2888 		free_opal_dev(dev->ctrl.opal_dev);
2889 		dev->ctrl.opal_dev = NULL;
2890 	}
2891 
2892 	if (dev->ctrl.oacs & NVME_CTRL_OACS_DBBUF_SUPP) {
2893 		result = nvme_dbbuf_dma_alloc(dev);
2894 		if (result)
2895 			dev_warn(dev->dev,
2896 				 "unable to allocate dma for dbbuf\n");
2897 	}
2898 
2899 	if (dev->ctrl.hmpre) {
2900 		result = nvme_setup_host_mem(dev);
2901 		if (result < 0)
2902 			goto out;
2903 	}
2904 
2905 	result = nvme_setup_io_queues(dev);
2906 	if (result)
2907 		goto out;
2908 
2909 	/*
2910 	 * Keep the controller around but remove all namespaces if we don't have
2911 	 * any working I/O queue.
2912 	 */
2913 	if (dev->online_queues < 2) {
2914 		dev_warn(dev->ctrl.device, "IO queues not created\n");
2915 		nvme_kill_queues(&dev->ctrl);
2916 		nvme_remove_namespaces(&dev->ctrl);
2917 		nvme_free_tagset(dev);
2918 	} else {
2919 		nvme_start_queues(&dev->ctrl);
2920 		nvme_wait_freeze(&dev->ctrl);
2921 		nvme_dev_add(dev);
2922 		nvme_unfreeze(&dev->ctrl);
2923 	}
2924 
2925 	/*
2926 	 * If only admin queue live, keep it to do further investigation or
2927 	 * recovery.
2928 	 */
2929 	if (!nvme_change_ctrl_state(&dev->ctrl, NVME_CTRL_LIVE)) {
2930 		dev_warn(dev->ctrl.device,
2931 			"failed to mark controller live state\n");
2932 		result = -ENODEV;
2933 		goto out;
2934 	}
2935 
2936 	if (!dev->attrs_added && !sysfs_create_group(&dev->ctrl.device->kobj,
2937 			&nvme_pci_attr_group))
2938 		dev->attrs_added = true;
2939 
2940 	nvme_start_ctrl(&dev->ctrl);
2941 	return;
2942 
2943  out_unlock:
2944 	mutex_unlock(&dev->shutdown_lock);
2945  out:
2946 	if (result)
2947 		dev_warn(dev->ctrl.device,
2948 			 "Removing after probe failure status: %d\n", result);
2949 	nvme_remove_dead_ctrl(dev);
2950 }
2951 
2952 static void nvme_remove_dead_ctrl_work(struct work_struct *work)
2953 {
2954 	struct nvme_dev *dev = container_of(work, struct nvme_dev, remove_work);
2955 	struct pci_dev *pdev = to_pci_dev(dev->dev);
2956 
2957 	if (pci_get_drvdata(pdev))
2958 		device_release_driver(&pdev->dev);
2959 	nvme_put_ctrl(&dev->ctrl);
2960 }
2961 
2962 static int nvme_pci_reg_read32(struct nvme_ctrl *ctrl, u32 off, u32 *val)
2963 {
2964 	*val = readl(to_nvme_dev(ctrl)->bar + off);
2965 	return 0;
2966 }
2967 
2968 static int nvme_pci_reg_write32(struct nvme_ctrl *ctrl, u32 off, u32 val)
2969 {
2970 	writel(val, to_nvme_dev(ctrl)->bar + off);
2971 	return 0;
2972 }
2973 
2974 static int nvme_pci_reg_read64(struct nvme_ctrl *ctrl, u32 off, u64 *val)
2975 {
2976 	*val = lo_hi_readq(to_nvme_dev(ctrl)->bar + off);
2977 	return 0;
2978 }
2979 
2980 static int nvme_pci_get_address(struct nvme_ctrl *ctrl, char *buf, int size)
2981 {
2982 	struct pci_dev *pdev = to_pci_dev(to_nvme_dev(ctrl)->dev);
2983 
2984 	return snprintf(buf, size, "%s\n", dev_name(&pdev->dev));
2985 }
2986 
2987 
2988 static void nvme_pci_print_device_info(struct nvme_ctrl *ctrl)
2989 {
2990 	struct pci_dev *pdev = to_pci_dev(to_nvme_dev(ctrl)->dev);
2991 	struct nvme_subsystem *subsys = ctrl->subsys;
2992 
2993 	dev_err(ctrl->device,
2994 		"VID:DID %04x:%04x model:%.*s firmware:%.*s\n",
2995 		pdev->vendor, pdev->device,
2996 		nvme_strlen(subsys->model, sizeof(subsys->model)),
2997 		subsys->model, nvme_strlen(subsys->firmware_rev,
2998 					   sizeof(subsys->firmware_rev)),
2999 		subsys->firmware_rev);
3000 }
3001 
3002 static const struct nvme_ctrl_ops nvme_pci_ctrl_ops = {
3003 	.name			= "pcie",
3004 	.module			= THIS_MODULE,
3005 	.flags			= NVME_F_METADATA_SUPPORTED |
3006 				  NVME_F_PCI_P2PDMA,
3007 	.reg_read32		= nvme_pci_reg_read32,
3008 	.reg_write32		= nvme_pci_reg_write32,
3009 	.reg_read64		= nvme_pci_reg_read64,
3010 	.free_ctrl		= nvme_pci_free_ctrl,
3011 	.submit_async_event	= nvme_pci_submit_async_event,
3012 	.get_address		= nvme_pci_get_address,
3013 	.print_device_info	= nvme_pci_print_device_info,
3014 };
3015 
3016 static int nvme_dev_map(struct nvme_dev *dev)
3017 {
3018 	struct pci_dev *pdev = to_pci_dev(dev->dev);
3019 
3020 	if (pci_request_mem_regions(pdev, "nvme"))
3021 		return -ENODEV;
3022 
3023 	if (nvme_remap_bar(dev, NVME_REG_DBS + 4096))
3024 		goto release;
3025 
3026 	return 0;
3027   release:
3028 	pci_release_mem_regions(pdev);
3029 	return -ENODEV;
3030 }
3031 
3032 static unsigned long check_vendor_combination_bug(struct pci_dev *pdev)
3033 {
3034 	if (pdev->vendor == 0x144d && pdev->device == 0xa802) {
3035 		/*
3036 		 * Several Samsung devices seem to drop off the PCIe bus
3037 		 * randomly when APST is on and uses the deepest sleep state.
3038 		 * This has been observed on a Samsung "SM951 NVMe SAMSUNG
3039 		 * 256GB", a "PM951 NVMe SAMSUNG 512GB", and a "Samsung SSD
3040 		 * 950 PRO 256GB", but it seems to be restricted to two Dell
3041 		 * laptops.
3042 		 */
3043 		if (dmi_match(DMI_SYS_VENDOR, "Dell Inc.") &&
3044 		    (dmi_match(DMI_PRODUCT_NAME, "XPS 15 9550") ||
3045 		     dmi_match(DMI_PRODUCT_NAME, "Precision 5510")))
3046 			return NVME_QUIRK_NO_DEEPEST_PS;
3047 	} else if (pdev->vendor == 0x144d && pdev->device == 0xa804) {
3048 		/*
3049 		 * Samsung SSD 960 EVO drops off the PCIe bus after system
3050 		 * suspend on a Ryzen board, ASUS PRIME B350M-A, as well as
3051 		 * within few minutes after bootup on a Coffee Lake board -
3052 		 * ASUS PRIME Z370-A
3053 		 */
3054 		if (dmi_match(DMI_BOARD_VENDOR, "ASUSTeK COMPUTER INC.") &&
3055 		    (dmi_match(DMI_BOARD_NAME, "PRIME B350M-A") ||
3056 		     dmi_match(DMI_BOARD_NAME, "PRIME Z370-A")))
3057 			return NVME_QUIRK_NO_APST;
3058 	} else if ((pdev->vendor == 0x144d && (pdev->device == 0xa801 ||
3059 		    pdev->device == 0xa808 || pdev->device == 0xa809)) ||
3060 		   (pdev->vendor == 0x1e0f && pdev->device == 0x0001)) {
3061 		/*
3062 		 * Forcing to use host managed nvme power settings for
3063 		 * lowest idle power with quick resume latency on
3064 		 * Samsung and Toshiba SSDs based on suspend behavior
3065 		 * on Coffee Lake board for LENOVO C640
3066 		 */
3067 		if ((dmi_match(DMI_BOARD_VENDOR, "LENOVO")) &&
3068 		     dmi_match(DMI_BOARD_NAME, "LNVNB161216"))
3069 			return NVME_QUIRK_SIMPLE_SUSPEND;
3070 	}
3071 
3072 	return 0;
3073 }
3074 
3075 static void nvme_async_probe(void *data, async_cookie_t cookie)
3076 {
3077 	struct nvme_dev *dev = data;
3078 
3079 	flush_work(&dev->ctrl.reset_work);
3080 	flush_work(&dev->ctrl.scan_work);
3081 	nvme_put_ctrl(&dev->ctrl);
3082 }
3083 
3084 static int nvme_probe(struct pci_dev *pdev, const struct pci_device_id *id)
3085 {
3086 	int node, result = -ENOMEM;
3087 	struct nvme_dev *dev;
3088 	unsigned long quirks = id->driver_data;
3089 	size_t alloc_size;
3090 
3091 	node = dev_to_node(&pdev->dev);
3092 	if (node == NUMA_NO_NODE)
3093 		set_dev_node(&pdev->dev, first_memory_node);
3094 
3095 	dev = kzalloc_node(sizeof(*dev), GFP_KERNEL, node);
3096 	if (!dev)
3097 		return -ENOMEM;
3098 
3099 	dev->nr_write_queues = write_queues;
3100 	dev->nr_poll_queues = poll_queues;
3101 	dev->nr_allocated_queues = nvme_max_io_queues(dev) + 1;
3102 	dev->queues = kcalloc_node(dev->nr_allocated_queues,
3103 			sizeof(struct nvme_queue), GFP_KERNEL, node);
3104 	if (!dev->queues)
3105 		goto free;
3106 
3107 	dev->dev = get_device(&pdev->dev);
3108 	pci_set_drvdata(pdev, dev);
3109 
3110 	result = nvme_dev_map(dev);
3111 	if (result)
3112 		goto put_pci;
3113 
3114 	INIT_WORK(&dev->ctrl.reset_work, nvme_reset_work);
3115 	INIT_WORK(&dev->remove_work, nvme_remove_dead_ctrl_work);
3116 	mutex_init(&dev->shutdown_lock);
3117 
3118 	result = nvme_setup_prp_pools(dev);
3119 	if (result)
3120 		goto unmap;
3121 
3122 	quirks |= check_vendor_combination_bug(pdev);
3123 
3124 	if (!noacpi && acpi_storage_d3(&pdev->dev)) {
3125 		/*
3126 		 * Some systems use a bios work around to ask for D3 on
3127 		 * platforms that support kernel managed suspend.
3128 		 */
3129 		dev_info(&pdev->dev,
3130 			 "platform quirk: setting simple suspend\n");
3131 		quirks |= NVME_QUIRK_SIMPLE_SUSPEND;
3132 	}
3133 
3134 	/*
3135 	 * Double check that our mempool alloc size will cover the biggest
3136 	 * command we support.
3137 	 */
3138 	alloc_size = nvme_pci_iod_alloc_size();
3139 	WARN_ON_ONCE(alloc_size > PAGE_SIZE);
3140 
3141 	dev->iod_mempool = mempool_create_node(1, mempool_kmalloc,
3142 						mempool_kfree,
3143 						(void *) alloc_size,
3144 						GFP_KERNEL, node);
3145 	if (!dev->iod_mempool) {
3146 		result = -ENOMEM;
3147 		goto release_pools;
3148 	}
3149 
3150 	result = nvme_init_ctrl(&dev->ctrl, &pdev->dev, &nvme_pci_ctrl_ops,
3151 			quirks);
3152 	if (result)
3153 		goto release_mempool;
3154 
3155 	dev_info(dev->ctrl.device, "pci function %s\n", dev_name(&pdev->dev));
3156 
3157 	nvme_reset_ctrl(&dev->ctrl);
3158 	async_schedule(nvme_async_probe, dev);
3159 
3160 	return 0;
3161 
3162  release_mempool:
3163 	mempool_destroy(dev->iod_mempool);
3164  release_pools:
3165 	nvme_release_prp_pools(dev);
3166  unmap:
3167 	nvme_dev_unmap(dev);
3168  put_pci:
3169 	put_device(dev->dev);
3170  free:
3171 	kfree(dev->queues);
3172 	kfree(dev);
3173 	return result;
3174 }
3175 
3176 static void nvme_reset_prepare(struct pci_dev *pdev)
3177 {
3178 	struct nvme_dev *dev = pci_get_drvdata(pdev);
3179 
3180 	/*
3181 	 * We don't need to check the return value from waiting for the reset
3182 	 * state as pci_dev device lock is held, making it impossible to race
3183 	 * with ->remove().
3184 	 */
3185 	nvme_disable_prepare_reset(dev, false);
3186 	nvme_sync_queues(&dev->ctrl);
3187 }
3188 
3189 static void nvme_reset_done(struct pci_dev *pdev)
3190 {
3191 	struct nvme_dev *dev = pci_get_drvdata(pdev);
3192 
3193 	if (!nvme_try_sched_reset(&dev->ctrl))
3194 		flush_work(&dev->ctrl.reset_work);
3195 }
3196 
3197 static void nvme_shutdown(struct pci_dev *pdev)
3198 {
3199 	struct nvme_dev *dev = pci_get_drvdata(pdev);
3200 
3201 	nvme_disable_prepare_reset(dev, true);
3202 }
3203 
3204 static void nvme_remove_attrs(struct nvme_dev *dev)
3205 {
3206 	if (dev->attrs_added)
3207 		sysfs_remove_group(&dev->ctrl.device->kobj,
3208 				   &nvme_pci_attr_group);
3209 }
3210 
3211 /*
3212  * The driver's remove may be called on a device in a partially initialized
3213  * state. This function must not have any dependencies on the device state in
3214  * order to proceed.
3215  */
3216 static void nvme_remove(struct pci_dev *pdev)
3217 {
3218 	struct nvme_dev *dev = pci_get_drvdata(pdev);
3219 
3220 	nvme_change_ctrl_state(&dev->ctrl, NVME_CTRL_DELETING);
3221 	pci_set_drvdata(pdev, NULL);
3222 
3223 	if (!pci_device_is_present(pdev)) {
3224 		nvme_change_ctrl_state(&dev->ctrl, NVME_CTRL_DEAD);
3225 		nvme_dev_disable(dev, true);
3226 	}
3227 
3228 	flush_work(&dev->ctrl.reset_work);
3229 	nvme_stop_ctrl(&dev->ctrl);
3230 	nvme_remove_namespaces(&dev->ctrl);
3231 	nvme_dev_disable(dev, true);
3232 	nvme_remove_attrs(dev);
3233 	nvme_free_host_mem(dev);
3234 	nvme_dev_remove_admin(dev);
3235 	nvme_free_queues(dev, 0);
3236 	nvme_release_prp_pools(dev);
3237 	nvme_dev_unmap(dev);
3238 	nvme_uninit_ctrl(&dev->ctrl);
3239 }
3240 
3241 #ifdef CONFIG_PM_SLEEP
3242 static int nvme_get_power_state(struct nvme_ctrl *ctrl, u32 *ps)
3243 {
3244 	return nvme_get_features(ctrl, NVME_FEAT_POWER_MGMT, 0, NULL, 0, ps);
3245 }
3246 
3247 static int nvme_set_power_state(struct nvme_ctrl *ctrl, u32 ps)
3248 {
3249 	return nvme_set_features(ctrl, NVME_FEAT_POWER_MGMT, ps, NULL, 0, NULL);
3250 }
3251 
3252 static int nvme_resume(struct device *dev)
3253 {
3254 	struct nvme_dev *ndev = pci_get_drvdata(to_pci_dev(dev));
3255 	struct nvme_ctrl *ctrl = &ndev->ctrl;
3256 
3257 	if (ndev->last_ps == U32_MAX ||
3258 	    nvme_set_power_state(ctrl, ndev->last_ps) != 0)
3259 		goto reset;
3260 	if (ctrl->hmpre && nvme_setup_host_mem(ndev))
3261 		goto reset;
3262 
3263 	return 0;
3264 reset:
3265 	return nvme_try_sched_reset(ctrl);
3266 }
3267 
3268 static int nvme_suspend(struct device *dev)
3269 {
3270 	struct pci_dev *pdev = to_pci_dev(dev);
3271 	struct nvme_dev *ndev = pci_get_drvdata(pdev);
3272 	struct nvme_ctrl *ctrl = &ndev->ctrl;
3273 	int ret = -EBUSY;
3274 
3275 	ndev->last_ps = U32_MAX;
3276 
3277 	/*
3278 	 * The platform does not remove power for a kernel managed suspend so
3279 	 * use host managed nvme power settings for lowest idle power if
3280 	 * possible. This should have quicker resume latency than a full device
3281 	 * shutdown.  But if the firmware is involved after the suspend or the
3282 	 * device does not support any non-default power states, shut down the
3283 	 * device fully.
3284 	 *
3285 	 * If ASPM is not enabled for the device, shut down the device and allow
3286 	 * the PCI bus layer to put it into D3 in order to take the PCIe link
3287 	 * down, so as to allow the platform to achieve its minimum low-power
3288 	 * state (which may not be possible if the link is up).
3289 	 */
3290 	if (pm_suspend_via_firmware() || !ctrl->npss ||
3291 	    !pcie_aspm_enabled(pdev) ||
3292 	    (ndev->ctrl.quirks & NVME_QUIRK_SIMPLE_SUSPEND))
3293 		return nvme_disable_prepare_reset(ndev, true);
3294 
3295 	nvme_start_freeze(ctrl);
3296 	nvme_wait_freeze(ctrl);
3297 	nvme_sync_queues(ctrl);
3298 
3299 	if (ctrl->state != NVME_CTRL_LIVE)
3300 		goto unfreeze;
3301 
3302 	/*
3303 	 * Host memory access may not be successful in a system suspend state,
3304 	 * but the specification allows the controller to access memory in a
3305 	 * non-operational power state.
3306 	 */
3307 	if (ndev->hmb) {
3308 		ret = nvme_set_host_mem(ndev, 0);
3309 		if (ret < 0)
3310 			goto unfreeze;
3311 	}
3312 
3313 	ret = nvme_get_power_state(ctrl, &ndev->last_ps);
3314 	if (ret < 0)
3315 		goto unfreeze;
3316 
3317 	/*
3318 	 * A saved state prevents pci pm from generically controlling the
3319 	 * device's power. If we're using protocol specific settings, we don't
3320 	 * want pci interfering.
3321 	 */
3322 	pci_save_state(pdev);
3323 
3324 	ret = nvme_set_power_state(ctrl, ctrl->npss);
3325 	if (ret < 0)
3326 		goto unfreeze;
3327 
3328 	if (ret) {
3329 		/* discard the saved state */
3330 		pci_load_saved_state(pdev, NULL);
3331 
3332 		/*
3333 		 * Clearing npss forces a controller reset on resume. The
3334 		 * correct value will be rediscovered then.
3335 		 */
3336 		ret = nvme_disable_prepare_reset(ndev, true);
3337 		ctrl->npss = 0;
3338 	}
3339 unfreeze:
3340 	nvme_unfreeze(ctrl);
3341 	return ret;
3342 }
3343 
3344 static int nvme_simple_suspend(struct device *dev)
3345 {
3346 	struct nvme_dev *ndev = pci_get_drvdata(to_pci_dev(dev));
3347 
3348 	return nvme_disable_prepare_reset(ndev, true);
3349 }
3350 
3351 static int nvme_simple_resume(struct device *dev)
3352 {
3353 	struct pci_dev *pdev = to_pci_dev(dev);
3354 	struct nvme_dev *ndev = pci_get_drvdata(pdev);
3355 
3356 	return nvme_try_sched_reset(&ndev->ctrl);
3357 }
3358 
3359 static const struct dev_pm_ops nvme_dev_pm_ops = {
3360 	.suspend	= nvme_suspend,
3361 	.resume		= nvme_resume,
3362 	.freeze		= nvme_simple_suspend,
3363 	.thaw		= nvme_simple_resume,
3364 	.poweroff	= nvme_simple_suspend,
3365 	.restore	= nvme_simple_resume,
3366 };
3367 #endif /* CONFIG_PM_SLEEP */
3368 
3369 static pci_ers_result_t nvme_error_detected(struct pci_dev *pdev,
3370 						pci_channel_state_t state)
3371 {
3372 	struct nvme_dev *dev = pci_get_drvdata(pdev);
3373 
3374 	/*
3375 	 * A frozen channel requires a reset. When detected, this method will
3376 	 * shutdown the controller to quiesce. The controller will be restarted
3377 	 * after the slot reset through driver's slot_reset callback.
3378 	 */
3379 	switch (state) {
3380 	case pci_channel_io_normal:
3381 		return PCI_ERS_RESULT_CAN_RECOVER;
3382 	case pci_channel_io_frozen:
3383 		dev_warn(dev->ctrl.device,
3384 			"frozen state error detected, reset controller\n");
3385 		nvme_dev_disable(dev, false);
3386 		return PCI_ERS_RESULT_NEED_RESET;
3387 	case pci_channel_io_perm_failure:
3388 		dev_warn(dev->ctrl.device,
3389 			"failure state error detected, request disconnect\n");
3390 		return PCI_ERS_RESULT_DISCONNECT;
3391 	}
3392 	return PCI_ERS_RESULT_NEED_RESET;
3393 }
3394 
3395 static pci_ers_result_t nvme_slot_reset(struct pci_dev *pdev)
3396 {
3397 	struct nvme_dev *dev = pci_get_drvdata(pdev);
3398 
3399 	dev_info(dev->ctrl.device, "restart after slot reset\n");
3400 	pci_restore_state(pdev);
3401 	nvme_reset_ctrl(&dev->ctrl);
3402 	return PCI_ERS_RESULT_RECOVERED;
3403 }
3404 
3405 static void nvme_error_resume(struct pci_dev *pdev)
3406 {
3407 	struct nvme_dev *dev = pci_get_drvdata(pdev);
3408 
3409 	flush_work(&dev->ctrl.reset_work);
3410 }
3411 
3412 static const struct pci_error_handlers nvme_err_handler = {
3413 	.error_detected	= nvme_error_detected,
3414 	.slot_reset	= nvme_slot_reset,
3415 	.resume		= nvme_error_resume,
3416 	.reset_prepare	= nvme_reset_prepare,
3417 	.reset_done	= nvme_reset_done,
3418 };
3419 
3420 static const struct pci_device_id nvme_id_table[] = {
3421 	{ PCI_VDEVICE(INTEL, 0x0953),	/* Intel 750/P3500/P3600/P3700 */
3422 		.driver_data = NVME_QUIRK_STRIPE_SIZE |
3423 				NVME_QUIRK_DEALLOCATE_ZEROES, },
3424 	{ PCI_VDEVICE(INTEL, 0x0a53),	/* Intel P3520 */
3425 		.driver_data = NVME_QUIRK_STRIPE_SIZE |
3426 				NVME_QUIRK_DEALLOCATE_ZEROES, },
3427 	{ PCI_VDEVICE(INTEL, 0x0a54),	/* Intel P4500/P4600 */
3428 		.driver_data = NVME_QUIRK_STRIPE_SIZE |
3429 				NVME_QUIRK_DEALLOCATE_ZEROES |
3430 				NVME_QUIRK_IGNORE_DEV_SUBNQN, },
3431 	{ PCI_VDEVICE(INTEL, 0x0a55),	/* Dell Express Flash P4600 */
3432 		.driver_data = NVME_QUIRK_STRIPE_SIZE |
3433 				NVME_QUIRK_DEALLOCATE_ZEROES, },
3434 	{ PCI_VDEVICE(INTEL, 0xf1a5),	/* Intel 600P/P3100 */
3435 		.driver_data = NVME_QUIRK_NO_DEEPEST_PS |
3436 				NVME_QUIRK_MEDIUM_PRIO_SQ |
3437 				NVME_QUIRK_NO_TEMP_THRESH_CHANGE |
3438 				NVME_QUIRK_DISABLE_WRITE_ZEROES, },
3439 	{ PCI_VDEVICE(INTEL, 0xf1a6),	/* Intel 760p/Pro 7600p */
3440 		.driver_data = NVME_QUIRK_IGNORE_DEV_SUBNQN, },
3441 	{ PCI_VDEVICE(INTEL, 0x5845),	/* Qemu emulated controller */
3442 		.driver_data = NVME_QUIRK_IDENTIFY_CNS |
3443 				NVME_QUIRK_DISABLE_WRITE_ZEROES |
3444 				NVME_QUIRK_BOGUS_NID, },
3445 	{ PCI_VDEVICE(REDHAT, 0x0010),	/* Qemu emulated controller */
3446 		.driver_data = NVME_QUIRK_BOGUS_NID, },
3447 	{ PCI_DEVICE(0x126f, 0x2263),	/* Silicon Motion unidentified */
3448 		.driver_data = NVME_QUIRK_NO_NS_DESC_LIST |
3449 				NVME_QUIRK_BOGUS_NID, },
3450 	{ PCI_DEVICE(0x1bb1, 0x0100),   /* Seagate Nytro Flash Storage */
3451 		.driver_data = NVME_QUIRK_DELAY_BEFORE_CHK_RDY |
3452 				NVME_QUIRK_NO_NS_DESC_LIST, },
3453 	{ PCI_DEVICE(0x1c58, 0x0003),	/* HGST adapter */
3454 		.driver_data = NVME_QUIRK_DELAY_BEFORE_CHK_RDY, },
3455 	{ PCI_DEVICE(0x1c58, 0x0023),	/* WDC SN200 adapter */
3456 		.driver_data = NVME_QUIRK_DELAY_BEFORE_CHK_RDY, },
3457 	{ PCI_DEVICE(0x1c5f, 0x0540),	/* Memblaze Pblaze4 adapter */
3458 		.driver_data = NVME_QUIRK_DELAY_BEFORE_CHK_RDY, },
3459 	{ PCI_DEVICE(0x144d, 0xa821),   /* Samsung PM1725 */
3460 		.driver_data = NVME_QUIRK_DELAY_BEFORE_CHK_RDY, },
3461 	{ PCI_DEVICE(0x144d, 0xa822),   /* Samsung PM1725a */
3462 		.driver_data = NVME_QUIRK_DELAY_BEFORE_CHK_RDY |
3463 				NVME_QUIRK_DISABLE_WRITE_ZEROES|
3464 				NVME_QUIRK_IGNORE_DEV_SUBNQN, },
3465 	{ PCI_DEVICE(0x1987, 0x5012),	/* Phison E12 */
3466 		.driver_data = NVME_QUIRK_BOGUS_NID, },
3467 	{ PCI_DEVICE(0x1987, 0x5016),	/* Phison E16 */
3468 		.driver_data = NVME_QUIRK_IGNORE_DEV_SUBNQN, },
3469 	{ PCI_DEVICE(0x1b4b, 0x1092),	/* Lexar 256 GB SSD */
3470 		.driver_data = NVME_QUIRK_NO_NS_DESC_LIST |
3471 				NVME_QUIRK_IGNORE_DEV_SUBNQN, },
3472 	{ PCI_DEVICE(0x10ec, 0x5762),   /* ADATA SX6000LNP */
3473 		.driver_data = NVME_QUIRK_IGNORE_DEV_SUBNQN, },
3474 	{ PCI_DEVICE(0x1cc1, 0x8201),   /* ADATA SX8200PNP 512GB */
3475 		.driver_data = NVME_QUIRK_NO_DEEPEST_PS |
3476 				NVME_QUIRK_IGNORE_DEV_SUBNQN, },
3477 	{ PCI_DEVICE(0x1c5c, 0x1504),   /* SK Hynix PC400 */
3478 		.driver_data = NVME_QUIRK_DISABLE_WRITE_ZEROES, },
3479 	{ PCI_DEVICE(0x1c5c, 0x174a),   /* SK Hynix P31 SSD */
3480 		.driver_data = NVME_QUIRK_BOGUS_NID, },
3481 	{ PCI_DEVICE(0x15b7, 0x2001),   /*  Sandisk Skyhawk */
3482 		.driver_data = NVME_QUIRK_DISABLE_WRITE_ZEROES, },
3483 	{ PCI_DEVICE(0x1d97, 0x2263),   /* SPCC */
3484 		.driver_data = NVME_QUIRK_DISABLE_WRITE_ZEROES, },
3485 	{ PCI_DEVICE(0x144d, 0xa80b),   /* Samsung PM9B1 256G and 512G */
3486 		.driver_data = NVME_QUIRK_DISABLE_WRITE_ZEROES, },
3487 	{ PCI_DEVICE(0x144d, 0xa809),   /* Samsung MZALQ256HBJD 256G */
3488 		.driver_data = NVME_QUIRK_DISABLE_WRITE_ZEROES, },
3489 	{ PCI_DEVICE(0x1cc4, 0x6303),   /* UMIS RPJTJ512MGE1QDY 512G */
3490 		.driver_data = NVME_QUIRK_DISABLE_WRITE_ZEROES, },
3491 	{ PCI_DEVICE(0x1cc4, 0x6302),   /* UMIS RPJTJ256MGE1QDY 256G */
3492 		.driver_data = NVME_QUIRK_DISABLE_WRITE_ZEROES, },
3493 	{ PCI_DEVICE(0x2646, 0x2262),   /* KINGSTON SKC2000 NVMe SSD */
3494 		.driver_data = NVME_QUIRK_NO_DEEPEST_PS, },
3495 	{ PCI_DEVICE(0x2646, 0x2263),   /* KINGSTON A2000 NVMe SSD  */
3496 		.driver_data = NVME_QUIRK_NO_DEEPEST_PS, },
3497 	{ PCI_DEVICE(0x1e4B, 0x1001),   /* MAXIO MAP1001 */
3498 		.driver_data = NVME_QUIRK_BOGUS_NID, },
3499 	{ PCI_DEVICE(0x1e4B, 0x1002),   /* MAXIO MAP1002 */
3500 		.driver_data = NVME_QUIRK_BOGUS_NID, },
3501 	{ PCI_DEVICE(0x1e4B, 0x1202),   /* MAXIO MAP1202 */
3502 		.driver_data = NVME_QUIRK_BOGUS_NID, },
3503 	{ PCI_DEVICE(0x1cc1, 0x5350),   /* ADATA XPG GAMMIX S50 */
3504 		.driver_data = NVME_QUIRK_BOGUS_NID, },
3505 	{ PCI_DEVICE(0x1e49, 0x0041),   /* ZHITAI TiPro7000 NVMe SSD */
3506 		.driver_data = NVME_QUIRK_NO_DEEPEST_PS, },
3507 	{ PCI_DEVICE(PCI_VENDOR_ID_AMAZON, 0x0061),
3508 		.driver_data = NVME_QUIRK_DMA_ADDRESS_BITS_48, },
3509 	{ PCI_DEVICE(PCI_VENDOR_ID_AMAZON, 0x0065),
3510 		.driver_data = NVME_QUIRK_DMA_ADDRESS_BITS_48, },
3511 	{ PCI_DEVICE(PCI_VENDOR_ID_AMAZON, 0x8061),
3512 		.driver_data = NVME_QUIRK_DMA_ADDRESS_BITS_48, },
3513 	{ PCI_DEVICE(PCI_VENDOR_ID_AMAZON, 0xcd00),
3514 		.driver_data = NVME_QUIRK_DMA_ADDRESS_BITS_48, },
3515 	{ PCI_DEVICE(PCI_VENDOR_ID_AMAZON, 0xcd01),
3516 		.driver_data = NVME_QUIRK_DMA_ADDRESS_BITS_48, },
3517 	{ PCI_DEVICE(PCI_VENDOR_ID_AMAZON, 0xcd02),
3518 		.driver_data = NVME_QUIRK_DMA_ADDRESS_BITS_48, },
3519 	{ PCI_DEVICE(PCI_VENDOR_ID_APPLE, 0x2001),
3520 		.driver_data = NVME_QUIRK_SINGLE_VECTOR },
3521 	{ PCI_DEVICE(PCI_VENDOR_ID_APPLE, 0x2003) },
3522 	{ PCI_DEVICE(PCI_VENDOR_ID_APPLE, 0x2005),
3523 		.driver_data = NVME_QUIRK_SINGLE_VECTOR |
3524 				NVME_QUIRK_128_BYTES_SQES |
3525 				NVME_QUIRK_SHARED_TAGS |
3526 				NVME_QUIRK_SKIP_CID_GEN },
3527 	{ PCI_DEVICE(0x144d, 0xa808),   /* Samsung X5 */
3528 		.driver_data =  NVME_QUIRK_DELAY_BEFORE_CHK_RDY|
3529 				NVME_QUIRK_NO_DEEPEST_PS |
3530 				NVME_QUIRK_IGNORE_DEV_SUBNQN, },
3531 	{ PCI_DEVICE_CLASS(PCI_CLASS_STORAGE_EXPRESS, 0xffffff) },
3532 	{ 0, }
3533 };
3534 MODULE_DEVICE_TABLE(pci, nvme_id_table);
3535 
3536 static struct pci_driver nvme_driver = {
3537 	.name		= "nvme",
3538 	.id_table	= nvme_id_table,
3539 	.probe		= nvme_probe,
3540 	.remove		= nvme_remove,
3541 	.shutdown	= nvme_shutdown,
3542 #ifdef CONFIG_PM_SLEEP
3543 	.driver		= {
3544 		.pm	= &nvme_dev_pm_ops,
3545 	},
3546 #endif
3547 	.sriov_configure = pci_sriov_configure_simple,
3548 	.err_handler	= &nvme_err_handler,
3549 };
3550 
3551 static int __init nvme_init(void)
3552 {
3553 	BUILD_BUG_ON(sizeof(struct nvme_create_cq) != 64);
3554 	BUILD_BUG_ON(sizeof(struct nvme_create_sq) != 64);
3555 	BUILD_BUG_ON(sizeof(struct nvme_delete_queue) != 64);
3556 	BUILD_BUG_ON(IRQ_AFFINITY_MAX_SETS < 2);
3557 
3558 	return pci_register_driver(&nvme_driver);
3559 }
3560 
3561 static void __exit nvme_exit(void)
3562 {
3563 	pci_unregister_driver(&nvme_driver);
3564 	flush_workqueue(nvme_wq);
3565 }
3566 
3567 MODULE_AUTHOR("Matthew Wilcox <willy@linux.intel.com>");
3568 MODULE_LICENSE("GPL");
3569 MODULE_VERSION("1.0");
3570 module_init(nvme_init);
3571 module_exit(nvme_exit);
3572