xref: /linux/drivers/nvme/host/pci.c (revision cb21777e86a33286badffe5fa76c6a0acb086cd6)
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/async.h>
9 #include <linux/blkdev.h>
10 #include <linux/blk-mq-dma.h>
11 #include <linux/blk-integrity.h>
12 #include <linux/dmi.h>
13 #include <linux/init.h>
14 #include <linux/interrupt.h>
15 #include <linux/io.h>
16 #include <linux/kstrtox.h>
17 #include <linux/memremap.h>
18 #include <linux/mm.h>
19 #include <linux/module.h>
20 #include <linux/mutex.h>
21 #include <linux/nodemask.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 
31 #include "trace.h"
32 #include "nvme.h"
33 #include "pci-dmabuf.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 /* Optimisation for I/Os between 4k and 128k */
39 #define NVME_SMALL_POOL_SIZE	256
40 
41 /*
42  * Arbitrary upper bound.
43  */
44 #define NVME_MAX_BYTES		SZ_8M
45 #define NVME_MAX_NR_DESCRIPTORS	5
46 
47 /*
48  * For data SGLs we support a single descriptors worth of SGL entries.
49  * For PRPs, segments don't matter at all.
50  */
51 #define NVME_MAX_SEGS \
52 	(NVME_CTRL_PAGE_SIZE / sizeof(struct nvme_sgl_desc))
53 
54 /*
55  * For metadata SGLs, only the small descriptor is supported, and the first
56  * entry is the segment descriptor, which for the data pointer sits in the SQE.
57  */
58 #define NVME_MAX_META_SEGS \
59 	((NVME_SMALL_POOL_SIZE / sizeof(struct nvme_sgl_desc)) - 1)
60 
61 /*
62  * The last entry is used to link to the next descriptor.
63  */
64 #define PRPS_PER_PAGE \
65 	(((NVME_CTRL_PAGE_SIZE / sizeof(__le64))) - 1)
66 
67 /*
68  * I/O could be non-aligned both at the beginning and end.
69  */
70 #define MAX_PRP_RANGE \
71 	(NVME_MAX_BYTES + 2 * (NVME_CTRL_PAGE_SIZE - 1))
72 
73 static_assert(MAX_PRP_RANGE / NVME_CTRL_PAGE_SIZE <=
74 	(1 /* prp1 */ + NVME_MAX_NR_DESCRIPTORS * PRPS_PER_PAGE));
75 
76 struct quirk_entry {
77 	u16 vendor_id;
78 	u16 dev_id;
79 	u32 enabled_quirks;
80 	u32 disabled_quirks;
81 };
82 
83 static int use_threaded_interrupts;
84 module_param(use_threaded_interrupts, int, 0444);
85 
86 static bool use_cmb_sqes = true;
87 module_param(use_cmb_sqes, bool, 0444);
88 MODULE_PARM_DESC(use_cmb_sqes, "use controller's memory buffer for I/O SQes");
89 
90 static unsigned int max_host_mem_size_mb = 128;
91 module_param(max_host_mem_size_mb, uint, 0444);
92 MODULE_PARM_DESC(max_host_mem_size_mb,
93 	"Maximum Host Memory Buffer (HMB) size per controller (in MiB)");
94 
95 static unsigned int sgl_threshold = SZ_32K;
96 module_param(sgl_threshold, uint, 0644);
97 MODULE_PARM_DESC(sgl_threshold,
98 		"Use SGLs when average request segment size is larger or equal to "
99 		"this size. Use 0 to disable SGLs.");
100 
101 #define NVME_PCI_MIN_QUEUE_SIZE 2
102 #define NVME_PCI_MAX_QUEUE_SIZE 4095
103 static int io_queue_depth_set(const char *val, const struct kernel_param *kp);
104 static const struct kernel_param_ops io_queue_depth_ops = {
105 	.set = io_queue_depth_set,
106 	.get = param_get_uint,
107 };
108 
109 static unsigned int io_queue_depth = 1024;
110 module_param_cb(io_queue_depth, &io_queue_depth_ops, &io_queue_depth, 0644);
111 MODULE_PARM_DESC(io_queue_depth, "set io queue depth, should >= 2 and < 4096");
112 
113 static struct quirk_entry *nvme_pci_quirk_list;
114 static unsigned int nvme_pci_quirk_count;
115 
116 /* Helper to parse individual quirk names */
117 static int nvme_parse_quirk_names(char *quirk_str, struct quirk_entry *entry)
118 {
119 	int i;
120 	size_t field_len;
121 	bool disabled, found;
122 	char *p = quirk_str, *field;
123 
124 	while ((field = strsep(&p, ",")) && *field) {
125 		disabled = false;
126 		found = false;
127 
128 		if (*field == '^') {
129 			/* Skip the '^' character */
130 			disabled = true;
131 			field++;
132 		}
133 
134 		field_len = strlen(field);
135 		for (i = 0; i < 32; i++) {
136 			unsigned int bit = 1U << i;
137 			char *q_name = nvme_quirk_name(bit);
138 			size_t q_len = strlen(q_name);
139 
140 			if (!strcmp(q_name, "unknown"))
141 				break;
142 
143 			if (!strcmp(q_name, field) &&
144 				    q_len == field_len) {
145 				if (disabled)
146 					entry->disabled_quirks |= bit;
147 				else
148 					entry->enabled_quirks |= bit;
149 				found = true;
150 				break;
151 			}
152 		}
153 
154 		if (!found) {
155 			pr_err("nvme: unrecognized quirk %s\n", field);
156 			return -EINVAL;
157 		}
158 	}
159 	return 0;
160 }
161 
162 /* Helper to parse a single VID:DID:quirk_names entry */
163 static int nvme_parse_quirk_entry(char *s, struct quirk_entry *entry)
164 {
165 	char *field;
166 
167 	field = strsep(&s, ":");
168 	if (!field || kstrtou16(field, 16, &entry->vendor_id))
169 		return -EINVAL;
170 
171 	field = strsep(&s, ":");
172 	if (!field || kstrtou16(field, 16, &entry->dev_id))
173 		return -EINVAL;
174 
175 	field = strsep(&s, ":");
176 	if (!field)
177 		return -EINVAL;
178 
179 	return nvme_parse_quirk_names(field, entry);
180 }
181 
182 static int quirks_param_set(const char *value, const struct kernel_param *kp)
183 {
184 	int count, err, i;
185 	struct quirk_entry *qlist;
186 	char *field, *val, *sep_ptr;
187 
188 	err = param_set_copystring(value, kp);
189 	if (err)
190 		return err;
191 
192 	val = kstrdup(value, GFP_KERNEL);
193 	if (!val)
194 		return -ENOMEM;
195 
196 	if (!*val)
197 		goto out_free_val;
198 
199 	count = 1;
200 	for (i = 0; val[i]; i++) {
201 		if (val[i] == '-')
202 			count++;
203 	}
204 
205 	qlist = kzalloc_objs(*qlist, count);
206 	if (!qlist) {
207 		err = -ENOMEM;
208 		goto out_free_val;
209 	}
210 
211 	i = 0;
212 	sep_ptr = val;
213 	while ((field = strsep(&sep_ptr, "-"))) {
214 		if (nvme_parse_quirk_entry(field, &qlist[i])) {
215 			pr_err("nvme: failed to parse quirk string %s\n",
216 				value);
217 			err = -EINVAL;
218 			goto out_free_qlist;
219 		}
220 
221 		i++;
222 	}
223 
224 	kfree(nvme_pci_quirk_list);
225 	nvme_pci_quirk_count = count;
226 	nvme_pci_quirk_list  = qlist;
227 	goto out_free_val;
228 
229 out_free_qlist:
230 	kfree(qlist);
231 out_free_val:
232 	kfree(val);
233 	return err;
234 }
235 
236 static char quirks_param[256];
237 static const struct kernel_param_ops quirks_param_ops = {
238 	.set = quirks_param_set,
239 	.get = param_get_string,
240 };
241 
242 static struct kparam_string quirks_param_string = {
243 	.maxlen = sizeof(quirks_param),
244 	.string = quirks_param,
245 };
246 
247 device_param_cb(quirks, &quirks_param_ops, &quirks_param_string, 0444);
248 MODULE_PARM_DESC(quirks, "Enable/disable NVMe quirks by specifying "
249 						"quirks=VID:DID:quirk_names");
250 
251 static int io_queue_count_set(const char *val, const struct kernel_param *kp)
252 {
253 	unsigned int n;
254 	int ret;
255 
256 	ret = kstrtouint(val, 10, &n);
257 	if (ret != 0 || n > blk_mq_num_possible_queues(0))
258 		return -EINVAL;
259 	return param_set_uint(val, kp);
260 }
261 
262 static const struct kernel_param_ops io_queue_count_ops = {
263 	.set = io_queue_count_set,
264 	.get = param_get_uint,
265 };
266 
267 static unsigned int write_queues;
268 module_param_cb(write_queues, &io_queue_count_ops, &write_queues, 0644);
269 MODULE_PARM_DESC(write_queues,
270 	"Number of queues to use for writes. If not set, reads and writes "
271 	"will share a queue set.");
272 
273 static unsigned int poll_queues;
274 module_param_cb(poll_queues, &io_queue_count_ops, &poll_queues, 0644);
275 MODULE_PARM_DESC(poll_queues, "Number of queues to use for polled IO.");
276 
277 static bool noacpi;
278 module_param(noacpi, bool, 0444);
279 MODULE_PARM_DESC(noacpi, "disable acpi bios quirks");
280 
281 struct nvme_dev;
282 struct nvme_queue;
283 
284 static void nvme_dev_disable(struct nvme_dev *dev, bool shutdown);
285 static void nvme_delete_io_queues(struct nvme_dev *dev);
286 static void nvme_update_attrs(struct nvme_dev *dev);
287 
288 struct nvme_descriptor_pools {
289 	struct dma_pool *large;
290 	struct dma_pool *small;
291 };
292 
293 /*
294  * Represents an NVM Express device.  Each nvme_dev is a PCI function.
295  */
296 struct nvme_dev {
297 	struct nvme_queue *queues;
298 	struct blk_mq_tag_set tagset;
299 	struct blk_mq_tag_set admin_tagset;
300 	u32 __iomem *dbs;
301 	struct device *dev;
302 	unsigned online_queues;
303 	unsigned max_qid;
304 	unsigned io_queues[HCTX_MAX_TYPES];
305 	unsigned int num_vecs;
306 	u32 q_depth;
307 	int io_sqes;
308 	u32 db_stride;
309 	void __iomem *bar;
310 	unsigned long bar_mapped_size;
311 	struct mutex shutdown_lock;
312 	bool subsystem;
313 	u64 cmb_size;
314 	bool cmb_use_sqes;
315 	u32 cmbsz;
316 	u32 cmbloc;
317 	struct nvme_ctrl ctrl;
318 	u32 last_ps;
319 	bool hmb;
320 	struct sg_table *hmb_sgt;
321 	mempool_t *dmavec_mempool;
322 	struct nvme_dma_buf_list dmabufs;
323 
324 	/* shadow doorbell buffer support: */
325 	__le32 *dbbuf_dbs;
326 	dma_addr_t dbbuf_dbs_dma_addr;
327 	__le32 *dbbuf_eis;
328 	dma_addr_t dbbuf_eis_dma_addr;
329 
330 	/* host memory buffer support: */
331 	u64 host_mem_size;
332 	u32 nr_host_mem_descs;
333 	u32 host_mem_descs_size;
334 	dma_addr_t host_mem_descs_dma;
335 	struct nvme_host_mem_buf_desc *host_mem_descs;
336 	void **host_mem_desc_bufs;
337 	unsigned int nr_allocated_queues;
338 	unsigned int nr_write_queues;
339 	unsigned int nr_poll_queues;
340 	struct nvme_descriptor_pools descriptor_pools[];
341 };
342 
343 static int io_queue_depth_set(const char *val, const struct kernel_param *kp)
344 {
345 	return param_set_uint_minmax(val, kp, NVME_PCI_MIN_QUEUE_SIZE,
346 			NVME_PCI_MAX_QUEUE_SIZE);
347 }
348 
349 static inline unsigned int sq_idx(unsigned int qid, u32 stride)
350 {
351 	return qid * 2 * stride;
352 }
353 
354 static inline unsigned int cq_idx(unsigned int qid, u32 stride)
355 {
356 	return (qid * 2 + 1) * stride;
357 }
358 
359 static inline struct nvme_dev *to_nvme_dev(struct nvme_ctrl *ctrl)
360 {
361 	return container_of(ctrl, struct nvme_dev, ctrl);
362 }
363 
364 /*
365  * An NVM Express queue.  Each device has at least two (one for admin
366  * commands and one for I/O commands).
367  */
368 struct nvme_queue {
369 	struct nvme_dev *dev;
370 	struct nvme_descriptor_pools descriptor_pools;
371 	spinlock_t sq_lock;
372 	void *sq_cmds
373 		__guarded_by(&sq_lock);
374 	 /* only used for poll queues: */
375 	spinlock_t cq_poll_lock ____cacheline_aligned_in_smp;
376 	struct nvme_completion *cqes;
377 	dma_addr_t sq_dma_addr;
378 	dma_addr_t cq_dma_addr;
379 	u32 __iomem *q_db;
380 	u32 q_depth;
381 	u16 cq_vector;
382 	u16 cq_head;
383 	u16 sq_tail
384 		__guarded_by(&sq_lock);
385 	u16 last_sq_tail
386 		__guarded_by(&sq_lock);
387 	u16 qid;
388 	u8 cq_phase;
389 	u8 sqes;
390 	unsigned long flags;
391 #define NVMEQ_ENABLED		0
392 #define NVMEQ_SQ_CMB		1
393 #define NVMEQ_POLLED		3
394 	__le32 *dbbuf_sq_db;
395 	__le32 *dbbuf_cq_db;
396 	__le32 *dbbuf_sq_ei;
397 	__le32 *dbbuf_cq_ei;
398 	struct completion delete_done;
399 };
400 
401 /* bits for iod->flags */
402 enum nvme_iod_flags {
403 	/* this command has been aborted by the timeout handler */
404 	IOD_ABORTED		= 1U << 0,
405 
406 	/* uses the small descriptor pool */
407 	IOD_SMALL_DESCRIPTOR	= 1U << 1,
408 
409 	/* single segment dma mapping */
410 	IOD_SINGLE_SEGMENT	= 1U << 2,
411 
412 	/* Data payload contains p2p memory */
413 	IOD_DATA_P2P		= 1U << 3,
414 
415 	/* Metadata contains p2p memory */
416 	IOD_META_P2P		= 1U << 4,
417 
418 	/* Data payload contains MMIO memory */
419 	IOD_DATA_MMIO		= 1U << 5,
420 
421 	/* Metadata contains MMIO memory */
422 	IOD_META_MMIO		= 1U << 6,
423 
424 	/* Metadata using non-coalesced MPTR */
425 	IOD_SINGLE_META_SEGMENT	= 1U << 7,
426 };
427 
428 struct nvme_dma_vec {
429 	dma_addr_t addr;
430 	unsigned int len;
431 };
432 
433 /*
434  * The nvme_iod describes the data in an I/O.
435  */
436 struct nvme_iod {
437 	struct nvme_request req;
438 	struct nvme_command cmd;
439 	u8 flags;
440 	u8 nr_descriptors;
441 
442 	size_t total_len;
443 	struct dma_iova_state dma_state;
444 	void *descriptors[NVME_MAX_NR_DESCRIPTORS];
445 	struct nvme_dma_vec *dma_vecs;
446 	unsigned int nr_dma_vecs;
447 
448 	dma_addr_t meta_dma;
449 	size_t meta_total_len;
450 	struct dma_iova_state meta_dma_state;
451 	struct nvme_sgl_desc *meta_descriptor;
452 };
453 
454 static inline unsigned int nvme_dbbuf_size(struct nvme_dev *dev)
455 {
456 	return dev->nr_allocated_queues * 8 * dev->db_stride;
457 }
458 
459 static void nvme_dbbuf_dma_alloc(struct nvme_dev *dev)
460 {
461 	unsigned int mem_size = nvme_dbbuf_size(dev);
462 
463 	if (!(dev->ctrl.oacs & NVME_CTRL_OACS_DBBUF_SUPP))
464 		return;
465 
466 	if (dev->dbbuf_dbs) {
467 		/*
468 		 * Clear the dbbuf memory so the driver doesn't observe stale
469 		 * values from the previous instantiation.
470 		 */
471 		memset(dev->dbbuf_dbs, 0, mem_size);
472 		memset(dev->dbbuf_eis, 0, mem_size);
473 		return;
474 	}
475 
476 	dev->dbbuf_dbs = dma_alloc_coherent(dev->dev, mem_size,
477 					    &dev->dbbuf_dbs_dma_addr,
478 					    GFP_KERNEL);
479 	if (!dev->dbbuf_dbs)
480 		goto fail;
481 	dev->dbbuf_eis = dma_alloc_coherent(dev->dev, mem_size,
482 					    &dev->dbbuf_eis_dma_addr,
483 					    GFP_KERNEL);
484 	if (!dev->dbbuf_eis)
485 		goto fail_free_dbbuf_dbs;
486 	return;
487 
488 fail_free_dbbuf_dbs:
489 	dma_free_coherent(dev->dev, mem_size, dev->dbbuf_dbs,
490 			  dev->dbbuf_dbs_dma_addr);
491 	dev->dbbuf_dbs = NULL;
492 fail:
493 	dev_warn(dev->dev, "unable to allocate dma for dbbuf\n");
494 }
495 
496 static void nvme_dbbuf_dma_free(struct nvme_dev *dev)
497 {
498 	unsigned int mem_size = nvme_dbbuf_size(dev);
499 
500 	if (dev->dbbuf_dbs) {
501 		dma_free_coherent(dev->dev, mem_size,
502 				  dev->dbbuf_dbs, dev->dbbuf_dbs_dma_addr);
503 		dev->dbbuf_dbs = NULL;
504 	}
505 	if (dev->dbbuf_eis) {
506 		dma_free_coherent(dev->dev, mem_size,
507 				  dev->dbbuf_eis, dev->dbbuf_eis_dma_addr);
508 		dev->dbbuf_eis = NULL;
509 	}
510 }
511 
512 static void nvme_dbbuf_init(struct nvme_dev *dev,
513 			    struct nvme_queue *nvmeq, int qid)
514 {
515 	if (!dev->dbbuf_dbs || !qid)
516 		return;
517 
518 	nvmeq->dbbuf_sq_db = &dev->dbbuf_dbs[sq_idx(qid, dev->db_stride)];
519 	nvmeq->dbbuf_cq_db = &dev->dbbuf_dbs[cq_idx(qid, dev->db_stride)];
520 	nvmeq->dbbuf_sq_ei = &dev->dbbuf_eis[sq_idx(qid, dev->db_stride)];
521 	nvmeq->dbbuf_cq_ei = &dev->dbbuf_eis[cq_idx(qid, dev->db_stride)];
522 }
523 
524 static void nvme_dbbuf_free(struct nvme_queue *nvmeq)
525 {
526 	if (!nvmeq->qid)
527 		return;
528 
529 	nvmeq->dbbuf_sq_db = NULL;
530 	nvmeq->dbbuf_cq_db = NULL;
531 	nvmeq->dbbuf_sq_ei = NULL;
532 	nvmeq->dbbuf_cq_ei = NULL;
533 }
534 
535 static void nvme_dbbuf_set(struct nvme_dev *dev)
536 {
537 	struct nvme_command c = { };
538 	unsigned int i;
539 
540 	if (!dev->dbbuf_dbs)
541 		return;
542 
543 	c.dbbuf.opcode = nvme_admin_dbbuf;
544 	c.dbbuf.prp1 = cpu_to_le64(dev->dbbuf_dbs_dma_addr);
545 	c.dbbuf.prp2 = cpu_to_le64(dev->dbbuf_eis_dma_addr);
546 
547 	if (nvme_submit_sync_cmd(dev->ctrl.admin_q, &c, NULL, 0)) {
548 		dev_warn(dev->ctrl.device, "unable to set dbbuf\n");
549 		/* Free memory and continue on */
550 		nvme_dbbuf_dma_free(dev);
551 
552 		for (i = 1; i < dev->online_queues; i++)
553 			nvme_dbbuf_free(&dev->queues[i]);
554 	}
555 }
556 
557 static inline int nvme_dbbuf_need_event(u16 event_idx, u16 new_idx, u16 old)
558 {
559 	return (u16)(new_idx - event_idx - 1) < (u16)(new_idx - old);
560 }
561 
562 /* Update dbbuf and return true if an MMIO is required */
563 static bool nvme_dbbuf_update_and_check_event(u16 value, __le32 *dbbuf_db,
564 					      volatile __le32 *dbbuf_ei)
565 {
566 	if (dbbuf_db) {
567 		u16 old_value, event_idx;
568 
569 		/*
570 		 * Ensure that the queue is written before updating
571 		 * the doorbell in memory
572 		 */
573 		wmb();
574 
575 		old_value = le32_to_cpu(*dbbuf_db);
576 		*dbbuf_db = cpu_to_le32(value);
577 
578 		/*
579 		 * Ensure that the doorbell is updated before reading the event
580 		 * index from memory.  The controller needs to provide similar
581 		 * ordering to ensure the event index is updated before reading
582 		 * the doorbell.
583 		 */
584 		mb();
585 
586 		event_idx = le32_to_cpu(*dbbuf_ei);
587 		if (!nvme_dbbuf_need_event(event_idx, value, old_value))
588 			return false;
589 	}
590 
591 	return true;
592 }
593 
594 static struct nvme_descriptor_pools *
595 nvme_setup_descriptor_pools(struct nvme_dev *dev, int numa_node)
596 {
597 	struct nvme_descriptor_pools *pools;
598 	size_t small_align = NVME_SMALL_POOL_SIZE;
599 
600 	if (numa_node == NUMA_NO_NODE)
601 		numa_node = 0;
602 
603 	pools = &dev->descriptor_pools[numa_node];
604 
605 	if (pools->small)
606 		return pools; /* already initialized */
607 
608 	pools->large = dma_pool_create_node("nvme descriptor page", dev->dev,
609 			NVME_CTRL_PAGE_SIZE, NVME_CTRL_PAGE_SIZE, 0, numa_node);
610 	if (!pools->large)
611 		return ERR_PTR(-ENOMEM);
612 
613 	if (dev->ctrl.quirks & NVME_QUIRK_DMAPOOL_ALIGN_512)
614 		small_align = 512;
615 
616 	pools->small = dma_pool_create_node("nvme descriptor small", dev->dev,
617 			NVME_SMALL_POOL_SIZE, small_align, 0, numa_node);
618 	if (!pools->small) {
619 		dma_pool_destroy(pools->large);
620 		pools->large = NULL;
621 		return ERR_PTR(-ENOMEM);
622 	}
623 
624 	return pools;
625 }
626 
627 static void nvme_release_descriptor_pools(struct nvme_dev *dev)
628 {
629 	unsigned i;
630 
631 	for (i = 0; i < nr_node_ids; i++) {
632 		struct nvme_descriptor_pools *pools = &dev->descriptor_pools[i];
633 
634 		dma_pool_destroy(pools->large);
635 		dma_pool_destroy(pools->small);
636 	}
637 }
638 
639 static int nvme_init_hctx_common(struct blk_mq_hw_ctx *hctx, void *data,
640 		unsigned qid)
641 {
642 	struct nvme_dev *dev = to_nvme_dev(data);
643 	struct nvme_queue *nvmeq = &dev->queues[qid];
644 	struct nvme_descriptor_pools *pools;
645 	struct blk_mq_tags *tags;
646 
647 	tags = qid ? dev->tagset.tags[qid - 1] : dev->admin_tagset.tags[0];
648 	WARN_ON(tags != hctx->tags);
649 	pools = nvme_setup_descriptor_pools(dev, hctx->numa_node);
650 	if (IS_ERR(pools))
651 		return PTR_ERR(pools);
652 
653 	nvmeq->descriptor_pools = *pools;
654 	hctx->driver_data = nvmeq;
655 	return 0;
656 }
657 
658 static int nvme_admin_init_hctx(struct blk_mq_hw_ctx *hctx, void *data,
659 				unsigned int hctx_idx)
660 {
661 	WARN_ON(hctx_idx != 0);
662 	return nvme_init_hctx_common(hctx, data, 0);
663 }
664 
665 static int nvme_init_hctx(struct blk_mq_hw_ctx *hctx, void *data,
666 			     unsigned int hctx_idx)
667 {
668 	return nvme_init_hctx_common(hctx, data, hctx_idx + 1);
669 }
670 
671 static int nvme_pci_init_request(struct blk_mq_tag_set *set,
672 		struct request *req, unsigned int hctx_idx,
673 		int numa_node)
674 {
675 	struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
676 
677 	nvme_req(req)->ctrl = set->driver_data;
678 	nvme_req(req)->cmd = &iod->cmd;
679 	return 0;
680 }
681 
682 static int queue_irq_offset(struct nvme_dev *dev)
683 {
684 	/* if we have more than 1 vec, admin queue offsets us by 1 */
685 	if (dev->num_vecs > 1)
686 		return 1;
687 
688 	return 0;
689 }
690 
691 static void nvme_pci_map_queues(struct blk_mq_tag_set *set)
692 {
693 	struct nvme_dev *dev = to_nvme_dev(set->driver_data);
694 	int i, qoff, offset;
695 
696 	offset = queue_irq_offset(dev);
697 	for (i = 0, qoff = 0; i < set->nr_maps; i++) {
698 		struct blk_mq_queue_map *map = &set->map[i];
699 
700 		map->nr_queues = dev->io_queues[i];
701 		if (!map->nr_queues) {
702 			BUG_ON(i == HCTX_TYPE_DEFAULT);
703 			continue;
704 		}
705 
706 		/*
707 		 * The poll queue(s) doesn't have an IRQ (and hence IRQ
708 		 * affinity), so use the regular blk-mq cpu mapping
709 		 */
710 		map->queue_offset = qoff;
711 		if (i != HCTX_TYPE_POLL && offset)
712 			blk_mq_map_hw_queues(map, dev->dev, offset);
713 		else
714 			blk_mq_map_queues(map);
715 		qoff += map->nr_queues;
716 		offset += map->nr_queues;
717 	}
718 }
719 
720 /*
721  * Write sq tail if we are asked to, or if the next command would wrap.
722  */
723 static inline void nvme_write_sq_db(struct nvme_queue *nvmeq, bool write_sq)
724 	__must_hold(&nvmeq->sq_lock)
725 {
726 	if (!write_sq) {
727 		u16 next_tail = nvmeq->sq_tail + 1;
728 
729 		if (next_tail == nvmeq->q_depth)
730 			next_tail = 0;
731 		if (next_tail != nvmeq->last_sq_tail)
732 			return;
733 	}
734 
735 	if (nvme_dbbuf_update_and_check_event(nvmeq->sq_tail,
736 			nvmeq->dbbuf_sq_db, nvmeq->dbbuf_sq_ei))
737 		writel(nvmeq->sq_tail, nvmeq->q_db);
738 	nvmeq->last_sq_tail = nvmeq->sq_tail;
739 }
740 
741 static inline void nvme_sq_copy_cmd(struct nvme_queue *nvmeq,
742 				    struct nvme_command *cmd)
743 	__must_hold(&nvmeq->sq_lock)
744 {
745 	memcpy(nvmeq->sq_cmds + (nvmeq->sq_tail << nvmeq->sqes),
746 		absolute_pointer(cmd), sizeof(*cmd));
747 	if (++nvmeq->sq_tail == nvmeq->q_depth)
748 		nvmeq->sq_tail = 0;
749 }
750 
751 static void nvme_commit_rqs(struct blk_mq_hw_ctx *hctx)
752 {
753 	struct nvme_queue *nvmeq = hctx->driver_data;
754 
755 	spin_lock(&nvmeq->sq_lock);
756 	if (nvmeq->sq_tail != nvmeq->last_sq_tail)
757 		nvme_write_sq_db(nvmeq, true);
758 	spin_unlock(&nvmeq->sq_lock);
759 }
760 
761 enum nvme_use_sgl {
762 	SGL_UNSUPPORTED,
763 	SGL_SUPPORTED,
764 	SGL_FORCED,
765 };
766 
767 static inline bool nvme_pci_metadata_use_sgls(struct request *req)
768 {
769 	struct nvme_queue *nvmeq = req->mq_hctx->driver_data;
770 	struct nvme_dev *dev = nvmeq->dev;
771 
772 	if (!nvme_ctrl_meta_sgl_supported(&dev->ctrl))
773 		return false;
774 	return req->nr_integrity_segments > 1 ||
775 		nvme_req(req)->flags & NVME_REQ_USERCMD;
776 }
777 
778 static inline enum nvme_use_sgl nvme_pci_use_sgls(struct nvme_dev *dev,
779 		struct request *req)
780 {
781 	struct nvme_queue *nvmeq = req->mq_hctx->driver_data;
782 
783 	if (nvmeq->qid && nvme_ctrl_sgl_supported(&dev->ctrl)) {
784 		/*
785 		 * When the controller is capable of using SGL, there are
786 		 * several conditions that we force to use it:
787 		 *
788 		 * 1. A request containing page gaps within the controller's
789 		 *    mask can not use the PRP format.
790 		 *
791 		 * 2. User commands use SGL because that lets the device
792 		 *    validate the requested transfer lengths.
793 		 *
794 		 * 3. Multiple integrity segments must use SGL as that's the
795 		 *    only way to describe such a command in NVMe.
796 		 */
797 		if (req_phys_gap_mask(req) & (NVME_CTRL_PAGE_SIZE - 1) ||
798 		    nvme_req(req)->flags & NVME_REQ_USERCMD ||
799 		    req->nr_integrity_segments > 1)
800 			return SGL_FORCED;
801 		return SGL_SUPPORTED;
802 	}
803 
804 	return SGL_UNSUPPORTED;
805 }
806 
807 static unsigned int nvme_pci_avg_seg_size(struct request *req)
808 {
809 	struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
810 	unsigned int nseg;
811 
812 	if (blk_rq_dma_map_coalesce(&iod->dma_state))
813 		nseg = 1;
814 	else
815 		nseg = blk_rq_nr_phys_segments(req);
816 	return DIV_ROUND_UP(blk_rq_payload_bytes(req), nseg);
817 }
818 
819 static inline struct dma_pool *nvme_dma_pool(struct nvme_queue *nvmeq,
820 		struct nvme_iod *iod)
821 {
822 	if (iod->flags & IOD_SMALL_DESCRIPTOR)
823 		return nvmeq->descriptor_pools.small;
824 	return nvmeq->descriptor_pools.large;
825 }
826 
827 static inline bool nvme_pci_cmd_use_meta_sgl(struct nvme_command *cmd)
828 {
829 	return (cmd->common.flags & NVME_CMD_SGL_ALL) == NVME_CMD_SGL_METASEG;
830 }
831 
832 static inline bool nvme_pci_cmd_use_sgl(struct nvme_command *cmd)
833 {
834 	return cmd->common.flags &
835 		(NVME_CMD_SGL_METABUF | NVME_CMD_SGL_METASEG);
836 }
837 
838 static inline dma_addr_t nvme_pci_first_desc_dma_addr(struct nvme_command *cmd)
839 {
840 	if (nvme_pci_cmd_use_sgl(cmd))
841 		return le64_to_cpu(cmd->common.dptr.sgl.addr);
842 	return le64_to_cpu(cmd->common.dptr.prp2);
843 }
844 
845 static void nvme_free_descriptors(struct request *req)
846 {
847 	struct nvme_queue *nvmeq = req->mq_hctx->driver_data;
848 	const int last_prp = NVME_CTRL_PAGE_SIZE / sizeof(__le64) - 1;
849 	struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
850 	dma_addr_t dma_addr = nvme_pci_first_desc_dma_addr(&iod->cmd);
851 	int i;
852 
853 	if (iod->nr_descriptors == 1) {
854 		dma_pool_free(nvme_dma_pool(nvmeq, iod), iod->descriptors[0],
855 				dma_addr);
856 		return;
857 	}
858 
859 	for (i = 0; i < iod->nr_descriptors; i++) {
860 		__le64 *prp_list = iod->descriptors[i];
861 		dma_addr_t next_dma_addr = le64_to_cpu(prp_list[last_prp]);
862 
863 		dma_pool_free(nvmeq->descriptor_pools.large, prp_list,
864 				dma_addr);
865 		dma_addr = next_dma_addr;
866 	}
867 }
868 
869 static blk_status_t nvme_rq_setup_dmabuf_map(struct request *req,
870 		struct nvme_queue *nvmeq)
871 {
872 	struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
873 	struct nvme_dmabuf_map *map = nvme_req_to_map(req);
874 	unsigned bvec_done = req->bio->bi_iter.bi_offset;
875 	unsigned map_idx = bvec_done / NVME_CTRL_PAGE_SIZE;
876 	unsigned offset = bvec_done & (NVME_CTRL_PAGE_SIZE - 1);
877 	int length = blk_rq_payload_bytes(req) - (NVME_CTRL_PAGE_SIZE - offset);
878 	dma_addr_t *dma_list = map->dma_list;
879 	u64 prp1_dma = dma_list[map_idx++] + offset;
880 	u64 dma_addr, prp2_dma;
881 	dma_addr_t prp_dma;
882 	__le64 *prp_list;
883 	unsigned i;
884 
885 	if (length <= 0) {
886 		prp2_dma = 0;
887 		goto done;
888 	}
889 
890 	if (length <= NVME_CTRL_PAGE_SIZE) {
891 		prp2_dma = dma_list[map_idx];
892 		goto done;
893 	}
894 
895 	if (DIV_ROUND_UP(length, NVME_CTRL_PAGE_SIZE) <=
896 	    NVME_SMALL_POOL_SIZE / sizeof(__le64))
897 		iod->flags |= IOD_SMALL_DESCRIPTOR;
898 
899 	prp_list = dma_pool_alloc(nvme_dma_pool(nvmeq, iod), GFP_ATOMIC,
900 			&prp_dma);
901 	if (!prp_list)
902 		return BLK_STS_RESOURCE;
903 
904 	iod->descriptors[iod->nr_descriptors++] = prp_list;
905 	prp2_dma = prp_dma;
906 	i = 0;
907 	for (;;) {
908 		if (i == NVME_CTRL_PAGE_SIZE >> 3) {
909 			__le64 *old_prp_list = prp_list;
910 
911 			prp_list = dma_pool_alloc(nvmeq->descriptor_pools.large,
912 					GFP_ATOMIC, &prp_dma);
913 			if (!prp_list)
914 				goto free_prps;
915 			iod->descriptors[iod->nr_descriptors++] = prp_list;
916 			prp_list[0] = old_prp_list[i - 1];
917 			old_prp_list[i - 1] = cpu_to_le64(prp_dma);
918 			i = 1;
919 		}
920 
921 		dma_addr = dma_list[map_idx++];
922 		prp_list[i++] = cpu_to_le64(dma_addr);
923 
924 		length -= NVME_CTRL_PAGE_SIZE;
925 		if (length <= 0)
926 			break;
927 	}
928 done:
929 	iod->cmd.common.dptr.prp1 = cpu_to_le64(prp1_dma);
930 	iod->cmd.common.dptr.prp2 = cpu_to_le64(prp2_dma);
931 	return BLK_STS_OK;
932 free_prps:
933 	iod->cmd.common.dptr.prp1 = cpu_to_le64(prp1_dma);
934 	iod->cmd.common.dptr.prp2 = cpu_to_le64(prp2_dma);
935 	nvme_free_descriptors(req);
936 	return BLK_STS_RESOURCE;
937 }
938 
939 static void nvme_free_prps(struct request *req, unsigned int attrs)
940 {
941 	struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
942 	struct nvme_queue *nvmeq = req->mq_hctx->driver_data;
943 	unsigned int i;
944 
945 	for (i = 0; i < iod->nr_dma_vecs; i++)
946 		dma_unmap_phys(nvmeq->dev->dev, iod->dma_vecs[i].addr,
947 			       iod->dma_vecs[i].len, rq_dma_dir(req), attrs);
948 	mempool_free(iod->dma_vecs, nvmeq->dev->dmavec_mempool);
949 }
950 
951 static void nvme_free_sgls(struct request *req, struct nvme_sgl_desc *sge,
952 		struct nvme_sgl_desc *sg_list, unsigned int attrs)
953 {
954 	struct nvme_queue *nvmeq = req->mq_hctx->driver_data;
955 	enum dma_data_direction dir = rq_dma_dir(req);
956 	unsigned int len = le32_to_cpu(sge->length);
957 	struct device *dma_dev = nvmeq->dev->dev;
958 	unsigned int i;
959 
960 	if (sge->type == (NVME_SGL_FMT_DATA_DESC << 4)) {
961 		dma_unmap_phys(dma_dev, le64_to_cpu(sge->addr), len, dir,
962 			       attrs);
963 		return;
964 	}
965 
966 	for (i = 0; i < len / sizeof(*sg_list); i++)
967 		dma_unmap_phys(dma_dev, le64_to_cpu(sg_list[i].addr),
968 			le32_to_cpu(sg_list[i].length), dir, attrs);
969 }
970 
971 static void nvme_unmap_metadata(struct request *req)
972 {
973 	struct nvme_queue *nvmeq = req->mq_hctx->driver_data;
974 	enum pci_p2pdma_map_type map = PCI_P2PDMA_MAP_NONE;
975 	enum dma_data_direction dir = rq_dma_dir(req);
976 	struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
977 	struct device *dma_dev = nvmeq->dev->dev;
978 	struct nvme_sgl_desc *sge = iod->meta_descriptor;
979 	unsigned int attrs = 0;
980 
981 	if (iod->flags & IOD_SINGLE_META_SEGMENT) {
982 		dma_unmap_page(dma_dev, iod->meta_dma,
983 			       rq_integrity_vec(req).bv_len,
984 			       rq_dma_dir(req));
985 		return;
986 	}
987 
988 	if (iod->flags & IOD_META_P2P)
989 		map = PCI_P2PDMA_MAP_BUS_ADDR;
990 	else if (iod->flags & IOD_META_MMIO) {
991 		map = PCI_P2PDMA_MAP_THRU_HOST_BRIDGE;
992 		attrs |= DMA_ATTR_MMIO;
993 	}
994 
995 	if (!blk_rq_dma_unmap(req, dma_dev, &iod->meta_dma_state,
996 			      iod->meta_total_len, map)) {
997 		if (nvme_pci_cmd_use_meta_sgl(&iod->cmd))
998 			nvme_free_sgls(req, sge, &sge[1], attrs);
999 		else
1000 			dma_unmap_phys(dma_dev, iod->meta_dma,
1001 				       iod->meta_total_len, dir, attrs);
1002 	}
1003 
1004 	if (iod->meta_descriptor)
1005 		dma_pool_free(nvmeq->descriptor_pools.small,
1006 			      iod->meta_descriptor, iod->meta_dma);
1007 }
1008 
1009 static void nvme_unmap_data(struct request *req)
1010 {
1011 	enum pci_p2pdma_map_type map = PCI_P2PDMA_MAP_NONE;
1012 	struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
1013 	struct nvme_queue *nvmeq = req->mq_hctx->driver_data;
1014 	struct device *dma_dev = nvmeq->dev->dev;
1015 	unsigned int attrs = 0;
1016 
1017 	if (blk_mq_rq_is_dmabuf(req))
1018 		goto out_free_desc;
1019 
1020 	if (iod->flags & IOD_SINGLE_SEGMENT) {
1021 		static_assert(offsetof(union nvme_data_ptr, prp1) ==
1022 				offsetof(union nvme_data_ptr, sgl.addr));
1023 		dma_unmap_page(dma_dev, le64_to_cpu(iod->cmd.common.dptr.prp1),
1024 				iod->total_len, rq_dma_dir(req));
1025 		return;
1026 	}
1027 
1028 	if (iod->flags & IOD_DATA_P2P)
1029 		map = PCI_P2PDMA_MAP_BUS_ADDR;
1030 	else if (iod->flags & IOD_DATA_MMIO) {
1031 		map = PCI_P2PDMA_MAP_THRU_HOST_BRIDGE;
1032 		attrs |= DMA_ATTR_MMIO;
1033 	}
1034 
1035 	if (!blk_rq_dma_unmap(req, dma_dev, &iod->dma_state, iod->total_len,
1036 			      map)) {
1037 		if (nvme_pci_cmd_use_sgl(&iod->cmd))
1038 			nvme_free_sgls(req, &iod->cmd.common.dptr.sgl,
1039 			               iod->descriptors[0], attrs);
1040 		else
1041 			nvme_free_prps(req, attrs);
1042 	}
1043 
1044 out_free_desc:
1045 	if (iod->nr_descriptors)
1046 		nvme_free_descriptors(req);
1047 }
1048 
1049 static bool nvme_pci_prp_save_mapping(struct request *req,
1050 				      struct device *dma_dev,
1051 				      struct blk_dma_iter *iter)
1052 {
1053 	struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
1054 
1055 	if (dma_use_iova(&iod->dma_state) || !dma_need_unmap(dma_dev) ||
1056 	    (iod->flags & IOD_DATA_P2P))
1057 		return true;
1058 
1059 	if (!iod->nr_dma_vecs) {
1060 		struct nvme_queue *nvmeq = req->mq_hctx->driver_data;
1061 
1062 		iod->dma_vecs = mempool_alloc(nvmeq->dev->dmavec_mempool,
1063 				GFP_ATOMIC);
1064 		if (!iod->dma_vecs) {
1065 			iter->status = BLK_STS_RESOURCE;
1066 			return false;
1067 		}
1068 	}
1069 
1070 	iod->dma_vecs[iod->nr_dma_vecs].addr = iter->addr;
1071 	iod->dma_vecs[iod->nr_dma_vecs].len = iter->len;
1072 	iod->nr_dma_vecs++;
1073 	return true;
1074 }
1075 
1076 static bool nvme_pci_prp_iter_next(struct request *req, struct device *dma_dev,
1077 		struct blk_dma_iter *iter)
1078 {
1079 	if (iter->len)
1080 		return true;
1081 	if (!blk_rq_dma_map_iter_next(req, dma_dev, iter))
1082 		return false;
1083 	return nvme_pci_prp_save_mapping(req, dma_dev, iter);
1084 }
1085 
1086 static void nvme_unmap_iter(struct request *req, struct blk_dma_iter *iter,
1087 			    struct dma_iova_state *state)
1088 {
1089 	struct nvme_queue *nvmeq = req->mq_hctx->driver_data;
1090 	struct device *dev = nvmeq->dev->dev;
1091 
1092 	if (!blk_rq_dma_unmap(req, dev, state, iter->len, iter->p2pdma.map)) {
1093 		unsigned int attrs = 0;
1094 
1095 		if (iter->p2pdma.map == PCI_P2PDMA_MAP_THRU_HOST_BRIDGE)
1096 			attrs |= DMA_ATTR_MMIO;
1097 
1098 		dma_unmap_phys(dev, iter->addr, iter->len, rq_dma_dir(req),
1099 			       attrs);
1100 	}
1101 }
1102 
1103 static blk_status_t nvme_pci_setup_data_prp(struct request *req,
1104 		struct blk_dma_iter *iter)
1105 {
1106 	struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
1107 	struct nvme_queue *nvmeq = req->mq_hctx->driver_data;
1108 	unsigned int length = blk_rq_payload_bytes(req);
1109 	dma_addr_t prp1_dma, prp2_dma = 0;
1110 	unsigned int prp_len, i;
1111 	__le64 *prp_list;
1112 
1113 	if (!nvme_pci_prp_save_mapping(req, nvmeq->dev->dev, iter)) {
1114 		nvme_unmap_iter(req, iter, &iod->dma_state);
1115 		return iter->status;
1116 	}
1117 
1118 	/*
1119 	 * PRP1 always points to the start of the DMA transfers.
1120 	 *
1121 	 * This is the only PRP (except for the list entries) that could be
1122 	 * non-aligned.
1123 	 */
1124 	prp1_dma = iter->addr;
1125 	prp_len = min(length, NVME_CTRL_PAGE_SIZE -
1126 			(iter->addr & (NVME_CTRL_PAGE_SIZE - 1)));
1127 	iod->total_len += prp_len;
1128 	iter->addr += prp_len;
1129 	iter->len -= prp_len;
1130 	length -= prp_len;
1131 	if (!length)
1132 		goto done;
1133 
1134 	if (!nvme_pci_prp_iter_next(req, nvmeq->dev->dev, iter)) {
1135 		if (WARN_ON_ONCE(!iter->status))
1136 			goto bad_sgl;
1137 		goto done;
1138 	}
1139 
1140 	/*
1141 	 * PRP2 is usually a list, but can point to data if all data to be
1142 	 * transferred fits into PRP1 + PRP2:
1143 	 */
1144 	if (length <= NVME_CTRL_PAGE_SIZE) {
1145 		prp2_dma = iter->addr;
1146 		iod->total_len += length;
1147 		goto done;
1148 	}
1149 
1150 	if (DIV_ROUND_UP(length, NVME_CTRL_PAGE_SIZE) <=
1151 	    NVME_SMALL_POOL_SIZE / sizeof(__le64))
1152 		iod->flags |= IOD_SMALL_DESCRIPTOR;
1153 
1154 	prp_list = dma_pool_alloc(nvme_dma_pool(nvmeq, iod), GFP_ATOMIC,
1155 			&prp2_dma);
1156 	if (!prp_list) {
1157 		iter->status = BLK_STS_RESOURCE;
1158 		goto done;
1159 	}
1160 	iod->descriptors[iod->nr_descriptors++] = prp_list;
1161 
1162 	i = 0;
1163 	for (;;) {
1164 		prp_list[i++] = cpu_to_le64(iter->addr);
1165 		prp_len = min(length, NVME_CTRL_PAGE_SIZE);
1166 		if (WARN_ON_ONCE(iter->len < prp_len))
1167 			goto bad_sgl;
1168 
1169 		iod->total_len += prp_len;
1170 		iter->addr += prp_len;
1171 		iter->len -= prp_len;
1172 		length -= prp_len;
1173 		if (!length)
1174 			break;
1175 
1176 		if (!nvme_pci_prp_iter_next(req, nvmeq->dev->dev, iter)) {
1177 			if (WARN_ON_ONCE(!iter->status))
1178 				goto bad_sgl;
1179 			goto done;
1180 		}
1181 
1182 		/*
1183 		 * If we've filled the entire descriptor, allocate a new that is
1184 		 * pointed to be the last entry in the previous PRP list.  To
1185 		 * accommodate for that move the last actual entry to the new
1186 		 * descriptor.
1187 		 */
1188 		if (i == NVME_CTRL_PAGE_SIZE >> 3) {
1189 			__le64 *old_prp_list = prp_list;
1190 			dma_addr_t prp_list_dma;
1191 
1192 			prp_list = dma_pool_alloc(nvmeq->descriptor_pools.large,
1193 					GFP_ATOMIC, &prp_list_dma);
1194 			if (!prp_list) {
1195 				iter->status = BLK_STS_RESOURCE;
1196 				goto done;
1197 			}
1198 			iod->descriptors[iod->nr_descriptors++] = prp_list;
1199 
1200 			prp_list[0] = old_prp_list[i - 1];
1201 			old_prp_list[i - 1] = cpu_to_le64(prp_list_dma);
1202 			i = 1;
1203 		}
1204 	}
1205 
1206 done:
1207 	/*
1208 	 * nvme_unmap_data uses the DPT field in the SQE to tear down the
1209 	 * mapping, so initialize it even for failures.
1210 	 */
1211 	iod->cmd.common.dptr.prp1 = cpu_to_le64(prp1_dma);
1212 	iod->cmd.common.dptr.prp2 = cpu_to_le64(prp2_dma);
1213 	if (unlikely(iter->status))
1214 		nvme_unmap_data(req);
1215 	return iter->status;
1216 
1217 bad_sgl:
1218 	dev_err_once(nvmeq->dev->dev,
1219 		"Incorrectly formed request for payload:%d nents:%d\n",
1220 		blk_rq_payload_bytes(req), blk_rq_nr_phys_segments(req));
1221 	nvme_unmap_data(req);
1222 	return BLK_STS_IOERR;
1223 }
1224 
1225 static void nvme_pci_sgl_set_data(struct nvme_sgl_desc *sge,
1226 		dma_addr_t addr, u32 len)
1227 {
1228 	sge->addr = cpu_to_le64(addr);
1229 	sge->length = cpu_to_le32(len);
1230 	sge->type = NVME_SGL_FMT_DATA_DESC << 4;
1231 }
1232 
1233 static void nvme_pci_dma_iter_set_sgl(struct nvme_sgl_desc *sge,
1234 		struct blk_dma_iter *iter)
1235 {
1236 	nvme_pci_sgl_set_data(sge, iter->addr, iter->len);
1237 }
1238 
1239 static void nvme_pci_sgl_set_seg(struct nvme_sgl_desc *sge,
1240 		dma_addr_t dma_addr, int entries)
1241 {
1242 	sge->addr = cpu_to_le64(dma_addr);
1243 	sge->length = cpu_to_le32(entries * sizeof(*sge));
1244 	sge->type = NVME_SGL_FMT_LAST_SEG_DESC << 4;
1245 }
1246 
1247 static unsigned int nvme_pci_dmabuf_sgl_nents(struct request *req)
1248 {
1249 	struct bio *bio = req->bio;
1250 	struct nvme_dmabuf_map *map = to_nvme_dmabuf_map(bio->bi_dmabuf_map);
1251 	struct scatterlist *sg;
1252 	unsigned long tmp;
1253 	size_t offset = bio->bi_iter.bi_offset;
1254 	size_t remaining = blk_rq_payload_bytes(req);
1255 	unsigned int nents = 0;
1256 
1257 	for_each_sgtable_dma_sg(map->sgt, sg, tmp) {
1258 		size_t sg_len = sg_dma_len(sg);
1259 
1260 		if (!remaining)
1261 			break;
1262 		if (offset >= sg_len) {
1263 			offset -= sg_len;
1264 			continue;
1265 		}
1266 
1267 		sg_len -= offset;
1268 		offset = 0;
1269 
1270 		do {
1271 			size_t chunk = min(remaining, sg_len);
1272 
1273 			nents++;
1274 			sg_len -= chunk;
1275 			remaining -= chunk;
1276 		} while (sg_len && remaining);
1277 	}
1278 
1279 	if (unlikely(remaining))
1280 		return 0;
1281 
1282 	return nents;
1283 }
1284 
1285 static blk_status_t nvme_rq_setup_dmabuf_sgl(struct request *req,
1286 		struct nvme_queue *nvmeq, unsigned int entries)
1287 {
1288 	struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
1289 	struct nvme_dmabuf_map *map = nvme_req_to_map(req);
1290 	size_t length = blk_rq_payload_bytes(req);
1291 	struct nvme_sgl_desc *sg_list = &iod->cmd.common.dptr.sgl;
1292 	size_t offset = req->bio->bi_iter.bi_offset;
1293 	bool pooled = entries > 1;
1294 	dma_addr_t sgl_dma = 0;
1295 	unsigned int mapped = 0;
1296 	unsigned long tmp;
1297 	struct scatterlist *sg;
1298 	size_t remaining = length;
1299 
1300 	if (!entries)
1301 		return BLK_STS_IOERR;
1302 
1303 	iod->cmd.common.flags = NVME_CMD_SGL_METABUF;
1304 	iod->total_len = length;
1305 
1306 	/* entries == 1 fits in the inline descriptor; more needs a pool. */
1307 	if (pooled) {
1308 		if (entries <= NVME_SMALL_POOL_SIZE / sizeof(*sg_list))
1309 			iod->flags |= IOD_SMALL_DESCRIPTOR;
1310 
1311 		sg_list = dma_pool_alloc(nvme_dma_pool(nvmeq, iod), GFP_ATOMIC,
1312 					 &sgl_dma);
1313 		if (!sg_list)
1314 			return BLK_STS_RESOURCE;
1315 		iod->descriptors[iod->nr_descriptors++] = sg_list;
1316 	}
1317 
1318 	for_each_sgtable_dma_sg(map->sgt, sg, tmp) {
1319 		size_t sg_len = sg_dma_len(sg);
1320 		dma_addr_t addr = sg_dma_address(sg);
1321 
1322 		if (!remaining)
1323 			break;
1324 		if (offset >= sg_len) {
1325 			offset -= sg_len;
1326 			continue;
1327 		}
1328 
1329 		addr += offset;
1330 		sg_len -= offset;
1331 		offset = 0;
1332 
1333 		do {
1334 			u32 chunk = min_t(size_t, remaining, sg_len);
1335 
1336 			if (WARN_ON_ONCE(mapped == entries))
1337 				goto err_free;
1338 			nvme_pci_sgl_set_data(&sg_list[mapped++], addr, chunk);
1339 
1340 			addr += chunk;
1341 			sg_len -= chunk;
1342 			remaining -= chunk;
1343 		} while (sg_len && remaining);
1344 	}
1345 
1346 	if (unlikely(remaining))
1347 		goto err_free;
1348 
1349 	if (pooled)
1350 		nvme_pci_sgl_set_seg(&iod->cmd.common.dptr.sgl, sgl_dma,
1351 				     mapped);
1352 	return BLK_STS_OK;
1353 err_free:
1354 	if (pooled) {
1355 		iod->nr_descriptors--;
1356 		dma_pool_free(nvme_dma_pool(nvmeq, iod), sg_list, sgl_dma);
1357 	}
1358 	return BLK_STS_IOERR;
1359 }
1360 
1361 static blk_status_t nvme_rq_setup_dmabuf(struct request *req,
1362 		struct nvme_queue *nvmeq, enum nvme_use_sgl use_sgl)
1363 {
1364 	unsigned int entries;
1365 	size_t avg_seg;
1366 
1367 	if (use_sgl == SGL_UNSUPPORTED)
1368 		return nvme_rq_setup_dmabuf_map(req, nvmeq);
1369 
1370 	entries = nvme_pci_dmabuf_sgl_nents(req);
1371 
1372 	if (use_sgl == SGL_FORCED)
1373 		return nvme_rq_setup_dmabuf_sgl(req, nvmeq, entries);
1374 
1375 	avg_seg = entries ?
1376 		DIV_ROUND_UP(blk_rq_payload_bytes(req), entries) : 0;
1377 	if (sgl_threshold && avg_seg >= sgl_threshold)
1378 		return nvme_rq_setup_dmabuf_sgl(req, nvmeq, entries);
1379 
1380 	return nvme_rq_setup_dmabuf_map(req, nvmeq);
1381 }
1382 
1383 static blk_status_t nvme_pci_setup_data_sgl(struct request *req,
1384 		struct blk_dma_iter *iter)
1385 {
1386 	struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
1387 	struct nvme_queue *nvmeq = req->mq_hctx->driver_data;
1388 	unsigned int entries = blk_rq_nr_phys_segments(req);
1389 	struct nvme_sgl_desc *sg_list;
1390 	dma_addr_t sgl_dma;
1391 	unsigned int mapped = 0;
1392 
1393 	/* set the transfer type as SGL */
1394 	iod->cmd.common.flags = NVME_CMD_SGL_METABUF;
1395 
1396 	if (entries == 1 || blk_rq_dma_map_coalesce(&iod->dma_state)) {
1397 		nvme_pci_dma_iter_set_sgl(&iod->cmd.common.dptr.sgl, iter);
1398 		iod->total_len += iter->len;
1399 		return BLK_STS_OK;
1400 	}
1401 
1402 	if (entries <= NVME_SMALL_POOL_SIZE / sizeof(*sg_list))
1403 		iod->flags |= IOD_SMALL_DESCRIPTOR;
1404 
1405 	sg_list = dma_pool_alloc(nvme_dma_pool(nvmeq, iod), GFP_ATOMIC,
1406 			&sgl_dma);
1407 	if (!sg_list) {
1408 		nvme_unmap_iter(req, iter, &iod->dma_state);
1409 		return BLK_STS_RESOURCE;
1410 	}
1411 
1412 	iod->descriptors[iod->nr_descriptors++] = sg_list;
1413 
1414 	do {
1415 		if (WARN_ON_ONCE(mapped == entries)) {
1416 			iter->status = BLK_STS_IOERR;
1417 			break;
1418 		}
1419 		nvme_pci_dma_iter_set_sgl(&sg_list[mapped++], iter);
1420 		iod->total_len += iter->len;
1421 	} while (blk_rq_dma_map_iter_next(req, nvmeq->dev->dev, iter));
1422 
1423 	nvme_pci_sgl_set_seg(&iod->cmd.common.dptr.sgl, sgl_dma, mapped);
1424 	if (unlikely(iter->status))
1425 		nvme_unmap_data(req);
1426 	return iter->status;
1427 }
1428 
1429 static blk_status_t nvme_pci_setup_data_simple(struct request *req,
1430 		enum nvme_use_sgl use_sgl)
1431 {
1432 	struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
1433 	struct nvme_queue *nvmeq = req->mq_hctx->driver_data;
1434 	struct bio_vec bv = req_bvec(req);
1435 	unsigned int prp1_offset = bv.bv_offset & (NVME_CTRL_PAGE_SIZE - 1);
1436 	bool prp_possible = prp1_offset + bv.bv_len <= NVME_CTRL_PAGE_SIZE * 2;
1437 	dma_addr_t dma_addr;
1438 
1439 	if (!use_sgl && !prp_possible)
1440 		return BLK_STS_AGAIN;
1441 	if (is_pci_p2pdma_page(bv.bv_page))
1442 		return BLK_STS_AGAIN;
1443 
1444 	dma_addr = dma_map_bvec(nvmeq->dev->dev, &bv, rq_dma_dir(req), 0);
1445 	if (dma_mapping_error(nvmeq->dev->dev, dma_addr))
1446 		return BLK_STS_RESOURCE;
1447 	iod->total_len = bv.bv_len;
1448 	iod->flags |= IOD_SINGLE_SEGMENT;
1449 
1450 	if (use_sgl == SGL_FORCED || !prp_possible) {
1451 		iod->cmd.common.flags = NVME_CMD_SGL_METABUF;
1452 		iod->cmd.common.dptr.sgl.addr = cpu_to_le64(dma_addr);
1453 		iod->cmd.common.dptr.sgl.length = cpu_to_le32(bv.bv_len);
1454 		iod->cmd.common.dptr.sgl.type = NVME_SGL_FMT_DATA_DESC << 4;
1455 	} else {
1456 		unsigned int first_prp_len = NVME_CTRL_PAGE_SIZE - prp1_offset;
1457 
1458 		iod->cmd.common.dptr.prp1 = cpu_to_le64(dma_addr);
1459 		iod->cmd.common.dptr.prp2 = 0;
1460 		if (bv.bv_len > first_prp_len)
1461 			iod->cmd.common.dptr.prp2 =
1462 				cpu_to_le64(dma_addr + first_prp_len);
1463 	}
1464 
1465 	return BLK_STS_OK;
1466 }
1467 
1468 static blk_status_t nvme_map_data(struct request *req)
1469 {
1470 	struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
1471 	struct nvme_queue *nvmeq = req->mq_hctx->driver_data;
1472 	struct nvme_dev *dev = nvmeq->dev;
1473 	enum nvme_use_sgl use_sgl = nvme_pci_use_sgls(dev, req);
1474 	struct blk_dma_iter iter;
1475 	blk_status_t ret;
1476 
1477 	if (blk_mq_rq_is_dmabuf(req))
1478 		return nvme_rq_setup_dmabuf(req, nvmeq, use_sgl);
1479 
1480 	/*
1481 	 * Try to skip the DMA iterator for single segment requests, as that
1482 	 * significantly improves performances for small I/O sizes.
1483 	 */
1484 	if (blk_rq_nr_phys_segments(req) == 1) {
1485 		ret = nvme_pci_setup_data_simple(req, use_sgl);
1486 		if (ret != BLK_STS_AGAIN)
1487 			return ret;
1488 	}
1489 
1490 	if (!blk_rq_dma_map_iter_start(req, dev->dev, &iod->dma_state, &iter))
1491 		return iter.status;
1492 
1493 	switch (iter.p2pdma.map) {
1494 	case PCI_P2PDMA_MAP_BUS_ADDR:
1495 		iod->flags |= IOD_DATA_P2P;
1496 		break;
1497 	case PCI_P2PDMA_MAP_THRU_HOST_BRIDGE:
1498 		iod->flags |= IOD_DATA_MMIO;
1499 		break;
1500 	case PCI_P2PDMA_MAP_NONE:
1501 		break;
1502 	default:
1503 		return BLK_STS_RESOURCE;
1504 	}
1505 
1506 	if (use_sgl == SGL_FORCED ||
1507 	    (use_sgl == SGL_SUPPORTED &&
1508 	     (sgl_threshold && nvme_pci_avg_seg_size(req) >= sgl_threshold)))
1509 		return nvme_pci_setup_data_sgl(req, &iter);
1510 	return nvme_pci_setup_data_prp(req, &iter);
1511 }
1512 
1513 static blk_status_t nvme_pci_setup_meta_iter(struct request *req)
1514 {
1515 	struct nvme_queue *nvmeq = req->mq_hctx->driver_data;
1516 	unsigned int entries = req->nr_integrity_segments;
1517 	struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
1518 	struct nvme_dev *dev = nvmeq->dev;
1519 	struct nvme_sgl_desc *sg_list;
1520 	struct blk_dma_iter iter;
1521 	dma_addr_t sgl_dma;
1522 	int i = 0;
1523 
1524 	if (!blk_rq_integrity_dma_map_iter_start(req, dev->dev,
1525 						&iod->meta_dma_state, &iter))
1526 		return iter.status;
1527 
1528 	switch (iter.p2pdma.map) {
1529 	case PCI_P2PDMA_MAP_BUS_ADDR:
1530 		iod->flags |= IOD_META_P2P;
1531 		break;
1532 	case PCI_P2PDMA_MAP_THRU_HOST_BRIDGE:
1533 		iod->flags |= IOD_META_MMIO;
1534 		break;
1535 	case PCI_P2PDMA_MAP_NONE:
1536 		break;
1537 	default:
1538 		return BLK_STS_RESOURCE;
1539 	}
1540 
1541 	if (blk_rq_dma_map_coalesce(&iod->meta_dma_state))
1542 		entries = 1;
1543 
1544 	/*
1545 	 * The NVMe MPTR descriptor has an implicit length that the host and
1546 	 * device must agree on to avoid data/memory corruption. We trust the
1547 	 * kernel allocated correctly based on the format's parameters, so use
1548 	 * the more efficient MPTR to avoid extra dma pool allocations for the
1549 	 * SGL indirection.
1550 	 *
1551 	 * But for user commands, we don't necessarily know what they do, so
1552 	 * the driver can't validate the metadata buffer size. The SGL
1553 	 * descriptor provides an explicit length, so we're relying on that
1554 	 * mechanism to catch any misunderstandings between the application and
1555 	 * device.
1556 	 *
1557 	 * P2P DMA also needs to use the blk_dma_iter method, so mptr setup
1558 	 * leverages this routine when that happens.
1559 	 */
1560 	if (!nvme_ctrl_meta_sgl_supported(&dev->ctrl) ||
1561 	    (entries == 1 && !(nvme_req(req)->flags & NVME_REQ_USERCMD))) {
1562 		iod->cmd.common.metadata = cpu_to_le64(iter.addr);
1563 		iod->meta_total_len = iter.len;
1564 		iod->meta_dma = iter.addr;
1565 		iod->meta_descriptor = NULL;
1566 		return BLK_STS_OK;
1567 	}
1568 
1569 	sg_list = dma_pool_alloc(nvmeq->descriptor_pools.small, GFP_ATOMIC,
1570 			&sgl_dma);
1571 	if (!sg_list) {
1572 		nvme_unmap_iter(req, &iter, &iod->meta_dma_state);
1573 		return BLK_STS_RESOURCE;
1574 	}
1575 
1576 	iod->meta_descriptor = sg_list;
1577 	iod->meta_dma = sgl_dma;
1578 	iod->cmd.common.flags = NVME_CMD_SGL_METASEG;
1579 	iod->cmd.common.metadata = cpu_to_le64(sgl_dma);
1580 	if (entries == 1) {
1581 		iod->meta_total_len = iter.len;
1582 		nvme_pci_dma_iter_set_sgl(sg_list, &iter);
1583 		return BLK_STS_OK;
1584 	}
1585 
1586 	sgl_dma += sizeof(*sg_list);
1587 	do {
1588 		nvme_pci_dma_iter_set_sgl(&sg_list[++i], &iter);
1589 		iod->meta_total_len += iter.len;
1590 	} while (blk_rq_integrity_dma_map_iter_next(req, dev->dev, &iter));
1591 
1592 	nvme_pci_sgl_set_seg(sg_list, sgl_dma, i);
1593 	if (unlikely(iter.status))
1594 		nvme_unmap_metadata(req);
1595 	return iter.status;
1596 }
1597 
1598 static blk_status_t nvme_pci_setup_meta_mptr(struct request *req)
1599 {
1600 	struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
1601 	struct nvme_queue *nvmeq = req->mq_hctx->driver_data;
1602 	struct bio_vec bv = rq_integrity_vec(req);
1603 
1604 	if (is_pci_p2pdma_page(bv.bv_page))
1605 		return nvme_pci_setup_meta_iter(req);
1606 
1607 	iod->meta_dma = dma_map_bvec(nvmeq->dev->dev, &bv, rq_dma_dir(req), 0);
1608 	if (dma_mapping_error(nvmeq->dev->dev, iod->meta_dma))
1609 		return BLK_STS_IOERR;
1610 	iod->cmd.common.metadata = cpu_to_le64(iod->meta_dma);
1611 	iod->flags |= IOD_SINGLE_META_SEGMENT;
1612 	return BLK_STS_OK;
1613 }
1614 
1615 static blk_status_t nvme_map_metadata(struct request *req)
1616 {
1617 	struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
1618 
1619 	if ((iod->cmd.common.flags & NVME_CMD_SGL_METABUF) &&
1620 	    nvme_pci_metadata_use_sgls(req))
1621 		return nvme_pci_setup_meta_iter(req);
1622 	return nvme_pci_setup_meta_mptr(req);
1623 }
1624 
1625 static blk_status_t nvme_prep_rq(struct request *req)
1626 {
1627 	struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
1628 	blk_status_t ret;
1629 
1630 	iod->flags = 0;
1631 	iod->nr_descriptors = 0;
1632 	iod->total_len = 0;
1633 	iod->meta_total_len = 0;
1634 	iod->nr_dma_vecs = 0;
1635 
1636 	ret = nvme_setup_cmd(req->q->queuedata, req);
1637 	if (ret)
1638 		return ret;
1639 
1640 	if (blk_rq_nr_phys_segments(req)) {
1641 		ret = nvme_map_data(req);
1642 		if (ret)
1643 			goto out_free_cmd;
1644 	}
1645 
1646 	if (blk_integrity_rq(req)) {
1647 		ret = nvme_map_metadata(req);
1648 		if (ret)
1649 			goto out_unmap_data;
1650 	}
1651 
1652 	nvme_start_request(req);
1653 	return BLK_STS_OK;
1654 out_unmap_data:
1655 	if (blk_rq_nr_phys_segments(req))
1656 		nvme_unmap_data(req);
1657 out_free_cmd:
1658 	nvme_cleanup_cmd(req);
1659 	return ret;
1660 }
1661 
1662 static blk_status_t nvme_queue_rq(struct blk_mq_hw_ctx *hctx,
1663 			 const struct blk_mq_queue_data *bd)
1664 {
1665 	struct nvme_queue *nvmeq = hctx->driver_data;
1666 	struct nvme_dev *dev = nvmeq->dev;
1667 	struct request *req = bd->rq;
1668 	struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
1669 	blk_status_t ret;
1670 
1671 	/*
1672 	 * We should not need to do this, but we're still using this to
1673 	 * ensure we can drain requests on a dying queue.
1674 	 */
1675 	if (unlikely(!test_bit(NVMEQ_ENABLED, &nvmeq->flags)))
1676 		return BLK_STS_IOERR;
1677 
1678 	if (unlikely(!nvme_check_ready(&dev->ctrl, req, true)))
1679 		return nvme_fail_nonready_command(&dev->ctrl, req);
1680 
1681 	ret = nvme_prep_rq(req);
1682 	if (unlikely(ret))
1683 		return ret;
1684 	spin_lock(&nvmeq->sq_lock);
1685 	nvme_sq_copy_cmd(nvmeq, &iod->cmd);
1686 	nvme_write_sq_db(nvmeq, bd->last);
1687 	spin_unlock(&nvmeq->sq_lock);
1688 	return BLK_STS_OK;
1689 }
1690 
1691 static void nvme_submit_cmds(struct nvme_queue *nvmeq, struct rq_list *rqlist)
1692 {
1693 	struct request *req;
1694 
1695 	if (rq_list_empty(rqlist))
1696 		return;
1697 
1698 	spin_lock(&nvmeq->sq_lock);
1699 	while ((req = rq_list_pop(rqlist))) {
1700 		struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
1701 
1702 		nvme_sq_copy_cmd(nvmeq, &iod->cmd);
1703 	}
1704 	nvme_write_sq_db(nvmeq, true);
1705 	spin_unlock(&nvmeq->sq_lock);
1706 }
1707 
1708 static bool nvme_prep_rq_batch(struct nvme_queue *nvmeq, struct request *req)
1709 {
1710 	/*
1711 	 * We should not need to do this, but we're still using this to
1712 	 * ensure we can drain requests on a dying queue.
1713 	 */
1714 	if (unlikely(!test_bit(NVMEQ_ENABLED, &nvmeq->flags)))
1715 		return false;
1716 	if (unlikely(!nvme_check_ready(&nvmeq->dev->ctrl, req, true)))
1717 		return false;
1718 
1719 	return nvme_prep_rq(req) == BLK_STS_OK;
1720 }
1721 
1722 static void nvme_queue_rqs(struct rq_list *rqlist)
1723 {
1724 	struct rq_list submit_list = { };
1725 	struct rq_list requeue_list = { };
1726 	struct nvme_queue *nvmeq = NULL;
1727 	struct request *req;
1728 
1729 	while ((req = rq_list_pop(rqlist))) {
1730 		if (nvmeq && nvmeq != req->mq_hctx->driver_data)
1731 			nvme_submit_cmds(nvmeq, &submit_list);
1732 		nvmeq = req->mq_hctx->driver_data;
1733 
1734 		if (nvme_prep_rq_batch(nvmeq, req))
1735 			rq_list_add_tail(&submit_list, req);
1736 		else
1737 			rq_list_add_tail(&requeue_list, req);
1738 	}
1739 
1740 	if (nvmeq)
1741 		nvme_submit_cmds(nvmeq, &submit_list);
1742 	*rqlist = requeue_list;
1743 }
1744 
1745 static __always_inline void nvme_pci_unmap_rq(struct request *req)
1746 {
1747 	if (blk_integrity_rq(req))
1748 		nvme_unmap_metadata(req);
1749 	if (blk_rq_nr_phys_segments(req))
1750 		nvme_unmap_data(req);
1751 }
1752 
1753 static void nvme_pci_complete_rq(struct request *req)
1754 {
1755 	nvme_pci_unmap_rq(req);
1756 	nvme_complete_rq(req);
1757 }
1758 
1759 static void nvme_pci_complete_batch(struct io_comp_batch *iob)
1760 {
1761 	nvme_complete_batch(iob, nvme_pci_unmap_rq);
1762 }
1763 
1764 /* We read the CQE phase first to check if the rest of the entry is valid */
1765 static inline bool nvme_cqe_pending(struct nvme_queue *nvmeq)
1766 {
1767 	struct nvme_completion *hcqe = &nvmeq->cqes[nvmeq->cq_head];
1768 
1769 	return (le16_to_cpu(READ_ONCE(hcqe->status)) & 1) == nvmeq->cq_phase;
1770 }
1771 
1772 static inline void nvme_ring_cq_doorbell(struct nvme_queue *nvmeq)
1773 {
1774 	u16 head = nvmeq->cq_head;
1775 
1776 	if (nvme_dbbuf_update_and_check_event(head, nvmeq->dbbuf_cq_db,
1777 					      nvmeq->dbbuf_cq_ei))
1778 		writel(head, nvmeq->q_db + nvmeq->dev->db_stride);
1779 }
1780 
1781 static inline struct blk_mq_tags *nvme_queue_tagset(struct nvme_queue *nvmeq)
1782 {
1783 	if (!nvmeq->qid)
1784 		return nvmeq->dev->admin_tagset.tags[0];
1785 	return nvmeq->dev->tagset.tags[nvmeq->qid - 1];
1786 }
1787 
1788 static inline void nvme_handle_cqe(struct nvme_queue *nvmeq,
1789 				   struct io_comp_batch *iob, u16 idx)
1790 {
1791 	struct nvme_completion *cqe = &nvmeq->cqes[idx];
1792 	__u16 command_id = READ_ONCE(cqe->command_id);
1793 	struct request *req;
1794 
1795 	/*
1796 	 * AEN requests are special as they don't time out and can
1797 	 * survive any kind of queue freeze and often don't respond to
1798 	 * aborts.  We don't even bother to allocate a struct request
1799 	 * for them but rather special case them here.
1800 	 */
1801 	if (unlikely(nvme_is_aen_req(nvmeq->qid, command_id))) {
1802 		nvme_complete_async_event(&nvmeq->dev->ctrl,
1803 				cqe->status, &cqe->result);
1804 		return;
1805 	}
1806 
1807 	req = nvme_find_rq(nvme_queue_tagset(nvmeq), command_id);
1808 	if (unlikely(!req)) {
1809 		dev_warn_ratelimited(nvmeq->dev->ctrl.device,
1810 				     "invalid id %d completed on queue %d\n",
1811 				     command_id, le16_to_cpu(cqe->sq_id));
1812 		return;
1813 	}
1814 
1815 	/*
1816 	 * Tracing only; annotate a lockless snapshot of nvmeq->sq_tail using
1817 	 * data_race(). This would also help suppress context analysis warning
1818 	 * while accessing nvmeq->sq_tail without acquiring ->sq_lock.
1819 	 */
1820 	trace_nvme_sq(req, cqe->sq_head, data_race(nvmeq->sq_tail));
1821 	if (!nvme_try_complete_req(req, cqe->status, cqe->result) &&
1822 	    !blk_mq_add_to_batch(req, iob,
1823 				 nvme_req(req)->status != NVME_SC_SUCCESS,
1824 				 nvme_pci_complete_batch))
1825 		nvme_pci_complete_rq(req);
1826 }
1827 
1828 static inline void nvme_update_cq_head(struct nvme_queue *nvmeq)
1829 {
1830 	u32 tmp = nvmeq->cq_head + 1;
1831 
1832 	if (tmp == nvmeq->q_depth) {
1833 		nvmeq->cq_head = 0;
1834 		nvmeq->cq_phase ^= 1;
1835 	} else {
1836 		nvmeq->cq_head = tmp;
1837 	}
1838 }
1839 
1840 static inline bool nvme_poll_cq(struct nvme_queue *nvmeq,
1841 			        struct io_comp_batch *iob)
1842 {
1843 	bool found = false;
1844 
1845 	while (nvme_cqe_pending(nvmeq)) {
1846 		found = true;
1847 		/*
1848 		 * load-load control dependency between phase and the rest of
1849 		 * the cqe requires a full read memory barrier
1850 		 */
1851 		dma_rmb();
1852 		nvme_handle_cqe(nvmeq, iob, nvmeq->cq_head);
1853 		nvme_update_cq_head(nvmeq);
1854 	}
1855 
1856 	if (found)
1857 		nvme_ring_cq_doorbell(nvmeq);
1858 	return found;
1859 }
1860 
1861 static irqreturn_t nvme_irq(int irq, void *data)
1862 {
1863 	struct nvme_queue *nvmeq = data;
1864 	DEFINE_IO_COMP_BATCH(iob);
1865 
1866 	if (nvme_poll_cq(nvmeq, &iob)) {
1867 		if (!rq_list_empty(&iob.req_list))
1868 			nvme_pci_complete_batch(&iob);
1869 		return IRQ_HANDLED;
1870 	}
1871 	return IRQ_NONE;
1872 }
1873 
1874 static irqreturn_t nvme_irq_check(int irq, void *data)
1875 {
1876 	struct nvme_queue *nvmeq = data;
1877 
1878 	if (nvme_cqe_pending(nvmeq))
1879 		return IRQ_WAKE_THREAD;
1880 	return IRQ_NONE;
1881 }
1882 
1883 /*
1884  * Poll for completions for any interrupt driven queue
1885  * Can be called from any context.
1886  */
1887 static void nvme_poll_irqdisable(struct nvme_queue *nvmeq)
1888 {
1889 	struct pci_dev *pdev = to_pci_dev(nvmeq->dev->dev);
1890 	int irq;
1891 
1892 	WARN_ON_ONCE(test_bit(NVMEQ_POLLED, &nvmeq->flags));
1893 
1894 	irq = pci_irq_vector(pdev, nvmeq->cq_vector);
1895 	disable_irq(irq);
1896 	spin_lock(&nvmeq->cq_poll_lock);
1897 	nvme_poll_cq(nvmeq, NULL);
1898 	spin_unlock(&nvmeq->cq_poll_lock);
1899 	enable_irq(irq);
1900 }
1901 
1902 static int nvme_poll(struct blk_mq_hw_ctx *hctx, struct io_comp_batch *iob)
1903 {
1904 	struct nvme_queue *nvmeq = hctx->driver_data;
1905 	bool found;
1906 
1907 	if (!test_bit(NVMEQ_POLLED, &nvmeq->flags) ||
1908 	    !nvme_cqe_pending(nvmeq))
1909 		return 0;
1910 
1911 	spin_lock(&nvmeq->cq_poll_lock);
1912 	found = nvme_poll_cq(nvmeq, iob);
1913 	spin_unlock(&nvmeq->cq_poll_lock);
1914 
1915 	return found;
1916 }
1917 
1918 static void nvme_pci_submit_async_event(struct nvme_ctrl *ctrl)
1919 {
1920 	struct nvme_dev *dev = to_nvme_dev(ctrl);
1921 	struct nvme_queue *nvmeq = &dev->queues[0];
1922 	struct nvme_command c = { };
1923 
1924 	c.common.opcode = nvme_admin_async_event;
1925 	c.common.command_id = NVME_AQ_BLK_MQ_DEPTH;
1926 
1927 	spin_lock(&nvmeq->sq_lock);
1928 	nvme_sq_copy_cmd(nvmeq, &c);
1929 	nvme_write_sq_db(nvmeq, true);
1930 	spin_unlock(&nvmeq->sq_lock);
1931 }
1932 
1933 static int nvme_pci_subsystem_reset(struct nvme_ctrl *ctrl)
1934 {
1935 	struct nvme_dev *dev = to_nvme_dev(ctrl);
1936 	int ret = 0;
1937 
1938 	/*
1939 	 * Taking the shutdown_lock ensures the BAR mapping is not being
1940 	 * altered by reset_work. Holding this lock before the RESETTING state
1941 	 * change, if successful, also ensures nvme_remove won't be able to
1942 	 * proceed to iounmap until we're done.
1943 	 */
1944 	mutex_lock(&dev->shutdown_lock);
1945 	if (!dev->bar_mapped_size) {
1946 		ret = -ENODEV;
1947 		goto unlock;
1948 	}
1949 
1950 	if (!nvme_change_ctrl_state(ctrl, NVME_CTRL_RESETTING)) {
1951 		ret = -EBUSY;
1952 		goto unlock;
1953 	}
1954 
1955 	writel(NVME_SUBSYS_RESET, dev->bar + NVME_REG_NSSR);
1956 
1957 	if (!nvme_change_ctrl_state(ctrl, NVME_CTRL_CONNECTING) ||
1958 	    !nvme_change_ctrl_state(ctrl, NVME_CTRL_LIVE))
1959 		goto unlock;
1960 
1961 	/*
1962 	 * Read controller status to flush the previous write and trigger a
1963 	 * pcie read error.
1964 	 */
1965 	readl(dev->bar + NVME_REG_CSTS);
1966 unlock:
1967 	mutex_unlock(&dev->shutdown_lock);
1968 	return ret;
1969 }
1970 
1971 static int adapter_delete_queue(struct nvme_dev *dev, u8 opcode, u16 id)
1972 {
1973 	struct nvme_command c = { };
1974 
1975 	c.delete_queue.opcode = opcode;
1976 	c.delete_queue.qid = cpu_to_le16(id);
1977 
1978 	return nvme_submit_sync_cmd(dev->ctrl.admin_q, &c, NULL, 0);
1979 }
1980 
1981 static int adapter_alloc_cq(struct nvme_dev *dev, u16 qid,
1982 		struct nvme_queue *nvmeq, s16 vector)
1983 {
1984 	struct nvme_command c = { };
1985 	int flags = NVME_QUEUE_PHYS_CONTIG;
1986 
1987 	if (!test_bit(NVMEQ_POLLED, &nvmeq->flags))
1988 		flags |= NVME_CQ_IRQ_ENABLED;
1989 
1990 	/*
1991 	 * Note: we (ab)use the fact that the prp fields survive if no data
1992 	 * is attached to the request.
1993 	 */
1994 	c.create_cq.opcode = nvme_admin_create_cq;
1995 	c.create_cq.prp1 = cpu_to_le64(nvmeq->cq_dma_addr);
1996 	c.create_cq.cqid = cpu_to_le16(qid);
1997 	c.create_cq.qsize = cpu_to_le16(nvmeq->q_depth - 1);
1998 	c.create_cq.cq_flags = cpu_to_le16(flags);
1999 	c.create_cq.irq_vector = cpu_to_le16(vector);
2000 
2001 	return nvme_submit_sync_cmd(dev->ctrl.admin_q, &c, NULL, 0);
2002 }
2003 
2004 static int adapter_alloc_sq(struct nvme_dev *dev, u16 qid,
2005 						struct nvme_queue *nvmeq)
2006 {
2007 	struct nvme_ctrl *ctrl = &dev->ctrl;
2008 	struct nvme_command c = { };
2009 	int flags = NVME_QUEUE_PHYS_CONTIG;
2010 
2011 	/*
2012 	 * Some drives have a bug that auto-enables WRRU if MEDIUM isn't
2013 	 * set. Since URGENT priority is zeroes, it makes all queues
2014 	 * URGENT.
2015 	 */
2016 	if (ctrl->quirks & NVME_QUIRK_MEDIUM_PRIO_SQ)
2017 		flags |= NVME_SQ_PRIO_MEDIUM;
2018 
2019 	/*
2020 	 * Note: we (ab)use the fact that the prp fields survive if no data
2021 	 * is attached to the request.
2022 	 */
2023 	c.create_sq.opcode = nvme_admin_create_sq;
2024 	c.create_sq.prp1 = cpu_to_le64(nvmeq->sq_dma_addr);
2025 	c.create_sq.sqid = cpu_to_le16(qid);
2026 	c.create_sq.qsize = cpu_to_le16(nvmeq->q_depth - 1);
2027 	c.create_sq.sq_flags = cpu_to_le16(flags);
2028 	c.create_sq.cqid = cpu_to_le16(qid);
2029 
2030 	return nvme_submit_sync_cmd(dev->ctrl.admin_q, &c, NULL, 0);
2031 }
2032 
2033 static int adapter_delete_cq(struct nvme_dev *dev, u16 cqid)
2034 {
2035 	return adapter_delete_queue(dev, nvme_admin_delete_cq, cqid);
2036 }
2037 
2038 static int adapter_delete_sq(struct nvme_dev *dev, u16 sqid)
2039 {
2040 	return adapter_delete_queue(dev, nvme_admin_delete_sq, sqid);
2041 }
2042 
2043 static enum rq_end_io_ret abort_endio(struct request *req, blk_status_t error,
2044 				      const struct io_comp_batch *iob)
2045 {
2046 	struct nvme_queue *nvmeq = req->mq_hctx->driver_data;
2047 
2048 	dev_warn(nvmeq->dev->ctrl.device,
2049 		 "Abort status: 0x%x", nvme_req(req)->status);
2050 	atomic_inc(&nvmeq->dev->ctrl.abort_limit);
2051 	blk_mq_free_request(req);
2052 	return RQ_END_IO_NONE;
2053 }
2054 
2055 static bool nvme_should_reset(struct nvme_dev *dev, u32 csts)
2056 {
2057 	/* If true, indicates loss of adapter communication, possibly by a
2058 	 * NVMe Subsystem reset.
2059 	 */
2060 	bool nssro = dev->subsystem && (csts & NVME_CSTS_NSSRO);
2061 
2062 	/* If there is a reset/reinit ongoing, we shouldn't reset again. */
2063 	switch (nvme_ctrl_state(&dev->ctrl)) {
2064 	case NVME_CTRL_RESETTING:
2065 	case NVME_CTRL_CONNECTING:
2066 		return false;
2067 	default:
2068 		break;
2069 	}
2070 
2071 	/* We shouldn't reset unless the controller is on fatal error state
2072 	 * _or_ if we lost the communication with it.
2073 	 */
2074 	if (!(csts & NVME_CSTS_CFS) && !nssro)
2075 		return false;
2076 
2077 	return true;
2078 }
2079 
2080 static void nvme_warn_reset(struct nvme_dev *dev, u32 csts)
2081 {
2082 	/* Read a config register to help see what died. */
2083 	u16 pci_status;
2084 	int result;
2085 
2086 	result = pci_read_config_word(to_pci_dev(dev->dev), PCI_STATUS,
2087 				      &pci_status);
2088 	if (result == PCIBIOS_SUCCESSFUL)
2089 		dev_warn(dev->ctrl.device,
2090 			 "controller is down; will reset: CSTS=0x%x, PCI_STATUS=0x%hx\n",
2091 			 csts, pci_status);
2092 	else
2093 		dev_warn(dev->ctrl.device,
2094 			 "controller is down; will reset: CSTS=0x%x, PCI_STATUS read failed (%d)\n",
2095 			 csts, result);
2096 
2097 	if (csts != ~0)
2098 		return;
2099 
2100 	dev_warn(dev->ctrl.device,
2101 		 "Does your device have a faulty power saving mode enabled?\n");
2102 	dev_warn(dev->ctrl.device,
2103 		 "Try \"nvme_core.default_ps_max_latency_us=0 pcie_aspm=off pcie_port_pm=off\" and report a bug\n");
2104 }
2105 
2106 static enum blk_eh_timer_return nvme_timeout(struct request *req)
2107 {
2108 	struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
2109 	struct nvme_queue *nvmeq = req->mq_hctx->driver_data;
2110 	struct nvme_dev *dev = nvmeq->dev;
2111 	struct request *abort_req;
2112 	struct nvme_command cmd = { };
2113 	struct pci_dev *pdev = to_pci_dev(dev->dev);
2114 	u32 csts = readl(dev->bar + NVME_REG_CSTS);
2115 	u8 opcode;
2116 
2117 	/*
2118 	 * Shutdown the device immediately if we see it is disconnected. This
2119 	 * unblocks PCIe error handling if the nvme driver is waiting in
2120 	 * error_resume for a device that has been removed. We can't unbind the
2121 	 * driver while the driver's error callback is waiting to complete, so
2122 	 * we're relying on a timeout to break that deadlock if a removal
2123 	 * occurs while reset work is running.
2124 	 */
2125 	if (pci_dev_is_disconnected(pdev))
2126 		nvme_change_ctrl_state(&dev->ctrl, NVME_CTRL_DELETING);
2127 	if (nvme_state_terminal(&dev->ctrl))
2128 		goto disable;
2129 
2130 	/* If PCI error recovery process is happening, we cannot reset or
2131 	 * the recovery mechanism will surely fail.
2132 	 */
2133 	mb();
2134 	if (pci_channel_offline(pdev))
2135 		return BLK_EH_RESET_TIMER;
2136 
2137 	/*
2138 	 * Reset immediately if the controller is failed
2139 	 */
2140 	if (nvme_should_reset(dev, csts)) {
2141 		nvme_warn_reset(dev, csts);
2142 		goto disable;
2143 	}
2144 
2145 	/*
2146 	 * Did we miss an interrupt?
2147 	 */
2148 	if (test_bit(NVMEQ_POLLED, &nvmeq->flags))
2149 		nvme_poll(req->mq_hctx, NULL);
2150 	else
2151 		nvme_poll_irqdisable(nvmeq);
2152 
2153 	if (blk_mq_rq_state(req) != MQ_RQ_IN_FLIGHT) {
2154 		dev_warn(dev->ctrl.device,
2155 			 "I/O tag %d (%04x) QID %d timeout, completion polled\n",
2156 			 req->tag, nvme_cid(req), nvmeq->qid);
2157 		return BLK_EH_DONE;
2158 	}
2159 
2160 	/*
2161 	 * Shutdown immediately if controller times out while starting. The
2162 	 * reset work will see the pci device disabled when it gets the forced
2163 	 * cancellation error. All outstanding requests are completed on
2164 	 * shutdown, so we return BLK_EH_DONE.
2165 	 */
2166 	switch (nvme_ctrl_state(&dev->ctrl)) {
2167 	case NVME_CTRL_CONNECTING:
2168 		nvme_change_ctrl_state(&dev->ctrl, NVME_CTRL_DELETING);
2169 		fallthrough;
2170 	case NVME_CTRL_DELETING:
2171 		dev_warn_ratelimited(dev->ctrl.device,
2172 			 "I/O tag %d (%04x) QID %d timeout, disable controller\n",
2173 			 req->tag, nvme_cid(req), nvmeq->qid);
2174 		nvme_req(req)->flags |= NVME_REQ_CANCELLED;
2175 		nvme_dev_disable(dev, true);
2176 		return BLK_EH_DONE;
2177 	case NVME_CTRL_RESETTING:
2178 		return BLK_EH_RESET_TIMER;
2179 	default:
2180 		break;
2181 	}
2182 
2183 	/*
2184 	 * Shutdown the controller immediately and schedule a reset if the
2185 	 * command was already aborted once before and still hasn't been
2186 	 * returned to the driver, or if this is the admin queue.
2187 	 */
2188 	opcode = nvme_req(req)->cmd->common.opcode;
2189 	if (!nvmeq->qid || (iod->flags & IOD_ABORTED)) {
2190 		dev_warn(dev->ctrl.device,
2191 			 "I/O tag %d (%04x) opcode %#x (%s) QID %d timeout, reset controller\n",
2192 			 req->tag, nvme_cid(req), opcode,
2193 			 nvme_opcode_str(nvmeq->qid, opcode), nvmeq->qid);
2194 		nvme_req(req)->flags |= NVME_REQ_CANCELLED;
2195 		goto disable;
2196 	}
2197 
2198 	if (atomic_dec_return(&dev->ctrl.abort_limit) < 0) {
2199 		atomic_inc(&dev->ctrl.abort_limit);
2200 		return BLK_EH_RESET_TIMER;
2201 	}
2202 	iod->flags |= IOD_ABORTED;
2203 
2204 	cmd.abort.opcode = nvme_admin_abort_cmd;
2205 	cmd.abort.cid = nvme_cid(req);
2206 	cmd.abort.sqid = cpu_to_le16(nvmeq->qid);
2207 
2208 	dev_warn(nvmeq->dev->ctrl.device,
2209 		 "I/O tag %d (%04x) opcode %#x (%s) QID %d timeout, aborting req_op:%s(%u) size:%u\n",
2210 		 req->tag, nvme_cid(req), opcode, nvme_get_opcode_str(opcode),
2211 		 nvmeq->qid, blk_op_str(req_op(req)), req_op(req),
2212 		 blk_rq_bytes(req));
2213 
2214 	abort_req = blk_mq_alloc_request(dev->ctrl.admin_q, nvme_req_op(&cmd),
2215 					 BLK_MQ_REQ_NOWAIT);
2216 	if (IS_ERR(abort_req)) {
2217 		atomic_inc(&dev->ctrl.abort_limit);
2218 		return BLK_EH_RESET_TIMER;
2219 	}
2220 	nvme_init_request(abort_req, &cmd);
2221 
2222 	abort_req->end_io = abort_endio;
2223 	abort_req->end_io_data = NULL;
2224 	blk_execute_rq_nowait(abort_req, false);
2225 
2226 	/*
2227 	 * The aborted req will be completed on receiving the abort req.
2228 	 * We enable the timer again. If hit twice, it'll cause a device reset,
2229 	 * as the device then is in a faulty state.
2230 	 */
2231 	return BLK_EH_RESET_TIMER;
2232 
2233 disable:
2234 	if (!nvme_change_ctrl_state(&dev->ctrl, NVME_CTRL_RESETTING)) {
2235 		if (nvme_state_terminal(&dev->ctrl))
2236 			nvme_dev_disable(dev, true);
2237 		return BLK_EH_DONE;
2238 	}
2239 
2240 	nvme_dev_disable(dev, false);
2241 	if (nvme_try_sched_reset(&dev->ctrl))
2242 		nvme_unquiesce_io_queues(&dev->ctrl);
2243 	return BLK_EH_DONE;
2244 }
2245 
2246 static void nvme_free_queue(struct nvme_queue *nvmeq)
2247 	__context_unsafe(/* frees queue which is no longer in use */)
2248 {
2249 	dma_free_coherent(nvmeq->dev->dev, CQ_SIZE(nvmeq),
2250 				(void *)nvmeq->cqes, nvmeq->cq_dma_addr);
2251 	if (!nvmeq->sq_cmds)
2252 		return;
2253 
2254 	if (test_and_clear_bit(NVMEQ_SQ_CMB, &nvmeq->flags)) {
2255 		pci_free_p2pmem(to_pci_dev(nvmeq->dev->dev),
2256 				nvmeq->sq_cmds, SQ_SIZE(nvmeq));
2257 	} else {
2258 		dma_free_coherent(nvmeq->dev->dev, SQ_SIZE(nvmeq),
2259 				nvmeq->sq_cmds, nvmeq->sq_dma_addr);
2260 	}
2261 }
2262 
2263 static void nvme_free_queues(struct nvme_dev *dev, int lowest)
2264 {
2265 	int i;
2266 
2267 	for (i = dev->ctrl.queue_count - 1; i >= lowest; i--) {
2268 		dev->ctrl.queue_count--;
2269 		nvme_free_queue(&dev->queues[i]);
2270 	}
2271 }
2272 
2273 static void nvme_suspend_queue(struct nvme_dev *dev, unsigned int qid)
2274 {
2275 	struct nvme_queue *nvmeq = &dev->queues[qid];
2276 
2277 	if (!test_and_clear_bit(NVMEQ_ENABLED, &nvmeq->flags))
2278 		return;
2279 
2280 	/* ensure that nvme_queue_rq() sees NVMEQ_ENABLED cleared */
2281 	mb();
2282 
2283 	nvmeq->dev->online_queues--;
2284 	if (!nvmeq->qid && nvmeq->dev->ctrl.admin_q)
2285 		nvme_quiesce_admin_queue(&nvmeq->dev->ctrl);
2286 	if (!test_and_clear_bit(NVMEQ_POLLED, &nvmeq->flags))
2287 		pci_free_irq(to_pci_dev(dev->dev), nvmeq->cq_vector, nvmeq);
2288 }
2289 
2290 static void nvme_suspend_io_queues(struct nvme_dev *dev)
2291 {
2292 	int i;
2293 
2294 	for (i = dev->ctrl.queue_count - 1; i > 0; i--)
2295 		nvme_suspend_queue(dev, i);
2296 }
2297 
2298 /*
2299  * Called only on a device that has been disabled and after all other threads
2300  * that can check this device's completion queues have synced, except
2301  * nvme_poll(). This is the last chance for the driver to see a natural
2302  * completion before nvme_cancel_request() terminates all incomplete requests.
2303  */
2304 static void nvme_reap_pending_cqes(struct nvme_dev *dev)
2305 {
2306 	int i;
2307 
2308 	for (i = dev->ctrl.queue_count - 1; i > 0; i--) {
2309 		spin_lock(&dev->queues[i].cq_poll_lock);
2310 		nvme_poll_cq(&dev->queues[i], NULL);
2311 		spin_unlock(&dev->queues[i].cq_poll_lock);
2312 	}
2313 }
2314 
2315 static int nvme_cmb_qdepth(struct nvme_dev *dev, int nr_io_queues,
2316 				int entry_size)
2317 {
2318 	int q_depth = dev->q_depth;
2319 	unsigned q_size_aligned = roundup(q_depth * entry_size,
2320 					  NVME_CTRL_PAGE_SIZE);
2321 
2322 	if (q_size_aligned * nr_io_queues > dev->cmb_size) {
2323 		u64 mem_per_q = div_u64(dev->cmb_size, nr_io_queues);
2324 
2325 		mem_per_q = round_down(mem_per_q, NVME_CTRL_PAGE_SIZE);
2326 		q_depth = div_u64(mem_per_q, entry_size);
2327 
2328 		/*
2329 		 * Ensure the reduced q_depth is above some threshold where it
2330 		 * would be better to map queues in system memory with the
2331 		 * original depth
2332 		 */
2333 		if (q_depth < 64)
2334 			return -ENOMEM;
2335 	}
2336 
2337 	return q_depth;
2338 }
2339 
2340 static int nvme_alloc_sq_cmds(struct nvme_dev *dev, struct nvme_queue *nvmeq,
2341 				int qid)
2342 	__context_unsafe(/* safe to allocate sq_cmds without any protection */)
2343 {
2344 	struct pci_dev *pdev = to_pci_dev(dev->dev);
2345 
2346 	if (qid && dev->cmb_use_sqes && (dev->cmbsz & NVME_CMBSZ_SQS)) {
2347 		nvmeq->sq_cmds = pci_alloc_p2pmem(pdev, SQ_SIZE(nvmeq));
2348 		if (nvmeq->sq_cmds) {
2349 			nvmeq->sq_dma_addr = pci_p2pmem_virt_to_bus(pdev,
2350 							nvmeq->sq_cmds);
2351 			if (nvmeq->sq_dma_addr) {
2352 				set_bit(NVMEQ_SQ_CMB, &nvmeq->flags);
2353 				return 0;
2354 			}
2355 
2356 			pci_free_p2pmem(pdev, nvmeq->sq_cmds, SQ_SIZE(nvmeq));
2357 		}
2358 	}
2359 
2360 	nvmeq->sq_cmds = dma_alloc_coherent(dev->dev, SQ_SIZE(nvmeq),
2361 				&nvmeq->sq_dma_addr, GFP_KERNEL);
2362 	if (!nvmeq->sq_cmds)
2363 		return -ENOMEM;
2364 	return 0;
2365 }
2366 
2367 static int nvme_alloc_queue(struct nvme_dev *dev, int qid, int depth)
2368 {
2369 	struct nvme_queue *nvmeq = &dev->queues[qid];
2370 
2371 	if (dev->ctrl.queue_count > qid)
2372 		return 0;
2373 
2374 	nvmeq->sqes = qid ? dev->io_sqes : NVME_ADM_SQES;
2375 	nvmeq->q_depth = depth;
2376 	nvmeq->cqes = dma_alloc_coherent(dev->dev, CQ_SIZE(nvmeq),
2377 					 &nvmeq->cq_dma_addr, GFP_KERNEL);
2378 	if (!nvmeq->cqes)
2379 		goto free_nvmeq;
2380 
2381 	if (nvme_alloc_sq_cmds(dev, nvmeq, qid))
2382 		goto free_cqdma;
2383 
2384 	nvmeq->dev = dev;
2385 	spin_lock_init(&nvmeq->sq_lock);
2386 	spin_lock_init(&nvmeq->cq_poll_lock);
2387 	nvmeq->cq_head = 0;
2388 	nvmeq->cq_phase = 1;
2389 	nvmeq->q_db = &dev->dbs[qid * 2 * dev->db_stride];
2390 	nvmeq->qid = qid;
2391 	dev->ctrl.queue_count++;
2392 
2393 	return 0;
2394 
2395  free_cqdma:
2396 	dma_free_coherent(dev->dev, CQ_SIZE(nvmeq), (void *)nvmeq->cqes,
2397 			  nvmeq->cq_dma_addr);
2398  free_nvmeq:
2399 	return -ENOMEM;
2400 }
2401 
2402 static int queue_request_irq(struct nvme_queue *nvmeq)
2403 {
2404 	struct pci_dev *pdev = to_pci_dev(nvmeq->dev->dev);
2405 	int nr = nvmeq->dev->ctrl.instance;
2406 
2407 	if (use_threaded_interrupts) {
2408 		return pci_request_irq(pdev, nvmeq->cq_vector, nvme_irq_check,
2409 				nvme_irq, nvmeq, "nvme%dq%d", nr, nvmeq->qid);
2410 	} else {
2411 		return pci_request_irq(pdev, nvmeq->cq_vector, nvme_irq,
2412 				NULL, nvmeq, "nvme%dq%d", nr, nvmeq->qid);
2413 	}
2414 }
2415 
2416 static void nvme_init_queue(struct nvme_queue *nvmeq, u16 qid)
2417 	__context_unsafe(/* initialize unpublished/lock-guarded variables */)
2418 {
2419 	struct nvme_dev *dev = nvmeq->dev;
2420 
2421 	nvmeq->sq_tail = 0;
2422 	nvmeq->last_sq_tail = 0;
2423 	nvmeq->cq_head = 0;
2424 	nvmeq->cq_phase = 1;
2425 	nvmeq->q_db = &dev->dbs[qid * 2 * dev->db_stride];
2426 	memset((void *)nvmeq->cqes, 0, CQ_SIZE(nvmeq));
2427 	nvme_dbbuf_init(dev, nvmeq, qid);
2428 	dev->online_queues++;
2429 	wmb(); /* ensure the first interrupt sees the initialization */
2430 }
2431 
2432 /*
2433  * Try getting shutdown_lock while setting up IO queues.
2434  */
2435 static int nvme_setup_io_queues_trylock(struct nvme_dev *dev)
2436 	__cond_acquires(0, &dev->shutdown_lock)
2437 {
2438 	/*
2439 	 * Give up if the lock is being held by nvme_dev_disable.
2440 	 */
2441 	if (!mutex_trylock(&dev->shutdown_lock))
2442 		return -ENODEV;
2443 
2444 	/*
2445 	 * Controller is in wrong state, fail early.
2446 	 */
2447 	if (nvme_ctrl_state(&dev->ctrl) != NVME_CTRL_CONNECTING) {
2448 		mutex_unlock(&dev->shutdown_lock);
2449 		return -ENODEV;
2450 	}
2451 
2452 	return 0;
2453 }
2454 
2455 static int nvme_create_queue(struct nvme_queue *nvmeq, int qid, bool polled)
2456 {
2457 	struct nvme_dev *dev = nvmeq->dev;
2458 	int result;
2459 	u16 vector = 0;
2460 
2461 	/*
2462 	 * A queue's vector matches the queue identifier unless the controller
2463 	 * has only one vector available.
2464 	 */
2465 	if (!polled)
2466 		vector = dev->num_vecs == 1 ? 0 : qid;
2467 	else
2468 		set_bit(NVMEQ_POLLED, &nvmeq->flags);
2469 
2470 	result = adapter_alloc_cq(dev, qid, nvmeq, vector);
2471 	if (result)
2472 		return result;
2473 
2474 	result = adapter_alloc_sq(dev, qid, nvmeq);
2475 	if (result < 0)
2476 		return result;
2477 	if (result)
2478 		goto release_cq;
2479 
2480 	nvmeq->cq_vector = vector;
2481 
2482 	result = nvme_setup_io_queues_trylock(dev);
2483 	if (result)
2484 		return result;
2485 	nvme_init_queue(nvmeq, qid);
2486 	if (!polled) {
2487 		result = queue_request_irq(nvmeq);
2488 		if (result < 0)
2489 			goto release_sq;
2490 	}
2491 
2492 	set_bit(NVMEQ_ENABLED, &nvmeq->flags);
2493 	mutex_unlock(&dev->shutdown_lock);
2494 	return result;
2495 
2496 release_sq:
2497 	dev->online_queues--;
2498 	mutex_unlock(&dev->shutdown_lock);
2499 	adapter_delete_sq(dev, qid);
2500 release_cq:
2501 	adapter_delete_cq(dev, qid);
2502 	return result;
2503 }
2504 
2505 static const struct blk_mq_ops nvme_mq_admin_ops = {
2506 	.queue_rq	= nvme_queue_rq,
2507 	.complete	= nvme_pci_complete_rq,
2508 	.commit_rqs	= nvme_commit_rqs,
2509 	.init_hctx	= nvme_admin_init_hctx,
2510 	.init_request	= nvme_pci_init_request,
2511 	.timeout	= nvme_timeout,
2512 };
2513 
2514 static const struct blk_mq_ops nvme_mq_ops = {
2515 	.queue_rq	= nvme_queue_rq,
2516 	.queue_rqs	= nvme_queue_rqs,
2517 	.complete	= nvme_pci_complete_rq,
2518 	.commit_rqs	= nvme_commit_rqs,
2519 	.init_hctx	= nvme_init_hctx,
2520 	.init_request	= nvme_pci_init_request,
2521 	.map_queues	= nvme_pci_map_queues,
2522 	.timeout	= nvme_timeout,
2523 	.poll		= nvme_poll,
2524 };
2525 
2526 static void nvme_dev_remove_admin(struct nvme_dev *dev)
2527 {
2528 	if (dev->ctrl.admin_q && !blk_queue_dying(dev->ctrl.admin_q)) {
2529 		/*
2530 		 * If the controller was reset during removal, it's possible
2531 		 * user requests may be waiting on a stopped queue. Start the
2532 		 * queue to flush these to completion.
2533 		 */
2534 		nvme_unquiesce_admin_queue(&dev->ctrl);
2535 		nvme_remove_admin_tag_set(&dev->ctrl);
2536 	}
2537 }
2538 
2539 static unsigned long db_bar_size(struct nvme_dev *dev, unsigned nr_io_queues)
2540 {
2541 	return NVME_REG_DBS + ((nr_io_queues + 1) * 8 * dev->db_stride);
2542 }
2543 
2544 static int nvme_remap_bar(struct nvme_dev *dev, unsigned long size)
2545 {
2546 	struct pci_dev *pdev = to_pci_dev(dev->dev);
2547 
2548 	if (size <= dev->bar_mapped_size)
2549 		return 0;
2550 	if (size > pci_resource_len(pdev, 0))
2551 		return -ENOMEM;
2552 	if (dev->bar)
2553 		iounmap(dev->bar);
2554 	dev->bar = ioremap(pci_resource_start(pdev, 0), size);
2555 	if (!dev->bar) {
2556 		dev->bar_mapped_size = 0;
2557 		return -ENOMEM;
2558 	}
2559 	dev->bar_mapped_size = size;
2560 	dev->dbs = dev->bar + NVME_REG_DBS;
2561 
2562 	return 0;
2563 }
2564 
2565 static int nvme_pci_configure_admin_queue(struct nvme_dev *dev)
2566 {
2567 	int result;
2568 	u32 aqa;
2569 	struct nvme_queue *nvmeq;
2570 
2571 	result = nvme_remap_bar(dev, db_bar_size(dev, 0));
2572 	if (result < 0)
2573 		return result;
2574 
2575 	dev->subsystem = readl(dev->bar + NVME_REG_VS) >= NVME_VS(1, 1, 0) ?
2576 				NVME_CAP_NSSRC(dev->ctrl.cap) : 0;
2577 
2578 	if (dev->subsystem &&
2579 	    (readl(dev->bar + NVME_REG_CSTS) & NVME_CSTS_NSSRO))
2580 		writel(NVME_CSTS_NSSRO, dev->bar + NVME_REG_CSTS);
2581 
2582 	/*
2583 	 * If the device has been passed off to us in an enabled state, just
2584 	 * clear the enabled bit.  The spec says we should set the 'shutdown
2585 	 * notification bits', but doing so may cause the device to complete
2586 	 * commands to the admin queue ... and we don't know what memory that
2587 	 * might be pointing at!
2588 	 */
2589 	result = nvme_disable_ctrl(&dev->ctrl, false);
2590 	if (result < 0) {
2591 		struct pci_dev *pdev = to_pci_dev(dev->dev);
2592 
2593 		/*
2594 		 * The NVMe Controller Reset method did not get an expected
2595 		 * CSTS.RDY transition, so something with the device appears to
2596 		 * be stuck. Use the lower level and bigger hammer PCIe
2597 		 * Function Level Reset to attempt restoring the device to its
2598 		 * initial state, and try again.
2599 		 */
2600 		result = pcie_reset_flr(pdev, false);
2601 		if (result < 0)
2602 			return result;
2603 
2604 		pci_restore_state(pdev);
2605 		result = nvme_disable_ctrl(&dev->ctrl, false);
2606 		if (result < 0)
2607 			return result;
2608 
2609 		dev_info(dev->ctrl.device,
2610 			"controller reset completed after pcie flr\n");
2611 	}
2612 
2613 	result = nvme_alloc_queue(dev, 0, NVME_AQ_DEPTH);
2614 	if (result)
2615 		return result;
2616 
2617 	dev->ctrl.numa_node = dev_to_node(dev->dev);
2618 
2619 	nvmeq = &dev->queues[0];
2620 	aqa = nvmeq->q_depth - 1;
2621 	aqa |= aqa << 16;
2622 
2623 	writel(aqa, dev->bar + NVME_REG_AQA);
2624 	lo_hi_writeq(nvmeq->sq_dma_addr, dev->bar + NVME_REG_ASQ);
2625 	lo_hi_writeq(nvmeq->cq_dma_addr, dev->bar + NVME_REG_ACQ);
2626 
2627 	result = nvme_enable_ctrl(&dev->ctrl);
2628 	if (result)
2629 		return result;
2630 
2631 	nvmeq->cq_vector = 0;
2632 	nvme_init_queue(nvmeq, 0);
2633 	result = queue_request_irq(nvmeq);
2634 	if (result) {
2635 		dev->online_queues--;
2636 		nvme_disable_ctrl(&dev->ctrl, false);
2637 		return result;
2638 	}
2639 
2640 	set_bit(NVMEQ_ENABLED, &nvmeq->flags);
2641 	return result;
2642 }
2643 
2644 static int nvme_create_io_queues(struct nvme_dev *dev)
2645 {
2646 	unsigned i, max, rw_queues;
2647 	int ret = 0;
2648 
2649 	for (i = dev->ctrl.queue_count; i <= dev->max_qid; i++) {
2650 		if (nvme_alloc_queue(dev, i, dev->q_depth)) {
2651 			ret = -ENOMEM;
2652 			break;
2653 		}
2654 	}
2655 
2656 	max = min(dev->max_qid, dev->ctrl.queue_count - 1);
2657 	if (max != 1 && dev->io_queues[HCTX_TYPE_POLL]) {
2658 		rw_queues = dev->io_queues[HCTX_TYPE_DEFAULT] +
2659 				dev->io_queues[HCTX_TYPE_READ];
2660 	} else {
2661 		rw_queues = max;
2662 	}
2663 
2664 	for (i = dev->online_queues; i <= max; i++) {
2665 		bool polled = i > rw_queues;
2666 
2667 		ret = nvme_create_queue(&dev->queues[i], i, polled);
2668 		if (ret)
2669 			break;
2670 	}
2671 
2672 	/*
2673 	 * Ignore failing Create SQ/CQ commands, we can continue with less
2674 	 * than the desired amount of queues, and even a controller without
2675 	 * I/O queues can still be used to issue admin commands.  This might
2676 	 * be useful to upgrade a buggy firmware for example.
2677 	 */
2678 	return ret >= 0 ? 0 : ret;
2679 }
2680 
2681 static u64 nvme_cmb_size_unit(struct nvme_dev *dev)
2682 {
2683 	u8 szu = (dev->cmbsz >> NVME_CMBSZ_SZU_SHIFT) & NVME_CMBSZ_SZU_MASK;
2684 
2685 	return 1ULL << (12 + 4 * szu);
2686 }
2687 
2688 static u32 nvme_cmb_size(struct nvme_dev *dev)
2689 {
2690 	return (dev->cmbsz >> NVME_CMBSZ_SZ_SHIFT) & NVME_CMBSZ_SZ_MASK;
2691 }
2692 
2693 static void nvme_map_cmb(struct nvme_dev *dev)
2694 {
2695 	u64 size, offset;
2696 	resource_size_t bar_size;
2697 	struct pci_dev *pdev = to_pci_dev(dev->dev);
2698 	int bar;
2699 
2700 	if (dev->cmb_size)
2701 		return;
2702 
2703 	if (NVME_CAP_CMBS(dev->ctrl.cap))
2704 		writel(NVME_CMBMSC_CRE, dev->bar + NVME_REG_CMBMSC);
2705 
2706 	dev->cmbsz = readl(dev->bar + NVME_REG_CMBSZ);
2707 	if (!dev->cmbsz)
2708 		return;
2709 	dev->cmbloc = readl(dev->bar + NVME_REG_CMBLOC);
2710 
2711 	size = nvme_cmb_size_unit(dev) * nvme_cmb_size(dev);
2712 	offset = nvme_cmb_size_unit(dev) * NVME_CMB_OFST(dev->cmbloc);
2713 	bar = NVME_CMB_BIR(dev->cmbloc);
2714 	bar_size = pci_resource_len(pdev, bar);
2715 
2716 	if (offset > bar_size)
2717 		return;
2718 
2719 	/*
2720 	 * Controllers may support a CMB size larger than their BAR, for
2721 	 * example, due to being behind a bridge. Reduce the CMB to the
2722 	 * reported size of the BAR
2723 	 */
2724 	size = min(size, bar_size - offset);
2725 
2726 	if (!IS_ALIGNED(size, memremap_compat_align()) ||
2727 	    !IS_ALIGNED(pci_resource_start(pdev, bar),
2728 			memremap_compat_align()))
2729 		return;
2730 
2731 	/*
2732 	 * Tell the controller about the host side address mapping the CMB,
2733 	 * and enable CMB decoding for the NVMe 1.4+ scheme:
2734 	 */
2735 	if (NVME_CAP_CMBS(dev->ctrl.cap)) {
2736 		hi_lo_writeq(NVME_CMBMSC_CRE | NVME_CMBMSC_CMSE |
2737 			     (pci_bus_address(pdev, bar) + offset),
2738 			     dev->bar + NVME_REG_CMBMSC);
2739 	}
2740 
2741 	if (pci_p2pdma_add_resource(pdev, bar, size, offset)) {
2742 		dev_warn(dev->ctrl.device,
2743 			 "failed to register the CMB\n");
2744 		hi_lo_writeq(0, dev->bar + NVME_REG_CMBMSC);
2745 		return;
2746 	}
2747 
2748 	dev->cmb_size = size;
2749 	dev->cmb_use_sqes = use_cmb_sqes && (dev->cmbsz & NVME_CMBSZ_SQS);
2750 
2751 	if ((dev->cmbsz & (NVME_CMBSZ_WDS | NVME_CMBSZ_RDS)) ==
2752 			(NVME_CMBSZ_WDS | NVME_CMBSZ_RDS))
2753 		pci_p2pmem_publish(pdev, true);
2754 }
2755 
2756 static int nvme_set_host_mem(struct nvme_dev *dev, u32 bits)
2757 {
2758 	u32 host_mem_size = dev->host_mem_size >> NVME_CTRL_PAGE_SHIFT;
2759 	u64 dma_addr = dev->host_mem_descs_dma;
2760 	struct nvme_command c = { };
2761 	int ret;
2762 
2763 	c.features.opcode	= nvme_admin_set_features;
2764 	c.features.fid		= cpu_to_le32(NVME_FEAT_HOST_MEM_BUF);
2765 	c.features.dword11	= cpu_to_le32(bits);
2766 	c.features.dword12	= cpu_to_le32(host_mem_size);
2767 	c.features.dword13	= cpu_to_le32(lower_32_bits(dma_addr));
2768 	c.features.dword14	= cpu_to_le32(upper_32_bits(dma_addr));
2769 	c.features.dword15	= cpu_to_le32(dev->nr_host_mem_descs);
2770 
2771 	ret = nvme_submit_sync_cmd(dev->ctrl.admin_q, &c, NULL, 0);
2772 	if (ret) {
2773 		dev_warn(dev->ctrl.device,
2774 			 "failed to set host mem (err %d, flags %#x).\n",
2775 			 ret, bits);
2776 	} else
2777 		dev->hmb = bits & NVME_HOST_MEM_ENABLE;
2778 
2779 	return ret;
2780 }
2781 
2782 static void nvme_free_host_mem_multi(struct nvme_dev *dev)
2783 {
2784 	int i;
2785 
2786 	for (i = 0; i < dev->nr_host_mem_descs; i++) {
2787 		struct nvme_host_mem_buf_desc *desc = &dev->host_mem_descs[i];
2788 		size_t size = le32_to_cpu(desc->size) * NVME_CTRL_PAGE_SIZE;
2789 
2790 		dma_free_attrs(dev->dev, size, dev->host_mem_desc_bufs[i],
2791 			       le64_to_cpu(desc->addr),
2792 			       DMA_ATTR_NO_KERNEL_MAPPING | DMA_ATTR_NO_WARN);
2793 	}
2794 
2795 	kfree(dev->host_mem_desc_bufs);
2796 	dev->host_mem_desc_bufs = NULL;
2797 }
2798 
2799 static void nvme_free_host_mem(struct nvme_dev *dev)
2800 {
2801 	if (dev->hmb_sgt) {
2802 		dma_free_noncontiguous(dev->dev, dev->host_mem_size,
2803 				dev->hmb_sgt, DMA_BIDIRECTIONAL);
2804 		dev->hmb_sgt = NULL;
2805 	} else {
2806 		nvme_free_host_mem_multi(dev);
2807 	}
2808 
2809 	dma_free_coherent(dev->dev, dev->host_mem_descs_size,
2810 			dev->host_mem_descs, dev->host_mem_descs_dma);
2811 	dev->host_mem_descs = NULL;
2812 	dev->host_mem_descs_size = 0;
2813 	dev->nr_host_mem_descs = 0;
2814 }
2815 
2816 static int nvme_alloc_host_mem_single(struct nvme_dev *dev, u64 size)
2817 {
2818 	dev->hmb_sgt = dma_alloc_noncontiguous(dev->dev, size,
2819 				DMA_BIDIRECTIONAL, GFP_KERNEL, 0);
2820 	if (!dev->hmb_sgt)
2821 		return -ENOMEM;
2822 
2823 	dev->host_mem_descs = dma_alloc_coherent(dev->dev,
2824 			sizeof(*dev->host_mem_descs), &dev->host_mem_descs_dma,
2825 			GFP_KERNEL);
2826 	if (!dev->host_mem_descs) {
2827 		dma_free_noncontiguous(dev->dev, size, dev->hmb_sgt,
2828 				DMA_BIDIRECTIONAL);
2829 		dev->hmb_sgt = NULL;
2830 		return -ENOMEM;
2831 	}
2832 	dev->host_mem_size = size;
2833 	dev->host_mem_descs_size = sizeof(*dev->host_mem_descs);
2834 	dev->nr_host_mem_descs = 1;
2835 
2836 	dev->host_mem_descs[0].addr =
2837 		cpu_to_le64(dev->hmb_sgt->sgl->dma_address);
2838 	dev->host_mem_descs[0].size = cpu_to_le32(size / NVME_CTRL_PAGE_SIZE);
2839 	return 0;
2840 }
2841 
2842 static int nvme_alloc_host_mem_multi(struct nvme_dev *dev, u64 preferred,
2843 		u32 chunk_size)
2844 {
2845 	struct nvme_host_mem_buf_desc *descs;
2846 	u32 max_entries, len, descs_size;
2847 	dma_addr_t descs_dma;
2848 	int i = 0;
2849 	void **bufs;
2850 	u64 size, tmp;
2851 
2852 	tmp = (preferred + chunk_size - 1);
2853 	do_div(tmp, chunk_size);
2854 	max_entries = tmp;
2855 
2856 	if (dev->ctrl.hmmaxd && dev->ctrl.hmmaxd < max_entries)
2857 		max_entries = dev->ctrl.hmmaxd;
2858 
2859 	descs_size = max_entries * sizeof(*descs);
2860 	descs = dma_alloc_coherent(dev->dev, descs_size, &descs_dma,
2861 			GFP_KERNEL);
2862 	if (!descs)
2863 		goto out;
2864 
2865 	bufs = kzalloc_objs(*bufs, max_entries);
2866 	if (!bufs)
2867 		goto out_free_descs;
2868 
2869 	for (size = 0; size < preferred && i < max_entries; size += len) {
2870 		dma_addr_t dma_addr;
2871 
2872 		len = min_t(u64, chunk_size, preferred - size);
2873 		bufs[i] = dma_alloc_attrs(dev->dev, len, &dma_addr, GFP_KERNEL,
2874 				DMA_ATTR_NO_KERNEL_MAPPING | DMA_ATTR_NO_WARN);
2875 		if (!bufs[i])
2876 			break;
2877 
2878 		descs[i].addr = cpu_to_le64(dma_addr);
2879 		descs[i].size = cpu_to_le32(len / NVME_CTRL_PAGE_SIZE);
2880 		i++;
2881 	}
2882 
2883 	if (!size)
2884 		goto out_free_bufs;
2885 
2886 	dev->nr_host_mem_descs = i;
2887 	dev->host_mem_size = size;
2888 	dev->host_mem_descs = descs;
2889 	dev->host_mem_descs_dma = descs_dma;
2890 	dev->host_mem_descs_size = descs_size;
2891 	dev->host_mem_desc_bufs = bufs;
2892 	return 0;
2893 
2894 out_free_bufs:
2895 	kfree(bufs);
2896 out_free_descs:
2897 	dma_free_coherent(dev->dev, descs_size, descs, descs_dma);
2898 out:
2899 	dev->host_mem_descs = NULL;
2900 	return -ENOMEM;
2901 }
2902 
2903 static int nvme_alloc_host_mem(struct nvme_dev *dev, u64 min, u64 preferred)
2904 {
2905 	unsigned long dma_merge_boundary = dma_get_merge_boundary(dev->dev);
2906 	u64 min_chunk = min_t(u64, preferred, PAGE_SIZE * MAX_ORDER_NR_PAGES);
2907 	u64 hmminds = max_t(u32, dev->ctrl.hmminds * 4096, PAGE_SIZE * 2);
2908 	u64 chunk_size;
2909 
2910 	/*
2911 	 * If there is an IOMMU that can merge pages, try a virtually
2912 	 * non-contiguous allocation for a single segment first.
2913 	 */
2914 	if (dma_merge_boundary && (PAGE_SIZE & dma_merge_boundary) == 0) {
2915 		if (!nvme_alloc_host_mem_single(dev, preferred))
2916 			return 0;
2917 	}
2918 
2919 	/* start big and work our way down */
2920 	for (chunk_size = min_chunk; chunk_size >= hmminds; chunk_size /= 2) {
2921 		if (!nvme_alloc_host_mem_multi(dev, preferred, chunk_size)) {
2922 			if (!min || dev->host_mem_size >= min)
2923 				return 0;
2924 			nvme_free_host_mem(dev);
2925 		}
2926 	}
2927 
2928 	return -ENOMEM;
2929 }
2930 
2931 static int nvme_setup_host_mem(struct nvme_dev *dev)
2932 {
2933 	u64 max = (u64)max_host_mem_size_mb * SZ_1M;
2934 	u64 preferred = (u64)dev->ctrl.hmpre * 4096;
2935 	u64 min = (u64)dev->ctrl.hmmin * 4096;
2936 	u32 enable_bits = NVME_HOST_MEM_ENABLE;
2937 	int ret;
2938 
2939 	if (!dev->ctrl.hmpre)
2940 		return 0;
2941 
2942 	preferred = min(preferred, max);
2943 	if (min > max) {
2944 		dev_warn(dev->ctrl.device,
2945 			"min host memory (%lld MiB) above limit (%d MiB).\n",
2946 			min >> ilog2(SZ_1M), max_host_mem_size_mb);
2947 		nvme_free_host_mem(dev);
2948 		return 0;
2949 	}
2950 
2951 	/*
2952 	 * If we already have a buffer allocated check if we can reuse it.
2953 	 */
2954 	if (dev->host_mem_descs) {
2955 		if (dev->host_mem_size >= min)
2956 			enable_bits |= NVME_HOST_MEM_RETURN;
2957 		else
2958 			nvme_free_host_mem(dev);
2959 	}
2960 
2961 	if (!dev->host_mem_descs) {
2962 		if (nvme_alloc_host_mem(dev, min, preferred)) {
2963 			dev_warn(dev->ctrl.device,
2964 				"failed to allocate host memory buffer.\n");
2965 			return 0; /* controller must work without HMB */
2966 		}
2967 
2968 		dev_info(dev->ctrl.device,
2969 			"allocated %lld MiB host memory buffer (%u segment%s).\n",
2970 			dev->host_mem_size >> ilog2(SZ_1M),
2971 			dev->nr_host_mem_descs,
2972 			str_plural(dev->nr_host_mem_descs));
2973 	}
2974 
2975 	ret = nvme_set_host_mem(dev, enable_bits);
2976 	if (ret)
2977 		nvme_free_host_mem(dev);
2978 	return ret;
2979 }
2980 
2981 static ssize_t cmb_show(struct device *dev, struct device_attribute *attr,
2982 		char *buf)
2983 {
2984 	struct nvme_dev *ndev = to_nvme_dev(dev_get_drvdata(dev));
2985 
2986 	return sysfs_emit(buf, "cmbloc : 0x%08x\ncmbsz  : 0x%08x\n",
2987 		       ndev->cmbloc, ndev->cmbsz);
2988 }
2989 static DEVICE_ATTR_RO(cmb);
2990 
2991 static ssize_t cmbloc_show(struct device *dev, struct device_attribute *attr,
2992 		char *buf)
2993 {
2994 	struct nvme_dev *ndev = to_nvme_dev(dev_get_drvdata(dev));
2995 
2996 	return sysfs_emit(buf, "%u\n", ndev->cmbloc);
2997 }
2998 static DEVICE_ATTR_RO(cmbloc);
2999 
3000 static ssize_t cmbsz_show(struct device *dev, struct device_attribute *attr,
3001 		char *buf)
3002 {
3003 	struct nvme_dev *ndev = to_nvme_dev(dev_get_drvdata(dev));
3004 
3005 	return sysfs_emit(buf, "%u\n", ndev->cmbsz);
3006 }
3007 static DEVICE_ATTR_RO(cmbsz);
3008 
3009 static ssize_t hmb_show(struct device *dev, struct device_attribute *attr,
3010 			char *buf)
3011 {
3012 	struct nvme_dev *ndev = to_nvme_dev(dev_get_drvdata(dev));
3013 
3014 	return sysfs_emit(buf, "%d\n", ndev->hmb);
3015 }
3016 
3017 static ssize_t hmb_store(struct device *dev, struct device_attribute *attr,
3018 			 const char *buf, size_t count)
3019 {
3020 	struct nvme_dev *ndev = to_nvme_dev(dev_get_drvdata(dev));
3021 	bool new;
3022 	int ret;
3023 
3024 	if (kstrtobool(buf, &new) < 0)
3025 		return -EINVAL;
3026 
3027 	if (new == ndev->hmb)
3028 		return count;
3029 
3030 	if (new) {
3031 		ret = nvme_setup_host_mem(ndev);
3032 	} else {
3033 		ret = nvme_set_host_mem(ndev, 0);
3034 		if (!ret)
3035 			nvme_free_host_mem(ndev);
3036 	}
3037 
3038 	if (ret < 0)
3039 		return ret;
3040 
3041 	return count;
3042 }
3043 static DEVICE_ATTR_RW(hmb);
3044 
3045 static umode_t nvme_pci_attrs_are_visible(struct kobject *kobj,
3046 		struct attribute *a, int n)
3047 {
3048 	struct nvme_ctrl *ctrl =
3049 		dev_get_drvdata(container_of(kobj, struct device, kobj));
3050 	struct nvme_dev *dev = to_nvme_dev(ctrl);
3051 
3052 	if (a == &dev_attr_cmb.attr ||
3053 	    a == &dev_attr_cmbloc.attr ||
3054 	    a == &dev_attr_cmbsz.attr) {
3055 	    	if (!dev->cmbsz)
3056 			return 0;
3057 	}
3058 	if (a == &dev_attr_hmb.attr && !ctrl->hmpre)
3059 		return 0;
3060 
3061 	return a->mode;
3062 }
3063 
3064 static struct attribute *nvme_pci_attrs[] = {
3065 	&dev_attr_cmb.attr,
3066 	&dev_attr_cmbloc.attr,
3067 	&dev_attr_cmbsz.attr,
3068 	&dev_attr_hmb.attr,
3069 	NULL,
3070 };
3071 
3072 static const struct attribute_group nvme_pci_dev_attrs_group = {
3073 	.attrs		= nvme_pci_attrs,
3074 	.is_visible	= nvme_pci_attrs_are_visible,
3075 };
3076 
3077 static const struct attribute_group *nvme_pci_dev_attr_groups[] = {
3078 	&nvme_dev_attrs_group,
3079 	&nvme_pci_dev_attrs_group,
3080 	&nvme_dev_diag_attrs_group,
3081 	NULL,
3082 };
3083 
3084 static void nvme_update_attrs(struct nvme_dev *dev)
3085 {
3086 	sysfs_update_group(&dev->ctrl.device->kobj, &nvme_pci_dev_attrs_group);
3087 }
3088 
3089 /*
3090  * nirqs is the number of interrupts available for write and read
3091  * queues. The core already reserved an interrupt for the admin queue.
3092  */
3093 static void nvme_calc_irq_sets(struct irq_affinity *affd, unsigned int nrirqs)
3094 {
3095 	struct nvme_dev *dev = affd->priv;
3096 	unsigned int nr_read_queues, nr_write_queues = dev->nr_write_queues;
3097 
3098 	/*
3099 	 * If there is no interrupt available for queues, ensure that
3100 	 * the default queue is set to 1. The affinity set size is
3101 	 * also set to one, but the irq core ignores it for this case.
3102 	 *
3103 	 * If only one interrupt is available or 'write_queue' == 0, combine
3104 	 * write and read queues.
3105 	 *
3106 	 * If 'write_queues' > 0, ensure it leaves room for at least one read
3107 	 * queue.
3108 	 */
3109 	if (!nrirqs) {
3110 		nrirqs = 1;
3111 		nr_read_queues = 0;
3112 	} else if (nrirqs == 1 || !nr_write_queues) {
3113 		nr_read_queues = 0;
3114 	} else if (nr_write_queues >= nrirqs) {
3115 		nr_read_queues = 1;
3116 	} else {
3117 		nr_read_queues = nrirqs - nr_write_queues;
3118 	}
3119 
3120 	dev->io_queues[HCTX_TYPE_DEFAULT] = nrirqs - nr_read_queues;
3121 	affd->set_size[HCTX_TYPE_DEFAULT] = nrirqs - nr_read_queues;
3122 	dev->io_queues[HCTX_TYPE_READ] = nr_read_queues;
3123 	affd->set_size[HCTX_TYPE_READ] = nr_read_queues;
3124 	affd->nr_sets = nr_read_queues ? 2 : 1;
3125 }
3126 
3127 static int nvme_setup_irqs(struct nvme_dev *dev, unsigned int nr_io_queues)
3128 {
3129 	struct pci_dev *pdev = to_pci_dev(dev->dev);
3130 	struct irq_affinity affd = {
3131 		.pre_vectors	= 1,
3132 		.calc_sets	= nvme_calc_irq_sets,
3133 		.priv		= dev,
3134 	};
3135 	unsigned int irq_queues, poll_queues;
3136 	unsigned int flags = PCI_IRQ_ALL_TYPES | PCI_IRQ_AFFINITY;
3137 
3138 	/*
3139 	 * Poll queues don't need interrupts, but we need at least one I/O queue
3140 	 * left over for non-polled I/O.
3141 	 */
3142 	poll_queues = min(dev->nr_poll_queues, nr_io_queues - 1);
3143 	dev->io_queues[HCTX_TYPE_POLL] = poll_queues;
3144 
3145 	/*
3146 	 * Initialize for the single interrupt case, will be updated in
3147 	 * nvme_calc_irq_sets().
3148 	 */
3149 	dev->io_queues[HCTX_TYPE_DEFAULT] = 1;
3150 	dev->io_queues[HCTX_TYPE_READ] = 0;
3151 
3152 	/*
3153 	 * We need interrupts for the admin queue and each non-polled I/O queue,
3154 	 * but some Apple controllers require all queues to use the first
3155 	 * vector.
3156 	 */
3157 	irq_queues = 1;
3158 	if (!(dev->ctrl.quirks & NVME_QUIRK_SINGLE_VECTOR))
3159 		irq_queues += (nr_io_queues - poll_queues);
3160 	if (dev->ctrl.quirks & NVME_QUIRK_BROKEN_MSI)
3161 		flags &= ~PCI_IRQ_MSI;
3162 	return pci_alloc_irq_vectors_affinity(pdev, 1, irq_queues, flags,
3163 					      &affd);
3164 }
3165 
3166 static unsigned int nvme_max_io_queues(struct nvme_dev *dev)
3167 {
3168 	/*
3169 	 * If tags are shared with admin queue (Apple bug), then
3170 	 * make sure we only use one IO queue.
3171 	 */
3172 	if (dev->ctrl.quirks & NVME_QUIRK_SHARED_TAGS)
3173 		return 1;
3174 	return blk_mq_num_possible_queues(0) + dev->nr_write_queues +
3175 		dev->nr_poll_queues;
3176 }
3177 
3178 static int nvme_setup_io_queues(struct nvme_dev *dev)
3179 {
3180 	struct nvme_queue *adminq = &dev->queues[0];
3181 	struct pci_dev *pdev = to_pci_dev(dev->dev);
3182 	unsigned int nr_io_queues;
3183 	unsigned long size;
3184 	int result;
3185 
3186 	/*
3187 	 * Sample the module parameters once at reset time so that we have
3188 	 * stable values to work with.
3189 	 */
3190 	dev->nr_write_queues = write_queues;
3191 	dev->nr_poll_queues = poll_queues;
3192 
3193 	if (dev->ctrl.tagset) {
3194 		/*
3195 		 * The set's maps are allocated only once at initialization
3196 		 * time. We can't add special queues later if their mq_map
3197 		 * wasn't preallocated.
3198 		 */
3199 		if (dev->ctrl.tagset->nr_maps < 3)
3200 			dev->nr_poll_queues = 0;
3201 		if (dev->ctrl.tagset->nr_maps < 2)
3202 			dev->nr_write_queues = 0;
3203 	}
3204 
3205 	/*
3206 	 * The initial number of allocated queue slots may be too large if the
3207 	 * user reduced the special queue parameters. Cap the value to the
3208 	 * number we need for this round.
3209 	 */
3210 	nr_io_queues = min(nvme_max_io_queues(dev),
3211 			   dev->nr_allocated_queues - 1);
3212 	result = nvme_set_queue_count(&dev->ctrl, &nr_io_queues);
3213 	if (result < 0)
3214 		return result;
3215 
3216 	if (nr_io_queues == 0)
3217 		return 0;
3218 
3219 	/*
3220 	 * Free IRQ resources as soon as NVMEQ_ENABLED bit transitions
3221 	 * from set to unset. If there is a window to it is truely freed,
3222 	 * pci_free_irq_vectors() jumping into this window will crash.
3223 	 * And take lock to avoid racing with pci_free_irq_vectors() in
3224 	 * nvme_dev_disable() path.
3225 	 */
3226 	result = nvme_setup_io_queues_trylock(dev);
3227 	if (result)
3228 		return result;
3229 	if (test_and_clear_bit(NVMEQ_ENABLED, &adminq->flags))
3230 		pci_free_irq(pdev, 0, adminq);
3231 
3232 	if (dev->cmb_use_sqes) {
3233 		result = nvme_cmb_qdepth(dev, nr_io_queues,
3234 				sizeof(struct nvme_command));
3235 		if (result > 0) {
3236 			dev->q_depth = result;
3237 			dev->ctrl.sqsize = result - 1;
3238 		} else {
3239 			dev->cmb_use_sqes = false;
3240 		}
3241 	}
3242 
3243 	do {
3244 		size = db_bar_size(dev, nr_io_queues);
3245 		result = nvme_remap_bar(dev, size);
3246 		if (!result)
3247 			break;
3248 		if (!--nr_io_queues) {
3249 			result = -ENOMEM;
3250 			goto out_unlock;
3251 		}
3252 	} while (1);
3253 	adminq->q_db = dev->dbs;
3254 
3255  retry:
3256 	/* Deregister the admin queue's interrupt */
3257 	if (test_and_clear_bit(NVMEQ_ENABLED, &adminq->flags))
3258 		pci_free_irq(pdev, 0, adminq);
3259 
3260 	/*
3261 	 * If we enable msix early due to not intx, disable it again before
3262 	 * setting up the full range we need.
3263 	 */
3264 	pci_free_irq_vectors(pdev);
3265 
3266 	result = nvme_setup_irqs(dev, nr_io_queues);
3267 	if (result <= 0) {
3268 		result = -EIO;
3269 		goto out_unlock;
3270 	}
3271 
3272 	dev->num_vecs = result;
3273 	result = max(result - 1, 1);
3274 	dev->max_qid = result + dev->io_queues[HCTX_TYPE_POLL];
3275 
3276 	/*
3277 	 * Should investigate if there's a performance win from allocating
3278 	 * more queues than interrupt vectors; it might allow the submission
3279 	 * path to scale better, even if the receive path is limited by the
3280 	 * number of interrupts.
3281 	 */
3282 	result = queue_request_irq(adminq);
3283 	if (result)
3284 		goto out_unlock;
3285 	set_bit(NVMEQ_ENABLED, &adminq->flags);
3286 	mutex_unlock(&dev->shutdown_lock);
3287 
3288 	result = nvme_create_io_queues(dev);
3289 	if (result || dev->online_queues < 2)
3290 		return result;
3291 
3292 	if (dev->online_queues - 1 < dev->max_qid) {
3293 		nr_io_queues = dev->online_queues - 1;
3294 		nvme_delete_io_queues(dev);
3295 		result = nvme_setup_io_queues_trylock(dev);
3296 		if (result)
3297 			return result;
3298 		nvme_suspend_io_queues(dev);
3299 		goto retry;
3300 	}
3301 	dev_info(dev->ctrl.device, "%d/%d/%d default/read/poll queues\n",
3302 					dev->io_queues[HCTX_TYPE_DEFAULT],
3303 					dev->io_queues[HCTX_TYPE_READ],
3304 					dev->io_queues[HCTX_TYPE_POLL]);
3305 	return 0;
3306 out_unlock:
3307 	mutex_unlock(&dev->shutdown_lock);
3308 	return result;
3309 }
3310 
3311 static enum rq_end_io_ret nvme_del_queue_end(struct request *req,
3312 					     blk_status_t error,
3313 					     const struct io_comp_batch *iob)
3314 {
3315 	struct nvme_queue *nvmeq = req->end_io_data;
3316 
3317 	blk_mq_free_request(req);
3318 	complete(&nvmeq->delete_done);
3319 	return RQ_END_IO_NONE;
3320 }
3321 
3322 static int nvme_delete_queue(struct nvme_queue *nvmeq, u8 opcode)
3323 {
3324 	struct request_queue *q = nvmeq->dev->ctrl.admin_q;
3325 	struct request *req;
3326 	struct nvme_command cmd = { };
3327 
3328 	cmd.delete_queue.opcode = opcode;
3329 	cmd.delete_queue.qid = cpu_to_le16(nvmeq->qid);
3330 
3331 	req = blk_mq_alloc_request(q, nvme_req_op(&cmd), BLK_MQ_REQ_NOWAIT);
3332 	if (IS_ERR(req))
3333 		return PTR_ERR(req);
3334 	nvme_init_request(req, &cmd);
3335 
3336 	req->end_io = nvme_del_queue_end;
3337 	req->end_io_data = nvmeq;
3338 
3339 	init_completion(&nvmeq->delete_done);
3340 	blk_execute_rq_nowait(req, false);
3341 	return 0;
3342 }
3343 
3344 static bool __nvme_delete_io_queues(struct nvme_dev *dev, u8 opcode)
3345 {
3346 	int nr_queues = dev->online_queues - 1, sent = 0;
3347 	unsigned long timeout;
3348 
3349  retry:
3350 	timeout = dev->ctrl.admin_timeout;
3351 	while (nr_queues > 0) {
3352 		if (nvme_delete_queue(&dev->queues[nr_queues], opcode))
3353 			break;
3354 		nr_queues--;
3355 		sent++;
3356 	}
3357 	while (sent) {
3358 		struct nvme_queue *nvmeq = &dev->queues[nr_queues + sent];
3359 
3360 		timeout = wait_for_completion_io_timeout(&nvmeq->delete_done,
3361 				timeout);
3362 		if (timeout == 0)
3363 			return false;
3364 
3365 		sent--;
3366 		if (nr_queues)
3367 			goto retry;
3368 	}
3369 	return true;
3370 }
3371 
3372 static void nvme_delete_io_queues(struct nvme_dev *dev)
3373 {
3374 	if (__nvme_delete_io_queues(dev, nvme_admin_delete_sq))
3375 		__nvme_delete_io_queues(dev, nvme_admin_delete_cq);
3376 }
3377 
3378 static unsigned int nvme_pci_nr_maps(struct nvme_dev *dev)
3379 {
3380 	if (dev->io_queues[HCTX_TYPE_POLL])
3381 		return 3;
3382 	if (dev->io_queues[HCTX_TYPE_READ])
3383 		return 2;
3384 	return 1;
3385 }
3386 
3387 static bool nvme_pci_update_nr_queues(struct nvme_dev *dev)
3388 {
3389 	if (!dev->ctrl.tagset) {
3390 		nvme_alloc_io_tag_set(&dev->ctrl, &dev->tagset, &nvme_mq_ops,
3391 				nvme_pci_nr_maps(dev), sizeof(struct nvme_iod));
3392 		return true;
3393 	}
3394 
3395 	/* Give up if we are racing with nvme_dev_disable() */
3396 	if (!mutex_trylock(&dev->shutdown_lock))
3397 		return false;
3398 
3399 	/* Check if nvme_dev_disable() has been executed already */
3400 	if (!dev->online_queues) {
3401 		mutex_unlock(&dev->shutdown_lock);
3402 		return false;
3403 	}
3404 
3405 	blk_mq_update_nr_hw_queues(&dev->tagset, dev->online_queues - 1);
3406 	/* free previously allocated queues that are no longer usable */
3407 	nvme_free_queues(dev, dev->online_queues);
3408 	mutex_unlock(&dev->shutdown_lock);
3409 	return true;
3410 }
3411 
3412 static int nvme_pci_enable(struct nvme_dev *dev)
3413 {
3414 	int result = -ENOMEM;
3415 	struct pci_dev *pdev = to_pci_dev(dev->dev);
3416 	unsigned int flags = PCI_IRQ_ALL_TYPES;
3417 
3418 	if (pci_enable_device_mem(pdev))
3419 		return result;
3420 
3421 	pci_set_master(pdev);
3422 
3423 	if (readl(dev->bar + NVME_REG_CSTS) == -1) {
3424 		dev_dbg(dev->ctrl.device, "reading CSTS register failed\n");
3425 		result = -ENODEV;
3426 		goto disable;
3427 	}
3428 
3429 	/*
3430 	 * Some devices and/or platforms don't advertise or work with INTx
3431 	 * interrupts. Pre-enable a single MSIX or MSI vec for setup. We'll
3432 	 * adjust this later.
3433 	 */
3434 	if (dev->ctrl.quirks & NVME_QUIRK_BROKEN_MSI)
3435 		flags &= ~PCI_IRQ_MSI;
3436 	result = pci_alloc_irq_vectors(pdev, 1, 1, flags);
3437 	if (result < 0)
3438 		goto disable;
3439 
3440 	dev->ctrl.cap = lo_hi_readq(dev->bar + NVME_REG_CAP);
3441 
3442 	dev->q_depth = min_t(u32, NVME_CAP_MQES(dev->ctrl.cap) + 1,
3443 				io_queue_depth);
3444 	dev->db_stride = 1 << NVME_CAP_STRIDE(dev->ctrl.cap);
3445 	dev->dbs = dev->bar + 4096;
3446 
3447 	/*
3448 	 * Some Apple controllers require a non-standard SQE size.
3449 	 * Interestingly they also seem to ignore the CC:IOSQES register
3450 	 * so we don't bother updating it here.
3451 	 */
3452 	if (dev->ctrl.quirks & NVME_QUIRK_128_BYTES_SQES)
3453 		dev->io_sqes = 7;
3454 	else
3455 		dev->io_sqes = NVME_NVM_IOSQES;
3456 
3457 	if (dev->ctrl.quirks & NVME_QUIRK_QDEPTH_ONE) {
3458 		dev->q_depth = 2;
3459 	} else if (pdev->vendor == PCI_VENDOR_ID_SAMSUNG &&
3460 		   (pdev->device == 0xa821 || pdev->device == 0xa822) &&
3461 		   NVME_CAP_MQES(dev->ctrl.cap) == 0) {
3462 		dev->q_depth = 64;
3463 		dev_err(dev->ctrl.device, "detected PM1725 NVMe controller, "
3464                         "set queue depth=%u\n", dev->q_depth);
3465 	}
3466 
3467 	/*
3468 	 * Controllers with the shared tags quirk need the IO queue to be
3469 	 * big enough so that we get 32 tags for the admin queue
3470 	 */
3471 	if ((dev->ctrl.quirks & NVME_QUIRK_SHARED_TAGS) &&
3472 	    (dev->q_depth < (NVME_AQ_DEPTH + 2))) {
3473 		dev->q_depth = NVME_AQ_DEPTH + 2;
3474 		dev_warn(dev->ctrl.device, "IO queue depth clamped to %d\n",
3475 			 dev->q_depth);
3476 	}
3477 	dev->ctrl.sqsize = dev->q_depth - 1; /* 0's based queue depth */
3478 
3479 	nvme_map_cmb(dev);
3480 
3481 	pci_save_state(pdev);
3482 
3483 	result = nvme_pci_configure_admin_queue(dev);
3484 	if (result)
3485 		goto free_irq;
3486 	return result;
3487 
3488  free_irq:
3489 	pci_free_irq_vectors(pdev);
3490  disable:
3491 	pci_disable_device(pdev);
3492 	return result;
3493 }
3494 
3495 static void nvme_dev_unmap(struct nvme_dev *dev)
3496 {
3497 	if (dev->bar)
3498 		iounmap(dev->bar);
3499 	pci_release_mem_regions(to_pci_dev(dev->dev));
3500 }
3501 
3502 static bool nvme_pci_ctrl_is_dead(struct nvme_dev *dev)
3503 {
3504 	struct pci_dev *pdev = to_pci_dev(dev->dev);
3505 	u32 csts;
3506 
3507 	if (!pci_is_enabled(pdev) || !pci_device_is_present(pdev))
3508 		return true;
3509 	if (pdev->error_state != pci_channel_io_normal)
3510 		return true;
3511 
3512 	csts = readl(dev->bar + NVME_REG_CSTS);
3513 	return (csts & NVME_CSTS_CFS) || !(csts & NVME_CSTS_RDY);
3514 }
3515 
3516 static void nvme_dev_disable(struct nvme_dev *dev, bool shutdown)
3517 {
3518 	enum nvme_ctrl_state state = nvme_ctrl_state(&dev->ctrl);
3519 	struct pci_dev *pdev = to_pci_dev(dev->dev);
3520 	bool dead;
3521 
3522 	mutex_lock(&dev->shutdown_lock);
3523 	dead = nvme_pci_ctrl_is_dead(dev);
3524 	if (state == NVME_CTRL_LIVE || state == NVME_CTRL_RESETTING) {
3525 		if (pci_is_enabled(pdev))
3526 			nvme_start_freeze(&dev->ctrl);
3527 		/*
3528 		 * Give the controller a chance to complete all entered requests
3529 		 * if doing a safe shutdown.
3530 		 */
3531 		if (!dead && shutdown)
3532 			nvme_wait_freeze_timeout(&dev->ctrl);
3533 	}
3534 
3535 	nvme_quiesce_io_queues(&dev->ctrl);
3536 
3537 	if (!dead && dev->ctrl.queue_count > 0) {
3538 		nvme_delete_io_queues(dev);
3539 		nvme_disable_ctrl(&dev->ctrl, shutdown);
3540 		nvme_poll_irqdisable(&dev->queues[0]);
3541 	}
3542 	nvme_suspend_io_queues(dev);
3543 	nvme_suspend_queue(dev, 0);
3544 	pci_free_irq_vectors(pdev);
3545 	if (pci_is_enabled(pdev))
3546 		pci_disable_device(pdev);
3547 	nvme_reap_pending_cqes(dev);
3548 
3549 	nvme_cancel_tagset(&dev->ctrl);
3550 	nvme_cancel_admin_tagset(&dev->ctrl);
3551 
3552 	/*
3553 	 * The driver will not be starting up queues again if shutting down so
3554 	 * must flush all entered requests to their failed completion to avoid
3555 	 * deadlocking blk-mq hot-cpu notifier.
3556 	 */
3557 	if (shutdown) {
3558 		nvme_unquiesce_io_queues(&dev->ctrl);
3559 		if (dev->ctrl.admin_q && !blk_queue_dying(dev->ctrl.admin_q))
3560 			nvme_unquiesce_admin_queue(&dev->ctrl);
3561 	}
3562 	mutex_unlock(&dev->shutdown_lock);
3563 }
3564 
3565 static int nvme_disable_prepare_reset(struct nvme_dev *dev, bool shutdown)
3566 {
3567 	if (!nvme_wait_reset(&dev->ctrl))
3568 		return -EBUSY;
3569 	nvme_dev_disable(dev, shutdown);
3570 	return 0;
3571 }
3572 
3573 static int nvme_pci_alloc_iod_mempool(struct nvme_dev *dev)
3574 {
3575 	size_t alloc_size = sizeof(struct nvme_dma_vec) * NVME_MAX_SEGS;
3576 
3577 	dev->dmavec_mempool = mempool_create_node(1,
3578 			mempool_kmalloc, mempool_kfree,
3579 			(void *)alloc_size, GFP_KERNEL,
3580 			dev_to_node(dev->dev));
3581 	if (!dev->dmavec_mempool)
3582 		return -ENOMEM;
3583 	return 0;
3584 }
3585 
3586 static void nvme_free_tagset(struct nvme_dev *dev)
3587 {
3588 	if (dev->tagset.tags)
3589 		nvme_remove_io_tag_set(&dev->ctrl);
3590 	dev->ctrl.tagset = NULL;
3591 }
3592 
3593 /* pairs with nvme_pci_alloc_dev */
3594 static void nvme_pci_free_ctrl(struct nvme_ctrl *ctrl)
3595 {
3596 	struct nvme_dev *dev = to_nvme_dev(ctrl);
3597 
3598 	nvme_free_tagset(dev);
3599 	put_device(dev->dev);
3600 	kfree(dev->queues);
3601 	kfree(dev);
3602 }
3603 
3604 static void nvme_reset_work(struct work_struct *work)
3605 {
3606 	struct nvme_dev *dev =
3607 		container_of(work, struct nvme_dev, ctrl.reset_work);
3608 	bool was_suspend = !!(dev->ctrl.ctrl_config & NVME_CC_SHN_NORMAL);
3609 	int result;
3610 
3611 	if (nvme_ctrl_state(&dev->ctrl) != NVME_CTRL_RESETTING) {
3612 		dev_warn(dev->ctrl.device, "ctrl state %d is not RESETTING\n",
3613 			 dev->ctrl.state);
3614 		result = -ENODEV;
3615 		goto out;
3616 	}
3617 
3618 	/*
3619 	 * If we're called to reset a live controller first shut it down before
3620 	 * moving on.
3621 	 */
3622 	if (dev->ctrl.ctrl_config & NVME_CC_ENABLE)
3623 		nvme_dev_disable(dev, false);
3624 	nvme_sync_queues(&dev->ctrl);
3625 
3626 	mutex_lock(&dev->shutdown_lock);
3627 	result = nvme_pci_enable(dev);
3628 	if (result)
3629 		goto out_unlock;
3630 	nvme_unquiesce_admin_queue(&dev->ctrl);
3631 	mutex_unlock(&dev->shutdown_lock);
3632 
3633 	/*
3634 	 * Introduce CONNECTING state from nvme-fc/rdma transports to mark the
3635 	 * initializing procedure here.
3636 	 */
3637 	if (!nvme_change_ctrl_state(&dev->ctrl, NVME_CTRL_CONNECTING)) {
3638 		dev_warn(dev->ctrl.device,
3639 			"failed to mark controller CONNECTING\n");
3640 		result = -EBUSY;
3641 		goto out;
3642 	}
3643 
3644 	result = nvme_init_ctrl_finish(&dev->ctrl, was_suspend);
3645 	if (result)
3646 		goto out;
3647 
3648 	if (nvme_ctrl_meta_sgl_supported(&dev->ctrl))
3649 		dev->ctrl.max_integrity_segments = NVME_MAX_META_SEGS;
3650 	else
3651 		dev->ctrl.max_integrity_segments = 1;
3652 
3653 	nvme_dbbuf_dma_alloc(dev);
3654 
3655 	result = nvme_setup_host_mem(dev);
3656 	if (result < 0)
3657 		goto out;
3658 
3659 	nvme_update_attrs(dev);
3660 
3661 	result = nvme_setup_io_queues(dev);
3662 	if (result)
3663 		goto out;
3664 
3665 	/*
3666 	 * Freeze and update the number of I/O queues as those might have
3667 	 * changed.  If there are no I/O queues left after this reset, keep the
3668 	 * controller around but remove all namespaces.
3669 	 */
3670 	if (dev->online_queues > 1) {
3671 		nvme_dbbuf_set(dev);
3672 		nvme_unquiesce_io_queues(&dev->ctrl);
3673 		nvme_wait_freeze(&dev->ctrl);
3674 		if (!nvme_pci_update_nr_queues(dev))
3675 			goto out;
3676 		nvme_unfreeze(&dev->ctrl);
3677 	} else {
3678 		dev_warn(dev->ctrl.device, "IO queues lost\n");
3679 		nvme_mark_namespaces_dead(&dev->ctrl);
3680 		nvme_unquiesce_io_queues(&dev->ctrl);
3681 		nvme_remove_namespaces(&dev->ctrl);
3682 		nvme_free_tagset(dev);
3683 	}
3684 
3685 	/*
3686 	 * If only admin queue live, keep it to do further investigation or
3687 	 * recovery.
3688 	 */
3689 	if (!nvme_change_ctrl_state(&dev->ctrl, NVME_CTRL_LIVE)) {
3690 		dev_warn(dev->ctrl.device,
3691 			"failed to mark controller live state\n");
3692 		result = -ENODEV;
3693 		goto out;
3694 	}
3695 
3696 	nvme_start_ctrl(&dev->ctrl);
3697 	return;
3698 
3699  out_unlock:
3700 	mutex_unlock(&dev->shutdown_lock);
3701  out:
3702 	/*
3703 	 * Set state to deleting now to avoid blocking nvme_wait_reset(), which
3704 	 * may be holding this pci_dev's device lock.
3705 	 */
3706 	dev_warn(dev->ctrl.device, "Disabling device after reset failure: %d\n",
3707 		 result);
3708 	nvme_change_ctrl_state(&dev->ctrl, NVME_CTRL_DELETING);
3709 	nvme_dev_disable(dev, true);
3710 	nvme_sync_queues(&dev->ctrl);
3711 	nvme_mark_namespaces_dead(&dev->ctrl);
3712 	nvme_unquiesce_io_queues(&dev->ctrl);
3713 	nvme_change_ctrl_state(&dev->ctrl, NVME_CTRL_DEAD);
3714 }
3715 
3716 static int nvme_pci_reg_read32(struct nvme_ctrl *ctrl, u32 off, u32 *val)
3717 {
3718 	*val = readl(to_nvme_dev(ctrl)->bar + off);
3719 	return 0;
3720 }
3721 
3722 static int nvme_pci_reg_write32(struct nvme_ctrl *ctrl, u32 off, u32 val)
3723 {
3724 	writel(val, to_nvme_dev(ctrl)->bar + off);
3725 	return 0;
3726 }
3727 
3728 static int nvme_pci_reg_read64(struct nvme_ctrl *ctrl, u32 off, u64 *val)
3729 {
3730 	*val = lo_hi_readq(to_nvme_dev(ctrl)->bar + off);
3731 	return 0;
3732 }
3733 
3734 static int nvme_pci_get_address(struct nvme_ctrl *ctrl, char *buf, int size)
3735 {
3736 	struct pci_dev *pdev = to_pci_dev(to_nvme_dev(ctrl)->dev);
3737 
3738 	return snprintf(buf, size, "%s\n", dev_name(&pdev->dev));
3739 }
3740 
3741 static void nvme_pci_print_device_info(struct nvme_ctrl *ctrl)
3742 {
3743 	struct pci_dev *pdev = to_pci_dev(to_nvme_dev(ctrl)->dev);
3744 	struct nvme_subsystem *subsys = ctrl->subsys;
3745 
3746 	dev_err(ctrl->device,
3747 		"VID:DID %04x:%04x model:%.*s firmware:%.*s\n",
3748 		pdev->vendor, pdev->device,
3749 		nvme_strlen(subsys->model, sizeof(subsys->model)),
3750 		subsys->model, nvme_strlen(subsys->firmware_rev,
3751 					   sizeof(subsys->firmware_rev)),
3752 		subsys->firmware_rev);
3753 }
3754 
3755 static bool nvme_pci_supports_pci_p2pdma(struct nvme_ctrl *ctrl)
3756 {
3757 	struct nvme_dev *dev = to_nvme_dev(ctrl);
3758 
3759 	return dma_pci_p2pdma_supported(dev->dev);
3760 }
3761 
3762 static unsigned long nvme_pci_get_virt_boundary(struct nvme_ctrl *ctrl,
3763 						bool is_admin)
3764 {
3765 	if (!nvme_ctrl_sgl_supported(ctrl) || is_admin)
3766 		return NVME_CTRL_PAGE_SIZE - 1;
3767 	return 0;
3768 }
3769 
3770 static int nvme_pci_init_dma_buf_io_ctx(struct nvme_ctrl *ctrl,
3771 		struct dma_buf_io_ctx *ctx)
3772 {
3773 	struct nvme_dev *dev = to_nvme_dev(ctrl);
3774 
3775 	return nvme_init_dma_buf_io_ctx(ctrl, &dev->dmabufs, ctx);
3776 }
3777 
3778 static const struct nvme_ctrl_ops nvme_pci_ctrl_ops = {
3779 	.name			= "pcie",
3780 	.module			= THIS_MODULE,
3781 	.flags			= NVME_F_METADATA_SUPPORTED,
3782 	.dev_attr_groups	= nvme_pci_dev_attr_groups,
3783 	.reg_read32		= nvme_pci_reg_read32,
3784 	.reg_write32		= nvme_pci_reg_write32,
3785 	.reg_read64		= nvme_pci_reg_read64,
3786 	.free_ctrl		= nvme_pci_free_ctrl,
3787 	.submit_async_event	= nvme_pci_submit_async_event,
3788 	.subsystem_reset	= nvme_pci_subsystem_reset,
3789 	.get_address		= nvme_pci_get_address,
3790 	.print_device_info	= nvme_pci_print_device_info,
3791 	.supports_pci_p2pdma	= nvme_pci_supports_pci_p2pdma,
3792 	.get_virt_boundary	= nvme_pci_get_virt_boundary,
3793 	.init_dma_buf_io_ctx	= nvme_pci_init_dma_buf_io_ctx,
3794 };
3795 
3796 static int nvme_dev_map(struct nvme_dev *dev)
3797 {
3798 	struct pci_dev *pdev = to_pci_dev(dev->dev);
3799 
3800 	if (pci_request_mem_regions(pdev, "nvme"))
3801 		return -ENODEV;
3802 
3803 	if (nvme_remap_bar(dev, NVME_REG_DBS + 4096))
3804 		goto release;
3805 
3806 	return 0;
3807   release:
3808 	pci_release_mem_regions(pdev);
3809 	return -ENODEV;
3810 }
3811 
3812 static unsigned long check_vendor_combination_bug(struct pci_dev *pdev)
3813 {
3814 	if (pdev->vendor == 0x144d && pdev->device == 0xa802) {
3815 		/*
3816 		 * Several Samsung devices seem to drop off the PCIe bus
3817 		 * randomly when APST is on and uses the deepest sleep state.
3818 		 * This has been observed on a Samsung "SM951 NVMe SAMSUNG
3819 		 * 256GB", a "PM951 NVMe SAMSUNG 512GB", and a "Samsung SSD
3820 		 * 950 PRO 256GB", but it seems to be restricted to two Dell
3821 		 * laptops.
3822 		 */
3823 		if (dmi_match(DMI_SYS_VENDOR, "Dell Inc.") &&
3824 		    (dmi_match(DMI_PRODUCT_NAME, "XPS 15 9550") ||
3825 		     dmi_match(DMI_PRODUCT_NAME, "Precision 5510")))
3826 			return NVME_QUIRK_NO_DEEPEST_PS;
3827 	} else if (pdev->vendor == 0x144d && pdev->device == 0xa804) {
3828 		/*
3829 		 * Samsung SSD 960 EVO drops off the PCIe bus after system
3830 		 * suspend on a Ryzen board, ASUS PRIME B350M-A, as well as
3831 		 * within few minutes after bootup on a Coffee Lake board -
3832 		 * ASUS PRIME Z370-A
3833 		 */
3834 		if (dmi_match(DMI_BOARD_VENDOR, "ASUSTeK COMPUTER INC.") &&
3835 		    (dmi_match(DMI_BOARD_NAME, "PRIME B350M-A") ||
3836 		     dmi_match(DMI_BOARD_NAME, "PRIME Z370-A")))
3837 			return NVME_QUIRK_NO_APST;
3838 	} else if ((pdev->vendor == 0x144d && (pdev->device == 0xa801 ||
3839 		    pdev->device == 0xa808 || pdev->device == 0xa809)) ||
3840 		   (pdev->vendor == 0x1e0f && pdev->device == 0x0001)) {
3841 		/*
3842 		 * Forcing to use host managed nvme power settings for
3843 		 * lowest idle power with quick resume latency on
3844 		 * Samsung and Toshiba SSDs based on suspend behavior
3845 		 * on Coffee Lake board for LENOVO C640
3846 		 */
3847 		if ((dmi_match(DMI_BOARD_VENDOR, "LENOVO")) &&
3848 		     dmi_match(DMI_BOARD_NAME, "LNVNB161216"))
3849 			return NVME_QUIRK_SIMPLE_SUSPEND;
3850 	} else if (pdev->vendor == 0x2646 && (pdev->device == 0x2263 ||
3851 		   pdev->device == 0x500f)) {
3852 		/*
3853 		 * Exclude some Kingston NV1 and A2000 devices from
3854 		 * NVME_QUIRK_SIMPLE_SUSPEND. Do a full suspend to save a
3855 		 * lot of energy with s2idle sleep on some TUXEDO platforms.
3856 		 */
3857 		if (dmi_match(DMI_BOARD_NAME, "NS5X_NS7XAU") ||
3858 		    dmi_match(DMI_BOARD_NAME, "NS5x_7xAU") ||
3859 		    dmi_match(DMI_BOARD_NAME, "NS5x_7xPU") ||
3860 		    dmi_match(DMI_BOARD_NAME, "PH4PRX1_PH6PRX1"))
3861 			return NVME_QUIRK_FORCE_NO_SIMPLE_SUSPEND;
3862 	} else if (pdev->vendor == 0x144d && pdev->device == 0xa80d) {
3863 		/*
3864 		 * Exclude Samsung 990 Evo from NVME_QUIRK_SIMPLE_SUSPEND
3865 		 * because of high power consumption (> 2 Watt) in s2idle
3866 		 * sleep. Only some boards with Intel CPU are affected.
3867 		 * (Note for testing: Samsung 990 Evo Plus has same PCI ID)
3868 		 */
3869 		if (dmi_match(DMI_BOARD_NAME, "DN50Z-140HC-YD") ||
3870 		    dmi_match(DMI_BOARD_NAME, "GMxPXxx") ||
3871 		    dmi_match(DMI_BOARD_NAME, "GXxMRXx") ||
3872 		    dmi_match(DMI_BOARD_NAME, "NS5X_NS7XAU") ||
3873 		    dmi_match(DMI_BOARD_NAME, "PH4PG31") ||
3874 		    dmi_match(DMI_BOARD_NAME, "PH4PRX1_PH6PRX1") ||
3875 		    dmi_match(DMI_BOARD_NAME, "PH6PG01_PH6PG71"))
3876 			return NVME_QUIRK_FORCE_NO_SIMPLE_SUSPEND;
3877 	}
3878 
3879 	/*
3880 	 * NVMe SSD drops off the PCIe bus after system idle
3881 	 * for 10 hours on a Lenovo N60z board.
3882 	 */
3883 	if (dmi_match(DMI_BOARD_NAME, "LXKT-ZXEG-N6"))
3884 		return NVME_QUIRK_NO_APST;
3885 
3886 	return 0;
3887 }
3888 
3889 static struct quirk_entry *detect_dynamic_quirks(struct pci_dev *pdev)
3890 {
3891 	int i;
3892 
3893 	for (i = 0; i < nvme_pci_quirk_count; i++)
3894 		if (pdev->vendor == nvme_pci_quirk_list[i].vendor_id &&
3895 		    pdev->device == nvme_pci_quirk_list[i].dev_id)
3896 			return &nvme_pci_quirk_list[i];
3897 
3898 	return NULL;
3899 }
3900 
3901 static struct nvme_dev *nvme_pci_alloc_dev(struct pci_dev *pdev,
3902 		const struct pci_device_id *id)
3903 {
3904 	unsigned long quirks = id->driver_data;
3905 	int node = dev_to_node(&pdev->dev);
3906 	struct nvme_dev *dev;
3907 	struct quirk_entry *qentry;
3908 	int ret = -ENOMEM;
3909 
3910 	dev = kzalloc_node(struct_size(dev, descriptor_pools, nr_node_ids),
3911 			GFP_KERNEL, node);
3912 	if (!dev)
3913 		return ERR_PTR(-ENOMEM);
3914 	INIT_WORK(&dev->ctrl.reset_work, nvme_reset_work);
3915 	mutex_init(&dev->shutdown_lock);
3916 	nvme_init_buf_list(&dev->dmabufs);
3917 
3918 	dev->nr_write_queues = write_queues;
3919 	dev->nr_poll_queues = poll_queues;
3920 	dev->nr_allocated_queues = nvme_max_io_queues(dev) + 1;
3921 	dev->queues = kcalloc_node(dev->nr_allocated_queues,
3922 			sizeof(struct nvme_queue), GFP_KERNEL, node);
3923 	if (!dev->queues)
3924 		goto out_free_dev;
3925 
3926 	dev->dev = get_device(&pdev->dev);
3927 
3928 	quirks |= check_vendor_combination_bug(pdev);
3929 	if (!noacpi &&
3930 	    !(quirks & NVME_QUIRK_FORCE_NO_SIMPLE_SUSPEND) &&
3931 	    acpi_storage_d3(&pdev->dev)) {
3932 		/*
3933 		 * Some systems use a bios work around to ask for D3 on
3934 		 * platforms that support kernel managed suspend.
3935 		 */
3936 		dev_info(&pdev->dev,
3937 			 "platform quirk: setting simple suspend\n");
3938 		quirks |= NVME_QUIRK_SIMPLE_SUSPEND;
3939 	}
3940 	qentry = detect_dynamic_quirks(pdev);
3941 	if (qentry) {
3942 		quirks |= qentry->enabled_quirks;
3943 		quirks &= ~qentry->disabled_quirks;
3944 	}
3945 	ret = nvme_init_ctrl(&dev->ctrl, &pdev->dev, &nvme_pci_ctrl_ops,
3946 			     quirks);
3947 	if (ret)
3948 		goto out_put_device;
3949 
3950 	if (dev->ctrl.quirks & NVME_QUIRK_DMA_ADDRESS_BITS_48)
3951 		dma_set_mask_and_coherent(&pdev->dev, DMA_BIT_MASK(48));
3952 	else
3953 		dma_set_mask_and_coherent(&pdev->dev, DMA_BIT_MASK(64));
3954 	dma_set_min_align_mask(&pdev->dev, NVME_CTRL_PAGE_SIZE - 1);
3955 	dma_set_max_seg_size(&pdev->dev, 0xffffffff);
3956 
3957 	/*
3958 	 * Limit the max command size to prevent iod->sg allocations going
3959 	 * over a single page.
3960 	 */
3961 	dev->ctrl.max_hw_sectors = min_t(u32,
3962 			NVME_MAX_BYTES >> SECTOR_SHIFT,
3963 			dma_max_mapping_size(&pdev->dev) >> SECTOR_SHIFT);
3964 	dev->ctrl.max_segments = NVME_MAX_SEGS;
3965 	dev->ctrl.max_integrity_segments = 1;
3966 	return dev;
3967 
3968 out_put_device:
3969 	put_device(dev->dev);
3970 	kfree(dev->queues);
3971 out_free_dev:
3972 	kfree(dev);
3973 	return ERR_PTR(ret);
3974 }
3975 
3976 static int nvme_probe(struct pci_dev *pdev, const struct pci_device_id *id)
3977 {
3978 	struct nvme_dev *dev;
3979 	int result = -ENOMEM;
3980 
3981 	dev = nvme_pci_alloc_dev(pdev, id);
3982 	if (IS_ERR(dev))
3983 		return PTR_ERR(dev);
3984 
3985 	result = nvme_add_ctrl(&dev->ctrl);
3986 	if (result)
3987 		goto out_put_ctrl;
3988 
3989 	result = nvme_dev_map(dev);
3990 	if (result)
3991 		goto out_uninit_ctrl;
3992 
3993 	result = nvme_pci_alloc_iod_mempool(dev);
3994 	if (result)
3995 		goto out_dev_unmap;
3996 
3997 	dev_info(dev->ctrl.device, "pci function %s\n", dev_name(&pdev->dev));
3998 
3999 	result = nvme_pci_enable(dev);
4000 	if (result)
4001 		goto out_release_iod_mempool;
4002 
4003 	result = nvme_alloc_admin_tag_set(&dev->ctrl, &dev->admin_tagset,
4004 				&nvme_mq_admin_ops, sizeof(struct nvme_iod));
4005 	if (result)
4006 		goto out_disable;
4007 
4008 	/*
4009 	 * Mark the controller as connecting before sending admin commands to
4010 	 * allow the timeout handler to do the right thing.
4011 	 */
4012 	if (!nvme_change_ctrl_state(&dev->ctrl, NVME_CTRL_CONNECTING)) {
4013 		dev_warn(dev->ctrl.device,
4014 			"failed to mark controller CONNECTING\n");
4015 		result = -EBUSY;
4016 		goto out_disable;
4017 	}
4018 
4019 	result = nvme_init_ctrl_finish(&dev->ctrl, false);
4020 	if (result)
4021 		goto out_disable;
4022 
4023 	if (nvme_ctrl_meta_sgl_supported(&dev->ctrl))
4024 		dev->ctrl.max_integrity_segments = NVME_MAX_META_SEGS;
4025 	else
4026 		dev->ctrl.max_integrity_segments = 1;
4027 
4028 	nvme_dbbuf_dma_alloc(dev);
4029 
4030 	result = nvme_setup_host_mem(dev);
4031 	if (result < 0)
4032 		goto out_disable;
4033 
4034 	nvme_update_attrs(dev);
4035 
4036 	result = nvme_setup_io_queues(dev);
4037 	if (result)
4038 		goto out_disable;
4039 
4040 	if (dev->online_queues > 1) {
4041 		nvme_alloc_io_tag_set(&dev->ctrl, &dev->tagset, &nvme_mq_ops,
4042 				nvme_pci_nr_maps(dev), sizeof(struct nvme_iod));
4043 		nvme_dbbuf_set(dev);
4044 	}
4045 
4046 	if (!dev->ctrl.tagset)
4047 		dev_warn(dev->ctrl.device, "IO queues not created\n");
4048 
4049 	if (!nvme_change_ctrl_state(&dev->ctrl, NVME_CTRL_LIVE)) {
4050 		dev_warn(dev->ctrl.device,
4051 			"failed to mark controller live state\n");
4052 		result = -ENODEV;
4053 		goto out_disable;
4054 	}
4055 
4056 	pci_set_drvdata(pdev, dev);
4057 
4058 	nvme_start_ctrl(&dev->ctrl);
4059 	nvme_put_ctrl(&dev->ctrl);
4060 	flush_work(&dev->ctrl.scan_work);
4061 	return 0;
4062 
4063 out_disable:
4064 	nvme_change_ctrl_state(&dev->ctrl, NVME_CTRL_DELETING);
4065 	nvme_dev_disable(dev, true);
4066 	nvme_free_host_mem(dev);
4067 	nvme_dev_remove_admin(dev);
4068 	nvme_dbbuf_dma_free(dev);
4069 	nvme_free_queues(dev, 0);
4070 	nvme_release_descriptor_pools(dev);
4071 out_release_iod_mempool:
4072 	mempool_destroy(dev->dmavec_mempool);
4073 out_dev_unmap:
4074 	nvme_dev_unmap(dev);
4075 out_uninit_ctrl:
4076 	nvme_uninit_ctrl(&dev->ctrl);
4077 out_put_ctrl:
4078 	nvme_put_ctrl(&dev->ctrl);
4079 	dev_err_probe(&pdev->dev, result, "probe failed\n");
4080 	return result;
4081 }
4082 
4083 static void nvme_reset_prepare(struct pci_dev *pdev)
4084 {
4085 	struct nvme_dev *dev = pci_get_drvdata(pdev);
4086 
4087 	/*
4088 	 * We don't need to check the return value from waiting for the reset
4089 	 * state as pci_dev device lock is held, making it impossible to race
4090 	 * with ->remove().
4091 	 */
4092 	nvme_disable_prepare_reset(dev, false);
4093 	nvme_sync_queues(&dev->ctrl);
4094 }
4095 
4096 static void nvme_reset_done(struct pci_dev *pdev)
4097 {
4098 	struct nvme_dev *dev = pci_get_drvdata(pdev);
4099 
4100 	if (!nvme_try_sched_reset(&dev->ctrl))
4101 		flush_work(&dev->ctrl.reset_work);
4102 }
4103 
4104 static void nvme_shutdown(struct pci_dev *pdev)
4105 {
4106 	struct nvme_dev *dev = pci_get_drvdata(pdev);
4107 
4108 	nvme_disable_prepare_reset(dev, true);
4109 }
4110 
4111 /*
4112  * The driver's remove may be called on a device in a partially initialized
4113  * state. This function must not have any dependencies on the device state in
4114  * order to proceed.
4115  */
4116 static void nvme_remove(struct pci_dev *pdev)
4117 {
4118 	struct nvme_dev *dev = pci_get_drvdata(pdev);
4119 
4120 	nvme_change_ctrl_state(&dev->ctrl, NVME_CTRL_DELETING);
4121 	pci_set_drvdata(pdev, NULL);
4122 
4123 	if (!pci_device_is_present(pdev)) {
4124 		nvme_change_ctrl_state(&dev->ctrl, NVME_CTRL_DEAD);
4125 		nvme_dev_disable(dev, true);
4126 	}
4127 
4128 	flush_work(&dev->ctrl.reset_work);
4129 	nvme_stop_ctrl(&dev->ctrl);
4130 	nvme_remove_namespaces(&dev->ctrl);
4131 	nvme_dev_disable(dev, true);
4132 	nvme_pci_remove_dmabuf(&dev->dmabufs);
4133 	nvme_free_host_mem(dev);
4134 	nvme_dev_remove_admin(dev);
4135 	nvme_dbbuf_dma_free(dev);
4136 	nvme_free_queues(dev, 0);
4137 	mempool_destroy(dev->dmavec_mempool);
4138 	nvme_release_descriptor_pools(dev);
4139 	nvme_dev_unmap(dev);
4140 	nvme_uninit_ctrl(&dev->ctrl);
4141 }
4142 
4143 #ifdef CONFIG_PM_SLEEP
4144 static int nvme_get_power_state(struct nvme_ctrl *ctrl, u32 *ps)
4145 {
4146 	return nvme_get_features(ctrl, NVME_FEAT_POWER_MGMT, 0, NULL, 0, ps);
4147 }
4148 
4149 static int nvme_set_power_state(struct nvme_ctrl *ctrl, u32 ps)
4150 {
4151 	return nvme_set_features(ctrl, NVME_FEAT_POWER_MGMT, ps, NULL, 0, NULL);
4152 }
4153 
4154 static int nvme_resume(struct device *dev)
4155 {
4156 	struct nvme_dev *ndev = pci_get_drvdata(to_pci_dev(dev));
4157 	struct nvme_ctrl *ctrl = &ndev->ctrl;
4158 
4159 	if (ndev->last_ps == U32_MAX ||
4160 	    nvme_set_power_state(ctrl, ndev->last_ps) != 0)
4161 		goto reset;
4162 	if (ctrl->hmpre && nvme_setup_host_mem(ndev))
4163 		goto reset;
4164 
4165 	return 0;
4166 reset:
4167 	return nvme_try_sched_reset(ctrl);
4168 }
4169 
4170 static int nvme_suspend(struct device *dev)
4171 {
4172 	struct pci_dev *pdev = to_pci_dev(dev);
4173 	struct nvme_dev *ndev = pci_get_drvdata(pdev);
4174 	struct nvme_ctrl *ctrl = &ndev->ctrl;
4175 	int ret = -EBUSY;
4176 
4177 	ndev->last_ps = U32_MAX;
4178 
4179 	/*
4180 	 * The platform does not remove power for a kernel managed suspend so
4181 	 * use host managed nvme power settings for lowest idle power if
4182 	 * possible. This should have quicker resume latency than a full device
4183 	 * shutdown.  But if the firmware is involved after the suspend or the
4184 	 * platform has any limitation in waking from low power states or the
4185 	 * device does not support any non-default power states, shut down the
4186 	 * device fully.
4187 	 *
4188 	 * If ASPM is not enabled for the device, shut down the device and allow
4189 	 * the PCI bus layer to put it into D3 in order to take the PCIe link
4190 	 * down, so as to allow the platform to achieve its minimum low-power
4191 	 * state (which may not be possible if the link is up).
4192 	 */
4193 	if (!pci_suspend_retains_context(pdev) || !ctrl->npss ||
4194 	    !pcie_aspm_enabled(pdev) ||
4195 	    (ndev->ctrl.quirks & NVME_QUIRK_SIMPLE_SUSPEND))
4196 		return nvme_disable_prepare_reset(ndev, true);
4197 
4198 	nvme_start_freeze(ctrl);
4199 	nvme_wait_freeze(ctrl);
4200 	nvme_sync_queues(ctrl);
4201 
4202 	if (nvme_ctrl_state(ctrl) != NVME_CTRL_LIVE)
4203 		goto unfreeze;
4204 
4205 	/*
4206 	 * Host memory access may not be successful in a system suspend state,
4207 	 * but the specification allows the controller to access memory in a
4208 	 * non-operational power state.
4209 	 */
4210 	if (ndev->hmb) {
4211 		ret = nvme_set_host_mem(ndev, 0);
4212 		if (ret < 0)
4213 			goto unfreeze;
4214 		if (ret) {
4215 			/*
4216 			 * The controller may still access the HMB. Shut it
4217 			 * down instead, which forces a reset on resume.
4218 			 */
4219 			ret = nvme_disable_prepare_reset(ndev, true);
4220 			goto unfreeze;
4221 		}
4222 	}
4223 
4224 	ret = nvme_get_power_state(ctrl, &ndev->last_ps);
4225 	if (ret < 0)
4226 		goto unfreeze;
4227 
4228 	/*
4229 	 * A saved state prevents pci pm from generically controlling the
4230 	 * device's power. If we're using protocol specific settings, we don't
4231 	 * want pci interfering.
4232 	 */
4233 	pci_save_state(pdev);
4234 
4235 	ret = nvme_set_power_state(ctrl, ctrl->npss);
4236 	if (ret < 0)
4237 		goto unfreeze;
4238 
4239 	if (ret) {
4240 		/* discard the saved state */
4241 		pci_load_saved_state(pdev, NULL);
4242 
4243 		/*
4244 		 * Clearing npss forces a controller reset on resume. The
4245 		 * correct value will be rediscovered then.
4246 		 */
4247 		ret = nvme_disable_prepare_reset(ndev, true);
4248 		ctrl->npss = 0;
4249 	}
4250 unfreeze:
4251 	nvme_unfreeze(ctrl);
4252 	return ret;
4253 }
4254 
4255 static int nvme_simple_suspend(struct device *dev)
4256 {
4257 	struct nvme_dev *ndev = pci_get_drvdata(to_pci_dev(dev));
4258 
4259 	return nvme_disable_prepare_reset(ndev, true);
4260 }
4261 
4262 static int nvme_simple_resume(struct device *dev)
4263 {
4264 	struct pci_dev *pdev = to_pci_dev(dev);
4265 	struct nvme_dev *ndev = pci_get_drvdata(pdev);
4266 
4267 	return nvme_try_sched_reset(&ndev->ctrl);
4268 }
4269 
4270 static const struct dev_pm_ops nvme_dev_pm_ops = {
4271 	.suspend	= nvme_suspend,
4272 	.resume		= nvme_resume,
4273 	.freeze		= nvme_simple_suspend,
4274 	.thaw		= nvme_simple_resume,
4275 	.poweroff	= nvme_simple_suspend,
4276 	.restore	= nvme_simple_resume,
4277 };
4278 #endif /* CONFIG_PM_SLEEP */
4279 
4280 static pci_ers_result_t nvme_error_detected(struct pci_dev *pdev,
4281 						pci_channel_state_t state)
4282 {
4283 	struct nvme_dev *dev = pci_get_drvdata(pdev);
4284 
4285 	/*
4286 	 * A frozen channel requires a reset. When detected, this method will
4287 	 * shutdown the controller to quiesce. The controller will be restarted
4288 	 * after the slot reset through driver's slot_reset callback.
4289 	 */
4290 	switch (state) {
4291 	case pci_channel_io_normal:
4292 		return PCI_ERS_RESULT_CAN_RECOVER;
4293 	case pci_channel_io_frozen:
4294 		dev_warn(dev->ctrl.device,
4295 			"frozen state error detected, reset controller\n");
4296 		if (!nvme_change_ctrl_state(&dev->ctrl, NVME_CTRL_RESETTING)) {
4297 			nvme_dev_disable(dev, true);
4298 			return PCI_ERS_RESULT_DISCONNECT;
4299 		}
4300 		nvme_dev_disable(dev, false);
4301 		return PCI_ERS_RESULT_NEED_RESET;
4302 	case pci_channel_io_perm_failure:
4303 		dev_warn(dev->ctrl.device,
4304 			"failure state error detected, request disconnect\n");
4305 		return PCI_ERS_RESULT_DISCONNECT;
4306 	}
4307 	return PCI_ERS_RESULT_NEED_RESET;
4308 }
4309 
4310 static pci_ers_result_t nvme_slot_reset(struct pci_dev *pdev)
4311 {
4312 	struct nvme_dev *dev = pci_get_drvdata(pdev);
4313 
4314 	dev_info(dev->ctrl.device, "restart after slot reset\n");
4315 	pci_restore_state(pdev);
4316 	if (nvme_try_sched_reset(&dev->ctrl))
4317 		nvme_unquiesce_io_queues(&dev->ctrl);
4318 	return PCI_ERS_RESULT_RECOVERED;
4319 }
4320 
4321 static void nvme_error_resume(struct pci_dev *pdev)
4322 {
4323 	struct nvme_dev *dev = pci_get_drvdata(pdev);
4324 
4325 	flush_work(&dev->ctrl.reset_work);
4326 }
4327 
4328 static const struct pci_error_handlers nvme_err_handler = {
4329 	.error_detected	= nvme_error_detected,
4330 	.slot_reset	= nvme_slot_reset,
4331 	.resume		= nvme_error_resume,
4332 	.reset_prepare	= nvme_reset_prepare,
4333 	.reset_done	= nvme_reset_done,
4334 };
4335 
4336 static const struct pci_device_id nvme_id_table[] = {
4337 	{ PCI_VDEVICE(INTEL, 0x0953),	/* Intel 750/P3500/P3600/P3700 */
4338 		.driver_data = NVME_QUIRK_STRIPE_SIZE |
4339 				NVME_QUIRK_DEALLOCATE_ZEROES, },
4340 	{ PCI_VDEVICE(INTEL, 0x0a53),	/* Intel P3520 */
4341 		.driver_data = NVME_QUIRK_STRIPE_SIZE |
4342 				NVME_QUIRK_DEALLOCATE_ZEROES, },
4343 	{ PCI_VDEVICE(INTEL, 0x0a54),	/* Intel P4500/P4600 */
4344 		.driver_data = NVME_QUIRK_STRIPE_SIZE |
4345 				NVME_QUIRK_IGNORE_DEV_SUBNQN |
4346 				NVME_QUIRK_BOGUS_NID, },
4347 	{ PCI_VDEVICE(INTEL, 0x0a55),	/* Dell Express Flash P4600 */
4348 		.driver_data = NVME_QUIRK_STRIPE_SIZE, },
4349 	{ PCI_VDEVICE(INTEL, 0xf1a5),	/* Intel 600P/P3100 */
4350 		.driver_data = NVME_QUIRK_NO_DEEPEST_PS |
4351 				NVME_QUIRK_MEDIUM_PRIO_SQ |
4352 				NVME_QUIRK_NO_TEMP_THRESH_CHANGE |
4353 				NVME_QUIRK_DISABLE_WRITE_ZEROES, },
4354 	{ PCI_VDEVICE(INTEL, 0xf1a6),	/* Intel 760p/Pro 7600p */
4355 		.driver_data = NVME_QUIRK_IGNORE_DEV_SUBNQN, },
4356 	{ PCI_VDEVICE(INTEL, 0x5845),	/* Qemu emulated controller */
4357 		.driver_data = NVME_QUIRK_IDENTIFY_CNS |
4358 				NVME_QUIRK_DISABLE_WRITE_ZEROES |
4359 				NVME_QUIRK_BOGUS_NID, },
4360 	{ PCI_VDEVICE(REDHAT, 0x0010),	/* Qemu emulated controller */
4361 		.driver_data = NVME_QUIRK_BOGUS_NID, },
4362 	{ PCI_DEVICE(0x1217, 0x8760), /* O2 Micro 64GB Steam Deck */
4363 		.driver_data = NVME_QUIRK_DMAPOOL_ALIGN_512, },
4364 	{ PCI_DEVICE(0x126f, 0x1001),	/* Silicon Motion generic */
4365 		.driver_data = NVME_QUIRK_NO_DEEPEST_PS |
4366 				NVME_QUIRK_IGNORE_DEV_SUBNQN, },
4367 	{ PCI_DEVICE(0x126f, 0x2262),	/* Silicon Motion generic */
4368 		.driver_data = NVME_QUIRK_NO_DEEPEST_PS |
4369 				NVME_QUIRK_BOGUS_NID, },
4370 	{ PCI_DEVICE(0x126f, 0x2263),	/* Silicon Motion unidentified */
4371 		.driver_data = NVME_QUIRK_NO_NS_DESC_LIST |
4372 				NVME_QUIRK_BOGUS_NID, },
4373 	{ PCI_DEVICE(0x1bb1, 0x0100),   /* Seagate Nytro Flash Storage */
4374 		.driver_data = NVME_QUIRK_DELAY_BEFORE_CHK_RDY |
4375 				NVME_QUIRK_NO_NS_DESC_LIST, },
4376 	{ PCI_DEVICE(0x1c58, 0x0003),	/* HGST adapter */
4377 		.driver_data = NVME_QUIRK_DELAY_BEFORE_CHK_RDY, },
4378 	{ PCI_DEVICE(0x1c58, 0x0023),	/* WDC SN200 adapter */
4379 		.driver_data = NVME_QUIRK_DELAY_BEFORE_CHK_RDY, },
4380 	{ PCI_DEVICE(0x1c5f, 0x0540),	/* Memblaze Pblaze4 adapter */
4381 		.driver_data = NVME_QUIRK_DELAY_BEFORE_CHK_RDY, },
4382 	{ PCI_DEVICE(0x1c5f, 0x0555),	/* Memblaze Pblaze5 adapter */
4383 		.driver_data = NVME_QUIRK_NO_NS_DESC_LIST, },
4384 	{ PCI_DEVICE(0x144d, 0xa821),   /* Samsung PM1725 */
4385 		.driver_data = NVME_QUIRK_DELAY_BEFORE_CHK_RDY, },
4386 	{ PCI_DEVICE(0x144d, 0xa822),   /* Samsung PM1725a */
4387 		.driver_data = NVME_QUIRK_DELAY_BEFORE_CHK_RDY |
4388 				NVME_QUIRK_DISABLE_WRITE_ZEROES|
4389 				NVME_QUIRK_IGNORE_DEV_SUBNQN, },
4390 	{ PCI_DEVICE(0x15b7, 0x5008),   /* Sandisk SN530 */
4391 		.driver_data = NVME_QUIRK_BROKEN_MSI },
4392 	{ PCI_DEVICE(0x15b7, 0x5009),   /* Sandisk SN550 */
4393 		.driver_data = NVME_QUIRK_BROKEN_MSI |
4394 				NVME_QUIRK_NO_DEEPEST_PS },
4395 	{ PCI_DEVICE(0x1987, 0x5012),	/* Phison E12 */
4396 		.driver_data = NVME_QUIRK_BOGUS_NID, },
4397 	{ PCI_DEVICE(0x1987, 0x5016),	/* Phison E16 */
4398 		.driver_data = NVME_QUIRK_IGNORE_DEV_SUBNQN |
4399 				NVME_QUIRK_BOGUS_NID, },
4400 	{ PCI_DEVICE(0x1987, 0x5019),  /* phison E19 */
4401 		.driver_data = NVME_QUIRK_DISABLE_WRITE_ZEROES, },
4402 	{ PCI_DEVICE(0x1987, 0x5021),   /* Phison E21 */
4403 		.driver_data = NVME_QUIRK_DISABLE_WRITE_ZEROES, },
4404 	{ PCI_DEVICE(0x1b4b, 0x1092),	/* Lexar 256 GB SSD */
4405 		.driver_data = NVME_QUIRK_NO_NS_DESC_LIST |
4406 				NVME_QUIRK_IGNORE_DEV_SUBNQN, },
4407 	{ PCI_DEVICE(0x1cc1, 0x33f8),   /* ADATA IM2P33F8ABR1 1 TB */
4408 		.driver_data = NVME_QUIRK_BOGUS_NID, },
4409 	{ PCI_DEVICE(0x10ec, 0x5762),   /* ADATA SX6000LNP */
4410 		.driver_data = NVME_QUIRK_IGNORE_DEV_SUBNQN |
4411 				NVME_QUIRK_BOGUS_NID, },
4412 	{ PCI_DEVICE(0x10ec, 0x5763),  /* ADATA SX6000PNP */
4413 		.driver_data = NVME_QUIRK_BOGUS_NID, },
4414 	{ PCI_DEVICE(0x1cc1, 0x8201),   /* ADATA SX8200PNP 512GB */
4415 		.driver_data = NVME_QUIRK_NO_DEEPEST_PS |
4416 				NVME_QUIRK_IGNORE_DEV_SUBNQN, },
4417 	 { PCI_DEVICE(0x1344, 0x5407), /* Micron Technology Inc NVMe SSD */
4418 		.driver_data = NVME_QUIRK_IGNORE_DEV_SUBNQN },
4419 	 { PCI_DEVICE(0x1344, 0x6001),   /* Micron Nitro NVMe */
4420 		 .driver_data = NVME_QUIRK_BOGUS_NID, },
4421 	{ PCI_DEVICE(0x1344, 0x6004),	/* Micron 4100AT */
4422 		.driver_data = NVME_QUIRK_DMAPOOL_ALIGN_512, },
4423 	{ PCI_DEVICE(0x1e0f, 0x0032),	/* KIOXIA XG10d */
4424 		.driver_data = NVME_QUIRK_DMAPOOL_ALIGN_512, },
4425 	{ PCI_DEVICE(0x1c5c, 0x1504),   /* SK Hynix PC400 */
4426 		.driver_data = NVME_QUIRK_DISABLE_WRITE_ZEROES, },
4427 	{ PCI_DEVICE(0x1c5c, 0x174a),   /* SK Hynix P31 SSD */
4428 		.driver_data = NVME_QUIRK_BOGUS_NID, },
4429 	{ PCI_DEVICE(0x1c5c, 0x1D59),   /* SK Hynix BC901 */
4430 		.driver_data = NVME_QUIRK_DISABLE_WRITE_ZEROES, },
4431 	{ PCI_DEVICE(0x15b7, 0x2001),   /*  Sandisk Skyhawk */
4432 		.driver_data = NVME_QUIRK_DISABLE_WRITE_ZEROES, },
4433 	{ PCI_DEVICE(0x1d97, 0x2263),   /* SPCC / Lexar NM610 */
4434 		.driver_data = NVME_QUIRK_DISABLE_WRITE_ZEROES |
4435 				NVME_QUIRK_BOGUS_NID, },
4436 	{ PCI_DEVICE(0x144d, 0xa80b),   /* Samsung PM9B1 256G and 512G */
4437 		.driver_data = NVME_QUIRK_DISABLE_WRITE_ZEROES |
4438 				NVME_QUIRK_BOGUS_NID, },
4439 	{ PCI_DEVICE(0x144d, 0xa809),   /* Samsung MZALQ256HBJD 256G */
4440 		.driver_data = NVME_QUIRK_DISABLE_WRITE_ZEROES, },
4441 	{ PCI_DEVICE(0x144d, 0xa802),   /* Samsung SM953 */
4442 		.driver_data = NVME_QUIRK_BOGUS_NID, },
4443 	{ PCI_DEVICE(0x1cc4, 0x6303),   /* UMIS RPJTJ512MGE1QDY 512G */
4444 		.driver_data = NVME_QUIRK_DISABLE_WRITE_ZEROES, },
4445 	{ PCI_DEVICE(0x1cc4, 0x6302),   /* UMIS RPJTJ256MGE1QDY 256G */
4446 		.driver_data = NVME_QUIRK_DISABLE_WRITE_ZEROES, },
4447 	{ PCI_DEVICE(0x2646, 0x2262),   /* KINGSTON SKC2000 NVMe SSD */
4448 		.driver_data = NVME_QUIRK_NO_DEEPEST_PS, },
4449 	{ PCI_DEVICE(0x2646, 0x2263),   /* KINGSTON A2000 NVMe SSD  */
4450 		.driver_data = NVME_QUIRK_NO_DEEPEST_PS, },
4451 	{ PCI_DEVICE(0x2646, 0x5013),   /* Kingston KC3000, Kingston FURY Renegade */
4452 		.driver_data = NVME_QUIRK_NO_SECONDARY_TEMP_THRESH, },
4453 	{ PCI_DEVICE(0x2646, 0x5018),   /* KINGSTON OM8SFP4xxxxP OS21012 NVMe SSD */
4454 		.driver_data = NVME_QUIRK_DISABLE_WRITE_ZEROES, },
4455 	{ PCI_DEVICE(0x2646, 0x5016),   /* KINGSTON OM3PGP4xxxxP OS21011 NVMe SSD */
4456 		.driver_data = NVME_QUIRK_DISABLE_WRITE_ZEROES, },
4457 	{ PCI_DEVICE(0x2646, 0x501A),   /* KINGSTON OM8PGP4xxxxP OS21005 NVMe SSD */
4458 		.driver_data = NVME_QUIRK_DISABLE_WRITE_ZEROES, },
4459 	{ PCI_DEVICE(0x2646, 0x501B),   /* KINGSTON OM8PGP4xxxxQ OS21005 NVMe SSD */
4460 		.driver_data = NVME_QUIRK_DISABLE_WRITE_ZEROES, },
4461 	{ PCI_DEVICE(0x2646, 0x501E),   /* KINGSTON OM3PGP4xxxxQ OS21011 NVMe SSD */
4462 		.driver_data = NVME_QUIRK_DISABLE_WRITE_ZEROES, },
4463 	{ PCI_DEVICE(0x2646, 0x502F),   /* KINGSTON OM3SGP4xxxxK NVMe SSD */
4464 		.driver_data = NVME_QUIRK_DISABLE_WRITE_ZEROES, },
4465 	{ PCI_DEVICE(0x1f40, 0x1202),   /* Netac Technologies Co. NV3000 NVMe SSD */
4466 		.driver_data = NVME_QUIRK_BOGUS_NID, },
4467 	{ PCI_DEVICE(0x1f40, 0x5236),   /* Netac Technologies Co. NV7000 NVMe SSD */
4468 		.driver_data = NVME_QUIRK_BOGUS_NID, },
4469 	{ PCI_DEVICE(0x1e4B, 0x1001),   /* MAXIO MAP1001 */
4470 		.driver_data = NVME_QUIRK_BOGUS_NID, },
4471 	{ PCI_DEVICE(0x1e4B, 0x1002),   /* MAXIO MAP1002 */
4472 		.driver_data = NVME_QUIRK_BOGUS_NID, },
4473 	{ PCI_DEVICE(0x1e4B, 0x1202),   /* MAXIO MAP1202 */
4474 		.driver_data = NVME_QUIRK_BOGUS_NID, },
4475 	{ PCI_DEVICE(0x1e4B, 0x1602),   /* MAXIO MAP1602 */
4476 		.driver_data = NVME_QUIRK_BOGUS_NID, },
4477 	{ PCI_DEVICE(0x1cc1, 0x5350),   /* ADATA XPG GAMMIX S50 */
4478 		.driver_data = NVME_QUIRK_BOGUS_NID, },
4479 	{ PCI_DEVICE(0x1cc1, 0x5370),   /* ADATA XPG GAMMIX S70 BLADE */
4480 		.driver_data = NVME_QUIRK_NO_FUA, },
4481 	{ PCI_DEVICE(0x1dbe, 0x5216),   /* Acer/INNOGRIT FA100/5216 NVMe SSD */
4482 		.driver_data = NVME_QUIRK_BOGUS_NID, },
4483 	{ PCI_DEVICE(0x1dbe, 0x5236),   /* ADATA XPG GAMMIX S70 */
4484 		.driver_data = NVME_QUIRK_BOGUS_NID, },
4485 	{ PCI_DEVICE(0x1e49, 0x0021),   /* ZHITAI TiPro5000 NVMe SSD */
4486 		.driver_data = NVME_QUIRK_NO_DEEPEST_PS, },
4487 	{ PCI_DEVICE(0x1e49, 0x0041),   /* ZHITAI TiPro7000 NVMe SSD */
4488 		.driver_data = NVME_QUIRK_NO_DEEPEST_PS, },
4489 	{ PCI_DEVICE(0x1fa0, 0x2283),   /* Wodposit WPBSNM8-256GTP */
4490 		.driver_data = NVME_QUIRK_NO_SECONDARY_TEMP_THRESH, },
4491 	{ PCI_DEVICE(0x025e, 0xf1ac),   /* SOLIDIGM  P44 pro SSDPFKKW020X7  */
4492 		.driver_data = NVME_QUIRK_NO_DEEPEST_PS, },
4493 	{ PCI_DEVICE(0xc0a9, 0x540a),   /* Crucial P2 */
4494 		.driver_data = NVME_QUIRK_BOGUS_NID, },
4495 	{ PCI_DEVICE(0x1d97, 0x1d97), /* Lexar NM620 */
4496 		.driver_data = NVME_QUIRK_BOGUS_NID, },
4497 	{ PCI_DEVICE(0x1d97, 0x2269), /* Lexar NM760 */
4498 		.driver_data = NVME_QUIRK_BOGUS_NID |
4499 				NVME_QUIRK_IGNORE_DEV_SUBNQN, },
4500 	{ PCI_DEVICE(0x10ec, 0x5763), /* TEAMGROUP T-FORCE CARDEA ZERO Z330 SSD */
4501 		.driver_data = NVME_QUIRK_BOGUS_NID, },
4502 	{ PCI_DEVICE(0x1e4b, 0x1602), /* HS-SSD-FUTURE 2048G  */
4503 		.driver_data = NVME_QUIRK_BOGUS_NID, },
4504 	{ PCI_DEVICE(0x10ec, 0x5765), /* TEAMGROUP MP33 2TB SSD */
4505 		.driver_data = NVME_QUIRK_BOGUS_NID, },
4506 	{ PCI_DEVICE(PCI_VENDOR_ID_AMAZON, 0x0061),
4507 		.driver_data = NVME_QUIRK_DMA_ADDRESS_BITS_48, },
4508 	{ PCI_DEVICE(PCI_VENDOR_ID_AMAZON, 0x0065),
4509 		.driver_data = NVME_QUIRK_DMA_ADDRESS_BITS_48, },
4510 	{ PCI_DEVICE(PCI_VENDOR_ID_AMAZON, 0x8061),
4511 		.driver_data = NVME_QUIRK_DMA_ADDRESS_BITS_48, },
4512 	{ PCI_DEVICE(PCI_VENDOR_ID_AMAZON, 0xcd00),
4513 		.driver_data = NVME_QUIRK_DMA_ADDRESS_BITS_48, },
4514 	{ PCI_DEVICE(PCI_VENDOR_ID_AMAZON, 0xcd01),
4515 		.driver_data = NVME_QUIRK_DMA_ADDRESS_BITS_48, },
4516 	{ PCI_DEVICE(PCI_VENDOR_ID_AMAZON, 0xcd02),
4517 		.driver_data = NVME_QUIRK_DMA_ADDRESS_BITS_48, },
4518 	{ PCI_DEVICE(PCI_VENDOR_ID_APPLE, 0x2001),
4519 		/*
4520 		 * Fix for the Apple controller found in the MacBook8,1 and
4521 		 * some MacBook7,1 to avoid controller resets and data loss.
4522 		 */
4523 		.driver_data = NVME_QUIRK_SINGLE_VECTOR |
4524 				NVME_QUIRK_QDEPTH_ONE },
4525 	{ PCI_DEVICE(PCI_VENDOR_ID_APPLE, 0x2003) },
4526 	{ PCI_DEVICE(PCI_VENDOR_ID_APPLE, 0x2005),
4527 		.driver_data = NVME_QUIRK_SINGLE_VECTOR |
4528 				NVME_QUIRK_128_BYTES_SQES |
4529 				NVME_QUIRK_SHARED_TAGS |
4530 				NVME_QUIRK_SKIP_CID_GEN |
4531 				NVME_QUIRK_IDENTIFY_CNS },
4532 	{ PCI_DEVICE_CLASS(PCI_CLASS_STORAGE_EXPRESS, 0xffffff) },
4533 	{ 0, }
4534 };
4535 MODULE_DEVICE_TABLE(pci, nvme_id_table);
4536 
4537 static struct pci_driver nvme_driver = {
4538 	.name		= "nvme",
4539 	.id_table	= nvme_id_table,
4540 	.probe		= nvme_probe,
4541 	.remove		= nvme_remove,
4542 	.shutdown	= nvme_shutdown,
4543 	.driver		= {
4544 		.probe_type	= PROBE_PREFER_ASYNCHRONOUS,
4545 #ifdef CONFIG_PM_SLEEP
4546 		.pm		= &nvme_dev_pm_ops,
4547 #endif
4548 	},
4549 	.sriov_configure = pci_sriov_configure_simple,
4550 	.err_handler	= &nvme_err_handler,
4551 };
4552 
4553 static int __init nvme_init(void)
4554 {
4555 	BUILD_BUG_ON(sizeof(struct nvme_create_cq) != 64);
4556 	BUILD_BUG_ON(sizeof(struct nvme_create_sq) != 64);
4557 	BUILD_BUG_ON(sizeof(struct nvme_delete_queue) != 64);
4558 	BUILD_BUG_ON(IRQ_AFFINITY_MAX_SETS < 2);
4559 
4560 	return pci_register_driver(&nvme_driver);
4561 }
4562 
4563 static void __exit nvme_exit(void)
4564 {
4565 	kfree(nvme_pci_quirk_list);
4566 	pci_unregister_driver(&nvme_driver);
4567 	flush_workqueue(nvme_wq);
4568 }
4569 
4570 MODULE_AUTHOR("Matthew Wilcox <willy@linux.intel.com>");
4571 MODULE_LICENSE("GPL");
4572 MODULE_VERSION("1.0");
4573 MODULE_DESCRIPTION("NVMe host PCIe transport driver");
4574 MODULE_IMPORT_NS("DMA_BUF");
4575 module_init(nvme_init);
4576 module_exit(nvme_exit);
4577