xref: /openbmc/u-boot/drivers/nvme/nvme.c (revision 01c541e0)
1 /*
2  * Copyright (C) 2017 NXP Semiconductors
3  * Copyright (C) 2017 Bin Meng <bmeng.cn@gmail.com>
4  *
5  * SPDX-License-Identifier:	GPL-2.0+
6  */
7 
8 #include <common.h>
9 #include <dm.h>
10 #include <errno.h>
11 #include <memalign.h>
12 #include <pci.h>
13 #include <dm/device-internal.h>
14 #include "nvme.h"
15 
16 #define NVME_Q_DEPTH		2
17 #define NVME_AQ_DEPTH		2
18 #define NVME_SQ_SIZE(depth)	(depth * sizeof(struct nvme_command))
19 #define NVME_CQ_SIZE(depth)	(depth * sizeof(struct nvme_completion))
20 #define ADMIN_TIMEOUT		60
21 #define IO_TIMEOUT		30
22 #define MAX_PRP_POOL		512
23 
24 enum nvme_queue_id {
25 	NVME_ADMIN_Q,
26 	NVME_IO_Q,
27 	NVME_Q_NUM,
28 };
29 
30 /*
31  * An NVM Express queue. Each device has at least two (one for admin
32  * commands and one for I/O commands).
33  */
34 struct nvme_queue {
35 	struct nvme_dev *dev;
36 	struct nvme_command *sq_cmds;
37 	struct nvme_completion *cqes;
38 	wait_queue_head_t sq_full;
39 	u32 __iomem *q_db;
40 	u16 q_depth;
41 	s16 cq_vector;
42 	u16 sq_head;
43 	u16 sq_tail;
44 	u16 cq_head;
45 	u16 qid;
46 	u8 cq_phase;
47 	u8 cqe_seen;
48 	unsigned long cmdid_data[];
49 };
50 
51 static int nvme_wait_ready(struct nvme_dev *dev, bool enabled)
52 {
53 	u32 bit = enabled ? NVME_CSTS_RDY : 0;
54 	int timeout;
55 	ulong start;
56 
57 	/* Timeout field in the CAP register is in 500 millisecond units */
58 	timeout = NVME_CAP_TIMEOUT(dev->cap) * 500;
59 
60 	start = get_timer(0);
61 	while (get_timer(start) < timeout) {
62 		if ((readl(&dev->bar->csts) & NVME_CSTS_RDY) == bit)
63 			return 0;
64 	}
65 
66 	return -ETIME;
67 }
68 
69 static int nvme_setup_prps(struct nvme_dev *dev, u64 *prp2,
70 			   int total_len, u64 dma_addr)
71 {
72 	u32 page_size = dev->page_size;
73 	int offset = dma_addr & (page_size - 1);
74 	u64 *prp_pool;
75 	int length = total_len;
76 	int i, nprps;
77 	length -= (page_size - offset);
78 
79 	if (length <= 0) {
80 		*prp2 = 0;
81 		return 0;
82 	}
83 
84 	if (length)
85 		dma_addr += (page_size - offset);
86 
87 	if (length <= page_size) {
88 		*prp2 = dma_addr;
89 		return 0;
90 	}
91 
92 	nprps = DIV_ROUND_UP(length, page_size);
93 
94 	if (nprps > dev->prp_entry_num) {
95 		free(dev->prp_pool);
96 		dev->prp_pool = malloc(nprps << 3);
97 		if (!dev->prp_pool) {
98 			printf("Error: malloc prp_pool fail\n");
99 			return -ENOMEM;
100 		}
101 		dev->prp_entry_num = nprps;
102 	}
103 
104 	prp_pool = dev->prp_pool;
105 	i = 0;
106 	while (nprps) {
107 		if (i == ((page_size >> 3) - 1)) {
108 			*(prp_pool + i) = cpu_to_le64((ulong)prp_pool +
109 					page_size);
110 			i = 0;
111 			prp_pool += page_size;
112 		}
113 		*(prp_pool + i++) = cpu_to_le64(dma_addr);
114 		dma_addr += page_size;
115 		nprps--;
116 	}
117 	*prp2 = (ulong)dev->prp_pool;
118 
119 	return 0;
120 }
121 
122 static __le16 nvme_get_cmd_id(void)
123 {
124 	static unsigned short cmdid;
125 
126 	return cpu_to_le16((cmdid < USHRT_MAX) ? cmdid++ : 0);
127 }
128 
129 static u16 nvme_read_completion_status(struct nvme_queue *nvmeq, u16 index)
130 {
131 	u64 start = (ulong)&nvmeq->cqes[index];
132 	u64 stop = start + sizeof(struct nvme_completion);
133 
134 	invalidate_dcache_range(start, stop);
135 
136 	return le16_to_cpu(readw(&(nvmeq->cqes[index].status)));
137 }
138 
139 /**
140  * nvme_submit_cmd() - copy a command into a queue and ring the doorbell
141  *
142  * @nvmeq:	The queue to use
143  * @cmd:	The command to send
144  */
145 static void nvme_submit_cmd(struct nvme_queue *nvmeq, struct nvme_command *cmd)
146 {
147 	u16 tail = nvmeq->sq_tail;
148 
149 	memcpy(&nvmeq->sq_cmds[tail], cmd, sizeof(*cmd));
150 	flush_dcache_range((ulong)&nvmeq->sq_cmds[tail],
151 			   (ulong)&nvmeq->sq_cmds[tail] + sizeof(*cmd));
152 
153 	if (++tail == nvmeq->q_depth)
154 		tail = 0;
155 	writel(tail, nvmeq->q_db);
156 	nvmeq->sq_tail = tail;
157 }
158 
159 static int nvme_submit_sync_cmd(struct nvme_queue *nvmeq,
160 				struct nvme_command *cmd,
161 				u32 *result, unsigned timeout)
162 {
163 	u16 head = nvmeq->cq_head;
164 	u16 phase = nvmeq->cq_phase;
165 	u16 status;
166 	ulong start_time;
167 	ulong timeout_us = timeout * 100000;
168 
169 	cmd->common.command_id = nvme_get_cmd_id();
170 	nvme_submit_cmd(nvmeq, cmd);
171 
172 	start_time = timer_get_us();
173 
174 	for (;;) {
175 		status = nvme_read_completion_status(nvmeq, head);
176 		if ((status & 0x01) == phase)
177 			break;
178 		if (timeout_us > 0 && (timer_get_us() - start_time)
179 		    >= timeout_us)
180 			return -ETIMEDOUT;
181 	}
182 
183 	status >>= 1;
184 	if (status) {
185 		printf("ERROR: status = %x, phase = %d, head = %d\n",
186 		       status, phase, head);
187 		status = 0;
188 		if (++head == nvmeq->q_depth) {
189 			head = 0;
190 			phase = !phase;
191 		}
192 		writel(head, nvmeq->q_db + nvmeq->dev->db_stride);
193 		nvmeq->cq_head = head;
194 		nvmeq->cq_phase = phase;
195 
196 		return -EIO;
197 	}
198 
199 	if (result)
200 		*result = le32_to_cpu(readl(&(nvmeq->cqes[head].result)));
201 
202 	if (++head == nvmeq->q_depth) {
203 		head = 0;
204 		phase = !phase;
205 	}
206 	writel(head, nvmeq->q_db + nvmeq->dev->db_stride);
207 	nvmeq->cq_head = head;
208 	nvmeq->cq_phase = phase;
209 
210 	return status;
211 }
212 
213 static int nvme_submit_admin_cmd(struct nvme_dev *dev, struct nvme_command *cmd,
214 				 u32 *result)
215 {
216 	return nvme_submit_sync_cmd(dev->queues[NVME_ADMIN_Q], cmd,
217 				    result, ADMIN_TIMEOUT);
218 }
219 
220 static struct nvme_queue *nvme_alloc_queue(struct nvme_dev *dev,
221 					   int qid, int depth)
222 {
223 	struct nvme_queue *nvmeq = malloc(sizeof(*nvmeq));
224 	if (!nvmeq)
225 		return NULL;
226 	memset(nvmeq, 0, sizeof(*nvmeq));
227 
228 	nvmeq->cqes = (void *)memalign(4096, NVME_CQ_SIZE(depth));
229 	if (!nvmeq->cqes)
230 		goto free_nvmeq;
231 	memset((void *)nvmeq->cqes, 0, NVME_CQ_SIZE(depth));
232 
233 	nvmeq->sq_cmds = (void *)memalign(4096, NVME_SQ_SIZE(depth));
234 	if (!nvmeq->sq_cmds)
235 		goto free_queue;
236 	memset((void *)nvmeq->sq_cmds, 0, NVME_SQ_SIZE(depth));
237 
238 	nvmeq->dev = dev;
239 
240 	nvmeq->cq_head = 0;
241 	nvmeq->cq_phase = 1;
242 	nvmeq->q_db = &dev->dbs[qid * 2 * dev->db_stride];
243 	nvmeq->q_depth = depth;
244 	nvmeq->qid = qid;
245 	dev->queue_count++;
246 	dev->queues[qid] = nvmeq;
247 
248 	return nvmeq;
249 
250  free_queue:
251 	free((void *)nvmeq->cqes);
252  free_nvmeq:
253 	free(nvmeq);
254 
255 	return NULL;
256 }
257 
258 static int nvme_delete_queue(struct nvme_dev *dev, u8 opcode, u16 id)
259 {
260 	struct nvme_command c;
261 
262 	memset(&c, 0, sizeof(c));
263 	c.delete_queue.opcode = opcode;
264 	c.delete_queue.qid = cpu_to_le16(id);
265 
266 	return nvme_submit_admin_cmd(dev, &c, NULL);
267 }
268 
269 static int nvme_delete_sq(struct nvme_dev *dev, u16 sqid)
270 {
271 	return nvme_delete_queue(dev, nvme_admin_delete_sq, sqid);
272 }
273 
274 static int nvme_delete_cq(struct nvme_dev *dev, u16 cqid)
275 {
276 	return nvme_delete_queue(dev, nvme_admin_delete_cq, cqid);
277 }
278 
279 static int nvme_enable_ctrl(struct nvme_dev *dev)
280 {
281 	dev->ctrl_config &= ~NVME_CC_SHN_MASK;
282 	dev->ctrl_config |= NVME_CC_ENABLE;
283 	writel(cpu_to_le32(dev->ctrl_config), &dev->bar->cc);
284 
285 	return nvme_wait_ready(dev, true);
286 }
287 
288 static int nvme_disable_ctrl(struct nvme_dev *dev)
289 {
290 	dev->ctrl_config &= ~NVME_CC_SHN_MASK;
291 	dev->ctrl_config &= ~NVME_CC_ENABLE;
292 	writel(cpu_to_le32(dev->ctrl_config), &dev->bar->cc);
293 
294 	return nvme_wait_ready(dev, false);
295 }
296 
297 static void nvme_free_queue(struct nvme_queue *nvmeq)
298 {
299 	free((void *)nvmeq->cqes);
300 	free(nvmeq->sq_cmds);
301 	free(nvmeq);
302 }
303 
304 static void nvme_free_queues(struct nvme_dev *dev, int lowest)
305 {
306 	int i;
307 
308 	for (i = dev->queue_count - 1; i >= lowest; i--) {
309 		struct nvme_queue *nvmeq = dev->queues[i];
310 		dev->queue_count--;
311 		dev->queues[i] = NULL;
312 		nvme_free_queue(nvmeq);
313 	}
314 }
315 
316 static void nvme_init_queue(struct nvme_queue *nvmeq, u16 qid)
317 {
318 	struct nvme_dev *dev = nvmeq->dev;
319 
320 	nvmeq->sq_tail = 0;
321 	nvmeq->cq_head = 0;
322 	nvmeq->cq_phase = 1;
323 	nvmeq->q_db = &dev->dbs[qid * 2 * dev->db_stride];
324 	memset((void *)nvmeq->cqes, 0, NVME_CQ_SIZE(nvmeq->q_depth));
325 	flush_dcache_range((ulong)nvmeq->cqes,
326 			   (ulong)nvmeq->cqes + NVME_CQ_SIZE(nvmeq->q_depth));
327 	dev->online_queues++;
328 }
329 
330 static int nvme_configure_admin_queue(struct nvme_dev *dev)
331 {
332 	int result;
333 	u32 aqa;
334 	u64 cap = dev->cap;
335 	struct nvme_queue *nvmeq;
336 	/* most architectures use 4KB as the page size */
337 	unsigned page_shift = 12;
338 	unsigned dev_page_min = NVME_CAP_MPSMIN(cap) + 12;
339 	unsigned dev_page_max = NVME_CAP_MPSMAX(cap) + 12;
340 
341 	if (page_shift < dev_page_min) {
342 		debug("Device minimum page size (%u) too large for host (%u)\n",
343 		      1 << dev_page_min, 1 << page_shift);
344 		return -ENODEV;
345 	}
346 
347 	if (page_shift > dev_page_max) {
348 		debug("Device maximum page size (%u) smaller than host (%u)\n",
349 		      1 << dev_page_max, 1 << page_shift);
350 		page_shift = dev_page_max;
351 	}
352 
353 	result = nvme_disable_ctrl(dev);
354 	if (result < 0)
355 		return result;
356 
357 	nvmeq = dev->queues[NVME_ADMIN_Q];
358 	if (!nvmeq) {
359 		nvmeq = nvme_alloc_queue(dev, 0, NVME_AQ_DEPTH);
360 		if (!nvmeq)
361 			return -ENOMEM;
362 	}
363 
364 	aqa = nvmeq->q_depth - 1;
365 	aqa |= aqa << 16;
366 	aqa |= aqa << 16;
367 
368 	dev->page_size = 1 << page_shift;
369 
370 	dev->ctrl_config = NVME_CC_CSS_NVM;
371 	dev->ctrl_config |= (page_shift - 12) << NVME_CC_MPS_SHIFT;
372 	dev->ctrl_config |= NVME_CC_ARB_RR | NVME_CC_SHN_NONE;
373 	dev->ctrl_config |= NVME_CC_IOSQES | NVME_CC_IOCQES;
374 
375 	writel(aqa, &dev->bar->aqa);
376 	nvme_writeq((ulong)nvmeq->sq_cmds, &dev->bar->asq);
377 	nvme_writeq((ulong)nvmeq->cqes, &dev->bar->acq);
378 
379 	result = nvme_enable_ctrl(dev);
380 	if (result)
381 		goto free_nvmeq;
382 
383 	nvmeq->cq_vector = 0;
384 
385 	nvme_init_queue(dev->queues[NVME_ADMIN_Q], 0);
386 
387 	return result;
388 
389  free_nvmeq:
390 	nvme_free_queues(dev, 0);
391 
392 	return result;
393 }
394 
395 static int nvme_alloc_cq(struct nvme_dev *dev, u16 qid,
396 			    struct nvme_queue *nvmeq)
397 {
398 	struct nvme_command c;
399 	int flags = NVME_QUEUE_PHYS_CONTIG | NVME_CQ_IRQ_ENABLED;
400 
401 	memset(&c, 0, sizeof(c));
402 	c.create_cq.opcode = nvme_admin_create_cq;
403 	c.create_cq.prp1 = cpu_to_le64((ulong)nvmeq->cqes);
404 	c.create_cq.cqid = cpu_to_le16(qid);
405 	c.create_cq.qsize = cpu_to_le16(nvmeq->q_depth - 1);
406 	c.create_cq.cq_flags = cpu_to_le16(flags);
407 	c.create_cq.irq_vector = cpu_to_le16(nvmeq->cq_vector);
408 
409 	return nvme_submit_admin_cmd(dev, &c, NULL);
410 }
411 
412 static int nvme_alloc_sq(struct nvme_dev *dev, u16 qid,
413 			    struct nvme_queue *nvmeq)
414 {
415 	struct nvme_command c;
416 	int flags = NVME_QUEUE_PHYS_CONTIG | NVME_SQ_PRIO_MEDIUM;
417 
418 	memset(&c, 0, sizeof(c));
419 	c.create_sq.opcode = nvme_admin_create_sq;
420 	c.create_sq.prp1 = cpu_to_le64((ulong)nvmeq->sq_cmds);
421 	c.create_sq.sqid = cpu_to_le16(qid);
422 	c.create_sq.qsize = cpu_to_le16(nvmeq->q_depth - 1);
423 	c.create_sq.sq_flags = cpu_to_le16(flags);
424 	c.create_sq.cqid = cpu_to_le16(qid);
425 
426 	return nvme_submit_admin_cmd(dev, &c, NULL);
427 }
428 
429 int nvme_identify(struct nvme_dev *dev, unsigned nsid,
430 		  unsigned cns, dma_addr_t dma_addr)
431 {
432 	struct nvme_command c;
433 	u32 page_size = dev->page_size;
434 	int offset = dma_addr & (page_size - 1);
435 	int length = sizeof(struct nvme_id_ctrl);
436 	int ret;
437 
438 	memset(&c, 0, sizeof(c));
439 	c.identify.opcode = nvme_admin_identify;
440 	c.identify.nsid = cpu_to_le32(nsid);
441 	c.identify.prp1 = cpu_to_le64(dma_addr);
442 
443 	length -= (page_size - offset);
444 	if (length <= 0) {
445 		c.identify.prp2 = 0;
446 	} else {
447 		dma_addr += (page_size - offset);
448 		c.identify.prp2 = cpu_to_le64(dma_addr);
449 	}
450 
451 	c.identify.cns = cpu_to_le32(cns);
452 
453 	ret = nvme_submit_admin_cmd(dev, &c, NULL);
454 	if (!ret)
455 		invalidate_dcache_range(dma_addr,
456 					dma_addr + sizeof(struct nvme_id_ctrl));
457 
458 	return ret;
459 }
460 
461 int nvme_get_features(struct nvme_dev *dev, unsigned fid, unsigned nsid,
462 		      dma_addr_t dma_addr, u32 *result)
463 {
464 	struct nvme_command c;
465 
466 	memset(&c, 0, sizeof(c));
467 	c.features.opcode = nvme_admin_get_features;
468 	c.features.nsid = cpu_to_le32(nsid);
469 	c.features.prp1 = cpu_to_le64(dma_addr);
470 	c.features.fid = cpu_to_le32(fid);
471 
472 	/*
473 	 * TODO: add cache invalidate operation when the size of
474 	 * the DMA buffer is known
475 	 */
476 
477 	return nvme_submit_admin_cmd(dev, &c, result);
478 }
479 
480 int nvme_set_features(struct nvme_dev *dev, unsigned fid, unsigned dword11,
481 		      dma_addr_t dma_addr, u32 *result)
482 {
483 	struct nvme_command c;
484 
485 	memset(&c, 0, sizeof(c));
486 	c.features.opcode = nvme_admin_set_features;
487 	c.features.prp1 = cpu_to_le64(dma_addr);
488 	c.features.fid = cpu_to_le32(fid);
489 	c.features.dword11 = cpu_to_le32(dword11);
490 
491 	/*
492 	 * TODO: add cache flush operation when the size of
493 	 * the DMA buffer is known
494 	 */
495 
496 	return nvme_submit_admin_cmd(dev, &c, result);
497 }
498 
499 static int nvme_create_queue(struct nvme_queue *nvmeq, int qid)
500 {
501 	struct nvme_dev *dev = nvmeq->dev;
502 	int result;
503 
504 	nvmeq->cq_vector = qid - 1;
505 	result = nvme_alloc_cq(dev, qid, nvmeq);
506 	if (result < 0)
507 		goto release_cq;
508 
509 	result = nvme_alloc_sq(dev, qid, nvmeq);
510 	if (result < 0)
511 		goto release_sq;
512 
513 	nvme_init_queue(nvmeq, qid);
514 
515 	return result;
516 
517  release_sq:
518 	nvme_delete_sq(dev, qid);
519  release_cq:
520 	nvme_delete_cq(dev, qid);
521 
522 	return result;
523 }
524 
525 static int nvme_set_queue_count(struct nvme_dev *dev, int count)
526 {
527 	int status;
528 	u32 result;
529 	u32 q_count = (count - 1) | ((count - 1) << 16);
530 
531 	status = nvme_set_features(dev, NVME_FEAT_NUM_QUEUES,
532 			q_count, 0, &result);
533 
534 	if (status < 0)
535 		return status;
536 	if (status > 1)
537 		return 0;
538 
539 	return min(result & 0xffff, result >> 16) + 1;
540 }
541 
542 static void nvme_create_io_queues(struct nvme_dev *dev)
543 {
544 	unsigned int i;
545 
546 	for (i = dev->queue_count; i <= dev->max_qid; i++)
547 		if (!nvme_alloc_queue(dev, i, dev->q_depth))
548 			break;
549 
550 	for (i = dev->online_queues; i <= dev->queue_count - 1; i++)
551 		if (nvme_create_queue(dev->queues[i], i))
552 			break;
553 }
554 
555 static int nvme_setup_io_queues(struct nvme_dev *dev)
556 {
557 	int nr_io_queues;
558 	int result;
559 
560 	nr_io_queues = 1;
561 	result = nvme_set_queue_count(dev, nr_io_queues);
562 	if (result <= 0)
563 		return result;
564 
565 	dev->max_qid = nr_io_queues;
566 
567 	/* Free previously allocated queues */
568 	nvme_free_queues(dev, nr_io_queues + 1);
569 	nvme_create_io_queues(dev);
570 
571 	return 0;
572 }
573 
574 static int nvme_get_info_from_identify(struct nvme_dev *dev)
575 {
576 	ALLOC_CACHE_ALIGN_BUFFER(char, buf, sizeof(struct nvme_id_ctrl));
577 	struct nvme_id_ctrl *ctrl = (struct nvme_id_ctrl *)buf;
578 	int ret;
579 	int shift = NVME_CAP_MPSMIN(dev->cap) + 12;
580 
581 	ret = nvme_identify(dev, 0, 1, (dma_addr_t)ctrl);
582 	if (ret)
583 		return -EIO;
584 
585 	dev->nn = le32_to_cpu(ctrl->nn);
586 	dev->vwc = ctrl->vwc;
587 	memcpy(dev->serial, ctrl->sn, sizeof(ctrl->sn));
588 	memcpy(dev->model, ctrl->mn, sizeof(ctrl->mn));
589 	memcpy(dev->firmware_rev, ctrl->fr, sizeof(ctrl->fr));
590 	if (ctrl->mdts)
591 		dev->max_transfer_shift = (ctrl->mdts + shift);
592 	else {
593 		/*
594 		 * Maximum Data Transfer Size (MDTS) field indicates the maximum
595 		 * data transfer size between the host and the controller. The
596 		 * host should not submit a command that exceeds this transfer
597 		 * size. The value is in units of the minimum memory page size
598 		 * and is reported as a power of two (2^n).
599 		 *
600 		 * The spec also says: a value of 0h indicates no restrictions
601 		 * on transfer size. But in nvme_blk_read/write() below we have
602 		 * the following algorithm for maximum number of logic blocks
603 		 * per transfer:
604 		 *
605 		 * u16 lbas = 1 << (dev->max_transfer_shift - ns->lba_shift);
606 		 *
607 		 * In order for lbas not to overflow, the maximum number is 15
608 		 * which means dev->max_transfer_shift = 15 + 9 (ns->lba_shift).
609 		 * Let's use 20 which provides 1MB size.
610 		 */
611 		dev->max_transfer_shift = 20;
612 	}
613 
614 	return 0;
615 }
616 
617 int nvme_scan_namespace(void)
618 {
619 	struct uclass *uc;
620 	struct udevice *dev;
621 	int ret;
622 
623 	ret = uclass_get(UCLASS_NVME, &uc);
624 	if (ret)
625 		return ret;
626 
627 	uclass_foreach_dev(dev, uc) {
628 		ret = device_probe(dev);
629 		if (ret)
630 			return ret;
631 	}
632 
633 	return 0;
634 }
635 
636 static int nvme_blk_probe(struct udevice *udev)
637 {
638 	struct nvme_dev *ndev = dev_get_priv(udev->parent);
639 	struct blk_desc *desc = dev_get_uclass_platdata(udev);
640 	struct nvme_ns *ns = dev_get_priv(udev);
641 	u8 flbas;
642 	ALLOC_CACHE_ALIGN_BUFFER(char, buf, sizeof(struct nvme_id_ns));
643 	struct nvme_id_ns *id = (struct nvme_id_ns *)buf;
644 	struct pci_child_platdata *pplat;
645 
646 	memset(ns, 0, sizeof(*ns));
647 	ns->dev = ndev;
648 	/* extract the namespace id from the block device name */
649 	ns->ns_id = trailing_strtol(udev->name) + 1;
650 	if (nvme_identify(ndev, ns->ns_id, 0, (dma_addr_t)id))
651 		return -EIO;
652 
653 	flbas = id->flbas & NVME_NS_FLBAS_LBA_MASK;
654 	ns->flbas = flbas;
655 	ns->lba_shift = id->lbaf[flbas].ds;
656 	ns->mode_select_num_blocks = le64_to_cpu(id->nsze);
657 	ns->mode_select_block_len = 1 << ns->lba_shift;
658 	list_add(&ns->list, &ndev->namespaces);
659 
660 	desc->lba = ns->mode_select_num_blocks;
661 	desc->log2blksz = ns->lba_shift;
662 	desc->blksz = 1 << ns->lba_shift;
663 	desc->bdev = udev;
664 	pplat = dev_get_parent_platdata(udev->parent);
665 	sprintf(desc->vendor, "0x%.4x", pplat->vendor);
666 	memcpy(desc->product, ndev->serial, sizeof(ndev->serial));
667 	memcpy(desc->revision, ndev->firmware_rev, sizeof(ndev->firmware_rev));
668 	part_init(desc);
669 
670 	return 0;
671 }
672 
673 static ulong nvme_blk_rw(struct udevice *udev, lbaint_t blknr,
674 			 lbaint_t blkcnt, void *buffer, bool read)
675 {
676 	struct nvme_ns *ns = dev_get_priv(udev);
677 	struct nvme_dev *dev = ns->dev;
678 	struct nvme_command c;
679 	struct blk_desc *desc = dev_get_uclass_platdata(udev);
680 	int status;
681 	u64 prp2;
682 	u64 total_len = blkcnt << desc->log2blksz;
683 	u64 temp_len = total_len;
684 
685 	u64 slba = blknr;
686 	u16 lbas = 1 << (dev->max_transfer_shift - ns->lba_shift);
687 	u64 total_lbas = blkcnt;
688 
689 	if (!read)
690 		flush_dcache_range((unsigned long)buffer,
691 				   (unsigned long)buffer + total_len);
692 
693 	c.rw.opcode = read ? nvme_cmd_read : nvme_cmd_write;
694 	c.rw.flags = 0;
695 	c.rw.nsid = cpu_to_le32(ns->ns_id);
696 	c.rw.control = 0;
697 	c.rw.dsmgmt = 0;
698 	c.rw.reftag = 0;
699 	c.rw.apptag = 0;
700 	c.rw.appmask = 0;
701 	c.rw.metadata = 0;
702 
703 	while (total_lbas) {
704 		if (total_lbas < lbas) {
705 			lbas = (u16)total_lbas;
706 			total_lbas = 0;
707 		} else {
708 			total_lbas -= lbas;
709 		}
710 
711 		if (nvme_setup_prps(dev, &prp2,
712 				    lbas << ns->lba_shift, (ulong)buffer))
713 			return -EIO;
714 		c.rw.slba = cpu_to_le64(slba);
715 		slba += lbas;
716 		c.rw.length = cpu_to_le16(lbas - 1);
717 		c.rw.prp1 = cpu_to_le64((ulong)buffer);
718 		c.rw.prp2 = cpu_to_le64(prp2);
719 		status = nvme_submit_sync_cmd(dev->queues[NVME_IO_Q],
720 				&c, NULL, IO_TIMEOUT);
721 		if (status)
722 			break;
723 		temp_len -= (u32)lbas << ns->lba_shift;
724 		buffer += lbas << ns->lba_shift;
725 	}
726 
727 	if (read)
728 		invalidate_dcache_range((unsigned long)buffer,
729 					(unsigned long)buffer + total_len);
730 
731 	return (total_len - temp_len) >> desc->log2blksz;
732 }
733 
734 static ulong nvme_blk_read(struct udevice *udev, lbaint_t blknr,
735 			   lbaint_t blkcnt, void *buffer)
736 {
737 	return nvme_blk_rw(udev, blknr, blkcnt, buffer, true);
738 }
739 
740 static ulong nvme_blk_write(struct udevice *udev, lbaint_t blknr,
741 			    lbaint_t blkcnt, const void *buffer)
742 {
743 	return nvme_blk_rw(udev, blknr, blkcnt, (void *)buffer, false);
744 }
745 
746 static const struct blk_ops nvme_blk_ops = {
747 	.read	= nvme_blk_read,
748 	.write	= nvme_blk_write,
749 };
750 
751 U_BOOT_DRIVER(nvme_blk) = {
752 	.name	= "nvme-blk",
753 	.id	= UCLASS_BLK,
754 	.probe	= nvme_blk_probe,
755 	.ops	= &nvme_blk_ops,
756 	.priv_auto_alloc_size = sizeof(struct nvme_ns),
757 };
758 
759 static int nvme_bind(struct udevice *udev)
760 {
761 	static int ndev_num;
762 	char name[20];
763 
764 	sprintf(name, "nvme#%d", ndev_num++);
765 
766 	return device_set_name(udev, name);
767 }
768 
769 static int nvme_probe(struct udevice *udev)
770 {
771 	int ret;
772 	struct nvme_dev *ndev = dev_get_priv(udev);
773 
774 	ndev->instance = trailing_strtol(udev->name);
775 
776 	INIT_LIST_HEAD(&ndev->namespaces);
777 	ndev->bar = dm_pci_map_bar(udev, PCI_BASE_ADDRESS_0,
778 			PCI_REGION_MEM);
779 	if (readl(&ndev->bar->csts) == -1) {
780 		ret = -ENODEV;
781 		printf("Error: %s: Out of memory!\n", udev->name);
782 		goto free_nvme;
783 	}
784 
785 	ndev->queues = malloc(NVME_Q_NUM * sizeof(struct nvme_queue *));
786 	if (!ndev->queues) {
787 		ret = -ENOMEM;
788 		printf("Error: %s: Out of memory!\n", udev->name);
789 		goto free_nvme;
790 	}
791 	memset(ndev->queues, 0, NVME_Q_NUM * sizeof(struct nvme_queue *));
792 
793 	ndev->prp_pool = malloc(MAX_PRP_POOL);
794 	if (!ndev->prp_pool) {
795 		ret = -ENOMEM;
796 		printf("Error: %s: Out of memory!\n", udev->name);
797 		goto free_nvme;
798 	}
799 	ndev->prp_entry_num = MAX_PRP_POOL >> 3;
800 
801 	ndev->cap = nvme_readq(&ndev->bar->cap);
802 	ndev->q_depth = min_t(int, NVME_CAP_MQES(ndev->cap) + 1, NVME_Q_DEPTH);
803 	ndev->db_stride = 1 << NVME_CAP_STRIDE(ndev->cap);
804 	ndev->dbs = ((void __iomem *)ndev->bar) + 4096;
805 
806 	ret = nvme_configure_admin_queue(ndev);
807 	if (ret)
808 		goto free_queue;
809 
810 	ret = nvme_setup_io_queues(ndev);
811 	if (ret)
812 		goto free_queue;
813 
814 	nvme_get_info_from_identify(ndev);
815 
816 	return 0;
817 
818 free_queue:
819 	free((void *)ndev->queues);
820 free_nvme:
821 	return ret;
822 }
823 
824 U_BOOT_DRIVER(nvme) = {
825 	.name	= "nvme",
826 	.id	= UCLASS_NVME,
827 	.bind	= nvme_bind,
828 	.probe	= nvme_probe,
829 	.priv_auto_alloc_size = sizeof(struct nvme_dev),
830 };
831 
832 struct pci_device_id nvme_supported[] = {
833 	{ PCI_DEVICE_CLASS(PCI_CLASS_STORAGE_EXPRESS, ~0) },
834 	{}
835 };
836 
837 U_BOOT_PCI_DEVICE(nvme, nvme_supported);
838