xref: /freebsd/sys/dev/nvme/nvme_ns.c (revision 148a8da8) 
1 /*-
2  * SPDX-License-Identifier: BSD-2-Clause-FreeBSD
3  *
4  * Copyright (C) 2012-2013 Intel Corporation
5  * All rights reserved.
6  *
7  * Redistribution and use in source and binary forms, with or without
8  * modification, are permitted provided that the following conditions
9  * are met:
10  * 1. Redistributions of source code must retain the above copyright
11  *    notice, this list of conditions and the following disclaimer.
12  * 2. Redistributions in binary form must reproduce the above copyright
13  *    notice, this list of conditions and the following disclaimer in the
14  *    documentation and/or other materials provided with the distribution.
15  *
16  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
17  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
18  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
19  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
20  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
21  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
22  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
23  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
24  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
25  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
26  * SUCH DAMAGE.
27  */
28 
29 #include <sys/cdefs.h>
30 __FBSDID("$FreeBSD$");
31 
32 #include <sys/param.h>
33 #include <sys/bio.h>
34 #include <sys/bus.h>
35 #include <sys/conf.h>
36 #include <sys/disk.h>
37 #include <sys/fcntl.h>
38 #include <sys/ioccom.h>
39 #include <sys/malloc.h>
40 #include <sys/module.h>
41 #include <sys/proc.h>
42 #include <sys/systm.h>
43 
44 #include <dev/pci/pcivar.h>
45 
46 #include <geom/geom.h>
47 
48 #include "nvme_private.h"
49 
50 static void		nvme_bio_child_inbed(struct bio *parent, int bio_error);
51 static void		nvme_bio_child_done(void *arg,
52 					    const struct nvme_completion *cpl);
53 static uint32_t		nvme_get_num_segments(uint64_t addr, uint64_t size,
54 					      uint32_t alignment);
55 static void		nvme_free_child_bios(int num_bios,
56 					     struct bio **child_bios);
57 static struct bio **	nvme_allocate_child_bios(int num_bios);
58 static struct bio **	nvme_construct_child_bios(struct bio *bp,
59 						  uint32_t alignment,
60 						  int *num_bios);
61 static int		nvme_ns_split_bio(struct nvme_namespace *ns,
62 					  struct bio *bp,
63 					  uint32_t alignment);
64 
65 static int
66 nvme_ns_ioctl(struct cdev *cdev, u_long cmd, caddr_t arg, int flag,
67     struct thread *td)
68 {
69 	struct nvme_namespace			*ns;
70 	struct nvme_controller			*ctrlr;
71 	struct nvme_pt_command			*pt;
72 
73 	ns = cdev->si_drv1;
74 	ctrlr = ns->ctrlr;
75 
76 	switch (cmd) {
77 	case NVME_IO_TEST:
78 	case NVME_BIO_TEST:
79 		nvme_ns_test(ns, cmd, arg);
80 		break;
81 	case NVME_PASSTHROUGH_CMD:
82 		pt = (struct nvme_pt_command *)arg;
83 		return (nvme_ctrlr_passthrough_cmd(ctrlr, pt, ns->id,
84 		    1 /* is_user_buffer */, 0 /* is_admin_cmd */));
85 	case DIOCGMEDIASIZE:
86 		*(off_t *)arg = (off_t)nvme_ns_get_size(ns);
87 		break;
88 	case DIOCGSECTORSIZE:
89 		*(u_int *)arg = nvme_ns_get_sector_size(ns);
90 		break;
91 	default:
92 		return (ENOTTY);
93 	}
94 
95 	return (0);
96 }
97 
98 static int
99 nvme_ns_open(struct cdev *dev __unused, int flags, int fmt __unused,
100     struct thread *td)
101 {
102 	int error = 0;
103 
104 	if (flags & FWRITE)
105 		error = securelevel_gt(td->td_ucred, 0);
106 
107 	return (error);
108 }
109 
110 static int
111 nvme_ns_close(struct cdev *dev __unused, int flags, int fmt __unused,
112     struct thread *td)
113 {
114 
115 	return (0);
116 }
117 
118 static void
119 nvme_ns_strategy_done(void *arg, const struct nvme_completion *cpl)
120 {
121 	struct bio *bp = arg;
122 
123 	/*
124 	 * TODO: add more extensive translation of NVMe status codes
125 	 *  to different bio error codes (i.e. EIO, EINVAL, etc.)
126 	 */
127 	if (nvme_completion_is_error(cpl)) {
128 		bp->bio_error = EIO;
129 		bp->bio_flags |= BIO_ERROR;
130 		bp->bio_resid = bp->bio_bcount;
131 	} else
132 		bp->bio_resid = 0;
133 
134 	biodone(bp);
135 }
136 
137 static void
138 nvme_ns_strategy(struct bio *bp)
139 {
140 	struct nvme_namespace	*ns;
141 	int			err;
142 
143 	ns = bp->bio_dev->si_drv1;
144 	err = nvme_ns_bio_process(ns, bp, nvme_ns_strategy_done);
145 
146 	if (err) {
147 		bp->bio_error = err;
148 		bp->bio_flags |= BIO_ERROR;
149 		bp->bio_resid = bp->bio_bcount;
150 		biodone(bp);
151 	}
152 
153 }
154 
155 static struct cdevsw nvme_ns_cdevsw = {
156 	.d_version =	D_VERSION,
157 	.d_flags =	D_DISK,
158 	.d_read =	physread,
159 	.d_write =	physwrite,
160 	.d_open =	nvme_ns_open,
161 	.d_close =	nvme_ns_close,
162 	.d_strategy =	nvme_ns_strategy,
163 	.d_ioctl =	nvme_ns_ioctl
164 };
165 
166 uint32_t
167 nvme_ns_get_max_io_xfer_size(struct nvme_namespace *ns)
168 {
169 	return ns->ctrlr->max_xfer_size;
170 }
171 
172 uint32_t
173 nvme_ns_get_sector_size(struct nvme_namespace *ns)
174 {
175 	uint8_t flbas_fmt, lbads;
176 
177 	flbas_fmt = (ns->data.flbas >> NVME_NS_DATA_FLBAS_FORMAT_SHIFT) &
178 		NVME_NS_DATA_FLBAS_FORMAT_MASK;
179 	lbads = (ns->data.lbaf[flbas_fmt] >> NVME_NS_DATA_LBAF_LBADS_SHIFT) &
180 		NVME_NS_DATA_LBAF_LBADS_MASK;
181 
182 	return (1 << lbads);
183 }
184 
185 uint64_t
186 nvme_ns_get_num_sectors(struct nvme_namespace *ns)
187 {
188 	return (ns->data.nsze);
189 }
190 
191 uint64_t
192 nvme_ns_get_size(struct nvme_namespace *ns)
193 {
194 	return (nvme_ns_get_num_sectors(ns) * nvme_ns_get_sector_size(ns));
195 }
196 
197 uint32_t
198 nvme_ns_get_flags(struct nvme_namespace *ns)
199 {
200 	return (ns->flags);
201 }
202 
203 const char *
204 nvme_ns_get_serial_number(struct nvme_namespace *ns)
205 {
206 	return ((const char *)ns->ctrlr->cdata.sn);
207 }
208 
209 const char *
210 nvme_ns_get_model_number(struct nvme_namespace *ns)
211 {
212 	return ((const char *)ns->ctrlr->cdata.mn);
213 }
214 
215 const struct nvme_namespace_data *
216 nvme_ns_get_data(struct nvme_namespace *ns)
217 {
218 
219 	return (&ns->data);
220 }
221 
222 uint32_t
223 nvme_ns_get_stripesize(struct nvme_namespace *ns)
224 {
225 
226 	return (ns->stripesize);
227 }
228 
229 static void
230 nvme_ns_bio_done(void *arg, const struct nvme_completion *status)
231 {
232 	struct bio	*bp = arg;
233 	nvme_cb_fn_t	bp_cb_fn;
234 
235 	bp_cb_fn = bp->bio_driver1;
236 
237 	if (bp->bio_driver2)
238 		free(bp->bio_driver2, M_NVME);
239 
240 	if (nvme_completion_is_error(status)) {
241 		bp->bio_flags |= BIO_ERROR;
242 		if (bp->bio_error == 0)
243 			bp->bio_error = EIO;
244 	}
245 
246 	if ((bp->bio_flags & BIO_ERROR) == 0)
247 		bp->bio_resid = 0;
248 	else
249 		bp->bio_resid = bp->bio_bcount;
250 
251 	bp_cb_fn(bp, status);
252 }
253 
254 static void
255 nvme_bio_child_inbed(struct bio *parent, int bio_error)
256 {
257 	struct nvme_completion	parent_cpl;
258 	int			children, inbed;
259 
260 	if (bio_error != 0) {
261 		parent->bio_flags |= BIO_ERROR;
262 		parent->bio_error = bio_error;
263 	}
264 
265 	/*
266 	 * atomic_fetchadd will return value before adding 1, so we still
267 	 *  must add 1 to get the updated inbed number.  Save bio_children
268 	 *  before incrementing to guard against race conditions when
269 	 *  two children bios complete on different queues.
270 	 */
271 	children = atomic_load_acq_int(&parent->bio_children);
272 	inbed = atomic_fetchadd_int(&parent->bio_inbed, 1) + 1;
273 	if (inbed == children) {
274 		bzero(&parent_cpl, sizeof(parent_cpl));
275 		if (parent->bio_flags & BIO_ERROR) {
276 			parent_cpl.status &= ~(NVME_STATUS_SC_MASK << NVME_STATUS_SC_SHIFT);
277 			parent_cpl.status |= (NVME_SC_DATA_TRANSFER_ERROR) << NVME_STATUS_SC_SHIFT;
278 		}
279 		nvme_ns_bio_done(parent, &parent_cpl);
280 	}
281 }
282 
283 static void
284 nvme_bio_child_done(void *arg, const struct nvme_completion *cpl)
285 {
286 	struct bio		*child = arg;
287 	struct bio		*parent;
288 	int			bio_error;
289 
290 	parent = child->bio_parent;
291 	g_destroy_bio(child);
292 	bio_error = nvme_completion_is_error(cpl) ? EIO : 0;
293 	nvme_bio_child_inbed(parent, bio_error);
294 }
295 
296 static uint32_t
297 nvme_get_num_segments(uint64_t addr, uint64_t size, uint32_t align)
298 {
299 	uint32_t	num_segs, offset, remainder;
300 
301 	if (align == 0)
302 		return (1);
303 
304 	KASSERT((align & (align - 1)) == 0, ("alignment not power of 2\n"));
305 
306 	num_segs = size / align;
307 	remainder = size & (align - 1);
308 	offset = addr & (align - 1);
309 	if (remainder > 0 || offset > 0)
310 		num_segs += 1 + (remainder + offset - 1) / align;
311 	return (num_segs);
312 }
313 
314 static void
315 nvme_free_child_bios(int num_bios, struct bio **child_bios)
316 {
317 	int i;
318 
319 	for (i = 0; i < num_bios; i++) {
320 		if (child_bios[i] != NULL)
321 			g_destroy_bio(child_bios[i]);
322 	}
323 
324 	free(child_bios, M_NVME);
325 }
326 
327 static struct bio **
328 nvme_allocate_child_bios(int num_bios)
329 {
330 	struct bio **child_bios;
331 	int err = 0, i;
332 
333 	child_bios = malloc(num_bios * sizeof(struct bio *), M_NVME, M_NOWAIT);
334 	if (child_bios == NULL)
335 		return (NULL);
336 
337 	for (i = 0; i < num_bios; i++) {
338 		child_bios[i] = g_new_bio();
339 		if (child_bios[i] == NULL)
340 			err = ENOMEM;
341 	}
342 
343 	if (err == ENOMEM) {
344 		nvme_free_child_bios(num_bios, child_bios);
345 		return (NULL);
346 	}
347 
348 	return (child_bios);
349 }
350 
351 static struct bio **
352 nvme_construct_child_bios(struct bio *bp, uint32_t alignment, int *num_bios)
353 {
354 	struct bio	**child_bios;
355 	struct bio	*child;
356 	uint64_t	cur_offset;
357 	caddr_t		data;
358 	uint32_t	rem_bcount;
359 	int		i;
360 	struct vm_page	**ma;
361 	uint32_t	ma_offset;
362 
363 	*num_bios = nvme_get_num_segments(bp->bio_offset, bp->bio_bcount,
364 	    alignment);
365 	child_bios = nvme_allocate_child_bios(*num_bios);
366 	if (child_bios == NULL)
367 		return (NULL);
368 
369 	bp->bio_children = *num_bios;
370 	bp->bio_inbed = 0;
371 	cur_offset = bp->bio_offset;
372 	rem_bcount = bp->bio_bcount;
373 	data = bp->bio_data;
374 	ma_offset = bp->bio_ma_offset;
375 	ma = bp->bio_ma;
376 
377 	for (i = 0; i < *num_bios; i++) {
378 		child = child_bios[i];
379 		child->bio_parent = bp;
380 		child->bio_cmd = bp->bio_cmd;
381 		child->bio_offset = cur_offset;
382 		child->bio_bcount = min(rem_bcount,
383 		    alignment - (cur_offset & (alignment - 1)));
384 		child->bio_flags = bp->bio_flags;
385 		if (bp->bio_flags & BIO_UNMAPPED) {
386 			child->bio_ma_offset = ma_offset;
387 			child->bio_ma = ma;
388 			child->bio_ma_n =
389 			    nvme_get_num_segments(child->bio_ma_offset,
390 				child->bio_bcount, PAGE_SIZE);
391 			ma_offset = (ma_offset + child->bio_bcount) &
392 			    PAGE_MASK;
393 			ma += child->bio_ma_n;
394 			if (ma_offset != 0)
395 				ma -= 1;
396 		} else {
397 			child->bio_data = data;
398 			data += child->bio_bcount;
399 		}
400 		cur_offset += child->bio_bcount;
401 		rem_bcount -= child->bio_bcount;
402 	}
403 
404 	return (child_bios);
405 }
406 
407 static int
408 nvme_ns_split_bio(struct nvme_namespace *ns, struct bio *bp,
409     uint32_t alignment)
410 {
411 	struct bio	*child;
412 	struct bio	**child_bios;
413 	int		err, i, num_bios;
414 
415 	child_bios = nvme_construct_child_bios(bp, alignment, &num_bios);
416 	if (child_bios == NULL)
417 		return (ENOMEM);
418 
419 	for (i = 0; i < num_bios; i++) {
420 		child = child_bios[i];
421 		err = nvme_ns_bio_process(ns, child, nvme_bio_child_done);
422 		if (err != 0) {
423 			nvme_bio_child_inbed(bp, err);
424 			g_destroy_bio(child);
425 		}
426 	}
427 
428 	free(child_bios, M_NVME);
429 	return (0);
430 }
431 
432 int
433 nvme_ns_bio_process(struct nvme_namespace *ns, struct bio *bp,
434 	nvme_cb_fn_t cb_fn)
435 {
436 	struct nvme_dsm_range	*dsm_range;
437 	uint32_t		num_bios;
438 	int			err;
439 
440 	bp->bio_driver1 = cb_fn;
441 
442 	if (ns->stripesize > 0 &&
443 	    (bp->bio_cmd == BIO_READ || bp->bio_cmd == BIO_WRITE)) {
444 		num_bios = nvme_get_num_segments(bp->bio_offset,
445 		    bp->bio_bcount, ns->stripesize);
446 		if (num_bios > 1)
447 			return (nvme_ns_split_bio(ns, bp, ns->stripesize));
448 	}
449 
450 	switch (bp->bio_cmd) {
451 	case BIO_READ:
452 		err = nvme_ns_cmd_read_bio(ns, bp, nvme_ns_bio_done, bp);
453 		break;
454 	case BIO_WRITE:
455 		err = nvme_ns_cmd_write_bio(ns, bp, nvme_ns_bio_done, bp);
456 		break;
457 	case BIO_FLUSH:
458 		err = nvme_ns_cmd_flush(ns, nvme_ns_bio_done, bp);
459 		break;
460 	case BIO_DELETE:
461 		dsm_range =
462 		    malloc(sizeof(struct nvme_dsm_range), M_NVME,
463 		    M_ZERO | M_WAITOK);
464 		if (!dsm_range) {
465 			err = ENOMEM;
466 			break;
467 		}
468 		dsm_range->length =
469 		    htole32(bp->bio_bcount/nvme_ns_get_sector_size(ns));
470 		dsm_range->starting_lba =
471 		    htole64(bp->bio_offset/nvme_ns_get_sector_size(ns));
472 		bp->bio_driver2 = dsm_range;
473 		err = nvme_ns_cmd_deallocate(ns, dsm_range, 1,
474 			nvme_ns_bio_done, bp);
475 		if (err != 0)
476 			free(dsm_range, M_NVME);
477 		break;
478 	default:
479 		err = EIO;
480 		break;
481 	}
482 
483 	return (err);
484 }
485 
486 int
487 nvme_ns_construct(struct nvme_namespace *ns, uint32_t id,
488     struct nvme_controller *ctrlr)
489 {
490 	struct make_dev_args                    md_args;
491 	struct nvme_completion_poll_status	status;
492 	int                                     res;
493 	int					unit;
494 	uint8_t					flbas_fmt;
495 	uint8_t					vwc_present;
496 
497 	ns->ctrlr = ctrlr;
498 	ns->id = id;
499 	ns->stripesize = 0;
500 
501 	/*
502 	 * Older Intel devices advertise in vendor specific space an alignment
503 	 * that improves performance.  If present use for the stripe size.  NVMe
504 	 * 1.3 standardized this as NOIOB, and newer Intel drives use that.
505 	 */
506 	switch (pci_get_devid(ctrlr->dev)) {
507 	case 0x09538086:		/* Intel DC PC3500 */
508 	case 0x0a538086:		/* Intel DC PC3520 */
509 	case 0x0a548086:		/* Intel DC PC4500 */
510 	case 0x0a558086:		/* Dell Intel P4600 */
511 		if (ctrlr->cdata.vs[3] != 0)
512 			ns->stripesize =
513 			    (1 << ctrlr->cdata.vs[3]) * ctrlr->min_page_size;
514 		break;
515 	default:
516 		break;
517 	}
518 
519 	/*
520 	 * Namespaces are reconstructed after a controller reset, so check
521 	 *  to make sure we only call mtx_init once on each mtx.
522 	 *
523 	 * TODO: Move this somewhere where it gets called at controller
524 	 *  construction time, which is not invoked as part of each
525 	 *  controller reset.
526 	 */
527 	if (!mtx_initialized(&ns->lock))
528 		mtx_init(&ns->lock, "nvme ns lock", NULL, MTX_DEF);
529 
530 	status.done = 0;
531 	nvme_ctrlr_cmd_identify_namespace(ctrlr, id, &ns->data,
532 	    nvme_completion_poll_cb, &status);
533 	while (!atomic_load_acq_int(&status.done))
534 		pause("nvme", 1);
535 	if (nvme_completion_is_error(&status.cpl)) {
536 		nvme_printf(ctrlr, "nvme_identify_namespace failed\n");
537 		return (ENXIO);
538 	}
539 
540 	/* Convert data to host endian */
541 	nvme_namespace_data_swapbytes(&ns->data);
542 
543 	/*
544 	 * If the size of is zero, chances are this isn't a valid
545 	 * namespace (eg one that's not been configured yet). The
546 	 * standard says the entire id will be zeros, so this is a
547 	 * cheap way to test for that.
548 	 */
549 	if (ns->data.nsze == 0)
550 		return (ENXIO);
551 
552 	flbas_fmt = (ns->data.flbas >> NVME_NS_DATA_FLBAS_FORMAT_SHIFT) &
553 		NVME_NS_DATA_FLBAS_FORMAT_MASK;
554 	/*
555 	 * Note: format is a 0-based value, so > is appropriate here,
556 	 *  not >=.
557 	 */
558 	if (flbas_fmt > ns->data.nlbaf) {
559 		printf("lba format %d exceeds number supported (%d)\n",
560 		    flbas_fmt, ns->data.nlbaf + 1);
561 		return (ENXIO);
562 	}
563 
564 	if (nvme_ctrlr_has_dataset_mgmt(&ctrlr->cdata))
565 		ns->flags |= NVME_NS_DEALLOCATE_SUPPORTED;
566 
567 	vwc_present = (ctrlr->cdata.vwc >> NVME_CTRLR_DATA_VWC_PRESENT_SHIFT) &
568 		NVME_CTRLR_DATA_VWC_PRESENT_MASK;
569 	if (vwc_present)
570 		ns->flags |= NVME_NS_FLUSH_SUPPORTED;
571 
572 	/*
573 	 * cdev may have already been created, if we are reconstructing the
574 	 *  namespace after a controller-level reset.
575 	 */
576 	if (ns->cdev != NULL)
577 		return (0);
578 
579 	/*
580 	 * Namespace IDs start at 1, so we need to subtract 1 to create a
581 	 *  correct unit number.
582 	 */
583 	unit = device_get_unit(ctrlr->dev) * NVME_MAX_NAMESPACES + ns->id - 1;
584 
585 	make_dev_args_init(&md_args);
586 	md_args.mda_devsw = &nvme_ns_cdevsw;
587 	md_args.mda_unit = unit;
588 	md_args.mda_mode = 0600;
589 	md_args.mda_si_drv1 = ns;
590 	res = make_dev_s(&md_args, &ns->cdev, "nvme%dns%d",
591 	    device_get_unit(ctrlr->dev), ns->id);
592 	if (res != 0)
593 		return (ENXIO);
594 
595 	ns->cdev->si_flags |= SI_UNMAPPED;
596 
597 	return (0);
598 }
599 
600 void nvme_ns_destruct(struct nvme_namespace *ns)
601 {
602 
603 	if (ns->cdev != NULL)
604 		destroy_dev(ns->cdev);
605 }
606