xref: /spdk/module/bdev/gpt/vbdev_gpt.c (revision 307b8c112ffd90a26d53dd15fad67bd9038ef526)
1 /*   SPDX-License-Identifier: BSD-3-Clause
2  *   Copyright (c) Intel Corporation.
3  *   All rights reserved.
4  */
5 
6 /*
7  * This driver reads a GPT partition table from a bdev and exposes a virtual block device for
8  * each partition.
9  */
10 
11 #include "gpt.h"
12 
13 #include "spdk/endian.h"
14 #include "spdk/env.h"
15 #include "spdk/thread.h"
16 #include "spdk/rpc.h"
17 #include "spdk/string.h"
18 #include "spdk/util.h"
19 
20 #include "spdk/bdev_module.h"
21 #include "spdk/log.h"
22 
23 static int vbdev_gpt_init(void);
24 static void vbdev_gpt_examine(struct spdk_bdev *bdev);
25 static int vbdev_gpt_get_ctx_size(void);
26 
27 static struct spdk_bdev_module gpt_if = {
28 	.name = "gpt",
29 	.module_init = vbdev_gpt_init,
30 	.get_ctx_size = vbdev_gpt_get_ctx_size,
31 	.examine_disk = vbdev_gpt_examine,
32 
33 };
34 SPDK_BDEV_MODULE_REGISTER(gpt, &gpt_if)
35 
36 /* Base block device gpt context */
37 struct gpt_base {
38 	struct spdk_gpt			gpt;
39 	struct spdk_bdev_part_base	*part_base;
40 	SPDK_BDEV_PART_TAILQ		parts;
41 
42 	/* This channel is only used for reading the partition table. */
43 	struct spdk_io_channel		*ch;
44 };
45 
46 /* Context for each gpt virtual bdev */
47 struct gpt_disk {
48 	struct spdk_bdev_part	part;
49 	uint32_t		partition_index;
50 };
51 
52 struct gpt_channel {
53 	struct spdk_bdev_part_channel	part_ch;
54 };
55 
56 struct gpt_io {
57 	struct spdk_io_channel *ch;
58 	struct spdk_bdev_io *bdev_io;
59 
60 	/* for bdev_io_wait */
61 	struct spdk_bdev_io_wait_entry bdev_io_wait;
62 };
63 
64 static void
65 gpt_base_free(void *ctx)
66 {
67 	struct gpt_base *gpt_base = ctx;
68 
69 	spdk_free(gpt_base->gpt.buf);
70 	free(gpt_base);
71 }
72 
73 static void
74 gpt_base_bdev_hotremove_cb(void *_part_base)
75 {
76 	struct spdk_bdev_part_base *part_base = _part_base;
77 	struct gpt_base *gpt_base = spdk_bdev_part_base_get_ctx(part_base);
78 
79 	spdk_bdev_part_base_hotremove(part_base, &gpt_base->parts);
80 }
81 
82 static int vbdev_gpt_destruct(void *ctx);
83 static void vbdev_gpt_submit_request(struct spdk_io_channel *_ch, struct spdk_bdev_io *bdev_io);
84 static int vbdev_gpt_dump_info_json(void *ctx, struct spdk_json_write_ctx *w);
85 static int vbdev_gpt_get_memory_domains(void *ctx, struct spdk_memory_domain **domains,
86 					int array_size);
87 
88 static struct spdk_bdev_fn_table vbdev_gpt_fn_table = {
89 	.destruct		= vbdev_gpt_destruct,
90 	.submit_request		= vbdev_gpt_submit_request,
91 	.dump_info_json		= vbdev_gpt_dump_info_json,
92 	.get_memory_domains	= vbdev_gpt_get_memory_domains,
93 };
94 
95 static struct gpt_base *
96 gpt_base_bdev_init(struct spdk_bdev *bdev)
97 {
98 	struct gpt_base *gpt_base;
99 	struct spdk_gpt *gpt;
100 	int rc;
101 
102 	gpt_base = calloc(1, sizeof(*gpt_base));
103 	if (!gpt_base) {
104 		SPDK_ERRLOG("Cannot alloc memory for gpt_base pointer\n");
105 		return NULL;
106 	}
107 
108 	TAILQ_INIT(&gpt_base->parts);
109 	rc = spdk_bdev_part_base_construct_ext(spdk_bdev_get_name(bdev),
110 					       gpt_base_bdev_hotremove_cb,
111 					       &gpt_if, &vbdev_gpt_fn_table,
112 					       &gpt_base->parts, gpt_base_free, gpt_base,
113 					       sizeof(struct gpt_channel), NULL, NULL, &gpt_base->part_base);
114 	if (rc != 0) {
115 		free(gpt_base);
116 		SPDK_ERRLOG("cannot construct gpt_base");
117 		return NULL;
118 	}
119 
120 	gpt = &gpt_base->gpt;
121 	gpt->parse_phase = SPDK_GPT_PARSE_PHASE_PRIMARY;
122 	gpt->buf_size = spdk_max(SPDK_GPT_BUFFER_SIZE, bdev->blocklen);
123 	gpt->buf = spdk_zmalloc(gpt->buf_size, spdk_bdev_get_buf_align(bdev), NULL,
124 				SPDK_ENV_LCORE_ID_ANY, SPDK_MALLOC_DMA);
125 	if (!gpt->buf) {
126 		SPDK_ERRLOG("Cannot alloc buf\n");
127 		spdk_bdev_part_base_free(gpt_base->part_base);
128 		return NULL;
129 	}
130 
131 	gpt->sector_size = bdev->blocklen;
132 	gpt->total_sectors = bdev->blockcnt;
133 	gpt->lba_start = 0;
134 	gpt->lba_end = gpt->total_sectors - 1;
135 
136 	return gpt_base;
137 }
138 
139 static int
140 vbdev_gpt_destruct(void *ctx)
141 {
142 	struct gpt_disk *gpt_disk = ctx;
143 
144 	return spdk_bdev_part_free(&gpt_disk->part);
145 }
146 
147 static void _vbdev_gpt_submit_request(struct spdk_io_channel *_ch, struct spdk_bdev_io *bdev_io);
148 
149 static void
150 vbdev_gpt_resubmit_request(void *arg)
151 {
152 	struct gpt_io *io = (struct gpt_io *)arg;
153 
154 	_vbdev_gpt_submit_request(io->ch, io->bdev_io);
155 }
156 
157 static void
158 vbdev_gpt_queue_io(struct gpt_io *io)
159 {
160 	struct gpt_channel *ch = spdk_io_channel_get_ctx(io->ch);
161 	int rc;
162 
163 	io->bdev_io_wait.bdev = io->bdev_io->bdev;
164 	io->bdev_io_wait.cb_fn = vbdev_gpt_resubmit_request;
165 	io->bdev_io_wait.cb_arg = io;
166 
167 	rc = spdk_bdev_queue_io_wait(io->bdev_io->bdev,
168 				     ch->part_ch.base_ch, &io->bdev_io_wait);
169 	if (rc != 0) {
170 		SPDK_ERRLOG("Queue io failed in vbdev_gpt_queue_io, rc=%d.\n", rc);
171 		spdk_bdev_io_complete(io->bdev_io, SPDK_BDEV_IO_STATUS_FAILED);
172 	}
173 }
174 
175 static void
176 vbdev_gpt_get_buf_cb(struct spdk_io_channel *ch, struct spdk_bdev_io *bdev_io, bool success)
177 {
178 	if (!success) {
179 		spdk_bdev_io_complete(bdev_io, SPDK_BDEV_IO_STATUS_FAILED);
180 		return;
181 	}
182 
183 	_vbdev_gpt_submit_request(ch, bdev_io);
184 }
185 
186 static void
187 _vbdev_gpt_submit_request(struct spdk_io_channel *_ch, struct spdk_bdev_io *bdev_io)
188 {
189 	struct gpt_channel *ch = spdk_io_channel_get_ctx(_ch);
190 	struct gpt_io *io = (struct gpt_io *)bdev_io->driver_ctx;
191 	int rc;
192 
193 	rc = spdk_bdev_part_submit_request(&ch->part_ch, bdev_io);
194 	if (rc) {
195 		if (rc == -ENOMEM) {
196 			SPDK_DEBUGLOG(vbdev_gpt, "gpt: no memory, queue io\n");
197 			io->ch = _ch;
198 			io->bdev_io = bdev_io;
199 			vbdev_gpt_queue_io(io);
200 		} else {
201 			SPDK_ERRLOG("gpt: error on bdev_io submission, rc=%d.\n", rc);
202 			spdk_bdev_io_complete(bdev_io, SPDK_BDEV_IO_STATUS_FAILED);
203 		}
204 	}
205 }
206 
207 static void
208 vbdev_gpt_submit_request(struct spdk_io_channel *_ch, struct spdk_bdev_io *bdev_io)
209 {
210 	switch (bdev_io->type) {
211 	case SPDK_BDEV_IO_TYPE_READ:
212 		spdk_bdev_io_get_buf(bdev_io, vbdev_gpt_get_buf_cb,
213 				     bdev_io->u.bdev.num_blocks * bdev_io->bdev->blocklen);
214 		break;
215 	default:
216 		_vbdev_gpt_submit_request(_ch, bdev_io);
217 		break;
218 	}
219 }
220 
221 static void
222 write_guid(struct spdk_json_write_ctx *w, const struct spdk_gpt_guid *guid)
223 {
224 	spdk_json_write_string_fmt(w, "%08x-%04x-%04x-%04x-%04x%08x",
225 				   from_le32(&guid->raw[0]),
226 				   from_le16(&guid->raw[4]),
227 				   from_le16(&guid->raw[6]),
228 				   from_be16(&guid->raw[8]),
229 				   from_be16(&guid->raw[10]),
230 				   from_be32(&guid->raw[12]));
231 }
232 
233 static void
234 write_string_utf16le(struct spdk_json_write_ctx *w, const uint16_t *str, size_t max_len)
235 {
236 	size_t len;
237 	const uint16_t *p;
238 
239 	for (len = 0, p = str; len < max_len && *p; p++) {
240 		len++;
241 	}
242 
243 	spdk_json_write_string_utf16le_raw(w, str, len);
244 }
245 
246 static int
247 vbdev_gpt_dump_info_json(void *ctx, struct spdk_json_write_ctx *w)
248 {
249 	struct gpt_disk *gpt_disk = SPDK_CONTAINEROF(ctx, struct gpt_disk, part);
250 	struct spdk_bdev_part_base *base_bdev = spdk_bdev_part_get_base(&gpt_disk->part);
251 	struct gpt_base *gpt_base = spdk_bdev_part_base_get_ctx(base_bdev);
252 	struct spdk_bdev *part_base_bdev = spdk_bdev_part_base_get_bdev(base_bdev);
253 	struct spdk_gpt *gpt = &gpt_base->gpt;
254 	struct spdk_gpt_partition_entry *gpt_entry = &gpt->partitions[gpt_disk->partition_index];
255 	uint64_t offset_blocks = spdk_bdev_part_get_offset_blocks(&gpt_disk->part);
256 
257 	spdk_json_write_named_object_begin(w, "gpt");
258 
259 	spdk_json_write_named_string(w, "base_bdev", spdk_bdev_get_name(part_base_bdev));
260 
261 	spdk_json_write_named_uint64(w, "offset_blocks", offset_blocks);
262 
263 	spdk_json_write_name(w, "partition_type_guid");
264 	write_guid(w, &gpt_entry->part_type_guid);
265 
266 	spdk_json_write_name(w, "unique_partition_guid");
267 	write_guid(w, &gpt_entry->unique_partition_guid);
268 
269 	spdk_json_write_name(w, "partition_name");
270 	write_string_utf16le(w, gpt_entry->partition_name, SPDK_COUNTOF(gpt_entry->partition_name));
271 
272 	spdk_json_write_object_end(w);
273 
274 	return 0;
275 }
276 
277 static int
278 vbdev_gpt_get_memory_domains(void *ctx, struct spdk_memory_domain **domains, int array_size)
279 {
280 	struct gpt_disk *gpt_disk = SPDK_CONTAINEROF(ctx, struct gpt_disk, part);
281 	struct spdk_bdev_part_base *part_base = spdk_bdev_part_get_base(&gpt_disk->part);
282 	struct spdk_bdev *part_base_bdev = spdk_bdev_part_base_get_bdev(part_base);
283 
284 	if (part_base_bdev->dif_check_flags & SPDK_DIF_FLAGS_REFTAG_CHECK) {
285 		/* bdev_part remaps reftag and touches metadata buffer, that means it can't support memory domains
286 		 * if dif is enabled */
287 		return 0;
288 	}
289 
290 	return spdk_bdev_get_memory_domains(part_base_bdev, domains, array_size);
291 }
292 
293 static int
294 vbdev_gpt_create_bdevs(struct gpt_base *gpt_base)
295 {
296 	uint32_t num_partition_entries;
297 	uint64_t i, head_lba_start, head_lba_end;
298 	uint32_t num_partitions;
299 	struct spdk_gpt_partition_entry *p;
300 	struct gpt_disk *d;
301 	struct spdk_gpt *gpt;
302 	char *name;
303 	struct spdk_bdev *base_bdev;
304 	int rc;
305 
306 	gpt = &gpt_base->gpt;
307 	num_partition_entries = from_le32(&gpt->header->num_partition_entries);
308 	head_lba_start = from_le64(&gpt->header->first_usable_lba);
309 	head_lba_end = from_le64(&gpt->header->last_usable_lba);
310 	num_partitions = 0;
311 
312 	for (i = 0; i < num_partition_entries; i++) {
313 		p = &gpt->partitions[i];
314 		uint64_t lba_start = from_le64(&p->starting_lba);
315 		uint64_t lba_end = from_le64(&p->ending_lba);
316 
317 		if (!SPDK_GPT_GUID_EQUAL(&gpt->partitions[i].part_type_guid,
318 					 &SPDK_GPT_PART_TYPE_GUID) ||
319 		    lba_start == 0) {
320 			continue;
321 		}
322 		if (lba_start < head_lba_start || lba_end > head_lba_end) {
323 			continue;
324 		}
325 
326 		d = calloc(1, sizeof(*d));
327 		if (!d) {
328 			SPDK_ERRLOG("Memory allocation failure\n");
329 			return -1;
330 		}
331 
332 		/* index start at 1 instead of 0 to match the existing style */
333 		base_bdev = spdk_bdev_part_base_get_bdev(gpt_base->part_base);
334 		name = spdk_sprintf_alloc("%sp%" PRIu64, spdk_bdev_get_name(base_bdev), i + 1);
335 		if (!name) {
336 			SPDK_ERRLOG("name allocation failure\n");
337 			free(d);
338 			return -1;
339 		}
340 
341 		rc = spdk_bdev_part_construct(&d->part, gpt_base->part_base, name,
342 					      lba_start, lba_end - lba_start, "GPT Disk");
343 		free(name);
344 		if (rc) {
345 			SPDK_ERRLOG("could not construct bdev part\n");
346 			/* spdk_bdev_part_construct will free name on failure */
347 			free(d);
348 			return -1;
349 		}
350 		num_partitions++;
351 		d->partition_index = i;
352 	}
353 
354 	return num_partitions;
355 }
356 
357 static void
358 gpt_read_secondary_table_complete(struct spdk_bdev_io *bdev_io, bool status, void *arg)
359 {
360 	struct gpt_base *gpt_base = (struct gpt_base *)arg;
361 	struct spdk_bdev *bdev = spdk_bdev_part_base_get_bdev(gpt_base->part_base);
362 	int rc, num_partitions = 0;
363 
364 	spdk_bdev_free_io(bdev_io);
365 	spdk_put_io_channel(gpt_base->ch);
366 	gpt_base->ch = NULL;
367 
368 	if (status != SPDK_BDEV_IO_STATUS_SUCCESS) {
369 		SPDK_ERRLOG("Gpt: bdev=%s io error status=%d\n",
370 			    spdk_bdev_get_name(bdev), status);
371 		goto end;
372 	}
373 
374 	rc = gpt_parse_partition_table(&gpt_base->gpt);
375 	if (rc) {
376 		SPDK_DEBUGLOG(vbdev_gpt, "Failed to parse secondary partition table\n");
377 		goto end;
378 	}
379 
380 	SPDK_WARNLOG("Gpt: bdev=%s primary partition table broken, use the secondary\n",
381 		     spdk_bdev_get_name(bdev));
382 
383 	num_partitions = vbdev_gpt_create_bdevs(gpt_base);
384 	if (num_partitions < 0) {
385 		SPDK_DEBUGLOG(vbdev_gpt, "Failed to split dev=%s by gpt table\n",
386 			      spdk_bdev_get_name(bdev));
387 	}
388 
389 end:
390 	spdk_bdev_module_examine_done(&gpt_if);
391 	if (num_partitions <= 0) {
392 		/* If no gpt_disk instances were created, free the base context */
393 		spdk_bdev_part_base_free(gpt_base->part_base);
394 	}
395 }
396 
397 static int
398 vbdev_gpt_read_secondary_table(struct gpt_base *gpt_base)
399 {
400 	struct spdk_gpt *gpt;
401 	struct spdk_bdev_desc *part_base_desc;
402 	uint64_t secondary_offset;
403 
404 	gpt = &gpt_base->gpt;
405 	gpt->parse_phase = SPDK_GPT_PARSE_PHASE_SECONDARY;
406 	gpt->header = NULL;
407 	gpt->partitions = NULL;
408 
409 	part_base_desc = spdk_bdev_part_base_get_desc(gpt_base->part_base);
410 
411 	secondary_offset = gpt->total_sectors * gpt->sector_size - gpt->buf_size;
412 	return spdk_bdev_read(part_base_desc, gpt_base->ch, gpt_base->gpt.buf, secondary_offset,
413 			      gpt_base->gpt.buf_size, gpt_read_secondary_table_complete,
414 			      gpt_base);
415 }
416 
417 static void
418 gpt_bdev_complete(struct spdk_bdev_io *bdev_io, bool status, void *arg)
419 {
420 	struct gpt_base *gpt_base = (struct gpt_base *)arg;
421 	struct spdk_bdev *bdev = spdk_bdev_part_base_get_bdev(gpt_base->part_base);
422 	int rc, num_partitions = 0;
423 
424 	spdk_bdev_free_io(bdev_io);
425 
426 	if (status != SPDK_BDEV_IO_STATUS_SUCCESS) {
427 		SPDK_ERRLOG("Gpt: bdev=%s io error status=%d\n",
428 			    spdk_bdev_get_name(bdev), status);
429 		goto end;
430 	}
431 
432 	rc = gpt_parse_mbr(&gpt_base->gpt);
433 	if (rc) {
434 		SPDK_DEBUGLOG(vbdev_gpt, "Failed to parse mbr\n");
435 		goto end;
436 	}
437 
438 	rc = gpt_parse_partition_table(&gpt_base->gpt);
439 	if (rc) {
440 		SPDK_DEBUGLOG(vbdev_gpt, "Failed to parse primary partition table\n");
441 		rc = vbdev_gpt_read_secondary_table(gpt_base);
442 		if (rc) {
443 			SPDK_ERRLOG("Failed to read secondary table\n");
444 			goto end;
445 		}
446 		return;
447 	}
448 
449 	num_partitions = vbdev_gpt_create_bdevs(gpt_base);
450 	if (num_partitions < 0) {
451 		SPDK_DEBUGLOG(vbdev_gpt, "Failed to split dev=%s by gpt table\n",
452 			      spdk_bdev_get_name(bdev));
453 	}
454 
455 end:
456 	spdk_put_io_channel(gpt_base->ch);
457 	gpt_base->ch = NULL;
458 	/*
459 	 * Notify the generic bdev layer that the actions related to the original examine
460 	 *  callback are now completed.
461 	 */
462 	spdk_bdev_module_examine_done(&gpt_if);
463 
464 	/*
465 	 * vbdev_gpt_create_bdevs returns the number of bdevs created upon success.
466 	 * We can branch on this value.
467 	 */
468 	if (num_partitions <= 0) {
469 		/* If no gpt_disk instances were created, free the base context */
470 		spdk_bdev_part_base_free(gpt_base->part_base);
471 	}
472 }
473 
474 static int
475 vbdev_gpt_read_gpt(struct spdk_bdev *bdev)
476 {
477 	struct gpt_base *gpt_base;
478 	struct spdk_bdev_desc *part_base_desc;
479 	int rc;
480 
481 	gpt_base = gpt_base_bdev_init(bdev);
482 	if (!gpt_base) {
483 		SPDK_ERRLOG("Cannot allocated gpt_base\n");
484 		return -1;
485 	}
486 
487 	part_base_desc = spdk_bdev_part_base_get_desc(gpt_base->part_base);
488 	gpt_base->ch = spdk_bdev_get_io_channel(part_base_desc);
489 	if (gpt_base->ch == NULL) {
490 		SPDK_ERRLOG("Failed to get an io_channel.\n");
491 		spdk_bdev_part_base_free(gpt_base->part_base);
492 		return -1;
493 	}
494 
495 	rc = spdk_bdev_read(part_base_desc, gpt_base->ch, gpt_base->gpt.buf, 0,
496 			    gpt_base->gpt.buf_size, gpt_bdev_complete, gpt_base);
497 	if (rc < 0) {
498 		spdk_put_io_channel(gpt_base->ch);
499 		spdk_bdev_part_base_free(gpt_base->part_base);
500 		SPDK_ERRLOG("Failed to send bdev_io command\n");
501 		return -1;
502 	}
503 
504 	return 0;
505 }
506 
507 static int
508 vbdev_gpt_init(void)
509 {
510 	return 0;
511 }
512 
513 static int
514 vbdev_gpt_get_ctx_size(void)
515 {
516 	return sizeof(struct gpt_io);
517 }
518 
519 static void
520 vbdev_gpt_examine(struct spdk_bdev *bdev)
521 {
522 	int rc;
523 
524 	/* A bdev with fewer than 2 blocks cannot have a GPT. Block 0 has
525 	 * the MBR and block 1 has the GPT header.
526 	 */
527 	if (spdk_bdev_get_num_blocks(bdev) < 2) {
528 		spdk_bdev_module_examine_done(&gpt_if);
529 		return;
530 	}
531 
532 	if (spdk_bdev_get_block_size(bdev) % 512 != 0) {
533 		SPDK_DEBUGLOG(vbdev_gpt,
534 			      "GPT module does not support block size %" PRIu32 " for bdev %s\n",
535 			      spdk_bdev_get_block_size(bdev), spdk_bdev_get_name(bdev));
536 		spdk_bdev_module_examine_done(&gpt_if);
537 		return;
538 	}
539 
540 	rc = vbdev_gpt_read_gpt(bdev);
541 	if (rc) {
542 		spdk_bdev_module_examine_done(&gpt_if);
543 		SPDK_ERRLOG("Failed to read info from bdev %s\n", spdk_bdev_get_name(bdev));
544 	}
545 }
546 
547 SPDK_LOG_REGISTER_COMPONENT(vbdev_gpt)
548