xref: /illumos-gate/usr/src/uts/common/fs/zfs/zio.c (revision 3bfb48feb84bb78924286a801c68f80198912fa7)
1 /*
2  * CDDL HEADER START
3  *
4  * The contents of this file are subject to the terms of the
5  * Common Development and Distribution License (the "License").
6  * You may not use this file except in compliance with the License.
7  *
8  * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9  * or http://www.opensolaris.org/os/licensing.
10  * See the License for the specific language governing permissions
11  * and limitations under the License.
12  *
13  * When distributing Covered Code, include this CDDL HEADER in each
14  * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15  * If applicable, add the following below this CDDL HEADER, with the
16  * fields enclosed by brackets "[]" replaced with your own identifying
17  * information: Portions Copyright [yyyy] [name of copyright owner]
18  *
19  * CDDL HEADER END
20  */
21 /*
22  * Copyright 2006 Sun Microsystems, Inc.  All rights reserved.
23  * Use is subject to license terms.
24  */
25 
26 #pragma ident	"%Z%%M%	%I%	%E% SMI"
27 
28 #include <sys/zfs_context.h>
29 #include <sys/fm/fs/zfs.h>
30 #include <sys/spa.h>
31 #include <sys/txg.h>
32 #include <sys/spa_impl.h>
33 #include <sys/vdev_impl.h>
34 #include <sys/zio_impl.h>
35 #include <sys/zio_compress.h>
36 #include <sys/zio_checksum.h>
37 
38 /*
39  * ==========================================================================
40  * I/O priority table
41  * ==========================================================================
42  */
43 uint8_t zio_priority_table[ZIO_PRIORITY_TABLE_SIZE] = {
44 	0,	/* ZIO_PRIORITY_NOW		*/
45 	0,	/* ZIO_PRIORITY_SYNC_READ	*/
46 	0,	/* ZIO_PRIORITY_SYNC_WRITE	*/
47 	6,	/* ZIO_PRIORITY_ASYNC_READ	*/
48 	4,	/* ZIO_PRIORITY_ASYNC_WRITE	*/
49 	4,	/* ZIO_PRIORITY_FREE		*/
50 	0,	/* ZIO_PRIORITY_CACHE_FILL	*/
51 	0,	/* ZIO_PRIORITY_LOG_WRITE	*/
52 	10,	/* ZIO_PRIORITY_RESILVER	*/
53 	20,	/* ZIO_PRIORITY_SCRUB		*/
54 };
55 
56 /*
57  * ==========================================================================
58  * I/O type descriptions
59  * ==========================================================================
60  */
61 char *zio_type_name[ZIO_TYPES] = {
62 	"null", "read", "write", "free", "claim", "ioctl" };
63 
64 /* At or above this size, force gang blocking - for testing */
65 uint64_t zio_gang_bang = SPA_MAXBLOCKSIZE + 1;
66 
67 typedef struct zio_sync_pass {
68 	int	zp_defer_free;		/* defer frees after this pass */
69 	int	zp_dontcompress;	/* don't compress after this pass */
70 	int	zp_rewrite;		/* rewrite new bps after this pass */
71 } zio_sync_pass_t;
72 
73 zio_sync_pass_t zio_sync_pass = {
74 	1,	/* zp_defer_free */
75 	4,	/* zp_dontcompress */
76 	1,	/* zp_rewrite */
77 };
78 
79 /*
80  * ==========================================================================
81  * I/O kmem caches
82  * ==========================================================================
83  */
84 kmem_cache_t *zio_buf_cache[SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT];
85 kmem_cache_t *zio_data_buf_cache[SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT];
86 
87 #ifdef _KERNEL
88 extern vmem_t *zio_alloc_arena;
89 #endif
90 
91 void
92 zio_init(void)
93 {
94 	size_t c;
95 	vmem_t *data_alloc_arena = NULL;
96 
97 #ifdef _KERNEL
98 	data_alloc_arena = zio_alloc_arena;
99 #endif
100 
101 	/*
102 	 * For small buffers, we want a cache for each multiple of
103 	 * SPA_MINBLOCKSIZE.  For medium-size buffers, we want a cache
104 	 * for each quarter-power of 2.  For large buffers, we want
105 	 * a cache for each multiple of PAGESIZE.
106 	 */
107 	for (c = 0; c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; c++) {
108 		size_t size = (c + 1) << SPA_MINBLOCKSHIFT;
109 		size_t p2 = size;
110 		size_t align = 0;
111 
112 		while (p2 & (p2 - 1))
113 			p2 &= p2 - 1;
114 
115 		if (size <= 4 * SPA_MINBLOCKSIZE) {
116 			align = SPA_MINBLOCKSIZE;
117 		} else if (P2PHASE(size, PAGESIZE) == 0) {
118 			align = PAGESIZE;
119 		} else if (P2PHASE(size, p2 >> 2) == 0) {
120 			align = p2 >> 2;
121 		}
122 
123 		if (align != 0) {
124 			char name[36];
125 			(void) sprintf(name, "zio_buf_%lu", (ulong_t)size);
126 			zio_buf_cache[c] = kmem_cache_create(name, size,
127 			    align, NULL, NULL, NULL, NULL, NULL, KMC_NODEBUG);
128 
129 			(void) sprintf(name, "zio_data_buf_%lu", (ulong_t)size);
130 			zio_data_buf_cache[c] = kmem_cache_create(name, size,
131 			    align, NULL, NULL, NULL, NULL, data_alloc_arena,
132 			    KMC_NODEBUG);
133 
134 			dprintf("creating cache for size %5lx align %5lx\n",
135 			    size, align);
136 		}
137 	}
138 
139 	while (--c != 0) {
140 		ASSERT(zio_buf_cache[c] != NULL);
141 		if (zio_buf_cache[c - 1] == NULL)
142 			zio_buf_cache[c - 1] = zio_buf_cache[c];
143 
144 		ASSERT(zio_data_buf_cache[c] != NULL);
145 		if (zio_data_buf_cache[c - 1] == NULL)
146 			zio_data_buf_cache[c - 1] = zio_data_buf_cache[c];
147 	}
148 
149 	zio_inject_init();
150 }
151 
152 void
153 zio_fini(void)
154 {
155 	size_t c;
156 	kmem_cache_t *last_cache = NULL;
157 	kmem_cache_t *last_data_cache = NULL;
158 
159 	for (c = 0; c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; c++) {
160 		if (zio_buf_cache[c] != last_cache) {
161 			last_cache = zio_buf_cache[c];
162 			kmem_cache_destroy(zio_buf_cache[c]);
163 		}
164 		zio_buf_cache[c] = NULL;
165 
166 		if (zio_data_buf_cache[c] != last_data_cache) {
167 			last_data_cache = zio_data_buf_cache[c];
168 			kmem_cache_destroy(zio_data_buf_cache[c]);
169 		}
170 		zio_data_buf_cache[c] = NULL;
171 	}
172 
173 	zio_inject_fini();
174 }
175 
176 /*
177  * ==========================================================================
178  * Allocate and free I/O buffers
179  * ==========================================================================
180  */
181 
182 /*
183  * Use zio_buf_alloc to allocate ZFS metadata.  This data will appear in a
184  * crashdump if the kernel panics, so use it judiciously.  Obviously, it's
185  * useful to inspect ZFS metadata, but if possible, we should avoid keeping
186  * excess / transient data in-core during a crashdump.
187  */
188 void *
189 zio_buf_alloc(size_t size)
190 {
191 	size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
192 
193 	ASSERT(c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
194 
195 	return (kmem_cache_alloc(zio_buf_cache[c], KM_SLEEP));
196 }
197 
198 /*
199  * Use zio_data_buf_alloc to allocate data.  The data will not appear in a
200  * crashdump if the kernel panics.  This exists so that we will limit the amount
201  * of ZFS data that shows up in a kernel crashdump.  (Thus reducing the amount
202  * of kernel heap dumped to disk when the kernel panics)
203  */
204 void *
205 zio_data_buf_alloc(size_t size)
206 {
207 	size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
208 
209 	ASSERT(c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
210 
211 	return (kmem_cache_alloc(zio_data_buf_cache[c], KM_SLEEP));
212 }
213 
214 void
215 zio_buf_free(void *buf, size_t size)
216 {
217 	size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
218 
219 	ASSERT(c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
220 
221 	kmem_cache_free(zio_buf_cache[c], buf);
222 }
223 
224 void
225 zio_data_buf_free(void *buf, size_t size)
226 {
227 	size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
228 
229 	ASSERT(c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
230 
231 	kmem_cache_free(zio_data_buf_cache[c], buf);
232 }
233 /*
234  * ==========================================================================
235  * Push and pop I/O transform buffers
236  * ==========================================================================
237  */
238 static void
239 zio_push_transform(zio_t *zio, void *data, uint64_t size, uint64_t bufsize)
240 {
241 	zio_transform_t *zt = kmem_alloc(sizeof (zio_transform_t), KM_SLEEP);
242 
243 	zt->zt_data = data;
244 	zt->zt_size = size;
245 	zt->zt_bufsize = bufsize;
246 
247 	zt->zt_next = zio->io_transform_stack;
248 	zio->io_transform_stack = zt;
249 
250 	zio->io_data = data;
251 	zio->io_size = size;
252 }
253 
254 static void
255 zio_pop_transform(zio_t *zio, void **data, uint64_t *size, uint64_t *bufsize)
256 {
257 	zio_transform_t *zt = zio->io_transform_stack;
258 
259 	*data = zt->zt_data;
260 	*size = zt->zt_size;
261 	*bufsize = zt->zt_bufsize;
262 
263 	zio->io_transform_stack = zt->zt_next;
264 	kmem_free(zt, sizeof (zio_transform_t));
265 
266 	if ((zt = zio->io_transform_stack) != NULL) {
267 		zio->io_data = zt->zt_data;
268 		zio->io_size = zt->zt_size;
269 	}
270 }
271 
272 static void
273 zio_clear_transform_stack(zio_t *zio)
274 {
275 	void *data;
276 	uint64_t size, bufsize;
277 
278 	ASSERT(zio->io_transform_stack != NULL);
279 
280 	zio_pop_transform(zio, &data, &size, &bufsize);
281 	while (zio->io_transform_stack != NULL) {
282 		zio_buf_free(data, bufsize);
283 		zio_pop_transform(zio, &data, &size, &bufsize);
284 	}
285 }
286 
287 /*
288  * ==========================================================================
289  * Create the various types of I/O (read, write, free)
290  * ==========================================================================
291  */
292 static zio_t *
293 zio_create(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp,
294     void *data, uint64_t size, zio_done_func_t *done, void *private,
295     zio_type_t type, int priority, int flags, uint8_t stage, uint32_t pipeline)
296 {
297 	zio_t *zio;
298 
299 	ASSERT3U(size, <=, SPA_MAXBLOCKSIZE);
300 	ASSERT(P2PHASE(size, SPA_MINBLOCKSIZE) == 0);
301 
302 	zio = kmem_zalloc(sizeof (zio_t), KM_SLEEP);
303 	zio->io_parent = pio;
304 	zio->io_spa = spa;
305 	zio->io_txg = txg;
306 	if (bp != NULL) {
307 		zio->io_bp = bp;
308 		zio->io_bp_copy = *bp;
309 		zio->io_bp_orig = *bp;
310 	}
311 	zio->io_done = done;
312 	zio->io_private = private;
313 	zio->io_type = type;
314 	zio->io_priority = priority;
315 	zio->io_stage = stage;
316 	zio->io_pipeline = pipeline;
317 	zio->io_async_stages = ZIO_ASYNC_PIPELINE_STAGES;
318 	zio->io_timestamp = lbolt64;
319 	zio->io_flags = flags;
320 	mutex_init(&zio->io_lock, NULL, MUTEX_DEFAULT, NULL);
321 	zio_push_transform(zio, data, size, size);
322 
323 	if (pio == NULL) {
324 		if (!(flags & ZIO_FLAG_CONFIG_HELD))
325 			spa_config_enter(zio->io_spa, RW_READER, zio);
326 		zio->io_root = zio;
327 	} else {
328 		zio->io_root = pio->io_root;
329 		if (!(flags & ZIO_FLAG_NOBOOKMARK))
330 			zio->io_logical = pio->io_logical;
331 		mutex_enter(&pio->io_lock);
332 		if (stage < ZIO_STAGE_READY)
333 			pio->io_children_notready++;
334 		pio->io_children_notdone++;
335 		zio->io_sibling_next = pio->io_child;
336 		zio->io_sibling_prev = NULL;
337 		if (pio->io_child != NULL)
338 			pio->io_child->io_sibling_prev = zio;
339 		pio->io_child = zio;
340 		zio->io_ndvas = pio->io_ndvas;
341 		mutex_exit(&pio->io_lock);
342 	}
343 
344 	return (zio);
345 }
346 
347 zio_t *
348 zio_null(zio_t *pio, spa_t *spa, zio_done_func_t *done, void *private,
349 	int flags)
350 {
351 	zio_t *zio;
352 
353 	zio = zio_create(pio, spa, 0, NULL, NULL, 0, done, private,
354 	    ZIO_TYPE_NULL, ZIO_PRIORITY_NOW, flags, ZIO_STAGE_OPEN,
355 	    ZIO_WAIT_FOR_CHILDREN_PIPELINE);
356 
357 	return (zio);
358 }
359 
360 zio_t *
361 zio_root(spa_t *spa, zio_done_func_t *done, void *private, int flags)
362 {
363 	return (zio_null(NULL, spa, done, private, flags));
364 }
365 
366 zio_t *
367 zio_read(zio_t *pio, spa_t *spa, blkptr_t *bp, void *data,
368     uint64_t size, zio_done_func_t *done, void *private,
369     int priority, int flags, zbookmark_t *zb)
370 {
371 	zio_t *zio;
372 
373 	ASSERT3U(size, ==, BP_GET_LSIZE(bp));
374 
375 	zio = zio_create(pio, spa, bp->blk_birth, bp, data, size, done, private,
376 	    ZIO_TYPE_READ, priority, flags | ZIO_FLAG_USER,
377 	    ZIO_STAGE_OPEN, ZIO_READ_PIPELINE);
378 	zio->io_bookmark = *zb;
379 
380 	zio->io_logical = zio;
381 
382 	/*
383 	 * Work off our copy of the bp so the caller can free it.
384 	 */
385 	zio->io_bp = &zio->io_bp_copy;
386 
387 	if (BP_GET_COMPRESS(bp) != ZIO_COMPRESS_OFF) {
388 		uint64_t csize = BP_GET_PSIZE(bp);
389 		void *cbuf = zio_buf_alloc(csize);
390 
391 		zio_push_transform(zio, cbuf, csize, csize);
392 		zio->io_pipeline |= 1U << ZIO_STAGE_READ_DECOMPRESS;
393 	}
394 
395 	if (BP_IS_GANG(bp)) {
396 		uint64_t gsize = SPA_GANGBLOCKSIZE;
397 		void *gbuf = zio_buf_alloc(gsize);
398 
399 		zio_push_transform(zio, gbuf, gsize, gsize);
400 		zio->io_pipeline |= 1U << ZIO_STAGE_READ_GANG_MEMBERS;
401 	}
402 
403 	return (zio);
404 }
405 
406 zio_t *
407 zio_write(zio_t *pio, spa_t *spa, int checksum, int compress, int ncopies,
408     uint64_t txg, blkptr_t *bp, void *data, uint64_t size,
409     zio_done_func_t *done, void *private, int priority, int flags,
410     zbookmark_t *zb)
411 {
412 	zio_t *zio;
413 
414 	ASSERT(checksum >= ZIO_CHECKSUM_OFF &&
415 	    checksum < ZIO_CHECKSUM_FUNCTIONS);
416 
417 	ASSERT(compress >= ZIO_COMPRESS_OFF &&
418 	    compress < ZIO_COMPRESS_FUNCTIONS);
419 
420 	zio = zio_create(pio, spa, txg, bp, data, size, done, private,
421 	    ZIO_TYPE_WRITE, priority, flags | ZIO_FLAG_USER,
422 	    ZIO_STAGE_OPEN, ZIO_WRITE_PIPELINE);
423 
424 	zio->io_bookmark = *zb;
425 
426 	zio->io_logical = zio;
427 
428 	zio->io_checksum = checksum;
429 	zio->io_compress = compress;
430 	zio->io_ndvas = ncopies;
431 
432 	if (compress != ZIO_COMPRESS_OFF)
433 		zio->io_async_stages |= 1U << ZIO_STAGE_WRITE_COMPRESS;
434 
435 	if (bp->blk_birth != txg) {
436 		/* XXX the bp usually (always?) gets re-zeroed later */
437 		BP_ZERO(bp);
438 		BP_SET_LSIZE(bp, size);
439 		BP_SET_PSIZE(bp, size);
440 	} else {
441 		/* Make sure someone doesn't change their mind on overwrites */
442 		ASSERT(MIN(zio->io_ndvas + BP_IS_GANG(bp),
443 		    spa_max_replication(spa)) == BP_GET_NDVAS(bp));
444 	}
445 
446 	return (zio);
447 }
448 
449 zio_t *
450 zio_rewrite(zio_t *pio, spa_t *spa, int checksum,
451     uint64_t txg, blkptr_t *bp, void *data, uint64_t size,
452     zio_done_func_t *done, void *private, int priority, int flags,
453     zbookmark_t *zb)
454 {
455 	zio_t *zio;
456 
457 	zio = zio_create(pio, spa, txg, bp, data, size, done, private,
458 	    ZIO_TYPE_WRITE, priority, flags | ZIO_FLAG_USER,
459 	    ZIO_STAGE_OPEN, ZIO_REWRITE_PIPELINE);
460 
461 	zio->io_bookmark = *zb;
462 	zio->io_checksum = checksum;
463 	zio->io_compress = ZIO_COMPRESS_OFF;
464 
465 	if (pio != NULL)
466 		ASSERT3U(zio->io_ndvas, <=, BP_GET_NDVAS(bp));
467 
468 	return (zio);
469 }
470 
471 static zio_t *
472 zio_write_allocate(zio_t *pio, spa_t *spa, int checksum,
473     uint64_t txg, blkptr_t *bp, void *data, uint64_t size,
474     zio_done_func_t *done, void *private, int priority, int flags)
475 {
476 	zio_t *zio;
477 
478 	BP_ZERO(bp);
479 	BP_SET_LSIZE(bp, size);
480 	BP_SET_PSIZE(bp, size);
481 	BP_SET_COMPRESS(bp, ZIO_COMPRESS_OFF);
482 
483 	zio = zio_create(pio, spa, txg, bp, data, size, done, private,
484 	    ZIO_TYPE_WRITE, priority, flags,
485 	    ZIO_STAGE_OPEN, ZIO_WRITE_ALLOCATE_PIPELINE);
486 
487 	zio->io_checksum = checksum;
488 	zio->io_compress = ZIO_COMPRESS_OFF;
489 
490 	return (zio);
491 }
492 
493 zio_t *
494 zio_free(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp,
495     zio_done_func_t *done, void *private)
496 {
497 	zio_t *zio;
498 
499 	ASSERT(!BP_IS_HOLE(bp));
500 
501 	if (txg == spa->spa_syncing_txg &&
502 	    spa->spa_sync_pass > zio_sync_pass.zp_defer_free) {
503 		bplist_enqueue_deferred(&spa->spa_sync_bplist, bp);
504 		return (zio_null(pio, spa, NULL, NULL, 0));
505 	}
506 
507 	zio = zio_create(pio, spa, txg, bp, NULL, 0, done, private,
508 	    ZIO_TYPE_FREE, ZIO_PRIORITY_FREE, ZIO_FLAG_USER,
509 	    ZIO_STAGE_OPEN, ZIO_FREE_PIPELINE);
510 
511 	zio->io_bp = &zio->io_bp_copy;
512 
513 	return (zio);
514 }
515 
516 zio_t *
517 zio_claim(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp,
518     zio_done_func_t *done, void *private)
519 {
520 	zio_t *zio;
521 
522 	/*
523 	 * A claim is an allocation of a specific block.  Claims are needed
524 	 * to support immediate writes in the intent log.  The issue is that
525 	 * immediate writes contain committed data, but in a txg that was
526 	 * *not* committed.  Upon opening the pool after an unclean shutdown,
527 	 * the intent log claims all blocks that contain immediate write data
528 	 * so that the SPA knows they're in use.
529 	 *
530 	 * All claims *must* be resolved in the first txg -- before the SPA
531 	 * starts allocating blocks -- so that nothing is allocated twice.
532 	 */
533 	ASSERT3U(spa->spa_uberblock.ub_rootbp.blk_birth, <, spa_first_txg(spa));
534 	ASSERT3U(spa_first_txg(spa), <=, txg);
535 
536 	zio = zio_create(pio, spa, txg, bp, NULL, 0, done, private,
537 	    ZIO_TYPE_CLAIM, ZIO_PRIORITY_NOW, 0,
538 	    ZIO_STAGE_OPEN, ZIO_CLAIM_PIPELINE);
539 
540 	zio->io_bp = &zio->io_bp_copy;
541 
542 	return (zio);
543 }
544 
545 zio_t *
546 zio_ioctl(zio_t *pio, spa_t *spa, vdev_t *vd, int cmd,
547     zio_done_func_t *done, void *private, int priority, int flags)
548 {
549 	zio_t *zio;
550 	int c;
551 
552 	if (vd->vdev_children == 0) {
553 		zio = zio_create(pio, spa, 0, NULL, NULL, 0, done, private,
554 		    ZIO_TYPE_IOCTL, priority, flags,
555 		    ZIO_STAGE_OPEN, ZIO_IOCTL_PIPELINE);
556 
557 		zio->io_vd = vd;
558 		zio->io_cmd = cmd;
559 	} else {
560 		zio = zio_null(pio, spa, NULL, NULL, flags);
561 
562 		for (c = 0; c < vd->vdev_children; c++)
563 			zio_nowait(zio_ioctl(zio, spa, vd->vdev_child[c], cmd,
564 			    done, private, priority, flags));
565 	}
566 
567 	return (zio);
568 }
569 
570 static void
571 zio_phys_bp_init(vdev_t *vd, blkptr_t *bp, uint64_t offset, uint64_t size,
572     int checksum)
573 {
574 	ASSERT(vd->vdev_children == 0);
575 
576 	ASSERT(size <= SPA_MAXBLOCKSIZE);
577 	ASSERT(P2PHASE(size, SPA_MINBLOCKSIZE) == 0);
578 	ASSERT(P2PHASE(offset, SPA_MINBLOCKSIZE) == 0);
579 
580 	ASSERT(offset + size <= VDEV_LABEL_START_SIZE ||
581 	    offset >= vd->vdev_psize - VDEV_LABEL_END_SIZE);
582 	ASSERT3U(offset + size, <=, vd->vdev_psize);
583 
584 	BP_ZERO(bp);
585 
586 	BP_SET_LSIZE(bp, size);
587 	BP_SET_PSIZE(bp, size);
588 
589 	BP_SET_CHECKSUM(bp, checksum);
590 	BP_SET_COMPRESS(bp, ZIO_COMPRESS_OFF);
591 	BP_SET_BYTEORDER(bp, ZFS_HOST_BYTEORDER);
592 
593 	if (checksum != ZIO_CHECKSUM_OFF)
594 		ZIO_SET_CHECKSUM(&bp->blk_cksum, offset, 0, 0, 0);
595 }
596 
597 zio_t *
598 zio_read_phys(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size,
599     void *data, int checksum, zio_done_func_t *done, void *private,
600     int priority, int flags)
601 {
602 	zio_t *zio;
603 	blkptr_t blk;
604 
605 	zio_phys_bp_init(vd, &blk, offset, size, checksum);
606 
607 	zio = zio_create(pio, vd->vdev_spa, 0, &blk, data, size, done, private,
608 	    ZIO_TYPE_READ, priority, flags | ZIO_FLAG_PHYSICAL,
609 	    ZIO_STAGE_OPEN, ZIO_READ_PHYS_PIPELINE);
610 
611 	zio->io_vd = vd;
612 	zio->io_offset = offset;
613 
614 	/*
615 	 * Work off our copy of the bp so the caller can free it.
616 	 */
617 	zio->io_bp = &zio->io_bp_copy;
618 
619 	return (zio);
620 }
621 
622 zio_t *
623 zio_write_phys(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size,
624     void *data, int checksum, zio_done_func_t *done, void *private,
625     int priority, int flags)
626 {
627 	zio_block_tail_t *zbt;
628 	void *wbuf;
629 	zio_t *zio;
630 	blkptr_t blk;
631 
632 	zio_phys_bp_init(vd, &blk, offset, size, checksum);
633 
634 	zio = zio_create(pio, vd->vdev_spa, 0, &blk, data, size, done, private,
635 	    ZIO_TYPE_WRITE, priority, flags | ZIO_FLAG_PHYSICAL,
636 	    ZIO_STAGE_OPEN, ZIO_WRITE_PHYS_PIPELINE);
637 
638 	zio->io_vd = vd;
639 	zio->io_offset = offset;
640 
641 	zio->io_bp = &zio->io_bp_copy;
642 	zio->io_checksum = checksum;
643 
644 	if (zio_checksum_table[checksum].ci_zbt) {
645 		/*
646 		 * zbt checksums are necessarily destructive -- they modify
647 		 * one word of the write buffer to hold the verifier/checksum.
648 		 * Therefore, we must make a local copy in case the data is
649 		 * being written to multiple places.
650 		 */
651 		wbuf = zio_buf_alloc(size);
652 		bcopy(data, wbuf, size);
653 		zio_push_transform(zio, wbuf, size, size);
654 
655 		zbt = (zio_block_tail_t *)((char *)wbuf + size) - 1;
656 		zbt->zbt_cksum = blk.blk_cksum;
657 	}
658 
659 	return (zio);
660 }
661 
662 /*
663  * Create a child I/O to do some work for us.  It has no associated bp.
664  */
665 zio_t *
666 zio_vdev_child_io(zio_t *zio, blkptr_t *bp, vdev_t *vd, uint64_t offset,
667 	void *data, uint64_t size, int type, int priority, int flags,
668 	zio_done_func_t *done, void *private)
669 {
670 	uint32_t pipeline = ZIO_VDEV_CHILD_PIPELINE;
671 	zio_t *cio;
672 
673 	if (type == ZIO_TYPE_READ && bp != NULL) {
674 		/*
675 		 * If we have the bp, then the child should perform the
676 		 * checksum and the parent need not.  This pushes error
677 		 * detection as close to the leaves as possible and
678 		 * eliminates redundant checksums in the interior nodes.
679 		 */
680 		pipeline |= 1U << ZIO_STAGE_CHECKSUM_VERIFY;
681 		zio->io_pipeline &= ~(1U << ZIO_STAGE_CHECKSUM_VERIFY);
682 	}
683 
684 	cio = zio_create(zio, zio->io_spa, zio->io_txg, bp, data, size,
685 	    done, private, type, priority,
686 	    (zio->io_flags & ZIO_FLAG_VDEV_INHERIT) | ZIO_FLAG_CANFAIL | flags,
687 	    ZIO_STAGE_VDEV_IO_START - 1, pipeline);
688 
689 	cio->io_vd = vd;
690 	cio->io_offset = offset;
691 
692 	return (cio);
693 }
694 
695 /*
696  * ==========================================================================
697  * Initiate I/O, either sync or async
698  * ==========================================================================
699  */
700 int
701 zio_wait(zio_t *zio)
702 {
703 	int error;
704 
705 	ASSERT(zio->io_stage == ZIO_STAGE_OPEN);
706 
707 	zio->io_waiter = curthread;
708 
709 	zio_next_stage_async(zio);
710 
711 	mutex_enter(&zio->io_lock);
712 	while (zio->io_stalled != ZIO_STAGE_DONE)
713 		cv_wait(&zio->io_cv, &zio->io_lock);
714 	mutex_exit(&zio->io_lock);
715 
716 	error = zio->io_error;
717 	mutex_destroy(&zio->io_lock);
718 	kmem_free(zio, sizeof (zio_t));
719 
720 	return (error);
721 }
722 
723 void
724 zio_nowait(zio_t *zio)
725 {
726 	zio_next_stage_async(zio);
727 }
728 
729 /*
730  * ==========================================================================
731  * I/O pipeline interlocks: parent/child dependency scoreboarding
732  * ==========================================================================
733  */
734 static void
735 zio_wait_for_children(zio_t *zio, uint32_t stage, uint64_t *countp)
736 {
737 	mutex_enter(&zio->io_lock);
738 	if (*countp == 0) {
739 		ASSERT(zio->io_stalled == 0);
740 		mutex_exit(&zio->io_lock);
741 		zio_next_stage(zio);
742 	} else {
743 		zio->io_stalled = stage;
744 		mutex_exit(&zio->io_lock);
745 	}
746 }
747 
748 static void
749 zio_notify_parent(zio_t *zio, uint32_t stage, uint64_t *countp)
750 {
751 	zio_t *pio = zio->io_parent;
752 
753 	mutex_enter(&pio->io_lock);
754 	if (pio->io_error == 0 && !(zio->io_flags & ZIO_FLAG_DONT_PROPAGATE))
755 		pio->io_error = zio->io_error;
756 	if (--*countp == 0 && pio->io_stalled == stage) {
757 		pio->io_stalled = 0;
758 		mutex_exit(&pio->io_lock);
759 		zio_next_stage_async(pio);
760 	} else {
761 		mutex_exit(&pio->io_lock);
762 	}
763 }
764 
765 static void
766 zio_wait_children_ready(zio_t *zio)
767 {
768 	zio_wait_for_children(zio, ZIO_STAGE_WAIT_CHILDREN_READY,
769 	    &zio->io_children_notready);
770 }
771 
772 void
773 zio_wait_children_done(zio_t *zio)
774 {
775 	zio_wait_for_children(zio, ZIO_STAGE_WAIT_CHILDREN_DONE,
776 	    &zio->io_children_notdone);
777 }
778 
779 static void
780 zio_ready(zio_t *zio)
781 {
782 	zio_t *pio = zio->io_parent;
783 
784 	if (pio != NULL)
785 		zio_notify_parent(zio, ZIO_STAGE_WAIT_CHILDREN_READY,
786 		    &pio->io_children_notready);
787 
788 	if (zio->io_bp)
789 		zio->io_bp_copy = *zio->io_bp;
790 
791 	zio_next_stage(zio);
792 }
793 
794 static void
795 zio_done(zio_t *zio)
796 {
797 	zio_t *pio = zio->io_parent;
798 	spa_t *spa = zio->io_spa;
799 	blkptr_t *bp = zio->io_bp;
800 	vdev_t *vd = zio->io_vd;
801 	char blkbuf[BP_SPRINTF_LEN];
802 
803 	ASSERT(zio->io_children_notready == 0);
804 	ASSERT(zio->io_children_notdone == 0);
805 
806 	if (bp != NULL) {
807 		ASSERT(bp->blk_pad[0] == 0);
808 		ASSERT(bp->blk_pad[1] == 0);
809 		ASSERT(bp->blk_pad[2] == 0);
810 		ASSERT(bcmp(bp, &zio->io_bp_copy, sizeof (blkptr_t)) == 0);
811 		if (zio->io_type == ZIO_TYPE_WRITE && !BP_IS_HOLE(bp) &&
812 		    !(zio->io_flags & ZIO_FLAG_IO_REPAIR)) {
813 			ASSERT(!BP_SHOULD_BYTESWAP(bp));
814 			if (zio->io_ndvas != 0)
815 				ASSERT3U(zio->io_ndvas, <=, BP_GET_NDVAS(bp));
816 			ASSERT(BP_COUNT_GANG(bp) == 0 ||
817 			    (BP_COUNT_GANG(bp) == BP_GET_NDVAS(bp)));
818 		}
819 	}
820 
821 	if (vd != NULL)
822 		vdev_stat_update(zio);
823 
824 	if (zio->io_error) {
825 		/*
826 		 * If this I/O is attached to a particular vdev,
827 		 * generate an error message describing the I/O failure
828 		 * at the block level.  We ignore these errors if the
829 		 * device is currently unavailable.
830 		 */
831 		if (zio->io_error != ECKSUM && vd != NULL && !vdev_is_dead(vd))
832 			zfs_ereport_post(FM_EREPORT_ZFS_IO,
833 			    zio->io_spa, vd, zio, 0, 0);
834 
835 		if ((zio->io_error == EIO ||
836 		    !(zio->io_flags & ZIO_FLAG_SPECULATIVE)) &&
837 		    zio->io_logical == zio) {
838 			/*
839 			 * For root I/O requests, tell the SPA to log the error
840 			 * appropriately.  Also, generate a logical data
841 			 * ereport.
842 			 */
843 			spa_log_error(zio->io_spa, zio);
844 
845 			zfs_ereport_post(FM_EREPORT_ZFS_DATA,
846 			    zio->io_spa, NULL, zio, 0, 0);
847 		}
848 
849 		/*
850 		 * For I/O requests that cannot fail, panic appropriately.
851 		 */
852 		if (!(zio->io_flags & ZIO_FLAG_CANFAIL)) {
853 			sprintf_blkptr(blkbuf, BP_SPRINTF_LEN,
854 			    bp ? bp : &zio->io_bp_copy);
855 			panic("ZFS: %s (%s on %s off %llx: zio %p %s): error "
856 			    "%d", zio->io_error == ECKSUM ?
857 			    "bad checksum" : "I/O failure",
858 			    zio_type_name[zio->io_type],
859 			    vdev_description(vd),
860 			    (u_longlong_t)zio->io_offset,
861 			    zio, blkbuf, zio->io_error);
862 		}
863 	}
864 
865 	zio_clear_transform_stack(zio);
866 
867 	if (zio->io_done)
868 		zio->io_done(zio);
869 
870 	ASSERT(zio->io_delegate_list == NULL);
871 	ASSERT(zio->io_delegate_next == NULL);
872 
873 	if (pio != NULL) {
874 		zio_t *next, *prev;
875 
876 		mutex_enter(&pio->io_lock);
877 		next = zio->io_sibling_next;
878 		prev = zio->io_sibling_prev;
879 		if (next != NULL)
880 			next->io_sibling_prev = prev;
881 		if (prev != NULL)
882 			prev->io_sibling_next = next;
883 		if (pio->io_child == zio)
884 			pio->io_child = next;
885 		mutex_exit(&pio->io_lock);
886 
887 		zio_notify_parent(zio, ZIO_STAGE_WAIT_CHILDREN_DONE,
888 		    &pio->io_children_notdone);
889 	}
890 
891 	if (pio == NULL && !(zio->io_flags & ZIO_FLAG_CONFIG_HELD))
892 		spa_config_exit(spa, zio);
893 
894 	if (zio->io_waiter != NULL) {
895 		mutex_enter(&zio->io_lock);
896 		ASSERT(zio->io_stage == ZIO_STAGE_DONE);
897 		zio->io_stalled = zio->io_stage;
898 		cv_broadcast(&zio->io_cv);
899 		mutex_exit(&zio->io_lock);
900 	} else {
901 		kmem_free(zio, sizeof (zio_t));
902 	}
903 }
904 
905 /*
906  * ==========================================================================
907  * Compression support
908  * ==========================================================================
909  */
910 static void
911 zio_write_compress(zio_t *zio)
912 {
913 	int compress = zio->io_compress;
914 	blkptr_t *bp = zio->io_bp;
915 	void *cbuf;
916 	uint64_t lsize = zio->io_size;
917 	uint64_t csize = lsize;
918 	uint64_t cbufsize = 0;
919 	int pass;
920 
921 	if (bp->blk_birth == zio->io_txg) {
922 		/*
923 		 * We're rewriting an existing block, which means we're
924 		 * working on behalf of spa_sync().  For spa_sync() to
925 		 * converge, it must eventually be the case that we don't
926 		 * have to allocate new blocks.  But compression changes
927 		 * the blocksize, which forces a reallocate, and makes
928 		 * convergence take longer.  Therefore, after the first
929 		 * few passes, stop compressing to ensure convergence.
930 		 */
931 		pass = spa_sync_pass(zio->io_spa);
932 		if (pass > zio_sync_pass.zp_dontcompress)
933 			compress = ZIO_COMPRESS_OFF;
934 	} else {
935 		ASSERT(BP_IS_HOLE(bp));
936 		pass = 1;
937 	}
938 
939 	if (compress != ZIO_COMPRESS_OFF)
940 		if (!zio_compress_data(compress, zio->io_data, zio->io_size,
941 		    &cbuf, &csize, &cbufsize))
942 			compress = ZIO_COMPRESS_OFF;
943 
944 	if (compress != ZIO_COMPRESS_OFF && csize != 0)
945 		zio_push_transform(zio, cbuf, csize, cbufsize);
946 
947 	/*
948 	 * The final pass of spa_sync() must be all rewrites, but the first
949 	 * few passes offer a trade-off: allocating blocks defers convergence,
950 	 * but newly allocated blocks are sequential, so they can be written
951 	 * to disk faster.  Therefore, we allow the first few passes of
952 	 * spa_sync() to reallocate new blocks, but force rewrites after that.
953 	 * There should only be a handful of blocks after pass 1 in any case.
954 	 */
955 	if (bp->blk_birth == zio->io_txg && BP_GET_PSIZE(bp) == csize &&
956 	    pass > zio_sync_pass.zp_rewrite) {
957 		ASSERT(csize != 0);
958 		BP_SET_LSIZE(bp, lsize);
959 		BP_SET_COMPRESS(bp, compress);
960 		zio->io_pipeline = ZIO_REWRITE_PIPELINE;
961 	} else {
962 		if (bp->blk_birth == zio->io_txg) {
963 			ASSERT3U(BP_GET_LSIZE(bp), ==, lsize);
964 			bzero(bp, sizeof (blkptr_t));
965 		}
966 		if (csize == 0) {
967 			BP_ZERO(bp);
968 			zio->io_pipeline = ZIO_WAIT_FOR_CHILDREN_PIPELINE;
969 		} else {
970 			ASSERT3U(BP_GET_NDVAS(bp), ==, 0);
971 			BP_SET_LSIZE(bp, lsize);
972 			BP_SET_PSIZE(bp, csize);
973 			BP_SET_COMPRESS(bp, compress);
974 			zio->io_pipeline = ZIO_WRITE_ALLOCATE_PIPELINE;
975 		}
976 	}
977 
978 	zio_next_stage(zio);
979 }
980 
981 static void
982 zio_read_decompress(zio_t *zio)
983 {
984 	blkptr_t *bp = zio->io_bp;
985 	void *data;
986 	uint64_t size;
987 	uint64_t bufsize;
988 	int compress = BP_GET_COMPRESS(bp);
989 
990 	ASSERT(compress != ZIO_COMPRESS_OFF);
991 
992 	zio_pop_transform(zio, &data, &size, &bufsize);
993 
994 	if (zio_decompress_data(compress, data, size,
995 	    zio->io_data, zio->io_size))
996 		zio->io_error = EIO;
997 
998 	zio_buf_free(data, bufsize);
999 
1000 	zio_next_stage(zio);
1001 }
1002 
1003 /*
1004  * ==========================================================================
1005  * Gang block support
1006  * ==========================================================================
1007  */
1008 static void
1009 zio_gang_pipeline(zio_t *zio)
1010 {
1011 	/*
1012 	 * By default, the pipeline assumes that we're dealing with a gang
1013 	 * block.  If we're not, strip out any gang-specific stages.
1014 	 */
1015 	if (!BP_IS_GANG(zio->io_bp))
1016 		zio->io_pipeline &= ~ZIO_GANG_STAGES;
1017 
1018 	zio_next_stage(zio);
1019 }
1020 
1021 static void
1022 zio_gang_byteswap(zio_t *zio)
1023 {
1024 	ASSERT(zio->io_size == SPA_GANGBLOCKSIZE);
1025 
1026 	if (BP_SHOULD_BYTESWAP(zio->io_bp))
1027 		byteswap_uint64_array(zio->io_data, zio->io_size);
1028 }
1029 
1030 static void
1031 zio_get_gang_header(zio_t *zio)
1032 {
1033 	blkptr_t *bp = zio->io_bp;
1034 	uint64_t gsize = SPA_GANGBLOCKSIZE;
1035 	void *gbuf = zio_buf_alloc(gsize);
1036 
1037 	ASSERT(BP_IS_GANG(bp));
1038 
1039 	zio_push_transform(zio, gbuf, gsize, gsize);
1040 
1041 	zio_nowait(zio_create(zio, zio->io_spa, bp->blk_birth, bp, gbuf, gsize,
1042 	    NULL, NULL, ZIO_TYPE_READ, zio->io_priority,
1043 	    zio->io_flags & ZIO_FLAG_GANG_INHERIT,
1044 	    ZIO_STAGE_OPEN, ZIO_READ_PIPELINE));
1045 
1046 	zio_wait_children_done(zio);
1047 }
1048 
1049 static void
1050 zio_read_gang_members(zio_t *zio)
1051 {
1052 	zio_gbh_phys_t *gbh;
1053 	uint64_t gsize, gbufsize, loff, lsize;
1054 	int i;
1055 
1056 	ASSERT(BP_IS_GANG(zio->io_bp));
1057 
1058 	zio_gang_byteswap(zio);
1059 	zio_pop_transform(zio, (void **)&gbh, &gsize, &gbufsize);
1060 
1061 	for (loff = 0, i = 0; loff != zio->io_size; loff += lsize, i++) {
1062 		blkptr_t *gbp = &gbh->zg_blkptr[i];
1063 		lsize = BP_GET_PSIZE(gbp);
1064 
1065 		ASSERT(BP_GET_COMPRESS(gbp) == ZIO_COMPRESS_OFF);
1066 		ASSERT3U(lsize, ==, BP_GET_LSIZE(gbp));
1067 		ASSERT3U(loff + lsize, <=, zio->io_size);
1068 		ASSERT(i < SPA_GBH_NBLKPTRS);
1069 		ASSERT(!BP_IS_HOLE(gbp));
1070 
1071 		zio_nowait(zio_read(zio, zio->io_spa, gbp,
1072 		    (char *)zio->io_data + loff, lsize, NULL, NULL,
1073 		    zio->io_priority, zio->io_flags & ZIO_FLAG_GANG_INHERIT,
1074 		    &zio->io_bookmark));
1075 	}
1076 
1077 	zio_buf_free(gbh, gbufsize);
1078 	zio_wait_children_done(zio);
1079 }
1080 
1081 static void
1082 zio_rewrite_gang_members(zio_t *zio)
1083 {
1084 	zio_gbh_phys_t *gbh;
1085 	uint64_t gsize, gbufsize, loff, lsize;
1086 	int i;
1087 
1088 	ASSERT(BP_IS_GANG(zio->io_bp));
1089 	ASSERT3U(zio->io_size, ==, SPA_GANGBLOCKSIZE);
1090 
1091 	zio_gang_byteswap(zio);
1092 	zio_pop_transform(zio, (void **)&gbh, &gsize, &gbufsize);
1093 
1094 	ASSERT(gsize == gbufsize);
1095 
1096 	for (loff = 0, i = 0; loff != zio->io_size; loff += lsize, i++) {
1097 		blkptr_t *gbp = &gbh->zg_blkptr[i];
1098 		lsize = BP_GET_PSIZE(gbp);
1099 
1100 		ASSERT(BP_GET_COMPRESS(gbp) == ZIO_COMPRESS_OFF);
1101 		ASSERT3U(lsize, ==, BP_GET_LSIZE(gbp));
1102 		ASSERT3U(loff + lsize, <=, zio->io_size);
1103 		ASSERT(i < SPA_GBH_NBLKPTRS);
1104 		ASSERT(!BP_IS_HOLE(gbp));
1105 
1106 		zio_nowait(zio_rewrite(zio, zio->io_spa, zio->io_checksum,
1107 		    zio->io_txg, gbp, (char *)zio->io_data + loff, lsize,
1108 		    NULL, NULL, zio->io_priority, zio->io_flags,
1109 		    &zio->io_bookmark));
1110 	}
1111 
1112 	zio_push_transform(zio, gbh, gsize, gbufsize);
1113 	zio_wait_children_ready(zio);
1114 }
1115 
1116 static void
1117 zio_free_gang_members(zio_t *zio)
1118 {
1119 	zio_gbh_phys_t *gbh;
1120 	uint64_t gsize, gbufsize;
1121 	int i;
1122 
1123 	ASSERT(BP_IS_GANG(zio->io_bp));
1124 
1125 	zio_gang_byteswap(zio);
1126 	zio_pop_transform(zio, (void **)&gbh, &gsize, &gbufsize);
1127 
1128 	for (i = 0; i < SPA_GBH_NBLKPTRS; i++) {
1129 		blkptr_t *gbp = &gbh->zg_blkptr[i];
1130 
1131 		if (BP_IS_HOLE(gbp))
1132 			continue;
1133 		zio_nowait(zio_free(zio, zio->io_spa, zio->io_txg,
1134 		    gbp, NULL, NULL));
1135 	}
1136 
1137 	zio_buf_free(gbh, gbufsize);
1138 	zio_next_stage(zio);
1139 }
1140 
1141 static void
1142 zio_claim_gang_members(zio_t *zio)
1143 {
1144 	zio_gbh_phys_t *gbh;
1145 	uint64_t gsize, gbufsize;
1146 	int i;
1147 
1148 	ASSERT(BP_IS_GANG(zio->io_bp));
1149 
1150 	zio_gang_byteswap(zio);
1151 	zio_pop_transform(zio, (void **)&gbh, &gsize, &gbufsize);
1152 
1153 	for (i = 0; i < SPA_GBH_NBLKPTRS; i++) {
1154 		blkptr_t *gbp = &gbh->zg_blkptr[i];
1155 		if (BP_IS_HOLE(gbp))
1156 			continue;
1157 		zio_nowait(zio_claim(zio, zio->io_spa, zio->io_txg,
1158 		    gbp, NULL, NULL));
1159 	}
1160 
1161 	zio_buf_free(gbh, gbufsize);
1162 	zio_next_stage(zio);
1163 }
1164 
1165 static void
1166 zio_write_allocate_gang_member_done(zio_t *zio)
1167 {
1168 	zio_t *pio = zio->io_parent;
1169 	dva_t *cdva = zio->io_bp->blk_dva;
1170 	dva_t *pdva = pio->io_bp->blk_dva;
1171 	uint64_t asize;
1172 	int d;
1173 
1174 	ASSERT3U(pio->io_ndvas, ==, zio->io_ndvas);
1175 	ASSERT3U(BP_GET_NDVAS(zio->io_bp), <=, BP_GET_NDVAS(pio->io_bp));
1176 	ASSERT3U(zio->io_ndvas, <=, BP_GET_NDVAS(zio->io_bp));
1177 	ASSERT3U(pio->io_ndvas, <=, BP_GET_NDVAS(pio->io_bp));
1178 
1179 	mutex_enter(&pio->io_lock);
1180 	for (d = 0; d < BP_GET_NDVAS(pio->io_bp); d++) {
1181 		ASSERT(DVA_GET_GANG(&pdva[d]));
1182 		asize = DVA_GET_ASIZE(&pdva[d]);
1183 		asize += DVA_GET_ASIZE(&cdva[d]);
1184 		DVA_SET_ASIZE(&pdva[d], asize);
1185 	}
1186 	mutex_exit(&pio->io_lock);
1187 }
1188 
1189 static void
1190 zio_write_allocate_gang_members(zio_t *zio)
1191 {
1192 	blkptr_t *bp = zio->io_bp;
1193 	dva_t *dva = bp->blk_dva;
1194 	spa_t *spa = zio->io_spa;
1195 	zio_gbh_phys_t *gbh;
1196 	uint64_t txg = zio->io_txg;
1197 	uint64_t resid = zio->io_size;
1198 	uint64_t maxalloc = P2ROUNDUP(zio->io_size >> 1, SPA_MINBLOCKSIZE);
1199 	uint64_t gsize, loff, lsize;
1200 	uint32_t gbps_left;
1201 	int ndvas = zio->io_ndvas;
1202 	int gbh_ndvas = MIN(ndvas + 1, spa_max_replication(spa));
1203 	int error;
1204 	int i, d;
1205 
1206 	gsize = SPA_GANGBLOCKSIZE;
1207 	gbps_left = SPA_GBH_NBLKPTRS;
1208 
1209 	error = metaslab_alloc(spa, gsize, bp, gbh_ndvas, txg, NULL, B_FALSE);
1210 	if (error == ENOSPC)
1211 		panic("can't allocate gang block header");
1212 	ASSERT(error == 0);
1213 
1214 	for (d = 0; d < gbh_ndvas; d++)
1215 		DVA_SET_GANG(&dva[d], 1);
1216 
1217 	bp->blk_birth = txg;
1218 
1219 	gbh = zio_buf_alloc(gsize);
1220 	bzero(gbh, gsize);
1221 
1222 	/* We need to test multi-level gang blocks */
1223 	if (maxalloc >= zio_gang_bang && (lbolt & 0x1) == 0)
1224 		maxalloc = MAX(maxalloc >> 2, SPA_MINBLOCKSIZE);
1225 
1226 	for (loff = 0, i = 0; loff != zio->io_size;
1227 	    loff += lsize, resid -= lsize, gbps_left--, i++) {
1228 		blkptr_t *gbp = &gbh->zg_blkptr[i];
1229 		dva = gbp->blk_dva;
1230 
1231 		ASSERT(gbps_left != 0);
1232 		maxalloc = MIN(maxalloc, resid);
1233 
1234 		while (resid <= maxalloc * gbps_left) {
1235 			error = metaslab_alloc(spa, maxalloc, gbp, ndvas,
1236 			    txg, bp, B_FALSE);
1237 			if (error == 0)
1238 				break;
1239 			ASSERT3U(error, ==, ENOSPC);
1240 			if (maxalloc == SPA_MINBLOCKSIZE)
1241 				panic("really out of space");
1242 			maxalloc = P2ROUNDUP(maxalloc >> 1, SPA_MINBLOCKSIZE);
1243 		}
1244 
1245 		if (resid <= maxalloc * gbps_left) {
1246 			lsize = maxalloc;
1247 			BP_SET_LSIZE(gbp, lsize);
1248 			BP_SET_PSIZE(gbp, lsize);
1249 			BP_SET_COMPRESS(gbp, ZIO_COMPRESS_OFF);
1250 			gbp->blk_birth = txg;
1251 			zio_nowait(zio_rewrite(zio, spa,
1252 			    zio->io_checksum, txg, gbp,
1253 			    (char *)zio->io_data + loff, lsize,
1254 			    zio_write_allocate_gang_member_done, NULL,
1255 			    zio->io_priority, zio->io_flags,
1256 			    &zio->io_bookmark));
1257 		} else {
1258 			lsize = P2ROUNDUP(resid / gbps_left, SPA_MINBLOCKSIZE);
1259 			ASSERT(lsize != SPA_MINBLOCKSIZE);
1260 			zio_nowait(zio_write_allocate(zio, spa,
1261 			    zio->io_checksum, txg, gbp,
1262 			    (char *)zio->io_data + loff, lsize,
1263 			    zio_write_allocate_gang_member_done, NULL,
1264 			    zio->io_priority, zio->io_flags));
1265 		}
1266 	}
1267 
1268 	ASSERT(resid == 0 && loff == zio->io_size);
1269 
1270 	zio->io_pipeline |= 1U << ZIO_STAGE_GANG_CHECKSUM_GENERATE;
1271 
1272 	zio_push_transform(zio, gbh, gsize, gsize);
1273 	/*
1274 	 * As much as we'd like this to be zio_wait_children_ready(),
1275 	 * updating our ASIZE doesn't happen until the io_done callback,
1276 	 * so we have to wait for that to finish in order for our BP
1277 	 * to be stable.
1278 	 */
1279 	zio_wait_children_done(zio);
1280 }
1281 
1282 /*
1283  * ==========================================================================
1284  * Allocate and free blocks
1285  * ==========================================================================
1286  */
1287 static void
1288 zio_dva_allocate(zio_t *zio)
1289 {
1290 	blkptr_t *bp = zio->io_bp;
1291 	int error;
1292 
1293 	ASSERT(BP_IS_HOLE(bp));
1294 	ASSERT3U(BP_GET_NDVAS(bp), ==, 0);
1295 	ASSERT3U(zio->io_ndvas, >, 0);
1296 	ASSERT3U(zio->io_ndvas, <=, spa_max_replication(zio->io_spa));
1297 
1298 	/* For testing, make some blocks above a certain size be gang blocks */
1299 	if (zio->io_size >= zio_gang_bang && (lbolt & 0x3) == 0) {
1300 		zio_write_allocate_gang_members(zio);
1301 		return;
1302 	}
1303 
1304 	ASSERT3U(zio->io_size, ==, BP_GET_PSIZE(bp));
1305 
1306 	error = metaslab_alloc(zio->io_spa, zio->io_size, bp, zio->io_ndvas,
1307 	    zio->io_txg, NULL, B_FALSE);
1308 
1309 	if (error == 0) {
1310 		bp->blk_birth = zio->io_txg;
1311 	} else if (error == ENOSPC) {
1312 		if (zio->io_size == SPA_MINBLOCKSIZE)
1313 			panic("really, truly out of space");
1314 		zio_write_allocate_gang_members(zio);
1315 		return;
1316 	} else {
1317 		zio->io_error = error;
1318 	}
1319 	zio_next_stage(zio);
1320 }
1321 
1322 static void
1323 zio_dva_free(zio_t *zio)
1324 {
1325 	blkptr_t *bp = zio->io_bp;
1326 
1327 	metaslab_free(zio->io_spa, bp, zio->io_txg, B_FALSE);
1328 
1329 	BP_ZERO(bp);
1330 
1331 	zio_next_stage(zio);
1332 }
1333 
1334 static void
1335 zio_dva_claim(zio_t *zio)
1336 {
1337 	zio->io_error = metaslab_claim(zio->io_spa, zio->io_bp, zio->io_txg);
1338 
1339 	zio_next_stage(zio);
1340 }
1341 
1342 /*
1343  * ==========================================================================
1344  * Read and write to physical devices
1345  * ==========================================================================
1346  */
1347 
1348 static void
1349 zio_vdev_io_start(zio_t *zio)
1350 {
1351 	vdev_t *vd = zio->io_vd;
1352 	vdev_t *tvd = vd ? vd->vdev_top : NULL;
1353 	blkptr_t *bp = zio->io_bp;
1354 	uint64_t align;
1355 
1356 	if (vd == NULL) {
1357 		/* The mirror_ops handle multiple DVAs in a single BP */
1358 		vdev_mirror_ops.vdev_op_io_start(zio);
1359 		return;
1360 	}
1361 
1362 	align = 1ULL << tvd->vdev_ashift;
1363 
1364 	if (zio->io_retries == 0 && vd == tvd)
1365 		zio->io_flags |= ZIO_FLAG_FAILFAST;
1366 
1367 	if (!(zio->io_flags & ZIO_FLAG_PHYSICAL) &&
1368 	    vd->vdev_children == 0) {
1369 		zio->io_flags |= ZIO_FLAG_PHYSICAL;
1370 		zio->io_offset += VDEV_LABEL_START_SIZE;
1371 	}
1372 
1373 	if (P2PHASE(zio->io_size, align) != 0) {
1374 		uint64_t asize = P2ROUNDUP(zio->io_size, align);
1375 		char *abuf = zio_buf_alloc(asize);
1376 		ASSERT(vd == tvd);
1377 		if (zio->io_type == ZIO_TYPE_WRITE) {
1378 			bcopy(zio->io_data, abuf, zio->io_size);
1379 			bzero(abuf + zio->io_size, asize - zio->io_size);
1380 		}
1381 		zio_push_transform(zio, abuf, asize, asize);
1382 		ASSERT(!(zio->io_flags & ZIO_FLAG_SUBBLOCK));
1383 		zio->io_flags |= ZIO_FLAG_SUBBLOCK;
1384 	}
1385 
1386 	ASSERT(P2PHASE(zio->io_offset, align) == 0);
1387 	ASSERT(P2PHASE(zio->io_size, align) == 0);
1388 	ASSERT(bp == NULL ||
1389 	    P2ROUNDUP(ZIO_GET_IOSIZE(zio), align) == zio->io_size);
1390 	ASSERT(zio->io_type != ZIO_TYPE_WRITE || (spa_mode & FWRITE));
1391 
1392 	vdev_io_start(zio);
1393 
1394 	/* zio_next_stage_async() gets called from io completion interrupt */
1395 }
1396 
1397 static void
1398 zio_vdev_io_done(zio_t *zio)
1399 {
1400 	if (zio->io_vd == NULL)
1401 		/* The mirror_ops handle multiple DVAs in a single BP */
1402 		vdev_mirror_ops.vdev_op_io_done(zio);
1403 	else
1404 		vdev_io_done(zio);
1405 }
1406 
1407 /* XXPOLICY */
1408 boolean_t
1409 zio_should_retry(zio_t *zio)
1410 {
1411 	vdev_t *vd = zio->io_vd;
1412 
1413 	if (zio->io_error == 0)
1414 		return (B_FALSE);
1415 	if (zio->io_delegate_list != NULL)
1416 		return (B_FALSE);
1417 	if (vd && vd != vd->vdev_top)
1418 		return (B_FALSE);
1419 	if (zio->io_flags & ZIO_FLAG_DONT_RETRY)
1420 		return (B_FALSE);
1421 	if (zio->io_retries > 0)
1422 		return (B_FALSE);
1423 
1424 	return (B_TRUE);
1425 }
1426 
1427 static void
1428 zio_vdev_io_assess(zio_t *zio)
1429 {
1430 	vdev_t *vd = zio->io_vd;
1431 	vdev_t *tvd = vd ? vd->vdev_top : NULL;
1432 
1433 	ASSERT(zio->io_vsd == NULL);
1434 
1435 	if (zio->io_flags & ZIO_FLAG_SUBBLOCK) {
1436 		void *abuf;
1437 		uint64_t asize;
1438 		ASSERT(vd == tvd);
1439 		zio_pop_transform(zio, &abuf, &asize, &asize);
1440 		if (zio->io_type == ZIO_TYPE_READ)
1441 			bcopy(abuf, zio->io_data, zio->io_size);
1442 		zio_buf_free(abuf, asize);
1443 		zio->io_flags &= ~ZIO_FLAG_SUBBLOCK;
1444 	}
1445 
1446 	if (zio_injection_enabled && !zio->io_error)
1447 		zio->io_error = zio_handle_fault_injection(zio, EIO);
1448 
1449 	/*
1450 	 * If the I/O failed, determine whether we should attempt to retry it.
1451 	 */
1452 	/* XXPOLICY */
1453 	if (zio_should_retry(zio)) {
1454 		ASSERT(tvd == vd);
1455 
1456 		zio->io_retries++;
1457 		zio->io_error = 0;
1458 		zio->io_flags &= ZIO_FLAG_VDEV_INHERIT;
1459 		/* XXPOLICY */
1460 		zio->io_flags &= ~ZIO_FLAG_FAILFAST;
1461 		zio->io_flags |= ZIO_FLAG_DONT_CACHE;
1462 		zio->io_stage = ZIO_STAGE_VDEV_IO_START - 1;
1463 
1464 		dprintf("retry #%d for %s to %s offset %llx\n",
1465 		    zio->io_retries, zio_type_name[zio->io_type],
1466 		    vdev_description(vd), zio->io_offset);
1467 
1468 		zio_next_stage_async(zio);
1469 		return;
1470 	}
1471 
1472 	if (zio->io_error != 0 && zio->io_error != ECKSUM &&
1473 	    !(zio->io_flags & ZIO_FLAG_SPECULATIVE) && vd) {
1474 		/*
1475 		 * Poor man's hotplug support.  Even if we're done retrying this
1476 		 * I/O, try to reopen the vdev to see if it's still attached.
1477 		 * To avoid excessive thrashing, we only try it once a minute.
1478 		 * This also has the effect of detecting when missing devices
1479 		 * have come back, by polling the device once a minute.
1480 		 *
1481 		 * We need to do this asynchronously because we can't grab
1482 		 * all the necessary locks way down here.
1483 		 */
1484 		if (gethrtime() - vd->vdev_last_try > 60ULL * NANOSEC) {
1485 			vd->vdev_last_try = gethrtime();
1486 			tvd->vdev_reopen_wanted = 1;
1487 			spa_async_request(vd->vdev_spa, SPA_ASYNC_REOPEN);
1488 		}
1489 	}
1490 
1491 	zio_next_stage(zio);
1492 }
1493 
1494 void
1495 zio_vdev_io_reissue(zio_t *zio)
1496 {
1497 	ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START);
1498 	ASSERT(zio->io_error == 0);
1499 
1500 	zio->io_stage--;
1501 }
1502 
1503 void
1504 zio_vdev_io_redone(zio_t *zio)
1505 {
1506 	ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_DONE);
1507 
1508 	zio->io_stage--;
1509 }
1510 
1511 void
1512 zio_vdev_io_bypass(zio_t *zio)
1513 {
1514 	ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START);
1515 	ASSERT(zio->io_error == 0);
1516 
1517 	zio->io_flags |= ZIO_FLAG_IO_BYPASS;
1518 	zio->io_stage = ZIO_STAGE_VDEV_IO_ASSESS - 1;
1519 }
1520 
1521 /*
1522  * ==========================================================================
1523  * Generate and verify checksums
1524  * ==========================================================================
1525  */
1526 static void
1527 zio_checksum_generate(zio_t *zio)
1528 {
1529 	int checksum = zio->io_checksum;
1530 	blkptr_t *bp = zio->io_bp;
1531 
1532 	ASSERT3U(zio->io_size, ==, BP_GET_PSIZE(bp));
1533 
1534 	BP_SET_CHECKSUM(bp, checksum);
1535 	BP_SET_BYTEORDER(bp, ZFS_HOST_BYTEORDER);
1536 
1537 	zio_checksum(checksum, &bp->blk_cksum, zio->io_data, zio->io_size);
1538 
1539 	zio_next_stage(zio);
1540 }
1541 
1542 static void
1543 zio_gang_checksum_generate(zio_t *zio)
1544 {
1545 	zio_cksum_t zc;
1546 	zio_gbh_phys_t *gbh = zio->io_data;
1547 
1548 	ASSERT(BP_IS_GANG(zio->io_bp));
1549 	ASSERT3U(zio->io_size, ==, SPA_GANGBLOCKSIZE);
1550 
1551 	zio_set_gang_verifier(zio, &gbh->zg_tail.zbt_cksum);
1552 
1553 	zio_checksum(ZIO_CHECKSUM_GANG_HEADER, &zc, zio->io_data, zio->io_size);
1554 
1555 	zio_next_stage(zio);
1556 }
1557 
1558 static void
1559 zio_checksum_verify(zio_t *zio)
1560 {
1561 	if (zio->io_bp != NULL) {
1562 		zio->io_error = zio_checksum_error(zio);
1563 		if (zio->io_error && !(zio->io_flags & ZIO_FLAG_SPECULATIVE))
1564 			zfs_ereport_post(FM_EREPORT_ZFS_CHECKSUM,
1565 			    zio->io_spa, zio->io_vd, zio, 0, 0);
1566 	}
1567 
1568 	zio_next_stage(zio);
1569 }
1570 
1571 /*
1572  * Called by RAID-Z to ensure we don't compute the checksum twice.
1573  */
1574 void
1575 zio_checksum_verified(zio_t *zio)
1576 {
1577 	zio->io_pipeline &= ~(1U << ZIO_STAGE_CHECKSUM_VERIFY);
1578 }
1579 
1580 /*
1581  * Set the external verifier for a gang block based on stuff in the bp
1582  */
1583 void
1584 zio_set_gang_verifier(zio_t *zio, zio_cksum_t *zcp)
1585 {
1586 	blkptr_t *bp = zio->io_bp;
1587 
1588 	zcp->zc_word[0] = DVA_GET_VDEV(BP_IDENTITY(bp));
1589 	zcp->zc_word[1] = DVA_GET_OFFSET(BP_IDENTITY(bp));
1590 	zcp->zc_word[2] = bp->blk_birth;
1591 	zcp->zc_word[3] = 0;
1592 }
1593 
1594 /*
1595  * ==========================================================================
1596  * Define the pipeline
1597  * ==========================================================================
1598  */
1599 typedef void zio_pipe_stage_t(zio_t *zio);
1600 
1601 static void
1602 zio_badop(zio_t *zio)
1603 {
1604 	panic("Invalid I/O pipeline stage %u for zio %p", zio->io_stage, zio);
1605 }
1606 
1607 zio_pipe_stage_t *zio_pipeline[ZIO_STAGE_DONE + 2] = {
1608 	zio_badop,
1609 	zio_wait_children_ready,
1610 	zio_write_compress,
1611 	zio_checksum_generate,
1612 	zio_gang_pipeline,
1613 	zio_get_gang_header,
1614 	zio_rewrite_gang_members,
1615 	zio_free_gang_members,
1616 	zio_claim_gang_members,
1617 	zio_dva_allocate,
1618 	zio_dva_free,
1619 	zio_dva_claim,
1620 	zio_gang_checksum_generate,
1621 	zio_ready,
1622 	zio_vdev_io_start,
1623 	zio_vdev_io_done,
1624 	zio_vdev_io_assess,
1625 	zio_wait_children_done,
1626 	zio_checksum_verify,
1627 	zio_read_gang_members,
1628 	zio_read_decompress,
1629 	zio_done,
1630 	zio_badop
1631 };
1632 
1633 /*
1634  * Move an I/O to the next stage of the pipeline and execute that stage.
1635  * There's no locking on io_stage because there's no legitimate way for
1636  * multiple threads to be attempting to process the same I/O.
1637  */
1638 void
1639 zio_next_stage(zio_t *zio)
1640 {
1641 	uint32_t pipeline = zio->io_pipeline;
1642 
1643 	ASSERT(!MUTEX_HELD(&zio->io_lock));
1644 
1645 	if (zio->io_error) {
1646 		dprintf("zio %p vdev %s offset %llx stage %d error %d\n",
1647 		    zio, vdev_description(zio->io_vd),
1648 		    zio->io_offset, zio->io_stage, zio->io_error);
1649 		if (((1U << zio->io_stage) & ZIO_VDEV_IO_PIPELINE) == 0)
1650 			pipeline &= ZIO_ERROR_PIPELINE_MASK;
1651 	}
1652 
1653 	while (((1U << ++zio->io_stage) & pipeline) == 0)
1654 		continue;
1655 
1656 	ASSERT(zio->io_stage <= ZIO_STAGE_DONE);
1657 	ASSERT(zio->io_stalled == 0);
1658 
1659 	zio_pipeline[zio->io_stage](zio);
1660 }
1661 
1662 void
1663 zio_next_stage_async(zio_t *zio)
1664 {
1665 	taskq_t *tq;
1666 	uint32_t pipeline = zio->io_pipeline;
1667 
1668 	ASSERT(!MUTEX_HELD(&zio->io_lock));
1669 
1670 	if (zio->io_error) {
1671 		dprintf("zio %p vdev %s offset %llx stage %d error %d\n",
1672 		    zio, vdev_description(zio->io_vd),
1673 		    zio->io_offset, zio->io_stage, zio->io_error);
1674 		if (((1U << zio->io_stage) & ZIO_VDEV_IO_PIPELINE) == 0)
1675 			pipeline &= ZIO_ERROR_PIPELINE_MASK;
1676 	}
1677 
1678 	while (((1U << ++zio->io_stage) & pipeline) == 0)
1679 		continue;
1680 
1681 	ASSERT(zio->io_stage <= ZIO_STAGE_DONE);
1682 	ASSERT(zio->io_stalled == 0);
1683 
1684 	/*
1685 	 * For performance, we'll probably want two sets of task queues:
1686 	 * per-CPU issue taskqs and per-CPU completion taskqs.  The per-CPU
1687 	 * part is for read performance: since we have to make a pass over
1688 	 * the data to checksum it anyway, we want to do this on the same CPU
1689 	 * that issued the read, because (assuming CPU scheduling affinity)
1690 	 * that thread is probably still there.  Getting this optimization
1691 	 * right avoids performance-hostile cache-to-cache transfers.
1692 	 *
1693 	 * Note that having two sets of task queues is also necessary for
1694 	 * correctness: if all of the issue threads get bogged down waiting
1695 	 * for dependent reads (e.g. metaslab freelist) to complete, then
1696 	 * there won't be any threads available to service I/O completion
1697 	 * interrupts.
1698 	 */
1699 	if ((1U << zio->io_stage) & zio->io_async_stages) {
1700 		if (zio->io_stage < ZIO_STAGE_VDEV_IO_DONE)
1701 			tq = zio->io_spa->spa_zio_issue_taskq[zio->io_type];
1702 		else
1703 			tq = zio->io_spa->spa_zio_intr_taskq[zio->io_type];
1704 		(void) taskq_dispatch(tq,
1705 		    (task_func_t *)zio_pipeline[zio->io_stage], zio, TQ_SLEEP);
1706 	} else {
1707 		zio_pipeline[zio->io_stage](zio);
1708 	}
1709 }
1710 
1711 /*
1712  * Try to allocate an intent log block.  Return 0 on success, errno on failure.
1713  */
1714 int
1715 zio_alloc_blk(spa_t *spa, uint64_t size, blkptr_t *new_bp, blkptr_t *old_bp,
1716     uint64_t txg)
1717 {
1718 	int error;
1719 
1720 	spa_config_enter(spa, RW_READER, FTAG);
1721 
1722 	/*
1723 	 * We were passed the previous log blocks dva_t in bp->blk_dva[0].
1724 	 */
1725 	error = metaslab_alloc(spa, size, new_bp, 1, txg, old_bp, B_TRUE);
1726 
1727 	if (error == 0) {
1728 		BP_SET_LSIZE(new_bp, size);
1729 		BP_SET_PSIZE(new_bp, size);
1730 		BP_SET_COMPRESS(new_bp, ZIO_COMPRESS_OFF);
1731 		BP_SET_CHECKSUM(new_bp, ZIO_CHECKSUM_ZILOG);
1732 		BP_SET_TYPE(new_bp, DMU_OT_INTENT_LOG);
1733 		BP_SET_LEVEL(new_bp, 0);
1734 		BP_SET_BYTEORDER(new_bp, ZFS_HOST_BYTEORDER);
1735 		new_bp->blk_birth = txg;
1736 	}
1737 
1738 	spa_config_exit(spa, FTAG);
1739 
1740 	return (error);
1741 }
1742 
1743 /*
1744  * Free an intent log block.  We know it can't be a gang block, so there's
1745  * nothing to do except metaslab_free() it.
1746  */
1747 void
1748 zio_free_blk(spa_t *spa, blkptr_t *bp, uint64_t txg)
1749 {
1750 	ASSERT(!BP_IS_GANG(bp));
1751 
1752 	spa_config_enter(spa, RW_READER, FTAG);
1753 
1754 	metaslab_free(spa, bp, txg, B_FALSE);
1755 
1756 	spa_config_exit(spa, FTAG);
1757 }
1758