xref: /openbmc/qemu/block/qed.c (revision e0091133)
1 /*
2  * QEMU Enhanced Disk Format
3  *
4  * Copyright IBM, Corp. 2010
5  *
6  * Authors:
7  *  Stefan Hajnoczi   <stefanha@linux.vnet.ibm.com>
8  *  Anthony Liguori   <aliguori@us.ibm.com>
9  *
10  * This work is licensed under the terms of the GNU LGPL, version 2 or later.
11  * See the COPYING.LIB file in the top-level directory.
12  *
13  */
14 
15 #include "qemu/osdep.h"
16 #include "block/qdict.h"
17 #include "qapi/error.h"
18 #include "qemu/timer.h"
19 #include "qemu/bswap.h"
20 #include "qemu/main-loop.h"
21 #include "qemu/module.h"
22 #include "qemu/option.h"
23 #include "qemu/memalign.h"
24 #include "trace.h"
25 #include "qed.h"
26 #include "sysemu/block-backend.h"
27 #include "qapi/qmp/qdict.h"
28 #include "qapi/qobject-input-visitor.h"
29 #include "qapi/qapi-visit-block-core.h"
30 
31 static QemuOptsList qed_create_opts;
32 
33 static int bdrv_qed_probe(const uint8_t *buf, int buf_size,
34                           const char *filename)
35 {
36     const QEDHeader *header = (const QEDHeader *)buf;
37 
38     if (buf_size < sizeof(*header)) {
39         return 0;
40     }
41     if (le32_to_cpu(header->magic) != QED_MAGIC) {
42         return 0;
43     }
44     return 100;
45 }
46 
47 /**
48  * Check whether an image format is raw
49  *
50  * @fmt:    Backing file format, may be NULL
51  */
52 static bool qed_fmt_is_raw(const char *fmt)
53 {
54     return fmt && strcmp(fmt, "raw") == 0;
55 }
56 
57 static void qed_header_le_to_cpu(const QEDHeader *le, QEDHeader *cpu)
58 {
59     cpu->magic = le32_to_cpu(le->magic);
60     cpu->cluster_size = le32_to_cpu(le->cluster_size);
61     cpu->table_size = le32_to_cpu(le->table_size);
62     cpu->header_size = le32_to_cpu(le->header_size);
63     cpu->features = le64_to_cpu(le->features);
64     cpu->compat_features = le64_to_cpu(le->compat_features);
65     cpu->autoclear_features = le64_to_cpu(le->autoclear_features);
66     cpu->l1_table_offset = le64_to_cpu(le->l1_table_offset);
67     cpu->image_size = le64_to_cpu(le->image_size);
68     cpu->backing_filename_offset = le32_to_cpu(le->backing_filename_offset);
69     cpu->backing_filename_size = le32_to_cpu(le->backing_filename_size);
70 }
71 
72 static void qed_header_cpu_to_le(const QEDHeader *cpu, QEDHeader *le)
73 {
74     le->magic = cpu_to_le32(cpu->magic);
75     le->cluster_size = cpu_to_le32(cpu->cluster_size);
76     le->table_size = cpu_to_le32(cpu->table_size);
77     le->header_size = cpu_to_le32(cpu->header_size);
78     le->features = cpu_to_le64(cpu->features);
79     le->compat_features = cpu_to_le64(cpu->compat_features);
80     le->autoclear_features = cpu_to_le64(cpu->autoclear_features);
81     le->l1_table_offset = cpu_to_le64(cpu->l1_table_offset);
82     le->image_size = cpu_to_le64(cpu->image_size);
83     le->backing_filename_offset = cpu_to_le32(cpu->backing_filename_offset);
84     le->backing_filename_size = cpu_to_le32(cpu->backing_filename_size);
85 }
86 
87 int qed_write_header_sync(BDRVQEDState *s)
88 {
89     QEDHeader le;
90 
91     qed_header_cpu_to_le(&s->header, &le);
92     return bdrv_pwrite(s->bs->file, 0, sizeof(le), &le, 0);
93 }
94 
95 /**
96  * Update header in-place (does not rewrite backing filename or other strings)
97  *
98  * This function only updates known header fields in-place and does not affect
99  * extra data after the QED header.
100  *
101  * No new allocating reqs can start while this function runs.
102  */
103 static int coroutine_fn qed_write_header(BDRVQEDState *s)
104 {
105     /* We must write full sectors for O_DIRECT but cannot necessarily generate
106      * the data following the header if an unrecognized compat feature is
107      * active.  Therefore, first read the sectors containing the header, update
108      * them, and write back.
109      */
110 
111     int nsectors = DIV_ROUND_UP(sizeof(QEDHeader), BDRV_SECTOR_SIZE);
112     size_t len = nsectors * BDRV_SECTOR_SIZE;
113     uint8_t *buf;
114     int ret;
115 
116     assert(s->allocating_acb || s->allocating_write_reqs_plugged);
117 
118     buf = qemu_blockalign(s->bs, len);
119 
120     ret = bdrv_co_pread(s->bs->file, 0, len, buf, 0);
121     if (ret < 0) {
122         goto out;
123     }
124 
125     /* Update header */
126     qed_header_cpu_to_le(&s->header, (QEDHeader *) buf);
127 
128     ret = bdrv_co_pwrite(s->bs->file, 0, len,  buf, 0);
129     if (ret < 0) {
130         goto out;
131     }
132 
133     ret = 0;
134 out:
135     qemu_vfree(buf);
136     return ret;
137 }
138 
139 static uint64_t qed_max_image_size(uint32_t cluster_size, uint32_t table_size)
140 {
141     uint64_t table_entries;
142     uint64_t l2_size;
143 
144     table_entries = (table_size * cluster_size) / sizeof(uint64_t);
145     l2_size = table_entries * cluster_size;
146 
147     return l2_size * table_entries;
148 }
149 
150 static bool qed_is_cluster_size_valid(uint32_t cluster_size)
151 {
152     if (cluster_size < QED_MIN_CLUSTER_SIZE ||
153         cluster_size > QED_MAX_CLUSTER_SIZE) {
154         return false;
155     }
156     if (cluster_size & (cluster_size - 1)) {
157         return false; /* not power of 2 */
158     }
159     return true;
160 }
161 
162 static bool qed_is_table_size_valid(uint32_t table_size)
163 {
164     if (table_size < QED_MIN_TABLE_SIZE ||
165         table_size > QED_MAX_TABLE_SIZE) {
166         return false;
167     }
168     if (table_size & (table_size - 1)) {
169         return false; /* not power of 2 */
170     }
171     return true;
172 }
173 
174 static bool qed_is_image_size_valid(uint64_t image_size, uint32_t cluster_size,
175                                     uint32_t table_size)
176 {
177     if (image_size % BDRV_SECTOR_SIZE != 0) {
178         return false; /* not multiple of sector size */
179     }
180     if (image_size > qed_max_image_size(cluster_size, table_size)) {
181         return false; /* image is too large */
182     }
183     return true;
184 }
185 
186 /**
187  * Read a string of known length from the image file
188  *
189  * @file:       Image file
190  * @offset:     File offset to start of string, in bytes
191  * @n:          String length in bytes
192  * @buf:        Destination buffer
193  * @buflen:     Destination buffer length in bytes
194  * @ret:        0 on success, -errno on failure
195  *
196  * The string is NUL-terminated.
197  */
198 static int qed_read_string(BdrvChild *file, uint64_t offset, size_t n,
199                            char *buf, size_t buflen)
200 {
201     int ret;
202     if (n >= buflen) {
203         return -EINVAL;
204     }
205     ret = bdrv_pread(file, offset, n, buf, 0);
206     if (ret < 0) {
207         return ret;
208     }
209     buf[n] = '\0';
210     return 0;
211 }
212 
213 /**
214  * Allocate new clusters
215  *
216  * @s:          QED state
217  * @n:          Number of contiguous clusters to allocate
218  * @ret:        Offset of first allocated cluster
219  *
220  * This function only produces the offset where the new clusters should be
221  * written.  It updates BDRVQEDState but does not make any changes to the image
222  * file.
223  *
224  * Called with table_lock held.
225  */
226 static uint64_t qed_alloc_clusters(BDRVQEDState *s, unsigned int n)
227 {
228     uint64_t offset = s->file_size;
229     s->file_size += n * s->header.cluster_size;
230     return offset;
231 }
232 
233 QEDTable *qed_alloc_table(BDRVQEDState *s)
234 {
235     /* Honor O_DIRECT memory alignment requirements */
236     return qemu_blockalign(s->bs,
237                            s->header.cluster_size * s->header.table_size);
238 }
239 
240 /**
241  * Allocate a new zeroed L2 table
242  *
243  * Called with table_lock held.
244  */
245 static CachedL2Table *qed_new_l2_table(BDRVQEDState *s)
246 {
247     CachedL2Table *l2_table = qed_alloc_l2_cache_entry(&s->l2_cache);
248 
249     l2_table->table = qed_alloc_table(s);
250     l2_table->offset = qed_alloc_clusters(s, s->header.table_size);
251 
252     memset(l2_table->table->offsets, 0,
253            s->header.cluster_size * s->header.table_size);
254     return l2_table;
255 }
256 
257 static bool coroutine_fn qed_plug_allocating_write_reqs(BDRVQEDState *s)
258 {
259     qemu_co_mutex_lock(&s->table_lock);
260 
261     /* No reentrancy is allowed.  */
262     assert(!s->allocating_write_reqs_plugged);
263     if (s->allocating_acb != NULL) {
264         /* Another allocating write came concurrently.  This cannot happen
265          * from bdrv_qed_co_drain_begin, but it can happen when the timer runs.
266          */
267         qemu_co_mutex_unlock(&s->table_lock);
268         return false;
269     }
270 
271     s->allocating_write_reqs_plugged = true;
272     qemu_co_mutex_unlock(&s->table_lock);
273     return true;
274 }
275 
276 static void coroutine_fn qed_unplug_allocating_write_reqs(BDRVQEDState *s)
277 {
278     qemu_co_mutex_lock(&s->table_lock);
279     assert(s->allocating_write_reqs_plugged);
280     s->allocating_write_reqs_plugged = false;
281     qemu_co_queue_next(&s->allocating_write_reqs);
282     qemu_co_mutex_unlock(&s->table_lock);
283 }
284 
285 static void coroutine_fn qed_need_check_timer_entry(void *opaque)
286 {
287     BDRVQEDState *s = opaque;
288     int ret;
289 
290     trace_qed_need_check_timer_cb(s);
291 
292     if (!qed_plug_allocating_write_reqs(s)) {
293         return;
294     }
295 
296     /* Ensure writes are on disk before clearing flag */
297     ret = bdrv_co_flush(s->bs->file->bs);
298     if (ret < 0) {
299         qed_unplug_allocating_write_reqs(s);
300         return;
301     }
302 
303     s->header.features &= ~QED_F_NEED_CHECK;
304     ret = qed_write_header(s);
305     (void) ret;
306 
307     qed_unplug_allocating_write_reqs(s);
308 
309     ret = bdrv_co_flush(s->bs);
310     (void) ret;
311 }
312 
313 static void qed_need_check_timer_cb(void *opaque)
314 {
315     Coroutine *co = qemu_coroutine_create(qed_need_check_timer_entry, opaque);
316     qemu_coroutine_enter(co);
317 }
318 
319 static void qed_start_need_check_timer(BDRVQEDState *s)
320 {
321     trace_qed_start_need_check_timer(s);
322 
323     /* Use QEMU_CLOCK_VIRTUAL so we don't alter the image file while suspended for
324      * migration.
325      */
326     timer_mod(s->need_check_timer, qemu_clock_get_ns(QEMU_CLOCK_VIRTUAL) +
327                    NANOSECONDS_PER_SECOND * QED_NEED_CHECK_TIMEOUT);
328 }
329 
330 /* It's okay to call this multiple times or when no timer is started */
331 static void qed_cancel_need_check_timer(BDRVQEDState *s)
332 {
333     trace_qed_cancel_need_check_timer(s);
334     timer_del(s->need_check_timer);
335 }
336 
337 static void bdrv_qed_detach_aio_context(BlockDriverState *bs)
338 {
339     BDRVQEDState *s = bs->opaque;
340 
341     qed_cancel_need_check_timer(s);
342     timer_free(s->need_check_timer);
343 }
344 
345 static void bdrv_qed_attach_aio_context(BlockDriverState *bs,
346                                         AioContext *new_context)
347 {
348     BDRVQEDState *s = bs->opaque;
349 
350     s->need_check_timer = aio_timer_new(new_context,
351                                         QEMU_CLOCK_VIRTUAL, SCALE_NS,
352                                         qed_need_check_timer_cb, s);
353     if (s->header.features & QED_F_NEED_CHECK) {
354         qed_start_need_check_timer(s);
355     }
356 }
357 
358 static void coroutine_fn bdrv_qed_co_drain_begin(BlockDriverState *bs)
359 {
360     BDRVQEDState *s = bs->opaque;
361 
362     /* Fire the timer immediately in order to start doing I/O as soon as the
363      * header is flushed.
364      */
365     if (s->need_check_timer && timer_pending(s->need_check_timer)) {
366         qed_cancel_need_check_timer(s);
367         qed_need_check_timer_entry(s);
368     }
369 }
370 
371 static void bdrv_qed_init_state(BlockDriverState *bs)
372 {
373     BDRVQEDState *s = bs->opaque;
374 
375     memset(s, 0, sizeof(BDRVQEDState));
376     s->bs = bs;
377     qemu_co_mutex_init(&s->table_lock);
378     qemu_co_queue_init(&s->allocating_write_reqs);
379 }
380 
381 /* Called with table_lock held.  */
382 static int coroutine_fn bdrv_qed_do_open(BlockDriverState *bs, QDict *options,
383                                          int flags, Error **errp)
384 {
385     BDRVQEDState *s = bs->opaque;
386     QEDHeader le_header;
387     int64_t file_size;
388     int ret;
389 
390     ret = bdrv_co_pread(bs->file, 0, sizeof(le_header), &le_header, 0);
391     if (ret < 0) {
392         error_setg(errp, "Failed to read QED header");
393         return ret;
394     }
395     qed_header_le_to_cpu(&le_header, &s->header);
396 
397     if (s->header.magic != QED_MAGIC) {
398         error_setg(errp, "Image not in QED format");
399         return -EINVAL;
400     }
401     if (s->header.features & ~QED_FEATURE_MASK) {
402         /* image uses unsupported feature bits */
403         error_setg(errp, "Unsupported QED features: %" PRIx64,
404                    s->header.features & ~QED_FEATURE_MASK);
405         return -ENOTSUP;
406     }
407     if (!qed_is_cluster_size_valid(s->header.cluster_size)) {
408         error_setg(errp, "QED cluster size is invalid");
409         return -EINVAL;
410     }
411 
412     /* Round down file size to the last cluster */
413     file_size = bdrv_getlength(bs->file->bs);
414     if (file_size < 0) {
415         error_setg(errp, "Failed to get file length");
416         return file_size;
417     }
418     s->file_size = qed_start_of_cluster(s, file_size);
419 
420     if (!qed_is_table_size_valid(s->header.table_size)) {
421         error_setg(errp, "QED table size is invalid");
422         return -EINVAL;
423     }
424     if (!qed_is_image_size_valid(s->header.image_size,
425                                  s->header.cluster_size,
426                                  s->header.table_size)) {
427         error_setg(errp, "QED image size is invalid");
428         return -EINVAL;
429     }
430     if (!qed_check_table_offset(s, s->header.l1_table_offset)) {
431         error_setg(errp, "QED table offset is invalid");
432         return -EINVAL;
433     }
434 
435     s->table_nelems = (s->header.cluster_size * s->header.table_size) /
436                       sizeof(uint64_t);
437     s->l2_shift = ctz32(s->header.cluster_size);
438     s->l2_mask = s->table_nelems - 1;
439     s->l1_shift = s->l2_shift + ctz32(s->table_nelems);
440 
441     /* Header size calculation must not overflow uint32_t */
442     if (s->header.header_size > UINT32_MAX / s->header.cluster_size) {
443         error_setg(errp, "QED header size is too large");
444         return -EINVAL;
445     }
446 
447     if ((s->header.features & QED_F_BACKING_FILE)) {
448         g_autofree char *backing_file_str = NULL;
449 
450         if ((uint64_t)s->header.backing_filename_offset +
451             s->header.backing_filename_size >
452             s->header.cluster_size * s->header.header_size) {
453             error_setg(errp, "QED backing filename offset is invalid");
454             return -EINVAL;
455         }
456 
457         backing_file_str = g_malloc(sizeof(bs->backing_file));
458         ret = qed_read_string(bs->file, s->header.backing_filename_offset,
459                               s->header.backing_filename_size,
460                               backing_file_str, sizeof(bs->backing_file));
461         if (ret < 0) {
462             error_setg(errp, "Failed to read backing filename");
463             return ret;
464         }
465 
466         if (!g_str_equal(backing_file_str, bs->backing_file)) {
467             pstrcpy(bs->backing_file, sizeof(bs->backing_file),
468                     backing_file_str);
469             pstrcpy(bs->auto_backing_file, sizeof(bs->auto_backing_file),
470                     backing_file_str);
471         }
472 
473         if (s->header.features & QED_F_BACKING_FORMAT_NO_PROBE) {
474             pstrcpy(bs->backing_format, sizeof(bs->backing_format), "raw");
475         }
476     }
477 
478     /* Reset unknown autoclear feature bits.  This is a backwards
479      * compatibility mechanism that allows images to be opened by older
480      * programs, which "knock out" unknown feature bits.  When an image is
481      * opened by a newer program again it can detect that the autoclear
482      * feature is no longer valid.
483      */
484     if ((s->header.autoclear_features & ~QED_AUTOCLEAR_FEATURE_MASK) != 0 &&
485         !bdrv_is_read_only(bs->file->bs) && !(flags & BDRV_O_INACTIVE)) {
486         s->header.autoclear_features &= QED_AUTOCLEAR_FEATURE_MASK;
487 
488         ret = qed_write_header_sync(s);
489         if (ret) {
490             error_setg(errp, "Failed to update header");
491             return ret;
492         }
493 
494         /* From here on only known autoclear feature bits are valid */
495         bdrv_co_flush(bs->file->bs);
496     }
497 
498     s->l1_table = qed_alloc_table(s);
499     qed_init_l2_cache(&s->l2_cache);
500 
501     ret = qed_read_l1_table_sync(s);
502     if (ret) {
503         error_setg(errp, "Failed to read L1 table");
504         goto out;
505     }
506 
507     /* If image was not closed cleanly, check consistency */
508     if (!(flags & BDRV_O_CHECK) && (s->header.features & QED_F_NEED_CHECK)) {
509         /* Read-only images cannot be fixed.  There is no risk of corruption
510          * since write operations are not possible.  Therefore, allow
511          * potentially inconsistent images to be opened read-only.  This can
512          * aid data recovery from an otherwise inconsistent image.
513          */
514         if (!bdrv_is_read_only(bs->file->bs) &&
515             !(flags & BDRV_O_INACTIVE)) {
516             BdrvCheckResult result = {0};
517 
518             ret = qed_check(s, &result, true);
519             if (ret) {
520                 error_setg(errp, "Image corrupted");
521                 goto out;
522             }
523         }
524     }
525 
526     bdrv_qed_attach_aio_context(bs, bdrv_get_aio_context(bs));
527 
528 out:
529     if (ret) {
530         qed_free_l2_cache(&s->l2_cache);
531         qemu_vfree(s->l1_table);
532     }
533     return ret;
534 }
535 
536 typedef struct QEDOpenCo {
537     BlockDriverState *bs;
538     QDict *options;
539     int flags;
540     Error **errp;
541     int ret;
542 } QEDOpenCo;
543 
544 static void coroutine_fn bdrv_qed_open_entry(void *opaque)
545 {
546     QEDOpenCo *qoc = opaque;
547     BDRVQEDState *s = qoc->bs->opaque;
548 
549     qemu_co_mutex_lock(&s->table_lock);
550     qoc->ret = bdrv_qed_do_open(qoc->bs, qoc->options, qoc->flags, qoc->errp);
551     qemu_co_mutex_unlock(&s->table_lock);
552 }
553 
554 static int bdrv_qed_open(BlockDriverState *bs, QDict *options, int flags,
555                          Error **errp)
556 {
557     QEDOpenCo qoc = {
558         .bs = bs,
559         .options = options,
560         .flags = flags,
561         .errp = errp,
562         .ret = -EINPROGRESS
563     };
564     int ret;
565 
566     ret = bdrv_open_file_child(NULL, options, "file", bs, errp);
567     if (ret < 0) {
568         return ret;
569     }
570 
571     bdrv_qed_init_state(bs);
572     if (qemu_in_coroutine()) {
573         bdrv_qed_open_entry(&qoc);
574     } else {
575         assert(qemu_get_current_aio_context() == qemu_get_aio_context());
576         qemu_coroutine_enter(qemu_coroutine_create(bdrv_qed_open_entry, &qoc));
577         BDRV_POLL_WHILE(bs, qoc.ret == -EINPROGRESS);
578     }
579     BDRV_POLL_WHILE(bs, qoc.ret == -EINPROGRESS);
580     return qoc.ret;
581 }
582 
583 static void bdrv_qed_refresh_limits(BlockDriverState *bs, Error **errp)
584 {
585     BDRVQEDState *s = bs->opaque;
586 
587     bs->bl.pwrite_zeroes_alignment = s->header.cluster_size;
588     bs->bl.max_pwrite_zeroes = QEMU_ALIGN_DOWN(INT_MAX, s->header.cluster_size);
589 }
590 
591 /* We have nothing to do for QED reopen, stubs just return
592  * success */
593 static int bdrv_qed_reopen_prepare(BDRVReopenState *state,
594                                    BlockReopenQueue *queue, Error **errp)
595 {
596     return 0;
597 }
598 
599 static void bdrv_qed_close(BlockDriverState *bs)
600 {
601     BDRVQEDState *s = bs->opaque;
602 
603     bdrv_qed_detach_aio_context(bs);
604 
605     /* Ensure writes reach stable storage */
606     bdrv_flush(bs->file->bs);
607 
608     /* Clean shutdown, no check required on next open */
609     if (s->header.features & QED_F_NEED_CHECK) {
610         s->header.features &= ~QED_F_NEED_CHECK;
611         qed_write_header_sync(s);
612     }
613 
614     qed_free_l2_cache(&s->l2_cache);
615     qemu_vfree(s->l1_table);
616 }
617 
618 static int coroutine_fn bdrv_qed_co_create(BlockdevCreateOptions *opts,
619                                            Error **errp)
620 {
621     BlockdevCreateOptionsQed *qed_opts;
622     BlockBackend *blk = NULL;
623     BlockDriverState *bs = NULL;
624 
625     QEDHeader header;
626     QEDHeader le_header;
627     uint8_t *l1_table = NULL;
628     size_t l1_size;
629     int ret = 0;
630 
631     assert(opts->driver == BLOCKDEV_DRIVER_QED);
632     qed_opts = &opts->u.qed;
633 
634     /* Validate options and set default values */
635     if (!qed_opts->has_cluster_size) {
636         qed_opts->cluster_size = QED_DEFAULT_CLUSTER_SIZE;
637     }
638     if (!qed_opts->has_table_size) {
639         qed_opts->table_size = QED_DEFAULT_TABLE_SIZE;
640     }
641 
642     if (!qed_is_cluster_size_valid(qed_opts->cluster_size)) {
643         error_setg(errp, "QED cluster size must be within range [%u, %u] "
644                          "and power of 2",
645                    QED_MIN_CLUSTER_SIZE, QED_MAX_CLUSTER_SIZE);
646         return -EINVAL;
647     }
648     if (!qed_is_table_size_valid(qed_opts->table_size)) {
649         error_setg(errp, "QED table size must be within range [%u, %u] "
650                          "and power of 2",
651                    QED_MIN_TABLE_SIZE, QED_MAX_TABLE_SIZE);
652         return -EINVAL;
653     }
654     if (!qed_is_image_size_valid(qed_opts->size, qed_opts->cluster_size,
655                                  qed_opts->table_size))
656     {
657         error_setg(errp, "QED image size must be a non-zero multiple of "
658                          "cluster size and less than %" PRIu64 " bytes",
659                    qed_max_image_size(qed_opts->cluster_size,
660                                       qed_opts->table_size));
661         return -EINVAL;
662     }
663 
664     /* Create BlockBackend to write to the image */
665     bs = bdrv_open_blockdev_ref(qed_opts->file, errp);
666     if (bs == NULL) {
667         return -EIO;
668     }
669 
670     blk = blk_new_with_bs(bs, BLK_PERM_WRITE | BLK_PERM_RESIZE, BLK_PERM_ALL,
671                           errp);
672     if (!blk) {
673         ret = -EPERM;
674         goto out;
675     }
676     blk_set_allow_write_beyond_eof(blk, true);
677 
678     /* Prepare image format */
679     header = (QEDHeader) {
680         .magic = QED_MAGIC,
681         .cluster_size = qed_opts->cluster_size,
682         .table_size = qed_opts->table_size,
683         .header_size = 1,
684         .features = 0,
685         .compat_features = 0,
686         .l1_table_offset = qed_opts->cluster_size,
687         .image_size = qed_opts->size,
688     };
689 
690     l1_size = header.cluster_size * header.table_size;
691 
692     /*
693      * The QED format associates file length with allocation status,
694      * so a new file (which is empty) must have a length of 0.
695      */
696     ret = blk_co_truncate(blk, 0, true, PREALLOC_MODE_OFF, 0, errp);
697     if (ret < 0) {
698         goto out;
699     }
700 
701     if (qed_opts->has_backing_file) {
702         header.features |= QED_F_BACKING_FILE;
703         header.backing_filename_offset = sizeof(le_header);
704         header.backing_filename_size = strlen(qed_opts->backing_file);
705 
706         if (qed_opts->has_backing_fmt) {
707             const char *backing_fmt = BlockdevDriver_str(qed_opts->backing_fmt);
708             if (qed_fmt_is_raw(backing_fmt)) {
709                 header.features |= QED_F_BACKING_FORMAT_NO_PROBE;
710             }
711         }
712     }
713 
714     qed_header_cpu_to_le(&header, &le_header);
715     ret = blk_co_pwrite(blk, 0, sizeof(le_header), &le_header, 0);
716     if (ret < 0) {
717         goto out;
718     }
719     ret = blk_co_pwrite(blk, sizeof(le_header), header.backing_filename_size,
720                      qed_opts->backing_file, 0);
721     if (ret < 0) {
722         goto out;
723     }
724 
725     l1_table = g_malloc0(l1_size);
726     ret = blk_co_pwrite(blk, header.l1_table_offset, l1_size, l1_table, 0);
727     if (ret < 0) {
728         goto out;
729     }
730 
731     ret = 0; /* success */
732 out:
733     g_free(l1_table);
734     blk_unref(blk);
735     bdrv_unref(bs);
736     return ret;
737 }
738 
739 static int coroutine_fn bdrv_qed_co_create_opts(BlockDriver *drv,
740                                                 const char *filename,
741                                                 QemuOpts *opts,
742                                                 Error **errp)
743 {
744     BlockdevCreateOptions *create_options = NULL;
745     QDict *qdict;
746     Visitor *v;
747     BlockDriverState *bs = NULL;
748     int ret;
749 
750     static const QDictRenames opt_renames[] = {
751         { BLOCK_OPT_BACKING_FILE,       "backing-file" },
752         { BLOCK_OPT_BACKING_FMT,        "backing-fmt" },
753         { BLOCK_OPT_CLUSTER_SIZE,       "cluster-size" },
754         { BLOCK_OPT_TABLE_SIZE,         "table-size" },
755         { NULL, NULL },
756     };
757 
758     /* Parse options and convert legacy syntax */
759     qdict = qemu_opts_to_qdict_filtered(opts, NULL, &qed_create_opts, true);
760 
761     if (!qdict_rename_keys(qdict, opt_renames, errp)) {
762         ret = -EINVAL;
763         goto fail;
764     }
765 
766     /* Create and open the file (protocol layer) */
767     ret = bdrv_create_file(filename, opts, errp);
768     if (ret < 0) {
769         goto fail;
770     }
771 
772     bs = bdrv_open(filename, NULL, NULL,
773                    BDRV_O_RDWR | BDRV_O_RESIZE | BDRV_O_PROTOCOL, errp);
774     if (bs == NULL) {
775         ret = -EIO;
776         goto fail;
777     }
778 
779     /* Now get the QAPI type BlockdevCreateOptions */
780     qdict_put_str(qdict, "driver", "qed");
781     qdict_put_str(qdict, "file", bs->node_name);
782 
783     v = qobject_input_visitor_new_flat_confused(qdict, errp);
784     if (!v) {
785         ret = -EINVAL;
786         goto fail;
787     }
788 
789     visit_type_BlockdevCreateOptions(v, NULL, &create_options, errp);
790     visit_free(v);
791     if (!create_options) {
792         ret = -EINVAL;
793         goto fail;
794     }
795 
796     /* Silently round up size */
797     assert(create_options->driver == BLOCKDEV_DRIVER_QED);
798     create_options->u.qed.size =
799         ROUND_UP(create_options->u.qed.size, BDRV_SECTOR_SIZE);
800 
801     /* Create the qed image (format layer) */
802     ret = bdrv_qed_co_create(create_options, errp);
803 
804 fail:
805     qobject_unref(qdict);
806     bdrv_unref(bs);
807     qapi_free_BlockdevCreateOptions(create_options);
808     return ret;
809 }
810 
811 static int coroutine_fn bdrv_qed_co_block_status(BlockDriverState *bs,
812                                                  bool want_zero,
813                                                  int64_t pos, int64_t bytes,
814                                                  int64_t *pnum, int64_t *map,
815                                                  BlockDriverState **file)
816 {
817     BDRVQEDState *s = bs->opaque;
818     size_t len = MIN(bytes, SIZE_MAX);
819     int status;
820     QEDRequest request = { .l2_table = NULL };
821     uint64_t offset;
822     int ret;
823 
824     qemu_co_mutex_lock(&s->table_lock);
825     ret = qed_find_cluster(s, &request, pos, &len, &offset);
826 
827     *pnum = len;
828     switch (ret) {
829     case QED_CLUSTER_FOUND:
830         *map = offset | qed_offset_into_cluster(s, pos);
831         status = BDRV_BLOCK_DATA | BDRV_BLOCK_OFFSET_VALID;
832         *file = bs->file->bs;
833         break;
834     case QED_CLUSTER_ZERO:
835         status = BDRV_BLOCK_ZERO;
836         break;
837     case QED_CLUSTER_L2:
838     case QED_CLUSTER_L1:
839         status = 0;
840         break;
841     default:
842         assert(ret < 0);
843         status = ret;
844         break;
845     }
846 
847     qed_unref_l2_cache_entry(request.l2_table);
848     qemu_co_mutex_unlock(&s->table_lock);
849 
850     return status;
851 }
852 
853 static BDRVQEDState *acb_to_s(QEDAIOCB *acb)
854 {
855     return acb->bs->opaque;
856 }
857 
858 /**
859  * Read from the backing file or zero-fill if no backing file
860  *
861  * @s:              QED state
862  * @pos:            Byte position in device
863  * @qiov:           Destination I/O vector
864  *
865  * This function reads qiov->size bytes starting at pos from the backing file.
866  * If there is no backing file then zeroes are read.
867  */
868 static int coroutine_fn qed_read_backing_file(BDRVQEDState *s, uint64_t pos,
869                                               QEMUIOVector *qiov)
870 {
871     if (s->bs->backing) {
872         BLKDBG_EVENT(s->bs->file, BLKDBG_READ_BACKING_AIO);
873         return bdrv_co_preadv(s->bs->backing, pos, qiov->size, qiov, 0);
874     }
875     qemu_iovec_memset(qiov, 0, 0, qiov->size);
876     return 0;
877 }
878 
879 /**
880  * Copy data from backing file into the image
881  *
882  * @s:          QED state
883  * @pos:        Byte position in device
884  * @len:        Number of bytes
885  * @offset:     Byte offset in image file
886  */
887 static int coroutine_fn qed_copy_from_backing_file(BDRVQEDState *s,
888                                                    uint64_t pos, uint64_t len,
889                                                    uint64_t offset)
890 {
891     QEMUIOVector qiov;
892     int ret;
893 
894     /* Skip copy entirely if there is no work to do */
895     if (len == 0) {
896         return 0;
897     }
898 
899     qemu_iovec_init_buf(&qiov, qemu_blockalign(s->bs, len), len);
900 
901     ret = qed_read_backing_file(s, pos, &qiov);
902 
903     if (ret) {
904         goto out;
905     }
906 
907     BLKDBG_EVENT(s->bs->file, BLKDBG_COW_WRITE);
908     ret = bdrv_co_pwritev(s->bs->file, offset, qiov.size, &qiov, 0);
909     if (ret < 0) {
910         goto out;
911     }
912     ret = 0;
913 out:
914     qemu_vfree(qemu_iovec_buf(&qiov));
915     return ret;
916 }
917 
918 /**
919  * Link one or more contiguous clusters into a table
920  *
921  * @s:              QED state
922  * @table:          L2 table
923  * @index:          First cluster index
924  * @n:              Number of contiguous clusters
925  * @cluster:        First cluster offset
926  *
927  * The cluster offset may be an allocated byte offset in the image file, the
928  * zero cluster marker, or the unallocated cluster marker.
929  *
930  * Called with table_lock held.
931  */
932 static void coroutine_fn qed_update_l2_table(BDRVQEDState *s, QEDTable *table,
933                                              int index, unsigned int n,
934                                              uint64_t cluster)
935 {
936     int i;
937     for (i = index; i < index + n; i++) {
938         table->offsets[i] = cluster;
939         if (!qed_offset_is_unalloc_cluster(cluster) &&
940             !qed_offset_is_zero_cluster(cluster)) {
941             cluster += s->header.cluster_size;
942         }
943     }
944 }
945 
946 /* Called with table_lock held.  */
947 static void coroutine_fn qed_aio_complete(QEDAIOCB *acb)
948 {
949     BDRVQEDState *s = acb_to_s(acb);
950 
951     /* Free resources */
952     qemu_iovec_destroy(&acb->cur_qiov);
953     qed_unref_l2_cache_entry(acb->request.l2_table);
954 
955     /* Free the buffer we may have allocated for zero writes */
956     if (acb->flags & QED_AIOCB_ZERO) {
957         qemu_vfree(acb->qiov->iov[0].iov_base);
958         acb->qiov->iov[0].iov_base = NULL;
959     }
960 
961     /* Start next allocating write request waiting behind this one.  Note that
962      * requests enqueue themselves when they first hit an unallocated cluster
963      * but they wait until the entire request is finished before waking up the
964      * next request in the queue.  This ensures that we don't cycle through
965      * requests multiple times but rather finish one at a time completely.
966      */
967     if (acb == s->allocating_acb) {
968         s->allocating_acb = NULL;
969         if (!qemu_co_queue_empty(&s->allocating_write_reqs)) {
970             qemu_co_queue_next(&s->allocating_write_reqs);
971         } else if (s->header.features & QED_F_NEED_CHECK) {
972             qed_start_need_check_timer(s);
973         }
974     }
975 }
976 
977 /**
978  * Update L1 table with new L2 table offset and write it out
979  *
980  * Called with table_lock held.
981  */
982 static int coroutine_fn qed_aio_write_l1_update(QEDAIOCB *acb)
983 {
984     BDRVQEDState *s = acb_to_s(acb);
985     CachedL2Table *l2_table = acb->request.l2_table;
986     uint64_t l2_offset = l2_table->offset;
987     int index, ret;
988 
989     index = qed_l1_index(s, acb->cur_pos);
990     s->l1_table->offsets[index] = l2_table->offset;
991 
992     ret = qed_write_l1_table(s, index, 1);
993 
994     /* Commit the current L2 table to the cache */
995     qed_commit_l2_cache_entry(&s->l2_cache, l2_table);
996 
997     /* This is guaranteed to succeed because we just committed the entry to the
998      * cache.
999      */
1000     acb->request.l2_table = qed_find_l2_cache_entry(&s->l2_cache, l2_offset);
1001     assert(acb->request.l2_table != NULL);
1002 
1003     return ret;
1004 }
1005 
1006 
1007 /**
1008  * Update L2 table with new cluster offsets and write them out
1009  *
1010  * Called with table_lock held.
1011  */
1012 static int coroutine_fn qed_aio_write_l2_update(QEDAIOCB *acb, uint64_t offset)
1013 {
1014     BDRVQEDState *s = acb_to_s(acb);
1015     bool need_alloc = acb->find_cluster_ret == QED_CLUSTER_L1;
1016     int index, ret;
1017 
1018     if (need_alloc) {
1019         qed_unref_l2_cache_entry(acb->request.l2_table);
1020         acb->request.l2_table = qed_new_l2_table(s);
1021     }
1022 
1023     index = qed_l2_index(s, acb->cur_pos);
1024     qed_update_l2_table(s, acb->request.l2_table->table, index, acb->cur_nclusters,
1025                          offset);
1026 
1027     if (need_alloc) {
1028         /* Write out the whole new L2 table */
1029         ret = qed_write_l2_table(s, &acb->request, 0, s->table_nelems, true);
1030         if (ret) {
1031             return ret;
1032         }
1033         return qed_aio_write_l1_update(acb);
1034     } else {
1035         /* Write out only the updated part of the L2 table */
1036         ret = qed_write_l2_table(s, &acb->request, index, acb->cur_nclusters,
1037                                  false);
1038         if (ret) {
1039             return ret;
1040         }
1041     }
1042     return 0;
1043 }
1044 
1045 /**
1046  * Write data to the image file
1047  *
1048  * Called with table_lock *not* held.
1049  */
1050 static int coroutine_fn qed_aio_write_main(QEDAIOCB *acb)
1051 {
1052     BDRVQEDState *s = acb_to_s(acb);
1053     uint64_t offset = acb->cur_cluster +
1054                       qed_offset_into_cluster(s, acb->cur_pos);
1055 
1056     trace_qed_aio_write_main(s, acb, 0, offset, acb->cur_qiov.size);
1057 
1058     BLKDBG_EVENT(s->bs->file, BLKDBG_WRITE_AIO);
1059     return bdrv_co_pwritev(s->bs->file, offset, acb->cur_qiov.size,
1060                            &acb->cur_qiov, 0);
1061 }
1062 
1063 /**
1064  * Populate untouched regions of new data cluster
1065  *
1066  * Called with table_lock held.
1067  */
1068 static int coroutine_fn qed_aio_write_cow(QEDAIOCB *acb)
1069 {
1070     BDRVQEDState *s = acb_to_s(acb);
1071     uint64_t start, len, offset;
1072     int ret;
1073 
1074     qemu_co_mutex_unlock(&s->table_lock);
1075 
1076     /* Populate front untouched region of new data cluster */
1077     start = qed_start_of_cluster(s, acb->cur_pos);
1078     len = qed_offset_into_cluster(s, acb->cur_pos);
1079 
1080     trace_qed_aio_write_prefill(s, acb, start, len, acb->cur_cluster);
1081     ret = qed_copy_from_backing_file(s, start, len, acb->cur_cluster);
1082     if (ret < 0) {
1083         goto out;
1084     }
1085 
1086     /* Populate back untouched region of new data cluster */
1087     start = acb->cur_pos + acb->cur_qiov.size;
1088     len = qed_start_of_cluster(s, start + s->header.cluster_size - 1) - start;
1089     offset = acb->cur_cluster +
1090              qed_offset_into_cluster(s, acb->cur_pos) +
1091              acb->cur_qiov.size;
1092 
1093     trace_qed_aio_write_postfill(s, acb, start, len, offset);
1094     ret = qed_copy_from_backing_file(s, start, len, offset);
1095     if (ret < 0) {
1096         goto out;
1097     }
1098 
1099     ret = qed_aio_write_main(acb);
1100     if (ret < 0) {
1101         goto out;
1102     }
1103 
1104     if (s->bs->backing) {
1105         /*
1106          * Flush new data clusters before updating the L2 table
1107          *
1108          * This flush is necessary when a backing file is in use.  A crash
1109          * during an allocating write could result in empty clusters in the
1110          * image.  If the write only touched a subregion of the cluster,
1111          * then backing image sectors have been lost in the untouched
1112          * region.  The solution is to flush after writing a new data
1113          * cluster and before updating the L2 table.
1114          */
1115         ret = bdrv_co_flush(s->bs->file->bs);
1116     }
1117 
1118 out:
1119     qemu_co_mutex_lock(&s->table_lock);
1120     return ret;
1121 }
1122 
1123 /**
1124  * Check if the QED_F_NEED_CHECK bit should be set during allocating write
1125  */
1126 static bool qed_should_set_need_check(BDRVQEDState *s)
1127 {
1128     /* The flush before L2 update path ensures consistency */
1129     if (s->bs->backing) {
1130         return false;
1131     }
1132 
1133     return !(s->header.features & QED_F_NEED_CHECK);
1134 }
1135 
1136 /**
1137  * Write new data cluster
1138  *
1139  * @acb:        Write request
1140  * @len:        Length in bytes
1141  *
1142  * This path is taken when writing to previously unallocated clusters.
1143  *
1144  * Called with table_lock held.
1145  */
1146 static int coroutine_fn qed_aio_write_alloc(QEDAIOCB *acb, size_t len)
1147 {
1148     BDRVQEDState *s = acb_to_s(acb);
1149     int ret;
1150 
1151     /* Cancel timer when the first allocating request comes in */
1152     if (s->allocating_acb == NULL) {
1153         qed_cancel_need_check_timer(s);
1154     }
1155 
1156     /* Freeze this request if another allocating write is in progress */
1157     if (s->allocating_acb != acb || s->allocating_write_reqs_plugged) {
1158         if (s->allocating_acb != NULL) {
1159             qemu_co_queue_wait(&s->allocating_write_reqs, &s->table_lock);
1160             assert(s->allocating_acb == NULL);
1161         }
1162         s->allocating_acb = acb;
1163         return -EAGAIN; /* start over with looking up table entries */
1164     }
1165 
1166     acb->cur_nclusters = qed_bytes_to_clusters(s,
1167             qed_offset_into_cluster(s, acb->cur_pos) + len);
1168     qemu_iovec_concat(&acb->cur_qiov, acb->qiov, acb->qiov_offset, len);
1169 
1170     if (acb->flags & QED_AIOCB_ZERO) {
1171         /* Skip ahead if the clusters are already zero */
1172         if (acb->find_cluster_ret == QED_CLUSTER_ZERO) {
1173             return 0;
1174         }
1175         acb->cur_cluster = 1;
1176     } else {
1177         acb->cur_cluster = qed_alloc_clusters(s, acb->cur_nclusters);
1178     }
1179 
1180     if (qed_should_set_need_check(s)) {
1181         s->header.features |= QED_F_NEED_CHECK;
1182         ret = qed_write_header(s);
1183         if (ret < 0) {
1184             return ret;
1185         }
1186     }
1187 
1188     if (!(acb->flags & QED_AIOCB_ZERO)) {
1189         ret = qed_aio_write_cow(acb);
1190         if (ret < 0) {
1191             return ret;
1192         }
1193     }
1194 
1195     return qed_aio_write_l2_update(acb, acb->cur_cluster);
1196 }
1197 
1198 /**
1199  * Write data cluster in place
1200  *
1201  * @acb:        Write request
1202  * @offset:     Cluster offset in bytes
1203  * @len:        Length in bytes
1204  *
1205  * This path is taken when writing to already allocated clusters.
1206  *
1207  * Called with table_lock held.
1208  */
1209 static int coroutine_fn qed_aio_write_inplace(QEDAIOCB *acb, uint64_t offset,
1210                                               size_t len)
1211 {
1212     BDRVQEDState *s = acb_to_s(acb);
1213     int r;
1214 
1215     qemu_co_mutex_unlock(&s->table_lock);
1216 
1217     /* Allocate buffer for zero writes */
1218     if (acb->flags & QED_AIOCB_ZERO) {
1219         struct iovec *iov = acb->qiov->iov;
1220 
1221         if (!iov->iov_base) {
1222             iov->iov_base = qemu_try_blockalign(acb->bs, iov->iov_len);
1223             if (iov->iov_base == NULL) {
1224                 r = -ENOMEM;
1225                 goto out;
1226             }
1227             memset(iov->iov_base, 0, iov->iov_len);
1228         }
1229     }
1230 
1231     /* Calculate the I/O vector */
1232     acb->cur_cluster = offset;
1233     qemu_iovec_concat(&acb->cur_qiov, acb->qiov, acb->qiov_offset, len);
1234 
1235     /* Do the actual write.  */
1236     r = qed_aio_write_main(acb);
1237 out:
1238     qemu_co_mutex_lock(&s->table_lock);
1239     return r;
1240 }
1241 
1242 /**
1243  * Write data cluster
1244  *
1245  * @opaque:     Write request
1246  * @ret:        QED_CLUSTER_FOUND, QED_CLUSTER_L2 or QED_CLUSTER_L1
1247  * @offset:     Cluster offset in bytes
1248  * @len:        Length in bytes
1249  *
1250  * Called with table_lock held.
1251  */
1252 static int coroutine_fn qed_aio_write_data(void *opaque, int ret,
1253                                            uint64_t offset, size_t len)
1254 {
1255     QEDAIOCB *acb = opaque;
1256 
1257     trace_qed_aio_write_data(acb_to_s(acb), acb, ret, offset, len);
1258 
1259     acb->find_cluster_ret = ret;
1260 
1261     switch (ret) {
1262     case QED_CLUSTER_FOUND:
1263         return qed_aio_write_inplace(acb, offset, len);
1264 
1265     case QED_CLUSTER_L2:
1266     case QED_CLUSTER_L1:
1267     case QED_CLUSTER_ZERO:
1268         return qed_aio_write_alloc(acb, len);
1269 
1270     default:
1271         g_assert_not_reached();
1272     }
1273 }
1274 
1275 /**
1276  * Read data cluster
1277  *
1278  * @opaque:     Read request
1279  * @ret:        QED_CLUSTER_FOUND, QED_CLUSTER_L2 or QED_CLUSTER_L1
1280  * @offset:     Cluster offset in bytes
1281  * @len:        Length in bytes
1282  *
1283  * Called with table_lock held.
1284  */
1285 static int coroutine_fn qed_aio_read_data(void *opaque, int ret,
1286                                           uint64_t offset, size_t len)
1287 {
1288     QEDAIOCB *acb = opaque;
1289     BDRVQEDState *s = acb_to_s(acb);
1290     BlockDriverState *bs = acb->bs;
1291     int r;
1292 
1293     qemu_co_mutex_unlock(&s->table_lock);
1294 
1295     /* Adjust offset into cluster */
1296     offset += qed_offset_into_cluster(s, acb->cur_pos);
1297 
1298     trace_qed_aio_read_data(s, acb, ret, offset, len);
1299 
1300     qemu_iovec_concat(&acb->cur_qiov, acb->qiov, acb->qiov_offset, len);
1301 
1302     /* Handle zero cluster and backing file reads, otherwise read
1303      * data cluster directly.
1304      */
1305     if (ret == QED_CLUSTER_ZERO) {
1306         qemu_iovec_memset(&acb->cur_qiov, 0, 0, acb->cur_qiov.size);
1307         r = 0;
1308     } else if (ret != QED_CLUSTER_FOUND) {
1309         r = qed_read_backing_file(s, acb->cur_pos, &acb->cur_qiov);
1310     } else {
1311         BLKDBG_EVENT(bs->file, BLKDBG_READ_AIO);
1312         r = bdrv_co_preadv(bs->file, offset, acb->cur_qiov.size,
1313                            &acb->cur_qiov, 0);
1314     }
1315 
1316     qemu_co_mutex_lock(&s->table_lock);
1317     return r;
1318 }
1319 
1320 /**
1321  * Begin next I/O or complete the request
1322  */
1323 static int coroutine_fn qed_aio_next_io(QEDAIOCB *acb)
1324 {
1325     BDRVQEDState *s = acb_to_s(acb);
1326     uint64_t offset;
1327     size_t len;
1328     int ret;
1329 
1330     qemu_co_mutex_lock(&s->table_lock);
1331     while (1) {
1332         trace_qed_aio_next_io(s, acb, 0, acb->cur_pos + acb->cur_qiov.size);
1333 
1334         acb->qiov_offset += acb->cur_qiov.size;
1335         acb->cur_pos += acb->cur_qiov.size;
1336         qemu_iovec_reset(&acb->cur_qiov);
1337 
1338         /* Complete request */
1339         if (acb->cur_pos >= acb->end_pos) {
1340             ret = 0;
1341             break;
1342         }
1343 
1344         /* Find next cluster and start I/O */
1345         len = acb->end_pos - acb->cur_pos;
1346         ret = qed_find_cluster(s, &acb->request, acb->cur_pos, &len, &offset);
1347         if (ret < 0) {
1348             break;
1349         }
1350 
1351         if (acb->flags & QED_AIOCB_WRITE) {
1352             ret = qed_aio_write_data(acb, ret, offset, len);
1353         } else {
1354             ret = qed_aio_read_data(acb, ret, offset, len);
1355         }
1356 
1357         if (ret < 0 && ret != -EAGAIN) {
1358             break;
1359         }
1360     }
1361 
1362     trace_qed_aio_complete(s, acb, ret);
1363     qed_aio_complete(acb);
1364     qemu_co_mutex_unlock(&s->table_lock);
1365     return ret;
1366 }
1367 
1368 static int coroutine_fn qed_co_request(BlockDriverState *bs, int64_t sector_num,
1369                                        QEMUIOVector *qiov, int nb_sectors,
1370                                        int flags)
1371 {
1372     QEDAIOCB acb = {
1373         .bs         = bs,
1374         .cur_pos    = (uint64_t) sector_num * BDRV_SECTOR_SIZE,
1375         .end_pos    = (sector_num + nb_sectors) * BDRV_SECTOR_SIZE,
1376         .qiov       = qiov,
1377         .flags      = flags,
1378     };
1379     qemu_iovec_init(&acb.cur_qiov, qiov->niov);
1380 
1381     trace_qed_aio_setup(bs->opaque, &acb, sector_num, nb_sectors, NULL, flags);
1382 
1383     /* Start request */
1384     return qed_aio_next_io(&acb);
1385 }
1386 
1387 static int coroutine_fn bdrv_qed_co_readv(BlockDriverState *bs,
1388                                           int64_t sector_num, int nb_sectors,
1389                                           QEMUIOVector *qiov)
1390 {
1391     return qed_co_request(bs, sector_num, qiov, nb_sectors, 0);
1392 }
1393 
1394 static int coroutine_fn bdrv_qed_co_writev(BlockDriverState *bs,
1395                                            int64_t sector_num, int nb_sectors,
1396                                            QEMUIOVector *qiov, int flags)
1397 {
1398     return qed_co_request(bs, sector_num, qiov, nb_sectors, QED_AIOCB_WRITE);
1399 }
1400 
1401 static int coroutine_fn bdrv_qed_co_pwrite_zeroes(BlockDriverState *bs,
1402                                                   int64_t offset,
1403                                                   int64_t bytes,
1404                                                   BdrvRequestFlags flags)
1405 {
1406     BDRVQEDState *s = bs->opaque;
1407 
1408     /*
1409      * Zero writes start without an I/O buffer.  If a buffer becomes necessary
1410      * then it will be allocated during request processing.
1411      */
1412     QEMUIOVector qiov = QEMU_IOVEC_INIT_BUF(qiov, NULL, bytes);
1413 
1414     /*
1415      * QED is not prepared for 63bit write-zero requests, so rely on
1416      * max_pwrite_zeroes.
1417      */
1418     assert(bytes <= INT_MAX);
1419 
1420     /* Fall back if the request is not aligned */
1421     if (qed_offset_into_cluster(s, offset) ||
1422         qed_offset_into_cluster(s, bytes)) {
1423         return -ENOTSUP;
1424     }
1425 
1426     return qed_co_request(bs, offset >> BDRV_SECTOR_BITS, &qiov,
1427                           bytes >> BDRV_SECTOR_BITS,
1428                           QED_AIOCB_WRITE | QED_AIOCB_ZERO);
1429 }
1430 
1431 static int coroutine_fn bdrv_qed_co_truncate(BlockDriverState *bs,
1432                                              int64_t offset,
1433                                              bool exact,
1434                                              PreallocMode prealloc,
1435                                              BdrvRequestFlags flags,
1436                                              Error **errp)
1437 {
1438     BDRVQEDState *s = bs->opaque;
1439     uint64_t old_image_size;
1440     int ret;
1441 
1442     if (prealloc != PREALLOC_MODE_OFF) {
1443         error_setg(errp, "Unsupported preallocation mode '%s'",
1444                    PreallocMode_str(prealloc));
1445         return -ENOTSUP;
1446     }
1447 
1448     if (!qed_is_image_size_valid(offset, s->header.cluster_size,
1449                                  s->header.table_size)) {
1450         error_setg(errp, "Invalid image size specified");
1451         return -EINVAL;
1452     }
1453 
1454     if ((uint64_t)offset < s->header.image_size) {
1455         error_setg(errp, "Shrinking images is currently not supported");
1456         return -ENOTSUP;
1457     }
1458 
1459     old_image_size = s->header.image_size;
1460     s->header.image_size = offset;
1461     ret = qed_write_header_sync(s);
1462     if (ret < 0) {
1463         s->header.image_size = old_image_size;
1464         error_setg_errno(errp, -ret, "Failed to update the image size");
1465     }
1466     return ret;
1467 }
1468 
1469 static int64_t bdrv_qed_getlength(BlockDriverState *bs)
1470 {
1471     BDRVQEDState *s = bs->opaque;
1472     return s->header.image_size;
1473 }
1474 
1475 static int bdrv_qed_get_info(BlockDriverState *bs, BlockDriverInfo *bdi)
1476 {
1477     BDRVQEDState *s = bs->opaque;
1478 
1479     memset(bdi, 0, sizeof(*bdi));
1480     bdi->cluster_size = s->header.cluster_size;
1481     bdi->is_dirty = s->header.features & QED_F_NEED_CHECK;
1482     return 0;
1483 }
1484 
1485 static int bdrv_qed_change_backing_file(BlockDriverState *bs,
1486                                         const char *backing_file,
1487                                         const char *backing_fmt)
1488 {
1489     BDRVQEDState *s = bs->opaque;
1490     QEDHeader new_header, le_header;
1491     void *buffer;
1492     size_t buffer_len, backing_file_len;
1493     int ret;
1494 
1495     /* Refuse to set backing filename if unknown compat feature bits are
1496      * active.  If the image uses an unknown compat feature then we may not
1497      * know the layout of data following the header structure and cannot safely
1498      * add a new string.
1499      */
1500     if (backing_file && (s->header.compat_features &
1501                          ~QED_COMPAT_FEATURE_MASK)) {
1502         return -ENOTSUP;
1503     }
1504 
1505     memcpy(&new_header, &s->header, sizeof(new_header));
1506 
1507     new_header.features &= ~(QED_F_BACKING_FILE |
1508                              QED_F_BACKING_FORMAT_NO_PROBE);
1509 
1510     /* Adjust feature flags */
1511     if (backing_file) {
1512         new_header.features |= QED_F_BACKING_FILE;
1513 
1514         if (qed_fmt_is_raw(backing_fmt)) {
1515             new_header.features |= QED_F_BACKING_FORMAT_NO_PROBE;
1516         }
1517     }
1518 
1519     /* Calculate new header size */
1520     backing_file_len = 0;
1521 
1522     if (backing_file) {
1523         backing_file_len = strlen(backing_file);
1524     }
1525 
1526     buffer_len = sizeof(new_header);
1527     new_header.backing_filename_offset = buffer_len;
1528     new_header.backing_filename_size = backing_file_len;
1529     buffer_len += backing_file_len;
1530 
1531     /* Make sure we can rewrite header without failing */
1532     if (buffer_len > new_header.header_size * new_header.cluster_size) {
1533         return -ENOSPC;
1534     }
1535 
1536     /* Prepare new header */
1537     buffer = g_malloc(buffer_len);
1538 
1539     qed_header_cpu_to_le(&new_header, &le_header);
1540     memcpy(buffer, &le_header, sizeof(le_header));
1541     buffer_len = sizeof(le_header);
1542 
1543     if (backing_file) {
1544         memcpy(buffer + buffer_len, backing_file, backing_file_len);
1545         buffer_len += backing_file_len;
1546     }
1547 
1548     /* Write new header */
1549     ret = bdrv_pwrite_sync(bs->file, 0, buffer_len, buffer, 0);
1550     g_free(buffer);
1551     if (ret == 0) {
1552         memcpy(&s->header, &new_header, sizeof(new_header));
1553     }
1554     return ret;
1555 }
1556 
1557 static void coroutine_fn bdrv_qed_co_invalidate_cache(BlockDriverState *bs,
1558                                                       Error **errp)
1559 {
1560     BDRVQEDState *s = bs->opaque;
1561     int ret;
1562 
1563     bdrv_qed_close(bs);
1564 
1565     bdrv_qed_init_state(bs);
1566     qemu_co_mutex_lock(&s->table_lock);
1567     ret = bdrv_qed_do_open(bs, NULL, bs->open_flags, errp);
1568     qemu_co_mutex_unlock(&s->table_lock);
1569     if (ret < 0) {
1570         error_prepend(errp, "Could not reopen qed layer: ");
1571     }
1572 }
1573 
1574 static int coroutine_fn bdrv_qed_co_check(BlockDriverState *bs,
1575                                           BdrvCheckResult *result,
1576                                           BdrvCheckMode fix)
1577 {
1578     BDRVQEDState *s = bs->opaque;
1579     int ret;
1580 
1581     qemu_co_mutex_lock(&s->table_lock);
1582     ret = qed_check(s, result, !!fix);
1583     qemu_co_mutex_unlock(&s->table_lock);
1584 
1585     return ret;
1586 }
1587 
1588 static QemuOptsList qed_create_opts = {
1589     .name = "qed-create-opts",
1590     .head = QTAILQ_HEAD_INITIALIZER(qed_create_opts.head),
1591     .desc = {
1592         {
1593             .name = BLOCK_OPT_SIZE,
1594             .type = QEMU_OPT_SIZE,
1595             .help = "Virtual disk size"
1596         },
1597         {
1598             .name = BLOCK_OPT_BACKING_FILE,
1599             .type = QEMU_OPT_STRING,
1600             .help = "File name of a base image"
1601         },
1602         {
1603             .name = BLOCK_OPT_BACKING_FMT,
1604             .type = QEMU_OPT_STRING,
1605             .help = "Image format of the base image"
1606         },
1607         {
1608             .name = BLOCK_OPT_CLUSTER_SIZE,
1609             .type = QEMU_OPT_SIZE,
1610             .help = "Cluster size (in bytes)",
1611             .def_value_str = stringify(QED_DEFAULT_CLUSTER_SIZE)
1612         },
1613         {
1614             .name = BLOCK_OPT_TABLE_SIZE,
1615             .type = QEMU_OPT_SIZE,
1616             .help = "L1/L2 table size (in clusters)"
1617         },
1618         { /* end of list */ }
1619     }
1620 };
1621 
1622 static BlockDriver bdrv_qed = {
1623     .format_name              = "qed",
1624     .instance_size            = sizeof(BDRVQEDState),
1625     .create_opts              = &qed_create_opts,
1626     .is_format                = true,
1627     .supports_backing         = true,
1628 
1629     .bdrv_probe               = bdrv_qed_probe,
1630     .bdrv_open                = bdrv_qed_open,
1631     .bdrv_close               = bdrv_qed_close,
1632     .bdrv_reopen_prepare      = bdrv_qed_reopen_prepare,
1633     .bdrv_child_perm          = bdrv_default_perms,
1634     .bdrv_co_create           = bdrv_qed_co_create,
1635     .bdrv_co_create_opts      = bdrv_qed_co_create_opts,
1636     .bdrv_has_zero_init       = bdrv_has_zero_init_1,
1637     .bdrv_co_block_status     = bdrv_qed_co_block_status,
1638     .bdrv_co_readv            = bdrv_qed_co_readv,
1639     .bdrv_co_writev           = bdrv_qed_co_writev,
1640     .bdrv_co_pwrite_zeroes    = bdrv_qed_co_pwrite_zeroes,
1641     .bdrv_co_truncate         = bdrv_qed_co_truncate,
1642     .bdrv_getlength           = bdrv_qed_getlength,
1643     .bdrv_get_info            = bdrv_qed_get_info,
1644     .bdrv_refresh_limits      = bdrv_qed_refresh_limits,
1645     .bdrv_change_backing_file = bdrv_qed_change_backing_file,
1646     .bdrv_co_invalidate_cache = bdrv_qed_co_invalidate_cache,
1647     .bdrv_co_check            = bdrv_qed_co_check,
1648     .bdrv_detach_aio_context  = bdrv_qed_detach_aio_context,
1649     .bdrv_attach_aio_context  = bdrv_qed_attach_aio_context,
1650     .bdrv_co_drain_begin      = bdrv_qed_co_drain_begin,
1651 };
1652 
1653 static void bdrv_qed_init(void)
1654 {
1655     bdrv_register(&bdrv_qed);
1656 }
1657 
1658 block_init(bdrv_qed_init);
1659