cachepc-qemu

Fork of AMDESE/qemu with changes for cachepc side-channel attack
git clone https://git.sinitax.com/sinitax/cachepc-qemu
Log | Files | Refs | Submodules | LICENSE | sfeed.txt

qed.c (49462B)


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