block/copy-on-read.c

   1 /*
   2  * Copy-on-read filter block driver
   3  *
   4  * Copyright (c) 2018 Red Hat, Inc.
   5  *
   6  * Author:
   7  *   Max Reitz <[email protected]>
   8  *
   9  * This program is free software; you can redistribute it and/or
  10  * modify it under the terms of the GNU General Public License as
  11  * published by the Free Software Foundation; either version 2 or
  12  * (at your option) version 3 of the License.
  13  *
  14  * This program is distributed in the hope that it will be useful,
  15  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  16  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  17  * GNU General Public License for more details.
  18  *
  19  * You should have received a copy of the GNU General Public License
  20  * along with this program; if not, see <http://www.gnu.org/licenses/>.
  21  */
  22
  23 #include "qemu/osdep.h"
  24 #include "block/block_int.h"
  25 #include "qemu/module.h"
  26 #include "qapi/error.h"
  27 #include "qapi/qmp/qdict.h"
  28 #include "block/copy-on-read.h"
  29
  30
  31 typedef struct BDRVStateCOR {
  32     bool active;
  33     BlockDriverState *bottom_bs;
  34     bool chain_frozen;
  35 } BDRVStateCOR;
  36
  37
  38 static int cor_open(BlockDriverState *bs, QDict *options, int flags,
  39                     Error **errp)
  40 {
  41     BlockDriverState *bottom_bs = NULL;
  42     BDRVStateCOR *state = bs->opaque;
  43     /* Find a bottom node name, if any */
  44     const char *bottom_node = qdict_get_try_str(options, "bottom");
  45
  46     bs->file = bdrv_open_child(NULL, options, "file", bs, &child_of_bds,
  47                                BDRV_CHILD_FILTERED | BDRV_CHILD_PRIMARY,
  48                                false, errp);
  49     if (!bs->file) {
  50         return -EINVAL;
  51     }
  52
  53     bs->supported_write_flags = BDRV_REQ_WRITE_UNCHANGED |
  54         (BDRV_REQ_FUA & bs->file->bs->supported_write_flags);
  55
  56     bs->supported_zero_flags = BDRV_REQ_WRITE_UNCHANGED |
  57         ((BDRV_REQ_FUA | BDRV_REQ_MAY_UNMAP | BDRV_REQ_NO_FALLBACK) &
  58             bs->file->bs->supported_zero_flags);
  59
  60     if (bottom_node) {
  61         bottom_bs = bdrv_find_node(bottom_node);
  62         if (!bottom_bs) {
  63             error_setg(errp, "Bottom node '%s' not found", bottom_node);
  64             qdict_del(options, "bottom");
  65             return -EINVAL;
  66         }
  67         qdict_del(options, "bottom");
  68
  69         if (!bottom_bs->drv) {
  70             error_setg(errp, "Bottom node '%s' not opened", bottom_node);
  71             return -EINVAL;
  72         }
  73
  74         if (bottom_bs->drv->is_filter) {
  75             error_setg(errp, "Bottom node '%s' is a filter", bottom_node);
  76             return -EINVAL;
  77         }
  78
  79         if (bdrv_freeze_backing_chain(bs, bottom_bs, errp) < 0) {
  80             return -EINVAL;
  81         }
  82         state->chain_frozen = true;
  83
  84         /*
  85          * We do freeze the chain, so it shouldn't be removed. Still, storing a
  86          * pointer worth bdrv_ref().
  87          */
  88         bdrv_ref(bottom_bs);
  89     }
  90     state->active = true;
  91     state->bottom_bs = bottom_bs;
  92
  93     /*
  94      * We don't need to call bdrv_child_refresh_perms() now as the permissions
  95      * will be updated later when the filter node gets its parent.
  96      */
  97
  98     return 0;
  99 }
 100
 101
 102 #define PERM_PASSTHROUGH (BLK_PERM_CONSISTENT_READ \
 103                           | BLK_PERM_WRITE \
 104                           | BLK_PERM_RESIZE)
 105 #define PERM_UNCHANGED (BLK_PERM_ALL & ~PERM_PASSTHROUGH)
 106
 107 static void cor_child_perm(BlockDriverState *bs, BdrvChild *c,
 108                            BdrvChildRole role,
 109                            BlockReopenQueue *reopen_queue,
 110                            uint64_t perm, uint64_t shared,
 111                            uint64_t *nperm, uint64_t *nshared)
 112 {
 113     BDRVStateCOR *s = bs->opaque;
 114
 115     if (!s->active) {
 116         /*
 117          * While the filter is being removed
 118          */
 119         *nperm = 0;
 120         *nshared = BLK_PERM_ALL;
 121         return;
 122     }
 123
 124     *nperm = perm & PERM_PASSTHROUGH;
 125     *nshared = (shared & PERM_PASSTHROUGH) | PERM_UNCHANGED;
 126
 127     /* We must not request write permissions for an inactive node, the child
 128      * cannot provide it. */
 129     if (!(bs->open_flags & BDRV_O_INACTIVE)) {
 130         *nperm |= BLK_PERM_WRITE_UNCHANGED;
 131     }
 132 }
 133
 134
 135 static int64_t cor_getlength(BlockDriverState *bs)
 136 {
 137     return bdrv_getlength(bs->file->bs);
 138 }
 139
 140
 141 static int coroutine_fn cor_co_preadv_part(BlockDriverState *bs,
 142                                            uint64_t offset, uint64_t bytes,
 143                                            QEMUIOVector *qiov,
 144                                            size_t qiov_offset,
 145                                            int flags)
 146 {
 147     int64_t n;
 148     int local_flags;
 149     int ret;
 150     BDRVStateCOR *state = bs->opaque;
 151
 152     if (!state->bottom_bs) {
 153         return bdrv_co_preadv_part(bs->file, offset, bytes, qiov, qiov_offset,
 154                                    flags | BDRV_REQ_COPY_ON_READ);
 155     }
 156
 157     while (bytes) {
 158         local_flags = flags;
 159
 160         /* In case of failure, try to copy-on-read anyway */
 161         ret = bdrv_is_allocated(bs->file->bs, offset, bytes, &n);
 162         if (ret <= 0) {
 163             ret = bdrv_is_allocated_above(bdrv_backing_chain_next(bs->file->bs),
 164                                           state->bottom_bs, true, offset,
 165                                           n, &n);
 166             if (ret > 0 || ret < 0) {
 167                 local_flags |= BDRV_REQ_COPY_ON_READ;
 168             }
 169             /* Finish earlier if the end of a backing file has been reached */
 170             if (n == 0) {
 171                 break;
 172             }
 173         }
 174
 175         ret = bdrv_co_preadv_part(bs->file, offset, n, qiov, qiov_offset,
 176                                   local_flags);
 177         if (ret < 0) {
 178             return ret;
 179         }
 180
 181         offset += n;
 182         qiov_offset += n;
 183         bytes -= n;
 184     }
 185
 186     return 0;
 187 }
 188
 189
 190 static int coroutine_fn cor_co_pwritev_part(BlockDriverState *bs,
 191                                             uint64_t offset,
 192                                             uint64_t bytes,
 193                                             QEMUIOVector *qiov,
 194                                             size_t qiov_offset, int flags)
 195 {
 196     return bdrv_co_pwritev_part(bs->file, offset, bytes, qiov, qiov_offset,
 197                                 flags);
 198 }
 199
 200
 201 static int coroutine_fn cor_co_pwrite_zeroes(BlockDriverState *bs,
 202                                              int64_t offset, int bytes,
 203                                              BdrvRequestFlags flags)
 204 {
 205     return bdrv_co_pwrite_zeroes(bs->file, offset, bytes, flags);
 206 }
 207
 208
 209 static int coroutine_fn cor_co_pdiscard(BlockDriverState *bs,
 210                                         int64_t offset, int bytes)
 211 {
 212     return bdrv_co_pdiscard(bs->file, offset, bytes);
 213 }
 214
 215
 216 static int coroutine_fn cor_co_pwritev_compressed(BlockDriverState *bs,
 217                                                   uint64_t offset,
 218                                                   uint64_t bytes,
 219                                                   QEMUIOVector *qiov)
 220 {
 221     return bdrv_co_pwritev(bs->file, offset, bytes, qiov,
 222                            BDRV_REQ_WRITE_COMPRESSED);
 223 }
 224
 225
 226 static void cor_eject(BlockDriverState *bs, bool eject_flag)
 227 {
 228     bdrv_eject(bs->file->bs, eject_flag);
 229 }
 230
 231
 232 static void cor_lock_medium(BlockDriverState *bs, bool locked)
 233 {
 234     bdrv_lock_medium(bs->file->bs, locked);
 235 }
 236
 237
 238 static void cor_close(BlockDriverState *bs)
 239 {
 240     BDRVStateCOR *s = bs->opaque;
 241
 242     if (s->chain_frozen) {
 243         s->chain_frozen = false;
 244         bdrv_unfreeze_backing_chain(bs, s->bottom_bs);
 245     }
 246
 247     bdrv_unref(s->bottom_bs);
 248 }
 249
 250
 251 static BlockDriver bdrv_copy_on_read = {
 252     .format_name                        = "copy-on-read",
 253     .instance_size                      = sizeof(BDRVStateCOR),
 254
 255     .bdrv_open                          = cor_open,
 256     .bdrv_close                         = cor_close,
 257     .bdrv_child_perm                    = cor_child_perm,
 258
 259     .bdrv_getlength                     = cor_getlength,
 260
 261     .bdrv_co_preadv_part                = cor_co_preadv_part,
 262     .bdrv_co_pwritev_part               = cor_co_pwritev_part,
 263     .bdrv_co_pwrite_zeroes              = cor_co_pwrite_zeroes,
 264     .bdrv_co_pdiscard                   = cor_co_pdiscard,
 265     .bdrv_co_pwritev_compressed         = cor_co_pwritev_compressed,
 266
 267     .bdrv_eject                         = cor_eject,
 268     .bdrv_lock_medium                   = cor_lock_medium,
 269
 270     .has_variable_length                = true,
 271     .is_filter                          = true,
 272 };
 273
 274
 275 void bdrv_cor_filter_drop(BlockDriverState *cor_filter_bs)
 276 {
 277     BdrvChild *child;
 278     BlockDriverState *bs;
 279     BDRVStateCOR *s = cor_filter_bs->opaque;
 280
 281     child = bdrv_filter_child(cor_filter_bs);
 282     if (!child) {
 283         return;
 284     }
 285     bs = child->bs;
 286
 287     /* Retain the BDS until we complete the graph change. */
 288     bdrv_ref(bs);
 289     /* Hold a guest back from writing while permissions are being reset. */
 290     bdrv_drained_begin(bs);
 291     /* Drop permissions before the graph change. */
 292     s->active = false;
 293     /* unfreeze, as otherwise bdrv_replace_node() will fail */
 294     if (s->chain_frozen) {
 295         s->chain_frozen = false;
 296         bdrv_unfreeze_backing_chain(cor_filter_bs, s->bottom_bs);
 297     }
 298     bdrv_child_refresh_perms(cor_filter_bs, child, &error_abort);
 299     bdrv_replace_node(cor_filter_bs, bs, &error_abort);
 300
 301     bdrv_drained_end(bs);
 302     bdrv_unref(bs);
 303     bdrv_unref(cor_filter_bs);
 304 }
 305
 306
 307 static void bdrv_copy_on_read_init(void)
 308 {
 309     bdrv_register(&bdrv_copy_on_read);
 310 }
 311
 312 block_init(bdrv_copy_on_read_init);