block: Framework for reopening files safely
authorJeff Cody <jcody@redhat.com>
Thu, 20 Sep 2012 19:13:19 +0000 (15:13 -0400)
committerKevin Wolf <kwolf@redhat.com>
Mon, 24 Sep 2012 13:15:11 +0000 (15:15 +0200)
This is based on Supriya Kannery's bdrv_reopen() patch series.

This provides a transactional method to reopen multiple
images files safely.

Image files are queue for reopen via bdrv_reopen_queue(), and the
reopen occurs when bdrv_reopen_multiple() is called.  Changes are
staged in bdrv_reopen_prepare() and in the equivalent driver level
functions.  If any of the staged images fails a prepare, then all
of the images left untouched, and the staged changes for each image
abandoned.

Block drivers are passed a reopen state structure, that contains:
    * BDS to reopen
    * flags for the reopen
    * opaque pointer for any driver-specific data that needs to be
      persistent from _prepare to _commit/_abort
    * reopen queue pointer, if the driver needs to queue additional
      BDS for a reopen

Signed-off-by: Jeff Cody <jcody@redhat.com>
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
block.c
block.h
block_int.h

diff --git a/block.c b/block.c
index 458bcc9b135d59d0272ab72d3919aab552afa0b9..c7c1a3bd331b508b0b3d0b0571de7ab9c0403dd3 100644 (file)
--- a/block.c
+++ b/block.c
@@ -863,6 +863,238 @@ unlink_and_fail:
     return ret;
 }
 
+typedef struct BlockReopenQueueEntry {
+     bool prepared;
+     BDRVReopenState state;
+     QSIMPLEQ_ENTRY(BlockReopenQueueEntry) entry;
+} BlockReopenQueueEntry;
+
+/*
+ * Adds a BlockDriverState to a simple queue for an atomic, transactional
+ * reopen of multiple devices.
+ *
+ * bs_queue can either be an existing BlockReopenQueue that has had QSIMPLE_INIT
+ * already performed, or alternatively may be NULL a new BlockReopenQueue will
+ * be created and initialized. This newly created BlockReopenQueue should be
+ * passed back in for subsequent calls that are intended to be of the same
+ * atomic 'set'.
+ *
+ * bs is the BlockDriverState to add to the reopen queue.
+ *
+ * flags contains the open flags for the associated bs
+ *
+ * returns a pointer to bs_queue, which is either the newly allocated
+ * bs_queue, or the existing bs_queue being used.
+ *
+ */
+BlockReopenQueue *bdrv_reopen_queue(BlockReopenQueue *bs_queue,
+                                    BlockDriverState *bs, int flags)
+{
+    assert(bs != NULL);
+
+    BlockReopenQueueEntry *bs_entry;
+    if (bs_queue == NULL) {
+        bs_queue = g_new0(BlockReopenQueue, 1);
+        QSIMPLEQ_INIT(bs_queue);
+    }
+
+    if (bs->file) {
+        bdrv_reopen_queue(bs_queue, bs->file, flags);
+    }
+
+    bs_entry = g_new0(BlockReopenQueueEntry, 1);
+    QSIMPLEQ_INSERT_TAIL(bs_queue, bs_entry, entry);
+
+    bs_entry->state.bs = bs;
+    bs_entry->state.flags = flags;
+
+    return bs_queue;
+}
+
+/*
+ * Reopen multiple BlockDriverStates atomically & transactionally.
+ *
+ * The queue passed in (bs_queue) must have been built up previous
+ * via bdrv_reopen_queue().
+ *
+ * Reopens all BDS specified in the queue, with the appropriate
+ * flags.  All devices are prepared for reopen, and failure of any
+ * device will cause all device changes to be abandonded, and intermediate
+ * data cleaned up.
+ *
+ * If all devices prepare successfully, then the changes are committed
+ * to all devices.
+ *
+ */
+int bdrv_reopen_multiple(BlockReopenQueue *bs_queue, Error **errp)
+{
+    int ret = -1;
+    BlockReopenQueueEntry *bs_entry, *next;
+    Error *local_err = NULL;
+
+    assert(bs_queue != NULL);
+
+    bdrv_drain_all();
+
+    QSIMPLEQ_FOREACH(bs_entry, bs_queue, entry) {
+        if (bdrv_reopen_prepare(&bs_entry->state, bs_queue, &local_err)) {
+            error_propagate(errp, local_err);
+            goto cleanup;
+        }
+        bs_entry->prepared = true;
+    }
+
+    /* If we reach this point, we have success and just need to apply the
+     * changes
+     */
+    QSIMPLEQ_FOREACH(bs_entry, bs_queue, entry) {
+        bdrv_reopen_commit(&bs_entry->state);
+    }
+
+    ret = 0;
+
+cleanup:
+    QSIMPLEQ_FOREACH_SAFE(bs_entry, bs_queue, entry, next) {
+        if (ret && bs_entry->prepared) {
+            bdrv_reopen_abort(&bs_entry->state);
+        }
+        g_free(bs_entry);
+    }
+    g_free(bs_queue);
+    return ret;
+}
+
+
+/* Reopen a single BlockDriverState with the specified flags. */
+int bdrv_reopen(BlockDriverState *bs, int bdrv_flags, Error **errp)
+{
+    int ret = -1;
+    Error *local_err = NULL;
+    BlockReopenQueue *queue = bdrv_reopen_queue(NULL, bs, bdrv_flags);
+
+    ret = bdrv_reopen_multiple(queue, &local_err);
+    if (local_err != NULL) {
+        error_propagate(errp, local_err);
+    }
+    return ret;
+}
+
+
+/*
+ * Prepares a BlockDriverState for reopen. All changes are staged in the
+ * 'opaque' field of the BDRVReopenState, which is used and allocated by
+ * the block driver layer .bdrv_reopen_prepare()
+ *
+ * bs is the BlockDriverState to reopen
+ * flags are the new open flags
+ * queue is the reopen queue
+ *
+ * Returns 0 on success, non-zero on error.  On error errp will be set
+ * as well.
+ *
+ * On failure, bdrv_reopen_abort() will be called to clean up any data.
+ * It is the responsibility of the caller to then call the abort() or
+ * commit() for any other BDS that have been left in a prepare() state
+ *
+ */
+int bdrv_reopen_prepare(BDRVReopenState *reopen_state, BlockReopenQueue *queue,
+                        Error **errp)
+{
+    int ret = -1;
+    Error *local_err = NULL;
+    BlockDriver *drv;
+
+    assert(reopen_state != NULL);
+    assert(reopen_state->bs->drv != NULL);
+    drv = reopen_state->bs->drv;
+
+    /* if we are to stay read-only, do not allow permission change
+     * to r/w */
+    if (!(reopen_state->bs->open_flags & BDRV_O_ALLOW_RDWR) &&
+        reopen_state->flags & BDRV_O_RDWR) {
+        error_set(errp, QERR_DEVICE_IS_READ_ONLY,
+                  reopen_state->bs->device_name);
+        goto error;
+    }
+
+
+    ret = bdrv_flush(reopen_state->bs);
+    if (ret) {
+        error_set(errp, ERROR_CLASS_GENERIC_ERROR, "Error (%s) flushing drive",
+                  strerror(-ret));
+        goto error;
+    }
+
+    if (drv->bdrv_reopen_prepare) {
+        ret = drv->bdrv_reopen_prepare(reopen_state, queue, &local_err);
+        if (ret) {
+            if (local_err != NULL) {
+                error_propagate(errp, local_err);
+            } else {
+                error_set(errp, QERR_OPEN_FILE_FAILED,
+                          reopen_state->bs->filename);
+            }
+            goto error;
+        }
+    } else {
+        /* It is currently mandatory to have a bdrv_reopen_prepare()
+         * handler for each supported drv. */
+        error_set(errp, QERR_BLOCK_FORMAT_FEATURE_NOT_SUPPORTED,
+                  drv->format_name, reopen_state->bs->device_name,
+                 "reopening of file");
+        ret = -1;
+        goto error;
+    }
+
+    ret = 0;
+
+error:
+    return ret;
+}
+
+/*
+ * Takes the staged changes for the reopen from bdrv_reopen_prepare(), and
+ * makes them final by swapping the staging BlockDriverState contents into
+ * the active BlockDriverState contents.
+ */
+void bdrv_reopen_commit(BDRVReopenState *reopen_state)
+{
+    BlockDriver *drv;
+
+    assert(reopen_state != NULL);
+    drv = reopen_state->bs->drv;
+    assert(drv != NULL);
+
+    /* If there are any driver level actions to take */
+    if (drv->bdrv_reopen_commit) {
+        drv->bdrv_reopen_commit(reopen_state);
+    }
+
+    /* set BDS specific flags now */
+    reopen_state->bs->open_flags         = reopen_state->flags;
+    reopen_state->bs->enable_write_cache = !!(reopen_state->flags &
+                                              BDRV_O_CACHE_WB);
+    reopen_state->bs->read_only = !(reopen_state->flags & BDRV_O_RDWR);
+}
+
+/*
+ * Abort the reopen, and delete and free the staged changes in
+ * reopen_state
+ */
+void bdrv_reopen_abort(BDRVReopenState *reopen_state)
+{
+    BlockDriver *drv;
+
+    assert(reopen_state != NULL);
+    drv = reopen_state->bs->drv;
+    assert(drv != NULL);
+
+    if (drv->bdrv_reopen_abort) {
+        drv->bdrv_reopen_abort(reopen_state);
+    }
+}
+
+
 void bdrv_close(BlockDriverState *bs)
 {
     bdrv_flush(bs);
diff --git a/block.h b/block.h
index 4d919c249fc625866e8a9ab96fc6b73873ff665d..b1095d8599a27943ba87ed7f5c5c230f4608a70b 100644 (file)
--- a/block.h
+++ b/block.h
@@ -97,6 +97,15 @@ typedef enum {
     BDRV_ACTION_REPORT, BDRV_ACTION_IGNORE, BDRV_ACTION_STOP
 } BlockQMPEventAction;
 
+typedef QSIMPLEQ_HEAD(BlockReopenQueue, BlockReopenQueueEntry) BlockReopenQueue;
+
+typedef struct BDRVReopenState {
+    BlockDriverState *bs;
+    int flags;
+    void *opaque;
+} BDRVReopenState;
+
+
 void bdrv_iostatus_enable(BlockDriverState *bs);
 void bdrv_iostatus_reset(BlockDriverState *bs);
 void bdrv_iostatus_disable(BlockDriverState *bs);
@@ -131,6 +140,14 @@ int bdrv_parse_cache_flags(const char *mode, int *flags);
 int bdrv_file_open(BlockDriverState **pbs, const char *filename, int flags);
 int bdrv_open(BlockDriverState *bs, const char *filename, int flags,
               BlockDriver *drv);
+BlockReopenQueue *bdrv_reopen_queue(BlockReopenQueue *bs_queue,
+                                    BlockDriverState *bs, int flags);
+int bdrv_reopen_multiple(BlockReopenQueue *bs_queue, Error **errp);
+int bdrv_reopen(BlockDriverState *bs, int bdrv_flags, Error **errp);
+int bdrv_reopen_prepare(BDRVReopenState *reopen_state,
+                        BlockReopenQueue *queue, Error **errp);
+void bdrv_reopen_commit(BDRVReopenState *reopen_state);
+void bdrv_reopen_abort(BDRVReopenState *reopen_state);
 void bdrv_close(BlockDriverState *bs);
 int bdrv_attach_dev(BlockDriverState *bs, void *dev);
 void bdrv_attach_dev_nofail(BlockDriverState *bs, void *dev);
index 4452f6f39810db78d7d5607e5d0ed62cc0765587..22b3d93d19b0de5719c55bd564c59e9b7bdadabe 100644 (file)
@@ -139,6 +139,13 @@ struct BlockDriver {
     int instance_size;
     int (*bdrv_probe)(const uint8_t *buf, int buf_size, const char *filename);
     int (*bdrv_probe_device)(const char *filename);
+
+    /* For handling image reopen for split or non-split files */
+    int (*bdrv_reopen_prepare)(BDRVReopenState *reopen_state,
+                               BlockReopenQueue *queue, Error **errp);
+    void (*bdrv_reopen_commit)(BDRVReopenState *reopen_state);
+    void (*bdrv_reopen_abort)(BDRVReopenState *reopen_state);
+
     int (*bdrv_open)(BlockDriverState *bs, int flags);
     int (*bdrv_file_open)(BlockDriverState *bs, const char *filename, int flags);
     int (*bdrv_read)(BlockDriverState *bs, int64_t sector_num,
@@ -336,6 +343,7 @@ struct BlockDriverState {
 
     /* long-running background operation */
     BlockJob *job;
+
 };
 
 int get_tmp_filename(char *filename, int size);