Re: [Qemu-block] [PATCH v3] block/vdi: Add locking for parallel requests

[Paolo Bonzini]

On 27/02/2015 20:54, Max Reitz wrote:
> When allocating a new cluster, the first write to it must be the one
> doing the allocation, because that one pads its write request to the
> cluster size; if another write to that cluster is executed before it,
> that write will be overwritten due to the padding.
> 
> See https://bugs.launchpad.net/qemu/+bug/1422307 for what can go wrong
> without this patch.
> 
> Cc: qemu-stable <address@hidden>
> Signed-off-by: Max Reitz <address@hidden>

Usage of CoMutex is tricky, but well commented.  So:

Reviewed-by: Paolo Bonzini <address@hidden>

> ---
> v3: Hopefully finally found the real issue which causes the problems
>     described in the bug report; at least it sounds very reasonable and
>     I can no longer reproduce any of the issues described there.
>     Thank you, Paolo and Stefan!
> ---
>  block/vdi.c | 25 +++++++++++++++++++++++++
>  1 file changed, 25 insertions(+)
> 
> diff --git a/block/vdi.c b/block/vdi.c
> index 74030c6..53bd02f 100644
> --- a/block/vdi.c
> +++ b/block/vdi.c
> @@ -53,6 +53,7 @@
>  #include "block/block_int.h"
>  #include "qemu/module.h"
>  #include "migration/migration.h"
> +#include "block/coroutine.h"
>  
>  #if defined(CONFIG_UUID)
>  #include <uuid/uuid.h>
> @@ -196,6 +197,8 @@ typedef struct {
>      /* VDI header (converted to host endianness). */
>      VdiHeader header;
>  
> +    CoMutex write_lock;
> +
>      Error *migration_blocker;
>  } BDRVVdiState;
>  
> @@ -504,6 +507,8 @@ static int vdi_open(BlockDriverState *bs, QDict *options, 
> int flags,
>                "vdi", bdrv_get_device_name(bs), "live migration");
>      migrate_add_blocker(s->migration_blocker);
>  
> +    qemu_co_mutex_init(&s->write_lock);
> +
>      return 0;
>  
>   fail_free_bmap:
> @@ -639,11 +644,31 @@ static int vdi_co_write(BlockDriverState *bs,
>                     buf, n_sectors * SECTOR_SIZE);
>              memset(block + (sector_in_block + n_sectors) * SECTOR_SIZE, 0,
>                     (s->block_sectors - n_sectors - sector_in_block) * 
> SECTOR_SIZE);
> +
> +            /* Note that this coroutine does not yield anywhere from reading 
> the
> +             * bmap entry until here, so in regards to all the coroutines 
> trying
> +             * to write to this cluster, the one doing the allocation will
> +             * always be the first to try to acquire the lock.
> +             * Therefore, it is also the first that will actually be able to
> +             * acquire the lock and thus the padded cluster is written before
> +             * the other coroutines can write to the affected area. */
> +            qemu_co_mutex_lock(&s->write_lock);
>              ret = bdrv_write(bs->file, offset, block, s->block_sectors);
> +            qemu_co_mutex_unlock(&s->write_lock);
>          } else {
>              uint64_t offset = s->header.offset_data / SECTOR_SIZE +
>                                (uint64_t)bmap_entry * s->block_sectors +
>                                sector_in_block;
> +            qemu_co_mutex_lock(&s->write_lock);
> +            /* This lock is only used to make sure the following write 
> operation
> +             * is executed after the write issued by the coroutine allocating
> +             * this cluster, therefore we do not need to keep it locked.
> +             * As stated above, the allocating coroutine will always try to 
> lock
> +             * the mutex before all the other concurrent accesses to that
> +             * cluster, therefore at this point we can be absolutely certain
> +             * that that write operation has returned (there may be other 
> writes
> +             * in flight, but they do not concern this very operation). */
> +            qemu_co_mutex_unlock(&s->write_lock);
>              ret = bdrv_write(bs->file, offset, buf, n_sectors);
>          }
>  
>