Dima, please review the patch. -- Best regards,
Konstantin Khorenko, Virtuozzo Linux Kernel Team On 03/24/2016 03:20 AM, Maxim Patlasov wrote:
While handling fallocate(), ext4 firstly calls ext4_update_inode_size(), then ext4_mark_inode_dirty(). Both update only in-core state. After fallocate(), ploop updates BAT bypassing vfs layer (submit bio to underlying block device). Hence, it's possible that BAT is already updated, but i_size in now written to the disk yet. If a crash happens in the middle, BAT points beyond end-of-file, i.e. ploop image is corrupted. The patch fixes the problem by ordinary fsync(). This is OK because we fallocate relatively seldom (PLOOP_MAX_PREALLOC=128MB) and also because no dirty pages to sync (all I/O is direct). https://jira.sw.ru/browse/PSBM-45326 Signed-off-by: Maxim Patlasov <[email protected]> --- drivers/block/ploop/io_direct.c | 6 ++++++ 1 file changed, 6 insertions(+) diff --git a/drivers/block/ploop/io_direct.c b/drivers/block/ploop/io_direct.c index 0c7e0c4..8032999 100644 --- a/drivers/block/ploop/io_direct.c +++ b/drivers/block/ploop/io_direct.c @@ -376,6 +376,7 @@ cached_submit(struct ploop_io *io, iblock_t iblk, struct ploop_request * preq, loff_t new_size; loff_t used_pos; bool may_fallocate = io->files.file->f_op->fallocate && + io->files.file->f_op->fsync && io->files.flags & EXT4_EXTENTS_FL; trace_cached_submit(preq); @@ -404,6 +405,11 @@ try_again: } } + /* flush new i_size to disk */ + err = io->files.file->f_op->FOP_FSYNC(io->files.file, 0); + if (err) + goto end_write; + io->prealloced_size = prealloc; } .
_______________________________________________ Devel mailing list [email protected] https://lists.openvz.org/mailman/listinfo/devel
