| Message ID | 1413993434-11816-2-git-send-email-mreitz@redhat.com |
|---|---|
| State | New |
| Headers | show |
On 10/22/2014 09:57 AM, Max Reitz wrote: > As its comment states, raw_co_get_block_status() should unconditionally > return 0 and set *pnum to 0 for after EOF. > > An assertion after lseek(..., SEEK_HOLE) tried to catch this case by > asserting that errno != -ENXIO (which would indicate a position after > the EOF); but it should be errno != ENXIO instead. Regardless of that, > there should be no such assertion at all. If bdrv_getlength() returned > an outdated value and the image has been resized outside of qemu, > lseek() will return with errno == ENXIO. Just return that value as an > error then. > > Setting *pnum to 0 and returning 0 should not be done here, as in that > case we should update the device length as well. So, from qemu's > perspective, the file has not been resized; it's just that there was an > error querying sectors beyond a certain point (the actual file size). > > Additionally, nb_sectors should be clamped against the image end. This > was probably not an issue if FIEMAP or SEEK_HOLE/SEEK_DATA worked, but > the fallback did not take this case into account. > > Reported-by: Kevin Wolf <kwolf@redhat.com> > Signed-off-by: Max Reitz <mreitz@redhat.com> > --- > block/raw-posix.c | 14 ++++++++++---- > 1 file changed, 10 insertions(+), 4 deletions(-) Reviewed-by: Eric Blake <eblake@redhat.com> > + if (total_size < 0) { > + return total_size; > + } else if (start >= total_size) { > + *pnum = 0; > + return 0; > + } else if (start + nb_sectors * BDRV_SECTOR_SIZE > total_size) { > + nb_sectors = (total_size - start) / BDRV_SECTOR_SIZE; Should this round up instead of truncate? But it would only matter for a file size that is not a multiple of sectors, where we probably have other issues, and where reporting just the full sectors also seems reasonable.
On 2014-10-22 at 18:57, Eric Blake wrote: > On 10/22/2014 09:57 AM, Max Reitz wrote: >> As its comment states, raw_co_get_block_status() should unconditionally >> return 0 and set *pnum to 0 for after EOF. >> >> An assertion after lseek(..., SEEK_HOLE) tried to catch this case by >> asserting that errno != -ENXIO (which would indicate a position after >> the EOF); but it should be errno != ENXIO instead. Regardless of that, >> there should be no such assertion at all. If bdrv_getlength() returned >> an outdated value and the image has been resized outside of qemu, >> lseek() will return with errno == ENXIO. Just return that value as an >> error then. >> >> Setting *pnum to 0 and returning 0 should not be done here, as in that >> case we should update the device length as well. So, from qemu's >> perspective, the file has not been resized; it's just that there was an >> error querying sectors beyond a certain point (the actual file size). >> >> Additionally, nb_sectors should be clamped against the image end. This >> was probably not an issue if FIEMAP or SEEK_HOLE/SEEK_DATA worked, but >> the fallback did not take this case into account. >> >> Reported-by: Kevin Wolf <kwolf@redhat.com> >> Signed-off-by: Max Reitz <mreitz@redhat.com> >> --- >> block/raw-posix.c | 14 ++++++++++---- >> 1 file changed, 10 insertions(+), 4 deletions(-) > Reviewed-by: Eric Blake <eblake@redhat.com> > >> + if (total_size < 0) { >> + return total_size; >> + } else if (start >= total_size) { >> + *pnum = 0; >> + return 0; >> + } else if (start + nb_sectors * BDRV_SECTOR_SIZE > total_size) { >> + nb_sectors = (total_size - start) / BDRV_SECTOR_SIZE; > Should this round up instead of truncate? But it would only matter for > a file size that is not a multiple of sectors, where we probably have > other issues, and where reporting just the full sectors also seems > reasonable. There already was a series (as far as I remember) that somehow tried to make all or at least some block drivers compatible with sizes which are not a multiple of the sector size, so I shouldn't be nullifying that work. Will use ROUND_UP(). Max
On 2014-10-23 at 09:27, Max Reitz wrote: > On 2014-10-22 at 18:57, Eric Blake wrote: >> On 10/22/2014 09:57 AM, Max Reitz wrote: >>> As its comment states, raw_co_get_block_status() should unconditionally >>> return 0 and set *pnum to 0 for after EOF. >>> >>> An assertion after lseek(..., SEEK_HOLE) tried to catch this case by >>> asserting that errno != -ENXIO (which would indicate a position after >>> the EOF); but it should be errno != ENXIO instead. Regardless of that, >>> there should be no such assertion at all. If bdrv_getlength() returned >>> an outdated value and the image has been resized outside of qemu, >>> lseek() will return with errno == ENXIO. Just return that value as an >>> error then. >>> >>> Setting *pnum to 0 and returning 0 should not be done here, as in that >>> case we should update the device length as well. So, from qemu's >>> perspective, the file has not been resized; it's just that there was an >>> error querying sectors beyond a certain point (the actual file size). >>> >>> Additionally, nb_sectors should be clamped against the image end. This >>> was probably not an issue if FIEMAP or SEEK_HOLE/SEEK_DATA worked, but >>> the fallback did not take this case into account. >>> >>> Reported-by: Kevin Wolf <kwolf@redhat.com> >>> Signed-off-by: Max Reitz <mreitz@redhat.com> >>> --- >>> block/raw-posix.c | 14 ++++++++++---- >>> 1 file changed, 10 insertions(+), 4 deletions(-) >> Reviewed-by: Eric Blake <eblake@redhat.com> >> >>> + if (total_size < 0) { >>> + return total_size; >>> + } else if (start >= total_size) { >>> + *pnum = 0; >>> + return 0; >>> + } else if (start + nb_sectors * BDRV_SECTOR_SIZE > total_size) { >>> + nb_sectors = (total_size - start) / BDRV_SECTOR_SIZE; >> Should this round up instead of truncate? But it would only matter for >> a file size that is not a multiple of sectors, where we probably have >> other issues, and where reporting just the full sectors also seems >> reasonable. > > There already was a series (as far as I remember) that somehow tried > to make all or at least some block drivers compatible with sizes which > are not a multiple of the sector size, so I shouldn't be nullifying > that work. Will use ROUND_UP(). Sorry, DIV_ROUND_UP(), of course. Max
diff --git a/block/raw-posix.c b/block/raw-posix.c index ee4ca3c..bd21fff 100644 --- a/block/raw-posix.c +++ b/block/raw-posix.c @@ -1519,10 +1519,6 @@ static int64_t try_seek_hole(BlockDriverState *bs, off_t start, off_t *data, *hole = lseek(s->fd, start, SEEK_HOLE); if (*hole == -1) { - /* -ENXIO indicates that sector_num was past the end of the file. - * There is a virtual hole there. */ - assert(errno != -ENXIO); - return -errno; } @@ -1562,6 +1558,7 @@ static int64_t coroutine_fn raw_co_get_block_status(BlockDriverState *bs, int nb_sectors, int *pnum) { off_t start, data = 0, hole = 0; + int64_t total_size; int64_t ret; ret = fd_open(bs); @@ -1570,6 +1567,15 @@ static int64_t coroutine_fn raw_co_get_block_status(BlockDriverState *bs, } start = sector_num * BDRV_SECTOR_SIZE; + total_size = bdrv_getlength(bs); + if (total_size < 0) { + return total_size; + } else if (start >= total_size) { + *pnum = 0; + return 0; + } else if (start + nb_sectors * BDRV_SECTOR_SIZE > total_size) { + nb_sectors = (total_size - start) / BDRV_SECTOR_SIZE; + } ret = try_seek_hole(bs, start, &data, &hole, pnum); if (ret < 0) {
As its comment states, raw_co_get_block_status() should unconditionally return 0 and set *pnum to 0 for after EOF. An assertion after lseek(..., SEEK_HOLE) tried to catch this case by asserting that errno != -ENXIO (which would indicate a position after the EOF); but it should be errno != ENXIO instead. Regardless of that, there should be no such assertion at all. If bdrv_getlength() returned an outdated value and the image has been resized outside of qemu, lseek() will return with errno == ENXIO. Just return that value as an error then. Setting *pnum to 0 and returning 0 should not be done here, as in that case we should update the device length as well. So, from qemu's perspective, the file has not been resized; it's just that there was an error querying sectors beyond a certain point (the actual file size). Additionally, nb_sectors should be clamped against the image end. This was probably not an issue if FIEMAP or SEEK_HOLE/SEEK_DATA worked, but the fallback did not take this case into account. Reported-by: Kevin Wolf <kwolf@redhat.com> Signed-off-by: Max Reitz <mreitz@redhat.com> --- block/raw-posix.c | 14 ++++++++++---- 1 file changed, 10 insertions(+), 4 deletions(-)