Re: [Cluster-devel] [GFS2 0/3] fallocate and quota fixes
- Original Message - This is a revised version of the patches required to properly fix the fallocate quota issue described in bz1174295 patch1: This patch supplies gfs2_quota_check() with the number of blocks the caller intends to allocate in the current operation, resulting in a more accurate quota check. patch2: gfs2_quota_check() and gfs2_inplace_reserve() return the number of blocks available subject to quota limits and rgrp size respectively. The difference from previous versions of this patch is that we return the available blocks even on success. patch3: The fallocate() function uses the features of patch2 to determine how many total blocks are available for allocation and uses them right away, instead of guessing inefficiently. The behavior of quota enforcement is altered by this patchset. i. Quotas are exceeded (a warning message is also issued to syslog) before the actual usage blocks exceed the imposed limit. In fact, the actual usage can never exceed the limit. Whenever it is determined that the completion of an operation will cause a quota to exceed its limit, such an operation is aborted with -EDQUOT and a 'quota exceeded' message is dispatched. ii. The gfs2_write_calc_reserv()/calc_max_reserv() functions are used to map between available blocks and the data bytes that can be written using them. Typically, for large files, some blocks are used up for metadata and only the remaining blocks can be used for data. Example: To write only a handful of bytes that would easily fit in one block, we might have to allocate an extra bunch of intermediate metadata blocks. If we had only 1 block left in our allotted quota, this operation would likely fail. The functions mentioned in ii. are not very efficient. They always compute the worst case number of extra blocks required and it is often the case that not all those extra blocks are used. We need to find a better algorithm to get a tighter estimate on the blocks needed for a given number of bytes. I've run some basic tests and things seem to be holding up. The failing case in bz1174295 is fixed using this patchset. I'll do test build and pass it on to Nate to test with. Abhi Das (3): gfs2: perform quota checks against allocation parameters gfs2: allow quota_check and inplace_reserve to return available blocks gfs2: allow fallocate to max out quotas/fs efficiently fs/gfs2/aops.c | 6 ++--- fs/gfs2/bmap.c | 2 +- fs/gfs2/file.c | 75 ++-- fs/gfs2/incore.h | 3 ++- fs/gfs2/inode.c | 18 -- fs/gfs2/quota.c | 40 ++ fs/gfs2/quota.h | 8 +++--- fs/gfs2/rgrp.c | 12 +++-- fs/gfs2/rgrp.h | 6 +++-- fs/gfs2/xattr.c | 2 +- 10 files changed, 117 insertions(+), 55 deletions(-) -- 1.8.1.4 Looks good to me. ACK to the series. Regards, Bob Peterson Red Hat File Systems
Re: [Cluster-devel] [GFS2 0/3] fallocate and quota fixes
Hi, On 25/02/15 05:17, Abhi Das wrote: This is a revised version of the patches required to properly fix the fallocate quota issue described in bz1174295 patch1: This patch supplies gfs2_quota_check() with the number of blocks the caller intends to allocate in the current operation, resulting in a more accurate quota check. patch2: gfs2_quota_check() and gfs2_inplace_reserve() return the number of blocks available subject to quota limits and rgrp size respectively. The difference from previous versions of this patch is that we return the available blocks even on success. patch3: The fallocate() function uses the features of patch2 to determine how many total blocks are available for allocation and uses them right away, instead of guessing inefficiently. The behavior of quota enforcement is altered by this patchset. i. Quotas are exceeded (a warning message is also issued to syslog) before the actual usage blocks exceed the imposed limit. In fact, the actual usage can never exceed the limit. Whenever it is determined that the completion of an operation will cause a quota to exceed its limit, such an operation is aborted with -EDQUOT and a 'quota exceeded' message is dispatched. ii. The gfs2_write_calc_reserv()/calc_max_reserv() functions are used to map between available blocks and the data bytes that can be written using them. Typically, for large files, some blocks are used up for metadata and only the remaining blocks can be used for data. Example: To write only a handful of bytes that would easily fit in one block, we might have to allocate an extra bunch of intermediate metadata blocks. If we had only 1 block left in our allotted quota, this operation would likely fail. The functions mentioned in ii. are not very efficient. They always compute the worst case number of extra blocks required and it is often the case that not all those extra blocks are used. We need to find a better algorithm to get a tighter estimate on the blocks needed for a given number of bytes. I've run some basic tests and things seem to be holding up. The failing case in bz1174295 is fixed using this patchset. I'll do test build and pass it on to Nate to test with. Abhi Das (3): gfs2: perform quota checks against allocation parameters gfs2: allow quota_check and inplace_reserve to return available blocks gfs2: allow fallocate to max out quotas/fs efficiently fs/gfs2/aops.c | 6 ++--- fs/gfs2/bmap.c | 2 +- fs/gfs2/file.c | 75 ++-- fs/gfs2/incore.h | 3 ++- fs/gfs2/inode.c | 18 -- fs/gfs2/quota.c | 40 ++ fs/gfs2/quota.h | 8 +++--- fs/gfs2/rgrp.c | 12 +++-- fs/gfs2/rgrp.h | 6 +++-- fs/gfs2/xattr.c | 2 +- 10 files changed, 117 insertions(+), 55 deletions(-) This looks good, but definitely needs careful testing, Steve.
