| Message ID | 20210623150143.188184-1-mreitz@redhat.com (mailing list archive) |
|---|---|
| Headers | show |
| Series | block: block-status cache for data regions | expand |
Am 23.06.2021 um 17:01 hat Max Reitz geschrieben: > Hi, > > See the cover letter from v1 for the general idea: > > https://lists.nongnu.org/archive/html/qemu-block/2021-06/msg00843.html > > > The biggest change here in v2 is that instead of having a common CoMutex > protect the block-status cache, we’re using RCU now. So to read from > the cache, or even to invalidate it, no lock is needed, only to update > it with new data. > > Disclaimer: I have no experience with RCU in practice so far, neither in > qemu nor anywhere else. So I hope I’ve used it correctly... This I hope as well. :-) With the missing qatomic_rcu_read() added: Reviewed-by: Kevin Wolf <kwolf@redhat.com>
Hi, See the cover letter from v1 for the general idea: https://lists.nongnu.org/archive/html/qemu-block/2021-06/msg00843.html The biggest change here in v2 is that instead of having a common CoMutex protect the block-status cache, we’re using RCU now. So to read from the cache, or even to invalidate it, no lock is needed, only to update it with new data. Disclaimer: I have no experience with RCU in practice so far, neither in qemu nor anywhere else. So I hope I’ve used it correctly... Differences to v1 in detail: - Patch 2: - Moved BdrvBlockStatusCache.lock up to BDS, it is now the RCU writer lock - BDS.block_status_cache is now a pointer, so it can be replaced with RCU - Moved all cache access functionality into helper functions (bdrv_bsc_is_data(), bdrv_bsc_invalidate_range(), bdrv_bsc_fill()) in block.c - Guard BSC accesses with RCU (BSC.valid is to be accessed atomically, which allows resetting it without taking an RCU write lock) - Check QLIST_EMPTY(&bs->children) not just when reading from the cache, but when filling it, too (so we don’t need an RCU update when it won’t make sense) - Patch 3: Added - Dropped the block/nbd patch (because it would make NBD query a larger range; the patch’s intent was to get more information for free, which this would not be) git-backport-diff against v1: Key: [----] : patches are identical [####] : number of functional differences between upstream/downstream patch [down] : patch is downstream-only The flags [FC] indicate (F)unctional and (C)ontextual differences, respectively 001/6:[----] [--] 'block: Drop BDS comment regarding bdrv_append()' 002/6:[0169] [FC] 'block: block-status cache for data regions' 003/6:[down] 'block: Clarify that @bytes is no limit on *pnum' 004/6:[----] [--] 'block/file-posix: Do not force-cap *pnum' 005/6:[----] [--] 'block/gluster: Do not force-cap *pnum' 006/6:[----] [--] 'block/iscsi: Do not force-cap *pnum' Max Reitz (6): block: Drop BDS comment regarding bdrv_append() block: block-status cache for data regions block: Clarify that @bytes is no limit on *pnum block/file-posix: Do not force-cap *pnum block/gluster: Do not force-cap *pnum block/iscsi: Do not force-cap *pnum include/block/block_int.h | 54 +++++++++++++++++++++++-- block.c | 84 +++++++++++++++++++++++++++++++++++++++ block/file-posix.c | 7 ++-- block/gluster.c | 7 ++-- block/io.c | 61 ++++++++++++++++++++++++++-- block/iscsi.c | 3 -- 6 files changed, 200 insertions(+), 16 deletions(-)