Re: [RFC PATCH] ceph: Write through cache support based on fscache
Hi Milosz, Thanks for your comments. We think SSD and fscache based write cache is definitely useful for Ceph, since to some extent, write amplification slow down the write performance of Ceph. Lustre has already introduced SSD based write cache. SSD can be treated as an outer big cache for page cache. It can reduce the requirement of network and OSD bandwidth. Write back cache is more performance useful, but more complicated to implement to meet the consistence and other correctness semantic demands of Ceph and POSIX, such as sync(). Write through cache is much simpler, which will not bother too much. So our goal is to implement both, we plan to submit it as a blueprint at the incoming CDS. It would be great if you could help review and give comments on our codes during the development. Again, thanks very much. Cheers, Li Wang On 11/02/2013 12:51 AM, Milosz Tanski wrote: Li, I think it would be fantastic to see a write cache. In many workloads you ended up writing out a file and then turning around and reading it right back in on the same node. There's a few things that I would like to see. First, an mount option to turn on/off write through caching. There are some workloads / user hardware configurations that will not benefit from this (it might be a net negative). Also, I think it's nice to have a fallback to disable it it's miss behaving. Second, for correctness I think you should only do write-through caching if you have an exclusive cap on the file. Currently as the code is written it only reads from fscache if the file is open in read only mode and has the cache cap. This would also have to change. Thanks, - Milosz P.S: Sorry for the second message Li, I fail at email and forgot to reply-all. On Fri, Nov 1, 2013 at 9:49 AM, Li Wang liw...@ubuntukylin.com wrote: Currently, fscache only plays as read cache for ceph, this patch enables it plays as the write through cache as well. A small trick to be discussed: if the writing to OSD finishes before the writing to fscache, the fscache writing is cancelled to avoid slow down the writepages() process. Signed-off-by: Min Chen minc...@ubuntukylin.com Signed-off-by: Li Wang liw...@ubuntukylin.com Signed-off-by: Yunchuan Wen yunchuan...@ubuntukylin.com --- fs/ceph/addr.c | 10 +++--- fs/ceph/cache.c | 29 + fs/ceph/cache.h | 13 + 3 files changed, 49 insertions(+), 3 deletions(-) diff --git a/fs/ceph/addr.c b/fs/ceph/addr.c index 6df8bd4..2465c49 100644 --- a/fs/ceph/addr.c +++ b/fs/ceph/addr.c @@ -506,7 +506,7 @@ static int writepage_nounlock(struct page *page, struct writeback_control *wbc) CONGESTION_ON_THRESH(fsc-mount_options-congestion_kb)) set_bdi_congested(fsc-backing_dev_info, BLK_RW_ASYNC); - ceph_readpage_to_fscache(inode, page); + ceph_writepage_to_fscache(inode, page); set_page_writeback(page); err = ceph_osdc_writepages(osdc, ceph_vino(inode), @@ -634,6 +634,7 @@ static void writepages_finish(struct ceph_osd_request *req, if ((issued (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO)) == 0) generic_error_remove_page(inode-i_mapping, page); + ceph_maybe_release_fscache_page(inode, page); unlock_page(page); } dout(%p wrote+cleaned %d pages\n, inode, wrote); @@ -746,7 +747,7 @@ retry: while (!done index = end) { int num_ops = do_sync ? 2 : 1; - unsigned i; + unsigned i, j; int first; pgoff_t next; int pvec_pages, locked_pages; @@ -894,7 +895,6 @@ get_more_pages: if (!locked_pages) goto release_pvec_pages; if (i) { - int j; BUG_ON(!locked_pages || first 0); if (pvec_pages i == pvec_pages @@ -924,6 +924,10 @@ get_more_pages: osd_req_op_extent_osd_data_pages(req, 0, pages, len, 0, !!pool, false); + for (j = 0; j locked_pages; j++) { + struct page *page = pages[j]; + ceph_writepage_to_fscache(inode, page); + } pages = NULL; /* request message now owns the pages array */ pool = NULL; diff --git a/fs/ceph/cache.c b/fs/ceph/cache.c index 6bfe65e..6f928c4 100644 --- a/fs/ceph/cache.c +++ b/fs/ceph/cache.c @@ -320,6 +320,24 @@ void ceph_readpage_to_fscache(struct inode *inode, struct page *page) fscache_uncache_page(ci-fscache, page); } +void ceph_writepage_to_fscache(struct inode *inode, struct page *page) +{ + struct ceph_inode_info *ci = ceph_inode(inode); + int ret; + + if (!cache_valid(ci)) + return; + + if (!PageFsCache(page
[RFC PATCH] ceph: Write through cache support based on fscache
Currently, fscache only plays as read cache for ceph, this patch enables it plays as the write through cache as well. A small trick to be discussed: if the writing to OSD finishes before the writing to fscache, the fscache writing is cancelled to avoid slow down the writepages() process. Signed-off-by: Min Chen minc...@ubuntukylin.com Signed-off-by: Li Wang liw...@ubuntukylin.com Signed-off-by: Yunchuan Wen yunchuan...@ubuntukylin.com --- fs/ceph/addr.c | 10 +++--- fs/ceph/cache.c | 29 + fs/ceph/cache.h | 13 + 3 files changed, 49 insertions(+), 3 deletions(-) diff --git a/fs/ceph/addr.c b/fs/ceph/addr.c index 6df8bd4..2465c49 100644 --- a/fs/ceph/addr.c +++ b/fs/ceph/addr.c @@ -506,7 +506,7 @@ static int writepage_nounlock(struct page *page, struct writeback_control *wbc) CONGESTION_ON_THRESH(fsc-mount_options-congestion_kb)) set_bdi_congested(fsc-backing_dev_info, BLK_RW_ASYNC); - ceph_readpage_to_fscache(inode, page); + ceph_writepage_to_fscache(inode, page); set_page_writeback(page); err = ceph_osdc_writepages(osdc, ceph_vino(inode), @@ -634,6 +634,7 @@ static void writepages_finish(struct ceph_osd_request *req, if ((issued (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO)) == 0) generic_error_remove_page(inode-i_mapping, page); + ceph_maybe_release_fscache_page(inode, page); unlock_page(page); } dout(%p wrote+cleaned %d pages\n, inode, wrote); @@ -746,7 +747,7 @@ retry: while (!done index = end) { int num_ops = do_sync ? 2 : 1; - unsigned i; + unsigned i, j; int first; pgoff_t next; int pvec_pages, locked_pages; @@ -894,7 +895,6 @@ get_more_pages: if (!locked_pages) goto release_pvec_pages; if (i) { - int j; BUG_ON(!locked_pages || first 0); if (pvec_pages i == pvec_pages @@ -924,6 +924,10 @@ get_more_pages: osd_req_op_extent_osd_data_pages(req, 0, pages, len, 0, !!pool, false); + for (j = 0; j locked_pages; j++) { + struct page *page = pages[j]; + ceph_writepage_to_fscache(inode, page); + } pages = NULL; /* request message now owns the pages array */ pool = NULL; diff --git a/fs/ceph/cache.c b/fs/ceph/cache.c index 6bfe65e..6f928c4 100644 --- a/fs/ceph/cache.c +++ b/fs/ceph/cache.c @@ -320,6 +320,24 @@ void ceph_readpage_to_fscache(struct inode *inode, struct page *page) fscache_uncache_page(ci-fscache, page); } +void ceph_writepage_to_fscache(struct inode *inode, struct page *page) +{ + struct ceph_inode_info *ci = ceph_inode(inode); + int ret; + + if (!cache_valid(ci)) + return; + + if (!PageFsCache(page)) { + if (fscache_alloc_page(ci-fscache, page, GFP_KERNEL)) + return; + } + + if (fscache_write_page(ci-fscache, page, GFP_KERNEL)) + fscache_uncache_page(ci-fscache, page); +} + + void ceph_invalidate_fscache_page(struct inode* inode, struct page *page) { struct ceph_inode_info *ci = ceph_inode(inode); @@ -328,6 +346,17 @@ void ceph_invalidate_fscache_page(struct inode* inode, struct page *page) fscache_uncache_page(ci-fscache, page); } +void ceph_maybe_release_fscache_page(struct inode *inode, struct page *page) +{ + struct ceph_inode_info *ci = ceph_inode(inode); + + if (PageFsCache(page)) { + if (!fscache_check_page_write(ci-fscache, page)) + fscache_maybe_release_page(ci-fscache, + page, GFP_KERNEL); + } +} + void ceph_fscache_unregister_fs(struct ceph_fs_client* fsc) { if (fsc-revalidate_wq) diff --git a/fs/ceph/cache.h b/fs/ceph/cache.h index ba94940..aa02b7a 100644 --- a/fs/ceph/cache.h +++ b/fs/ceph/cache.h @@ -45,7 +45,9 @@ int ceph_readpages_from_fscache(struct inode *inode, struct list_head *pages, unsigned *nr_pages); void ceph_readpage_to_fscache(struct inode *inode, struct page *page); +void ceph_writepage_to_fscache(struct inode *inode, struct page *page); void ceph_invalidate_fscache_page(struct inode* inode, struct page *page); +void ceph_maybe_release_fscache_page(struct inode *inode, struct page *page); void ceph_queue_revalidate(struct inode *inode); static inline void ceph_fscache_invalidate(struct inode *inode) @@ -127,6 +129,11 @@ static inline void ceph_readpage_to_fscache(struct inode *inode, { } +static inline void
Re: [RFC PATCH] ceph: Write through cache support based on fscache
Li, I think it would be fantastic to see a write cache. In many workloads you ended up writing out a file and then turning around and reading it right back in on the same node. There's a few things that I would like to see. First, an mount option to turn on/off write through caching. There are some workloads / user hardware configurations that will not benefit from this (it might be a net negative). Also, I think it's nice to have a fallback to disable it it's miss behaving. Second, for correctness I think you should only do write-through caching if you have an exclusive cap on the file. Currently as the code is written it only reads from fscache if the file is open in read only mode and has the cache cap. This would also have to change. Thanks, - Milosz P.S: Sorry for the second message Li, I fail at email and forgot to reply-all. On Fri, Nov 1, 2013 at 9:49 AM, Li Wang liw...@ubuntukylin.com wrote: Currently, fscache only plays as read cache for ceph, this patch enables it plays as the write through cache as well. A small trick to be discussed: if the writing to OSD finishes before the writing to fscache, the fscache writing is cancelled to avoid slow down the writepages() process. Signed-off-by: Min Chen minc...@ubuntukylin.com Signed-off-by: Li Wang liw...@ubuntukylin.com Signed-off-by: Yunchuan Wen yunchuan...@ubuntukylin.com --- fs/ceph/addr.c | 10 +++--- fs/ceph/cache.c | 29 + fs/ceph/cache.h | 13 + 3 files changed, 49 insertions(+), 3 deletions(-) diff --git a/fs/ceph/addr.c b/fs/ceph/addr.c index 6df8bd4..2465c49 100644 --- a/fs/ceph/addr.c +++ b/fs/ceph/addr.c @@ -506,7 +506,7 @@ static int writepage_nounlock(struct page *page, struct writeback_control *wbc) CONGESTION_ON_THRESH(fsc-mount_options-congestion_kb)) set_bdi_congested(fsc-backing_dev_info, BLK_RW_ASYNC); - ceph_readpage_to_fscache(inode, page); + ceph_writepage_to_fscache(inode, page); set_page_writeback(page); err = ceph_osdc_writepages(osdc, ceph_vino(inode), @@ -634,6 +634,7 @@ static void writepages_finish(struct ceph_osd_request *req, if ((issued (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO)) == 0) generic_error_remove_page(inode-i_mapping, page); + ceph_maybe_release_fscache_page(inode, page); unlock_page(page); } dout(%p wrote+cleaned %d pages\n, inode, wrote); @@ -746,7 +747,7 @@ retry: while (!done index = end) { int num_ops = do_sync ? 2 : 1; - unsigned i; + unsigned i, j; int first; pgoff_t next; int pvec_pages, locked_pages; @@ -894,7 +895,6 @@ get_more_pages: if (!locked_pages) goto release_pvec_pages; if (i) { - int j; BUG_ON(!locked_pages || first 0); if (pvec_pages i == pvec_pages @@ -924,6 +924,10 @@ get_more_pages: osd_req_op_extent_osd_data_pages(req, 0, pages, len, 0, !!pool, false); + for (j = 0; j locked_pages; j++) { + struct page *page = pages[j]; + ceph_writepage_to_fscache(inode, page); + } pages = NULL; /* request message now owns the pages array */ pool = NULL; diff --git a/fs/ceph/cache.c b/fs/ceph/cache.c index 6bfe65e..6f928c4 100644 --- a/fs/ceph/cache.c +++ b/fs/ceph/cache.c @@ -320,6 +320,24 @@ void ceph_readpage_to_fscache(struct inode *inode, struct page *page) fscache_uncache_page(ci-fscache, page); } +void ceph_writepage_to_fscache(struct inode *inode, struct page *page) +{ + struct ceph_inode_info *ci = ceph_inode(inode); + int ret; + + if (!cache_valid(ci)) + return; + + if (!PageFsCache(page)) { + if (fscache_alloc_page(ci-fscache, page, GFP_KERNEL)) + return; + } + + if (fscache_write_page(ci-fscache, page, GFP_KERNEL)) + fscache_uncache_page(ci-fscache, page); +} + + void ceph_invalidate_fscache_page(struct inode* inode, struct page *page) { struct ceph_inode_info *ci = ceph_inode(inode); @@ -328,6 +346,17 @@ void ceph_invalidate_fscache_page(struct inode* inode, struct page *page) fscache_uncache_page(ci-fscache, page); } +void ceph_maybe_release_fscache_page(struct inode *inode, struct page *page) +{ + struct ceph_inode_info *ci = ceph_inode(inode); + + if (PageFsCache(page)) { + if (!fscache_check_page_write(ci-fscache, page)) + fscache_maybe_release_page(ci-fscache
write-through cache
I would like to test the effect of using the new write-through cache on RBD volumes mounted to Openstack VMs. What, precisely, are the changes I need to make to the volume XML in order to do so? -Mandell -- To unsubscribe from this list: send the line unsubscribe ceph-devel in the body of a message to majord...@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
Re: write-through cache
On 05/23/2012 04:44 PM, Mandell Degerness wrote: I would like to test the effect of using the new write-through cache on RBD volumes mounted to Openstack VMs. What, precisely, are the changes I need to make to the volume XML in order to do so? If your only volumes are rbd, you can append :rbd_cache=true:rbd_cache_max_dirty=0 to the image name in the volume xml, so it would include something like source protocol=rbd name=pool/image:rbd_cache=true:rbd_cache_max_dirty=0 Josh -- To unsubscribe from this list: send the line unsubscribe ceph-devel in the body of a message to majord...@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html