ceph: Update the pages in fscache in writepages() path
PATCH 1/1 ceph: Update the pages in fscache in writepages() path
Currently, the pages in fscache only are updated in writepage() path, add the process in writepages().
--- fs/ceph/addr.c | 8 +++++--- 1 file changed, 5 insertions(+), 3 deletions(-)
fs/ceph/addr.c +5 -3
@@ -746,7 +746,7 @@ 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 +894,6 @@ 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 &&
@@ -924,7 +923,10 @@ 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_readpage_to_fscache (inode, page); + } pages =NULL ;/* request message now owns the pages array */ pool =NULL ;
Signature · 1 line
1.7.9.5
Li,
First, sorry for the late reply on this.
Currently fscache is only supported for files that are open in read only mode. I originally was going to let fscache cache in the write path as well as long as the file was open in with O_LAZY. I abandoned that idea. When a user opens the file in O_LAZY we can cache things locally with the assumption that the user will care of the synchronization in some other manner. There is no way of invalidating a subset of the pages in object cached by fscache, there is no way we can make O_LAZY work well.
The ceph_readpage_to_fscache() in writepage has no effect and it should be removed. ceph_readpage_to_fscache() calls cache_valid() to see if it should perform the page save, and since the file can't have a CACHE cap at the point in time it doesn't do it.
Thanks,
- Milosz
Quoted 50 lines · Li Wang
Currently, the pages in fscache only are updated in
writepage()path, add the process inwritepages().Signed-off-by Min Chen <[email protected]>Signed-off-by Li Wang <[email protected]>Signed-off-by Yunchuan Wen <[email protected]>--- fs/ceph/addr.c | 8 +++++--- 1 file changed, 5 insertions(+), 3 deletions(-)fs/ceph/addr.c +5
-4@@ -746,7 +746,7 @@ 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 +894,6 @@ 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 &&@@ -924,7 +923,10 @@ 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_readpage_to_fscache (inode, page); + } pages =NULL ;/* request message now owns the pages array */ pool =NULL ; --1.7.9.5
--
To unsubscribe from this list: send the line "unsubscribe ceph-devel" in the body of a message to [email protected]
More majordomo info at http://vger.kernel.org/majordomo-info.html
Signature · 7 lines
Milosz Tanski CTO 10 East 53rd Street, 37th floor New York, NY 10022 p: 646-253-9055 e: [email protected]
Quoted 13 lines · Milosz Tanski
Li,
First, sorry for the late reply on this.
Currently fscache is only supported for files that are open in read only mode. I originally was going to let fscache cache in the write path as well as long as the file was open in with
O_LAZY. I abandoned that idea. When a user opens the file inO_LAZYwe can cache things locally with the assumption that the user will care of the synchronization in some other manner. There is no way of invalidating a subset of the pages in object cached by fscache, there is no way we can makeO_LAZYwork well.
Quoted 4 lines
The
ceph_readpage_to_fscache()in writepage has no effect and it should be removed.ceph_readpage_to_fscache()callscache_valid()to see if it should perform the page save, and since the file can't have a CACHE cap at the point in time it doesn't do it.
(Hmm, Dusting off my understanding of fscache and reading fs/ceph/cache.c; watch out!) It looks like cache_valid is
static inline int cache_valid(struct ceph_inode_info *ci)
{
return ((ceph_caps_issued(ci) & CEPH_CAP_FILE_CACHE) &&
(ci->i_fscache_gen == ci->i_rdcache_gen));
}and in the FILE_EXCL case, the MDS will issue CACHE|BUFFER caps. But I think the aux key (size+mtime) will prevent any use of the cache as soon as the first write happens and mtime changes, right?
I think that in order to make this work, we need to fix/create a file_version (or something similar) field in the (mds) inode_t to have some useful value. I.e., increment it any time
- a different client/writer comes along
- a file is modified by the mds (e.g., truncated or recovered)
but allow it to otherwise remain the same as long as only a single client is working with the file exclusively. This will be more precise than the (size, mtime) check that is currently used, and would remain valid when a single client opens the same file for exclusive read/write multiple times but there are no other intervening changes.
Milosz, if that were in place, is there any reason not to wire up writepage and allow the fscache to be used write-through?
sage
Quoted 68 lines
Thanks,
- MiloszQuoted 50 lines · Li Wang
Currently, the pages in fscache only are updated in
writepage()path, add the process inwritepages().Signed-off-by Min Chen <[email protected]>Signed-off-by Li Wang <[email protected]>Signed-off-by Yunchuan Wen <[email protected]>--- fs/ceph/addr.c | 8 +++++--- 1 file changed, 5 insertions(+), 3 deletions(-)fs/ceph/addr.c +5
-4@@ -746,7 +746,7 @@ 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 +894,6 @@ 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 &&@@ -924,7 +923,10 @@ 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_readpage_to_fscache (inode, page); + } pages =NULL ;/* request message now owns the pages array */ pool =NULL ; --1.7.9.5
--
To unsubscribe from this list: send the line "unsubscribe ceph-devel" in the body of a message to [email protected]
More majordomo info at http://vger.kernel.org/majordomo-info.htmlSignature · 7 lines
Milosz Tanski CTO 10 East 53rd Street, 37th floor New York, NY 10022 p: 646-253-9055 e: [email protected]
Sage,
I think the incrementing version counter on the whole is a neater solution then using size and mtime. If nothing else it's more explicit in the the read cache version. With what you suggested plus additional changes to the open code (where the cookie gets created) the write-through scenario should be correct.
Sadly, my understanding of the MDS protocol is still not great. So when doing this in the first place I erred on the side of using what was already in place.
In a kind of un-related question. Is there a debug hook in the kclient (or MDS for that matter) to dump the current file inodes (names) with issues caps and to which hosts. This would be very helpful for debugging, since from time to time I see a one of the clients get stuck in getattr (via mdsc debug log).
Thanks,
- Milosz
Quoted 121 lines · Sage Weil
Quoted 13 lines · Milosz Tanski
Li,
First, sorry for the late reply on this.
Currently fscache is only supported for files that are open in read only mode. I originally was going to let fscache cache in the write path as well as long as the file was open in with
O_LAZY. I abandoned that idea. When a user opens the file inO_LAZYwe can cache things locally with the assumption that the user will care of the synchronization in some other manner. There is no way of invalidating a subset of the pages in object cached by fscache, there is no way we can makeO_LAZYwork well.Quoted 4 lines
The
ceph_readpage_to_fscache()in writepage has no effect and it should be removed.ceph_readpage_to_fscache()callscache_valid()to see if it should perform the page save, and since the file can't have a CACHE cap at the point in time it doesn't do it.(Hmm, Dusting off my understanding of fscache and reading
fs/ceph/cache.c; watch out!) It looks likecache_validisstatic inline int
cache_valid(struct ceph_inode_info *ci){ return ((ceph_caps_issued(ci) & CEPH_CAP_FILE_CACHE) && (ci->i_fscache_gen == ci->i_rdcache_gen)); }and in the
FILE_EXCLcase, the MDS will issue CACHE|BUFFER caps. But I think the aux key (size+mtime) will prevent any use of the cache as soon as the first write happens and mtime changes, right?I think that in order to make this work, we need to fix/create a
file_version(or something similar) field in the (mds)inode_tto have some useful value. I.e., increment it any time- a different client/writer comes along
- a file is modified by the mds (e.g., truncated or recovered)but allow it to otherwise remain the same as long as only a single client is working with the file exclusively. This will be more precise than the (size, mtime) check that is currently used, and would remain valid when a single client opens the same file for exclusive read/write multiple times but there are no other intervening changes.
Milosz, if that were in place, is there any reason not to wire up writepage and allow the fscache to be used write-through?
sage
Quoted 68 lines
Thanks,
- MiloszQuoted 46 lines · Li Wang
Currently, the pages in fscache only are updated in
writepage()path, add the process inwritepages().Signed-off-by Min Chen <[email protected]>Signed-off-by Li Wang <[email protected]>Signed-off-by Yunchuan Wen <[email protected]>--- fs/ceph/addr.c | 8 +++++--- 1 file changed, 5 insertions(+), 3 deletions(-)fs/ceph/addr.c +5
-4@@ -746,7 +746,7 @@ 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 +894,6 @@ 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 &&@@ -924,7 +923,10 @@ 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_readpage_to_fscache (inode, page); + } pages =NULL ;/* request message now owns the pages array */ pool =NULL ; --1.7.9.5
Quoted 4 lines
--
To unsubscribe from this list: send the line "unsubscribe ceph-devel" in the body of a message to [email protected]
More majordomo info at http://vger.kernel.org/majordomo-info.html--
Milosz Tanski
CTO
10 East 53rd Street, 37th floor
New York, NY 10022p: 646-253-9055
e: [email protected]
Signature · 7 lines
Milosz Tanski CTO 10 East 53rd Street, 37th floor New York, NY 10022 p: 646-253-9055 e: [email protected]
Quoted 12 lines · Milosz Tanski
Sage,
I think the incrementing version counter on the whole is a neater solution then using size and mtime. If nothing else it's more explicit in the the read cache version. With what you suggested plus additional changes to the open code (where the cookie gets created) the write-through scenario should be correct.
Sadly, my understanding of the MDS protocol is still not great. So when doing this in the first place I erred on the side of using what was already in place.
Quoted 6 lines
In a kind of un-related question. Is there a debug hook in the kclient (or MDS for that matter) to dump the current file inodes (names) with issues caps and to which hosts. This would be very helpful for debugging, since from time to time I see a one of the clients get stuck in getattr (via mdsc debug log).
"ceph mds tell \* dumpcache" dump the mds cache to a file. the dump file contains caps information.
Regards
Yan, Zheng
Quoted 141 lines
Thanks,
- MiloszQuoted 45 lines · Sage Weil
Quoted 13 lines · Milosz Tanski
Li,
First, sorry for the late reply on this.
Currently fscache is only supported for files that are open in read only mode. I originally was going to let fscache cache in the write path as well as long as the file was open in with
O_LAZY. I abandoned that idea. When a user opens the file inO_LAZYwe can cache things locally with the assumption that the user will care of the synchronization in some other manner. There is no way of invalidating a subset of the pages in object cached by fscache, there is no way we can makeO_LAZYwork well.Quoted 4 lines
The
ceph_readpage_to_fscache()in writepage has no effect and it should be removed.ceph_readpage_to_fscache()callscache_valid()to see if it should perform the page save, and since the file can't have a CACHE cap at the point in time it doesn't do it.(Hmm, Dusting off my understanding of fscache and reading
fs/ceph/cache.c; watch out!) It looks likecache_validisstatic inline int
cache_valid(struct ceph_inode_info *ci){ return ((ceph_caps_issued(ci) & CEPH_CAP_FILE_CACHE) && (ci->i_fscache_gen == ci->i_rdcache_gen)); }and in the
FILE_EXCLcase, the MDS will issue CACHE|BUFFER caps. But I think the aux key (size+mtime) will prevent any use of the cache as soon as the first write happens and mtime changes, right?I think that in order to make this work, we need to fix/create a
file_version(or something similar) field in the (mds)inode_tto have some useful value. I.e., increment it any time- a different client/writer comes along
- a file is modified by the mds (e.g., truncated or recovered)but allow it to otherwise remain the same as long as only a single client is working with the file exclusively. This will be more precise than the (size, mtime) check that is currently used, and would remain valid when a single client opens the same file for exclusive read/write multiple times but there are no other intervening changes.
Quoted 76 lines
Milosz, if that were in place, is there any reason not to wire up writepage and allow the fscache to be used write-through?
sage
Quoted 68 lines
Thanks,
- MiloszQuoted 46 lines · Li Wang
Currently, the pages in fscache only are updated in
writepage()path, add the process inwritepages().Signed-off-by Min Chen <[email protected]>Signed-off-by Li Wang <[email protected]>Signed-off-by Yunchuan Wen <[email protected]>--- fs/ceph/addr.c | 8 +++++--- 1 file changed, 5 insertions(+), 3 deletions(-)fs/ceph/addr.c +5
-4@@ -746,7 +746,7 @@ 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 +894,6 @@ 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 &&@@ -924,7 +923,10 @@ 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_readpage_to_fscache (inode, page); + } pages =NULL ;/* request message now owns the pages array */ pool =NULL ; --1.7.9.5
Quoted 4 lines
--
To unsubscribe from this list: send the line "unsubscribe ceph-devel" in the body of a message to [email protected]
More majordomo info at http://vger.kernel.org/majordomo-info.html--
Milosz Tanski
CTO
10 East 53rd Street, 37th floor
New York, NY 10022p: 646-253-9055
e: [email protected]--
Milosz Tanski
CTO
10 East 53rd Street, 37th floor
New York, NY 10022p: 646-253-9055
e: [email protected]
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to [email protected]
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/
Yan and Sage,
I've ran into this issue again on my test cluster. The client hangs all requests for a particular inode, I did a dump cache to see what's going... but I don't understand to enough to be able to read this line well enough.
Can you guys help me read this, so I can further track down and hopefully fix this issue.
[inode 10000346eed [2,head] /petabucket/beta/17511b3d12466609785b6a0e34597431721d177240371c0a1a4e347a1605381b/advertiser_id.dict auth v214 ap=5+0 dirtyparent s=925 n(v0 b925 1=1+0) (ifile sync->mix)
(iversion lock) cr={59947=0-4194304@1} caps={59947=pAsLsXsFr/pAsxXsxFxwb@26,60001=pAsLsXsFr/-@1,60655=pAsLsXsFr/pAsLsXsFscr/pFscr@36} | ptrwaiter=0 request=4 lock=1 caps=1 dirtyparent=1 dirty=1 waiter=1
authpin=1 0x17dd6b70]
root@bnode-16a1ed7d:~# cat /sys/kernel/debug/ceph/e23a1bfc-8328-46bf-bc59-1209df3f5434.client60655/mdsc 15659 mds0 getattr #10000346eed 15679 mds0 getattr #10000346eed 15710 mds0 getattr #10000346eed 15922 mds0 getattr #10000346eed
Quoted 167 lines · Zheng
Quoted 12 lines · Milosz Tanski
Sage,
I think the incrementing version counter on the whole is a neater solution then using size and mtime. If nothing else it's more explicit in the the read cache version. With what you suggested plus additional changes to the open code (where the cookie gets created) the write-through scenario should be correct.
Sadly, my understanding of the MDS protocol is still not great. So when doing this in the first place I erred on the side of using what was already in place.
Quoted 6 lines
In a kind of un-related question. Is there a debug hook in the kclient (or MDS for that matter) to dump the current file inodes (names) with issues caps and to which hosts. This would be very helpful for debugging, since from time to time I see a one of the clients get stuck in getattr (via mdsc debug log).
"ceph mds tell \* dumpcache" dump the mds cache to a file. the dump file contains caps information.
Regards
Yan, ZhengQuoted 141 lines
Thanks,
- MiloszQuoted 45 lines · Sage Weil
Quoted 13 lines · Milosz Tanski
Li,
First, sorry for the late reply on this.
Currently fscache is only supported for files that are open in read only mode. I originally was going to let fscache cache in the write path as well as long as the file was open in with
O_LAZY. I abandoned that idea. When a user opens the file inO_LAZYwe can cache things locally with the assumption that the user will care of the synchronization in some other manner. There is no way of invalidating a subset of the pages in object cached by fscache, there is no way we can makeO_LAZYwork well.Quoted 4 lines
The
ceph_readpage_to_fscache()in writepage has no effect and it should be removed.ceph_readpage_to_fscache()callscache_valid()to see if it should perform the page save, and since the file can't have a CACHE cap at the point in time it doesn't do it.(Hmm, Dusting off my understanding of fscache and reading
fs/ceph/cache.c; watch out!) It looks likecache_validisstatic inline int
cache_valid(struct ceph_inode_info *ci){ return ((ceph_caps_issued(ci) & CEPH_CAP_FILE_CACHE) && (ci->i_fscache_gen == ci->i_rdcache_gen)); }and in the
FILE_EXCLcase, the MDS will issue CACHE|BUFFER caps. But I think the aux key (size+mtime) will prevent any use of the cache as soon as the first write happens and mtime changes, right?I think that in order to make this work, we need to fix/create a
file_version(or something similar) field in the (mds)inode_tto have some useful value. I.e., increment it any time- a different client/writer comes along
- a file is modified by the mds (e.g., truncated or recovered)but allow it to otherwise remain the same as long as only a single client is working with the file exclusively. This will be more precise than the (size, mtime) check that is currently used, and would remain valid when a single client opens the same file for exclusive read/write multiple times but there are no other intervening changes.
Quoted 76 lines
Milosz, if that were in place, is there any reason not to wire up writepage and allow the fscache to be used write-through?
sage
Quoted 68 lines
Thanks,
- MiloszQuoted 46 lines · Li Wang
Currently, the pages in fscache only are updated in
writepage()path, add the process inwritepages().Signed-off-by Min Chen <[email protected]>Signed-off-by Li Wang <[email protected]>Signed-off-by Yunchuan Wen <[email protected]>--- fs/ceph/addr.c | 8 +++++--- 1 file changed, 5 insertions(+), 3 deletions(-)fs/ceph/addr.c +5
-4@@ -746,7 +746,7 @@ 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 +894,6 @@ 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 &&@@ -924,7 +923,10 @@ 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_readpage_to_fscache (inode, page); + } pages =NULL ;/* request message now owns the pages array */ pool =NULL ; --1.7.9.5
Quoted 4 lines
--
To unsubscribe from this list: send the line "unsubscribe ceph-devel" in the body of a message to [email protected]
More majordomo info at http://vger.kernel.org/majordomo-info.html--
Milosz Tanski
CTO
10 East 53rd Street, 37th floor
New York, NY 10022p: 646-253-9055
e: [email protected]--
Milosz Tanski
CTO
10 East 53rd Street, 37th floor
New York, NY 10022p: 646-253-9055
e: [email protected]
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to [email protected]
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/
Signature · 7 lines
Milosz Tanski CTO 10 East 53rd Street, 37th floor New York, NY 10022 p: 646-253-9055 e: [email protected]
Quoted 18 lines · Milosz Tanski
Yan and Sage,
I've ran into this issue again on my test cluster. The client hangs all requests for a particular inode, I did a dump cache to see what's going... but I don't understand to enough to be able to read this line well enough.
Can you guys help me read this, so I can further track down and hopefully fix this issue.
[inode
10000346eed[2,head] /petabucket/beta/17511b3d12466609785b6a0e34597431721d177240371c0a1a4e347a1605381b/advertiser_id.dict auth v214 ap=5+0 dirtyparent s=925 n(v0 b925 1=1+0) (ifilesync->mix)
(iversion lock) cr={59947=0-4194304@1} caps={59947=pAsLsXsFr/pAsxXsxFxwb@26,60001=pAsLsXsFr/-@1,60655=pAsLsXsFr/pAsLsXsFscr/pFscr@36} | ptrwaiter=0 request=4 lock=1 caps=1 dirtyparent=1 dirty=1 waiter=1
authpin=10x17dd6b70]
Quoted 6 lines
root@bnode-16a1ed7d:~# cat /sys/kernel/debug/ceph/e23a1bfc-8328-46bf-bc59-1209df3f5434.client60655/mdsc 15659 mds0 getattr #10000346eed 15679 mds0 getattr #10000346eed 15710 mds0 getattr #10000346eed 15922 mds0 getattr #10000346eed
which kernel do you use? is there any blocked process (echo w > /proc/sysrq-trigger) on client.60655 ? 3.12 kernel contains few fixes for similar hang.
Regards
Yan, Zheng
Quoted 180 lines · Zheng
Quoted 20 lines · Milosz Tanski
Quoted 12 lines
Sage,
I think the incrementing version counter on the whole is a neater solution then using size and mtime. If nothing else it's more explicit in the the read cache version. With what you suggested plus additional changes to the open code (where the cookie gets created) the write-through scenario should be correct.
Sadly, my understanding of the MDS protocol is still not great. So when doing this in the first place I erred on the side of using what was already in place.
Quoted 6 lines
In a kind of un-related question. Is there a debug hook in the kclient (or MDS for that matter) to dump the current file inodes (names) with issues caps and to which hosts. This would be very helpful for debugging, since from time to time I see a one of the clients get stuck in getattr (via mdsc debug log).
Quoted 147 lines
"ceph mds tell \* dumpcache" dump the mds cache to a file. the dump file contains caps information.
Regards
Yan, ZhengQuoted 141 lines
Thanks,
- MiloszQuoted 45 lines · Sage Weil
Quoted 13 lines · Milosz Tanski
Li,
First, sorry for the late reply on this.
Currently fscache is only supported for files that are open in read only mode. I originally was going to let fscache cache in the write path as well as long as the file was open in with
O_LAZY. I abandoned that idea. When a user opens the file inO_LAZYwe can cache things locally with the assumption that the user will care of the synchronization in some other manner. There is no way of invalidating a subset of the pages in object cached by fscache, there is no way we can makeO_LAZYwork well.Quoted 4 lines
The
ceph_readpage_to_fscache()in writepage has no effect and it should be removed.ceph_readpage_to_fscache()callscache_valid()to see if it should perform the page save, and since the file can't have a CACHE cap at the point in time it doesn't do it.(Hmm, Dusting off my understanding of fscache and reading
fs/ceph/cache.c; watch out!) It looks likecache_validisstatic inline int
cache_valid(struct ceph_inode_info *ci){ return ((ceph_caps_issued(ci) & CEPH_CAP_FILE_CACHE) && (ci->i_fscache_gen == ci->i_rdcache_gen)); }and in the
FILE_EXCLcase, the MDS will issue CACHE|BUFFER caps. But I think the aux key (size+mtime) will prevent any use of the cache as soon as the first write happens and mtime changes, right?I think that in order to make this work, we need to fix/create a
file_version(or something similar) field in the (mds)inode_tto have some useful value. I.e., increment it any time- a different client/writer comes along
- a file is modified by the mds (e.g., truncated or recovered)but allow it to otherwise remain the same as long as only a single client is working with the file exclusively. This will be more precise than the (size, mtime) check that is currently used, and would remain valid when a single client opens the same file for exclusive read/write multiple times but there are no other intervening changes.
Quoted 76 lines
Milosz, if that were in place, is there any reason not to wire up writepage and allow the fscache to be used write-through?
sage
Quoted 68 lines
Thanks,
- MiloszQuoted 46 lines · Li Wang
Currently, the pages in fscache only are updated in
writepage()path, add the process inwritepages().Signed-off-by Min Chen <[email protected]>Signed-off-by Li Wang <[email protected]>Signed-off-by Yunchuan Wen <[email protected]>--- fs/ceph/addr.c | 8 +++++--- 1 file changed, 5 insertions(+), 3 deletions(-)fs/ceph/addr.c +5
-4@@ -746,7 +746,7 @@ 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 +894,6 @@ 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 &&@@ -924,7 +923,10 @@ 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_readpage_to_fscache (inode, page); + } pages =NULL ;/* request message now owns the pages array */ pool =NULL ; --1.7.9.5
Quoted 4 lines
--
To unsubscribe from this list: send the line "unsubscribe ceph-devel" in the body of a message to [email protected]
More majordomo info at http://vger.kernel.org/majordomo-info.html--
Milosz Tanski
CTO
10 East 53rd Street, 37th floor
New York, NY 10022p: 646-253-9055
e: [email protected]--
Milosz Tanski
CTO
10 East 53rd Street, 37th floor
New York, NY 10022p: 646-253-9055
e: [email protected]
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to [email protected]
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/--
Milosz Tanski
CTO
10 East 53rd Street, 37th floor
New York, NY 10022p: 646-253-9055
e: [email protected]
Yan,
I'll use this trick next time around. I did dump the kernel stacks for my process. 4 threads were blocked on SYS_newfstat (and the mds request further up the stack).
I ended up restarting MDS after a few hours of trying to track it down. It resolved it self following that.
This machine is running a pretty recent kernel on there -- 3.12 + merged testing ceph -- the other machines are running a slightly older 3.12-rc. I have observed the issue previously on random nodes infrequently for months (maybe once a week or two).
Thanks again,
- Milosz
Quoted 213 lines · Zheng
Quoted 18 lines · Milosz Tanski
Yan and Sage,
I've ran into this issue again on my test cluster. The client hangs all requests for a particular inode, I did a dump cache to see what's going... but I don't understand to enough to be able to read this line well enough.
Can you guys help me read this, so I can further track down and hopefully fix this issue.
[inode
10000346eed[2,head] /petabucket/beta/17511b3d12466609785b6a0e34597431721d177240371c0a1a4e347a1605381b/advertiser_id.dict auth v214 ap=5+0 dirtyparent s=925 n(v0 b925 1=1+0) (ifilesync->mix)
(iversion lock) cr={59947=0-4194304@1} caps={59947=pAsLsXsFr/pAsxXsxFxwb@26,60001=pAsLsXsFr/-@1,60655=pAsLsXsFr/pAsLsXsFscr/pFscr@36} | ptrwaiter=0 request=4 lock=1 caps=1 dirtyparent=1 dirty=1 waiter=1
authpin=10x17dd6b70]Quoted 6 lines
root@bnode-16a1ed7d:~# cat /sys/kernel/debug/ceph/e23a1bfc-8328-46bf-bc59-1209df3f5434.client60655/mdsc 15659 mds0 getattr #10000346eed 15679 mds0 getattr #10000346eed 15710 mds0 getattr #10000346eed 15922 mds0 getattr #10000346eedwhich kernel do you use? is there any blocked process (echo w >
/proc/sysrq-trigger) on client.60655 ? 3.12 kernel contains few fixes for similar hang.Regards
Yan, ZhengQuoted 180 lines · Zheng
Quoted 20 lines · Milosz Tanski
Quoted 12 lines
Sage,
I think the incrementing version counter on the whole is a neater solution then using size and mtime. If nothing else it's more explicit in the the read cache version. With what you suggested plus additional changes to the open code (where the cookie gets created) the write-through scenario should be correct.
Sadly, my understanding of the MDS protocol is still not great. So when doing this in the first place I erred on the side of using what was already in place.
Quoted 6 lines
In a kind of un-related question. Is there a debug hook in the kclient (or MDS for that matter) to dump the current file inodes (names) with issues caps and to which hosts. This would be very helpful for debugging, since from time to time I see a one of the clients get stuck in getattr (via mdsc debug log).
Quoted 147 lines
"ceph mds tell \* dumpcache" dump the mds cache to a file. the dump file contains caps information.
Regards
Yan, ZhengQuoted 141 lines
Thanks,
- MiloszQuoted 45 lines · Sage Weil
Quoted 13 lines · Milosz Tanski
Li,
First, sorry for the late reply on this.
Currently fscache is only supported for files that are open in read only mode. I originally was going to let fscache cache in the write path as well as long as the file was open in with
O_LAZY. I abandoned that idea. When a user opens the file inO_LAZYwe can cache things locally with the assumption that the user will care of the synchronization in some other manner. There is no way of invalidating a subset of the pages in object cached by fscache, there is no way we can makeO_LAZYwork well.Quoted 4 lines
The
ceph_readpage_to_fscache()in writepage has no effect and it should be removed.ceph_readpage_to_fscache()callscache_valid()to see if it should perform the page save, and since the file can't have a CACHE cap at the point in time it doesn't do it.(Hmm, Dusting off my understanding of fscache and reading
fs/ceph/cache.c; watch out!) It looks likecache_validisstatic inline int
cache_valid(struct ceph_inode_info *ci){ return ((ceph_caps_issued(ci) & CEPH_CAP_FILE_CACHE) && (ci->i_fscache_gen == ci->i_rdcache_gen)); }and in the
FILE_EXCLcase, the MDS will issue CACHE|BUFFER caps. But I think the aux key (size+mtime) will prevent any use of the cache as soon as the first write happens and mtime changes, right?I think that in order to make this work, we need to fix/create a
file_version(or something similar) field in the (mds)inode_tto have some useful value. I.e., increment it any time- a different client/writer comes along
- a file is modified by the mds (e.g., truncated or recovered)but allow it to otherwise remain the same as long as only a single client is working with the file exclusively. This will be more precise than the (size, mtime) check that is currently used, and would remain valid when a single client opens the same file for exclusive read/write multiple times but there are no other intervening changes.
Quoted 76 lines
Milosz, if that were in place, is there any reason not to wire up writepage and allow the fscache to be used write-through?
sage
Quoted 68 lines
Thanks,
- MiloszQuoted 46 lines · Li Wang
Currently, the pages in fscache only are updated in
writepage()path, add the process inwritepages().Signed-off-by Min Chen <[email protected]>Signed-off-by Li Wang <[email protected]>Signed-off-by Yunchuan Wen <[email protected]>--- fs/ceph/addr.c | 8 +++++--- 1 file changed, 5 insertions(+), 3 deletions(-)fs/ceph/addr.c +5
-4@@ -746,7 +746,7 @@ 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 +894,6 @@ 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 &&@@ -924,7 +923,10 @@ 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_readpage_to_fscache (inode, page); + } pages =NULL ;/* request message now owns the pages array */ pool =NULL ; --1.7.9.5
Quoted 4 lines
--
To unsubscribe from this list: send the line "unsubscribe ceph-devel" in the body of a message to [email protected]
More majordomo info at http://vger.kernel.org/majordomo-info.html--
Milosz Tanski
CTO
10 East 53rd Street, 37th floor
New York, NY 10022p: 646-253-9055
e: [email protected]--
Milosz Tanski
CTO
10 East 53rd Street, 37th floor
New York, NY 10022p: 646-253-9055
e: [email protected]
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to [email protected]
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/--
Milosz Tanski
CTO
10 East 53rd Street, 37th floor
New York, NY 10022p: 646-253-9055
e: [email protected]
Signature · 7 lines
Milosz Tanski CTO 10 East 53rd Street, 37th floor New York, NY 10022 p: 646-253-9055 e: [email protected]
A message that is not in the archive (sent off-list or not yet received): CANP1eJHETFZ+oBoz+XP4hnOc20UNECNoW5jfY-C-6q8Z7ExgEg@mail.gmail.com
A message that is not in the archive (sent off-list or not yet received): [email protected]
Quoted 11 lines · Sage Weil
Quoted 7 lines · Milosz Tanski
Sage,
I think the incrementing version counter on the whole is a neater solution then using size and mtime. If nothing else it's more explicit in the the read cache version. With what you suggested plus additional changes to the open code (where the cookie gets created) the write-through scenario should be correct.
Great to hear. I don't think it will be much work on the MDS; it'll just take a bit of care to catch the relevant cases.
Just let me know any way I can help that along.
Quoted 13 lines
Quoted 4 lines
Sadly, my understanding of the MDS protocol is still not great. So when doing this in the first place I erred on the side of using what was already in place.
Of course!
Quoted 5 lines
In a kind of un-related question. Is there a debug hook in the kclient (or MDS for that matter) to dump the current file inodes (names) with issues caps and to which hosts. This would be very helpful for debugging, since from time to time I see a one of the clients get stuck in getattr (via mdsc debug log).
Quoted 5 lines
cat
/sys/kernel/debug/ceph/*/mdscwill show requests in flight, but nothing right now shows the caps. That would be super valuable, thoguh. I'm not sure if there is a conveient way to iterate over the
kmem_cache.. :/
That's how I can confirm that I am experiencing the issue. I cat mdsc and I end up seeing one (or more) getattr requests that sit there stubbornly. If I figure out which file it is on the filesystem and try to access its metadata on the filesystem those requests hang in a similar fashion (no matter what node it is).
I've been trying to figure out the root cause of it for a while now. And so far I can't tell if one of the applications misbehaving, MDS or kclient. To make matters worse it can take weeks to encounter it again. The dumpcache feature that Yan pointed me at should give me a new tool in debugging it.
Being able to see the same same kind of view from the client side would be supper helpful in debugging. I'll take a look at the code to see if we can iterate through the inodes for a mount point in a strait-forward manner to make it happen.
Quoted 142 lines
s
Quoted 139 lines
Thanks,
- MiloszQuoted 45 lines · Sage Weil
Quoted 13 lines · Milosz Tanski
Li,
First, sorry for the late reply on this.
Currently fscache is only supported for files that are open in read only mode. I originally was going to let fscache cache in the write path as well as long as the file was open in with
O_LAZY. I abandoned that idea. When a user opens the file inO_LAZYwe can cache things locally with the assumption that the user will care of the synchronization in some other manner. There is no way of invalidating a subset of the pages in object cached by fscache, there is no way we can makeO_LAZYwork well.Quoted 4 lines
The
ceph_readpage_to_fscache()in writepage has no effect and it should be removed.ceph_readpage_to_fscache()callscache_valid()to see if it should perform the page save, and since the file can't have a CACHE cap at the point in time it doesn't do it.(Hmm, Dusting off my understanding of fscache and reading
fs/ceph/cache.c; watch out!) It looks likecache_validisstatic inline int
cache_valid(struct ceph_inode_info *ci){ return ((ceph_caps_issued(ci) & CEPH_CAP_FILE_CACHE) && (ci->i_fscache_gen == ci->i_rdcache_gen)); }and in the
FILE_EXCLcase, the MDS will issue CACHE|BUFFER caps. But I think the aux key (size+mtime) will prevent any use of the cache as soon as the first write happens and mtime changes, right?I think that in order to make this work, we need to fix/create a
file_version(or something similar) field in the (mds)inode_tto have some useful value. I.e., increment it any time- a different client/writer comes along
- a file is modified by the mds (e.g., truncated or recovered)but allow it to otherwise remain the same as long as only a single client is working with the file exclusively. This will be more precise than the (size, mtime) check that is currently used, and would remain valid when a single client opens the same file for exclusive read/write multiple times but there are no other intervening changes.
Quoted 76 lines
Milosz, if that were in place, is there any reason not to wire up writepage and allow the fscache to be used write-through?
sage
Quoted 68 lines
Thanks,
- MiloszQuoted 46 lines · Li Wang
Currently, the pages in fscache only are updated in
writepage()path, add the process inwritepages().Signed-off-by Min Chen <[email protected]>Signed-off-by Li Wang <[email protected]>Signed-off-by Yunchuan Wen <[email protected]>--- fs/ceph/addr.c | 8 +++++--- 1 file changed, 5 insertions(+), 3 deletions(-)fs/ceph/addr.c +5
-4@@ -746,7 +746,7 @@ 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 +894,6 @@ 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 &&@@ -924,7 +923,10 @@ 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_readpage_to_fscache (inode, page); + } pages =NULL ;/* request message now owns the pages array */ pool =NULL ; --1.7.9.5
Quoted 4 lines
--
To unsubscribe from this list: send the line "unsubscribe ceph-devel" in the body of a message to [email protected]
More majordomo info at http://vger.kernel.org/majordomo-info.html--
Milosz Tanski
CTO
10 East 53rd Street, 37th floor
New York, NY 10022p: 646-253-9055
e: [email protected]--
Milosz Tanski
CTO
10 East 53rd Street, 37th floor
New York, NY 10022p: 646-253-9055
e: [email protected]
Signature · 7 lines
Milosz Tanski CTO 10 East 53rd Street, 37th floor New York, NY 10022 p: 646-253-9055 e: [email protected]