Gang He
2015-Dec-17 15:08 UTC
[Ocfs2-devel] The root cause analysis about buffer read getting starvation
Hello Mark and all,
In the past days, I and Eric were looking at a customer issue, the customer is
complaining that buffer reading sometimes lasts too much time ( 1 - 10 seconds)
in case reading/writing the same file from different nodes concurrently, some
day ago I sent a mail to the list for some discussions, you can read some
details via the link
https://oss.oracle.com/pipermail/ocfs2-devel/2015-December/011389.html.
But, this problem does not happen under SLES10 (sp1 - sp4), the customer
upgraded his Linux OS to SLES11(sp3 or sp4), the problem happened, this is why
the customer complains, he hope we can give a investigation, to see how to make
OCFS2 buffer reading/writing behavior be consistent with SLES10.
According to our code reviewing and some testings, we found that the root cause
to let buffer read get starvation.
The suspicious code in aops.c
274 static int ocfs2_readpage(struct file *file, struct page *page)
275 {
276 struct inode *inode = page->mapping->host;
277 struct ocfs2_inode_info *oi = OCFS2_I(inode);
278 loff_t start = (loff_t)page->index << PAGE_CACHE_SHIFT;
279 int ret, unlock = 1;
280 long delta;
281 struct timespec t_enter, t_mid1, t_mid2, t_exit;
282
283 trace_ocfs2_readpage((unsigned long long)oi->ip_blkno,
284 (page ? page->index : 0));
285
286 ret = ocfs2_inode_lock_with_page(inode, NULL, 0, page); <<=
here, using nonblock way to get lock will bring many times retry, spend too much
time
287 if (ret != 0) {
288 if (ret == AOP_TRUNCATED_PAGE)
289 unlock = 0;
290 mlog_errno(ret);
291 goto out;
292 }
293
294 if (down_read_trylock(&oi->ip_alloc_sem) == 0) { <<=
here, the same problem with above
295 /*
296 * Unlock the page and cycle ip_alloc_sem so that we
don't
297 * busyloop waiting for ip_alloc_sem to unlock
298 */
299 ret = AOP_TRUNCATED_PAGE;
300 unlock_page(page);
301 unlock = 0;
302 down_read(&oi->ip_alloc_sem);
303 up_read(&oi->ip_alloc_sem);
304 goto out_inode_unlock;
305 }
As you can see, using nonblock way to get lock will bring many time retry, spend
too much time.
We can't modify the code to using block way to get the lock, as this will
bring a dead lock.
Actually, we did some testing when trying to use block way to get the lock here,
the deadlock problems were encountered.
But, in SLES10 source code, there is not any using nonblock way to get lock in
buffer reading/writing, this is why buffer reading/writing are very fair to get
IO when reading/writing the same file from multiple nodes.
Why the dead locks happen on SLES11? you can see the source code, there are some
code change, especially inode alloc_sem lock.
On SLES11, to get inode alloc_sem lock is moved into ocfs2_readpage and
ocfs2_write_begin, why we need to do that? this will let us to bring a dead lock
factor, to avoid the dead locks, we will have to use nonblocking way to get some
locks in ocfs2_readpage, the result will let buffer reading be unfair to get IO.
and that, to avoid CPU busy loop, add some code to get the lock with block way
in case can't get a lock in nonblock way, waste too much time also.
Finally, I want to discuss with your guys, how to fix this issue? could we move
the inode alloc_sem lock back to ocfs2_file_aio_read/ocfs2_file_aio_write?
we can get the inode alloc_sem lock before calling into
ocfs2_readpage/ocfs2_write_begin, just like SLES10.
Since I have not enough background behind these code changes, hope you can give
some comments.
Thanks a lot.
Gang
Eric Ren
2015-Dec-18 09:09 UTC
[Ocfs2-devel] The root cause analysis about buffer read getting starvation
Hi all, On Thu, Dec 17, 2015 at 08:08:42AM -0700, He Gang wrote:> Hello Mark and all, > In the past days, I and Eric were looking at a customer issue, the customer is complaining that buffer reading sometimes lasts too much time ( 1 - 10 seconds) in case reading/writing the same file from different nodes concurrently, some day ago I sent a mail to the list for some discussions, you can read some details via the link https://oss.oracle.com/pipermail/ocfs2-devel/2015-December/011389.html. > But, this problem does not happen under SLES10 (sp1 - sp4), the customer upgraded his Linux OS to SLES11(sp3 or sp4), the problem happened, this is why the customer complains, he hope we can give a investigation, to see how to make OCFS2 buffer reading/writing behavior be consistent with SLES10. > According to our code reviewing and some testings, we found that the root cause to let buffer read get starvation. > The suspicious code in aops.c > 274 static int ocfs2_readpage(struct file *file, struct page *page) > 275 { > 276 struct inode *inode = page->mapping->host; > 277 struct ocfs2_inode_info *oi = OCFS2_I(inode); > 278 loff_t start = (loff_t)page->index << PAGE_CACHE_SHIFT; > 279 int ret, unlock = 1; > 280 long delta; > 281 struct timespec t_enter, t_mid1, t_mid2, t_exit; > 282 > 283 trace_ocfs2_readpage((unsigned long long)oi->ip_blkno, > 284 (page ? page->index : 0)); > 285 > 286 ret = ocfs2_inode_lock_with_page(inode, NULL, 0, page); <<= here, using nonblock way to get lock will bring many times retry, spend too much time > 287 if (ret != 0) { > 288 if (ret == AOP_TRUNCATED_PAGE) > 289 unlock = 0; > 290 mlog_errno(ret); > 291 goto out; > 292 } > 293 > 294 if (down_read_trylock(&oi->ip_alloc_sem) == 0) { <<= here, the same problem with above > 295 /* > 296 * Unlock the page and cycle ip_alloc_sem so that we don't > 297 * busyloop waiting for ip_alloc_sem to unlock > 298 */ > 299 ret = AOP_TRUNCATED_PAGE; > 300 unlock_page(page); > 301 unlock = 0; > 302 down_read(&oi->ip_alloc_sem); > 303 up_read(&oi->ip_alloc_sem); > 304 goto out_inode_unlock; > 305 } > > > As you can see, using nonblock way to get lock will bring many time retry, spend too much time. > We can't modify the code to using block way to get the lock, as this will bring a dead lock. > Actually, we did some testing when trying to use block way to get the lock here, the deadlock problems were encountered. > But, in SLES10 source code, there is not any using nonblock way to get lock in buffer reading/writing, this is why buffer reading/writing are very fair to get IO when reading/writing the same file from multiple nodes.SLES10 with kernel version about 2.6.16.x, used blocking way, i.e. down_read(), wich has the potential deaklock between page lock / ip_alloc_sem when one node get the cluster lock and does writing and reading on same file on it. This deadlock was fixed by this commit: --- commit e9dfc0b2bc42761410e8db6c252c6c5889e178b8 Author: Mark Fasheh <mark.fasheh at oracle.com> Date: Mon May 14 11:38:51 2007 -0700 ocfs2: trylock in ocfs2_readpage() Similarly to the page lock / cluster lock inversion in ocfs2_readpage, we can deadlock on ip_alloc_sem. We can down_read_trylock() instead and just return AOP_TRUNCATED_PAGE if the operation fails. Signed-off-by: Mark Fasheh <mark.fasheh at oracle.com> diff --git a/fs/ocfs2/aops.c b/fs/ocfs2/aops.c index 8e7cafb..3030670 100644 --- a/fs/ocfs2/aops.c +++ b/fs/ocfs2/aops.c @@ -222,7 +222,10 @@ static int ocfs2_readpage(struct file *file, struct page *page) goto out; } - down_read(&OCFS2_I(inode)->ip_alloc_sem); + if (down_read_trylock(&OCFS2_I(inode)->ip_alloc_sem) == 0) { + ret = AOP_TRUNCATED_PAGE; + goto out_meta_unlock; + } /* * i_size might have just been updated as we grabed the meta lock. We @@ -258,6 +261,7 @@ static int ocfs2_readpage(struct file *file, struct page *page) ocfs2_data_unlock(inode, 0); out_alloc: up_read(&OCFS2_I(inode)->ip_alloc_sem); +out_meta_unlock: ocfs2_meta_unlock(inode, 0); out: if (unlock) --- But somehow with this patch, performance in the scenario become very bad. I don't how this could happen? because the reading node just has only one thread reading the shared file, then down_read_trylock() should always get ip_alloc_sem successfully, right? if not, who else may race ip_alloc_sem? Thanks, Eric> Why the dead locks happen on SLES11? you can see the source code, there are some code change, especially inode alloc_sem lock. > On SLES11, to get inode alloc_sem lock is moved into ocfs2_readpage and ocfs2_write_begin, why we need to do that? this will let us to bring a dead lock factor, to avoid the dead locks, we will have to use nonblocking way to get some locks in ocfs2_readpage, the result will let buffer reading be unfair to get IO. and that, to avoid CPU busy loop, add some code to get the lock with block way in case can't get a lock in nonblock way, waste too much time also. > Finally, I want to discuss with your guys, how to fix this issue? could we move the inode alloc_sem lock back to ocfs2_file_aio_read/ocfs2_file_aio_write? > we can get the inode alloc_sem lock before calling into ocfs2_readpage/ocfs2_write_begin, just like SLES10. > Since I have not enough background behind these code changes, hope you can give some comments. > > Thanks a lot. > Gang > > > > _______________________________________________ > Ocfs2-devel mailing list > Ocfs2-devel at oss.oracle.com > https://oss.oracle.com/mailman/listinfo/ocfs2-devel >
Eric Ren
2015-Dec-25 09:05 UTC
[Ocfs2-devel] The root cause analysis about buffer read getting starvation
Hi Mark and Gang,> In the past days, I and Eric were looking at a customer issue, the customer is complaining that buffer reading sometimes lasts too much time ( 1 - 10 seconds) in case reading/writing the same file from different nodes concurrently, some day ago I sent a mail to the list for some discussions, you can read some details via the link https://oss.oracle.com/pipermail/ocfs2-devel/2015-December/011389.html. > But, this problem does not happen under SLES10 (sp1 - sp4), the customer upgraded his Linux OS to SLES11(sp3 or sp4), the problem happened, this is why the customer complains, he hope we can give a investigation, to see how to make OCFS2 buffer reading/writing behavior be consistent with SLES10.Sorry for the previous false testing report, which must mislead you somehow. Actually, sle10 sp4(not sure how about sp2/3) is free of the problem we're discussing, according to both code analysis and new more careful testing result. As Gang said, the cluster lock and local lock has changed a lot. The 2 biggest change in my head is: 1. for read, cluster lock is put into *_readpage(s), leaving the rest part of read() such as copying data from page cache to user space. This can result in page cache being truncated by downconvert thread on behalf of another node before copied to user space, especially in fierce contention of cluster lock. 2. The order of getting cluster lock and local lock is changed, in order to avoid deadlock, then the unblocked way to get cluster lock was introduced. Gang concern about this more. So I won't repeat it;-) Here I want to focus on the 1st case, because there's only one thread, i.e either reading or writing on each node in this case, so local lock may not matter too much, besides the lock inversion problem. OK, just some thoughts on 1st issue. Please correct me if anything wrong. 1. to be convinient for discussion, I'd like to give read() and write() path here: vfs_read() ocfs2_file_aio_read() ocfs2_inode_lock_atime ocfs2_inode_unlock() generic_file_aio_read() do_generic_file_read() page_cache_sync_readahead() or page_cache_async_readahead() ocfs2_readpages() ocfs2_inode_lock_full(NOBLOCK) down_read_trylock(ip_alloc_sem) ocfs2_readpage() /* continue to retry until to succeed */ ocfs2_inode_lock_with_page(NOBLOCK) down_read_trylock(ip_alloc_sem) ocfs2_inode_unlock PAGE OK: COPY PAGE TO USER SPACE vfs_write() ocfs2_file_aio_write() ocfs2_rw_lock(EX) generic_buffered_write() generic_perform_write() ocfs2_write_begin() ocfs2_inode_lock_full() dow_write(ip_alloc_sem) COPY FROM USER SPACE ocfs2_write_end() up_write(ip_alloc_sem) ocfs2_inode_unlock ocfs2_rw_unlock() 2. There is a race window to truncate page cache, making the effort of readahead and *_readpage stuff go to waste! For read(), the cluster lock just be held when doing readpage() stuff, to get data from disk to page cache. When the contention of cluster lock for same file is rather fierce, so it's likely EX lock from another node has been already queued. Once returning from readpage(), cluster lock will been released, which will kick up downconvert thread to work. Then ocfs2dc will truncate all cached pages: --- 3583 static int ocfs2_data_convert_worker(struct ocfs2_lock_res *lockres, 3584 int blocking) 3585 { ... 3616 sync_mapping_buffers(mapping); 3617 if (blocking == DLM_LOCK_EX) { 3618 truncate_inode_pages(mapping, 0); 3619 } else { .... --- 3. testing I'd like to elaborate how I tested and the result data if you have intersts. Not here, or this thead would extend too long. Overall, it looks that the more fiercer the contention is, the more unfair read() is treated... 4. kernel version info Linux sle10sp1 2.6.16.46-0.12-smp #1 SMP Thu May 17 14:00:09 UTC 2007 x86_64 x86_64 x86_64 GNU/Linux sle10sp4:~ # uname -a Linux sle10sp4 2.6.16.60-0.85.1-smp #1 SMP Thu Mar 17 11:45:06 UTC 2011 x86_64 x86_64 x86_64 GNU/Linux 5. readpage() on sle10 sp1 is like: --- 211 static int ocfs2_readpage(struct file *file, struct page *page) 212 { 213 int ret; 214 215 mlog_entry("(0x%p, %lu)\n", file, (page ? page->index : 0)); 216 217 ret = block_read_full_page(page, ocfs2_get_block); 218 219 mlog_exit(ret); 220 221 return ret; 222 } --- It's not long time for me on ocfs2, so I'm not really sure if I'm reporting stupid issue again ;-) Look forward your comments on this! Thanks, Eric> According to our code reviewing and some testings, we found that the root cause to let buffer read get starvation. > The suspicious code in aops.c > 274 static int ocfs2_readpage(struct file *file, struct page *page) > 275 { > 276 struct inode *inode = page->mapping->host; > 277 struct ocfs2_inode_info *oi = OCFS2_I(inode); > 278 loff_t start = (loff_t)page->index << PAGE_CACHE_SHIFT; > 279 int ret, unlock = 1; > 280 long delta; > 281 struct timespec t_enter, t_mid1, t_mid2, t_exit; > 282 > 283 trace_ocfs2_readpage((unsigned long long)oi->ip_blkno, > 284 (page ? page->index : 0)); > 285 > 286 ret = ocfs2_inode_lock_with_page(inode, NULL, 0, page); <<= here, using nonblock way to get lock will bring many times retry, spend too much time > 287 if (ret != 0) { > 288 if (ret == AOP_TRUNCATED_PAGE) > 289 unlock = 0; > 290 mlog_errno(ret); > 291 goto out; > 292 } > 293 > 294 if (down_read_trylock(&oi->ip_alloc_sem) == 0) { <<= here, the same problem with above > 295 /* > 296 * Unlock the page and cycle ip_alloc_sem so that we don't > 297 * busyloop waiting for ip_alloc_sem to unlock > 298 */ > 299 ret = AOP_TRUNCATED_PAGE; > 300 unlock_page(page); > 301 unlock = 0; > 302 down_read(&oi->ip_alloc_sem); > 303 up_read(&oi->ip_alloc_sem); > 304 goto out_inode_unlock; > 305 } > > > As you can see, using nonblock way to get lock will bring many time retry, spend too much time. > We can't modify the code to using block way to get the lock, as this will bring a dead lock. > Actually, we did some testing when trying to use block way to get the lock here, the deadlock problems were encountered. > But, in SLES10 source code, there is not any using nonblock way to get lock in buffer reading/writing, this is why buffer reading/writing are very fair to get IO when reading/writing the same file from multiple nodes. > Why the dead locks happen on SLES11? you can see the source code, there are some code change, especially inode alloc_sem lock. > On SLES11, to get inode alloc_sem lock is moved into ocfs2_readpage and ocfs2_write_begin, why we need to do that? this will let us to bring a dead lock factor, to avoid the dead locks, we will have to use nonblocking way to get some locks in ocfs2_readpage, the result will let buffer reading be unfair to get IO. and that, to avoid CPU busy loop, add some code to get the lock with block way in case can't get a lock in nonblock way, waste too much time also. > Finally, I want to discuss with your guys, how to fix this issue? could we move the inode alloc_sem lock back to ocfs2_file_aio_read/ocfs2_file_aio_write? > we can get the inode alloc_sem lock before calling into ocfs2_readpage/ocfs2_write_begin, just like SLES10. > Since I have not enough background behind these code changes, hope you can give some comments. > > Thanks a lot. > Gang > > > > _______________________________________________ > Ocfs2-devel mailing list > Ocfs2-devel at oss.oracle.com > https://oss.oracle.com/mailman/listinfo/ocfs2-devel >