multipathd: Don't keep starting TUR threads, if they always hang.
Benjamin Marzinski
maintainerRFC PATCH 1/1 multipathd: Don't keep starting TUR threads, if they always hang.
If tur thead hangs, multipathd was simply creating a new thread, and assuming that the old thread would get cleaned up eventually. I have seen a case recently where there were 26000 multipathd threads on a system, all stuck trying to send TUR commands to path devices. The root cause of the issue was a scsi kernel issue, but it shows that the way multipathd currently deals with stuck threads could use some refinement.
Now, when one tur thread hangs, multipathd will act as it did before. If a second one in a row hangs, multipathd will instead wait for it to complete before starting another thread. Once the thread completes, the count is reset.
--- libmultipath/checkers/tur.c | 23 +++++++++++++++++++++-- 1 file changed, 21 insertions(+), 2 deletions(-)
libmultipath/checkers/tur.c +21 -2
@@ -26,6 +26,7 @@#define TUR_CMD_LEN 6#define HEAVY_CHECK_COUNT 10 +#define MAX_NR_TIMEOUTS 1enum {MSG_TUR_RUNNING =CHECKER_FIRST_MSGID ,
@@ -54,6 +55,7 @@ struct tur_checker_context {int holders;/* uatomic access only */ int msgid;struct checker_context ctx; +unsigned int nr_timeouts; };int libcheck_init (struct checker * c)
@@ -358,8 +360,23 @@ int libcheck_check(struct checker * c) } }else {if (uatomic_read (&ct->holders) >1 ) { +/* The thread has been cancelled but hasn't quit. */ +if (ct->nr_timeouts ==MAX_NR_TIMEOUTS ) { +condlog (2 ,"%d:%d : waiting for stalled tur thread to finish" , +major (ct->devt),minor (ct->devt)); + ct->nr_timeouts++; + }/* -* The thread has been cancelled but hasn't quit. +* Don't start new threads until the last once has +* finished. +*/ +if (ct->nr_timeouts >MAX_NR_TIMEOUTS ) { + c->msgid =MSG_TUR_TIMEOUT ; +return PATH_TIMEOUT ; + } + ct->nr_timeouts++; +/* +* Start a new thread while the old one is stalled. * We have to prevent it from interfering with the new * thread. We create a new context and leave the old * one with the stale thread, hoping it will clean up
@@ -375,13 +392,15 @@ int libcheck_check(struct checker * c)*/ if (libcheck_init (c) !=0 )return PATH_UNCHECKED ; + ((struct tur_checker_context *)c->context)->nr_timeouts = ct->nr_timeouts;if (!uatomic_sub_return (&ct->holders,1 ))/* It did terminate, eventually */ cleanup_context (ct); ct = c->context; - } + }else + ct->nr_timeouts =0 ;/* Start new TUR checker */ pthread_mutex_lock (&ct->lock); tur_status = ct->state =PATH_PENDING ;
Signature · 6 lines
2.17.2 -- dm-devel mailing list [email protected] https://listman.redhat.com/mailman/listinfo/dm-devel
Quoted 8 lines · Benjamin Marzinski
If tur thead hangs, multipathd was simply creating a new thread, and assuming that the old thread would get cleaned up eventually. I have seen a case recently where there were 26000 multipathd threads on a system, all stuck trying to send TUR commands to path devices. The root
cause of the issue was a scsi kernel issue, but it shows that the way multipathd currently deals with stuck threads could use some refinement.
Alright, let's start the discussion about dead TUR threads again ;-) For my own records, here are links to some previous ML threads:
https://listman.redhat.com/archives/dm-devel/2018-September/036102.html
https://listman.redhat.com/archives/dm-devel/2018-October/036156.html
https://listman.redhat.com/archives/dm-devel/2018-October/036159.html
https://listman.redhat.com/archives/dm-devel/2018-October/036240.html
https://listman.redhat.com/archives/dm-devel/2018-October/036362.html
Have you assessed why the SCSI issues were causing indefinite hangs? Was it just the TUR command hanging forever in the kernel, or some other issue?
I have seen similar issues in the past; they were also related to SCSI problems, IIRC. I used to think that this is just a symptom: The "dead" threads are detached and consume a minimal amount of non-shared memory (they are a pain when analyzing a crash dump, though).
But I agree, just continuing to start new threads forever isn't optimal.
Quoted 19 lines
Now, when one tur thread hangs, multipathd will act as it did before. If a second one in a row hangs, multipathd will instead wait for it to
complete before starting another thread. Once the thread completes, the
count is reset.Signed-off-by Benjamin Marzinski <[email protected]>---
libmultipath/checkers/tur.c | 23 +++++++++++++++++++++--1 file changed, 21 insertions(+), 2 deletions(-)libmultipath/checkers/tur.c +0
-0@@ -26,6 +26,7 @@
Quoted 4 lines
#defineTUR_CMD_LEN6
#define HEAVY_CHECK_COUNT 10
+#defineMAX_NR_TIMEOUTS1
Why did you choose 1? Perhaps we should make a few more attempts?
Other than that, this looks ok to me (assuming you've tested with the code from bdf55d6, or something similar).
Martin
Quoted 55 lines
enum {
MSG_TUR_RUNNING =CHECKER_FIRST_MSGID,libmultipath/checkers/tur.c +0
-0@@ -54,6 +55,7 @@ struct tur_checker_context {int holders; /* uatomic access only */
int msgid;
structchecker_contextctx;
+ unsigned intnr_timeouts;
};
intlibcheck_init(struct checker * c)libmultipath/checkers/tur.c +1
-0@@ -358,8 +360,23 @@ int libcheck_check(struct checker * c) } } else { if (uatomic_read(&ct->holders) > 1) { + /* The thread has been cancelled but hasn'tquit. */
+ if (ct->nr_timeouts==MAX_NR_TIMEOUTS) { + condlog(2, "%d:%d : waiting for stalled tur thread to finish",
+ major(ct->devt), minor(ct-Quoted 1 line
devt));
+ ct->nr_timeouts++;
+ }
/*
- * The thread has been cancelled but hasn't quit.
+ * Don't start new threads until the last once has
+ * finished.
+ */
+ if (ct->nr_timeouts>MAX_NR_TIMEOUTS) { + c->msgid =MSG_TUR_TIMEOUT; + returnPATH_TIMEOUT;
+ }
+ ct->nr_timeouts++;
+ /*
+ * Start a new thread while the old one is stalled.
* We have to prevent it from interfering with the new
* thread. We create a new context and leave the old
* one with the stale thread, hoping it will clean uplibmultipath/checkers/tur.c +1
-0@@ -375,13 +392,15 @@ int libcheck_check(struct checker * c)*/ if (libcheck_init(c) != 0) return PATH_UNCHECKED; + ((struct tur_checker_context *)c->context)- >nr_timeouts = ct->nr_timeouts;
if (!uatomic_sub_return(&ct->holders, 1)) /* It did terminate, eventually */cleanup_context(ct);
Quoted 8 lines
ct =c->context;
- }
+ } else
+ ct->nr_timeouts = 0; /* Start new TUR checker */pthread_mutex_lock(&ct->lock);tur_status=ct->state=PATH_PENDING;
--
dm-devel mailing list
[email protected]
https://listman.redhat.com/mailman/listinfo/dm-devel
Benjamin Marzinski
maintainerQuoted 19 lines · Martin Wilck
Quoted 8 lines · Benjamin Marzinski
If tur thead hangs, multipathd was simply creating a new thread, and assuming that the old thread would get cleaned up eventually. I have seen a case recently where there were 26000 multipathd threads on a system, all stuck trying to send TUR commands to path devices. The root
cause of the issue was a scsi kernel issue, but it shows that the way multipathd currently deals with stuck threads could use some refinement.Alright, let's start the discussion about dead TUR threads again ;-) For my own records, here are links to some previous ML threads:
https://listman.redhat.com/archives/dm-devel/2018-September/036102.html
https://listman.redhat.com/archives/dm-devel/2018-October/036156.html
https://listman.redhat.com/archives/dm-devel/2018-October/036159.html
https://listman.redhat.com/archives/dm-devel/2018-October/036240.html
https://listman.redhat.com/archives/dm-devel/2018-October/036362.html
Quoted 3 lines
Have you assessed why the SCSI issues were causing indefinite hangs? Was it just the TUR command hanging forever in the kernel, or some other issue?
Yeah, all the threads are stuck in the kernel, waiting for either a response to their command, or to be timed out, which isn't happening for some reason yet to be determined.
Quoted 21 lines
I have seen similar issues in the past; they were also related to SCSI problems, IIRC. I used to think that this is just a symptom: The "dead" threads are detached and consume a minimal amount of non-shared memory (they are a pain when analyzing a crash dump, though).
But I agree, just continuing to start new threads forever isn't optimal.
Quoted 13 lines
Now, when one tur thread hangs, multipathd will act as it did before. If a second one in a row hangs, multipathd will instead wait for it to
complete before starting another thread. Once the thread completes, the
count is reset.Signed-off-by Benjamin Marzinski <[email protected]>---
libmultipath/checkers/tur.c | 23 +++++++++++++++++++++--1 file changed, 21 insertions(+), 2 deletions(-)
Quoted 12 lines
Quoted 10 lines
libmultipath/checkers/tur.c +0
-0@@ -26,6 +26,7 @@
#defineTUR_CMD_LEN6
#define HEAVY_CHECK_COUNT 10
+#defineMAX_NR_TIMEOUTS1Why did you choose 1? Perhaps we should make a few more attempts?
Oops. I forgot to include the Notes when formatting my patch (I need to make that the default). Here they are:
Notes:
I'd be interested in knowing what people think of this solution. I'm open to making multipathd start more threads before it gives up. We also could make multipathd save the contexts from the stalled threads, so that when it stops creating new ones, instead of just waiting for the last thread to complete, it could start up again as soon as any of the outstanding threads completed. We could also consider not stopping creating new threads entirely, but instead having a delay before we create a new checker thread, where we wait for the last thread to complete. If the delay doubled after evey consecutive timeout, the number of threads created would stay at a more reasonable level until someone got around to looking into what was going wrong. Thoughts?
But to answer your question, there was no obvious number to choose, and you can make the case that if it fails once, that's a fluke. If it fails twice in a row, then it's likely to keep faiing. But I'm fine with picking a bigger number, or any of the other possibilities I listed. I just wanted to throw something out as a starting point.
-Ben
Quoted 2 lines
Other than that, this looks ok to me (assuming you've tested with the code from bdf55d6, or something similar).
Yep. I tested it with the zombie tur checker tests, with different sleep seconds and intervals.
Quoted 66 lines
Martin
Quoted 63 lines
enum {
MSG_TUR_RUNNING =CHECKER_FIRST_MSGID,libmultipath/checkers/tur.c +0
-0@@ -54,6 +55,7 @@ struct tur_checker_context {int holders; /* uatomic access only */
int msgid;
structchecker_contextctx;
+ unsigned intnr_timeouts;
};
intlibcheck_init(struct checker * c)libmultipath/checkers/tur.c +1
-0@@ -358,8 +360,23 @@ int libcheck_check(struct checker * c) } } else { if (uatomic_read(&ct->holders) > 1) { + /* The thread has been cancelled but hasn'tquit. */
+ if (ct->nr_timeouts==MAX_NR_TIMEOUTS) { + condlog(2, "%d:%d : waiting for stalled tur thread to finish",
+ major(ct->devt), minor(ct-Quoted 1 line
devt));
+ ct->nr_timeouts++;
+ }
/*
- * The thread has been cancelled but hasn't quit.
+ * Don't start new threads until the last once has
+ * finished.
+ */
+ if (ct->nr_timeouts>MAX_NR_TIMEOUTS) { + c->msgid =MSG_TUR_TIMEOUT; + returnPATH_TIMEOUT;
+ }
+ ct->nr_timeouts++;
+ /*
+ * Start a new thread while the old one is stalled.
* We have to prevent it from interfering with the new
* thread. We create a new context and leave the old
* one with the stale thread, hoping it will clean uplibmultipath/checkers/tur.c +1
-0@@ -375,13 +392,15 @@ int libcheck_check(struct checker * c)*/ if (libcheck_init(c) != 0) return PATH_UNCHECKED; + ((struct tur_checker_context *)c->context)- >nr_timeouts = ct->nr_timeouts;
if (!uatomic_sub_return(&ct->holders, 1)) /* It did terminate, eventually */cleanup_context(ct);
ct =c->context;
- }
+ } else
+ ct->nr_timeouts = 0;
/* Start new TUR checker */pthread_mutex_lock(&ct->lock);tur_status=ct->state=PATH_PENDING;
--
dm-devel mailing list
[email protected]
https://listman.redhat.com/mailman/listinfo/dm-devel
Martin Wilck
Reviewed-byQuoted 11 lines · Benjamin Marzinski
Quoted 2 lines · Martin Wilck
Why did you choose 1? Perhaps we should make a few more attempts?
Oops. I forgot to include the Notes when formatting my patch (I need to
make that the default). Here they are:Notes:
Quoted 7 lines
I'd be interested in knowing what people think of this solution. I'm
open to making multipathd start more threads before it gives up. We
also could make multipathd save the contexts from the stalled threads, so that when it stops creating new ones, instead of just waiting for the last thread to complete, it could start up again
I wouldn't recommend that. We drop all references to the old context for a good reason: to be sure there are no races when the hanging thread eventually does exit. Keeping such references would re- open a Pandora's box which we sealed and closed in 2018.
Quoted 6 lines
as
soon as any of the outstanding threads completed. We could also consider not stopping creating new threads entirely, but instead having a delay before we create a new checker thread, where we wait
for the last thread to complete.
I wouldn't do this, either. The hang check is not done immediately after cancelling the thread, but in the following libcheck_check() invocation. That means there has been some delay already when we do the check, at least a second. And this cancellation happened after the SCSI timeout expired, anyway. How long are we going to wait for the normally instanteneous cancellation to complete? I like the "just forget about this thread" attitude which has saved us a lot of trouble lately, IMO.
Quoted 7 lines
If the delay doubled after evey
consecutive timeout, the number of threads created would stay at a
more reasonable level until someone got around to looking into what
was going wrong. Thoughts?
Quoted 5 lines
But to answer your question, there was no obvious number to choose, and
you can make the case that if it fails once, that's a fluke. If it fails
twice in a row, then it's likely to keep faiing.
That makes sense. I'm fine with the patch.
Quoted 8 lines
But I'm fine with
picking a bigger number, or any of the other possibilities I listed. I
just wanted to throw something out as a starting point.-Ben
Quoted 6 lines
Quoted 3 lines
Other than that, this looks ok to me (assuming you've tested with the
code from bdf55d6, or something similar).Yep. I tested it with the zombie tur checker tests, with different sleep seconds and intervals.
--
dm-devel mailing list
[email protected]
https://listman.redhat.com/mailman/listinfo/dm-devel