-o intr mount option prevents core dumps on 2.4 kernel

PATCH In review Steve Dickson 2 participants 14 messages Jan 6, 2005 – Jan 14, 2005

Subsystems (5)

Networking VFS NFS client NFS server +1 more subsystems

lore

Share & tools

lore

Steve Dickson

author

2005-01-06 18:31 UTC #0 lore raw

PATCH 1/1 -o intr mount option prevents core dumps on 2.4 kernel

Here is a 2.4.28 patch that allows core dumps to be written on fileystems mounted with the 'intr' option.

By setting the PF_DUMPCORE bit in the current->flags (via do_coredump()), I was able to signal the RPC and NFS code that the current task is trying to dump a core. So the trick was to have the rpc code (temporarily) ignore the fact the task got signalled() and let the write/commit path complete....

I do this by checking the PF_DUMPCORE bit in both __rpc_execute() and nfs_wait_event(). One side effect is the dropping core process becomes uninterruptable.... I combat this by only allowing the process three retries before giving up...

Finally there are two placing in the NFS code where I don't ignore the signal and do fail the writing of the core. One, is in nfs_create_request() when there are no available pages and in nfs3_rpc_wrapper() when the server returns EJUKEBOX. But I did place printks so it will be clear as to why it failed....

Comments... Suggestions... Crimes against humanity that I'm committing???

steved.

fs/nfs/pagelist.c renamed from fs/nfs/pagelist.c.orig +6 -1

@@ -94,8 +94,13 @@ nfs_create_request(struct rpc_cred *cred
 		 */
 		if (nfs_try_to_free_pages(server))
 			continue;
-		if (signalled() && (server->flags & NFS_MOUNT_INTR))
+
+		if (signalled() && (server->flags & NFS_MOUNT_INTR)) {
+			if (task_core_dumping())
+				printk(KERN_WARNING \
+					"NFS: Core Dump Aborted due to lack of resources\n");
 			return ERR_PTR(-ERESTARTSYS);
+		}
 		yield();
 	}

fs/nfs/nfs3proc.c renamed from fs/nfs/nfs3proc.c.orig +3 -0

@@ -31,6 +31,9 @@ nfs3_rpc_wrapper(struct rpc_clnt *clnt, 
 		set_current_state(TASK_INTERRUPTIBLE);
 		schedule_timeout(NFS_JUKEBOX_RETRY_TIME);
 		res = -ERESTARTSYS;
+		if (signalled() && task_core_dumping())
+			printk(KERN_WARNING \
+				"NFS: Core Dump Aborted due to lack of resources\n");
 	} while (!signalled());
 	rpc_clnt_sigunmask(clnt, &oldset);
 	return res;

fs/exec.c renamed from fs/exec.c.orig +1 -0

@@ -1125,6 +1125,7 @@ int do_coredump(long signr, struct pt_re
 	if (current->rlim[RLIMIT_CORE].rlim_cur < binfmt->min_coredump)
 		goto fail;
 
+	current->flags |= PF_DUMPCORE;
  	format_corename(corename, core_pattern, signr);
 	file = filp_open(corename, O_CREAT | 2 | O_NOFOLLOW, 0600);
 	if (IS_ERR(file))

include/linux/sunrpc/clnt.h renamed from include/linux/sunrpc/clnt.h.orig +2 -0

@@ -108,6 +108,8 @@ struct rpc_procinfo {
 
 #ifdef __KERNEL__
 
+#define task_core_dumping() (current->flags & PF_DUMPCORE)
+
 struct rpc_clnt *rpc_create_client(struct rpc_xprt *xprt, char *servname,
 				struct rpc_program *info,
 				u32 version, int authflavor);

include/linux/nfs_fs.h renamed from include/linux/nfs_fs.h.orig +1 -1

@@ -345,7 +345,7 @@ extern void * nfs_root_data(void);
 #define nfs_wait_event(clnt, wq, condition)				\
 ({									\
 	int __retval = 0;						\
-	if (clnt->cl_intr) {						\
+	if (clnt->cl_intr && !task_core_dumping()) { \
 		sigset_t oldmask;					\
 		rpc_clnt_sigmask(clnt, &oldmask);			\
 		__retval = wait_event_interruptible(wq, condition);	\

net/sunrpc/sched.c renamed from net/sunrpc/sched.c.orig +13 -2

@@ -498,7 +498,7 @@ __rpc_atrun(struct rpc_task *task)
 static int
 __rpc_execute(struct rpc_task *task)
 {
-	int		status = 0;
+	int		status = 0, core_retry = 0;
 
 	dprintk("RPC: %4d rpc_execute flgs %x\n",
 				task->tk_pid, task->tk_flags);

@@ -572,9 +572,20 @@ __rpc_execute(struct rpc_task *task)
 			 * -ERESTARTSYS. In order to catch any callbacks that
 			 * clean up after sleeping on some queue, we don't
 			 * break the loop here, but go around once more.
-			 */
+			 */	
 			if (task->tk_client->cl_intr && signalled()) {
 				dprintk("RPC: %4d got signal\n", task->tk_pid);
+				/*
+				 * If we are dropping a core and 'intr' is set,
+				 * we want to make an attempted at writting
+				 * out the core, but not get hung up in it.
+				 */
+				if (task_core_dumping()) {
+					dprintk("RPC: %4d dropping core retries %d\n", 
+						task->tk_pid, core_retry);
+					if (++core_retry < 3)
+						continue;
+				}
 				task->tk_flags |= RPC_TASK_KILLED;
 				rpc_exit(task, -ERESTARTSYS);
 				rpc_wake_up_task(task);

Trond Myklebust

2005-01-07 17:14 UTC #1 lore raw

to den 06.01.2005 Klokka 13:31 (-0500) skreiv Steve Dickson:

Quoted 1 line

Comments... Suggestions... Crimes against humanity that I'm committing???

Firstly, how is "cl_intr" preventing core dumps? Normally, we should be masking all signals except SIGKILL, SIGINT and SIGQUIT. I can see that masking SIGINT and SIGQUIT might be useful here, but not SIGKILL...

Secondly: Why not do all this in rpc_clnt_sigmask()?

Cheers,
Trond

Signature · 11 lines

Trond Myklebust <[email protected]>



-------------------------------------------------------
The SF.Net email is sponsored by: Beat the post-holiday blues
Get a FREE limited edition SourceForge.net t-shirt from ThinkGeek.
It's fun and FREE -- well, almost....http://www.thinkgeek.com/sfshirt
_______________________________________________
NFS maillist  -  [email protected]
https://lists.sourceforge.net/lists/listinfo/nfs

Steve Dickson

author

2005-01-11 19:33 UTC #2 lore raw

Quoted 4 lines · Trond Myklebust

to den 06.01.2005 Klokka 13:31 (-0500) skreiv Steve Dickson:

Quoted 10 lines

Quoted 3 lines
Comments... Suggestions... Crimes against humanity that I'm committing???
Firstly, how is "cl_intr" preventing core dumps? Normally, we should be masking all signals except SIGKILL, SIGINT and SIGQUIT. I can see that masking SIGINT and SIGQUIT might be useful here, but not SIGKILL...

Because __rpc_execute() notices that signalled() is set and aborts the write by returning -ERESTARTSYS makes the core dumping code stop dumping.

Quoted 3 lines

Secondly: Why not do all this in rpc_clnt_sigmask()?

I looked at that.... but I really was not sure if could hack it up to turn off signalled() and plus I was sure how I could turn signalled() back on at the end of the write.... The way I did just seemed less risky and seem to work...

steved.

-------------------------------------------------------
The SF.Net email is sponsored by: Beat the post-holiday blues Get a FREE limited edition SourceForge.net t-shirt from ThinkGeek. It's fun and FREE -- well, almost....http://www.thinkgeek.com/sfshirt
_______________________________________________
NFS maillist - [email protected]
https://lists.sourceforge.net/lists/listinfo/nfs

Trond Myklebust

2005-01-11 20:11 UTC #3 lore raw

ty den 11.01.2005 Klokka 14:36 (-0500) skreiv Steve Dickson:

Quoted 8 lines

Quoted 5 lines
Firstly, how is "cl_intr" preventing core dumps? Normally, we should be masking all signals except SIGKILL, SIGINT and SIGQUIT. I can see that masking SIGINT and SIGQUIT might be useful here, but not SIGKILL...
Because __rpc_execute() notices that signalled() is set and aborts the write by
returning -ERESTARTSYS makes the core dumping code stop dumping.

Err... That's what the sigmask is for. signalled() only looks at those signals that are enabled by the sigmask.

Quoted 7 lines

Quoted 3 lines
Secondly: Why not do all this in rpc_clnt_sigmask()?
I looked at that.... but I really was not sure if could hack it up to turn off signalled() and plus I was sure how I could turn signalled() back on at the end of the write.... The way I did just seemed less risky and seem to work...

See above.

Cheers,
Trond

Signature · 11 lines

Trond Myklebust <[email protected]>



-------------------------------------------------------
The SF.Net email is sponsored by: Beat the post-holiday blues
Get a FREE limited edition SourceForge.net t-shirt from ThinkGeek.
It's fun and FREE -- well, almost....http://www.thinkgeek.com/sfshirt
_______________________________________________
NFS maillist  -  [email protected]
https://lists.sourceforge.net/lists/listinfo/nfs

Steve Dickson

author

2005-01-11 20:54 UTC #4 lore raw

Quoted 11 lines · Trond Myklebust

ty den 11.01.2005 Klokka 14:36 (-0500) skreiv Steve Dickson:
Quoted 7 lines
Quoted 7 lines
Firstly, how is "cl_intr" preventing core dumps? Normally, we should be masking all signals except SIGKILL, SIGINT and SIGQUIT. I can see that masking SIGINT and SIGQUIT might be useful here, but not SIGKILL...

Quoted 9 lines

Quoted 4 lines
Because __rpc_execute() notices that signalled() is set and aborts the write by returning -ERESTARTSYS makes the core dumping code stop dumping.
Err... That's what the sigmask is for. signalled() only looks at those signals that are enabled by the sigmask.

hmm... I'll admit kernel signal process is not an area I've ventured into much so correct me if I'm wrong.... But if SIGKILL is masked off, would that stop the process from getting the signal and in turn stop the core dropping as well???

steved.

Trond Myklebust

2005-01-11 21:59 UTC #5 lore raw

ty den 11.01.2005 Klokka 15:58 (-0500) skreiv Steve Dickson:

Quoted 5 lines

hmm... I'll admit kernel signal process is not an area I've ventured into much so
correct me if I'm wrong.... But if SIGKILL is masked off, would that stop the process from getting the signal and in turn stop the core dropping as well???

Yes, but I'm not sure that it makes sense to mask off SIGKILL as that makes core dumping completely uninterruptible. People do expect "kill -9" to always work when the "intr" flag is set.

You should, however, mask off all other signals, including SIGINT and SIGQUIT. The latter two (+ sigkill) are the only ones that rpc_clnt_sigmask() does not currently mask out when "intr" is set.

Cheers,
Trond

Signature · 11 lines

Trond Myklebust <[email protected]>



-------------------------------------------------------
The SF.Net email is sponsored by: Beat the post-holiday blues
Get a FREE limited edition SourceForge.net t-shirt from ThinkGeek.
It's fun and FREE -- well, almost....http://www.thinkgeek.com/sfshirt
_______________________________________________
NFS maillist  -  [email protected]
https://lists.sourceforge.net/lists/listinfo/nfs

Steve Dickson

author

2005-01-12 18:18 UTC #6 lore raw

Quoted 5 lines · Trond Myklebust

Yes, but I'm not sure that it makes sense to mask off SIGKILL as that makes core dumping completely uninterruptible. People do expect "kill -9" to always work when the "intr" flag is set.

True... and it will continue to work...

Quoted 5 lines

You should, however, mask off all other signals, including SIGINT and SIGQUIT. The latter two (+ sigkill) are the only ones that rpc_clnt_sigmask() does not currently mask out when "intr" is set.

Good point... that attached patch does just that....

steved

fs/nfs/pagelist.c renamed from fs/nfs/pagelist.c.orig +6 -1

@@ -94,8 +94,13 @@ nfs_create_request(struct rpc_cred *cred
 		 */
 		if (nfs_try_to_free_pages(server))
 			continue;
-		if (signalled() && (server->flags & NFS_MOUNT_INTR))
+
+		if (signalled() && (server->flags & NFS_MOUNT_INTR)) {
+			if (task_core_dumping())
+				printk(KERN_WARNING \
+					"NFS: Core Dump Aborted due to lack of resources\n");
 			return ERR_PTR(-ERESTARTSYS);
+		}
 		yield();
 	}

fs/nfs/nfs3proc.c renamed from fs/nfs/nfs3proc.c.orig +3 -0

@@ -31,6 +31,9 @@ nfs3_rpc_wrapper(struct rpc_clnt *clnt, 
 		set_current_state(TASK_INTERRUPTIBLE);
 		schedule_timeout(NFS_JUKEBOX_RETRY_TIME);
 		res = -ERESTARTSYS;
+		if (signalled() && task_core_dumping())
+			printk(KERN_WARNING \
+				"NFS: Core Dump Aborted due to lack of resources\n");
 	} while (!signalled());
 	rpc_clnt_sigunmask(clnt, &oldset);
 	return res;

fs/exec.c renamed from fs/exec.c.orig +1 -0

@@ -1125,6 +1125,7 @@ int do_coredump(long signr, struct pt_re
 	if (current->rlim[RLIMIT_CORE].rlim_cur < binfmt->min_coredump)
 		goto fail;
 
+	current->flags |= PF_DUMPCORE;
  	format_corename(corename, core_pattern, signr);
 	file = filp_open(corename, O_CREAT | 2 | O_NOFOLLOW, 0600);
 	if (IS_ERR(file))

include/linux/sunrpc/clnt.h renamed from include/linux/sunrpc/clnt.h.orig +2 -0

@@ -108,6 +108,8 @@ struct rpc_procinfo {
 
 #ifdef __KERNEL__
 
+#define task_core_dumping() (current->flags & PF_DUMPCORE)
+
 struct rpc_clnt *rpc_create_client(struct rpc_xprt *xprt, char *servname,
 				struct rpc_program *info,
 				u32 version, int authflavor);

include/linux/nfs_fs.h renamed from include/linux/nfs_fs.h.orig +1 -1

@@ -345,7 +345,7 @@ extern void * nfs_root_data(void);
 #define nfs_wait_event(clnt, wq, condition)				\
 ({									\
 	int __retval = 0;						\
-	if (clnt->cl_intr) {						\
+	if (clnt->cl_intr && !task_core_dumping()) { \
 		sigset_t oldmask;					\
 		rpc_clnt_sigmask(clnt, &oldmask);			\
 		__retval = wait_event_interruptible(wq, condition);	\

net/sunrpc/sched.c renamed from net/sunrpc/sched.c.orig +13 -2

@@ -498,7 +498,7 @@ __rpc_atrun(struct rpc_task *task)
 static int
 __rpc_execute(struct rpc_task *task)
 {
-	int		status = 0;
+	int		status = 0, core_retry = 0;
 
 	dprintk("RPC: %4d rpc_execute flgs %x\n",
 				task->tk_pid, task->tk_flags);

@@ -572,9 +572,20 @@ __rpc_execute(struct rpc_task *task)
 			 * -ERESTARTSYS. In order to catch any callbacks that
 			 * clean up after sleeping on some queue, we don't
 			 * break the loop here, but go around once more.
-			 */
+			 */	
 			if (task->tk_client->cl_intr && signalled()) {
 				dprintk("RPC: %4d got signal\n", task->tk_pid);
+				/*
+				 * If we are dropping a core and 'intr' is set,
+				 * we want to make an attempted at writting
+				 * out the core, but not get hung up in it.
+				 */
+				if (task_core_dumping()) {
+					dprintk("RPC: %4d dropping core retries %d\n", 
+						task->tk_pid, core_retry);
+					if (++core_retry < 3)
+						continue;
+				}
 				task->tk_flags |= RPC_TASK_KILLED;
 				rpc_exit(task, -ERESTARTSYS);
 				rpc_wake_up_task(task);

net/sunrpc/clnt.c renamed from net/sunrpc/clnt.c.orig +1 -1

@@ -209,7 +209,7 @@ void rpc_clnt_sigmask(struct rpc_clnt *c
 	unsigned long	irqflags;
 	
 	/* Turn off various signals */
-	if (clnt->cl_intr) {
+	if (clnt->cl_intr && !task_core_dumping()) {
 		struct k_sigaction *action = current->sig->action;
 		if (action[SIGINT-1].sa.sa_handler == SIG_DFL)
 			sigallow |= sigmask(SIGINT);

Trond Myklebust

2005-01-12 18:25 UTC #7 lore raw

on den 12.01.2005 Klokka 13:18 (-0500) skreiv Steve Dickson:

Quoted 6 lines

Quoted 5 lines
You should, however, mask off all other signals, including SIGINT and SIGQUIT. The latter two (+ sigkill) are the only ones that rpc_clnt_sigmask() does not currently mask out when "intr" is set.
Good point... that attached patch does just that....

I meant something more like the following.

Cheers,
Trond

Signature · 1 line

Trond Myklebust <[email protected]>

 fs/exec.c         |    1 +
 net/sunrpc/clnt.c |    2 +-
 2 files changed, 2 insertions(+), 1 deletion(-)

fs/exec.c +1 -0

Index: linux-2.4.28-rc1/fs/exec.c
===================================================================
@@ -1125,6 +1125,7 @@ int do_coredump(long signr, struct pt_re
 	if (current->rlim[RLIMIT_CORE].rlim_cur < binfmt->min_coredump)
 		goto fail;
 
+	current->flags |= PF_DUMPCORE;
  	format_corename(corename, core_pattern, signr);
 	file = filp_open(corename, O_CREAT | 2 | O_NOFOLLOW, 0600);
 	if (IS_ERR(file))

net/sunrpc/clnt.c +1 -1

Index: linux-2.4.28-rc1/net/sunrpc/clnt.c
===================================================================
@@ -209,7 +209,7 @@ void rpc_clnt_sigmask(struct rpc_clnt *c
 	unsigned long	irqflags;
 	
 	/* Turn off various signals */
-	if (clnt->cl_intr) {
+	if (clnt->cl_intr && !(current->flags & PF_DUMPCORE)) {
 		struct k_sigaction *action = current->sig->action;
 		if (action[SIGINT-1].sa.sa_handler == SIG_DFL)
 			sigallow |= sigmask(SIGINT);

Steve Dickson

author

2005-01-12 19:12 UTC #8 lore raw

Quoted 6 lines · Trond Myklebust

===================================================================

net/sunrpc/clnt.c void rpc_clnt_sigmask(struct rpc_clnt *c · line 215

Quoted 8 lines

net/sunrpc/clnt.c +1 -1

@@ -209,7 +209,7 @@ void rpc_clnt_sigmask(struct rpc_clnt *c
 	unsigned long	irqflags;
 
 	/* Turn off various signals */
-	if (clnt->cl_intr) {
+	if (clnt->cl_intr && !(current->flags & PF_DUMPCORE)) {
 		struct k_sigaction *action = current->sig->action;
 		if (action[SIGINT-1].sa.sa_handler == SIG_DFL)
 			sigallow |= sigmask(SIGINT);

Well did you try just this? ;-)

Here is what I found.... just adding the PF_DUMPCORE rpc_clnt_sigmask() no core was dropped because __rpc_execute returned -ERESTARTSYS due to signalled() == TRUE.

When I added back the PF_DUMPCORE check to __rpc_execute(), only the header of the core was dropped because nfs_wait_event() returned -ERESTARTSYS because signalled() == TRUE.

When I added back the PF_DUMPCORE to nfs_wait_event(), the entire core was dropped.

So it appears to me that you need both checks.....

steved

Trond Myklebust

2005-01-13 00:38 UTC #9 lore raw

on den 12.01.2005 Klokka 14:11 (-0500) skreiv Steve Dickson:

Quoted 9 lines

Here is what I found.... just adding the PF_DUMPCORE rpc_clnt_sigmask() no core was dropped
because __rpc_execute returned -ERESTARTSYS due to signalled() == TRUE.
When I added back the PF_DUMPCORE check to __rpc_execute(), only the header of the core
was dropped because nfs_wait_event() returned -ERESTARTSYS because signalled() == TRUE.

Quoted 4 lines

When I added back the PF_DUMPCORE to nfs_wait_event(), the entire core was dropped.
So it appears to me that you need both checks.....

No! Those extra checks are neither necessary, nor are they even correct! I repeat what I said in my earlier mail:

- The change to nfs_wait_event() converts it into an unconditional uninterruptible sleep. That means you can never "kill -9" out of waiting for the core dump to finish in case the server crashes!

- The loop you add to __rpc_execute() is pointless: The signal that caused your process to wake up is neither cleared nor masked, so when it later tries to block (such as when waiting for a reply from the server), that process will just find itself immediately woken up again by the same signal.
IOW: any milage you are getting out of this change will be 100% timing-dependent. Not being able to wait for replies from the server, you are depending entirely on the server to reply so quickly that it appears to be "instantaneous".
...

As for myself, I'm getting full coredumps when I apply the 2 line patch I sent you. The appended testcase works every time:

[trondmy@trondheim trondmy]$ uname -a
Linux trondheim.citi.umich.edu 2.4.29-rc2 #3 on jan 12 19:12:57 EST 2005 i686 athlon i386 GNU/Linux [trondmy@trondheim trondmy]$ ./a.out
Quit (core dumped)
[trondmy@trondheim trondmy]$ ls -l core.1657
-rw------- 1 trondmy users 16953344 jan 12 19:36 core.1657

Cheers,
Trond

Signature · 44 lines

Trond Myklebust <[email protected]>


---------------------------------------
#include <signal.h>
#include <stdlib.h>
#include <sys/time.h>
#include <sys/resource.h>
#include <unistd.h>


#define CORESIZE (1024*1024*16)
int main()
{
	void *ptr;
	struct rlimit rlimit = {
		.rlim_cur = RLIM_INFINITY,
		.rlim_max = RLIM_INFINITY
	};

	/* Ensure we are set to create core dump files */
	if (setrlimit(RLIMIT_CORE, &rlimit) < 0) {
		perror("setrlimit failed!");
		exit(1);
	}

	/* Ensure coredump is large! */
	ptr = malloc(CORESIZE);
	memset(ptr, '0', CORESIZE);

	/* raise hell */
	raise(SIGQUIT);
}




-------------------------------------------------------
The SF.Net email is sponsored by: Beat the post-holiday blues
Get a FREE limited edition SourceForge.net t-shirt from ThinkGeek.
It's fun and FREE -- well, almost....http://www.thinkgeek.com/sfshirt
_______________________________________________
NFS maillist  -  [email protected]
https://lists.sourceforge.net/lists/listinfo/nfs

Steve Dickson

author

2005-01-13 14:01 UTC #10 lore raw

Quoted 16 lines · Trond Myklebust

on den 12.01.2005 Klokka 14:11 (-0500) skreiv Steve Dickson:
Quoted 11 lines
Here is what I found.... just adding the PF_DUMPCORE rpc_clnt_sigmask() no core was dropped because __rpc_execute returned -ERESTARTSYS due to signalled() == TRUE.
When I added back the PF_DUMPCORE check to __rpc_execute(), only the header of the core was dropped because nfs_wait_event() returned -ERESTARTSYS because signalled() == TRUE.
When I added back the PF_DUMPCORE to nfs_wait_event(), the entire core was dropped.
So it appears to me that you need both checks.....

Quoted 9 lines

No! Those extra checks are neither necessary, nor are they even correct! I repeat what I said in my earlier mail:
- The change to nfs_wait_event() converts it into an unconditional uninterruptible sleep. That means you can never "kill -9" out of waiting for the core dump to finish in case the server crashes!

By adding the PF_DUMPCORE check to rpc_clnt_sigmask() (as you suggested)
I thought I had taken care of this problem... but now I realize the check in nfs_wait_event() is the real issue... sorry for making you swing that clue bat twice!! :)

Quoted 7 lines

- The loop you add to __rpc_execute() is pointless: The signal that caused your process to wake up is neither cleared nor masked, so when it later tries to block (such as when waiting for a reply from the server), that process will just find itself immediately woken up again by the same signal.

I do understand this point.... and I *thought* I had addressed it by making __rpc_execute() temporary ignoring signals and having nfs_wait_event () call wait_event() instead of wait_event_interruptible().....

Quoted 4 lines

As for myself, I'm getting full coredumps when I apply the 2 line patch I sent you. The appended testcase works every time:

hmm.... something is amiss.... because I'm definitely not see this.... but I'll keep digging...

thanks!

steved.

Steve Dickson

author

2005-01-13 21:37 UTC #11 lore raw

Quoted 6 lines · Steve Dickson

Quoted 5 lines
As for myself, I'm getting full coredumps when I apply the 2 line patch I sent you. The appended testcase works every time:
hmm.... something is amiss.... because I'm definitely not see this.... but I'll keep digging...

after further review.... it seems the reason your patch works on a 2.4.28 kernel, is
because signalled() never returns true, even when the process is sent the SIGSEGV signal. Why signalled() never returns true, I don't know.... but the reason your patch did not work on a RHEL3 kernel is because signalled() does
return true and calls to recalc_sigpending() do not reset that pending signal....
which seems to be a bug in the NTPL-related code.... code that is not in the
2.4.28 kernel....

steved.

Trond Myklebust

2005-01-14 02:57 UTC #12 lore raw

to den 13.01.2005 Klokka 16:37 (-0500) skreiv Steve Dickson:

Quoted 11 lines

after further review.... it seems the reason your patch works on a 2.4.28 kernel, is
because signalled() never returns true, even when the process is sent the SIGSEGV signal. Why signalled() never returns true, I don't know.... but the
reason your patch did not work on a RHEL3 kernel is because signalled() does
return true and calls to recalc_sigpending() do not reset that pending signal....
which seems to be a bug in the NTPL-related code.... code that is not in the 2.4.28 kernel....

recalc_sigpending() failing to honour the sigmask sounds like a pretty major bug! Is there no way you can backport some of the 2.6 signal code or something like that in order to work around it?

It sounds to me as if that bug will affect more than just coredumps...

Cheers,
Trond

Signature · 11 lines

Trond Myklebust <[email protected]>



-------------------------------------------------------
The SF.Net email is sponsored by: Beat the post-holiday blues
Get a FREE limited edition SourceForge.net t-shirt from ThinkGeek.
It's fun and FREE -- well, almost....http://www.thinkgeek.com/sfshirt
_______________________________________________
NFS maillist  -  [email protected]
https://lists.sourceforge.net/lists/listinfo/nfs

Steve Dickson

author

2005-01-14 19:52 UTC #13 lore raw

Quoted 13 lines · Trond Myklebust

to den 13.01.2005 Klokka 16:37 (-0500) skreiv Steve Dickson:
Quoted 8 lines
after further review.... it seems the reason your patch works on a 2.4.28 kernel, is
because signalled() never returns true, even when the process is sent the SIGSEGV signal. Why signalled() never returns true, I don't know.... but the reason your patch did not work on a RHEL3 kernel is because signalled() does
return true and calls to recalc_sigpending() do not reset that pending signal....
which seems to be a bug in the NTPL-related code.... code that is not in the
2.4.28 kernel....

Quoted 3 lines

recalc_sigpending() failing to honour the sigmask sounds like a pretty major bug!

Well it turns it was a bug in the kernel signaling code which was causing this issue.... which means there are no changes needed to the NFS code, which is always a good thing... :)

See bz 132162 if interested....

steved.