Re: + fs-epoll-add-procfs-fdinfo-helper.patch added to -mm tree

Re: + fs-epoll-add-procfs-fdinfo-helper.patch added to -mm tree

[Oleg Nesterov]

>  struct signalfd_ctx {
> +	seqcount_t cnt;
>  	sigset_t sigmask;
>  };
> ...
> @@ -278,7 +302,9 @@ SYSCALL_DEFINE4(signalfd4, int, ufd, sig
>  			return -EINVAL;
>  		}
>  		spin_lock_irq(&current->sighand->siglock);
> +		write_seqcount_begin(&ctx->cnt);
>  		ctx->sigmask = sigmask;
> +		write_seqcount_end(&ctx->cnt);
>  		spin_unlock_irq(&current->sighand->siglock);

This doesn't look right.

The problem is, the current locking is broken, ->siglock can not serialize
->sigmask changes. Just suppose the the child inherits sigfd from parent
and they both do sys_signalfd4() at the same time.

Nothing really bad can happen, that is why nobody bothers to fix this.
But this patch makes the thing worse, write_seqcount_begin() must be
serialized correctly.

Oleg.

Re: + fs-epoll-add-procfs-fdinfo-helper.patch added to -mm tree

[Cyrill Gorcunov]

On Tue, Nov 13, 2012 at 03:50:50PM +0100, Oleg Nesterov wrote:
> This doesn't look right.
> 
> The problem is, the current locking is broken, ->siglock can not serialize
> ->sigmask changes. Just suppose the the child inherits sigfd from parent
> and they both do sys_signalfd4() at the same time.
> 
> Nothing really bad can happen, that is why nobody bothers to fix this.
> But this patch makes the thing worse, write_seqcount_begin() must be
> serialized correctly.

Thanks a lot, Oleg! I'll update.

Re: + fs-epoll-add-procfs-fdinfo-helper.patch added to -mm tree

[Cyrill Gorcunov]

On Tue, Nov 13, 2012 at 07:53:13PM +0400, Cyrill Gorcunov wrote:
> On Tue, Nov 13, 2012 at 03:50:50PM +0100, Oleg Nesterov wrote:
> > This doesn't look right.
> > 
> > The problem is, the current locking is broken, ->siglock can not serialize
> > ->sigmask changes. Just suppose the the child inherits sigfd from parent
> > and they both do sys_signalfd4() at the same time.
> > 
> > Nothing really bad can happen, that is why nobody bothers to fix this.
> > But this patch makes the thing worse, write_seqcount_begin() must be
> > serialized correctly.
> 
> Thanks a lot, Oleg! I'll update.

Something like below?
---
From: Cyrill Gorcunov <gorcunov@openvz.org>
Subject: fdinfo: Show sigmask for signalfd fd v3

Signed-off-by: Cyrill Gorcunov <gorcunov@openvz.org>
CC: Pavel Emelyanov <xemul@parallels.com>
CC: Al Viro <viro@ZenIV.linux.org.uk>
CC: Alexey Dobriyan <adobriyan@gmail.com>
CC: Andrew Morton <akpm@linux-foundation.org>
CC: James Bottomley <jbottomley@parallels.com>
CC: "Aneesh Kumar K.V" <aneesh.kumar@linux.vnet.ibm.com>
CC: Alexey Dobriyan <adobriyan@gmail.com>
CC: Matthew Helsley <matt.helsley@gmail.com>
CC: "J. Bruce Fields" <bfields@fieldses.org>
CC: "Aneesh Kumar K.V" <aneesh.kumar@linux.vnet.ibm.com>
---
 fs/proc/array.c         |    2 +-
 fs/signalfd.c           |   24 ++++++++++++++++++++++++
 include/linux/proc_fs.h |    3 +++
 3 files changed, 28 insertions(+), 1 deletion(-)

Index: linux-2.6.git/fs/proc/array.c
===================================================================
--- linux-2.6.git.orig/fs/proc/array.c
+++ linux-2.6.git/fs/proc/array.c
@@ -220,7 +220,7 @@ static inline void task_state(struct seq
 	seq_putc(m, '\n');
 }
 
-static void render_sigset_t(struct seq_file *m, const char *header,
+void render_sigset_t(struct seq_file *m, const char *header,
 				sigset_t *set)
 {
 	int i;
Index: linux-2.6.git/fs/signalfd.c
===================================================================
--- linux-2.6.git.orig/fs/signalfd.c
+++ linux-2.6.git/fs/signalfd.c
@@ -29,6 +29,7 @@
 #include <linux/anon_inodes.h>
 #include <linux/signalfd.h>
 #include <linux/syscalls.h>
+#include <linux/proc_fs.h>
 
 void signalfd_cleanup(struct sighand_struct *sighand)
 {
@@ -46,6 +47,7 @@ void signalfd_cleanup(struct sighand_str
 }
 
 struct signalfd_ctx {
+	rwlock_t lock;
 	sigset_t sigmask;
 };
 
@@ -227,7 +229,26 @@ static ssize_t signalfd_read(struct file
 	return total ? total: ret;
 }
 
+#ifdef CONFIG_PROC_FS
+static int signalfd_show_fdinfo(struct seq_file *m, struct file *f)
+{
+	struct signalfd_ctx *ctx = f->private_data;
+	sigset_t sigmask;
+
+	read_lock(&ctx->lock);
+	sigmask = ctx->sigmask;
+	read_unlock(&ctx->lock);
+
+	signotset(&sigmask);
+	render_sigset_t(m, "sigmask:\t", &sigmask);
+	return 0;
+}
+#endif
+
 static const struct file_operations signalfd_fops = {
+#ifdef CONFIG_PROC_FS
+	.show_fdinfo	= signalfd_show_fdinfo,
+#endif
 	.release	= signalfd_release,
 	.poll		= signalfd_poll,
 	.read		= signalfd_read,
@@ -259,6 +280,7 @@ SYSCALL_DEFINE4(signalfd4, int, ufd, sig
 			return -ENOMEM;
 
 		ctx->sigmask = sigmask;
+		rwlock_init(&ctx->lock);
 
 		/*
 		 * When we call this, the initialization must be complete, since
@@ -278,7 +300,9 @@ SYSCALL_DEFINE4(signalfd4, int, ufd, sig
 			return -EINVAL;
 		}
 		spin_lock_irq(&current->sighand->siglock);
+		write_lock(&ctx->lock);
 		ctx->sigmask = sigmask;
+		write_unlock(&ctx->lock);
 		spin_unlock_irq(&current->sighand->siglock);
 
 		wake_up(&current->sighand->signalfd_wqh);
Index: linux-2.6.git/include/linux/proc_fs.h
===================================================================
--- linux-2.6.git.orig/include/linux/proc_fs.h
+++ linux-2.6.git/include/linux/proc_fs.h
@@ -290,4 +290,7 @@ static inline struct net *PDE_NET(struct
 	return pde->parent->data;
 }
 
+#include <linux/signal.h>
+
+void render_sigset_t(struct seq_file *m, const char *header, sigset_t *set);
 #endif /* _LINUX_PROC_FS_H */

Re: + fs-epoll-add-procfs-fdinfo-helper.patch added to -mm tree

[Oleg Nesterov]

On 11/13, Cyrill Gorcunov wrote:
>
>  struct signalfd_ctx {
> +	rwlock_t lock;
>  	sigset_t sigmask;

Oh, I don't think.

rwlock_t is horrible in general, and what it can buy for signalfd?
A plain spinlock would be better. Or seqlock_t.

Whatever you do, you are trying to introduce the lock which should
serialize the access to ->sigmask correctly. In this case I think
you should split this change into 2 patches. The first one should
fix the locking, imo. sys_signalfd4() should not use ->siglock at
all, and the users which take ->siglock to read ->sigmask should be
updated.

Or,

> +#ifdef CONFIG_PROC_FS
> +static int signalfd_show_fdinfo(struct seq_file *m, struct file *f)
> +{
> +	struct signalfd_ctx *ctx = f->private_data;
> +	sigset_t sigmask;
> +
> +	read_lock(&ctx->lock);
> +	sigmask = ctx->sigmask;
> +	read_unlock(&ctx->lock);

Just read ctx->sigmask lockless. Do we really care if show_fdinfo()
reads the value "in between" ?

Oleg.

Re: + fs-epoll-add-procfs-fdinfo-helper.patch added to -mm tree

[Cyrill Gorcunov]

On Tue, Nov 13, 2012 at 05:49:51PM +0100, Oleg Nesterov wrote:
> On 11/13, Cyrill Gorcunov wrote:
> >
> >  struct signalfd_ctx {
> > +	rwlock_t lock;
> >  	sigset_t sigmask;
> 
> Oh, I don't think.
> 
> rwlock_t is horrible in general, and what it can buy for signalfd?
> A plain spinlock would be better. Or seqlock_t.
> 
> Whatever you do, you are trying to introduce the lock which should
> serialize the access to ->sigmask correctly. In this case I think
> you should split this change into 2 patches. The first one should
> fix the locking, imo. sys_signalfd4() should not use ->siglock at
> all, and the users which take ->siglock to read ->sigmask should be
> updated.

I see

> 
> Or,
> 
> > +#ifdef CONFIG_PROC_FS
> > +static int signalfd_show_fdinfo(struct seq_file *m, struct file *f)
> > +{
> > +	struct signalfd_ctx *ctx = f->private_data;
> > +	sigset_t sigmask;
> > +
> > +	read_lock(&ctx->lock);
> > +	sigmask = ctx->sigmask;
> > +	read_unlock(&ctx->lock);
> 
> Just read ctx->sigmask lockless. Do we really care if show_fdinfo()
> reads the value "in between" ?

As from c/r patch I think we can read it lockless (since we do stop
tasks anyway before doing checkpoint). So I would prefer to provide
it without locks at all.

Re: + fs-epoll-add-procfs-fdinfo-helper.patch added to -mm tree

[Cyrill Gorcunov]

On Tue, Nov 13, 2012 at 09:01:59PM +0400, Cyrill Gorcunov wrote:
> 
> As from c/r patch I think we can read it lockless (since we do stop
> tasks anyway before doing checkpoint). So I would prefer to provide
> it without locks at all.

Something like this. (Also I wonder where the documentation about
fdinfo should go?)
---
From: Cyrill Gorcunov <gorcunov@openvz.org>
Subject: fdinfo: Show sigmask for signalfd fd v3

The sigmask is read in lockless manner for a sake of
code simplicity, thus if precise data needed here
the tasks which refer to the signalfd should be
stopped before read.

Signed-off-by: Cyrill Gorcunov <gorcunov@openvz.org>
CC: Pavel Emelyanov <xemul@parallels.com>
CC: Al Viro <viro@ZenIV.linux.org.uk>
CC: Alexey Dobriyan <adobriyan@gmail.com>
CC: Andrew Morton <akpm@linux-foundation.org>
CC: James Bottomley <jbottomley@parallels.com>
CC: "Aneesh Kumar K.V" <aneesh.kumar@linux.vnet.ibm.com>
CC: Alexey Dobriyan <adobriyan@gmail.com>
CC: Matthew Helsley <matt.helsley@gmail.com>
CC: "J. Bruce Fields" <bfields@fieldses.org>
CC: "Aneesh Kumar K.V" <aneesh.kumar@linux.vnet.ibm.com>
---
 fs/proc/array.c         |    2 +-
 fs/signalfd.c           |   18 ++++++++++++++++++
 include/linux/proc_fs.h |    3 +++
 3 files changed, 22 insertions(+), 1 deletion(-)

Index: linux-2.6.git/fs/proc/array.c
===================================================================
--- linux-2.6.git.orig/fs/proc/array.c
+++ linux-2.6.git/fs/proc/array.c
@@ -220,7 +220,7 @@ static inline void task_state(struct seq
 	seq_putc(m, '\n');
 }
 
-static void render_sigset_t(struct seq_file *m, const char *header,
+void render_sigset_t(struct seq_file *m, const char *header,
 				sigset_t *set)
 {
 	int i;
Index: linux-2.6.git/fs/signalfd.c
===================================================================
--- linux-2.6.git.orig/fs/signalfd.c
+++ linux-2.6.git/fs/signalfd.c
@@ -29,6 +29,7 @@
 #include <linux/anon_inodes.h>
 #include <linux/signalfd.h>
 #include <linux/syscalls.h>
+#include <linux/proc_fs.h>
 
 void signalfd_cleanup(struct sighand_struct *sighand)
 {
@@ -227,7 +228,24 @@ static ssize_t signalfd_read(struct file
 	return total ? total: ret;
 }
 
+#ifdef CONFIG_PROC_FS
+static int signalfd_show_fdinfo(struct seq_file *m, struct file *f)
+{
+	struct signalfd_ctx *ctx = f->private_data;
+	sigset_t sigmask;
+
+	sigmask = ctx->sigmask;
+	signotset(&sigmask);
+	render_sigset_t(m, "sigmask:\t", &sigmask);
+
+	return 0;
+}
+#endif
+
 static const struct file_operations signalfd_fops = {
+#ifdef CONFIG_PROC_FS
+	.show_fdinfo	= signalfd_show_fdinfo,
+#endif
 	.release	= signalfd_release,
 	.poll		= signalfd_poll,
 	.read		= signalfd_read,
Index: linux-2.6.git/include/linux/proc_fs.h
===================================================================
--- linux-2.6.git.orig/include/linux/proc_fs.h
+++ linux-2.6.git/include/linux/proc_fs.h
@@ -290,4 +290,7 @@ static inline struct net *PDE_NET(struct
 	return pde->parent->data;
 }
 
+#include <linux/signal.h>
+
+void render_sigset_t(struct seq_file *m, const char *header, sigset_t *set);
 #endif /* _LINUX_PROC_FS_H */

Re: + fs-epoll-add-procfs-fdinfo-helper.patch added to -mm tree

[Oleg Nesterov]

On 11/13, Cyrill Gorcunov wrote:
>
> On Tue, Nov 13, 2012 at 09:01:59PM +0400, Cyrill Gorcunov wrote:
> >
> The sigmask is read in lockless manner for a sake of
> code simplicity, thus if precise data needed here
> the tasks which refer to the signalfd should be
> stopped before read.

Yes, I think this is fine, and this patch should replace
fs-epoll-add-procfs-fdinfo-helper.patch in -mm.




>  static const struct file_operations signalfd_fops = {
> +#ifdef CONFIG_PROC_FS
> +	.show_fdinfo	= signalfd_show_fdinfo,
> +#endif

I am just curious and I can't find the patch which adds .show_fdinfo.
Could you please send me link/patch offlist?

Oleg.

Re: + fs-epoll-add-procfs-fdinfo-helper.patch added to -mm tree

[Cyrill Gorcunov]

On Tue, Nov 13, 2012 at 06:47:06PM +0100, Oleg Nesterov wrote:
> On 11/13, Cyrill Gorcunov wrote:
> >
> > On Tue, Nov 13, 2012 at 09:01:59PM +0400, Cyrill Gorcunov wrote:
> > >
> > The sigmask is read in lockless manner for a sake of
> > code simplicity, thus if precise data needed here
> > the tasks which refer to the signalfd should be
> > stopped before read.
> 
> Yes, I think this is fine, and this patch should replace
> fs-epoll-add-procfs-fdinfo-helper.patch in -mm.

No, epoll is different one, this patch is for
fdinfo-show-sigmask-for-signalfd-fd.patch, I'll send
the updates as new series.
 
> >  static const struct file_operations signalfd_fops = {
> > +#ifdef CONFIG_PROC_FS
> > +	.show_fdinfo	= signalfd_show_fdinfo,
> > +#endif
> 
> I am just curious and I can't find the patch which adds .show_fdinfo.
> Could you please send me link/patch offlist?

It is titled as procfs-add-ability-to-plug-in-auxiliary-fdinfo-providers.patch
Sure I'll resend you it offlist in a minute.

	Cyrill

Re: + fs-epoll-add-procfs-fdinfo-helper.patch added to -mm tree

[Oleg Nesterov]

Damn, sorry for noise...

On 11/13, Oleg Nesterov wrote:
>
> On 11/13, Cyrill Gorcunov wrote:
> >
> > On Tue, Nov 13, 2012 at 09:01:59PM +0400, Cyrill Gorcunov wrote:
> > >
> > The sigmask is read in lockless manner for a sake of
> > code simplicity, thus if precise data needed here
> > the tasks which refer to the signalfd should be
> > stopped before read.
>
> Yes, I think this is fine, and this patch should replace
> fs-epoll-add-procfs-fdinfo-helper.patch in -mm.

I meant fdinfo-show-sigmask-for-signalfd-fd.patch

And the subject was wrong from the very beginning, sorry for confusion.

Oleg.

+ fs-epoll-add-procfs-fdinfo-helper.patch added to -mm tree

[akpm]

The patch titled
     Subject: fs, epoll: add procfs fdinfo helper
has been added to the -mm tree.  Its filename is
     fs-epoll-add-procfs-fdinfo-helper.patch

Before you just go and hit "reply", please:
   a) Consider who else should be cc'ed
   b) Prefer to cc a suitable mailing list as well
   c) Ideally: find the original patch on the mailing list and do a
      reply-to-all to that, adding suitable additional cc's

*** Remember to use Documentation/SubmitChecklist when testing your code ***

The -mm tree is included into linux-next and is updated
there every 3-4 working days

------------------------------------------------------
From: Cyrill Gorcunov <gorcunov@openvz.org>
Subject: fs, epoll: add procfs fdinfo helper

This allow us to print out eventpoll target file descriptor,
events and data, the /proc/pid/fdinfo/fd consists of

 | pos:	0
 | flags:	02
 | tfd:        5 events:       1d data: ffffffffffffffff enabled: 1

[avagin@: fix for unitialized ret variable]

Signed-off-by: Cyrill Gorcunov <gorcunov@openvz.org>
Cc: Pavel Emelyanov <xemul@parallels.com>
Cc: Al Viro <viro@ZenIV.linux.org.uk>
Cc: Alexey Dobriyan <adobriyan@gmail.com>
Cc: James Bottomley <jbottomley@parallels.com>
Cc: "Aneesh Kumar K.V" <aneesh.kumar@linux.vnet.ibm.com>
Cc: Alexey Dobriyan <adobriyan@gmail.com>
Cc: Matthew Helsley <matt.helsley@gmail.com>
Cc: "J. Bruce Fields" <bfields@fieldses.org>
Cc: "Aneesh Kumar K.V" <aneesh.kumar@linux.vnet.ibm.com>
Cc: Andrey Vagin <avagin@openvz.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
---

 fs/eventpoll.c |   29 +++++++++++++++++++++++++++++
 1 file changed, 29 insertions(+)

diff -puN fs/eventpoll.c~fs-epoll-add-procfs-fdinfo-helper fs/eventpoll.c
--- a/fs/eventpoll.c~fs-epoll-add-procfs-fdinfo-helper
+++ a/fs/eventpoll.c
@@ -38,6 +38,8 @@
 #include <asm/io.h>
 #include <asm/mman.h>
 #include <linux/atomic.h>
+#include <linux/proc_fs.h>
+#include <linux/seq_file.h>
 
 /*
  * LOCKING:
@@ -813,8 +815,35 @@ static unsigned int ep_eventpoll_poll(st
 	return pollflags != -1 ? pollflags : 0;
 }
 
+#ifdef CONFIG_PROC_FS
+static int ep_show_fdinfo(struct seq_file *m, struct file *f)
+{
+	struct eventpoll *ep = f->private_data;
+	struct rb_node *rbp;
+	int ret = 0;
+
+	mutex_lock(&ep->mtx);
+	for (rbp = rb_first(&ep->rbr); rbp; rbp = rb_next(rbp)) {
+		struct epitem *epi = rb_entry(rbp, struct epitem, rbn);
+
+		ret = seq_printf(m, "tfd: %8d events: %8x data: %16llx enabled: %d\n",
+				 epi->ffd.fd, epi->event.events,
+				 (long long)epi->event.data,
+				 ep_is_linked(&epi->rdllink));
+		if (ret)
+			break;
+	}
+	mutex_unlock(&ep->mtx);
+
+	return ret;
+}
+#endif
+
 /* File callbacks that implement the eventpoll file behaviour */
 static const struct file_operations eventpoll_fops = {
+#ifdef CONFIG_PROC_FS
+	.show_fdinfo	= ep_show_fdinfo,
+#endif
 	.release	= ep_eventpoll_release,
 	.poll		= ep_eventpoll_poll,
 	.llseek		= noop_llseek,
_

Patches currently in -mm which might be from gorcunov@openvz.org are

linux-next.patch
proc-check-vma-vm_file-before-dereferencing.patch
procfs-add-vmflags-field-in-smaps-output-v4.patch
procfs-add-vmflags-field-in-smaps-output-v4-fix.patch
proc-pid-status-add-seccomp-field.patch
procfs-add-ability-to-plug-in-auxiliary-fdinfo-providers.patch
procfs-add-ability-to-plug-in-auxiliary-fdinfo-providers-fix.patch
fs-exportfs-avoid-nil-dereference-if-no-s_export_op-present.patch
fs-notify-add-file-handle-entry-into-inotify_inode_mark.patch
fs-notify-add-procfs-fdinfo-helper.patch
fs-eventfd-add-procfs-fdinfo-helper.patch
fs-epoll-add-procfs-fdinfo-helper.patch
fdinfo-show-sigmask-for-signalfd-fd.patch
tools-testing-selftests-kcmp-kcmp_testc-print-reason-for-failure-in-kcmp_test.patch