Re: [PATCH] proc: #68991 Fix GPF in /proc/$PID/map_files

Re: [PATCH] proc: #68991 Fix GPF in /proc/$PID/map_files

[Pavel Emelyanov]

On 03/08/2014 06:20 AM, Artem Fetishev wrote:
> https://bugzilla.kernel.org/show_bug.cgi?id=68991
> 
> The expected logic of proc_map_files_get_link() is either to return 0 and
> initialize 'path' or return an error and leave 'path' uninitialized. By the time
> dname_to_vma_addr() returns 0 the corresponding vma may have already be gone. In
> this case the path is not initialized but the return value is still 0. This
> results in 'general protection fault' inside d_path().
> 
> Steps to reproduce:
> CONFIG_CHECKPOINT_RESTORE=y
> 
> fd = open(...);
> while (1) {
> 	mmap(fd, ...);
> 	munmap(fd, ...);
> }
> 
> ls -la /proc/$PID/map_files
> 
> Signed-off-by: Artem Fetishev <artem_fetishev@epam.com>
> Signed-off-by: Aleksandr Terekhov <aleksandr_terekhov@epam.com>

Acked-by: Pavel Emelyanov <xemul@parallels.com>


> ---
> diff --git a/fs/proc/base.c b/fs/proc/base.c
> index 5150706..b976062 100644
> --- a/fs/proc/base.c
> +++ b/fs/proc/base.c
> @@ -1824,6 +1824,7 @@ static int proc_map_files_get_link(struct dentry *dentry, struct path *path)
>  	if (rc)
>  		goto out_mmput;
>  
> +	rc = -ENOENT;
>  	down_read(&mm->mmap_sem);
>  	vma = find_exact_vma(mm, vm_start, vm_end);
>  	if (vma && vma->vm_file) {
> .
> 

Re: [PATCH] proc: #68991 Fix GPF in /proc/$PID/map_files

[Cyrill Gorcunov]

On Fri, Mar 07, 2014 at 06:29:13PM +0400, Pavel Emelyanov wrote:
> On 03/08/2014 06:20 AM, Artem Fetishev wrote:
> > https://bugzilla.kernel.org/show_bug.cgi?id=68991
> > 
> > The expected logic of proc_map_files_get_link() is either to return 0 and
> > initialize 'path' or return an error and leave 'path' uninitialized. By the time
> > dname_to_vma_addr() returns 0 the corresponding vma may have already be gone. In
> > this case the path is not initialized but the return value is still 0. This
> > results in 'general protection fault' inside d_path().
> > 
> > Steps to reproduce:
> > CONFIG_CHECKPOINT_RESTORE=y
> > 
> > fd = open(...);
> > while (1) {
> > 	mmap(fd, ...);
> > 	munmap(fd, ...);
> > }
> > 
> > ls -la /proc/$PID/map_files
> > 
> > Signed-off-by: Artem Fetishev <artem_fetishev@epam.com>
> > Signed-off-by: Aleksandr Terekhov <aleksandr_terekhov@epam.com>
> 
> Acked-by: Pavel Emelyanov <xemul@parallels.com>

Good catch, thanks!

Acked-by: Cyrill Gorcunov <gorcunov@openvz.org>

I suppose this is a stable@ material.

Re: [PATCH] proc: #68991 Fix GPF in /proc/$PID/map_files

[Eric W. Biederman]

Artem Fetishev <artem_fetishev@epam.com> writes:

> https://bugzilla.kernel.org/show_bug.cgi?id=68991
>
> The expected logic of proc_map_files_get_link() is either to return 0 and
> initialize 'path' or return an error and leave 'path' uninitialized. By the time
> dname_to_vma_addr() returns 0 the corresponding vma may have already be gone. In
> this case the path is not initialized but the return value is still 0. This
> results in 'general protection fault' inside d_path().
>
> Steps to reproduce:
> CONFIG_CHECKPOINT_RESTORE=y
>
> fd = open(...);
> while (1) {
> 	mmap(fd, ...);
> 	munmap(fd, ...);
> }
>
> ls -la /proc/$PID/map_files
>
> Signed-off-by: Artem Fetishev <artem_fetishev@epam.com>
> Signed-off-by: Aleksandr Terekhov <aleksandr_terekhov@epam.com>
Reviewed-by: "Eric W. Biederman" <ebiederm@xmission.com>
> ---
> diff --git a/fs/proc/base.c b/fs/proc/base.c
> index 5150706..b976062 100644
> --- a/fs/proc/base.c
> +++ b/fs/proc/base.c
> @@ -1824,6 +1824,7 @@ static int proc_map_files_get_link(struct dentry *dentry, struct path *path)
>  	if (rc)
>  		goto out_mmput;
>  
> +	rc = -ENOENT;
>  	down_read(&mm->mmap_sem);
>  	vma = find_exact_vma(mm, vm_start, vm_end);
>  	if (vma && vma->vm_file) {
> _______________________________________________
> Containers mailing list
> Containers@lists.linux-foundation.org
> https://lists.linuxfoundation.org/mailman/listinfo/containers

[PATCH] proc: #68991 Fix GPF in /proc/$PID/map_files

[Artem Fetishev]

https://bugzilla.kernel.org/show_bug.cgi?id=68991

The expected logic of proc_map_files_get_link() is either to return 0 and
initialize 'path' or return an error and leave 'path' uninitialized. By the time
dname_to_vma_addr() returns 0 the corresponding vma may have already be gone. In
this case the path is not initialized but the return value is still 0. This
results in 'general protection fault' inside d_path().

Steps to reproduce:
CONFIG_CHECKPOINT_RESTORE=y

fd = open(...);
while (1) {
	mmap(fd, ...);
	munmap(fd, ...);
}

ls -la /proc/$PID/map_files

Signed-off-by: Artem Fetishev <artem_fetishev@epam.com>
Signed-off-by: Aleksandr Terekhov <aleksandr_terekhov@epam.com>
---
diff --git a/fs/proc/base.c b/fs/proc/base.c
index 5150706..b976062 100644
--- a/fs/proc/base.c
+++ b/fs/proc/base.c
@@ -1824,6 +1824,7 @@ static int proc_map_files_get_link(struct dentry *dentry, struct path *path)
 	if (rc)
 		goto out_mmput;
 
+	rc = -ENOENT;
 	down_read(&mm->mmap_sem);
 	vma = find_exact_vma(mm, vm_start, vm_end);
 	if (vma && vma->vm_file) {