lkml.org 
[lkml]   [2013]   [Mar]   [15]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
From
SubjectRe: [PATCH 1/1] poweroff: change orderly_poweroff() to use schedule_work()
On 03/14, Andrew Morton wrote:
>
> On Wed, 13 Mar 2013 18:47:05 +0100 Oleg Nesterov <oleg@redhat.com> wrote:
>
> > This means that orderly_poweroff() becomes async even if we do not
> > run the command and always succeeds, schedule_work() can only fail
> > if the work is already pending. We can export __orderly_poweroff()
> > and change the non-atomic callers which want the old semantics.
> >
> > ...
> >
> > @@ -2218,21 +2237,9 @@ static int __orderly_poweroff(void)
> > */
> > int orderly_poweroff(bool force)
> > {
> > + if (force) /* do not override the pending "true" */
> > + poweroff_force = true;
> > + schedule_work(&poweroff_work);
> > + return 0;
> > }
>
> afaict the current version of orderly_poweroff() will never return -
> either __orderly_poweroff() will block until the machine shuts down or
> kernel_power_off() will do so.

Note that __orderly_poweroff() uses UMH_WAIT_EXEC, not UMH_WAIT_PROC,
so it returns right after /sbin/poweroff starts to execute.

So it is already asynchronous unless execve() fails.

> However with this patch there is a path via which orderly_poweroff()
> can return to its caller, I think?

See above, but please also read the changelog.

With this patch orderly_poweroff() is always async, even if exec fails,
but

> If so, the caller might be rather
> surprised and we're exercising never-before-used code paths. In fact
> if the surprised caller goes oops, the poweroff might not occur at all.

This should not happen.

Anyway. Please also note that now we can export __orderly_poweroff() and
probably change it, it can have another argument "bool use_UMH_WAIT_PROC".

int __orderly_poweroff(bool force, bool sync)
{
int wait = sync ? UMH_WAIT_EXEC : UMH_WAIT_EXEC;

ret = call_usermodehelper(argv[0], argv, envp, wait);

if (force) {
// EXEC failed or /sbin/poweroff didn't do its work
if (ret || sync)
kernel_power_off();
}
}

The non-atomic callers can use __orderly_poweroff(sync => true).

---------------------------------------------------------------------------
And, Andrew, et all... Could you help with another mentioned problem? It is
really trivial, but exactly because it is trivial I do not know what should
I do.

To remind, say, argv_split(poweroff_cmd) can race with sysctl changing this
string, in this case it can write to the memory after argv[] array. We can
fix this, or we can rewrite argv_split/free:

void argv_free(char **argv)
{
kfree(argv[-1]);
kfree(argv);
}

char **argv_split(gfp_t gfp, const char *str, int *argcp)
{
char *argv_str;
bool was_space;
char **argv, **argv_ret;
int argc;

argv_str = kstrndup(str, KMALLOC_MAX_SIZE, gfp);
if (!argv_str)
return NULL;

argc = count_argc(argv_str);
argv = kmalloc(sizeof(*argv) * (argc + 2), gfp);
if (!argv) {
kfree(argv_str);
return NULL;
}

*argv = argv_str;
argv_ret = ++argv;
for (was_space = true; *argv_str; argv_str++) {
if (isspace(*argv_str)) {
was_space = true;
*argv_str = 0;
} else if (was_space) {
was_space = false;
*argv++ = argv_str;
}
}
*argv = NULL;

if (argcp)
*argcp = argc;
return argv_ret;
}

This way it uses a single kstrndup() to keep the arguments and it is
always safe.

But, whatever we do with argv_split(), it can hit the string "in between".
Personally I think we do not really care, but...

Perhaps we should add proc_dostring_lock() which takes some lock and
modify the callers of argv_split() (or add argv_split_lock) ?

Or perhaps we should introduce the rwsem which should protect every
sysctl-string and proc_dostring() should take this lock?

Help! I'd prefer to rewrite argv_split(), but I agree with any suggestion
in advance.

Oleg.



\
 
 \ /
  Last update: 2013-03-15 18:21    [W:0.463 / U:0.120 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site