mm: avoid 'might_sleep()' in get_mmap_lock_carefully()

This might_sleep() goes back a long time: it was originally introduced
way back when by commit 010060741a ("x86: add might_sleep() to
do_page_fault()"), and made it into the generic VM code when the x86
fault path got re-organized and generalized in commit c2508ec5a5 ("mm:
introduce new 'lock_mm_and_find_vma()' page fault helper").

However, it turns out that the placement of that might_sleep() has
always been rather questionable simply because it's not only a debug
statement to warn about sleeping in contexts that shouldn't sleep (which
was the original reason for adding it), but it also implies a voluntary
scheduling point.

That, in turn, is less than desirable for two reasons:

 (a) it ends up being done after we successfully got the mmap_lock, so
     just as we got the lock we will now eagerly schedule away and
     increase lock contention

and

 (b) this is all very possibly part of the "oops, things went horribly
     wrong" path and we just haven't figured that out yet

After all, the whole _reason_ for having that get_mmap_lock_carefully()
rather than just doing the obvious mmap_read_lock() is because this code
wants to deal somewhat gracefully with potential kernel wild pointer
bugs.

So then a voluntary scheduling point here is simply not a good idea.

We could certainly turn the 'might_sleep()' into a '__might_sleep()' and
make it be just the debug check that it was originally intended to be.

But even that seems questionable in the wild kernel pointer case - which
again is part of the whole point of this code.  The problem wouldn't be
about the _sleeping_ part of the page fault, but about a bad kernel
access.  The fact that that bad kernel access might happen in a section
that you shouldn't sleep in is secondary.

So it really ends up being the case that this is simply entirely the
wrong place to do this debug check and related scheduling point at all.

So let's just remove the check entirely.  It's been around for over a
decade, it has served its purpose.

The re-schedule will happen at return to user space anyway for the
normal case, and the warning - if we even need it - might be better off
done as a special case for "page fault from kernel mode" once we've
dealt with any potential kernel oopses where the oops is the relevant
thing, not some artificial "scheduling while atomic" test.

Reported-by: Mateusz Guzik <mjguzik@gmail.com>
Link: https://lore.kernel.org/lkml/20230820104303.2083444-1-mjguzik@gmail.com/
Cc: Matthew Wilcox <willy@infradead.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
This commit is contained in:
Linus Torvalds 2023-08-21 06:11:33 +02:00
parent 706a741595
commit 4542057e18

View File

@ -5257,11 +5257,8 @@ EXPORT_SYMBOL_GPL(handle_mm_fault);
static inline bool get_mmap_lock_carefully(struct mm_struct *mm, struct pt_regs *regs)
{
/* Even if this succeeds, make it clear we *might* have slept */
if (likely(mmap_read_trylock(mm))) {
might_sleep();
if (likely(mmap_read_trylock(mm)))
return true;
}
if (regs && !user_mode(regs)) {
unsigned long ip = instruction_pointer(regs);