On Thu, 1 Apr 2021 17:05:19 -0700
Andrew Morton <akpm@xxxxxxxxxxxxxxxxxxxx> wrote:
> On Thu, 1 Apr 2021 23:30:10 +0100 Sergei Trofimovich <slyfox@xxxxxxxxxx> wrote:
>
> > Before the change page_owner recursion was detected via fetching
> > backtrace and inspecting it for current instruction pointer.
> > It has a few problems:
> > - it is slightly slow as it requires extra backtrace and a linear
> > stack scan of the result
> > - it is too late to check if backtrace fetching required memory
> > allocation itself (ia64's unwinder requires it).
> >
> > To simplify recursion tracking let's use page_owner recursion depth
> > as a counter in 'struct task_struct'.
>
> Seems like a better approach.
>
> > The change make page_owner=on work on ia64 bu avoiding infinite
> > recursion in:
> > kmalloc()
> > -> __set_page_owner()
> > -> save_stack()
> > -> unwind() [ia64-specific]
> > -> build_script()
> > -> kmalloc()
> > -> __set_page_owner() [we short-circuit here]
> > -> save_stack()
> > -> unwind() [recursion]
> >
> > ...
> >
> > --- a/include/linux/sched.h
> > +++ b/include/linux/sched.h
> > @@ -1371,6 +1371,15 @@ struct task_struct {
> > struct llist_head kretprobe_instances;
> > #endif
> >
> > +#ifdef CONFIG_PAGE_OWNER
> > + /*
> > + * Used by page_owner=on to detect recursion in page tracking.
> > + * Is it fine to have non-atomic ops here if we ever access
> > + * this variable via current->page_owner_depth?
>
> Yes, it is fine. This part of the comment can be removed.
Cool! Will do.
> > + */
> > + unsigned int page_owner_depth;
> > +#endif
>
> Adding to the task_struct has a cost. But I don't expect that
> PAGE_OWNER is commonly used in prodction builds (correct?).
Yeah, PAGE_OWNER should not be enabled for production kernels.
Not having extra memory overhead (or layout disruption) is a nice
benefit though. I'll switch to "Unserialized, strictly 'current'" bitfield.
> > --- a/init/init_task.c
> > +++ b/init/init_task.c
> > @@ -213,6 +213,9 @@ struct task_struct init_task
> > #ifdef CONFIG_SECCOMP
> > .seccomp = { .filter_count = ATOMIC_INIT(0) },
> > #endif
> > +#ifdef CONFIG_PAGE_OWNER
> > + .page_owner_depth = 0,
> > +#endif
> > };
> > EXPORT_SYMBOL(init_task);
>
> It will be initialized to zero by the compiler. We can omit this hunk
> entirely.
>
> > --- a/mm/page_owner.c
> > +++ b/mm/page_owner.c
> > @@ -20,6 +20,16 @@
> > */
> > #define PAGE_OWNER_STACK_DEPTH (16)
> >
> > +/*
> > + * How many reenters we allow to page_owner.
> > + *
> > + * Sometimes metadata allocation tracking requires more memory to be allocated:
> > + * - when new stack trace is saved to stack depot
> > + * - when backtrace itself is calculated (ia64)
> > + * Instead of falling to infinite recursion give it a chance to recover.
> > + */
> > +#define PAGE_OWNER_MAX_RECURSION_DEPTH (1)
>
> So this is presently a boolean. Is there any expectation that
> PAGE_OWNER_MAX_RECURSION_DEPTH will ever be greater than 1? If not, we
> could use a single bit in the task_struct. Add it to the
> "Unserialized, strictly 'current'" bitfields. Could make it a 2-bit field if we want
> to permit PAGE_OWNER_MAX_RECURSION_DEPTH=larger.
Let's settle on depth=1. depth>1 is not trivial for other reasons I don't
completely understand.
Follow-up patch incoming.
--
Sergei
