lkml.org 
[lkml]   [2016]   [May]   [3]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
From
SubjectRe: [PATCH 6/6] mm/page_owner: use stackdepot to store stacktrace
On Tue 03-05-16 14:23:04, Joonsoo Kim wrote:
> From: Joonsoo Kim <iamjoonsoo.kim@lge.com>
>
> Currently, we store each page's allocation stacktrace on corresponding
> page_ext structure and it requires a lot of memory. This causes the problem
> that memory tight system doesn't work well if page_owner is enabled.
> Moreover, even with this large memory consumption, we cannot get full
> stacktrace because we allocate memory at boot time and just maintain
> 8 stacktrace slots to balance memory consumption. We could increase it
> to more but it would make system unusable or change system behaviour.
>
> To solve the problem, this patch uses stackdepot to store stacktrace.
> It obviously provides memory saving but there is a drawback that
> stackdepot could fail.
>
> stackdepot allocates memory at runtime so it could fail if system has
> not enough memory. But, most of allocation stack are generated at very
> early time and there are much memory at this time. So, failure would not
> happen easily. And, one failure means that we miss just one page's
> allocation stacktrace so it would not be a big problem. In this patch,
> when memory allocation failure happens, we store special stracktrace
> handle to the page that is failed to save stacktrace. With it, user
> can guess memory usage properly even if failure happens.
>
> Memory saving looks as following. (Boot 4GB memory system with page_owner)
>
> 92274688 bytes -> 25165824 bytes

It is not clear to me whether this is after a fresh boot or some workload
which would grow the stack depot as well. What is a usual cap for the
memory consumption.

> 72% reduction in static allocation size. Even if we should add up size of
> dynamic allocation memory, it would not that big because stacktrace is
> mostly duplicated.
>
> Note that implementation looks complex than someone would imagine because
> there is recursion issue. stackdepot uses page allocator and page_owner
> is called at page allocation. Using stackdepot in page_owner could re-call
> page allcator and then page_owner. That is a recursion.

This is rather fragile. How do we check there is no lock dependency
introduced later on - e.g. split_page called from a different
locking/reclaim context than alloc_pages? Would it be safer to
use ~__GFP_DIRECT_RECLAIM for those stack allocations? Or do you think
there would be too many failed allocations? This alone wouldn't remove a
need for the recursion detection but it sounds less tricky.

> To detect and
> avoid it, whenever we obtain stacktrace, recursion is checked and
> page_owner is set to dummy information if found. Dummy information means
> that this page is allocated for page_owner feature itself
> (such as stackdepot) and it's understandable behavior for user.
>
> Signed-off-by: Joonsoo Kim <iamjoonsoo.kim@lge.com>

I like the idea in general I just wish this would be less subtle. Few
more comments below.

[...]
> -void __set_page_owner(struct page *page, unsigned int order, gfp_t gfp_mask)
> +static inline bool check_recursive_alloc(struct stack_trace *trace,
> + unsigned long ip)
> {
> - struct page_ext *page_ext = lookup_page_ext(page);
> + int i, count;
> +
> + if (!trace->nr_entries)
> + return false;
> +
> + for (i = 0, count = 0; i < trace->nr_entries; i++) {
> + if (trace->entries[i] == ip && ++count == 2)
> + return true;
> + }

This would deserve a comment I guess. Btw, don't we have a better and
more robust way to detect the recursion? Per task_struct flag or
something like that?

[...]
> +static noinline depot_stack_handle_t save_stack(gfp_t flags)
> +{
> + unsigned long entries[PAGE_OWNER_STACK_DEPTH];
> struct stack_trace trace = {
> .nr_entries = 0,
> - .max_entries = ARRAY_SIZE(page_ext->trace_entries),
> - .entries = &page_ext->trace_entries[0],
> - .skip = 3,
> + .entries = entries,
> + .max_entries = PAGE_OWNER_STACK_DEPTH,
> + .skip = 0
> };
[...]
> void __dump_page_owner(struct page *page)
> {
> struct page_ext *page_ext = lookup_page_ext(page);
> + unsigned long entries[PAGE_OWNER_STACK_DEPTH];

This is worrying because of the excessive stack consumption while we
might be in a deep call chain already. Can we preallocate a hash table
for few buffers when the feature is enabled? This would require locking
of course but chances are that contention wouldn't be that large.

> struct stack_trace trace = {
> - .nr_entries = page_ext->nr_entries,
> - .entries = &page_ext->trace_entries[0],
> + .nr_entries = 0,
> + .entries = entries,
> + .max_entries = PAGE_OWNER_STACK_DEPTH,
> + .skip = 0
> };
> + depot_stack_handle_t handle;
> gfp_t gfp_mask = page_ext->gfp_mask;
> int mt = gfpflags_to_migratetype(gfp_mask);
>

Thanks!
--
Michal Hocko
SUSE Labs

\
 
 \ /
  Last update: 2016-05-03 11:21    [W:0.197 / U:8.996 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site