Re: [PATCH] mm/page_owner: Record timestamp and pid

[Vlastimil Babka]

On 11/27/20 8:23 PM, Souptick Joarder wrote:
> On Sat, Nov 28, 2020 at 12:36 AM Vlastimil Babka <vbabka@xxxxxxx> wrote:
> > On 11/27/20 7:57 PM, Georgi Djakov wrote:
> > > Hi Vlastimil,
> > >
> > > Thanks for the comment!
> > >
> > > On 11/27/20 19:52, Vlastimil Babka wrote:
> > >> On 11/12/20 8:14 PM, Andrew Morton wrote:
> > >>> On Thu, 12 Nov 2020 20:41:06 +0200 Georgi Djakov <georgi.djakov@xxxxxxxxxx>
> > >>> wrote:
> > >>>
> > >>>> From: Liam Mark <lmark@xxxxxxxxxxxxxx>
> > >>>>
> > >>>> Collect the time for each allocation recorded in page owner so that
> > >>>> allocation "surges" can be measured.
> > >>>>
> > >>>> Record the pid for each allocation recorded in page owner so that
> > >>>> the source of allocation "surges" can be better identified.
> > >>>
> > >>> Please provide a description of why this is considered useful.  What
> > >>> has it been used for, what problems has it been used to solve?
> > >>
> > >> Worth noting that on x86_64 it doubles the size of struct page_owner
> > >> from 16 bytes to 32, so it better be justified:
> > >
> > > Well, that's true. But for debug options there is almost always some penalty.
> > > The timestamp and pid information is very useful for me (and others, i believe)
> > > when doing memory analysis. On a crash for example, we can get this information
> > > from kdump (or RAM-dump) and look into it to catch memory allocation problems
> > > more easily.
> >
> > Right. Btw, you should add printing the info to __dump_page_owner().
> >
> > > If you find the above argument not strong enough, how about a separate config
> > > option for this? Maybe something like CONFIG_PAGE_OWNER_EXTENDED, which could
> > > be enabled in addition to CONFIG_PAGE_OWNER?
> >
> > It might be strong enough if it's mentioned in changelog, and also what exactly
> > the space tradeoff is :)
>
> Just a thought ... putting it inside CONFIG_PAGE_OWNER_DEBUG might be
> better if it is used
> purely for debugging purposes.

I don't think we need to introduce new config just yet, until someone makes the 
case for it. Even then, it could be instead doable as an extension to 
"page_owner=on" boot option.
I mean let's add those fields, but improve the changelog.

> > You can also mention that SLUB object tracking has also pid+timestamp.
> >
> > > Thanks,
> > > Georgi
> > >
> > >>
> > >> struct page_owner {
> > >>          short unsigned int         order;                /*     0     2 */
> > >>          short int                  last_migrate_reason;  /*     2     2 */
> > >>          gfp_t                      gfp_mask;             /*     4     4 */
> > >>          depot_stack_handle_t       handle;               /*     8     4 */
> > >>          depot_stack_handle_t       free_handle;          /*    12     4 */
> > >>          u64                        ts_nsec;              /*    16     8 */
> > >>          int                        pid;                  /*    24     4 */
> > >>
> > >>          /* size: 32, cachelines: 1, members: 7 */
> > >>          /* padding: 4 */
> > >>          /* last cacheline: 32 bytes */
> > >> };
> > >>
> > >