| Message ID | 20200201034029.4063170-11-jhubbard@nvidia.com (mailing list archive) |
|---|---|
| State | Superseded |
| Headers | show |
| Series | mm/gup: track FOLL_PIN pages | expand |
On Fri, Jan 31, 2020 at 07:40:27PM -0800, John Hubbard wrote: > diff --git a/mm/gup.c b/mm/gup.c > index c10d0d051c5b..9fe61d15fc0e 100644 > --- a/mm/gup.c > +++ b/mm/gup.c > @@ -29,6 +29,19 @@ struct follow_page_context { > unsigned int page_mask; > }; > > +#ifdef CONFIG_DEBUG_VM Why under CONFIG_DEBUG_VM? There's nothing about this in the cover letter. > +static inline void __update_proc_vmstat(struct page *page, > + enum node_stat_item item, int count) > +{ > + mod_node_page_state(page_pgdat(page), item, count); > +} > +#else > +static inline void __update_proc_vmstat(struct page *page, > + enum node_stat_item item, int count) > +{ > +} > +#endif > + > static void hpage_pincount_add(struct page *page, int refs) > { > VM_BUG_ON_PAGE(!hpage_pincount_available(page), page);
On 2/3/20 5:53 AM, Kirill A. Shutemov wrote: > On Fri, Jan 31, 2020 at 07:40:27PM -0800, John Hubbard wrote: >> diff --git a/mm/gup.c b/mm/gup.c >> index c10d0d051c5b..9fe61d15fc0e 100644 >> --- a/mm/gup.c >> +++ b/mm/gup.c >> @@ -29,6 +29,19 @@ struct follow_page_context { >> unsigned int page_mask; >> }; >> >> +#ifdef CONFIG_DEBUG_VM > > Why under CONFIG_DEBUG_VM? There's nothing about this in the cover letter. > Early on, gup_benchmark showed a really significant slowdown from using these counters. And I don't doubt that it's still the case. I'll re-measure and add a short summary and a few numbers to the patch commit description, and to the v4 cover letter. thanks,
On Mon, Feb 03, 2020 at 01:04:04PM -0800, John Hubbard wrote: > On 2/3/20 5:53 AM, Kirill A. Shutemov wrote: > > On Fri, Jan 31, 2020 at 07:40:27PM -0800, John Hubbard wrote: > >> diff --git a/mm/gup.c b/mm/gup.c > >> index c10d0d051c5b..9fe61d15fc0e 100644 > >> --- a/mm/gup.c > >> +++ b/mm/gup.c > >> @@ -29,6 +29,19 @@ struct follow_page_context { > >> unsigned int page_mask; > >> }; > >> > >> +#ifdef CONFIG_DEBUG_VM > > > > Why under CONFIG_DEBUG_VM? There's nothing about this in the cover letter. > > > > Early on, gup_benchmark showed a really significant slowdown from using these > counters. And I don't doubt that it's still the case. > > I'll re-measure and add a short summary and a few numbers to the patch commit > description, and to the v4 cover letter. Looks like you'll show zeros for these counters if debug is off. It can be confusing to the user. I think these counters should go away if you don't count them.
On 2/3/20 1:30 PM, Kirill A. Shutemov wrote: > On Mon, Feb 03, 2020 at 01:04:04PM -0800, John Hubbard wrote: >> On 2/3/20 5:53 AM, Kirill A. Shutemov wrote: >>> On Fri, Jan 31, 2020 at 07:40:27PM -0800, John Hubbard wrote: >>>> diff --git a/mm/gup.c b/mm/gup.c >>>> index c10d0d051c5b..9fe61d15fc0e 100644 >>>> --- a/mm/gup.c >>>> +++ b/mm/gup.c >>>> @@ -29,6 +29,19 @@ struct follow_page_context { >>>> unsigned int page_mask; >>>> }; >>>> >>>> +#ifdef CONFIG_DEBUG_VM >>> >>> Why under CONFIG_DEBUG_VM? There's nothing about this in the cover letter. >>> >> >> Early on, gup_benchmark showed a really significant slowdown from using these >> counters. And I don't doubt that it's still the case. >> >> I'll re-measure and add a short summary and a few numbers to the patch commit >> description, and to the v4 cover letter. > > Looks like you'll show zeros for these counters if debug is off. It can be > confusing to the user. I think these counters should go away if you don't > count them. > OK, that's a good point. (And in fact, the counters==0 situation already led me astray briefly while debugging with Leon R, even. heh.) I'll remove them entirely for the !CONFIG_DEBUG_VM case. thanks,
On 2/3/20 1:34 PM, John Hubbard wrote: > On 2/3/20 1:30 PM, Kirill A. Shutemov wrote: >> On Mon, Feb 03, 2020 at 01:04:04PM -0800, John Hubbard wrote: >>> On 2/3/20 5:53 AM, Kirill A. Shutemov wrote: >>>> On Fri, Jan 31, 2020 at 07:40:27PM -0800, John Hubbard wrote: >>>>> diff --git a/mm/gup.c b/mm/gup.c >>>>> index c10d0d051c5b..9fe61d15fc0e 100644 >>>>> --- a/mm/gup.c >>>>> +++ b/mm/gup.c >>>>> @@ -29,6 +29,19 @@ struct follow_page_context { >>>>> unsigned int page_mask; >>>>> }; >>>>> >>>>> +#ifdef CONFIG_DEBUG_VM >>>> >>>> Why under CONFIG_DEBUG_VM? There's nothing about this in the cover letter. >>>> >>> >>> Early on, gup_benchmark showed a really significant slowdown from using these >>> counters. And I don't doubt that it's still the case. >>> >>> I'll re-measure and add a short summary and a few numbers to the patch commit >>> description, and to the v4 cover letter. >> >> Looks like you'll show zeros for these counters if debug is off. It can be >> confusing to the user. I think these counters should go away if you don't >> count them. >> > > OK, that's a good point. (And in fact, the counters==0 situation already led me > astray briefly while debugging with Leon R, even. heh.) I'll remove them entirely for > the !CONFIG_DEBUG_VM case. > On second thought, let me do some more careful performance testing. I don't recall now if I was just removing every possible perf slowdown item, when I made this decision. It could be that the perf is not affected, and I could just leave this feature enabled at all times, which would be nicer. And after all, these counters were designed for pretty hot-path items. I'll report back with results... thanks,
On 2/3/20 3:16 PM, John Hubbard wrote: > On 2/3/20 1:34 PM, John Hubbard wrote: >> On 2/3/20 1:30 PM, Kirill A. Shutemov wrote: >>> On Mon, Feb 03, 2020 at 01:04:04PM -0800, John Hubbard wrote: >>>> On 2/3/20 5:53 AM, Kirill A. Shutemov wrote: >>>>> On Fri, Jan 31, 2020 at 07:40:27PM -0800, John Hubbard wrote: >>>>>> diff --git a/mm/gup.c b/mm/gup.c >>>>>> index c10d0d051c5b..9fe61d15fc0e 100644 >>>>>> --- a/mm/gup.c >>>>>> +++ b/mm/gup.c >>>>>> @@ -29,6 +29,19 @@ struct follow_page_context { >>>>>> unsigned int page_mask; >>>>>> }; >>>>>> >>>>>> +#ifdef CONFIG_DEBUG_VM >>>>> >>>>> Why under CONFIG_DEBUG_VM? There's nothing about this in the cover letter. >>>>> >>>> >>>> Early on, gup_benchmark showed a really significant slowdown from using these >>>> counters. And I don't doubt that it's still the case. >>>> >>>> I'll re-measure and add a short summary and a few numbers to the patch commit >>>> description, and to the v4 cover letter. >>> >>> Looks like you'll show zeros for these counters if debug is off. It can be >>> confusing to the user. I think these counters should go away if you don't >>> count them. >>> >> >> OK, that's a good point. (And in fact, the counters==0 situation already led me >> astray briefly while debugging with Leon R, even. heh.) I'll remove them entirely for >> the !CONFIG_DEBUG_VM case. >> > > On second thought, let me do some more careful performance testing. I don't recall > now if I was just removing every possible perf slowdown item, when I made this decision. > It could be that the perf is not affected, and I could just leave this feature enabled > at all times, which would be nicer. > > And after all, these counters were designed for pretty hot-path items. I'll report back > with results... Sure enough, any perf effects are hidden in the noise. I'll just remove the CONFIG_DEBUG_VM checks. Glad you asked about this! thanks,
diff --git a/include/linux/mmzone.h b/include/linux/mmzone.h index c2bc309d1634..01d690586206 100644 --- a/include/linux/mmzone.h +++ b/include/linux/mmzone.h @@ -243,6 +243,8 @@ enum node_stat_item { NR_DIRTIED, /* page dirtyings since bootup */ NR_WRITTEN, /* page writings since bootup */ NR_KERNEL_MISC_RECLAIMABLE, /* reclaimable non-slab kernel pages */ + NR_FOLL_PIN_REQUESTED, /* via: pin_user_page(), gup flag: FOLL_PIN */ + NR_FOLL_PIN_RETURNED, /* pages returned via unpin_user_page() */ NR_VM_NODE_STAT_ITEMS }; diff --git a/mm/gup.c b/mm/gup.c index c10d0d051c5b..9fe61d15fc0e 100644 --- a/mm/gup.c +++ b/mm/gup.c @@ -29,6 +29,19 @@ struct follow_page_context { unsigned int page_mask; }; +#ifdef CONFIG_DEBUG_VM +static inline void __update_proc_vmstat(struct page *page, + enum node_stat_item item, int count) +{ + mod_node_page_state(page_pgdat(page), item, count); +} +#else +static inline void __update_proc_vmstat(struct page *page, + enum node_stat_item item, int count) +{ +} +#endif + static void hpage_pincount_add(struct page *page, int refs) { VM_BUG_ON_PAGE(!hpage_pincount_available(page), page); @@ -86,6 +99,8 @@ static __maybe_unused struct page *try_grab_compound_head(struct page *page, if (flags & FOLL_GET) return try_get_compound_head(page, refs); else if (flags & FOLL_PIN) { + int orig_refs = refs; + /* * When pinning a compound page of order > 1 (which is what * hpage_pincount_available() checks for), use an exact count to @@ -104,6 +119,7 @@ static __maybe_unused struct page *try_grab_compound_head(struct page *page, if (hpage_pincount_available(page)) hpage_pincount_add(page, refs); + __update_proc_vmstat(page, NR_FOLL_PIN_REQUESTED, orig_refs); return page; } @@ -158,6 +174,8 @@ bool __must_check try_grab_page(struct page *page, unsigned int flags) * once, so that the page really is pinned. */ page_ref_add(page, refs); + + __update_proc_vmstat(page, NR_FOLL_PIN_REQUESTED, 1); } return true; @@ -178,6 +196,7 @@ static bool __unpin_devmap_managed_user_page(struct page *page) count = page_ref_sub_return(page, refs); + __update_proc_vmstat(page, NR_FOLL_PIN_RETURNED, 1); /* * devmap page refcounts are 1-based, rather than 0-based: if * refcount is 1, then the page is free and the refcount is @@ -228,6 +247,8 @@ void unpin_user_page(struct page *page) if (page_ref_sub_and_test(page, refs)) __put_page(page); + + __update_proc_vmstat(page, NR_FOLL_PIN_RETURNED, 1); } EXPORT_SYMBOL(unpin_user_page); diff --git a/mm/vmstat.c b/mm/vmstat.c index 78d53378db99..b56808bae1b4 100644 --- a/mm/vmstat.c +++ b/mm/vmstat.c @@ -1168,6 +1168,8 @@ const char * const vmstat_text[] = { "nr_dirtied", "nr_written", "nr_kernel_misc_reclaimable", + "nr_foll_pin_requested", + "nr_foll_pin_returned", /* enum writeback_stat_item counters */ "nr_dirty_threshold",
Now that pages are "DMA-pinned" via pin_user_page*(), and unpinned via unpin_user_pages*(), we need some visibility into whether all of this is working correctly. Add two new fields to /proc/vmstat: nr_foll_pin_requested nr_foll_pin_returned These are documented in Documentation/core-api/pin_user_pages.rst. They represent the number of pages (since boot time) that have been pinned ("nr_foll_pin_requested") and unpinned ("nr_foll_pin_returned"), via pin_user_pages*() and unpin_user_pages*(). In the absence of long-running DMA or RDMA operations that hold pages pinned, the above two fields will normally be equal to each other. Signed-off-by: John Hubbard <jhubbard@nvidia.com> --- include/linux/mmzone.h | 2 ++ mm/gup.c | 21 +++++++++++++++++++++ mm/vmstat.c | 2 ++ 3 files changed, 25 insertions(+)