On Thu, Feb 20, 2020 at 5:11 AM Kirill A. Shutemov <kirill@xxxxxxxxxxxxx> wrote:
>
> On Wed, Feb 19, 2020 at 09:01:23AM -0800, Yang Shi wrote:
> > On Tue, Feb 18, 2020 at 7:44 PM David Rientjes <rientjes@xxxxxxxxxx> wrote:
> > >
> > > On Tue, 18 Feb 2020, Yang Shi wrote:
> > >
> > > > > diff --git a/mm/shmem.c b/mm/shmem.c
> > > > > --- a/mm/shmem.c
> > > > > +++ b/mm/shmem.c
> > > > > @@ -1502,9 +1502,8 @@ static struct page *shmem_alloc_page(gfp_t gfp,
> > > > > return page;
> > > > > }
> > > > >
> > > > > -static struct page *shmem_alloc_and_acct_page(gfp_t gfp,
> > > > > - struct inode *inode,
> > > > > - pgoff_t index, bool huge)
> > > > > +static struct page *shmem_alloc_and_acct_page(gfp_t gfp, struct inode *inode,
> > > > > + pgoff_t index, bool fault, bool huge)
> > > > > {
> > > > > struct shmem_inode_info *info = SHMEM_I(inode);
> > > > > struct page *page;
> > > > > @@ -1518,9 +1517,11 @@ static struct page *shmem_alloc_and_acct_page(gfp_t gfp,
> > > > > if (!shmem_inode_acct_block(inode, nr))
> > > > > goto failed;
> > > > >
> > > > > - if (huge)
> > > > > + if (huge) {
> > > > > page = shmem_alloc_hugepage(gfp, info, index);
> > > > > - else
> > > > > + if (!page && fault)
> > > > > + count_vm_event(THP_FAULT_FALLBACK);
> > > > > + } else
> > > > > page = shmem_alloc_page(gfp, info, index);
> > > > > if (page) {
> > > > > __SetPageLocked(page);
> > > > > @@ -1832,11 +1833,10 @@ static int shmem_getpage_gfp(struct inode *inode, pgoff_t index,
> > > > > }
> > > > >
> > > > > alloc_huge:
> > > > > - page = shmem_alloc_and_acct_page(gfp, inode, index, true);
> > > > > + page = shmem_alloc_and_acct_page(gfp, inode, index, vmf, true);
> > > > > if (IS_ERR(page)) {
> > > > > alloc_nohuge:
> > > > > - page = shmem_alloc_and_acct_page(gfp, inode,
> > > > > - index, false);
> > > > > + page = shmem_alloc_and_acct_page(gfp, inode, index, vmf, false);
> > > > > }
> > > > > if (IS_ERR(page)) {
> > > > > int retry = 5;
> > > > > @@ -1871,8 +1871,11 @@ static int shmem_getpage_gfp(struct inode *inode, pgoff_t index,
> > > > >
> > > > > error = mem_cgroup_try_charge_delay(page, charge_mm, gfp, &memcg,
> > > > > PageTransHuge(page));
> > > > > - if (error)
> > > > > + if (error) {
> > > > > + if (vmf && PageTransHuge(page))
> > > > > + count_vm_event(THP_FAULT_FALLBACK);
> > > > > goto unacct;
> > > > > + }
> > > > > error = shmem_add_to_page_cache(page, mapping, hindex,
> > > > > NULL, gfp & GFP_RECLAIM_MASK);
> > > > > if (error) {
> > > > > @@ -1883,6 +1886,8 @@ static int shmem_getpage_gfp(struct inode *inode, pgoff_t index,
> > > > > mem_cgroup_commit_charge(page, memcg, false,
> > > > > PageTransHuge(page));
> > > > > lru_cache_add_anon(page);
> > > > > + if (vmf && PageTransHuge(page))
> > > > > + count_vm_event(THP_FAULT_ALLOC);
> > > >
> > > > I think shmem THP alloc is accounted to THP_FILE_ALLOC. And it has
> > > > been accounted by shmem_add_to_page_cache(). So, it sounds like a
> > > > double count.
> > > >
> > >
> > > I think we can choose to either include file allocations into both
> > > thp_fault_alloc and thp_fault_fallback or we can exclude them from both of
> > > them. I don't think we can account for only one of them.
> >
> > How's about the 3rd option, adding THP_FILE_FALLBACK.
>
> I like this option.
>
> Problem with THP_FAULT_* is that shmem_getpage_gfp() is called not only
> from fault path, but also from syscalls.
I found another usecase for THP_FILE_FALLBACK. I wanted to measure
file THP allocation success rate in our uecase. It looks nr_file_alloc
/ (nr_file_alloc + nr_file_fallback) is the most simple way.
David, are you still working on this patch?
>
> --
> Kirill A. Shutemov
