Re: [PATCH v8 6/9] secretmem: add memcg accounting

From: Mike Rapoport
Date: Sun Nov 15 2020 - 04:18:48 EST


On Fri, Nov 13, 2020 at 03:42:25PM -0800, Roman Gushchin wrote:
> вт, 10 нояб. 2020 г. в 07:16, Mike Rapoport <rppt@xxxxxxxxxx>:
> >
> > From: Mike Rapoport <rppt@xxxxxxxxxxxxx>
> >
> > Account memory consumed by secretmem to memcg. The accounting is updated
> > when the memory is actually allocated and freed.
> >
> > Signed-off-by: Mike Rapoport <rppt@xxxxxxxxxxxxx>
> > ---
> > mm/filemap.c | 2 +-
> > mm/secretmem.c | 42 +++++++++++++++++++++++++++++++++++++++++-
> > 2 files changed, 42 insertions(+), 2 deletions(-)
> >
> > diff --git a/mm/filemap.c b/mm/filemap.c
> > index 249cf489f5df..11387a077373 100644
> > --- a/mm/filemap.c
> > +++ b/mm/filemap.c
> > @@ -844,7 +844,7 @@ static noinline int __add_to_page_cache_locked(struct page *page,
> > page->mapping = mapping;
> > page->index = offset;
> >
> > - if (!huge) {
> > + if (!huge && !page->memcg_data) {
> > error = mem_cgroup_charge(page, current->mm, gfp);
> > if (error)
> > goto error;
> > diff --git a/mm/secretmem.c b/mm/secretmem.c
> > index 1aa2b7cffe0d..1eb7667016fa 100644
> > --- a/mm/secretmem.c
> > +++ b/mm/secretmem.c
> > @@ -17,6 +17,7 @@
> > #include <linux/syscalls.h>
> > #include <linux/memblock.h>
> > #include <linux/pseudo_fs.h>
> > +#include <linux/memcontrol.h>
> > #include <linux/set_memory.h>
> > #include <linux/sched/signal.h>
> >
> > @@ -49,6 +50,38 @@ struct secretmem_ctx {
> >
> > static struct cma *secretmem_cma;
> >
>
> Hi Mike!
>
> > +static int secretmem_memcg_charge(struct page *page, gfp_t gfp, int order)
> > +{
> > + unsigned long nr_pages = (1 << order);
> > + int i, err;
> > +
> > + err = memcg_kmem_charge_page(page, gfp, order);
> > + if (err)
> > + return err;
> > +
> > + for (i = 1; i < nr_pages; i++) {
> > + struct page *p = page + i;
> > +
> > + p->memcg_data = page->memcg_data;
> > + }
>
> Hm, it looks very strange to me. Why do we need to copy memcg_data?
> What about css reference counting?

I need to copy memcg_data to mark a page as being accounted so it won't
be charged again when it is added to page cache.

What happens here is that I allocate a large page and then use it as a
local cache for allocations in secretmem_fault(). I charge the large
page as kmem.

During secretmem_fault() a small sub-page from that large page goes into
page cache and there I skip its memcg accounting.

In the end, when the large page is freed, the memcg_data for all its
sub-pages is cleared and I uncharge memcg with the order of large page.

An alternative would be to uncharge a small page from kmem in
secretmem_fault() and make this page charged in add_to_page_cache(), but
that would complicate the release path as I would need to re-charge the
small page back to kmem at secretmem_freepage() and track all the
participating memcgs till the large page is freed.

> And what about statistics?

Hmm, that's probably won't be accurate :-/

> I'm sorry for being late.
>
> Thank you!
>
> > +
> > + return 0;
> > +}
> > +
> > +static void secretmem_memcg_uncharge(struct page *page, int order)
> > +{
> > + unsigned long nr_pages = (1 << order);
> > + int i;
> > +
> > + for (i = 1; i < nr_pages; i++) {
> > + struct page *p = page + i;
> > +
> > + p->memcg_data = 0;
> > + }
> > +
> > + memcg_kmem_uncharge_page(page, PMD_PAGE_ORDER);
> > +}
> > +
> > static int secretmem_pool_increase(struct secretmem_ctx *ctx, gfp_t gfp)
> > {
> > unsigned long nr_pages = (1 << PMD_PAGE_ORDER);
> > @@ -61,10 +94,14 @@ static int secretmem_pool_increase(struct secretmem_ctx *ctx, gfp_t gfp)
> > if (!page)
> > return -ENOMEM;
> >
> > - err = set_direct_map_invalid_noflush(page, nr_pages);
> > + err = secretmem_memcg_charge(page, gfp, PMD_PAGE_ORDER);
> > if (err)
> > goto err_cma_release;
> >
> > + err = set_direct_map_invalid_noflush(page, nr_pages);
> > + if (err)
> > + goto err_memcg_uncharge;
> > +
> > addr = (unsigned long)page_address(page);
> > err = gen_pool_add(pool, addr, PMD_SIZE, NUMA_NO_NODE);
> > if (err)
> > @@ -81,6 +118,8 @@ static int secretmem_pool_increase(struct secretmem_ctx *ctx, gfp_t gfp)
> > * won't fail
> > */
> > set_direct_map_default_noflush(page, nr_pages);
> > +err_memcg_uncharge:
> > + secretmem_memcg_uncharge(page, PMD_PAGE_ORDER);
> > err_cma_release:
> > cma_release(secretmem_cma, page, nr_pages);
> > return err;
> > @@ -310,6 +349,7 @@ static void secretmem_cleanup_chunk(struct gen_pool *pool,
> > int i;
> >
> > set_direct_map_default_noflush(page, nr_pages);
> > + secretmem_memcg_uncharge(page, PMD_PAGE_ORDER);
> >
> > for (i = 0; i < nr_pages; i++)
> > clear_highpage(page + i);
> > --
> > 2.28.0
> >
> >

--
Sincerely yours,
Mike.