On Sun, Aug 4, 2019 at 9:28 PM Minchan Kim <minc...@kernel.org> wrote:
>
> Hi Henry,
>
> On Thu, Aug 01, 2019 at 06:53:32PM -0700, Henry Burns wrote:
> > In zs_destroy_pool() we call flush_work(&pool->free_work). However, we
> > have no guarantee that migration isn't happening in the background
> > at that time.
> >
> > Since migration can't directly free pages, it relies on free_work
> > being scheduled to free the pages. But there's nothing preventing an
> > in-progress migrate from queuing the work *after*
> > zs_unregister_migration() has called flush_work(). Which would mean
> > pages still pointing at the inode when we free it.
>
> We already unregister shrinker so there is no upcoming async free call
> via shrinker so the only concern is zs_compact API direct call from
> the user. Is that what what you desribe from the description?
What I am describing is a call to zsmalloc_aops->migratepage() by
kcompactd (which can call schedule work in either
zs_page_migrate() or zs_page_putback should the zspage become empty).
While we are migrating a page, we remove it from the class. Suppose
zs_free() loses a race with migration. We would schedule
async_free_zspage() to handle freeing that zspage, however we have no
guarantee that migration has finished
by the time we finish flush_work(&pool->work). In that case we then
call iput(inode), and now we have a page
pointing to a non-existent inode. (At which point something like
kcompactd would potentially BUG() if it tries to get a page
(from the inode) that doesn't exist anymore)
>
> If so, can't we add a flag to indicate destroy of the pool and
> global counter to indicate how many of zs_compact was nested?
>
> So, zs_unregister_migration in zs_destroy_pool can set the flag to
> prevent upcoming zs_compact call and wait until the global counter
> will be zero. Once it's done, finally flush the work.
>
> My point is it's not a per-class granuarity but global.
We could have a pool level counter of isolated pages, and wait for
that to finish before starting flush_work(&pool->work); However,
that would require an atomic_long in zs_pool, and we would have to eat
the cost of any contention over that lock. Still, it might be
preferable to a per-class granularity.
>
> Thanks.
>
> >
> > Since we know at destroy time all objects should be free, no new
> > migrations can come in (since zs_page_isolate() fails for fully-free
> > zspages). This means it is sufficient to track a "# isolated zspages"
> > count by class, and have the destroy logic ensure all such pages have
> > drained before proceeding. Keeping that state under the class
> > spinlock keeps the logic straightforward.
> >
> > Signed-off-by: Henry Burns <henrybu...@google.com>
> > ---
> > mm/zsmalloc.c | 68 ++++++++++++++++++++++++++++++++++++++++++++++++---
> > 1 file changed, 65 insertions(+), 3 deletions(-)
> >
> > diff --git a/mm/zsmalloc.c b/mm/zsmalloc.c
> > index efa660a87787..1f16ed4d6a13 100644
> > --- a/mm/zsmalloc.c
> > +++ b/mm/zsmalloc.c
> > @@ -53,6 +53,7 @@
> > #include <linux/zpool.h>
> > #include <linux/mount.h>
> > #include <linux/migrate.h>
> > +#include <linux/wait.h>
> > #include <linux/pagemap.h>
> > #include <linux/fs.h>
> >
> > @@ -206,6 +207,10 @@ struct size_class {
> > int objs_per_zspage;
> > /* Number of PAGE_SIZE sized pages to combine to form a 'zspage' */
> > int pages_per_zspage;
> > +#ifdef CONFIG_COMPACTION
> > + /* Number of zspages currently isolated by compaction */
> > + int isolated;
> > +#endif
> >
> > unsigned int index;
> > struct zs_size_stat stats;
> > @@ -267,6 +272,8 @@ struct zs_pool {
> > #ifdef CONFIG_COMPACTION
> > struct inode *inode;
> > struct work_struct free_work;
> > + /* A workqueue for when migration races with async_free_zspage() */
> > + struct wait_queue_head migration_wait;
> > #endif
> > };
> >
> > @@ -1917,6 +1924,21 @@ static void putback_zspage_deferred(struct zs_pool
> > *pool,
> >
> > }
> >
> > +static inline void zs_class_dec_isolated(struct zs_pool *pool,
> > + struct size_class *class)
> > +{
> > + assert_spin_locked(&class->lock);
> > + VM_BUG_ON(class->isolated <= 0);
> > + class->isolated--;
> > + /*
> > + * There's no possibility of racing, since wait_for_isolated_drain()
> > + * checks the isolated count under &class->lock after enqueuing
> > + * on migration_wait.
> > + */
> > + if (class->isolated == 0 && waitqueue_active(&pool->migration_wait))
> > + wake_up_all(&pool->migration_wait);
> > +}
> > +
> > static void replace_sub_page(struct size_class *class, struct zspage
> > *zspage,
> > struct page *newpage, struct page *oldpage)
> > {
> > @@ -1986,6 +2008,7 @@ static bool zs_page_isolate(struct page *page,
> > isolate_mode_t mode)
> > */
> > if (!list_empty(&zspage->list) && !is_zspage_isolated(zspage)) {
> > get_zspage_mapping(zspage, &class_idx, &fullness);
> > + class->isolated++;
> > remove_zspage(class, zspage, fullness);
> > }
> >
> > @@ -2085,8 +2108,14 @@ static int zs_page_migrate(struct address_space
> > *mapping, struct page *newpage,
> > * Page migration is done so let's putback isolated zspage to
> > * the list if @page is final isolated subpage in the zspage.
> > */
> > - if (!is_zspage_isolated(zspage))
> > + if (!is_zspage_isolated(zspage)) {
> > + /*
> > + * We still hold the class lock while all of this is
> > happening,
> > + * so we cannot race with zs_destroy_pool()
> > + */
> > putback_zspage_deferred(pool, class, zspage);
> > + zs_class_dec_isolated(pool, class);
> > + }
> >
> > reset_page(page);
> > put_page(page);
> > @@ -2131,9 +2160,11 @@ static void zs_page_putback(struct page *page)
> >
> > spin_lock(&class->lock);
> > dec_zspage_isolation(zspage);
> > - if (!is_zspage_isolated(zspage))
> > - putback_zspage_deferred(pool, class, zspage);
> >
> > + if (!is_zspage_isolated(zspage)) {
> > + putback_zspage_deferred(pool, class, zspage);
> > + zs_class_dec_isolated(pool, class);
> > + }
> > spin_unlock(&class->lock);
> > }
> >
> > @@ -2156,8 +2187,36 @@ static int zs_register_migration(struct zs_pool
> > *pool)
> > return 0;
> > }
> >
> > +static bool class_isolated_are_drained(struct size_class *class)
> > +{
> > + bool ret;
> > +
> > + spin_lock(&class->lock);
> > + ret = class->isolated == 0;
> > + spin_unlock(&class->lock);
> > + return ret;
> > +}
> > +
> > +/* Function for resolving migration */
> > +static void wait_for_isolated_drain(struct zs_pool *pool)
> > +{
> > + int i;
> > +
> > + /*
> > + * We're in the process of destroying the pool, so there are no
> > + * active allocations. zs_page_isolate() fails for completely free
> > + * zspages, so we need only wait for each size_class's isolated
> > + * count to hit zero.
> > + */
> > + for (i = 0; i < ZS_SIZE_CLASSES; i++) {
> > + wait_event(pool->migration_wait,
> > + class_isolated_are_drained(pool->size_class[i]));
> > + }
> > +}
> > +
> > static void zs_unregister_migration(struct zs_pool *pool)
> > {
> > + wait_for_isolated_drain(pool); /* This can block */
> > flush_work(&pool->free_work);
> > iput(pool->inode);
> > }
> > @@ -2401,6 +2460,8 @@ struct zs_pool *zs_create_pool(const char *name)
> > if (!pool->name)
> > goto err;
> >
> > + init_waitqueue_head(&pool->migration_wait);
> > +
> > if (create_cache(pool))
> > goto err;
> >
> > @@ -2466,6 +2527,7 @@ struct zs_pool *zs_create_pool(const char *name)
> > class->index = i;
> > class->pages_per_zspage = pages_per_zspage;
> > class->objs_per_zspage = objs_per_zspage;
> > + class->isolated = 0;
> > spin_lock_init(&class->lock);
> > pool->size_class[i] = class;
> > for (fullness = ZS_EMPTY; fullness < NR_ZS_FULLNESS;
> > --
> > 2.22.0.770.g0f2c4a37fd-goog
> >
