Re: [PATCH v1 4/5] zram: add swap full hint

[Minchan Kim]

On Tue, Oct 07, 2014 at 08:36:08AM +0900, Minchan Kim wrote:
> Hello Dan,
> 
> Sorry for the delay. I had internal works which should be handled
> urgent. I hope you don't lose your interest due to my bad response
> latency.
> 
> On Thu, Sep 25, 2014 at 11:52:22AM -0400, Dan Streetman wrote:
> > On Wed, Sep 24, 2014 at 9:02 PM, Minchan Kim <minchan@xxxxxxxxxx> wrote:
> > > On Wed, Sep 24, 2014 at 10:01:03AM -0400, Dan Streetman wrote:
> > >> On Sun, Sep 21, 2014 at 8:03 PM, Minchan Kim <minchan@xxxxxxxxxx> wrote:
> > >> > This patch implement SWAP_FULL handler in zram so that VM can
> > >> > know whether zram is full or not and use it to stop anonymous
> > >> > page reclaim.
> > >> >
> > >> > How to judge fullness is below,
> > >> >
> > >> > fullness = (100 * used space / total space)
> > >> >
> > >> > It means the higher fullness is, the slower we reach zram full.
> > >> > Now, default of fullness is 80 so that it biased more momory
> > >> > consumption rather than early OOM kill.
> > >> >
> > >> > Above logic works only when used space of zram hit over the limit
> > >> > but zram also pretend to be full once 32 consecutive allocation
> > >> > fail happens. It's safe guard to prevent system hang caused by
> > >> > fragment uncertainty.
> > >> >
> > >> > Signed-off-by: Minchan Kim <minchan@xxxxxxxxxx>
> > >> > ---
> > >> >  drivers/block/zram/zram_drv.c | 60 ++++++++++++++++++++++++++++++++++++++++---
> > >> >  drivers/block/zram/zram_drv.h |  1 +
> > >> >  2 files changed, 57 insertions(+), 4 deletions(-)
> > >> >
> > >> > diff --git a/drivers/block/zram/zram_drv.c b/drivers/block/zram/zram_drv.c
> > >> > index 22a37764c409..649cad9d0b1c 100644
> > >> > --- a/drivers/block/zram/zram_drv.c
> > >> > +++ b/drivers/block/zram/zram_drv.c
> > >> > @@ -43,6 +43,20 @@ static const char *default_compressor = "lzo";
> > >> >  /* Module params (documentation at end) */
> > >> >  static unsigned int num_devices = 1;
> > >> >
> > >> > +/*
> > >> > + * If (100 * used_pages / total_pages) >= ZRAM_FULLNESS_PERCENT),
> > >> > + * we regards it as zram-full. It means that the higher
> > >> > + * ZRAM_FULLNESS_PERCENT is, the slower we reach zram full.
> > >> > + */
> > >> > +#define ZRAM_FULLNESS_PERCENT 80
> > >>
> > >> As Andrew said, this (or the user-configurable fullness param from the
> > >> next patch) should have more detail about exactly why it's needed and
> > >> what it does.  The details of how zram considers itself "full" should
> > >> be clear, which probably includes explaining zsmalloc fragmentation.
> > >> It should be also clear this param only matters when limit_pages is
> > >> set, and this param is only checked when zsmalloc's total size has
> > >> reached that limit.
> > >
> > > Sure, How about this?
> > >
> > >                 The fullness file is read/write and specifies how easily
> > >                 zram become full state. Normally, we can think "full"
> > >                 once all of memory is consumed but it's not simple with
> > >                 zram because zsmalloc has some issue by internal design
> > >                 so that write could fail once consumed *page* by zram
> > >                 reaches the mem_limit and zsmalloc cannot have a empty
> > >                 slot for the compressed object's size on fragmenet space
> > >                 although it has more empty slots for other sizes.
> > 
> > I understand that, but it might be confusing or unclear to anyone
> > who's not familiar with how zsmalloc works.
> > 
> > Maybe it could be explained by referencing the existing
> > compr_data_size and mem_used_total?  In addition to some or all of the
> > above, you could add something like:
> > 
> > This controls when zram decides that it is "full".  It is a percent
> > value, checked against compr_data_size / mem_used_total.  When
> > mem_used_total is equal to mem_limit, the fullness is checked and if
> > the compr_data_size / mem_used_total percentage is higher than this
> > specified fullness value, zram is considered "full".
> 
> Better than my verbose version.
> 
> > 
> > 
> > >
> > >                 We regard zram as full once consumed *page* reaches the
> > >                 mem_limit and consumed memory until now is higher the value
> > >                 resulted from the knob. So, if you set the value high,
> > >                 you can squeeze more pages into fragment space so you could
> > >                 avoid early OOM while you could see more write-fail warning,
> > >                 overhead to fail-write recovering by VM and reclaim latency.
> > >                 If you set the value low, you can see OOM kill easily
> > >                 even though there are memory space in zram but you could
> > >                 avoid shortcomings mentioned above.
> > 
> > You should clarify also that this is currently only used by
> > swap-on-zram, and this value prevents swap from writing to zram once
> > it is "full".  This setting has no effect when using zram for a
> > mounted filesystem.
> 
> Sure.
> 
> > 
> > >
> > >                 This knobs is valid ony if you set mem_limit.
> > >                 Currently, initial value is 80% but it could be changed.
> > >
> > > I didn't decide how to change it from percent.
> > > Decimal fraction Jerome mentioned does make sense to me so please ignore
> > > percent part in above.
> > >
> > >>
> > >> Also, since the next patch changes it to be used only as a default,
> > >> shouldn't it be DEFAULT_ZRAM_FULLNESS_PERCENT or similar?
> > >
> > > Okay, I will do it in 5/5.
> > >
> > >>
> > >> > +
> > >> > +/*
> > >> > + * If zram fails to allocate memory consecutively up to this,
> > >> > + * we regard it as zram-full. It's safe guard to prevent too
> > >> > + * many swap write fail due to lack of fragmentation uncertainty.
> > >> > + */
> > >> > +#define ALLOC_FAIL_MAX 32
> > >> > +
> > >> >  #define ZRAM_ATTR_RO(name)                                             \
> > >> >  static ssize_t zram_attr_##name##_show(struct device *d,               \
> > >> >                                 struct device_attribute *attr, char *b) \
> > >> > @@ -148,6 +162,7 @@ static ssize_t mem_limit_store(struct device *dev,
> > >> >
> > >> >         down_write(&zram->init_lock);
> > >> >         zram->limit_pages = PAGE_ALIGN(limit) >> PAGE_SHIFT;
> > >> > +       atomic_set(&zram->alloc_fail, 0);
> > >> >         up_write(&zram->init_lock);
> > >> >
> > >> >         return len;
> > >> > @@ -410,6 +425,7 @@ static void zram_free_page(struct zram *zram, size_t index)
> > >> >         atomic64_sub(zram_get_obj_size(meta, index),
> > >> >                         &zram->stats.compr_data_size);
> > >> >         atomic64_dec(&zram->stats.pages_stored);
> > >> > +       atomic_set(&zram->alloc_fail, 0);
> > >> >
> > >> >         meta->table[index].handle = 0;
> > >> >         zram_set_obj_size(meta, index, 0);
> > >> > @@ -597,10 +613,15 @@ static int zram_bvec_write(struct zram *zram, struct bio_vec *bvec, u32 index,
> > >> >         }
> > >> >
> > >> >         alloced_pages = zs_get_total_pages(meta->mem_pool);
> > >> > -       if (zram->limit_pages && alloced_pages > zram->limit_pages) {
> > >> > -               zs_free(meta->mem_pool, handle);
> > >> > -               ret = -ENOMEM;
> > >> > -               goto out;
> > >> > +       if (zram->limit_pages) {
> > >> > +               if (alloced_pages > zram->limit_pages) {
> > >> > +                       zs_free(meta->mem_pool, handle);
> > >> > +                       atomic_inc(&zram->alloc_fail);
> > >> > +                       ret = -ENOMEM;
> > >> > +                       goto out;
> > >> > +               } else {
> > >> > +                       atomic_set(&zram->alloc_fail, 0);
> > >> > +               }
> > >>
> > >> So, with zram_full() checking for alloced_pages >= limit_pages, this
> > >> will need to be changed; the way it is now it prevents that from ever
> > >> being true.
> > >>
> > >> Instead I believe this check has to be moved to before zs_malloc(), so
> > >> that alloced_pages > limit_pages is true.
> > >
> > > I don't get it why you said "it prevents that from ever being true".
> > > Now, zram can use up until limit_pages (ie, used memory == zram->limit_pages)
> > > and trying to get more is failed. so zram_full checks it as
> > > toal_pages >= zram->limit_pages so what is problem?
> > > If I miss your point, could you explain more?
> > 
> > ok, that's true, it's possible for alloc_pages == limit_pages, but
> > since zsmalloc will increase its size by a full zspage, and those can
> > be anywhere between 1 and 4 pages in size, it's only a (very roughly)
> > 25% chance that an alloc will cause alloc_pages == limit_pages, it's
> > more likely that an alloc will cause alloc_pages > limit_pages.  Now,
> > after some number of write failures, that 25% (-ish) probability will
> > be met, and alloc_pages == limit_pages will happen, but there's a
> > rather high chance that there will be some number of write failures
> > first.
> > 
> > To summarize or restate that, I guess what I'm saying is that for
> > users who don't care about some write failures and/or users with no
> > other swap devices except zram, it probably does not matter.  However
> > for them, they probably will rely on the 32 write failure limit, and
> > not the fullness limit.  For users where zram is only the primary swap
> > device, and there is a backup swap device, they probably will want
> > zram to fail over to the backup fairly quickly, with as few write
> > failures as possible (preferably, none, I would think).  And this
> > situation makes that highly unlikely - since there's only about a 25%
> > chance of alloc_pages == limit_pages with no previous write failures,
> > it's almost a certainty that there will be write failures before zram
> > is decided to be "full", even if "fullness" is set to 0.
> > 
> > With that said, you're right that it will eventually work, and those
> > few write failures while trying to get to alloc_pages == limit_pages
> > would probably not be noticable.  However, do remember that zram won't
> > stay full forever, so if it is only the primary swap device, it's
> > likely it will move between "full" and "not full" quite a lot, and
> > those few write failures may start adding up.
> 
> Fair enough.
> 
> But it is possible to see write-failure even though we correct
> it because there is potential chance for zram to fail to allocate
> order-0 page by a few reason which one of them is CMA I got several
> reports because zRAM cannot allocate a movable page due to lack of
> migration while usersapce goes with it well. I have a plan to fix it
> with zsmalloc migration work but there are another chances to make
> fail order-0 page by serval ways so I don't think we cannot prevent
> write-failure completely unless we have reserved memory for zram.
> 
> Having said that, I agree it would be better to reduce such fails
> with small code piece so I will check zram_full as follows,
> 
> /*
>  * XXX: zsmalloc_maxpages check should be removed when zsmalloc
>  * implement using of fragmented spaces in last page of zspage.
>  */
> if (total_pages >= zram->limit_pages - zsmalloc_maxpages()) {
>         ...
> }
> 

How about this?

diff --git a/drivers/block/zram/zram_drv.c b/drivers/block/zram/zram_drv.c
index 19da34aaf4f5..f03a94d7aa17 100644
--- a/drivers/block/zram/zram_drv.c
+++ b/drivers/block/zram/zram_drv.c
@@ -978,7 +978,7 @@ static void zram_full(struct block_device *bdev, bool *full)
 	meta = zram->meta;
 	total_pages = zs_get_total_pages(meta->mem_pool);
 
-	if (total_pages >= zram->limit_pages) {
+	if (total_pages > zram->limit_pages - zs_get_maxpages_per_zspage()) {
 
 		compr_pages = atomic64_read(&zram->stats.compr_data_size)
 					>> PAGE_SHIFT;
diff --git a/include/linux/zsmalloc.h b/include/linux/zsmalloc.h
index 05c214760977..73eb87bc5a4e 100644
--- a/include/linux/zsmalloc.h
+++ b/include/linux/zsmalloc.h
@@ -48,4 +48,5 @@ void zs_unmap_object(struct zs_pool *pool, unsigned long handle);
 
 unsigned long zs_get_total_pages(struct zs_pool *pool);
 
+int zs_get_maxpages_per_zspage(void);
 #endif
diff --git a/mm/zsmalloc.c b/mm/zsmalloc.c
index 839a48c3ca27..6b6653455573 100644
--- a/mm/zsmalloc.c
+++ b/mm/zsmalloc.c
@@ -316,6 +316,12 @@ static struct zpool_driver zs_zpool_driver = {
 MODULE_ALIAS("zpool-zsmalloc");
 #endif /* CONFIG_ZPOOL */
 
+int zs_get_maxpages_per_zspage(void)
+{
+	return ZS_MAX_PAGES_PER_ZSPAGE;
+}
+EXPORT_SYMBOL_GPL(zs_get_maxpages_per_zspage);
+
 /* per-cpu VM mapping areas for zspage accesses that cross page boundaries */
 static DEFINE_PER_CPU(struct mapping_area, zs_map_area);
 
-- 
Kind regards,
Minchan Kim
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/