[PATCH] zram: Avoid duplicate slots in batched writeback

[PATCH] zram: Avoid duplicate slots in batched writeback

[raoxu]

From: Xu Rao <raoxu@uniontech.com>

The writeback interface accepts multiple page_index and page_indexes
parameters in one operation. Their ranges can overlap, but
scan_slots_for_writeback() only excludes slots that are already being
written back or marked as same-filled.

A slot selected by an earlier range is placed on the post-processing list
and marked with ZRAM_PP_SLOT. Without checking that flag, an overlapping
range can enqueue the same slot again and submit it for writeback more
than once.

Skip slots already marked ZRAM_PP_SLOT while holding the slot lock. This
keeps each slot in a writeback scan session represented by at most one
post-processing entry and leaves the existing release and reuse paths
unchanged.

Signed-off-by: Xu Rao <raoxu@uniontech.com>
---
 drivers/block/zram/zram_drv.c | 3 +++
 1 file changed, 3 insertions(+)

diff --git a/drivers/block/zram/zram_drv.c b/drivers/block/zram/zram_drv.c
index ab064c84b092..50a25989a23b 100644
--- a/drivers/block/zram/zram_drv.c
+++ b/drivers/block/zram/zram_drv.c
@@ -1215,6 +1215,9 @@ static void scan_slots_for_writeback(struct zram *zram, u32 mode,
 		if (!slot_allocated(zram, index))
 			goto next;

+		if (test_slot_flag(zram, index, ZRAM_PP_SLOT))
+			goto next;
+
 		if (test_slot_flag(zram, index, ZRAM_WB) ||
 		    test_slot_flag(zram, index, ZRAM_SAME))
 			goto next;
--
2.50.1

Re: [PATCH] zram: Avoid duplicate slots in batched writeback

[Sergey Senozhatsky]

On (26/11/10 14:18), raoxu wrote:
> The writeback interface accepts multiple page_index and page_indexes
> parameters in one operation. Their ranges can overlap, but
> scan_slots_for_writeback() only excludes slots that are already being
> written back or marked as same-filled.
> 
> A slot selected by an earlier range is placed on the post-processing list
> and marked with ZRAM_PP_SLOT. Without checking that flag, an overlapping
> range can enqueue the same slot again and submit it for writeback more
> than once.
> 
> Skip slots already marked ZRAM_PP_SLOT while holding the slot lock. This
> keeps each slot in a writeback scan session represented by at most one
> post-processing entry and leaves the existing release and reuse paths
> unchanged.
> 
> Signed-off-by: Xu Rao <raoxu@uniontech.com>

LGTM.
Reviewed-by: Sergey Senozhatsky <senozhatsky@chromium.org>

Jens, do you want to pick it up and route via block tree or should
we route via "more traditional" tree (Andrew's tree)?

Re: [PATCH] zram: Avoid duplicate slots in batched writeback

[Andrew Morton]

On Tue, 10 Nov 2026 15:40:13 +0900 Sergey Senozhatsky <senozhatsky@chromium.org> wrote:

> On (26/11/10 14:18), raoxu wrote:
> > The writeback interface accepts multiple page_index and page_indexes
> > parameters in one operation. Their ranges can overlap, but
> > scan_slots_for_writeback() only excludes slots that are already being
> > written back or marked as same-filled.
> > 
> > A slot selected by an earlier range is placed on the post-processing list
> > and marked with ZRAM_PP_SLOT. Without checking that flag, an overlapping
> > range can enqueue the same slot again and submit it for writeback more
> > than once.
> > 
> > Skip slots already marked ZRAM_PP_SLOT while holding the slot lock. This
> > keeps each slot in a writeback scan session represented by at most one
> > post-processing entry and leaves the existing release and reuse paths
> > unchanged.
> > 
> > Signed-off-by: Xu Rao <raoxu@uniontech.com>
> 
> LGTM.
> Reviewed-by: Sergey Senozhatsky <senozhatsky@chromium.org>
> 
> Jens, do you want to pick it up and route via block tree or should
> we route via "more traditional" tree (Andrew's tree)?

I grabbed it.

I did recently attempt to pull zram out of the mm project, by moving it
from mm.git's mm-unstable branch and into mm.git's mm-nonmm-unstable
branch.  Because "why does the mm project contain a block device
driver".

I undid that change after about a week - it didn't work out because
zram is joined at the hip with mm/zsmalloc.c.  zsmalloc is also used by
mm/zswap.c.

Oh well.  afaik the zram story has been going smoothly for a long
time(?).

Re: [PATCH] zram: Avoid duplicate slots in batched writeback

[David Hildenbrand (Arm)]

On 11/13/26 08:26, Andrew Morton wrote:
> On Tue, 10 Nov 2026 15:40:13 +0900 Sergey Senozhatsky <senozhatsky@chromium.org> wrote:
> 
>> On (26/11/10 14:18), raoxu wrote:
>>> The writeback interface accepts multiple page_index and page_indexes
>>> parameters in one operation. Their ranges can overlap, but
>>> scan_slots_for_writeback() only excludes slots that are already being
>>> written back or marked as same-filled.
>>>
>>> A slot selected by an earlier range is placed on the post-processing list
>>> and marked with ZRAM_PP_SLOT. Without checking that flag, an overlapping
>>> range can enqueue the same slot again and submit it for writeback more
>>> than once.
>>>
>>> Skip slots already marked ZRAM_PP_SLOT while holding the slot lock. This
>>> keeps each slot in a writeback scan session represented by at most one
>>> post-processing entry and leaves the existing release and reuse paths
>>> unchanged.
>>>
>>> Signed-off-by: Xu Rao <raoxu@uniontech.com>
>>
>> LGTM.
>> Reviewed-by: Sergey Senozhatsky <senozhatsky@chromium.org>
>>
>> Jens, do you want to pick it up and route via block tree or should
>> we route via "more traditional" tree (Andrew's tree)?
> 
> I grabbed it.
> 
> I did recently attempt to pull zram out of the mm project, by moving it
> from mm.git's mm-unstable branch and into mm.git's mm-nonmm-unstable
> branch.  Because "why does the mm project contain a block device
> driver".
> 
> I undid that change after about a week - it didn't work out because
> zram is joined at the hip with mm/zsmalloc.c.  zsmalloc is also used by
> mm/zswap.c.

Are there cleaner ways of isolating both things?

Like, many drivers use kmalloc() etc, and they are not part of core-mm :)

-- 
Cheers,

David