On Fri, Apr 07, 2023 at 06:37:43AM +0200, Christoph Hellwig wrote:
> On Thu, Apr 06, 2023 at 05:23:51PM -0700, Minchan Kim wrote:
> > rw_page path so that bio comes next to serve the rw_page failure.
> > In the case, zram will always do chained bio so we are fine with
> > asynchronous IO.
> >
> > diff --git a/drivers/block/zram/zram_drv.c b/drivers/block/zram/zram_drv.c
> > index b8549c61ff2c..23fa0e03cdc1 100644
> > --- a/drivers/block/zram/zram_drv.c
> > +++ b/drivers/block/zram/zram_drv.c
> > @@ -1264,6 +1264,8 @@ static int __zram_bvec_read(struct zram *zram, struct page *page, u32 index,
> > struct bio_vec bvec;
> >
> > zram_slot_unlock(zram, index);
> > + if (partial_io)
> > + return -EAGAIN;
> >
> > bvec.bv_page = page;
> > bvec.bv_len = PAGE_SIZE;
>
> What tree is this supposed to apply to? 6.2 already has the
> zram_bvec_read_from_bdev helper.
It was for prior to 6.1 where doesn't have (94541bc3fbde4, zram: always
expose rw_page). Since 6.1, with returning -EOPNOTSUPP on the rw_page,
it will defer the request to be handled via submit_bio so every IO to
read data from backing device will have parent bio and chained.
>
>
> But either way partial_io can be true when called from zram_bvec_write,
> where we can't just -EAGAIN as ->submit_bio is not allowed to return
> that except for REQ_NOWAIT bios, and even then it needs to be handle
> them when submitted without REQ_NOWAIT.
The case returning -EAGAIN is only rw_page path not the submit_bio path
so the request will be deferred to submit_bio path. However, yeah,
I see it's still broken since read_from_bdev_sync never wait the read
complettion again.
Then, as you suggested, just disable the feature for non-4K would be
the simple option.
I see only this way to achieve it and not sure it will cover all the
cases. If you have better idea, let me know.
Thank you!
diff --git a/drivers/block/zram/Kconfig b/drivers/block/zram/Kconfig
index 0386b7da02aa..ae7662430789 100644
--- a/drivers/block/zram/Kconfig
+++ b/drivers/block/zram/Kconfig
@@ -58,6 +58,12 @@ config ZRAM_DEF_COMP
config ZRAM_WRITEBACK
bool "Write back incompressible or idle page to backing device"
depends on ZRAM
+ depends on !PAGE_SIZE_8KB
+ depends on !PAGE_SIZE_16KB
+ depends on !PAGE_SIZE_32KB
+ depends on !PAGE_SIZE_64KB
+ depends on !PAGE_SIZE_256KB
+ depends on !PAGE_SIZE_1MB
help
With incompressible page, there is no memory saving to keep it
in memory. Instead, write it out to backing device.
