On Tue, 15 Apr 2025 18:58:51 -0700, akpm@xxxxxxxxxxxxxxxxxxxx wrote:
> On Tue, 15 Apr 2025 17:02:32 +0800 alexjlzheng@xxxxxxxxx wrote:
>
> > From: Jinliang Zheng <alexjlzheng@xxxxxxxxxxx>
> >
> > In the dirty_ratio_handler() function, vm_dirty_bytes must be set to
> > zero before calling writeback_set_ratelimit(), as global_dirty_limits()
> > always prioritizes the value of vm_dirty_bytes.
>
> Can you please tell us precisely where global_dirty_limits()
> prioritizes vm_dirty_bytes? I spent a while chasing code and didn't
> see how global_dirty_limits() gets to node_dirty_ok()(?).
Thank you for your reply.
It's domain_dirty_limits() that's relevant here, not node_dirty_ok:
dirty_ratio_handler
writeback_set_ratelimit
global_dirty_limits(&dirty_thresh) <- ratelimit_pages based on dirty_thresh
domain_dirty_limits
if (bytes) <- bytes = vm_dirty_bytes <--------+
thresh = f1(bytes) <- prioritizes vm_dirty_bytes |
else |
thresh = f2(ratio) |
ratelimit_pages = f3(dirty_thresh) |
vm_dirty_bytes = 0 <- it's late! ---------------------+
>
> > That causes ratelimit_pages to still use the value calculated based on
> > vm_dirty_bytes, which is wrong now.
> >
> > Fixes: 9d823e8f6b1b ("writeback: per task dirty rate limit")
> > Signed-off-by: Jinliang Zheng <alexjlzheng@xxxxxxxxxxx>
> > Reviewed-by: MengEn Sun <mengensun@xxxxxxxxxxx>
> > Cc: stable@xxxxxxxxxxxxxxx
>
> Please, as always, provide a description of the userspace-visible
> effects of this bug?
The impact visible to userspace is difficult to capture directly because there is no
procfs/sysfs interface exported to user space. However, it will have a real impact
on the balance of dirty pages.
For example:
1. On default, we have vm_dirty_ratio=40, vm_dirty_bytes=0
2. echo 8192 > dirty_bytes, then vm_dirty_bytes=8192, vm_dirty_ratio=0, and ratelimit_pages
is calculated based on vm_dirty_bytes now.
3. echo 20 > dirty_ratio, then since vm_dirty_bytes is not reset to zero when
writeback_set_ratelimit() -> global_dirty_limits() -> domain_dirty_limits() is called,
reallimit_pages is still calculated based on vm_dirty_bytes instead of vm_dirty_ratio.
This does not conform to the actual intention of the user.
thanks,
Jinliang Zheng :)
