On 16.06.23 09:37, Yosry Ahmed wrote:
On Thu, Jun 15, 2023 at 9:41 PM 贺中坤 <hezhongkun.hzk@xxxxxxxxxxxxx> wrote:
Thanks Fabian for tagging me.
I am not familiar with #1, so I will speak to #2. Zhongkun, There are
a few parts that I do not understand -- hopefully you can help me out
here:
(1) If I understand correctly in this patch we set the active memcg
trying to charge any pages allocated in a zspage to the current memcg,
yet that zspage will contain multiple compressed object slots, not
just the one used by this memcg. Aren't we overcharging the memcg?
Basically the first memcg that happens to allocate the zspage will pay
for all the objects in this zspage, even after it stops using the
zspage completely?
It will not overcharge. As you said below, we are not using
__GFP_ACCOUNT and charging the compressed slots to the memcgs.
(2) Patch 3 seems to be charging the compressed slots to the memcgs,
yet this patch is trying to charge the entire zspage. Aren't we double
charging the zspage? I am guessing this isn't happening because (as
Michal pointed out) we are not using __GFP_ACCOUNT here anyway, so
this patch may be NOP, and the actual charging is coming from patch 3
only.
YES， the actual charging is coming from patch 3. This patch just
delivers the BIO page's memcg to the current task which is not the
consumer.
(3) Zswap recently implemented per-memcg charging of compressed
objects in a much simpler way. If your main interest is #2 (which is
what I understand from the commit log), it seems like zswap might be
providing this already? Why can't you use zswap? Is it the fact that
zswap requires a backing swapfile?
Thanks for your reply and review. Yes, the zswap requires a backing
swapfile. The I/O path is very complex, sometimes it will throttle the
whole system if some resources are short , so we hope to use zram.
Is the only problem with zswap for you the requirement of a backing swapfile?
If yes, I am in the early stages of developing a solution to make
zswap work without a backing swapfile. This was discussed in LSF/MM
[1]. Would this make zswap usable in for your use case?
Out of curiosity, are there any other known pros/cons when using
zswap-without-swap instead of zram?
I know that zram requires sizing (size of the virtual block device) and
consumes metadata, zswap doesn't.
--
Cheers,
David / dhildenb
