On 2023/4/27 11:47, Hui Wang wrote:
...
Hi Xiang and Michal,
Is it tested with a pure ext4 without any other fs background?
Basically yes. Maybe there is a squashfs mounted for python3 in my test environment. But stress-ng and its needed sharing libs are in the ext4.
I don't think it's true that "ext4->readahead() doesn't call
alloc_page()" since I think even ext2/ext4 uses buffer head
interfaces to read metadata (extents or old block mapping)
from its bd_inode for readahead, which indirectly allocates
some extra pages to page cache as well.
Calling alloc_page() or allocating memory in the readahead() is not a problem, suppose we have 4 processes (A, B, C and D). Process A, B and C are entering out_of_memory() because of allocating memory in the readahead(), they are looping and waiting for some memory be released. And process D could enter out_of_memory() with __GFP_FS, then it could trigger oom killer, so A, B and C could get the memory and return to the readahead(), there is no system hang issue.
But if all 4 processes enter out_of_memory() from readahead(), they will loop and wait endlessly, there is no process to trigger oom killer, so the users will think the system is getting hang.
I applied my change for ext4->readahead to linux-next, and tested it on my ubuntu classic server for arm64, I could reproduce the hang issue within 1 minutes with 100% rate. I guess it is easy to reproduce the issue because it is an embedded environment, the total number of processes in the system is very limited, nearly all userspace processes will finally reach out_of_memory() from ext4_readahead(), and nearly all kthreads will not reach out_of_memory() for long time, that makes the system in a state like hang (not real hang).
And this is why I wrote a patch to let a specific kthread trigger oom killer forcibly (my initial patch).
The difference only here is the total number of pages to be
allocated here, but many extra compressed data takeing extra
allocation causes worse. So I think it much depends on how
stressful does your stress workload work like, and I'm even
not sure it's a real issue since if you stop the stress
workload, it will immediately recover (only it may not oom
directly).
Yes, it is not a real hang. All userspace processes are looping and waiting for other processes to release or reclaim memory. And in this case, we can't stop the stress workload since users can't control the system through console.
So Michal,
Don't know if you read the "[PATCH 0/1] mm/oom_kill: system enters a state something like hang when running stress-ng", do you know why out_of_memory() will return immediately if there is no __GFP_FS, could we drop these lines directly:
/*
* The OOM killer does not compensate for IO-less reclaim.
* pagefault_out_of_memory lost its gfp context so we have to
* make sure exclude 0 mask - all other users should have at least
* ___GFP_DIRECT_RECLAIM to get here. But mem_cgroup_oom() has to
* invoke the OOM killer even if it is a GFP_NOFS allocation.
*/
if (oc->gfp_mask && !(oc->gfp_mask & __GFP_FS) && !is_memcg_oom(oc))
return true;
I'm curious about this as well.
Also apart from that we also have some stacked virtual block devices
as well, mostly they will use GFP_NOIO allocations. So if they
allocate extra pages with GFP_NOIO like this, oom kill won't happen
as well. And mostly such block drivers cannot fail out I/Os (IO
error) since it will have worse results to upper fses and end users.
Generally I don't think reserving more pages makes sense since it
still depends on stressed workload (and it makes more pages privately
and unuseable to other subsystems and user programs), so we still have
two results:
- wait (hang) for previous requests to get free pages;
- return ENOMEM instead.
Thanks,
Gao Xiang
Thanks,
Hui.
Thanks,
Gao Xiang
