On Fri, Aug 30, 2024 at 11:25 PM Vlastimil Babka <vbabka@xxxxxxx> wrote: > > On 8/30/24 11:14, Yafang Shao wrote: > > On Thu, Aug 29, 2024 at 10:29 PM Dave Chinner <david@xxxxxxxxxxxxx> wrote: > > > > Hello Dave, > > > > I've noticed that XFS has increasingly replaced kmem_alloc() with > > __GFP_NOFAIL. For example, in kernel 4.19.y, there are 0 instances of > > __GFP_NOFAIL under fs/xfs, but in kernel 6.1.y, there are 41 > > occurrences. In kmem_alloc(), there's an explicit > > memalloc_retry_wait() to throttle the allocator under heavy memory > > pressure, which aligns with your filesystem design. However, using > > __GFP_NOFAIL removes this throttling mechanism, potentially causing > > issues when the system is under heavy memory load. I'm concerned that > > this shift might not be a beneficial trend. > > > > We have been using XFS for our big data servers for years, and it has > > consistently performed well with older kernels like 4.19.y. However, > > after upgrading all our servers from 4.19.y to 6.1.y over the past two > > years, we have frequently encountered livelock issues caused by memory > > exhaustion. To mitigate this, we've had to limit the RSS of > > applications, which isn't an ideal solution and represents a worrying > > trend. > > By "livelock issues caused by memory exhaustion" you mean the long-standing > infamous issue that the system might become thrashing for the remaining > small amount of page cache, and anonymous memory being swapped out/in, > instead of issuing OOM, because there's always just enough progress of the > reclaim to keep going, but the system isn't basically doing anything else? > Exactly > I think that's related to near-exhausted memory by userspace, If user space is the root cause, the appropriate response should be to terminate the offending user tasks. However, this doesn't happen at all. > so I'm not > sure why XFS would be to blame here. Honestly, I'm not sure what to blame, as I don't have a clear understanding of what's happening during memory allocation. One server among tens of thousands in production randomly experiences a livelock within days, making it extremely difficult to pinpoint the root cause. > > That said, if memalloc_retry_wait() is indeed a useful mechanism, maybe we > could perform it inside the page allocator itself for __GFP_NOFAIL? Perhaps an additional wait or exit mechanism should be implemented specifically for __GFP_NOFAIL. -- Regards Yafang
