On Thu, Nov 25, 2021 at 12:39 AM David Hildenbrand <david@xxxxxxxxxx> wrote: > > On 25.11.21 03:45, Shakeel Butt wrote: > > Many applications do sophisticated management of their heap memory for > > better performance but with low cost. We have a bunch of such > > applications running on our production and examples include caching and > > data storage services. These applications keep their hot data on the > > THPs for better performance and release the cold data through > > MADV_DONTNEED to keep the memory cost low. > > > > The kernel defers the split and release of THPs until there is memory > > pressure. This complicates the memory management of these sophisticated > > applications which then needs to look into low level kernel handling of > > THPs to better gauge their headroom for expansion. > > > > More specifically these applications monitor their cgroup usage to decide > > if they can expand the memory footprint or release some (unneeded/cold) > > buffer. They uses madvise(MADV_DONTNEED) to release the memory which > > basically puts the THP into defer list. These deferred THPs are still > > charged to the cgroup which leads to bloated usage read by the application > > and making wrong decisions. In addition these applications are very > > latency sensitive and would prefer to not face memory reclaim due to > > non-deterministic nature of reclaim. > > > > Internally we added a cgroup interface to trigger the split of deferred > > THPs for that cgroup but this is hacky and exposing kernel internals to > > users. This patch solves this problem in a more general way for the users > > by splitting the THPS synchronously on MADV_DONTNEED. This patch does > > the same for munmap() too. > > > > I'll have to defer diving into the code. > > Just a comment: It might be good to add that there are still cases where > splitting the compound page can fail -- for example, if the page is > still pinned/referenced. > > So if you have a THP and intended to only pin/reference e.g., the first > 4k of it (e.g., O_DIRECT, io_uring fixed buffers), MADV_DONTNEED/unmap > e.g., the last 4k of it will not split synchronously. > > In addition to explicit user action on a compound page; I remember there > might be other kernel-internal temporary references that could > theoretically block splitting, but maybe most of them are at least for > now limited to !compound pages. > Hi David, Thanks for your time (and apologies) but I have to rescind this patch for now due to mistaken performance impact. Let's move the discussion to the other thread and decide next steps there. thanks, Shakeel
