On 10/30/24 18:02, Joanne Koong wrote: > On Wed, Oct 30, 2024 at 9:21 AM Bernd Schubert > <bernd.schubert@xxxxxxxxxxx> wrote: >> >> On 10/30/24 17:04, Joanne Koong wrote: >>> On Wed, Oct 30, 2024 at 2:32 AM Bernd Schubert >>> <bernd.schubert@xxxxxxxxxxx> wrote: >>>> >>>> On 10/28/24 22:58, Joanne Koong wrote: >>>>> On Fri, Oct 25, 2024 at 3:40 PM Joanne Koong <joannelkoong@xxxxxxxxx> wrote: >>>>>> >>>>>>> Same here, I need to look some more into the compaction / page >>>>>>> migration paths. I'm planning to do this early next week and will >>>>>>> report back with what I find. >>>>>>> >>>>>> >>>>>> These are my notes so far: >>>>>> >>>>>> * We hit the folio_wait_writeback() path when callers call >>>>>> migrate_pages() with mode MIGRATE_SYNC >>>>>> ... -> migrate_pages() -> migrate_pages_sync() -> >>>>>> migrate_pages_batch() -> migrate_folio_unmap() -> >>>>>> folio_wait_writeback() >>>>>> >>>>>> * These are the places where we call migrate_pages(): >>>>>> 1) demote_folio_list() >>>>>> Can ignore this. It calls migrate_pages() in MIGRATE_ASYNC mode >>>>>> >>>>>> 2) __damon_pa_migrate_folio_list() >>>>>> Can ignore this. It calls migrate_pages() in MIGRATE_ASYNC mode >>>>>> >>>>>> 3) migrate_misplaced_folio() >>>>>> Can ignore this. It calls migrate_pages() in MIGRATE_ASYNC mode >>>>>> >>>>>> 4) do_move_pages_to_node() >>>>>> Can ignore this. This calls migrate_pages() in MIGRATE_SYNC mode but >>>>>> this path is only invoked by the move_pages() syscall. It's fine to >>>>>> wait on writeback for the move_pages() syscall since the user would >>>>>> have to deliberately invoke this on the fuse server for this to apply >>>>>> to the server's fuse folios >>>>>> >>>>>> 5) migrate_to_node() >>>>>> Can ignore this for the same reason as in 4. This path is only invoked >>>>>> by the migrate_pages() syscall. >>>>>> >>>>>> 6) do_mbind() >>>>>> Can ignore this for the same reason as 4 and 5. This path is only >>>>>> invoked by the mbind() syscall. >>>>>> >>>>>> 7) soft_offline_in_use_page() >>>>>> Can skip soft offlining fuse folios (eg folios with the >>>>>> AS_NO_WRITEBACK_WAIT mapping flag set). >>>>>> The path for this is soft_offline_page() -> soft_offline_in_use_page() >>>>>> -> migrate_pages(). soft_offline_page() only invokes this for in-use >>>>>> pages in a well-defined state (see ret value of get_hwpoison_page()). >>>>>> My understanding of soft offlining pages is that it's a mitigation >>>>>> strategy for handling pages that are experiencing errors but are not >>>>>> yet completely unusable, and its main purpose is to prevent future >>>>>> issues. It seems fine to skip this for fuse folios. >>>>>> >>>>>> 8) do_migrate_range() >>>>>> 9) compact_zone() >>>>>> 10) migrate_longterm_unpinnable_folios() >>>>>> 11) __alloc_contig_migrate_range() >>>>>> >>>>>> 8 to 11 needs more investigation / thinking about. I don't see a good >>>>>> way around these tbh. I think we have to operate under the assumption >>>>>> that the fuse server running is malicious or benevolently but >>>>>> incorrectly written and could possibly never complete writeback. So we >>>>>> definitely can't wait on these but it also doesn't seem like we can >>>>>> skip waiting on these, especially for the case where the server uses >>>>>> spliced pages, nor does it seem like we can just fail these with >>>>>> -EBUSY or something. >>>> >>>> I see some code paths with -EAGAIN in migration. Could you explain why >>>> we can't just fail migration for fuse write-back pages? >>>> >> >> Hi Joanne, >> >> thanks a lot for your quick reply (especially as my reviews come in very >> late). >> > > Thanks for your comments/reviews, Bernd! I always appreciate them. > >>> >>> My understanding (and please correct me here Shakeel if I'm wrong) is >>> that this could block system optimizations, especially since if an >>> unprivileged malicious fuse server never replies to the writeback >>> request, then this completely stalls progress. In the best case >>> scenario, -EAGAIN could be used because the server might just be slow >>> in serving the writeback, but I think we need to also account for >>> servers that never complete the writeback. For >>> __alloc_contig_migrate_range() for example, my understanding is that >>> this is used to migrate pages so that there are more physically >>> contiguous ranges of memory freed up. If fuse writeback blocks that, >>> then that hurts system health overall. >> >> Hmm, I wonder what is worse - tmp page copies or missing compaction. >> Especially if we expect a low range of in-writeback pages/folios. >> One could argue that an evil user might spawn many fuse server >> processes to work around the default low fuse write-back limits, but >> does that make any difference with tmp pages? And these cannot be >> compacted either? > > My understanding (and Shakeel please jump in here if this isn't right) > is that tmp pages can be migrated/compacted. I think it's only pages > marked as under writeback that are considered to be non-movable. > >> >> And with timeouts that would be so far totally uncritical, I >> think. >> >> >> You also mentioned >> >>> especially for the case where the server uses spliced pages >> >> could you provide more details for that? >> 7> > For the page migration / compaction paths, I don't think we can do the > workaround we could do for sync where we skip waiting on writeback for > fuse folios and continue on with the operation, because the migration > / compaction paths operate on the pages. For the splice case, we > assign the page to the pipebuffer (fuse_ref_page()), so if the > migration/compaction happens on the page before the server has read > this page from the pipebuffer, it'll be incorrect data or maybe crash > the kernel. > >> >> >>> >>>>>> >>>>> >>>>> I'm still not seeing a good way around this. >>>>> >>>>> What about this then? We add a new fuse sysctl called something like >>>>> "/proc/sys/fs/fuse/writeback_optimization_timeout" where if the sys >>>>> admin sets this, then it opts into optimizing writeback to be as fast >>>>> as possible (eg skipping the page copies) and if the server doesn't >>>>> fulfill the writeback by the set timeout value, then the connection is >>>>> aborted. >>>>> >>>>> Alternatively, we could also repurpose >>>>> /proc/sys/fs/fuse/max_request_timeout from the request timeout >>>>> patchset [1] but I like the additional flexibility and explicitness >>>>> having the "writeback_optimization_timeout" sysctl gives. >>>>> >>>>> Any thoughts on this? >>>> >>>> >>>> I'm a bit worried that we might lock up the system until time out is >>>> reached - not ideal. Especially as timeouts are in minutes now. But >>>> even a slightly stuttering video system not be great. I think we >>>> should give users/admin the choice then, if they prefer slow page >>>> copies or fast, but possibly shortly unresponsive system. >>>> >>> I was thinking the /proc/sys/fs/fuse/writeback_optimization_timeout >>> would be in seconds, where the sys admin would probably set something >>> more reasonable like 5 seconds or so. >>> If this syctl value is set, then servers who want writebacks to be >>> fast can opt into it at mount time (and by doing so agree that they >>> will service writeback requests by the timeout or their connection >>> will be aborted). >> >> >> I think your current patch set has it in minutes? (Should be easy >> enough to change that.) Though I'm more worried about the impact >> of _frequent_ timeout scanning through the different fuse lists >> on performance, than about missing compaction for folios that are >> currently in write-back. Hmm, if tmp pages can be compacted, isn't that a problem for splice? I.e. I don't understand what the difference between tmp page and write-back page for migration. >> > > Ah, for this the " /proc/sys/fs/fuse/writeback_optimization_timeout" > would be a separate thing from the > "/proc/sys/fs/fuse/max_request_timeout". The > "/proc/sys/fs/fuse/writeback_optimization_timeout" would only apply > for writeback requests. I was thinking implementation-wise, for > writebacks we could just have a timer associated with each request > (instead of having to grab locks with the fuse lists), since they > won't be super common. Ah, thank you! I had missed that this is another variable. Issue with too short timeouts would probably be network hick-up that would immediately kill fuse-server. I.e. if it just the missing page compaction/migration, maybe larger time outs would be acceptable. Thanks, Bernd
