| Message ID | 1685531374-6091-1-git-send-email-quic_charante@quicinc.com (mailing list archive) |
|---|---|
| State | New |
| Headers | show |
| Series | mm: madvise: fix uneven accounting of psi | expand |
Hi Charan, On Tue, Jun 6, 2023 at 7:54 AM Charan Teja Kalla <quic_charante@quicinc.com> wrote: > > Thanks Johannes for the detailed review comments... > > On 6/5/2023 11:30 PM, Johannes Weiner wrote: > >> Agree that we shouldn't be really silence the thrashing. My point is we > >> shouldn't be considering the folios as thrashing If those were getting > >> reclaim by the user him self through MADV_PAGEOUT under the assumption > >> that __user knows they are not real working set__. Please let me know > >> if I am not making sense here. > > I'm not sure I agree with this. I think it misses the point of what > > the madvise is actually for. > > > > The workingset is defined based on access frequency and available > > memory. Thrashing is defined as having to read pages back shortly > > after their eviction. > > > > MADV_PAGEOUT is for the application to inform the kernel that it's > > done accessing the pages, so that the kernel can accelerate their > > eviction over other pages that may still be in use. This is ultimately > > meant to REDUCE reclaim and paging. > > > > However, in this case, the MADVISE_PAGEOUT evicts pages that are > > reused after and then refault. It INCREASED reclaim and paging. > > > I agree here... > > Surely that's a problem? And the system would have behaved better > > without the madvise() in the first place? > > > Yes, the system behavior could be much better without this PAGEOUT > operation... > > In fact, I would argue that the pressure spike is a great signal for > > detecting overzealous madvising. If you're redefining the workingset > > from access frequency to "whatever the user is saying", that will take > > away an important mechanism to detect advise bugs and unnecessary IO. > currently wanted to do the PAGEOUT operation but what information lacks > is if I am really operating on the workingset pages. Had the client > knows that he is operating on the workingset pages, he could have backed > off from madvising. > > I now note that I shouldn't be defining the workingset from "whatever > user is saying". But then, IMO, there should be a way from the kernel to > the user that his madvise operation is being performed on the workingset > pages. > > One way the user can do is monitoring the PSI events while PAGEOUT is > being performed and he may exclude those VMA's from next time. > > Alternatively kernel itself can support it may be through like > MADV_PAGEOUT_INACTIVE which doesn't pageout the Workingset pages. > > Please let me know your opinion about this interface. > > This has the usecase on android where it just assumes that 2nd > background app will most likely to be not used in the future thus > reclaim those app pages. It works well for most of the times but such > assumption will go wrong with the usecase I had mentioned. Hi Folks. Sorry for being late to the party. Yeah, userspace does not have a crystal ball to predict future user behavior, so there will always be pathological cases when usual assumptions and resulting madvise() would make things worse. I think this discussion can be split into several questions/issues: 1. Inconsistency in how madvise(MADV_PAGEOUT) would affect PSI calculation when the page is refaulted, based on the path it took before being evicted by madvise(). In your initial description case (a) is inconsistent with (b) and (c) and it's probably worth fixing. IMHO (a) should be made consistent with others, not the other way around. My reasoning is that page was expelled from the active list, so it was part of the active workingset. 2. Whether refaults caused by incorrect madvise(MADV_PAGEOUT) should be counted as workingset refault and affect PSI. This one I think is trickier. IMHO it should be counted as workingset refault simply because it was refaulted and it was part of the workingset. Whether it should affect PSI, which is supposed to be an indicator of "pressure" is, I think, debatable. With madvise() in the mix, refault might happen without any real memory pressure... So, the answer is not obvious to me. 3. Should refaults caused by incorrect madvise(MADV_PAGEOUT) be distinguished from the ones which were evicted by kernel reclaim mechanisms. I can see use for that from userspace to detect incorrect madvise() and adjust its aggressiveness. I think the API might get a bit complex because of the need to associate refaults with specific madvise()/VMAs to understand which hint was incorrect and adjust the behavior. Hope my feedback is useful and if we can improve Android's userspace behavior, I'm happy to help make that happen. Thanks, Suren. > > --Thanks.
Thanks Suren & Johannes, On 6/7/2023 1:18 AM, Suren Baghdasaryan wrote: > Hi Folks. Sorry for being late to the party. > Yeah, userspace does not have a crystal ball to predict future user > behavior, so there will always be pathological cases when usual > assumptions and resulting madvise() would make things worse. > > I think this discussion can be split into several questions/issues: > 1. Inconsistency in how madvise(MADV_PAGEOUT) would affect PSI > calculation when the page is refaulted, based on the path it took > before being evicted by madvise(). In your initial description case > (a) is inconsistent with (b) and (c) and it's probably worth fixing. > IMHO (a) should be made consistent with others, not the other way > around. My reasoning is that page was expelled from the active list, > so it was part of the active workingset. > That means we should be setting Workingset on the page while it is on the active list and when it is being pageout through madvising. Right? I see, this makes it consistent. On the same note, discussing with Suren offline, Should the refaulted madvise pages start always at the inactive list? If they are really active, they get promoted anyway.. > 2. Whether refaults caused by incorrect madvise(MADV_PAGEOUT) should > be counted as workingset refault and affect PSI. > This one I think is trickier. IMHO it should be counted as workingset > refault simply because it was refaulted and it was part of the > workingset. Whether it should affect PSI, which is supposed to be an > indicator of "pressure" is, I think, debatable. With madvise() in the > mix, refault might happen without any real memory pressure... So, the > answer is not obvious to me. > > 3. Should refaults caused by incorrect madvise(MADV_PAGEOUT) be > distinguished from the ones which were evicted by kernel reclaim > mechanisms. > I can see use for that from userspace to detect incorrect madvise() > and adjust its aggressiveness. I think the API might get a bit complex > because of the need to associate refaults with specific madvise()/VMAs > to understand which hint was incorrect and adjust the behavior. > Instead what is the opinion about giving an MADV_PAGEOUT_INACTIVE interface which does operate on a page only If it is on the inactive list and !PageWorkingset ? > Hope my feedback is useful and if we can improve Android's userspace > behavior, I'm happy to help make that happen. Thanks...
On Fri, Jun 9, 2023 at 5:42 AM Charan Teja Kalla <quic_charante@quicinc.com> wrote: > > Thanks Suren & Johannes, > > On 6/7/2023 1:18 AM, Suren Baghdasaryan wrote: > > Hi Folks. Sorry for being late to the party. > > Yeah, userspace does not have a crystal ball to predict future user > > behavior, so there will always be pathological cases when usual > > assumptions and resulting madvise() would make things worse. > > > > I think this discussion can be split into several questions/issues: > > 1. Inconsistency in how madvise(MADV_PAGEOUT) would affect PSI > > calculation when the page is refaulted, based on the path it took > > before being evicted by madvise(). In your initial description case > > (a) is inconsistent with (b) and (c) and it's probably worth fixing. > > IMHO (a) should be made consistent with others, not the other way > > around. My reasoning is that page was expelled from the active list, > > so it was part of the active workingset. > > > That means we should be setting Workingset on the page while it is on > the active list and when it is being pageout through madvising. Right? I > see, this makes it consistent. This was my opinion but others might think otherwise, like I found out in some recent conversations. So, it would be great to get some more feedback before making the change. > > On the same note, discussing with Suren offline, Should the refaulted > madvise pages start always at the inactive list? If they are really > active, they get promoted anyway.. > > > 2. Whether refaults caused by incorrect madvise(MADV_PAGEOUT) should > > be counted as workingset refault and affect PSI. > > This one I think is trickier. IMHO it should be counted as workingset > > refault simply because it was refaulted and it was part of the > > workingset. Whether it should affect PSI, which is supposed to be an > > indicator of "pressure" is, I think, debatable. With madvise() in the > > mix, refault might happen without any real memory pressure... So, the > > answer is not obvious to me. > > > > 3. Should refaults caused by incorrect madvise(MADV_PAGEOUT) be > > distinguished from the ones which were evicted by kernel reclaim > > mechanisms. > > I can see use for that from userspace to detect incorrect madvise() > > and adjust its aggressiveness. I think the API might get a bit complex > > because of the need to associate refaults with specific madvise()/VMAs > > to understand which hint was incorrect and adjust the behavior. > > Instead what is the opinion about giving an MADV_PAGEOUT_INACTIVE > interface which does operate on a page only If it is on the inactive > list and !PageWorkingset ? IOW you want a less aggressive mechanism which can be used by the userspace to tell the kernel "I think these pages won't be used but I'm not 100% sure, so drop them only if they are inactive"? I don't know how much that will help when the madvise() ends up being wrong but maybe you can quickly experiment and tell us if the difference is substantial? > > > Hope my feedback is useful and if we can improve Android's userspace > > behavior, I'm happy to help make that happen. > Thanks...
On Fri, Jun 09, 2023 at 04:13:14PM -0700, Suren Baghdasaryan wrote: > On Fri, Jun 9, 2023 at 5:42 AM Charan Teja Kalla > <quic_charante@quicinc.com> wrote: > > > > Thanks Suren & Johannes, > > > > On 6/7/2023 1:18 AM, Suren Baghdasaryan wrote: > > > Hi Folks. Sorry for being late to the party. > > > Yeah, userspace does not have a crystal ball to predict future user > > > behavior, so there will always be pathological cases when usual > > > assumptions and resulting madvise() would make things worse. > > > > > > I think this discussion can be split into several questions/issues: > > > 1. Inconsistency in how madvise(MADV_PAGEOUT) would affect PSI > > > calculation when the page is refaulted, based on the path it took > > > before being evicted by madvise(). In your initial description case > > > (a) is inconsistent with (b) and (c) and it's probably worth fixing. > > > IMHO (a) should be made consistent with others, not the other way > > > around. My reasoning is that page was expelled from the active list, > > > so it was part of the active workingset. > > > > > That means we should be setting Workingset on the page while it is on > > the active list and when it is being pageout through madvising. Right? I > > see, this makes it consistent. > > This was my opinion but others might think otherwise, like I found out > in some recent conversations. So, it would be great to get some more > feedback before making the change. I also agree with the consistency fix: it should set Workingset when madvise zaps pages from the active list.
On Fri, Jun 09, 2023 at 06:12:28PM +0530, Charan Teja Kalla wrote: > Thanks Suren & Johannes, > > On 6/7/2023 1:18 AM, Suren Baghdasaryan wrote: > > Hi Folks. Sorry for being late to the party. > > Yeah, userspace does not have a crystal ball to predict future user > > behavior, so there will always be pathological cases when usual > > assumptions and resulting madvise() would make things worse. > > > > I think this discussion can be split into several questions/issues: > > 1. Inconsistency in how madvise(MADV_PAGEOUT) would affect PSI > > calculation when the page is refaulted, based on the path it took > > before being evicted by madvise(). In your initial description case > > (a) is inconsistent with (b) and (c) and it's probably worth fixing. > > IMHO (a) should be made consistent with others, not the other way > > around. My reasoning is that page was expelled from the active list, > > so it was part of the active workingset. > > > That means we should be setting Workingset on the page while it is on > the active list and when it is being pageout through madvising. Right? I > see, this makes it consistent. > > On the same note, discussing with Suren offline, Should the refaulted > madvise pages start always at the inactive list? If they are really > active, they get promoted anyway.. > Can you elaborate on the rationale why refaulted madvise pages needs to be on inactive list? If it had not been paged out via madvise, it would have been activated no? Thanks, Pavan
Hi Suren, On 6/10/2023 4:43 AM, Suren Baghdasaryan wrote: >>> I can see use for that from userspace to detect incorrect madvise() >>> and adjust its aggressiveness. I think the API might get a bit complex >>> because of the need to associate refaults with specific madvise()/VMAs >>> to understand which hint was incorrect and adjust the behavior. >>> Instead what is the opinion about giving an MADV_PAGEOUT_INACTIVE >> interface which does operate on a page only If it is on the inactive >> list and !PageWorkingset ? > IOW you want a less aggressive mechanism which can be used by the > userspace to tell the kernel "I think these pages won't be used but > I'm not 100% sure, so drop them only if they are inactive"? > I don't know how much that will help when the madvise() ends up being > wrong but maybe you can quickly experiment and tell us if the > difference is substantial? We did some extensive testing on Android and this ask is not helping us much. I am really not sure if there is some other usecase that can benefit from this. So, for now I just stick to your suggestion of making the pages on the Active list as the Workingset at the time of pageout. > Thanks.
On Mon, Jun 26, 2023 at 7:31 AM Charan Teja Kalla <quic_charante@quicinc.com> wrote: > > Hi Suren, > > On 6/10/2023 4:43 AM, Suren Baghdasaryan wrote: > >>> I can see use for that from userspace to detect incorrect madvise() > >>> and adjust its aggressiveness. I think the API might get a bit complex > >>> because of the need to associate refaults with specific madvise()/VMAs > >>> to understand which hint was incorrect and adjust the behavior. > >>> Instead what is the opinion about giving an MADV_PAGEOUT_INACTIVE > >> interface which does operate on a page only If it is on the inactive > >> list and !PageWorkingset ? > > IOW you want a less aggressive mechanism which can be used by the > > userspace to tell the kernel "I think these pages won't be used but > > I'm not 100% sure, so drop them only if they are inactive"? > > I don't know how much that will help when the madvise() ends up being > > wrong but maybe you can quickly experiment and tell us if the > > difference is substantial? > > We did some extensive testing on Android and this ask is not helping us > much. I am really not sure if there is some other usecase that can > benefit from this. So, for now I just stick to your suggestion of making > the pages on the Active list as the Workingset at the time of pageout. > > Thanks for checking that. Your plan SGTM. > > Thanks. >
diff --git a/mm/madvise.c b/mm/madvise.c index 340125d..3410c39 100644 --- a/mm/madvise.c +++ b/mm/madvise.c @@ -409,8 +409,10 @@ static int madvise_cold_or_pageout_pte_range(pmd_t *pmd, if (folio_isolate_lru(folio)) { if (folio_test_unevictable(folio)) folio_putback_lru(folio); - else + else { + folio_clear_workingset(folio); list_add(&folio->lru, &folio_list); + } } } else folio_deactivate(folio); @@ -503,8 +505,10 @@ static int madvise_cold_or_pageout_pte_range(pmd_t *pmd, if (folio_isolate_lru(folio)) { if (folio_test_unevictable(folio)) folio_putback_lru(folio); - else + else { + folio_clear_workingset(folio); list_add(&folio->lru, &folio_list); + } } } else folio_deactivate(folio);
A folio turns into a Workingset during: 1) shrink_active_list() placing the folio from active to inactive list. 2) When a workingset transition is happening during the folio refault. And when Workingset is set on a folio, PSI for memory can be accounted during a) That folio is being reclaimed and b) Refault of that folio. There exists clients who can do the proactive reclaim using the system calls like madvise(), whose folios can be safely treated as inactive folios assuming the client knows that these folios are not needed in the near future thus wanted to reclaim them. For such folios psi is not accounted uniformly: a) A folio started at inactive and moved to active as part of accesses. Workingset is absent on the folio thus madvise(MADV_PAGEOUT) don't account such folios for PSI. b) When the same folio transition from inactive->active and then to inactive through shrink_active_list(). Workingset is set on the folio thus madvise(MADV_PAGEOUT) account such folios for PSI. c) When the same folio is part of active list directly as a result of folio refault and this was a workingset folio prior to eviction. Workingset is set on the folio thus madvise(MADV_PAGEOUT) account such folios for PSI. As said about the MADV_PAGEOUT on a folio is accounted in b) and c) but not in a) which is inconsistent. Remove this inconsistency by always not considering the PSI for folios that are getting reclaimed through madvise(MADV_PAGEOUT) by clearing the Workingset on a folio. This consistency of clearing the workingset was chosen under the assumption that client knows these folios are not in active use thus reclaiming them hence not eligible as workingset folios. Probably it is the same reason why workingset is not set on a folio through MADV_COLD but during the shrink_active_list() though both the actions make the folio put onto the inactive list. This patch is tested on Android, Snapdragon SoC with 8Gb RAM, 4GB swap mounted on zram which has 2GB of backingdev. The test case involved launching some memory hungry apps in an order and do the proactive reclaim for the app that went to background using madvise(MADV_PAGEOUT). We are seeing ~40% less total values of psi mem some and full when this patch is combined with [1]. [1]https://lore.kernel.org/all/20220214214921.419687-1-hannes@cmpxchg.org/T/#u Signed-off-by: Charan Teja Kalla <quic_charante@quicinc.com> --- mm/madvise.c | 8 ++++++-- 1 file changed, 6 insertions(+), 2 deletions(-)