| Message ID | 20181113055252.79406-8-sashal@kernel.org (mailing list archive) |
|---|---|
| State | New, archived |
| Headers | show |
| Series | None | expand |
On Tue, 13 Nov 2018 00:52:51 -0500 Sasha Levin <sashal@kernel.org> wrote: > From: Jann Horn <jannh@google.com> > > [ Upstream commit f0ecf25a093fc0589f0a6bc4c1ea068bbb67d220 ] > > Having two gigantic arrays that must manually be kept in sync, including > ifdefs, isn't exactly robust. To make it easier to catch such issues in > the future, add a BUILD_BUG_ON(). > > ... > > --- a/mm/vmstat.c > +++ b/mm/vmstat.c > @@ -1189,6 +1189,8 @@ static void *vmstat_start(struct seq_file *m, loff_t *pos) > stat_items_size += sizeof(struct vm_event_state); > #endif > > + BUILD_BUG_ON(stat_items_size != > + ARRAY_SIZE(vmstat_text) * sizeof(unsigned long)); > v = kmalloc(stat_items_size, GFP_KERNEL); > m->private = v; > if (!v) I don't think there's any way in which this can make a -stable kernel more stable! Generally, I consider -stable in every patch I merge, so for each patch which doesn't have cc:stable, that tag is missing for a reason. In other words, your criteria for -stable addition are different from mine. And I think your criteria differ from those described in Documentation/process/stable-kernel-rules.rst. So... what is your overall thinking on patch selection?
On Thu, Nov 15, 2018 at 02:08:10PM -0800, Andrew Morton wrote: >On Tue, 13 Nov 2018 00:52:51 -0500 Sasha Levin <sashal@kernel.org> wrote: > >> From: Jann Horn <jannh@google.com> >> >> [ Upstream commit f0ecf25a093fc0589f0a6bc4c1ea068bbb67d220 ] >> >> Having two gigantic arrays that must manually be kept in sync, including >> ifdefs, isn't exactly robust. To make it easier to catch such issues in >> the future, add a BUILD_BUG_ON(). >> >> ... >> >> --- a/mm/vmstat.c >> +++ b/mm/vmstat.c >> @@ -1189,6 +1189,8 @@ static void *vmstat_start(struct seq_file *m, loff_t *pos) >> stat_items_size += sizeof(struct vm_event_state); >> #endif >> >> + BUILD_BUG_ON(stat_items_size != >> + ARRAY_SIZE(vmstat_text) * sizeof(unsigned long)); >> v = kmalloc(stat_items_size, GFP_KERNEL); >> m->private = v; >> if (!v) > >I don't think there's any way in which this can make a -stable kernel >more stable! > > >Generally, I consider -stable in every patch I merge, so for each patch >which doesn't have cc:stable, that tag is missing for a reason. > >In other words, your criteria for -stable addition are different from >mine. > >And I think your criteria differ from those described in >Documentation/process/stable-kernel-rules.rst. > >So... what is your overall thinking on patch selection? Indeed, this doesn't fix anything. My concern is that in the future, we will pull a patch that will cause the issue described here, and that issue will only be relevant on stable. It is very hard to debug this, and I suspect that stable kernels will still pass all their tests with flying colors. As an example, consider the case where commit 28e2c4bb99aa ("mm/vmstat.c: fix outdated vmstat_text") is backported to a kernel that doesn't have commit 7a9cdebdcc17 ("mm: get rid of vmacache_flush_all() entirely"). I also felt safe with this patch since it adds a single BUILD_BUG_ON() which does nothing during runtime, so the chances it introduces anything beyond a build regression seemed to be slim to none. -- Thanks, Sasha
On Thu, 15 Nov 2018 17:37:18 -0500 Sasha Levin <sashal@kernel.org> wrote: > On Thu, Nov 15, 2018 at 02:08:10PM -0800, Andrew Morton wrote: > >On Tue, 13 Nov 2018 00:52:51 -0500 Sasha Levin <sashal@kernel.org> wrote: > > > >> From: Jann Horn <jannh@google.com> > >> > >> [ Upstream commit f0ecf25a093fc0589f0a6bc4c1ea068bbb67d220 ] > >> > >> Having two gigantic arrays that must manually be kept in sync, including > >> ifdefs, isn't exactly robust. To make it easier to catch such issues in > >> the future, add a BUILD_BUG_ON(). > >> > >> ... > >> > >> --- a/mm/vmstat.c > >> +++ b/mm/vmstat.c > >> @@ -1189,6 +1189,8 @@ static void *vmstat_start(struct seq_file *m, loff_t *pos) > >> stat_items_size += sizeof(struct vm_event_state); > >> #endif > >> > >> + BUILD_BUG_ON(stat_items_size != > >> + ARRAY_SIZE(vmstat_text) * sizeof(unsigned long)); > >> v = kmalloc(stat_items_size, GFP_KERNEL); > >> m->private = v; > >> if (!v) > > > >I don't think there's any way in which this can make a -stable kernel > >more stable! > > > > > >Generally, I consider -stable in every patch I merge, so for each patch > >which doesn't have cc:stable, that tag is missing for a reason. > > > >In other words, your criteria for -stable addition are different from > >mine. > > > >And I think your criteria differ from those described in > >Documentation/process/stable-kernel-rules.rst. > > > >So... what is your overall thinking on patch selection? > > Indeed, this doesn't fix anything. > > My concern is that in the future, we will pull a patch that will cause > the issue described here, and that issue will only be relevant on > stable. It is very hard to debug this, and I suspect that stable kernels > will still pass all their tests with flying colors. > > As an example, consider the case where commit 28e2c4bb99aa ("mm/vmstat.c: > fix outdated vmstat_text") is backported to a kernel that doesn't have > commit 7a9cdebdcc17 ("mm: get rid of vmacache_flush_all() entirely"). > > I also felt safe with this patch since it adds a single BUILD_BUG_ON() > which does nothing during runtime, so the chances it introduces anything > beyond a build regression seemed to be slim to none. Well OK. But my question was general and covers basically every autosel patch which originated in -mm.
On Thu, Nov 15, 2018 at 02:47:19PM -0800, Andrew Morton wrote: >On Thu, 15 Nov 2018 17:37:18 -0500 Sasha Levin <sashal@kernel.org> wrote: > >> On Thu, Nov 15, 2018 at 02:08:10PM -0800, Andrew Morton wrote: >> >On Tue, 13 Nov 2018 00:52:51 -0500 Sasha Levin <sashal@kernel.org> wrote: >> > >> >> From: Jann Horn <jannh@google.com> >> >> >> >> [ Upstream commit f0ecf25a093fc0589f0a6bc4c1ea068bbb67d220 ] >> >> >> >> Having two gigantic arrays that must manually be kept in sync, including >> >> ifdefs, isn't exactly robust. To make it easier to catch such issues in >> >> the future, add a BUILD_BUG_ON(). >> >> >> >> ... >> >> >> >> --- a/mm/vmstat.c >> >> +++ b/mm/vmstat.c >> >> @@ -1189,6 +1189,8 @@ static void *vmstat_start(struct seq_file *m, loff_t *pos) >> >> stat_items_size += sizeof(struct vm_event_state); >> >> #endif >> >> >> >> + BUILD_BUG_ON(stat_items_size != >> >> + ARRAY_SIZE(vmstat_text) * sizeof(unsigned long)); >> >> v = kmalloc(stat_items_size, GFP_KERNEL); >> >> m->private = v; >> >> if (!v) >> > >> >I don't think there's any way in which this can make a -stable kernel >> >more stable! >> > >> > >> >Generally, I consider -stable in every patch I merge, so for each patch >> >which doesn't have cc:stable, that tag is missing for a reason. >> > >> >In other words, your criteria for -stable addition are different from >> >mine. >> > >> >And I think your criteria differ from those described in >> >Documentation/process/stable-kernel-rules.rst. >> > >> >So... what is your overall thinking on patch selection? >> >> Indeed, this doesn't fix anything. >> >> My concern is that in the future, we will pull a patch that will cause >> the issue described here, and that issue will only be relevant on >> stable. It is very hard to debug this, and I suspect that stable kernels >> will still pass all their tests with flying colors. >> >> As an example, consider the case where commit 28e2c4bb99aa ("mm/vmstat.c: >> fix outdated vmstat_text") is backported to a kernel that doesn't have >> commit 7a9cdebdcc17 ("mm: get rid of vmacache_flush_all() entirely"). >> >> I also felt safe with this patch since it adds a single BUILD_BUG_ON() >> which does nothing during runtime, so the chances it introduces anything >> beyond a build regression seemed to be slim to none. > >Well OK. But my question was general and covers basically every >autosel patch which originated in -mm. Sure. I picked 3 patches that show up on top when I google for AUTOSEL in linux-mm, maybe they'll be a good example to help me understand why they were not selected. This one fixes a case where too few struct pages are allocated when using mirrorred memory: https://marc.info/?l=linux-mm&m=154211933211147&w=2 Race condition with memory hotplug due to missing locks: https://marc.info/?l=linux-mm&m=154211934011188&w=2 Raising an OOM event that causes issues in userspace when no OOM has actually occured: https://marc.info/?l=linux-mm&m=154211939811582&w=2 I think that all 3 cases represent a "real" bug users can hit, and I honestly don't know why they were not tagged for stable. -- Thanks, Sasha
On Thu 15-11-18 18:01:18, Sasha Levin wrote: > On Thu, Nov 15, 2018 at 02:47:19PM -0800, Andrew Morton wrote: > > On Thu, 15 Nov 2018 17:37:18 -0500 Sasha Levin <sashal@kernel.org> wrote: > > > > > On Thu, Nov 15, 2018 at 02:08:10PM -0800, Andrew Morton wrote: > > > >On Tue, 13 Nov 2018 00:52:51 -0500 Sasha Levin <sashal@kernel.org> wrote: > > > > > > > >> From: Jann Horn <jannh@google.com> > > > >> > > > >> [ Upstream commit f0ecf25a093fc0589f0a6bc4c1ea068bbb67d220 ] > > > >> > > > >> Having two gigantic arrays that must manually be kept in sync, including > > > >> ifdefs, isn't exactly robust. To make it easier to catch such issues in > > > >> the future, add a BUILD_BUG_ON(). > > > >> > > > >> ... > > > >> > > > >> --- a/mm/vmstat.c > > > >> +++ b/mm/vmstat.c > > > >> @@ -1189,6 +1189,8 @@ static void *vmstat_start(struct seq_file *m, loff_t *pos) > > > >> stat_items_size += sizeof(struct vm_event_state); > > > >> #endif > > > >> > > > >> + BUILD_BUG_ON(stat_items_size != > > > >> + ARRAY_SIZE(vmstat_text) * sizeof(unsigned long)); > > > >> v = kmalloc(stat_items_size, GFP_KERNEL); > > > >> m->private = v; > > > >> if (!v) > > > > > > > >I don't think there's any way in which this can make a -stable kernel > > > >more stable! > > > > > > > > > > > >Generally, I consider -stable in every patch I merge, so for each patch > > > >which doesn't have cc:stable, that tag is missing for a reason. > > > > > > > >In other words, your criteria for -stable addition are different from > > > >mine. > > > > > > > >And I think your criteria differ from those described in > > > >Documentation/process/stable-kernel-rules.rst. > > > > > > > >So... what is your overall thinking on patch selection? > > > > > > Indeed, this doesn't fix anything. > > > > > > My concern is that in the future, we will pull a patch that will cause > > > the issue described here, and that issue will only be relevant on > > > stable. It is very hard to debug this, and I suspect that stable kernels > > > will still pass all their tests with flying colors. > > > > > > As an example, consider the case where commit 28e2c4bb99aa ("mm/vmstat.c: > > > fix outdated vmstat_text") is backported to a kernel that doesn't have > > > commit 7a9cdebdcc17 ("mm: get rid of vmacache_flush_all() entirely"). > > > > > > I also felt safe with this patch since it adds a single BUILD_BUG_ON() > > > which does nothing during runtime, so the chances it introduces anything > > > beyond a build regression seemed to be slim to none. > > > > Well OK. But my question was general and covers basically every > > autosel patch which originated in -mm. > > Sure. I picked 3 patches that show up on top when I google for AUTOSEL > in linux-mm, maybe they'll be a good example to help me understand why > they were not selected. > > This one fixes a case where too few struct pages are allocated when > using mirrorred memory: > > https://marc.info/?l=linux-mm&m=154211933211147&w=2 Let me quote "I found this bug by reading the code." I do not think anybody has ever seen this in practice. > Race condition with memory hotplug due to missing locks: > > https://marc.info/?l=linux-mm&m=154211934011188&w=2 Memory hotplug locking is dubious at best and this patch doesn't really fix it. It fixes a theoretical problem. I am not aware anybody would be hitting in practice. We need to rework the locking quite extensively. > Raising an OOM event that causes issues in userspace when no OOM has > actually occured: > > https://marc.info/?l=linux-mm&m=154211939811582&w=2 The patch makes sense I just do not think this is a stable material. The semantic of the event was and still is suboptimal. > I think that all 3 cases represent a "real" bug users can hit, and I > honestly don't know why they were not tagged for stable. It would be much better to ask in the respective email thread rather than spamming mailing with AUTOSEL patches which rarely get any attention. We have been through this discussion several times already and I thought we have agreed that those subsystems which are seriously considering stable are opted out from the AUTOSEL automagic. Has anything changed in that regards.
On Fri, Nov 16, 2018 at 09:55:25AM +0100, Michal Hocko wrote: >On Thu 15-11-18 18:01:18, Sasha Levin wrote: >> On Thu, Nov 15, 2018 at 02:47:19PM -0800, Andrew Morton wrote: >> > On Thu, 15 Nov 2018 17:37:18 -0500 Sasha Levin <sashal@kernel.org> wrote: >> > >> > > On Thu, Nov 15, 2018 at 02:08:10PM -0800, Andrew Morton wrote: >> > > >On Tue, 13 Nov 2018 00:52:51 -0500 Sasha Levin <sashal@kernel.org> wrote: >> > > > >> > > >> From: Jann Horn <jannh@google.com> >> > > >> >> > > >> [ Upstream commit f0ecf25a093fc0589f0a6bc4c1ea068bbb67d220 ] >> > > >> >> > > >> Having two gigantic arrays that must manually be kept in sync, including >> > > >> ifdefs, isn't exactly robust. To make it easier to catch such issues in >> > > >> the future, add a BUILD_BUG_ON(). >> > > >> >> > > >> ... >> > > >> >> > > >> --- a/mm/vmstat.c >> > > >> +++ b/mm/vmstat.c >> > > >> @@ -1189,6 +1189,8 @@ static void *vmstat_start(struct seq_file *m, loff_t *pos) >> > > >> stat_items_size += sizeof(struct vm_event_state); >> > > >> #endif >> > > >> >> > > >> + BUILD_BUG_ON(stat_items_size != >> > > >> + ARRAY_SIZE(vmstat_text) * sizeof(unsigned long)); >> > > >> v = kmalloc(stat_items_size, GFP_KERNEL); >> > > >> m->private = v; >> > > >> if (!v) >> > > > >> > > >I don't think there's any way in which this can make a -stable kernel >> > > >more stable! >> > > > >> > > > >> > > >Generally, I consider -stable in every patch I merge, so for each patch >> > > >which doesn't have cc:stable, that tag is missing for a reason. >> > > > >> > > >In other words, your criteria for -stable addition are different from >> > > >mine. >> > > > >> > > >And I think your criteria differ from those described in >> > > >Documentation/process/stable-kernel-rules.rst. >> > > > >> > > >So... what is your overall thinking on patch selection? >> > > >> > > Indeed, this doesn't fix anything. >> > > >> > > My concern is that in the future, we will pull a patch that will cause >> > > the issue described here, and that issue will only be relevant on >> > > stable. It is very hard to debug this, and I suspect that stable kernels >> > > will still pass all their tests with flying colors. >> > > >> > > As an example, consider the case where commit 28e2c4bb99aa ("mm/vmstat.c: >> > > fix outdated vmstat_text") is backported to a kernel that doesn't have >> > > commit 7a9cdebdcc17 ("mm: get rid of vmacache_flush_all() entirely"). >> > > >> > > I also felt safe with this patch since it adds a single BUILD_BUG_ON() >> > > which does nothing during runtime, so the chances it introduces anything >> > > beyond a build regression seemed to be slim to none. >> > >> > Well OK. But my question was general and covers basically every >> > autosel patch which originated in -mm. >> >> Sure. I picked 3 patches that show up on top when I google for AUTOSEL >> in linux-mm, maybe they'll be a good example to help me understand why >> they were not selected. >> >> This one fixes a case where too few struct pages are allocated when >> using mirrorred memory: >> >> https://marc.info/?l=linux-mm&m=154211933211147&w=2 > >Let me quote "I found this bug by reading the code." I do not think >anybody has ever seen this in practice. > >> Race condition with memory hotplug due to missing locks: >> >> https://marc.info/?l=linux-mm&m=154211934011188&w=2 > >Memory hotplug locking is dubious at best and this patch doesn't really >fix it. It fixes a theoretical problem. I am not aware anybody would be >hitting in practice. We need to rework the locking quite extensively. The word "theoretical" used in the stable rules file does not mean that we need to have actual reports of users hitting bugs before we start backporting the relevant patch, it simply suggests that there needs to be a reasonable explanation of how this issue can be hit. For this memory hotplug patch in particular, I use the hv_balloon driver at this very moment (running a linux guest on windows, with "dynamic memory" enabled). Should I wait for it to crash before I can fix it? Is the upstream code perfect? No, but that doesn't mean that it's not working at all, and if there are users they expect to see fixes going in and not just sitting idly waiting for a big rewrite that will come in a few years. Memory hotplug fixes are not something you think should go to stable? Andrew sent a few of them to stable, so that can't be the case. >> Raising an OOM event that causes issues in userspace when no OOM has >> actually occured: >> >> https://marc.info/?l=linux-mm&m=154211939811582&w=2 > >The patch makes sense I just do not think this is a stable material. The >semantic of the event was and still is suboptimal. I really fail to understand your reasoning about -stable here. This patch is something people actually hit in the field, spent time on triaging and analysing it, and submitting a fix which looks reasonably straightforward. That fix was acked by quite a few folks (including yourself) and merged in. And as far as we can tell, it actually fixed the problem. Why is it not stable material? My understanding is that you're concerned with the patch itself being "suboptimal", but in that case - why did you ack it? >> I think that all 3 cases represent a "real" bug users can hit, and I >> honestly don't know why they were not tagged for stable. > >It would be much better to ask in the respective email thread rather >than spamming mailing with AUTOSEL patches which rarely get any >attention. I actually tried it, but the comments I got is that it gets in the way and people preferred something they can filter. >We have been through this discussion several times already and I thought >we have agreed that those subsystems which are seriously considering stable >are opted out from the AUTOSEL automagic. Has anything changed in that >regards. I checked in with Andrew to get his input on this, he suggested that these patches should be sent to linux-mm and he'll give it a close look. Ultimately this is the subsystem's decision, yes, but I was under the impression that this decision wasn't made yet. I guess that I'm really failing to understand why patches like the third one here (the OOM one) are being kept out. -- Thanks, Sasha
On Fri 16-11-18 13:19:04, Sasha Levin wrote: > On Fri, Nov 16, 2018 at 09:55:25AM +0100, Michal Hocko wrote: [...] > > > Race condition with memory hotplug due to missing locks: > > > > > > https://marc.info/?l=linux-mm&m=154211934011188&w=2 > > > > Memory hotplug locking is dubious at best and this patch doesn't really > > fix it. It fixes a theoretical problem. I am not aware anybody would be > > hitting in practice. We need to rework the locking quite extensively. > > The word "theoretical" used in the stable rules file does not mean > that we need to have actual reports of users hitting bugs before we > start backporting the relevant patch, it simply suggests that there > needs to be a reasonable explanation of how this issue can be hit. > > For this memory hotplug patch in particular, I use the hv_balloon driver > at this very moment (running a linux guest on windows, with "dynamic > memory" enabled). Should I wait for it to crash before I can fix it? > > Is the upstream code perfect? No, but that doesn't mean that it's not > working at all, and if there are users they expect to see fixes going in > and not just sitting idly waiting for a big rewrite that will come in a > few years. > > Memory hotplug fixes are not something you think should go to stable? > Andrew sent a few of them to stable, so that can't be the case. I am not arguing about hotplug fixes in general. I was arguing that this particular one is a theoretical one and hotplug locking is quite subtle. E.g. 381eab4a6ee mm/memory_hotplug: fix online/offline_pages called w.o. mem_hotplug_lock http://lkml.kernel.org/r/20181114070909.GB2653@MiWiFi-R3L-srv So in general unless the issue is really triggered easily I am rather conservative. > > > Raising an OOM event that causes issues in userspace when no OOM has > > > actually occured: > > > > > > https://marc.info/?l=linux-mm&m=154211939811582&w=2 > > > > The patch makes sense I just do not think this is a stable material. The > > semantic of the event was and still is suboptimal. > > I really fail to understand your reasoning about -stable here. This > patch is something people actually hit in the field, spent time on > triaging and analysing it, and submitting a fix which looks reasonably > straightforward. > > That fix was acked by quite a few folks (including yourself) and merged > in. And as far as we can tell, it actually fixed the problem. > > Why is it not stable material? Because the semantic of the OOM event is quite tricky itself. We have discussed this patch and concluded that the updated one is more sensible. But it is not yet clear whether this is actually what other users expect as well. That to me does sound quite risky for a stable kernel. > My understanding is that you're concerned with the patch itself being > "suboptimal", but in that case - why did you ack it? > > > > I think that all 3 cases represent a "real" bug users can hit, and I > > > honestly don't know why they were not tagged for stable. > > > > It would be much better to ask in the respective email thread rather > > than spamming mailing with AUTOSEL patches which rarely get any > > attention. > > I actually tried it, but the comments I got is that it gets in the way > and people preferred something they can filter. which means that AUTOSEL just goes to /dev/null... > > We have been through this discussion several times already and I thought > > we have agreed that those subsystems which are seriously considering stable > > are opted out from the AUTOSEL automagic. Has anything changed in that > > regards. > > I checked in with Andrew to get his input on this, he suggested that > these patches should be sent to linux-mm and he'll give it a close look. If Andrew is happy to get AUTOSEL patches then I will not object of course but let's not merge these patches without and expclicit OK.
On Fri, Nov 16, 2018 at 07:44:57PM +0100, Michal Hocko wrote: >On Fri 16-11-18 13:19:04, Sasha Levin wrote: >> On Fri, Nov 16, 2018 at 09:55:25AM +0100, Michal Hocko wrote: >[...] >> > > Race condition with memory hotplug due to missing locks: >> > > >> > > https://marc.info/?l=linux-mm&m=154211934011188&w=2 >> > >> > Memory hotplug locking is dubious at best and this patch doesn't really >> > fix it. It fixes a theoretical problem. I am not aware anybody would be >> > hitting in practice. We need to rework the locking quite extensively. >> >> The word "theoretical" used in the stable rules file does not mean >> that we need to have actual reports of users hitting bugs before we >> start backporting the relevant patch, it simply suggests that there >> needs to be a reasonable explanation of how this issue can be hit. >> >> For this memory hotplug patch in particular, I use the hv_balloon driver >> at this very moment (running a linux guest on windows, with "dynamic >> memory" enabled). Should I wait for it to crash before I can fix it? >> >> Is the upstream code perfect? No, but that doesn't mean that it's not >> working at all, and if there are users they expect to see fixes going in >> and not just sitting idly waiting for a big rewrite that will come in a >> few years. >> >> Memory hotplug fixes are not something you think should go to stable? >> Andrew sent a few of them to stable, so that can't be the case. > >I am not arguing about hotplug fixes in general. I was arguing that this >particular one is a theoretical one and hotplug locking is quite subtle. >E.g. 381eab4a6ee mm/memory_hotplug: fix online/offline_pages called w.o. mem_hotplug_lock >http://lkml.kernel.org/r/20181114070909.GB2653@MiWiFi-R3L-srv >So in general unless the issue is really triggered easily I am rather >conservative. We have millions of machines running linux, everything is triggered "easily" at that scale. >> > > Raising an OOM event that causes issues in userspace when no OOM has >> > > actually occured: >> > > >> > > https://marc.info/?l=linux-mm&m=154211939811582&w=2 >> > >> > The patch makes sense I just do not think this is a stable material. The >> > semantic of the event was and still is suboptimal. >> >> I really fail to understand your reasoning about -stable here. This >> patch is something people actually hit in the field, spent time on >> triaging and analysing it, and submitting a fix which looks reasonably >> straightforward. >> >> That fix was acked by quite a few folks (including yourself) and merged >> in. And as far as we can tell, it actually fixed the problem. >> >> Why is it not stable material? > >Because the semantic of the OOM event is quite tricky itself. We have >discussed this patch and concluded that the updated one is more >sensible. But it is not yet clear whether this is actually what other >users expect as well. That to me does sound quite risky for a stable >kernel. So there's another patch following this one that fixes it? Sure - can I take both? Users expect to not have their containers die randomly, if you're saying that you're still working on a fix for that then that is a different story than saying "we fixed it, but it should not go to stable". And let's also draw a line there, users will not wait for the OOM event logic to be perfect before they can expect their workloads to run without issues. >> My understanding is that you're concerned with the patch itself being >> "suboptimal", but in that case - why did you ack it? >> >> > > I think that all 3 cases represent a "real" bug users can hit, and I >> > > honestly don't know why they were not tagged for stable. >> > >> > It would be much better to ask in the respective email thread rather >> > than spamming mailing with AUTOSEL patches which rarely get any >> > attention. >> >> I actually tried it, but the comments I got is that it gets in the way >> and people preferred something they can filter. > >which means that AUTOSEL just goes to /dev/null... Or just not get mixed with the process? for some people it's easier to see AUTOSEL mails with the way it works now rather than if they suddenly show up as a continuation of a weeks old thread. >> > We have been through this discussion several times already and I thought >> > we have agreed that those subsystems which are seriously considering stable >> > are opted out from the AUTOSEL automagic. Has anything changed in that >> > regards. >> >> I checked in with Andrew to get his input on this, he suggested that >> these patches should be sent to linux-mm and he'll give it a close look. > >If Andrew is happy to get AUTOSEL patches then I will not object of >course but let's not merge these patches without and expclicit OK. This is fair. I think that the process has caused some unnecessary friction: we all want the same result but just disagree on the means :) I won't merge any mm/ AUTOSEL patches until this gets clearer. -- Thanks, Sasha
On Fri 16-11-18 14:19:10, Sasha Levin wrote: > On Fri, Nov 16, 2018 at 07:44:57PM +0100, Michal Hocko wrote: > > On Fri 16-11-18 13:19:04, Sasha Levin wrote: > > > On Fri, Nov 16, 2018 at 09:55:25AM +0100, Michal Hocko wrote: > > [...] > > > > > Race condition with memory hotplug due to missing locks: > > > > > > > > > > https://marc.info/?l=linux-mm&m=154211934011188&w=2 > > > > > > > > Memory hotplug locking is dubious at best and this patch doesn't really > > > > fix it. It fixes a theoretical problem. I am not aware anybody would be > > > > hitting in practice. We need to rework the locking quite extensively. > > > > > > The word "theoretical" used in the stable rules file does not mean > > > that we need to have actual reports of users hitting bugs before we > > > start backporting the relevant patch, it simply suggests that there > > > needs to be a reasonable explanation of how this issue can be hit. > > > > > > For this memory hotplug patch in particular, I use the hv_balloon driver > > > at this very moment (running a linux guest on windows, with "dynamic > > > memory" enabled). Should I wait for it to crash before I can fix it? > > > > > > Is the upstream code perfect? No, but that doesn't mean that it's not > > > working at all, and if there are users they expect to see fixes going in > > > and not just sitting idly waiting for a big rewrite that will come in a > > > few years. > > > > > > Memory hotplug fixes are not something you think should go to stable? > > > Andrew sent a few of them to stable, so that can't be the case. > > > > I am not arguing about hotplug fixes in general. I was arguing that this > > particular one is a theoretical one and hotplug locking is quite subtle. > > E.g. 381eab4a6ee mm/memory_hotplug: fix online/offline_pages called w.o. mem_hotplug_lock > > http://lkml.kernel.org/r/20181114070909.GB2653@MiWiFi-R3L-srv > > So in general unless the issue is really triggered easily I am rather > > conservative. > > We have millions of machines running linux, everything is triggered > "easily" at that scale. yet a zero report... > > > > > Raising an OOM event that causes issues in userspace when no OOM has > > > > > actually occured: > > > > > > > > > > https://marc.info/?l=linux-mm&m=154211939811582&w=2 > > > > > > > > The patch makes sense I just do not think this is a stable material. The > > > > semantic of the event was and still is suboptimal. > > > > > > I really fail to understand your reasoning about -stable here. This > > > patch is something people actually hit in the field, spent time on > > > triaging and analysing it, and submitting a fix which looks reasonably > > > straightforward. > > > > > > That fix was acked by quite a few folks (including yourself) and merged > > > in. And as far as we can tell, it actually fixed the problem. > > > > > > Why is it not stable material? > > > > Because the semantic of the OOM event is quite tricky itself. We have > > discussed this patch and concluded that the updated one is more > > sensible. But it is not yet clear whether this is actually what other > > users expect as well. That to me does sound quite risky for a stable > > kernel. > > So there's another patch following this one that fixes it? Sure - can I > take both? No. There is no known bug. I am arguing that such a change needs some time to settle. I am quite skeptical that this will actually trigger any bug. I will not _object_ if this was merged if somebody explicitly asks for it. I am saying that I am not convinced it is a stable material. So I guess our views on what is stable material differ. As I have said several times already, I think the volume of patches flowing to the stable tree is really high. To the point that taking stable trees for our SLES kernels become problematic. I have heard the similar from others. More is not always better. But let's not repeat this discussion again. If Andrew doesn't mind then keep sending AUTOSEL emails but please let's not apply those patches automatically. Thanks!
diff --git a/mm/vmstat.c b/mm/vmstat.c index 4590aa42b6cd..792374f7088f 100644 --- a/mm/vmstat.c +++ b/mm/vmstat.c @@ -1189,6 +1189,8 @@ static void *vmstat_start(struct seq_file *m, loff_t *pos) stat_items_size += sizeof(struct vm_event_state); #endif + BUILD_BUG_ON(stat_items_size != + ARRAY_SIZE(vmstat_text) * sizeof(unsigned long)); v = kmalloc(stat_items_size, GFP_KERNEL); m->private = v; if (!v)