| Message ID | 20200819101157.12723-12-david@redhat.com (mailing list archive) |
|---|---|
| State | New, archived |
| Headers | show |
| Series | mm/memory_hotplug: online_pages()/offline_pages() cleanups | expand |
On Wed 19-08-20 12:11:57, David Hildenbrand wrote: > Currently, it can happen that pages are allocated (and freed) via the buddy > before we finished basic memory onlining. > > For example, pages are exposed to the buddy and can be allocated before > we actually mark the sections online. Allocated pages could suddenly > fail pfn_to_online_page() checks. We had similar issues with pcp > handling, when pages are allocated+freed before we reach > zone_pcp_update() in online_pages() [1]. > > Instead, mark all pageblocks MIGRATE_ISOLATE, such that allocations are > impossible. Once done with the heavy lifting, use > undo_isolate_page_range() to move the pages to the MIGRATE_MOVABLE > freelist, marking them ready for allocation. Similar to offline_pages(), > we have to manually adjust zone->nr_isolate_pageblock. > > [1] https://lkml.kernel.org/r/1597150703-19003-1-git-send-email-charante@codeaurora.org > > Cc: Andrew Morton <akpm@linux-foundation.org> > Cc: Michal Hocko <mhocko@suse.com> > Cc: Wei Yang <richard.weiyang@linux.alibaba.com> > Cc: Baoquan He <bhe@redhat.com> > Cc: Pankaj Gupta <pankaj.gupta.linux@gmail.com> > Cc: Oscar Salvador <osalvador@suse.de> > Cc: Charan Teja Reddy <charante@codeaurora.org> > Signed-off-by: David Hildenbrand <david@redhat.com> Acked-by: Michal Hocko <mhocko@suse.com> Yes this looks very sensible and we should have done that from the beginning. I just have one minor comment below > @@ -816,6 +816,14 @@ int __ref online_pages(unsigned long pfn, unsigned long nr_pages, > if (ret) > goto failed_addition; > > + /* > + * Fixup the number of isolated pageblocks before marking the sections > + * onlining, such that undo_isolate_page_range() works correctly. > + */ > + spin_lock_irqsave(&zone->lock, flags); > + zone->nr_isolate_pageblock += nr_pages / pageblock_nr_pages; > + spin_unlock_irqrestore(&zone->lock, flags); > + I am not entirely happy about this. I am wondering whether it would make more sense to keep the counter in sync already in memmap_init_zone. Sure we add a branch to the boot time initialization - and it always fails there - but the code would be cleaner and we wouldn't have to do tricks like this in caller(s).
On 19.08.20 15:16, Michal Hocko wrote: > On Wed 19-08-20 12:11:57, David Hildenbrand wrote: >> Currently, it can happen that pages are allocated (and freed) via the buddy >> before we finished basic memory onlining. >> >> For example, pages are exposed to the buddy and can be allocated before >> we actually mark the sections online. Allocated pages could suddenly >> fail pfn_to_online_page() checks. We had similar issues with pcp >> handling, when pages are allocated+freed before we reach >> zone_pcp_update() in online_pages() [1]. >> >> Instead, mark all pageblocks MIGRATE_ISOLATE, such that allocations are >> impossible. Once done with the heavy lifting, use >> undo_isolate_page_range() to move the pages to the MIGRATE_MOVABLE >> freelist, marking them ready for allocation. Similar to offline_pages(), >> we have to manually adjust zone->nr_isolate_pageblock. >> >> [1] https://lkml.kernel.org/r/1597150703-19003-1-git-send-email-charante@codeaurora.org >> >> Cc: Andrew Morton <akpm@linux-foundation.org> >> Cc: Michal Hocko <mhocko@suse.com> >> Cc: Wei Yang <richard.weiyang@linux.alibaba.com> >> Cc: Baoquan He <bhe@redhat.com> >> Cc: Pankaj Gupta <pankaj.gupta.linux@gmail.com> >> Cc: Oscar Salvador <osalvador@suse.de> >> Cc: Charan Teja Reddy <charante@codeaurora.org> >> Signed-off-by: David Hildenbrand <david@redhat.com> > > Acked-by: Michal Hocko <mhocko@suse.com> > > Yes this looks very sensible and we should have done that from the > beginning. I just have one minor comment below >> @@ -816,6 +816,14 @@ int __ref online_pages(unsigned long pfn, unsigned long nr_pages, >> if (ret) >> goto failed_addition; >> >> + /* >> + * Fixup the number of isolated pageblocks before marking the sections >> + * onlining, such that undo_isolate_page_range() works correctly. >> + */ >> + spin_lock_irqsave(&zone->lock, flags); >> + zone->nr_isolate_pageblock += nr_pages / pageblock_nr_pages; >> + spin_unlock_irqrestore(&zone->lock, flags); >> + > > I am not entirely happy about this. I am wondering whether it would make > more sense to keep the counter in sync already in memmap_init_zone. Sure > we add a branch to the boot time initialization - and it always fails > there - but the code would be cleaner and we wouldn't have to do tricks > like this in caller(s). I had that in mind initially. The issue is that we have to fixup in case onlining fails, which I consider even more ugly. Also 1. It's the complete reverse of the offlining path now. 2. pageblock flags are essentially stale unless the section is online, my approach moves the handling to the point where nothing else will go wrong and we are just about to mark sections online. That looks a little cleaner to me. Unless there are strong opinions, I'd prefer to keep it like this. Thanks for the very fast review Michal!
diff --git a/mm/Kconfig b/mm/Kconfig index 6c974888f86f9..85a16ce1dbc49 100644 --- a/mm/Kconfig +++ b/mm/Kconfig @@ -152,6 +152,7 @@ config HAVE_BOOTMEM_INFO_NODE # eventually, we can have this option just 'select SPARSEMEM' config MEMORY_HOTPLUG bool "Allow for memory hot-add" + select MEMORY_ISOLATION depends on SPARSEMEM || X86_64_ACPI_NUMA depends on ARCH_ENABLE_MEMORY_HOTPLUG depends on 64BIT || BROKEN @@ -178,7 +179,6 @@ config MEMORY_HOTPLUG_DEFAULT_ONLINE config MEMORY_HOTREMOVE bool "Allow for memory hot remove" - select MEMORY_ISOLATION select HAVE_BOOTMEM_INFO_NODE if (X86_64 || PPC64) depends on MEMORY_HOTPLUG && ARCH_ENABLE_MEMORY_HOTREMOVE depends on MIGRATION diff --git a/mm/memory_hotplug.c b/mm/memory_hotplug.c index 1c16a5def781e..35d56cbd3e45b 100644 --- a/mm/memory_hotplug.c +++ b/mm/memory_hotplug.c @@ -805,7 +805,7 @@ int __ref online_pages(unsigned long pfn, unsigned long nr_pages, /* associate pfn range with the zone */ zone = zone_for_pfn_range(online_type, nid, pfn, nr_pages); - move_pfn_range_to_zone(zone, pfn, nr_pages, NULL, MIGRATE_MOVABLE); + move_pfn_range_to_zone(zone, pfn, nr_pages, NULL, MIGRATE_ISOLATE); arg.start_pfn = pfn; arg.nr_pages = nr_pages; @@ -816,6 +816,14 @@ int __ref online_pages(unsigned long pfn, unsigned long nr_pages, if (ret) goto failed_addition; + /* + * Fixup the number of isolated pageblocks before marking the sections + * onlining, such that undo_isolate_page_range() works correctly. + */ + spin_lock_irqsave(&zone->lock, flags); + zone->nr_isolate_pageblock += nr_pages / pageblock_nr_pages; + spin_unlock_irqrestore(&zone->lock, flags); + /* * If this zone is not populated, then it is not in zonelist. * This means the page allocator ignores this zone. @@ -833,21 +841,25 @@ int __ref online_pages(unsigned long pfn, unsigned long nr_pages, zone->zone_pgdat->node_present_pages += nr_pages; pgdat_resize_unlock(zone->zone_pgdat, &flags); + node_states_set_node(nid, &arg); + if (need_zonelists_rebuild) + build_all_zonelists(NULL); + zone_pcp_update(zone); + + /* Basic onlining is complete, allow allocation of onlined pages. */ + undo_isolate_page_range(pfn, pfn + nr_pages, MIGRATE_MOVABLE); + /* * When exposing larger, physically contiguous memory areas to the * buddy, shuffling in the buddy (when freeing onlined pages, putting * them either to the head or the tail of the freelist) is only helpful * for maintaining the shuffle, but not for creating the initial * shuffle. Shuffle the whole zone to make sure the just onlined pages - * are properly distributed across the whole freelist. + * are properly distributed across the whole freelist. Make sure to + * shuffle once pageblocks are no longer isolated. */ shuffle_zone(zone); - node_states_set_node(nid, &arg); - if (need_zonelists_rebuild) - build_all_zonelists(NULL); - zone_pcp_update(zone); - init_per_zone_wmark_min(); kswapd_run(nid); @@ -1550,9 +1562,9 @@ int __ref offline_pages(unsigned long start_pfn, unsigned long nr_pages) pr_info("Offlined Pages %ld\n", nr_pages); /* - * Onlining will reset pagetype flags and makes migrate type - * MOVABLE, so just need to decrease the number of isolated - * pageblocks zone counter here. + * The memory sections are marked offline, and the pageblock flags + * effectively stale; nobody should be touching them. Fixup the number + * of isolated pageblocks, memory onlining will properly revert this. */ spin_lock_irqsave(&zone->lock, flags); zone->nr_isolate_pageblock -= nr_pages / pageblock_nr_pages;
Currently, it can happen that pages are allocated (and freed) via the buddy before we finished basic memory onlining. For example, pages are exposed to the buddy and can be allocated before we actually mark the sections online. Allocated pages could suddenly fail pfn_to_online_page() checks. We had similar issues with pcp handling, when pages are allocated+freed before we reach zone_pcp_update() in online_pages() [1]. Instead, mark all pageblocks MIGRATE_ISOLATE, such that allocations are impossible. Once done with the heavy lifting, use undo_isolate_page_range() to move the pages to the MIGRATE_MOVABLE freelist, marking them ready for allocation. Similar to offline_pages(), we have to manually adjust zone->nr_isolate_pageblock. [1] https://lkml.kernel.org/r/1597150703-19003-1-git-send-email-charante@codeaurora.org Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Michal Hocko <mhocko@suse.com> Cc: Wei Yang <richard.weiyang@linux.alibaba.com> Cc: Baoquan He <bhe@redhat.com> Cc: Pankaj Gupta <pankaj.gupta.linux@gmail.com> Cc: Oscar Salvador <osalvador@suse.de> Cc: Charan Teja Reddy <charante@codeaurora.org> Signed-off-by: David Hildenbrand <david@redhat.com> --- mm/Kconfig | 2 +- mm/memory_hotplug.c | 32 ++++++++++++++++++++++---------- 2 files changed, 23 insertions(+), 11 deletions(-)