summaryrefslogtreecommitdiff
path: root/mm
diff options
context:
space:
mode:
authorBen Widawsky <ben.widawsky@intel.com>2020-06-25 20:30:51 -0700
committerLinus Torvalds <torvalds@linux-foundation.org>2020-06-26 00:27:38 -0700
commitb7e3debdd0408c0dca5d4750371afa5003f792dc (patch)
treee578d0411704cb82aede9b19210ebeeb35432075 /mm
parent7a0e27b2a0ce2735e27e21ebc8b777550fe0ed81 (diff)
downloadlwn-b7e3debdd0408c0dca5d4750371afa5003f792dc.tar.gz
lwn-b7e3debdd0408c0dca5d4750371afa5003f792dc.zip
mm/memory_hotplug.c: fix false softlockup during pfn range removal
When working with very large nodes, poisoning the struct pages (for which there will be very many) can take a very long time. If the system is using voluntary preemptions, the software watchdog will not be able to detect forward progress. This patch addresses this issue by offering to give up time like __remove_pages() does. This behavior was introduced in v5.6 with: commit d33695b16a9f ("mm/memory_hotplug: poison memmap in remove_pfn_range_from_zone()") Alternately, init_page_poison could do this cond_resched(), but it seems to me that the caller of init_page_poison() is what actually knows whether or not it should relax its own priority. Based on Dan's notes, I think this is perfectly safe: commit f931ab479dd2 ("mm: fix devm_memremap_pages crash, use mem_hotplug_{begin, done}") Aside from fixing the lockup, it is also a friendlier thing to do on lower core systems that might wipe out large chunks of hotplug memory (probably not a very common case). Fixes this kind of splat: watchdog: BUG: soft lockup - CPU#46 stuck for 22s! [daxctl:9922] irq event stamp: 138450 hardirqs last enabled at (138449): [<ffffffffa1001f26>] trace_hardirqs_on_thunk+0x1a/0x1c hardirqs last disabled at (138450): [<ffffffffa1001f42>] trace_hardirqs_off_thunk+0x1a/0x1c softirqs last enabled at (138448): [<ffffffffa1e00347>] __do_softirq+0x347/0x456 softirqs last disabled at (138443): [<ffffffffa10c416d>] irq_exit+0x7d/0xb0 CPU: 46 PID: 9922 Comm: daxctl Not tainted 5.7.0-BEN-14238-g373c6049b336 #30 Hardware name: Intel Corporation PURLEY/PURLEY, BIOS PLYXCRB1.86B.0578.D07.1902280810 02/28/2019 RIP: 0010:memset_erms+0x9/0x10 Code: c1 e9 03 40 0f b6 f6 48 b8 01 01 01 01 01 01 01 01 48 0f af c6 f3 48 ab 89 d1 f3 aa 4c 89 c8 c3 90 49 89 f9 40 88 f0 48 89 d1 <f3> aa 4c 89 c8 c3 90 49 89 fa 40 0f b6 ce 48 b8 01 01 01 01 01 01 Call Trace: remove_pfn_range_from_zone+0x3a/0x380 memunmap_pages+0x17f/0x280 release_nodes+0x22a/0x260 __device_release_driver+0x172/0x220 device_driver_detach+0x3e/0xa0 unbind_store+0x113/0x130 kernfs_fop_write+0xdc/0x1c0 vfs_write+0xde/0x1d0 ksys_write+0x58/0xd0 do_syscall_64+0x5a/0x120 entry_SYSCALL_64_after_hwframe+0x49/0xb3 Built 2 zonelists, mobility grouping on. Total pages: 49050381 Policy zone: Normal Built 3 zonelists, mobility grouping on. Total pages: 49312525 Policy zone: Normal David said: "It really only is an issue for devmem. Ordinary hotplugged system memory is not affected (onlined/offlined in memory block granularity)." Link: http://lkml.kernel.org/r/20200619231213.1160351-1-ben.widawsky@intel.com Fixes: commit d33695b16a9f ("mm/memory_hotplug: poison memmap in remove_pfn_range_from_zone()") Signed-off-by: Ben Widawsky <ben.widawsky@intel.com> Reported-by: "Scargall, Steve" <steve.scargall@intel.com> Reported-by: Ben Widawsky <ben.widawsky@intel.com> Acked-by: David Hildenbrand <david@redhat.com> Cc: Dan Williams <dan.j.williams@intel.com> Cc: Vishal Verma <vishal.l.verma@intel.com> Cc: <stable@vger.kernel.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Diffstat (limited to 'mm')
-rw-r--r--mm/memory_hotplug.c13
1 files changed, 11 insertions, 2 deletions
diff --git a/mm/memory_hotplug.c b/mm/memory_hotplug.c
index 9b34e03e730a..da374cd3d45b 100644
--- a/mm/memory_hotplug.c
+++ b/mm/memory_hotplug.c
@@ -471,11 +471,20 @@ void __ref remove_pfn_range_from_zone(struct zone *zone,
unsigned long start_pfn,
unsigned long nr_pages)
{
+ const unsigned long end_pfn = start_pfn + nr_pages;
struct pglist_data *pgdat = zone->zone_pgdat;
- unsigned long flags;
+ unsigned long pfn, cur_nr_pages, flags;
/* Poison struct pages because they are now uninitialized again. */
- page_init_poison(pfn_to_page(start_pfn), sizeof(struct page) * nr_pages);
+ for (pfn = start_pfn; pfn < end_pfn; pfn += cur_nr_pages) {
+ cond_resched();
+
+ /* Select all remaining pages up to the next section boundary */
+ cur_nr_pages =
+ min(end_pfn - pfn, SECTION_ALIGN_UP(pfn + 1) - pfn);
+ page_init_poison(pfn_to_page(pfn),
+ sizeof(struct page) * cur_nr_pages);
+ }
#ifdef CONFIG_ZONE_DEVICE
/*