[SRU][X/Y/Z/A][PATCH 0/5] Fixes for LP:#1680513

[Gavin Guo]

BugLink: http://bugs.launchpad.net/bugs/1680513

[Impact]
After numad is enabled and there are several VMs running on the same
host machine(host kernel version: 4.4.0-72-generic #93), the
softlockup messages can be observed inside the VMs' dmesg.

First, the crashdump was captured when the symptom was observed. At
the first glance, it looks like an IPI lost issue. The numad process
initiates a migration of memory, and as part of this, needs to flush
the TLB cache of another CPU. When the crash dump was taken, that
other CPU has the TLB flush pending, but not executed.

The numad kernel task is holding a semaphore lock mmap_sem(for the
VM's memory) to do the migration, and the tasks that actually end up
being blocked are other virtual CPUs for the same VM. These tasks need
to access or make changes to the memory map for the VM because of the
VM page fault, but cannot acquire the semaphore lock.

However, the original thoughts on the root cause (unhandled IPI or csd
lock issue) are incorrect.

We originally suspected an issue with a lost IPI (inter processor
interrupt) that performs remote CPU cache flushes during page
migration, or a known issue with the "csd" lock used to synchronize
the remote CPU cache flush. A lost IPI would be a function of the
system firmware or chipset (it is not a CPU issue), but the known csd
issue is hardware independent.

Gavin created the hotfix kernel with changes in the csd_lock_wait
function that would time out if the unlock never happens (the end
result of either cause), and print messages to the console when that
timeout occurred. The messages look like:

csd_lock_wait called %d times

csd: Detected non-responsive CSD lock (#%d) on CPU#%02d, waiting %Ld.%03Ld secs for CPU#%02d

However, the VMs are still experiencing the hangs, but the
csd_lock_wait timeout is not happening. This suggests that the csd
lock / lost IPI is not the actual cause.

In the crash dump, the numad task has induced a migration, and the
stack is as follows:

#1 [ffff885f8fb4fb78] smp_call_function_many
#2 [ffff885f8fb4fbc0] native_flush_tlb_others
#3 [ffff885f8fb4fc08] flush_tlb_page
#4 [ffff885f8fb4fc30] ptep_clear_flush
#5 [ffff885f8fb4fc60] try_to_unmap_one
#6 [ffff885f8fb4fcd0] rmap_walk_ksm
#7 [ffff885f8fb4fd28] rmap_walk
#8 [ffff885f8fb4fd80] try_to_unmap
#9 [ffff885f8fb4fdc8] migrate_pages
#10 [ffff885f8fb4fe80] do_migrate_pages

The frame #1 is actually in the csd_lock_wait function mentioned
above, but the compiler has optimized that call and it does not appear
in the stack.

What happens here is that do_migrate_pages (frame #10) acquires the
semaphore that everything else is waiting for (and that eventually
produce the hang warnings), and it holds that semaphore for the
duration of the page migration. This strongly suggests that this
single do_migrate_pages call is taking in excess of 10 seconds, and if
the csd lock is not stuck, then something else within its call path is
not functioning correctly.

We originally suspected that the lost IPI/csd lock hang was
responsible for the hung task timeouts, but in the absence of the csd
warning messages, the cause presumably lies elsewhere.

A KSM function appears in frame #6; this is the function that will
search out the merged pages to handle them for the migration.

Gavin have tried to disassemble the code and finally find the
stable_node->hlist is as long as 2306920 entries:

rmap_item list(stable_node->hlist):
stable_node: 0xffff881f836ba000 stable_node->hlist->first = 0xffff883f3e5746b0

struct hlist_head {
[0] struct hlist_node *first;
}
struct hlist_node {
[0] struct hlist_node *next;
[8] struct hlist_node **pprev;
}

crash> list hlist_node.next 0xffff883f3e5746b0 > rmap_item.lst

$ wc -l rmap_item.lst
2306920 rmap_item.lst

This is roughly 9 GB of pages. The theory is that KSM has merged a
very large number of pages that are empty (the value of all locations
in the page are zero).

The bug can be observed by the perf flame graph[1]:

[1]. http://kernel.ubuntu.com/~gavinguo/sf00131845/numa-131845.svg

[Fix]
Andrea Arcangeli already sent out the patch[2] in the 2015/11/10.
Andrew Morton also said he will apply the patch. However, the patch
finally disappears from the mmtom tree in April 2016. Andrea suggested
apply the 3 patches[3].

[2]. [PATCH 1/1] ksm: introduce ksm_max_page_sharing per page
deduplication limit
http://www.spinics.net/lists/linux-mm/msg96866.html

[3]. Re: [PATCH 1/1] ksm: introduce ksm_max_page_sharing per page
deduplication limit
https://www.spinics.net/lists/linux-mm/msg113829.html

[Test Case]
The patches has been tested with 9 VMs and each has 32GB ram and 16
VCPUs. Numad/KSM are also enabled in the machine. After running for
6 days, the system is stable and unstable CPU loading cannot be
observed inside the virtual appliances monitor[4]. The numad cpu
utilization rate is normal and guest hang also cannot be observed.

Machine type: Dell PowerEdge R920
Memory: 528GB with 4 NUMA nodes
CPU: 120 cores

[4]. http://kernel.ubuntu.com/~gavinguo/sf00131845/virtual_appliances_loading.png

Andrea Arcangeli (5):
  ksm: introduce ksm_max_page_sharing per page deduplication limit
  ksm: fix use after free with merge_across_nodes = 0
  ksm: cleanup stable_node chain collapse case
  ksm: swap the two output parameters of chain/chain_prune
  ksm: optimize refile of stable_node_dup at the head of the chain

 Documentation/vm/ksm.txt |  63 ++++
 mm/ksm.c                 | 820 +++++++++++++++++++++++++++++++++++++++++++----
 2 files changed, 817 insertions(+), 66 deletions(-)

-- 
2.7.4