[PATCHv3 2/2] powerpc/pseries: update device tree before ejecting hotplug uevents

[Nathan Lynch]

Pingfan Liu <kernelfans at gmail.com> writes:
> A bug is observed on pseries by taking the following steps on rhel:
> -1. drmgr -c mem -r -q 5
> -2. echo c > /proc/sysrq-trigger
>
> And then, the failure looks like:
> kdump: saving to /sysroot//var/crash/127.0.0.1-2020-01-16-02:06:14/
> kdump: saving vmcore-dmesg.txt
> kdump: saving vmcore-dmesg.txt complete
> kdump: saving vmcore
>  Checking for memory holes                         : [  0.0 %] /                   Checking for memory holes                         : [100.0 %] |                   Excluding unnecessary pages                       : [100.0 %] \                   Copying data                                      : [  0.3 %] -          eta: 38s[   44.337636] hash-mmu: mm: Hashing failure ! EA=0x7fffba400000 access=0x8000000000000004 current=makedumpfile
> [   44.337663] hash-mmu:     trap=0x300 vsid=0x13a109c ssize=1 base psize=2 psize 2 pte=0xc000000050000504
> [   44.337677] hash-mmu: mm: Hashing failure ! EA=0x7fffba400000 access=0x8000000000000004 current=makedumpfile
> [   44.337692] hash-mmu:     trap=0x300 vsid=0x13a109c ssize=1 base psize=2 psize 2 pte=0xc000000050000504
> [   44.337708] makedumpfile[469]: unhandled signal 7 at 00007fffba400000 nip 00007fffbbc4d7fc lr 000000011356ca3c code 2
> [   44.338548] Core dump to |/bin/false pipe failed
> /lib/kdump-lib-initramfs.sh: line 98:   469 Bus error               $CORE_COLLECTOR /proc/vmcore $_mp/$KDUMP_PATH/$HOST_IP-$DATEDIR/vmcore-incomplete
> kdump: saving vmcore failed
>
> * Root cause *
>   After analyzing, it turns out that in the current implementation,
> when hot-removing lmb, the KOBJ_REMOVE event ejects before the dt updating as
> the code __remove_memory() comes before drmem_update_dt().
> So in kdump kernel, when read_from_oldmem() resorts to
> pSeries_lpar_hpte_insert() to install hpte, but fails with -2 due to
> non-exist pfn. And finally, low_hash_fault() raise SIGBUS to process, as it
> can be observed "Bus error"
>
> From a viewpoint of listener and publisher, the publisher notifies the
> listener before data is ready.  This introduces a problem where udev
> launches kexec-tools (due to KOBJ_REMOVE) and loads a stale dt before
> updating. And in capture kernel, makedumpfile will access the memory based
> on the stale dt info, and hit a SIGBUS error due to an un-existed lmb.
>
> * Fix *
>   In order to fix this issue, update dt before __remove_memory(), and
> accordingly the same rule in hot-add path.
>
> This will introduce extra dt updating payload for each involved lmb when hotplug.
> But it should be fine since drmem_update_dt() is memory based operation and
> hotplug is not a hot path.

This is great analysis but the performance implications of the change
are grave. The add/remove paths here are already O(n) where n is the
quantity of memory assigned to the LP, this change would make it O(n^2):

dlpar_memory_add_by_count
  for_each_drmem_lmb             <--
    dlpar_add_lmb
      drmem_update_dt(_v1|_v2)
        for_each_drmem_lmb       <--

Memory add/remove isn't a hot path but quadratic runtime complexity
isn't acceptable. Its current performance is bad enough that I have
internal bugs open on it.

Not to mention we leak memory every time drmem_update_dt is called
because we can't safely free device tree properties :-(

Also note that this sort of reverts (fixes?) 063b8b1251fd
("powerpc/pseries/memory-hotplug: Only update DT once per memory DLPAR
request").