On Thu, Apr 30, 2020 at 04:11:20PM +0200, Joerg Roedel wrote:
> The page-fault handler calls a tracing function which again ends up in
> trace_event_ignore_this_pid(), where it faults again. From here on the CPU is
> in
> a page-fault loop, which continues until the stack overflows (with
> CONFIG_VMAP_STACK).
Did some more testing to find out what this issue has to do with
763802b53a42 x86/mm: split vmalloc_sync_all()
Above commit removes a call to vmalloc_sync_all() from the vmalloc
unmapping path, because that call caused severe performance regressions
on some workloads and was not needed on x86-64 anyway.
But that call caused vmalloc_sync_all() to be called regularily on
x86-64 machines, so that all page-tables were more likely to be in sync.
The call was introduced by commit
3f8fd02b1bf1 mm/vmalloc: Sync unmappings in __purge_vmap_area_lazy()
to fix a correctness issue on x86-32 PAE systems, which also need
unmappings of large pages in the vmalloc area to be synchronized.
This additional call to vmalloc_sync_all() did hide the problem. I
verified it by reverting both of the above commits on v5.7-rc3 and
testing on that kernel. The problem is reproducible there too, the box
hangs hard.
So the underlying problem is that a vmalloc()'ed tracing buffer is used
to trace the page-fault handler, so that it has no chance of faulting in
the buffer address to poking_mm and maybe other PGDs.
The right fix is to call vmalloc_sync_mappings() right after allocating
tracing or perf buffers via v[zm]alloc().
Regards,
Joerg
