Re: CHANGE_PV_LIST_LOCK_TO_PHYS is not correct when !NUMA ?

From: Konstantin Belousov <kostikbel_at_gmail.com>
Date: Sat, 9 May 2020 23:47:43 +0300
On Sat, May 09, 2020 at 11:33:40PM +0300, Andriy Gapon wrote:
> On 09/05/2020 19:50, Konstantin Belousov wrote:
> > On Sat, May 09, 2020 at 07:16:27PM +0300, Andriy Gapon wrote:
> >> On 09/05/2020 19:13, Konstantin Belousov wrote:
> >>> On Sat, May 09, 2020 at 06:52:24PM +0300, Andriy Gapon wrote:
> >>>> On 08/05/2020 19:15, Konstantin Belousov wrote:
> >>>>> On Fri, May 08, 2020 at 06:53:24PM +0300, Andriy Gapon wrote:
> >>>>>>
> >>>>>> I have a reproducible panic with a custom kernel without option NUMA while using
> >>>>>> amdgpu driver from linuxkpi-based drm:
> >>>>>>
> >>>>>> panic: address 41ec00000 beyond the last segment
> >>>>>>
> >>>>>> I did some quick debugging and the panic happens when Xorg server tries to
> >>>>>> access a frame buffer (or something like that).  There is a page fault that gets
> >>>>>> satisfied by ttm with a fictitious page.
> >>>>>>
> >>>>>> The stack trace is:
> >>>>>> #11 0xffffffff808031a3 in panic (fmt=0xffffffff8119a998 <cnputs_mtx>
> >>>>>> "5\003ʀ\377\377\377\377") at /usr/devel/git/motil/sys/kern/kern_shutdown.c:839
> >>>>>> #12 0xffffffff80bbc552 in pmap_enter (pmap=<optimized out>, va=34504441856,
> >>>>>> m=<optimized out>, prot=<optimized out>, flags=<optimized out>, psind=<optimized
> >>>>>> out>) at /usr/devel/git/motil/sys/amd64/amd64/pmap.c:6035
> >>>>>> #13 0xffffffff80b288be in vm_fault_populate (fs=<optimized out>) at
> >>>>>> /usr/devel/git/motil/sys/vm/vm_fault.c:519
> >>>>>> #14 vm_fault_allocate (fs=<optimized out>) at
> >>>>>> /usr/devel/git/motil/sys/vm/vm_fault.c:1032
> >>>>>> #15 vm_fault (map=<optimized out>, vaddr=<optimized out>, fault_type=<optimized
> >>>>>> out>, fault_flags=<optimized out>, m_hold=<optimized out>) at
> >>>>>> /usr/devel/git/motil/sys/vm/vm_fault.c:1342
> >>>>>> #16 0xffffffff80b26e7e in vm_fault_trap (map=0xfffffe0017cd39e8,
> >>>>>> vaddr=<optimized out>, fault_type=<optimized out>, fault_flags=0,
> >>>>>> signo=0xfffffe00a810dbc4, ucode=0xfffffe00a810dbc0) at
> >>>>>> /usr/devel/git/motil/sys/vm/vm_fault.c:589
> >>>>>> #17 0xffffffff80bcf89c in trap_pfault (frame=0xfffffe00a810dc00,
> >>>>>> usermode=<optimized out>, signo=<optimized out>, ucode=0xffffffff80853250
> >>>>>> <putchar>) at /usr/devel/git/motil/sys/amd64/amd64/trap.c:821
> >>>>>> #18 0xffffffff80bceeec in trap (frame=0xfffffe00a810dc00) at
> >>>>>> /usr/devel/git/motil/sys/amd64/amd64/trap.c:34
> >>>>>>
> >>>>>>
> >>>>>> The line number in pmap_enter() is incorrect, I guess because of optimizations.
> >>>>>> The assert seems to be reached via pmap_enter -> CHANGE_PV_LIST_LOCK_TO_PHYS ->
> >>>>>> PHYS_TO_PV_LIST_LOCK -> pa_index().
> >>>>>>
> >>>>>> The panic in correct in that the page is fictitious and its physical address is
> >>>>>> beyond the end of real physical memory.
> >>>>>> It seems that NUMA PHYS_TO_PV_LIST_LOCK() is aware of such pages, but !NUMA one
> >>>>>> is not.
> >>>>>
> >>>>> I think you can remove this assert.  pa_index() is always taken by
> >>>>> % NVP_LIST_LOCKS, because fictitious mappings are not promoted.
> >>>>>
> >>>>> Try that and commit if it works for you.
> >>>>
> >>>> I tried this change:
> >>>> diff --git a/sys/amd64/amd64/pmap.c b/sys/amd64/amd64/pmap.c
> >>>> index 4deed86a76d1a..b834b7f0388b7 100644
> >>>> --- a/sys/amd64/amd64/pmap.c
> >>>> +++ b/sys/amd64/amd64/pmap.c
> >>>> _at__at_ -345,7 +345,7 _at__at_ pmap_pku_mask_bit(pmap_t pmap)
> >>>>  #define	NPV_LIST_LOCKS	MAXCPU
> >>>>
> >>>>  #define	PHYS_TO_PV_LIST_LOCK(pa)	\
> >>>> -			(&pv_list_locks[pa_index(pa) % NPV_LIST_LOCKS])
> >>>> +			(&pv_list_locks[((pa) >> PDRSHIFT) % NPV_LIST_LOCKS])
> >>>>  #endif
> >>>>
> >>>>  #define	CHANGE_PV_LIST_LOCK_TO_PHYS(lockp, pa)	do {	\
> >>>>
> >>>> It fixed the original problem, but I got a new panic.
> >>>> "DI already started" in pmap_remove() -> pmap_delayed_invl_start_u().
> >>>> I guess that !NUMA variant does not get much testing, so I'll probably just
> >>>> stick with the default.
> >>> Why didn't you just removed the KASSERT from pa_index ?
> >>
> >> Well, I thought it might be useful in the NUMA case.
> >> pa_index() definition is shared between both cases.
> > Might be define the macro two times, for NUMA/non-NUMA.  non-NUMA case
> > does not need the assert, because users take it mod NPV_LIST_LOCKS.
> > 
> 
> I still don't see how that could help with "DI already started" panic.

Might be not, might be it would help due to pmap_delayed_invl_genp().
But I would more worry about this 'already started' issue, because
this must not happen.  Can you remove the assert from the macro and
provide backtrace of 'DI already started' panic ?
Received on Sat May 09 2020 - 18:47:51 UTC

This archive was generated by hypermail 2.4.0 : Wed May 19 2021 - 11:41:24 UTC