Re: [PATCH] powerpc/mm: Limit allocation of SWIOTLB on server machines

[Thiago Jung Bauermann]

Konrad Rzeszutek Wilk <konrad@xxxxxxxxxx> writes:

> On Fri, Jan 08, 2021 at 09:27:01PM -0300, Thiago Jung Bauermann wrote:
>> 
>> Ram Pai <linuxram@xxxxxxxxxx> writes:
>> 
>> > On Wed, Dec 23, 2020 at 09:06:01PM -0300, Thiago Jung Bauermann wrote:
>> >> 
>> >> Hi Ram,
>> >> 
>> >> Thanks for reviewing this patch.
>> >> 
>> >> Ram Pai <linuxram@xxxxxxxxxx> writes:
>> >> 
>> >> > On Fri, Dec 18, 2020 at 03:21:03AM -0300, Thiago Jung Bauermann wrote:
>> >> >> On server-class POWER machines, we don't need the SWIOTLB unless we're a
>> >> >> secure VM. Nevertheless, if CONFIG_SWIOTLB is enabled we unconditionally
>> >> >> allocate it.
>> >> >> 
>> >> >> In most cases this is harmless, but on a few machine configurations (e.g.,
>> >> >> POWER9 powernv systems with 4 GB area reserved for crashdump kernel) it can
>> >> >> happen that memblock can't find a 64 MB chunk of memory for the SWIOTLB and
>> >> >> fails with a scary-looking WARN_ONCE:
>> >> >> 
>> >> >>  ------------[ cut here ]------------
>> >> >>  memblock: bottom-up allocation failed, memory hotremove may be affected
>> >> >>  WARNING: CPU: 0 PID: 0 at mm/memblock.c:332 memblock_find_in_range_node+0x328/0x340
>> >> >>  Modules linked in:
>> >> >>  CPU: 0 PID: 0 Comm: swapper Not tainted 5.10.0-rc2-orig+ #6
>> >> >>  NIP:  c000000000442f38 LR: c000000000442f34 CTR: c0000000001e0080
>> >> >>  REGS: c000000001def900 TRAP: 0700   Not tainted  (5.10.0-rc2-orig+)
>> >> >>  MSR:  9000000002021033 <SF,HV,VEC,ME,IR,DR,RI,LE>  CR: 28022222  XER: 20040000
>> >> >>  CFAR: c00000000014b7b4 IRQMASK: 1
>> >> >>  GPR00: c000000000442f34 c000000001defba0 c000000001deff00 0000000000000047
>> >> >>  GPR04: 00000000ffff7fff c000000001def828 c000000001def820 0000000000000000
>> >> >>  GPR08: 0000001ffc3e0000 c000000001b75478 c000000001b75478 0000000000000001
>> >> >>  GPR12: 0000000000002000 c000000002030000 0000000000000000 0000000000000000
>> >> >>  GPR16: 0000000000000000 0000000000000000 0000000000000000 0000000002030000
>> >> >>  GPR20: 0000000000000000 0000000000010000 0000000000010000 c000000001defc10
>> >> >>  GPR24: c000000001defc08 c000000001c91868 c000000001defc18 c000000001c91890
>> >> >>  GPR28: 0000000000000000 ffffffffffffffff 0000000004000000 00000000ffffffff
>> >> >>  NIP [c000000000442f38] memblock_find_in_range_node+0x328/0x340
>> >> >>  LR [c000000000442f34] memblock_find_in_range_node+0x324/0x340
>> >> >>  Call Trace:
>> >> >>  [c000000001defba0] [c000000000442f34] memblock_find_in_range_node+0x324/0x340 (unreliable)
>> >> >>  [c000000001defc90] [c0000000015ac088] memblock_alloc_range_nid+0xec/0x1b0
>> >> >>  [c000000001defd40] [c0000000015ac1f8] memblock_alloc_internal+0xac/0x110
>> >> >>  [c000000001defda0] [c0000000015ac4d0] memblock_alloc_try_nid+0x94/0xcc
>> >> >>  [c000000001defe30] [c00000000159c3c8] swiotlb_init+0x78/0x104
>> >> >>  [c000000001defea0] [c00000000158378c] mem_init+0x4c/0x98
>> >> >>  [c000000001defec0] [c00000000157457c] start_kernel+0x714/0xac8
>> >> >>  [c000000001deff90] [c00000000000d244] start_here_common+0x1c/0x58
>> >> >>  Instruction dump:
>> >> >>  2c230000 4182ffd4 ea610088 ea810090 4bfffe84 39200001 3d42fff4 3c62ff60
>> >> >>  3863c560 992a8bfc 4bd0881d 60000000 <0fe00000> ea610088 4bfffd94 60000000
>> >> >>  random: get_random_bytes called from __warn+0x128/0x184 with crng_init=0
>> >> >>  ---[ end trace 0000000000000000 ]---
>> >> >>  software IO TLB: Cannot allocate buffer
>> >> >> 
>> >> >> Unless this is a secure VM the message can actually be ignored, because the
>> >> >> SWIOTLB isn't needed. Therefore, let's avoid the SWIOTLB in those cases.
>> >> >
>> >> > The above warn_on is conveying a genuine warning. Should it be silenced?
>> >> 
>> >> Not sure I understand your point. This patch doesn't silence the
>> >> warning, it avoids the problem it is warning about.
>> >
>> > Sorry, I should have explained it better. My point is...  
>> >
>> > 	If CONFIG_SWIOTLB is enabled, it means that the kernel is
>> > 	promising the bounce buffering capability. I know, currently we
>> > 	do not have any kernel subsystems that use bounce buffers on
>> > 	non-secure-pseries-kernel or powernv-kernel.  But that does not
>> > 	mean, there wont be any. In case there is such a third-party
>> > 	module needing bounce buffering, it wont be able to operate,
>> > 	because of the proposed change in your patch.
>> >
>> > 	Is that a good thing or a bad thing, I do not know. I will let
>> > 	the experts opine.
>> 
>> Ping? Does anyone else has an opinion on this? The other option I can
>> think of is changing the crashkernel code to not reserve so much memory
>> below 4 GB. Other people are considering this option, but it's not
>> planned for the near future.
>
> That seems a more suitable solution regardless, but there is always
> the danger of not being enough or being too big.
>
> There was some autocrashkernel allocation patches going around
> for x86 and ARM that perhaps could be re-used?

Do you mean this?

https://lore.kernel.org/lkml/20201118232431.21832-1-saeed.mirzamohammadi@xxxxxxxxxx/

I agree it would be great to have crashkernel=auto upstream.

>> Also, there's a patch currently in linux-next which removes the scary
>> warning because of unrelated reasons:
>> 
>> https://lore.kernel.org/lkml/20201217201214.3414100-2-guro@xxxxxx
>> 
>> So assuming that the patch above goes in and keeping the assumption that
>> the swiotlb won't be needed in the powernv machines where I've seen the
>> warning happen, we can just leave things as they are now.
>
> If that solves the problem, then that is OK.

Awesome! Today Mike Rapoport mentioned that the patch above uncovers a
few issues with late memory reservations on on x86 that he is fixing:

https://lore.kernel.org/lkml/20210115083255.12744-1-rppt@xxxxxxxxxx/

So that would be the last piece of the puzzle.

-- 
Thiago Jung Bauermann
IBM Linux Technology Center