Re: panic in uma_startup for many-core amd64 system

on 18/10/2010 16:40 Giovanni Trematerra said the following:
> On Mon, Oct 18, 2010 at 8:40 AM, Andriy Gapon <avg_at_freebsd.org> wrote:
>> on 23/09/2010 21:25 Andriy Gapon said the following:
>>>
>>> Jeff,
>>>
>>> just for the kicks I tried to emulate a machine with 64 logical CPUs using
>>> qemu-devel port:
>>> qemu-system-x86_64 -smp sockets=4,cores=8,threads=2 ...
>>>
>>> It seems that FreeBSD agreed to recognize only first 32 CPUs, but it paniced anyway.
>>>
>>> Here's a backtrace:
>>> #34 0xffffffff804fe7f5 in zone_alloc_item (zone=0xffffffff80be1554,
>>> udata=0xffffffff80be1550, flags=1924) at /usr/src/sys/vm/uma_core.c:2506
>>> #35 0xffffffff804ff35d in hash_alloc (hash=0xffffff001ffdb030) at
>>> /usr/src/sys/vm/uma_core.c:483
>>> #36 0xffffffff804ff642 in keg_ctor (mem=Variable "mem" is not available.
>>> ) at /usr/src/sys/vm/uma_core.c:1396
>>> #37 0xffffffff804fe91b in zone_alloc_item (zone=0xffffffff80a1f300,
>>> udata=0xffffffff80be1b60, flags=2) at /usr/src/sys/vm/uma_core.c:2544
>>> #38 0xffffffff804ff92e in zone_ctor (mem=Variable "mem" is not available.
>>> ) at /usr/src/sys/vm/uma_core.c:1832
>>> #39 0xffffffff804ffca4 in uma_startup (bootmem=0xffffff001ffac000, boot_pages=48)
>>> at /usr/src/sys/vm/uma_core.c:1741
>>> #40 0xffffffff80514822 in vm_page_startup (vaddr=18446744071576817664) at
>>> /usr/src/sys/vm/vm_page.c:360
>>> #41 0xffffffff805060c5 in vm_mem_init (dummy=Variable "dummy" is not available.
>>> ) at /usr/src/sys/vm/vm_init.c:118
>>> #42 0xffffffff803258b9 in mi_startup () at /usr/src/sys/kern/init_main.c:253
>>> #43 0xffffffff8017177c in btext () at /usr/src/sys/amd64/amd64/locore.S:81
>>> [[[
>>> Note:
>>> 1. Frame numbers are high because the backtrace is obtained via gdb remotely
>>> connected to qemu and also there is bunch of extra frames from DDB, etc.
>>> 2. Line numbers in uma_core. won't match those in FreeBSD tree, because I've doing
>>> some unrelated hacking in the file.
>>> ]]]
>>>
>>> The problem seems to be with creation of "UMA Zones" zone and keg.
>>> Because of the large number of processors, size argument in the following snippet
>>> is set to a value of 4480:
>>>
>>> args.name = "UMA Zones";
>>> args.size = sizeof(struct uma_zone) +
>>>     (sizeof(struct uma_cache) * (mp_maxid + 1));
>>>
>>> Because of this, keg_ctor() calls keg_large_init():
>>>
>>> else if ((keg->uk_size+UMA_FRITM_SZ) >
>>>     (UMA_SLAB_SIZE - sizeof(struct uma_slab)))
>>>         keg_large_init(keg);
>>> else
>>>         keg_small_init(keg);
>>>
>>> keg_large_init sets UMA_ZONE_OFFPAGE and UMA_ZONE_HASH flags for this keg.
>>> This leads to hash_alloc() being invoked from keg_ctor():
>>>
>>> if (keg->uk_flags & UMA_ZONE_HASH)
>>>         hash_alloc(&keg->uk_hash);
>>>
>>> But the problem is that "UMA Hash" zone is not created yet and thus the call leads
>>> to the panic.  "UMA Hash" zone is the last of system zones created.
>>>
>>> Not sure what the proper fix here could/should be.
>>> Would it work to simply not set UMA_ZONE_HASH flag when UMA_ZFLAG_INTERNAL is set?
>>>
>>>
>>> And some final calculations.
>>> On the test system sizeof(struct uma_cache) is 128 bytes and (mp_maxid + 1) is 32,
>>> so it's already UMA_SLAB_SIZE = PAGE_SIZE = 4096.
>>>
>>
>> Here is a simple solution that seems to work:
>> http://people.freebsd.org/~avg/uma-many-cpus.diff
>> Not sure if it's the best we can do.
>>
> 
> I don't know if it makes sense I only want to raise a flag.
> Is it safe to call kmem_malloc() before bucket_init() during
> uma_startup() to reserve room for CPU caches?

Hmm, not sure what exactly you mean.

> Reading the top uma_int.h comment, it seems that the best way to
> handle this issue
> would be to implement and allow for dynamic slab sizes.

Again, not sure if I follow you, I don't see relation between per-cpu caches and
dynamic slab size.

> I'm also afraid that memory footprint will be larger than now.

Of course, but only by sizeof(pointer) per zone.

-- 
Andriy Gapon