svn commit: r326758 - in head/sys/i386: conf include

[John Baldwin]

On 12/11/17 5:26 AM, Eugene Grosbein wrote:
> On 11.12.2017 16:19, Konstantin Belousov wrote:
>> On Mon, Dec 11, 2017 at 04:32:37AM +0000, Conrad Meyer wrote:
>>> Author: cem
>>> Date: Mon Dec 11 04:32:37 2017
>>> New Revision: 326758
>>> URL: https://svnweb.freebsd.org/changeset/base/326758
>>>
>>> Log:
>>>   i386: Bump KSTACK_PAGES default to match amd64
>> i386 is not amd64, the change is wrong.
>>
>> i386 has the word size two times smaller than amd64, which makes typical
>> frame smaller by 30-40% over same code on amd64. Also i386 has much
>> smaller available KVA size (tens of MB) and KVA fragmentation is both
>> more severe and more fatal due to this. I expect that your change will
>> make any non-trivial load which creates enough threads to either fail
>> randomly or deadlock.
>>
>> If somebody tries to fit large load onto i386 machine, he must know what to
>> do and how to configure the kernel to adapt to the load (which does not
>> require the recompilation).
> 
> Its very easy to get kernel stack overflow with 11-STABLE/i386
> without any significant load due to abuse of kernel stack in many kernel subsystems
> as shown in the https://bugs.freebsd.org/bugzilla/show_bug.cgi?id=219476
> 
> Contrary, "enough threads" seems to be very non-trivial number of threads
> and pretty unusual load pattern for i386 as I run several such systems
> with kern.kstack_pages=4 for quite long time and have no problems.
> No random fails, no deadlocks. And with 2 pages only 11-STABLE/i386 is just unusable
> if one utilizes SCTP, IPv6, some WiFi connectivity, IPSEC or even very small ZFS pool.
> 
> I still wonder if there is really such load pattern that creates "enough threads"
> for i386 to make 4-pages stack troublesome.

I suspect two things are at play in running out of stack in 10.x and later.
1) Usage of XSAVE for AVX, etc. state uses more of the kstack (as kib@ alludes
to), and 2) clang.  Certainly for MIPS I have found that compiling with clang
instead of gcc for mips64 gives a kernel that panics for stack overflow for any
use of NFS.  It might be that this is due to something MIPS-specific, but it
might be worthwhile retesting with kstack_pages=2 and building the kernel
with CROSS_TOOLCHAIN=i386-gcc after installing the appropriate package.

For people wanting to use 32-bit binaries on a modern x86 box, we should probably
add support for the x32 ABI permitting you to use 32-bit pointers but with the
full complement of 64-bit registers (and still using the 64-bit kernel so
there is not KVA pressure for the kernel itself).

-- 
John Baldwin