math_state_restore and kernel_fpu_end disable interrupts?

[Nate Eldredge <nate@thatsmathematics.com>]

In trying to track down a bug (see below), I noticed that 
math_state_restore() in arch/x86/kernel/traps.c appears to unconditionally 
disable interrupts when called.  Is this intended behavior or a bug?

The bug in question is triggered by dumping core on an ecryptfs file 
system when aesni-intel is loaded.  (See 
https://bugs.launchpad.net/ubuntu/+source/linux/+bug/1265841 for the 
original report.)  The symptom is that __find_get_block() gets called with 
interrupts disabled, causing a BUG().  I tried to find where interrupts 
were getting disabled and wound up in aes_set_key_common() in 
arch/x86/crypto/aesni-intel_glue.c.  It calls aesni_set_key(), and since 
that uses the FPU, it wraps it in kernel_fpu_begin()/kernel_fpu_end(). 
But kernel_fpu_end() calls math_state_restore() which disables interrupts. 
I've verified that interrupts are still enabled just before the call to 
kernel_fpu_end().

math_state_restore() does:

local_irq_enable();
init_fpu(tsk);
local_irq_disable();

with the result that interrupts are disabled when it finishes, even if 
they were enabled to begin with.  That looks strange to me; are we sure it 
shouldn't just save and restore the interrupt flag?  Or are we not 
supposed to call it with interrupts enabled?

Given the intimidating comment preceding math_state_restore() ("Don't 
touch unless you *really* know how it works"), it's entirely possible I am 
missing something...

Any suggestions appreciated.  Thanks!

-- 
Nate Eldredge
nate@thatsmathematics.com