On 2020-09-17 17:54:10 [+0200], pet...@infradead.org wrote:
> I'm not sure what the problem with FPU was, I was throwing alternatives
> at tglx to see what would stick, in part to (re)discover the design
> constraints of this thing.
was this recent or distant in the time line?
> One reason for not allowing migrate_disable() to sleep was: FPU code.
>
> Could it be it does something like:
>
> preempt_disable();
> spin_lock();
>
> spin_unlock();
> preempt_enable();
>
> Where we'll never get preempted while migrate_disable()'d and thus never
> trigger any of the sleep paths?
I try to get rid of something like that. This doesn't work either way
because the spin_lock() may block which it can't with disabled
preemption.
Looking at my queue, FPU related is crypto. And here we break the loops
mostly due to the construct:
kernel_fpu_begin();
while (bytes)
crypto_thingy();
skcipher_walk_done()
and skcipher_walk_done() could allocate/free/map memory. This is
independent.
Ah. We used to have migrate_disable() in pagefault_disable(). The x86
FPU code does
preempt_disable();
…
pagefault_disable();
but that migrate_disable() was moved from pagefault_disable() to
kmap_atomic(). We shouldn't have
preempt_disable(); || local_irq_disable();
kmap_atomic();
on RT. I've been running around removing those. See
a10dcebacdb0c ("fs/ntfs/aops.c: don't disable interrupts during
kmap_atomic()")
ce1e518190ea7 ("ide: don't disable interrupts during kmap_atomic()")
f3a1075e5fc34 ("block: don't disable interrupts during kmap_atomic()")
Sebastian
