Currently, we allow kernel mode NEON in softirq or hardirq context by
stacking and unstacking a slice of the NEON register file for each call
to kernel_neon_begin() and kernel_neon_end(), respectively.
Given that
a) a CPU typically spends most of its time in userland, during which time
no kernel mode NEON in process context is in progress,
b) a CPU spends most of its time in the kernel doing other things than
kernel mode NEON when it gets interrupted to perform kernel mode NEON
in softirq context
the stacking and subsequent unstacking is only necessary if we are
interrupting a thread while it is performing kernel mode NEON in process
context, which means that in all other cases, we can simply preserve the
userland FPSIMD state once, and only restore it upon return to userland,
even if we are being invoked from softirq or hardirq context.
So instead of checking whether we are running in interrupt context, keep
track of the level of nested kernel mode NEON calls in progress, and only
perform the eager stack/unstack if the level exceeds 1.
Signed-off-by: Ard Biesheuvel <redacted>
---
v6:
- use a spinlock instead of disabling interrupts
v5:
- perform the test-and-set and the fpsimd_save_state with interrupts disabled,
to prevent nested kernel_neon_begin()/_end() pairs to clobber the state
while it is being preserved
v4:
- use this_cpu_inc/dec, which give sufficient guarantees regarding
concurrency, but do not imply SMP barriers, which are not needed here
v3:
- avoid corruption by concurrent invocations of kernel_neon_begin()/_end()
v2:
- BUG() on unexpected values of the nesting level
- relax the BUG() on num_regs>32 to a WARN, given that nothing actually
breaks in that case
arch/arm64/include/asm/fpsimd.h | 3 +
arch/arm64/kernel/fpsimd.c | 77 ++++++++++++++------
2 files changed, 58 insertions(+), 22 deletions(-)
@@ -37,6 +38,8 @@ struct fpsimd_state {u32fpcr;};};+/* lock protecting the preserved state while it is being saved */+spinlock_tlock;/* the id of the last cpu to have restored this state */unsignedintcpu;};
On Tue, Dec 13, 2016 at 12:35:29PM +0000, Ard Biesheuvel wrote:
Currently, we allow kernel mode NEON in softirq or hardirq context by
stacking and unstacking a slice of the NEON register file for each call
to kernel_neon_begin() and kernel_neon_end(), respectively.
Given that
a) a CPU typically spends most of its time in userland, during which time
no kernel mode NEON in process context is in progress,
b) a CPU spends most of its time in the kernel doing other things than
kernel mode NEON when it gets interrupted to perform kernel mode NEON
in softirq context
the stacking and subsequent unstacking is only necessary if we are
interrupting a thread while it is performing kernel mode NEON in process
context, which means that in all other cases, we can simply preserve the
userland FPSIMD state once, and only restore it upon return to userland,
even if we are being invoked from softirq or hardirq context.
So instead of checking whether we are running in interrupt context, keep
track of the level of nested kernel mode NEON calls in progress, and only
perform the eager stack/unstack if the level exceeds 1.
Hang on ... we could be in the middle of fpsimd_save_state() for other
reasons, say on the context switch path. Wouldn't we need to take the
lock for those too?
Also, a spinlock can't protect a critical section from code running on
the same CPU... wouldn't it just deadlock in that case?
I still tend to prefer v4 -- there we do a redundant double-save only if
one kernel_neon_begin() interrupts the actual task context save.
Cheers
---Dave
quoted hunk
Signed-off-by: Ard Biesheuvel <redacted>
---
v6:
- use a spinlock instead of disabling interrupts
v5:
- perform the test-and-set and the fpsimd_save_state with interrupts disabled,
to prevent nested kernel_neon_begin()/_end() pairs to clobber the state
while it is being preserved
v4:
- use this_cpu_inc/dec, which give sufficient guarantees regarding
concurrency, but do not imply SMP barriers, which are not needed here
v3:
- avoid corruption by concurrent invocations of kernel_neon_begin()/_end()
v2:
- BUG() on unexpected values of the nesting level
- relax the BUG() on num_regs>32 to a WARN, given that nothing actually
breaks in that case
arch/arm64/include/asm/fpsimd.h | 3 +
arch/arm64/kernel/fpsimd.c | 77 ++++++++++++++------
2 files changed, 58 insertions(+), 22 deletions(-)
@@ -37,6 +38,8 @@ struct fpsimd_state {u32fpcr;};};+/* lock protecting the preserved state while it is being saved */+spinlock_tlock;/* the id of the last cpu to have restored this state */unsignedintcpu;};
--
2.7.4
_______________________________________________
linux-arm-kernel mailing list
linux-arm-kernel at lists.infradead.org
http://lists.infradead.org/mailman/listinfo/linux-arm-kernel
On 13 December 2016 at 13:49, Dave Martin [off-list ref] wrote:
On Tue, Dec 13, 2016 at 12:35:29PM +0000, Ard Biesheuvel wrote:
quoted
Currently, we allow kernel mode NEON in softirq or hardirq context by
stacking and unstacking a slice of the NEON register file for each call
to kernel_neon_begin() and kernel_neon_end(), respectively.
Given that
a) a CPU typically spends most of its time in userland, during which time
no kernel mode NEON in process context is in progress,
b) a CPU spends most of its time in the kernel doing other things than
kernel mode NEON when it gets interrupted to perform kernel mode NEON
in softirq context
the stacking and subsequent unstacking is only necessary if we are
interrupting a thread while it is performing kernel mode NEON in process
context, which means that in all other cases, we can simply preserve the
userland FPSIMD state once, and only restore it upon return to userland,
even if we are being invoked from softirq or hardirq context.
So instead of checking whether we are running in interrupt context, keep
track of the level of nested kernel mode NEON calls in progress, and only
perform the eager stack/unstack if the level exceeds 1.
Hang on ... we could be in the middle of fpsimd_save_state() for other
reasons, say on the context switch path. Wouldn't we need to take the
lock for those too?
Yes. Saving the state should never be interrupted, and now that I
think of it, this is an issue for your original patch as well: even if
you always preserve the state first in kernel_neon_begin(), any
occurrence of fpsimd_save_state() could be interrupted by a
kernel_neon_begin()/_end() pair, after which the SVE state is nuked
anyway (regardless of whether a patch like this one is applied) So I
think we do need the lock in all cases where the FP/SIMD is copied to
memory and the flag set.
Also, a spinlock can't protect a critical section from code running on
the same CPU... wouldn't it just deadlock in that case?
That is why I use spin_trylock(). But I realise now that the patch is
incorrect: the nesting level needs to be incremented first, or the
interrupter will still clobber the context while it is being
preserved.
I still tend to prefer v4 -- there we do a redundant double-save only if
one kernel_neon_begin() interrupts the actual task context save.
I think we're not at the bottom of this rabbit hole yet:-)
On Tue, Dec 13, 2016 at 02:17:14PM +0000, Ard Biesheuvel wrote:
On 13 December 2016 at 13:49, Dave Martin [off-list ref] wrote:
quoted
On Tue, Dec 13, 2016 at 12:35:29PM +0000, Ard Biesheuvel wrote:
quoted
Currently, we allow kernel mode NEON in softirq or hardirq context by
stacking and unstacking a slice of the NEON register file for each call
to kernel_neon_begin() and kernel_neon_end(), respectively.
Given that
a) a CPU typically spends most of its time in userland, during which time
no kernel mode NEON in process context is in progress,
b) a CPU spends most of its time in the kernel doing other things than
kernel mode NEON when it gets interrupted to perform kernel mode NEON
in softirq context
the stacking and subsequent unstacking is only necessary if we are
interrupting a thread while it is performing kernel mode NEON in process
context, which means that in all other cases, we can simply preserve the
userland FPSIMD state once, and only restore it upon return to userland,
even if we are being invoked from softirq or hardirq context.
So instead of checking whether we are running in interrupt context, keep
track of the level of nested kernel mode NEON calls in progress, and only
perform the eager stack/unstack if the level exceeds 1.
Hang on ... we could be in the middle of fpsimd_save_state() for other
reasons, say on the context switch path. Wouldn't we need to take the
lock for those too?
Yes. Saving the state should never be interrupted, and now that I
think of it, this is an issue for your original patch as well: even if
you always preserve the state first in kernel_neon_begin(), any
occurrence of fpsimd_save_state() could be interrupted by a
kernel_neon_begin()/_end() pair, after which the SVE state is nuked
anyway (regardless of whether a patch like this one is applied) So I
think we do need the lock in all cases where the FP/SIMD is copied to
memory and the flag set.
Yes, that's true. I didn't really trust my hack that much yet (for
development I was typically just putting a BUG_ON() on
kernel_neon_begin() to make sure it's not being called...)
quoted
Also, a spinlock can't protect a critical section from code running on
the same CPU... wouldn't it just deadlock in that case?
That is why I use spin_trylock(). But I realise now that the patch is
incorrect: the nesting level needs to be incremented first, or the
Ah, I missed the significance of that.
interrupter will still clobber the context while it is being
preserved.
Hmmm, yes.
quoted
I still tend to prefer v4 -- there we do a redundant double-save only if
one kernel_neon_begin() interrupts the actual task context save.
I think we're not at the bottom of this rabbit hole yet:-)
Seemingly not...
I think it's better to ignore SVE to begin with and focus on what
changes are desirable for the NEON case by itself. That's likely to
give cleaner code which can then be extended for SVE as appropriate.
For SVE, the basic change is that if task state saving isn't finished,
then we need to save (partial) SVE state in a nested
kernel_neon_begin(), not just NEON state. I'm hoping we can detect
this case as TIF_SVE && !TIF_FOREIGN_FPSTATE, if we defer setting
TIF_FOREIGN_FPSTATE until after the task state save is complete ...
unless there's some other reason this won't work.
Cheers
---Dave
On Tue, Dec 13, 2016 at 12:35:29PM +0000, Ard Biesheuvel wrote:
Currently, we allow kernel mode NEON in softirq or hardirq context by
stacking and unstacking a slice of the NEON register file for each call
to kernel_neon_begin() and kernel_neon_end(), respectively.
Given that
a) a CPU typically spends most of its time in userland, during which time
no kernel mode NEON in process context is in progress,
b) a CPU spends most of its time in the kernel doing other things than
kernel mode NEON when it gets interrupted to perform kernel mode NEON
in softirq context
the stacking and subsequent unstacking is only necessary if we are
interrupting a thread while it is performing kernel mode NEON in process
context, which means that in all other cases, we can simply preserve the
userland FPSIMD state once, and only restore it upon return to userland,
even if we are being invoked from softirq or hardirq context.
So instead of checking whether we are running in interrupt context, keep
track of the level of nested kernel mode NEON calls in progress, and only
perform the eager stack/unstack if the level exceeds 1.
Signed-off-by: Ard Biesheuvel <redacted>
---
v6:
- use a spinlock instead of disabling interrupts
My concern with the spinlock or disabling interrupts is the latency if
we ever get some insanely long SVE vectors.
--
Catalin
On 13 December 2016 at 17:34, Catalin Marinas [off-list ref] wrote:
On Tue, Dec 13, 2016 at 12:35:29PM +0000, Ard Biesheuvel wrote:
quoted
Currently, we allow kernel mode NEON in softirq or hardirq context by
stacking and unstacking a slice of the NEON register file for each call
to kernel_neon_begin() and kernel_neon_end(), respectively.
Given that
a) a CPU typically spends most of its time in userland, during which time
no kernel mode NEON in process context is in progress,
b) a CPU spends most of its time in the kernel doing other things than
kernel mode NEON when it gets interrupted to perform kernel mode NEON
in softirq context
the stacking and subsequent unstacking is only necessary if we are
interrupting a thread while it is performing kernel mode NEON in process
context, which means that in all other cases, we can simply preserve the
userland FPSIMD state once, and only restore it upon return to userland,
even if we are being invoked from softirq or hardirq context.
So instead of checking whether we are running in interrupt context, keep
track of the level of nested kernel mode NEON calls in progress, and only
perform the eager stack/unstack if the level exceeds 1.
Signed-off-by: Ard Biesheuvel <redacted>
---
v6:
- use a spinlock instead of disabling interrupts
My concern with the spinlock or disabling interrupts is the latency if
we ever get some insanely long SVE vectors.
Thinking about this a bit more, and trying a couple of things in code,
I think this looks like it could be a nasty problem.
The primary issue is that *any* code that handles the FP/SIMD state,
i.e., preserve it, restore it, etc, could be interrupted, and if
kernel_neon_begin()/_end() was used during that time, the top SVE end
of the registers is just blown away if we don't make the nested
save/restore SVE aware.
For instance, looking at
void fpsimd_restore_current_state(void)
{
preempt_disable();
if (test_and_clear_thread_flag(TIF_FOREIGN_FPSTATE)) {
struct fpsimd_state *st = ¤t->thread.fpsimd_state;
fpsimd_load_state(st);
this_cpu_write(fpsimd_last_state, st);
st->cpu = smp_processor_id();
}
preempt_enable();
}
if fpsimd_load_state() is interrupted and the NEON is used, the
registers that were restored before the interruption will have
incorrect values if SVE is being used by userland.
On the preserve side, we can deal with this by preserving into a temp
buffer, and only store the value that was recorded when the flag was
set, i.e.,
static void safe_preserve_fpsimd_state(void)
{
union __fpsimd_percpu_state *s = this_cpu_ptr(nested_fpsimdstate);
if (in_irq()) {
/* No interruptions possible, so just proceed */
if (!test_and_set_thread_flag(TIF_FOREIGN_FPSTATE))
fpsimd_save_state(¤t->thread.fpsimd_state);
} else {
struct fpsimd_state *fp;
fp = in_interrupt() ? &s[0].full : &s[1].full;
fpsimd_save_state(fp);
if (!test_and_set_thread_flag(TIF_FOREIGN_FPSTATE))
current->thread.fpsimd_state = *fp;
}
}
which is already cumbersome if the full FP/SIMD state grows to 64 KB
However, on the restore side, I fail to see how we can tolerate
interruptions in a similar way.
--
Ard.
On 13 December 2016 at 19:17, Ard Biesheuvel [off-list ref] wrote:
On 13 December 2016 at 17:34, Catalin Marinas [off-list ref] wrote:
quoted
On Tue, Dec 13, 2016 at 12:35:29PM +0000, Ard Biesheuvel wrote:
quoted
Currently, we allow kernel mode NEON in softirq or hardirq context by
stacking and unstacking a slice of the NEON register file for each call
to kernel_neon_begin() and kernel_neon_end(), respectively.
Given that
a) a CPU typically spends most of its time in userland, during which time
no kernel mode NEON in process context is in progress,
b) a CPU spends most of its time in the kernel doing other things than
kernel mode NEON when it gets interrupted to perform kernel mode NEON
in softirq context
the stacking and subsequent unstacking is only necessary if we are
interrupting a thread while it is performing kernel mode NEON in process
context, which means that in all other cases, we can simply preserve the
userland FPSIMD state once, and only restore it upon return to userland,
even if we are being invoked from softirq or hardirq context.
So instead of checking whether we are running in interrupt context, keep
track of the level of nested kernel mode NEON calls in progress, and only
perform the eager stack/unstack if the level exceeds 1.
Signed-off-by: Ard Biesheuvel <redacted>
---
v6:
- use a spinlock instead of disabling interrupts
My concern with the spinlock or disabling interrupts is the latency if
we ever get some insanely long SVE vectors.
Thinking about this a bit more, and trying a couple of things in code,
I think this looks like it could be a nasty problem.
The primary issue is that *any* code that handles the FP/SIMD state,
i.e., preserve it, restore it, etc, could be interrupted, and if
kernel_neon_begin()/_end() was used during that time, the top SVE end
of the registers is just blown away if we don't make the nested
save/restore SVE aware.
For instance, looking at
void fpsimd_restore_current_state(void)
{
preempt_disable();
if (test_and_clear_thread_flag(TIF_FOREIGN_FPSTATE)) {
struct fpsimd_state *st = ¤t->thread.fpsimd_state;
fpsimd_load_state(st);
this_cpu_write(fpsimd_last_state, st);
st->cpu = smp_processor_id();
}
preempt_enable();
}
if fpsimd_load_state() is interrupted and the NEON is used, the
registers that were restored before the interruption will have
incorrect values if SVE is being used by userland.
On the preserve side, we can deal with this by preserving into a temp
buffer, and only store the value that was recorded when the flag was
set, i.e.,
static void safe_preserve_fpsimd_state(void)
{
union __fpsimd_percpu_state *s = this_cpu_ptr(nested_fpsimdstate);
if (in_irq()) {
/* No interruptions possible, so just proceed */
if (!test_and_set_thread_flag(TIF_FOREIGN_FPSTATE))
fpsimd_save_state(¤t->thread.fpsimd_state);
The indentation is slightly off here: the 'else' treats the !in_irq() case
} else {
struct fpsimd_state *fp;
fp = in_interrupt() ? &s[0].full : &s[1].full;
fpsimd_save_state(fp);
if (!test_and_set_thread_flag(TIF_FOREIGN_FPSTATE))
current->thread.fpsimd_state = *fp;
}
}
which is already cumbersome if the full FP/SIMD state grows to 64 KB
However, on the restore side, I fail to see how we can tolerate
interruptions in a similar way.
--
Ard.