mbox series

[v2,0/9] arm64: rework NEON yielding to avoid scheduling from asm code

Message ID 20210203113626.220151-1-ardb@kernel.org
Headers show
Series arm64: rework NEON yielding to avoid scheduling from asm code | expand

Message

Ard Biesheuvel Feb. 3, 2021, 11:36 a.m. UTC
Given how kernel mode NEON code disables preemption (to ensure that the
FP/SIMD register state is protected without having to context switch it),
we need to take care not to let those algorithms operate on unbounded
input data, or we may end up with excessive scheduling blackouts on
CONFIG_PREEMPT kernels.

This is currently handled by the cond_yield_neon macros, which check the
preempt count and the TIF_NEED_RESCHED flag from assembler code, and call
into kernel_neon_end()+kernel_neon_begin(), triggering a reschedule.
This works as expected, but is a bit messy, given how much of the state
preserve/restore code in the algorithm needs to be duplicated, as well as
causing the need to manage the stack frame explicitly. All of this is better
handled by the compiler, especially now that we have enabled features such
as the shadow call stack and BTI, and are working to improve call stack
validation.

In some cases, yielding is not necessary at all: algoritms that implement
skciphers and use the skcipher walk API will be invoked at page granularity,
which is granular enough for our purpose.

In other cases, it is better to simply return early from the assembler
routine if a reschedule is pending, and let the C code handle with this, by
retrying the call until it completes. This removes any voluntary schedule()
calls from the call stack, making the code much easier to reason about in
the context of stack validation, rcu_tasks synchronization, etc.

Practical note: assuming there are no objections to these changes, it may
be the most convenient to take patch #1 into the arm64 tree for v5.12,
and postpone the rest for merging via the crypto tree. (Note that this
series was created against the cryptodev tree, and so the arm64 maintainers
are also welcome to take the whole set if it applies cleanly to the arm64
tree)

Will: if you stick #1 on a separate branch, please base it on v5.11-rc1

Changes since v1:
- use sub+cbz instead of cmp+b.eq to avoid clobbering the flags in cond_yield
  (patch #1)

Cc: Dave Martin <dave.martin@arm.com>
Cc: Eric Biggers <ebiggers@google.com>

Ard Biesheuvel (9):
  arm64: assembler: add cond_yield macro
  crypto: arm64/sha1-ce - simplify NEON yield
  crypto: arm64/sha2-ce - simplify NEON yield
  crypto: arm64/sha3-ce - simplify NEON yield
  crypto: arm64/sha512-ce - simplify NEON yield
  crypto: arm64/aes-neonbs - remove NEON yield calls
  crypto: arm64/aes-ce-mac - simplify NEON yield
  crypto: arm64/crc-t10dif - move NEON yield to C code
  arm64: assembler: remove conditional NEON yield macros

 arch/arm64/crypto/aes-glue.c          | 21 +++--
 arch/arm64/crypto/aes-modes.S         | 52 +++++--------
 arch/arm64/crypto/aes-neonbs-core.S   |  8 +-
 arch/arm64/crypto/crct10dif-ce-core.S | 43 +++--------
 arch/arm64/crypto/crct10dif-ce-glue.c | 30 ++++++--
 arch/arm64/crypto/sha1-ce-core.S      | 47 ++++--------
 arch/arm64/crypto/sha1-ce-glue.c      | 22 +++---
 arch/arm64/crypto/sha2-ce-core.S      | 38 ++++-----
 arch/arm64/crypto/sha2-ce-glue.c      | 22 +++---
 arch/arm64/crypto/sha3-ce-core.S      | 81 ++++++++------------
 arch/arm64/crypto/sha3-ce-glue.c      | 14 ++--
 arch/arm64/crypto/sha512-ce-core.S    | 29 ++-----
 arch/arm64/crypto/sha512-ce-glue.c    | 53 +++++++------
 arch/arm64/include/asm/assembler.h    | 78 +++----------------
 14 files changed, 209 insertions(+), 329 deletions(-)

Comments

Will Deacon Feb. 4, 2021, 10:33 a.m. UTC | #1
On Wed, Feb 03, 2021 at 09:31:31PM +0000, Will Deacon wrote:
> On Wed, 3 Feb 2021 12:36:17 +0100, Ard Biesheuvel wrote:
> > Given how kernel mode NEON code disables preemption (to ensure that the
> > FP/SIMD register state is protected without having to context switch it),
> > we need to take care not to let those algorithms operate on unbounded
> > input data, or we may end up with excessive scheduling blackouts on
> > CONFIG_PREEMPT kernels.
> > 
> > This is currently handled by the cond_yield_neon macros, which check the
> > preempt count and the TIF_NEED_RESCHED flag from assembler code, and call
> > into kernel_neon_end()+kernel_neon_begin(), triggering a reschedule.
> > This works as expected, but is a bit messy, given how much of the state
> > preserve/restore code in the algorithm needs to be duplicated, as well as
> > causing the need to manage the stack frame explicitly. All of this is better
> > handled by the compiler, especially now that we have enabled features such
> > as the shadow call stack and BTI, and are working to improve call stack
> > validation.
> > 
> > [...]
> 
> Applied first patch only to arm64 (for-next/crypto), thanks!

Oops, looks like I typo'd the external branch (for-next/crypo). No offense
intended! I'll rename it now; SHAs will stay the same.

Will
Will Deacon Feb. 4, 2021, 1:03 p.m. UTC | #2
On Thu, Feb 04, 2021 at 10:10:26PM +1100, Herbert Xu wrote:
> On Thu, Feb 04, 2021 at 09:29:16AM +0100, Ard Biesheuvel wrote:
> >
> > Half of the patches in this series conflict with
> > 
> > 0df07d8117c3 crypto: arm64/sha - add missing module aliases
> > 
> > in the cryptodev tree, so that won't work.
> 
> Fair enough.  I'll take the patches.

Cheers, Herbert. Please just leave the final patch ("arm64: assembler:
remove conditional NEON yield macro"), as we can come back to that for
5.13.

Will
Herbert Xu Feb. 10, 2021, 7:23 a.m. UTC | #3
On Wed, Feb 03, 2021 at 12:36:17PM +0100, Ard Biesheuvel wrote:
> Given how kernel mode NEON code disables preemption (to ensure that the

> FP/SIMD register state is protected without having to context switch it),

> we need to take care not to let those algorithms operate on unbounded

> input data, or we may end up with excessive scheduling blackouts on

> CONFIG_PREEMPT kernels.

> 

> This is currently handled by the cond_yield_neon macros, which check the

> preempt count and the TIF_NEED_RESCHED flag from assembler code, and call

> into kernel_neon_end()+kernel_neon_begin(), triggering a reschedule.

> This works as expected, but is a bit messy, given how much of the state

> preserve/restore code in the algorithm needs to be duplicated, as well as

> causing the need to manage the stack frame explicitly. All of this is better

> handled by the compiler, especially now that we have enabled features such

> as the shadow call stack and BTI, and are working to improve call stack

> validation.

> 

> In some cases, yielding is not necessary at all: algoritms that implement

> skciphers and use the skcipher walk API will be invoked at page granularity,

> which is granular enough for our purpose.

> 

> In other cases, it is better to simply return early from the assembler

> routine if a reschedule is pending, and let the C code handle with this, by

> retrying the call until it completes. This removes any voluntary schedule()

> calls from the call stack, making the code much easier to reason about in

> the context of stack validation, rcu_tasks synchronization, etc.

> 

> Practical note: assuming there are no objections to these changes, it may

> be the most convenient to take patch #1 into the arm64 tree for v5.12,

> and postpone the rest for merging via the crypto tree. (Note that this

> series was created against the cryptodev tree, and so the arm64 maintainers

> are also welcome to take the whole set if it applies cleanly to the arm64

> tree)

> 

> Will: if you stick #1 on a separate branch, please base it on v5.11-rc1

> 

> Changes since v1:

> - use sub+cbz instead of cmp+b.eq to avoid clobbering the flags in cond_yield

>   (patch #1)

> 

> Cc: Dave Martin <dave.martin@arm.com>

> Cc: Eric Biggers <ebiggers@google.com>

> 

> Ard Biesheuvel (9):

>   arm64: assembler: add cond_yield macro

>   crypto: arm64/sha1-ce - simplify NEON yield

>   crypto: arm64/sha2-ce - simplify NEON yield

>   crypto: arm64/sha3-ce - simplify NEON yield

>   crypto: arm64/sha512-ce - simplify NEON yield

>   crypto: arm64/aes-neonbs - remove NEON yield calls

>   crypto: arm64/aes-ce-mac - simplify NEON yield

>   crypto: arm64/crc-t10dif - move NEON yield to C code

>   arm64: assembler: remove conditional NEON yield macros

> 

>  arch/arm64/crypto/aes-glue.c          | 21 +++--

>  arch/arm64/crypto/aes-modes.S         | 52 +++++--------

>  arch/arm64/crypto/aes-neonbs-core.S   |  8 +-

>  arch/arm64/crypto/crct10dif-ce-core.S | 43 +++--------

>  arch/arm64/crypto/crct10dif-ce-glue.c | 30 ++++++--

>  arch/arm64/crypto/sha1-ce-core.S      | 47 ++++--------

>  arch/arm64/crypto/sha1-ce-glue.c      | 22 +++---

>  arch/arm64/crypto/sha2-ce-core.S      | 38 ++++-----

>  arch/arm64/crypto/sha2-ce-glue.c      | 22 +++---

>  arch/arm64/crypto/sha3-ce-core.S      | 81 ++++++++------------

>  arch/arm64/crypto/sha3-ce-glue.c      | 14 ++--

>  arch/arm64/crypto/sha512-ce-core.S    | 29 ++-----

>  arch/arm64/crypto/sha512-ce-glue.c    | 53 +++++++------

>  arch/arm64/include/asm/assembler.h    | 78 +++----------------

>  14 files changed, 209 insertions(+), 329 deletions(-)


Patches 2-8 applied.  Thanks.
-- 
Email: Herbert Xu <herbert@gondor.apana.org.au>
Home Page: http://gondor.apana.org.au/~herbert/
PGP Key: http://gondor.apana.org.au/~herbert/pubkey.txt