drivers/net/b44: Change to non-atomic bit operations
Subsystems (3)
PATCH 1/1 drivers/net/b44: Change to non-atomic bit operations
Since "pwol_mask" is local and never exposed to concurrency, there is no need to set bit in pwol_mask by costly atomic operations. On x86, accessing data across two cache lines in one atomic bit operation (aka split lock) can take over 1000 cycles.
Directly operate on the byte which contains the bit instead of using __set_bit() to avoid any big endian concern due to type cast to unsigned long in __set_bit().
--- drivers/net/ethernet/broadcom/b44.c | 9 ++++++--- 1 file changed, 6 insertions(+), 3 deletions(-)
drivers/net/ethernet/broadcom/b44.c +6 -3
@@ -1516,8 +1516,10 @@ static int b44_magic_pattern(u8 *macaddr, u8 *ppattern, u8 *pmask, int offset)int ethaddr_bytes =ETH_ALEN ;memset (ppattern + offset,0xff , magicsync); -for (j =0 ; j < magicsync; j++) -set_bit (len++, (unsigned long *) pmask); +for (j =0 ; j < magicsync; j++) { + pmask[len >>3 ] |=BIT (len &7 ); + len++; + }for (j =0 ; j <B44_MAX_PATTERNS ; j++) {if ((B44_PATTERN_SIZE - len) >=ETH_ALEN )
@@ -1529,7 +1531,8 @@ static int b44_magic_pattern(u8 *macaddr, u8 *ppattern, u8 *pmask, int offset)for (k =0 ; k< ethaddr_bytes; k++) { ppattern[offset + magicsync + (j *ETH_ALEN ) + k] = macaddr[k]; -set_bit (len++, (unsigned long *) pmask); + pmask[len >>3 ] |=BIT (len &7 ); + len++; } }return len -1 ;
Signature · 1 line
2.19.1
David Miller
maintainerFrom: Fenghua Yu <[email protected]>
Date: Fri, 20 Dec 2019 15:29:11 -0800
Quoted 2 lines
On x86, accessing data across two cache lines in one atomic bit operation (aka split lock) can take over 1000 cycles.
This happens during configuration of WOL, nobody cares that the atomic operations done in this function take 1000 cycles each.
I'm not applying this patch. It is gratuitous, and the commit message talks about "performance" considuations (cycle counts) that completely don't matter here.
If you are merely just arbitrarily trying to remove locked atomic operations across the tree for it's own sake, then you should be completely honest about that in your commit message.
Quoted 9 lines · David Miller
From: Fenghua Yu <[email protected]>
Date: Fri, 20 Dec 2019 15:29:11 -0800Quoted 2 lines
On x86, accessing data across two cache lines in one atomic bit operation (aka split lock) can take over 1000 cycles.
This happens during configuration of WOL, nobody cares that the atomic operations done in this function take 1000 cycles each.
Quoted 7 lines
I'm not applying this patch. It is gratuitous, and the commit message talks about "performance" considuations (cycle counts) that completely don't matter here.
If you are merely just arbitrarily trying to remove locked atomic operations across the tree for it's own sake, then you should be completely honest about that in your commit message.
We are enabling split lock in the kernel (by default):
https://lkml.org/lkml/2019/12/12/1129
After applying the split lock detection patch, the set_bit() in b44.c may cause split lock and kernel dies.
So should I change the commit message to add the above info?
Thanks.
-Fenghua
David Miller
maintainerFrom: Fenghua Yu <[email protected]>
Date: Tue, 24 Dec 2019 17:10:20 -0800
Quoted 18 lines · David Miller
Quoted 9 lines
From: Fenghua Yu <[email protected]>
Date: Fri, 20 Dec 2019 15:29:11 -0800Quoted 2 lines
On x86, accessing data across two cache lines in one atomic bit operation (aka split lock) can take over 1000 cycles.
This happens during configuration of WOL, nobody cares that the atomic operations done in this function take 1000 cycles each.
Quoted 7 lines
I'm not applying this patch. It is gratuitous, and the commit message talks about "performance" considuations (cycle counts) that completely don't matter here.
If you are merely just arbitrarily trying to remove locked atomic operations across the tree for it's own sake, then you should be completely honest about that in your commit message.
Quoted 7 lines
We are enabling split lock in the kernel (by default):
https://lkml.org/lkml/2019/12/12/1129After applying the split lock detection patch, the
set_bit()inb44.cmay cause split lock and kernel dies.So should I change the commit message to add the above info?
You're asking me if you should make your commit message accurate?
PATCH v2 1/1 drivers/net/b44: Change to non-atomic bit operations on pwol_mask
From: Fenghua Yu <[email protected]>
Atomic operations that span cache lines are super-expensive on x86 (not just to the current processor, but also to other processes as all memory operations are blocked until the operation completes). Upcoming x86 processors have a switch to cause such operations to generate a #AC trap. It is expected that some real time systems will enable this mode in BIOS.
In preparation for this, it is necessary to fix code that may execute atomic instructions with operands that cross cachelines because the #AC trap will crash the kernel.
Since "pwol_mask" is local and never exposed to concurrency, there is no need to set bits in pwol_mask using atomic operations.
Directly operate on the byte which contains the bit instead of using __set_bit() to avoid any big endian concern due to type cast to unsigned long in __set_bit().
---
Tony: Updated commit comment with background information motivating this change. No changes to code.
drivers/net/ethernet/broadcom/b44.c | 9 ++++++--- 1 file changed, 6 insertions(+), 3 deletions(-)
drivers/net/ethernet/broadcom/b44.c +6 -3
@@ -1516,8 +1516,10 @@ static int b44_magic_pattern(u8 *macaddr, u8 *ppattern, u8 *pmask, int offset)int ethaddr_bytes =ETH_ALEN ;memset (ppattern + offset,0xff , magicsync); -for (j =0 ; j < magicsync; j++) -set_bit (len++, (unsigned long *) pmask); +for (j =0 ; j < magicsync; j++) { + pmask[len >>3 ] |=BIT (len &7 ); + len++; + }for (j =0 ; j <B44_MAX_PATTERNS ; j++) {if ((B44_PATTERN_SIZE - len) >=ETH_ALEN )
@@ -1529,7 +1531,8 @@ static int b44_magic_pattern(u8 *macaddr, u8 *ppattern, u8 *pmask, int offset)for (k =0 ; k< ethaddr_bytes; k++) { ppattern[offset + magicsync + (j *ETH_ALEN ) + k] = macaddr[k]; -set_bit (len++, (unsigned long *) pmask); + pmask[len >>3 ] |=BIT (len &7 ); + len++; } }return len -1 ;
Signature · 1 line
2.21.0
Re: [Patch v2] drivers/net/b44: Change to non-atomic bit operations on pwol_mask
From: "Luck, Tony" <[email protected]>
Date: Thu, 2 Jan 2020 13:27:06 -0800
Quoted 16 lines
From: Fenghua Yu <[email protected]>
Atomic operations that span cache lines are super-expensive on x86 (not just to the current processor, but also to other processes as all memory operations are blocked until the operation completes). Upcoming x86 processors have a switch to cause such operations to generate a #AC trap. It is expected that some real time systems will enable this mode in BIOS.
In preparation for this, it is necessary to fix code that may execute atomic instructions with operands that cross cachelines because the #AC trap will crash the kernel.
Since "
pwol_mask" is local and never exposed to concurrency, there is no need to set bits inpwol_maskusing atomic operations.
Quoted 7 lines
Directly operate on the byte which contains the bit instead of using
__set_bit()to avoid any big endian concern due to type cast to unsigned long in__set_bit().Suggested-by Peter Zijlstra <[email protected]>Signed-off-by Fenghua Yu <[email protected]>Signed-off-by Tony Luck <[email protected]>
Applied, thanks.
I wonder if this is being used in an endian safe way. Maybe the way the filter is written into the chip makes it work out, I don't know.
Quoted 21 lines · On Tue, 24 Dec 2019 17:10:20 -0800 Fenghua Yu <[email protected]> wrote:
Quoted 9 lines · David Miller
From: Fenghua Yu <[email protected]>
Date: Fri, 20 Dec 2019 15:29:11 -0800Quoted 2 lines
On x86, accessing data across two cache lines in one atomic bit operation (aka split lock) can take over 1000 cycles.
This happens during configuration of WOL, nobody cares that the atomic operations done in this function take 1000 cycles each.
Quoted 7 lines
I'm not applying this patch. It is gratuitous, and the commit message talks about "performance" considuations (cycle counts) that completely don't matter here.
If you are merely just arbitrarily trying to remove locked atomic operations across the tree for it's own sake, then you should be completely honest about that in your commit message.
We are enabling split lock in the kernel (by default):
https://lkml.org/lkml/2019/12/12/1129
Quoted 8 lines
After applying the split lock detection patch, the
set_bit()inb44.cmay cause split lock and kernel dies.So should I change the commit message to add the above info?
Thanks.
-Fenghua
Why not just make pwol_pattern aligned and choose the right word to do the operation on?
Quoted 2 lines
Why not just make
pwol_patternaligned and choose the right word to do the operation on?
We use that approach for places where the operation needs to be atomic.
But this one doesn't need an atomic operation since there can be no other entity operating on the same bitmap in parallel.
-Tony
From: Luck, Tony
Quoted 9 lines
Sent: 02 January 2020 18:23
Quoted 2 lines
Why not just make
pwol_patternaligned and choose the right word to do the operation on?We use that approach for places where the operation needs to be atomic.
But this one doesn't need an atomic operation since there can be no other entity operating on the same bitmap in parallel.
From what I remember this code is setting up a bitmap that is transferred to the hardware (it might be the multicast filter).
As such it shouldn't be relying on the implementation of the bitmap functions (which operate on long[]) to generate the required byte map required by the hardware.
The code is horrid.
David
-
Registered Address Lakeside, Bramley Road, Mount Farm, Milton Keynes, MK1 1PT, UK Registration No: 1397386 (Wales)