Skip to contentSkip to conversation

powerpc: reduce number of WATCHDOG_RESET calls from flush_cache

PATCH New Rasmus Villemoes 5 participants 9 messages Jun 4, 2020 – Sep 24, 2020 PowerPC

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 powerpc: reduce number of WATCHDOG_RESET calls from flush_cache

Calling WATCHDOG_RESET for each and every cache line is overkill.

In our case, the kernel image is a little over 7MB, and the almost 500000 calls of WATCHDOG_RESET() adds about one second to the boottime.

I very highly doubt there's any real hardware where flushing 64K from cache to memory takes more than a few milliseconds, so this should be completely safe. Since it reduces the number of WATCHDOG_RESET() calls by roughly a factor of 1000, the overhead from those is practically eliminated. (Just in case the range flushed is so small that it doesn't cross a 64K boundary, add a single WATCHDOG_RESET() between the loops).

64K is chosen because that's also the default chunk size used by the hashing algorithms, and when, say, a sha256 digest of a kernel image of a few MB is being verified, that's almost guaranteed to be cache-cold, so apart from the computations being done, the hashing is also bounded by memory speed - so if 64K works for those cases, it should certainly also work when memory access is the only thing being done.

---
 arch/powerpc/lib/cache.c | 8 ++++++--
 1 file changed, 6 insertions(+), 2 deletions(-)
arch/powerpc/lib/cache.c +6 -2
index 528361e972..df2310f4e2 100644
@@ -8,6 +8,7 @@
 #include <cpu_func.h>
 #include <asm/cache.h>
 #include <watchdog.h>
+#include <linux/sizes.h>
 
 void flush_cache(ulong start_addr, ulong size)
 {
@@ -21,15 +22,18 @@ void flush_cache(ulong start_addr, ulong size)
 	for (addr = start; (addr <= end) && (addr >= start);
 			addr += CONFIG_SYS_CACHELINE_SIZE) {
 		asm volatile("dcbst 0,%0" : : "r" (addr) : "memory");
-		WATCHDOG_RESET();
+		if ((addr & (SZ_64K - 1)) == 0)
+			WATCHDOG_RESET();
 	}
 	/* wait for all dcbst to complete on bus */
 	asm volatile("sync" : : : "memory");
+	WATCHDOG_RESET();
 
 	for (addr = start; (addr <= end) && (addr >= start);
 			addr += CONFIG_SYS_CACHELINE_SIZE) {
 		asm volatile("icbi 0,%0" : : "r" (addr) : "memory");
-		WATCHDOG_RESET();
+		if ((addr & (SZ_64K - 1)) == 0)
+			WATCHDOG_RESET();
 	}
 	asm volatile("sync" : : : "memory");
 	/* flush prefetch queue */
Signature · 1 line
2.23.0
#1 lore raw

PATCH 1/1 powerpc: reduce number of WATCHDOG_RESET calls from flush_cache

Quoted 14 lines · Rasmus Villemoes

Calling WATCHDOG_RESET for each and every cache line is overkill.

In our case, the kernel image is a little over 7MB, and the almost 500000 calls of WATCHDOG_RESET() adds about one second to the boottime.

I very highly doubt there's any real hardware where flushing 64K from cache to memory takes more than a few milliseconds, so this should be completely safe. Since it reduces the number of WATCHDOG_RESET() calls by roughly a factor of 1000, the overhead from those is practically eliminated. (Just in case the range flushed is so small that it doesn't cross a 64K boundary, add a single WATCHDOG_RESET() between the loops).

Quoted 8 lines

64K is chosen because that's also the default chunk size used by the hashing algorithms, and when, say, a sha256 digest of a kernel image of a few MB is being verified, that's almost guaranteed to be cache-cold, so apart from the computations being done, the hashing is also bounded by memory speed - so if 64K works for those cases, it should certainly also work when memory access is the only thing being done.

Thanks,
Stefan

Quoted 38 lines
---
  arch/powerpc/lib/cache.c | 8 ++++++--
  1 file changed, 6 insertions(+), 2 deletions(-)
arch/powerpc/lib/cache.c +6 -2
index 528361e972..df2310f4e2 100644
@@ -8,6 +8,7 @@
  #include <cpu_func.h>
  #include <asm/cache.h>
  #include <watchdog.h>
+#include <linux/sizes.h>
 
  void flush_cache(ulong start_addr, ulong size)
  {
@@ -21,15 +22,18 @@ void flush_cache(ulong start_addr, ulong size)
  	for (addr = start; (addr <= end) && (addr >= start);
  			addr += CONFIG_SYS_CACHELINE_SIZE) {
  		asm volatile("dcbst 0,%0" : : "r" (addr) : "memory");
-		WATCHDOG_RESET();
+		if ((addr & (SZ_64K - 1)) == 0)
+			WATCHDOG_RESET();
  	}
  	/* wait for all dcbst to complete on bus */
  	asm volatile("sync" : : : "memory");
+	WATCHDOG_RESET();
 
  	for (addr = start; (addr <= end) && (addr >= start);
  			addr += CONFIG_SYS_CACHELINE_SIZE) {
  		asm volatile("icbi 0,%0" : : "r" (addr) : "memory");
-		WATCHDOG_RESET();
+		if ((addr & (SZ_64K - 1)) == 0)
+			WATCHDOG_RESET();
  	}
  	asm volatile("sync" : : : "memory");
  	/* flush prefetch queue */
#2 lore raw

PATCH 1/1 powerpc: reduce number of WATCHDOG_RESET calls from flush_cache

Quoted 22 lines · Stefan Roese
Quoted 21 lines · Rasmus Villemoes

Calling WATCHDOG_RESET for each and every cache line is overkill.

In our case, the kernel image is a little over 7MB, and the almost 500000 calls of WATCHDOG_RESET() adds about one second to the boottime.

I very highly doubt there's any real hardware where flushing 64K from cache to memory takes more than a few milliseconds, so this should be completely safe. Since it reduces the number of WATCHDOG_RESET() calls by roughly a factor of 1000, the overhead from those is practically eliminated. (Just in case the range flushed is so small that it doesn't cross a 64K boundary, add a single WATCHDOG_RESET() between the loops).

64K is chosen because that's also the default chunk size used by the hashing algorithms, and when, say, a sha256 digest of a kernel image of a few MB is being verified, that's almost guaranteed to be cache-cold, so apart from the computations being done, the hashing is also bounded by memory speed - so if 64K works for those cases, it should certainly also work when memory access is the only thing being done.

Quoted 3 lines
Quoted 1 line

Ping.

Quoted 39 lines

Thanks,
Stefan

Quoted 36 lines

---
? arch/powerpc/lib/cache.c | 8 ++++++--
? 1 file changed, 6 insertions(+), 2 deletions(-)

arch/powerpc/lib/cache.c +0 -0
index 528361e972..df2310f4e2 100644
@@ -8,6 +8,7 @@

? #include <cpu_func.h>
? #include <asm/cache.h>
? #include <watchdog.h>
+#include <linux/sizes.h>
? ? void flush_cache(ulong start_addr, ulong size) ? {

arch/powerpc/lib/cache.c +2 -1
@@ -21,15 +22,18 @@ void flush_cache(ulong start_addr, ulong size)
 ????? for (addr = start; (addr <= end) && (addr >= start);
 ????????????? addr += CONFIG_SYS_CACHELINE_SIZE) {
 ????????? asm volatile("dcbst 0,%0" : : "r" (addr) : "memory");
-??????? WATCHDOG_RESET();
+??????? if ((addr & (SZ_64K - 1)) == 0)
+??????????? WATCHDOG_RESET();
 ????? }

????? /* wait for all dcbst to complete on bus */ ????? asm volatile("sync" : : : "memory");
+??? WATCHDOG_RESET();
? ????? for (addr = start; (addr <= end) && (addr >= start); ????????????? addr += CONFIG_SYS_CACHELINE_SIZE) { ????????? asm volatile("icbi 0,%0" : : "r" (addr) : "memory"); -??????? WATCHDOG_RESET();
+??????? if ((addr & (SZ_64K - 1)) == 0) +??????????? WATCHDOG_RESET();
????? }
????? asm volatile("sync" : : : "memory");
????? /* flush prefetch queue */

Signature · 8 lines
Rasmus Villemoes
Software Developer
Prevas A/S
Hedeager 3
DK-8200 Aarhus N
+45 51210274
rasmus.villemoes at prevas.dk
www.prevas.dk
#3 lore raw

PATCH 1/1 powerpc: reduce number of WATCHDOG_RESET calls from flush_cache

Quoted 23 lines · Rasmus Villemoes
Quoted 22 lines · Stefan Roese
Quoted 21 lines · Rasmus Villemoes

Calling WATCHDOG_RESET for each and every cache line is overkill.

In our case, the kernel image is a little over 7MB, and the almost 500000 calls of WATCHDOG_RESET() adds about one second to the boottime.

I very highly doubt there's any real hardware where flushing 64K from cache to memory takes more than a few milliseconds, so this should be completely safe. Since it reduces the number of WATCHDOG_RESET() calls by roughly a factor of 1000, the overhead from those is practically eliminated. (Just in case the range flushed is so small that it doesn't cross a 64K boundary, add a single WATCHDOG_RESET() between the loops).

64K is chosen because that's also the default chunk size used by the hashing algorithms, and when, say, a sha256 digest of a kernel image of a few MB is being verified, that's almost guaranteed to be cache-cold, so apart from the computations being done, the hashing is also bounded by memory speed - so if 64K works for those cases, it should certainly also work when memory access is the only thing being done.

Quoted 5 lines
Quoted 3 lines
Quoted 1 line

Ping.

This too should end up in -next "soon". Thanks.

Signature · 8 lines
Tom
-------------- next part --------------
A non-text attachment was scrubbed...
Name: signature.asc
Type: application/pgp-signature
Size: 659 bytes
Desc: not available
URL: <https://lists.denx.de/pipermail/u-boot/attachments/20200918/b78f4180/attachment.sig>
#4 lore raw

PATCH 1/1 powerpc: reduce number of WATCHDOG_RESET calls from flush_cache

In reply to Rasmus Villemoes

Dear Rasmus,

Quoted 11 lines · In message <[email protected]> you wrote:
Quoted 10 lines · Stefan Roese
Quoted 9 lines · Rasmus Villemoes

Calling WATCHDOG_RESET for each and every cache line is overkill.

In our case, the kernel image is a little over 7MB, and the almost 500000 calls of WATCHDOG_RESET() adds about one second to the boottime.

I very highly doubt there's any real hardware where flushing 64K from cache to memory takes more than a few milliseconds, so this should be completely safe. Since it reduces the number of

Maybe it is, maybe it is not. I remember boards where the watchdog trigger had to happen in pretty tight intervals, something like min 20, max 60 milliseconds.

Quoted 1 line
Quoted 1 line
Quoted 1 line

+??????? if ((addr & (SZ_64K - 1)) == 0)

In any case, please write readable code. The use of SZ_ here is not a good idea. Also, you might want to make the actual parameter configurable, so boards with specific timing requirements can adjust this as needed.

Best regards,

Wolfgang Denk

Signature · 6 lines
DENX Software Engineering GmbH,      Managing Director: Wolfgang Denk
HRB 165235 Munich, Office: Kirchenstr.5, D-82194 Groebenzell, Germany
Phone: (+49)-8142-66989-10 Fax: (+49)-8142-66989-80 Email: wd at denx.de
Perfection is reached, not when there is no longer anything  to  add,
but when there is no longer anything to take away.
                                           - Antoine de Saint-Exupery
#5 lore raw

PATCH 1/1 powerpc: reduce number of WATCHDOG_RESET calls from flush_cache

Quoted 19 lines · Wolfgang Denk

Dear Rasmus,

Quoted 11 lines · In message <[email protected]> you wrote:
Quoted 10 lines · Stefan Roese
Quoted 9 lines · Rasmus Villemoes

Calling WATCHDOG_RESET for each and every cache line is overkill.

In our case, the kernel image is a little over 7MB, and the almost 500000 calls of WATCHDOG_RESET() adds about one second to the boottime.

I very highly doubt there's any real hardware where flushing 64K from cache to memory takes more than a few milliseconds, so this should be completely safe. Since it reduces the number of

Maybe it is, maybe it is not. I remember boards where the watchdog trigger had to happen in pretty tight intervals, something like min 20, max 60 milliseconds.

Quoted 6 lines
Quoted 1 line
Quoted 1 line
Quoted 1 line

+??????? if ((addr & (SZ_64K - 1)) == 0)

In any case, please write readable code. The use of SZ_ here is not a good idea. Also, you might want to make the actual parameter configurable, so boards with specific timing requirements can adjust this as needed.

We should add a comment somewhere here to match the commit message, but since we're flushing on 64K chunks, SZ_64K makes sense. If we make it build-time configurable then it's not a concern. I assume we don't want to make this run time configurable.

Signature · 8 lines
Tom
-------------- next part --------------
A non-text attachment was scrubbed...
Name: signature.asc
Type: application/pgp-signature
Size: 659 bytes
Desc: not available
URL: <https://lists.denx.de/pipermail/u-boot/attachments/20200924/486dbce4/attachment.sig>
#6 lore raw

PATCH 1/1 powerpc: reduce number of WATCHDOG_RESET calls from flush_cache

In reply to Rasmus Villemoes
Quoted 26 lines

-----Original Message-----
From: U-Boot <[email protected]> On Behalf Of Rasmus Villemoes
Sent: Friday, September 18, 2020 1:51 PM
To: Stefan Roese <[email protected]>; u-boot at lists.denx.de
Cc: Christophe Leroy <[email protected]>; Wolfgang Denk <[email protected]>; Simon Glass <[email protected]>; Tom Rini <[email protected]>
Subject: Re: [PATCH] powerpc: reduce number of WATCHDOG_RESET calls from flush_cache

Quoted 15 lines · Stefan Roese
Quoted 14 lines · Rasmus Villemoes

Calling WATCHDOG_RESET for each and every cache line is overkill.

In our case, the kernel image is a little over 7MB, and the almost 500000 calls of WATCHDOG_RESET() adds about one second to the boottime.

I very highly doubt there's any real hardware where flushing 64K from cache to memory takes more than a few milliseconds, so this should be completely safe. Since it reduces the number of WATCHDOG_RESET() calls by roughly a factor of 1000, the overhead from those is practically eliminated. (Just in case the range flushed is so small that it doesn't cross a 64K boundary, add a single WATCHDOG_RESET() between the loops).

Quoted 12 lines
Quoted 6 lines
Quoted 6 lines

64K is chosen because that's also the default chunk size used by the hashing algorithms, and when, say, a sha256 digest of a kernel image of a few MB is being verified, that's almost guaranteed to be cache-cold, so apart from the computations being done, the hashing is also bounded by memory speed - so if 64K works for those cases, it should certainly also work when memory access is the only thing being

done.

Quoted 4 lines
Quoted 2 lines

<snip>
Applied to u-boot-mpc85xx. Awaiting upstream

Thanks
Priyanka

#7 lore raw

PATCH 1/1 powerpc: reduce number of WATCHDOG_RESET calls from flush_cache

Dear Priyanka,

Quoted 7 lines · In message <VE1PR04MB649416C255B3379B43E7F7FDE6380@VE1PR04MB6494.eurprd04.prod.…
Quoted 5 lines
Quoted 4 lines
Quoted 2 lines

<snip>
Applied to u-boot-mpc85xx. Awaiting upstream

I object. This must at least be configurable, so that boards with tight timing requirements can still work.

Best regards,

Wolfgang Denk

Signature · 5 lines
DENX Software Engineering GmbH,      Managing Director: Wolfgang Denk
HRB 165235 Munich, Office: Kirchenstr.5, D-82194 Groebenzell, Germany
Phone: (+49)-8142-66989-10 Fax: (+49)-8142-66989-80 Email: wd at denx.de
The human mind  ordinarily  operates  at  only  ten  percent  of  its
capacity. The rest is overhead for the operating system.
#8 lore raw

PATCH 1/1 powerpc: reduce number of WATCHDOG_RESET calls from flush_cache

Quoted 6 lines · Wolfgang Denk

Dear Priyanka,

Quoted 3 lines · In message <VE1PR04MB649416C255B3379B43E7F7FDE6380@VE1PR04MB6494.eurprd04.prod.…
Quoted 3 lines
Quoted 3 lines
Quoted 2 lines
Quoted 7 lines
Quoted 4 lines
Quoted 2 lines
Quoted 1 line

<snip>
Applied to u-boot-mpc85xx. Awaiting upstream

I object. This must at least be configurable, so that boards with tight timing requirements can still work.

I'm removing just this patch from the -next pull request.

Signature · 8 lines
Tom
-------------- next part --------------
A non-text attachment was scrubbed...
Name: signature.asc
Type: application/pgp-signature
Size: 659 bytes
Desc: not available
URL: <https://lists.denx.de/pipermail/u-boot/attachments/20200924/d8dbaf12/attachment.sig>