[PATCH] lzma: Fix decompression speed regression
Uncompressing a 1.7Mbytes FIT image on U-boot 2023.04 takes approx 7s on a powerpc 8xx. The same on U-boot 2023.07-rc6 takes approx 28s unless watchdog is disabled.
During that decompression, LzmaDec_DecodeReal() calls schedule 1.6 million times, that is every 4µs in average.
In the past it used to be a call to WATCHDOG_RESET() which was just calling hw_watchdog_reset().
But the combination of commit 29caf9305b6 ("cyclic: Use schedule() instead of WATCHDOG_RESET()") and commit 26e8ebcd7cb ("watchdog: mpc8xxx: Make it generic") results in an heavier processing.
However, there is absolutely no point in calling schedule() that often.
By moving and keeping only one call to schedule() in the main loop the number of calls is reduced to 1.2 million which is still too much. So add logic to only call schedule every 1024 times. That leads to a call to schedule approx every 6ms which is still far enough to entertain the watchdog which has a 1s timeout on powerpc 8xx.
powerpc 8xx being one of the slowest targets we have today in U-boot, and most other watchdogs having a timeout of one minutes instead of one second like the 8xx, this fix should not have negative impact on other targets.
Fixes: 29caf9305b6 ("cyclic: Use schedule() instead of WATCHDOG_RESET()") Signed-off-by: Christophe Leroy christophe.leroy@csgroup.eu --- lib/lzma/LzmaDec.c | 18 ++++-------------- 1 file changed, 4 insertions(+), 14 deletions(-)
diff --git a/lib/lzma/LzmaDec.c b/lib/lzma/LzmaDec.c index 341149f766..a90b35c6a9 100644 --- a/lib/lzma/LzmaDec.c +++ b/lib/lzma/LzmaDec.c @@ -152,8 +152,7 @@ static int MY_FAST_CALL LzmaDec_DecodeReal(CLzmaDec *p, SizeT limit, const Byte const Byte *buf = p->buf; UInt32 range = p->range; UInt32 code = p->code; - - schedule(); + unsigned int loop = 0;
do { @@ -162,6 +161,9 @@ static int MY_FAST_CALL LzmaDec_DecodeReal(CLzmaDec *p, SizeT limit, const Byte unsigned ttt; unsigned posState = processedPos & pbMask;
+ if (!(loop++ & 1023)) + schedule(); + prob = probs + IsMatch + (state << kNumPosBitsMax) + posState; IF_BIT_0(prob) { @@ -177,8 +179,6 @@ static int MY_FAST_CALL LzmaDec_DecodeReal(CLzmaDec *p, SizeT limit, const Byte state -= (state < 4) ? state : 3; symbol = 1;
- schedule(); - do { GET_BIT(prob + symbol, symbol) } while (symbol < 0x100); } else @@ -188,8 +188,6 @@ static int MY_FAST_CALL LzmaDec_DecodeReal(CLzmaDec *p, SizeT limit, const Byte state -= (state < 10) ? 3 : 6; symbol = 1;
- schedule(); - do { unsigned bit; @@ -321,8 +319,6 @@ static int MY_FAST_CALL LzmaDec_DecodeReal(CLzmaDec *p, SizeT limit, const Byte UInt32 mask = 1; unsigned i = 1;
- schedule(); - do { GET_BIT2(prob + i, i, ; , distance |= mask); @@ -335,8 +331,6 @@ static int MY_FAST_CALL LzmaDec_DecodeReal(CLzmaDec *p, SizeT limit, const Byte { numDirectBits -= kNumAlignBits;
- schedule(); - do { NORMALIZE @@ -409,8 +403,6 @@ static int MY_FAST_CALL LzmaDec_DecodeReal(CLzmaDec *p, SizeT limit, const Byte const Byte *lim = dest + curLen; dicPos += curLen;
- schedule(); - do *(dest) = (Byte)*(dest + src); while (++dest != lim); @@ -418,8 +410,6 @@ static int MY_FAST_CALL LzmaDec_DecodeReal(CLzmaDec *p, SizeT limit, const Byte else {
- schedule(); - do { dic[dicPos++] = dic[pos];
On Wed, 5 Jul 2023 at 09:54, Christophe Leroy christophe.leroy@csgroup.eu wrote:
Uncompressing a 1.7Mbytes FIT image on U-boot 2023.04 takes approx 7s on a powerpc 8xx. The same on U-boot 2023.07-rc6 takes approx 28s unless watchdog is disabled.
During that decompression, LzmaDec_DecodeReal() calls schedule 1.6 million times, that is every 4µs in average.
In the past it used to be a call to WATCHDOG_RESET() which was just calling hw_watchdog_reset().
But the combination of commit 29caf9305b6 ("cyclic: Use schedule() instead of WATCHDOG_RESET()") and commit 26e8ebcd7cb ("watchdog: mpc8xxx: Make it generic") results in an heavier processing.
However, there is absolutely no point in calling schedule() that often.
By moving and keeping only one call to schedule() in the main loop the number of calls is reduced to 1.2 million which is still too much. So add logic to only call schedule every 1024 times. That leads to a call to schedule approx every 6ms which is still far enough to entertain the watchdog which has a 1s timeout on powerpc 8xx.
powerpc 8xx being one of the slowest targets we have today in U-boot, and most other watchdogs having a timeout of one minutes instead of one second like the 8xx, this fix should not have negative impact on other targets.
Fixes: 29caf9305b6 ("cyclic: Use schedule() instead of WATCHDOG_RESET()") Signed-off-by: Christophe Leroy christophe.leroy@csgroup.eu
lib/lzma/LzmaDec.c | 18 ++++-------------- 1 file changed, 4 insertions(+), 14 deletions(-)
Reviewed-by: Simon Glass sjg@chromium.org
+Stefan Roese I wonder if we need a more general fix?
Hi Simon, Hi Christophe,
On 7/6/23 17:57, Simon Glass wrote:
On Wed, 5 Jul 2023 at 09:54, Christophe Leroy christophe.leroy@csgroup.eu wrote:
Uncompressing a 1.7Mbytes FIT image on U-boot 2023.04 takes approx 7s on a powerpc 8xx. The same on U-boot 2023.07-rc6 takes approx 28s unless watchdog is disabled.
During that decompression, LzmaDec_DecodeReal() calls schedule 1.6 million times, that is every 4µs in average.
In the past it used to be a call to WATCHDOG_RESET() which was just calling hw_watchdog_reset().
But the combination of commit 29caf9305b6 ("cyclic: Use schedule() instead of WATCHDOG_RESET()") and commit 26e8ebcd7cb ("watchdog: mpc8xxx: Make it generic") results in an heavier processing.
However, there is absolutely no point in calling schedule() that often.
By moving and keeping only one call to schedule() in the main loop the number of calls is reduced to 1.2 million which is still too much. So add logic to only call schedule every 1024 times. That leads to a call to schedule approx every 6ms which is still far enough to entertain the watchdog which has a 1s timeout on powerpc 8xx.
powerpc 8xx being one of the slowest targets we have today in U-boot, and most other watchdogs having a timeout of one minutes instead of one second like the 8xx, this fix should not have negative impact on other targets.
Fixes: 29caf9305b6 ("cyclic: Use schedule() instead of WATCHDOG_RESET()") Signed-off-by: Christophe Leroy christophe.leroy@csgroup.eu
lib/lzma/LzmaDec.c | 18 ++++-------------- 1 file changed, 4 insertions(+), 14 deletions(-)
Reviewed-by: Simon Glass sjg@chromium.org
+Stefan Roese I wonder if we need a more general fix?
IIRC, we already have some code in the watchdog IF making sure, that the WDT reset is not called too often. I need to re-check here. And yes, we should probably either move this code to the common schedule() function or add it here. Best by adding a new Kconfig option, configuring the max schedule frequency, e.g. 1000 Hz.
Thanks, Stefan
Le 07/07/2023 à 11:09, Stefan Roese a écrit :
[Vous ne recevez pas souvent de courriers de sr@denx.de. Découvrez pourquoi ceci est important à https://aka.ms/LearnAboutSenderIdentification ]
Hi Simon, Hi Christophe,
On 7/6/23 17:57, Simon Glass wrote:
On Wed, 5 Jul 2023 at 09:54, Christophe Leroy christophe.leroy@csgroup.eu wrote:
Uncompressing a 1.7Mbytes FIT image on U-boot 2023.04 takes approx 7s on a powerpc 8xx. The same on U-boot 2023.07-rc6 takes approx 28s unless watchdog is disabled.
During that decompression, LzmaDec_DecodeReal() calls schedule 1.6 million times, that is every 4µs in average.
In the past it used to be a call to WATCHDOG_RESET() which was just calling hw_watchdog_reset().
But the combination of commit 29caf9305b6 ("cyclic: Use schedule() instead of WATCHDOG_RESET()") and commit 26e8ebcd7cb ("watchdog: mpc8xxx: Make it generic") results in an heavier processing.
However, there is absolutely no point in calling schedule() that often.
By moving and keeping only one call to schedule() in the main loop the number of calls is reduced to 1.2 million which is still too much. So add logic to only call schedule every 1024 times. That leads to a call to schedule approx every 6ms which is still far enough to entertain the watchdog which has a 1s timeout on powerpc 8xx.
powerpc 8xx being one of the slowest targets we have today in U-boot, and most other watchdogs having a timeout of one minutes instead of one second like the 8xx, this fix should not have negative impact on other targets.
Fixes: 29caf9305b6 ("cyclic: Use schedule() instead of WATCHDOG_RESET()") Signed-off-by: Christophe Leroy christophe.leroy@csgroup.eu
lib/lzma/LzmaDec.c | 18 ++++-------------- 1 file changed, 4 insertions(+), 14 deletions(-)
Reviewed-by: Simon Glass sjg@chromium.org
+Stefan Roese I wonder if we need a more general fix?
IIRC, we already have some code in the watchdog IF making sure, that the WDT reset is not called too often. I need to re-check here. And yes, we should probably either move this code to the common schedule() function or add it here. Best by adding a new Kconfig option, configuring the max schedule frequency, e.g. 1000 Hz.
Yes we may add a ratelimitation inside schedule() but if we only do that it will only be a plaster. It will still use a huge amount of CPU just for reading the time and checking the limit.
A caller shouldn't call schedule() million of times per second, I think the root cause need to be fixed anyway. LZMA has this problem, ZLIB has not. I didn't check other decompressors.
The best would be that schedule() throw a WARN_ONCE() when that happens so that we can identify frantic callers.
Christophe
Hi Christophe,
On 7/7/23 11:27, Christophe Leroy wrote:
Le 07/07/2023 à 11:09, Stefan Roese a écrit :
[Vous ne recevez pas souvent de courriers de sr@denx.de. Découvrez pourquoi ceci est important à https://aka.ms/LearnAboutSenderIdentification ]
Hi Simon, Hi Christophe,
On 7/6/23 17:57, Simon Glass wrote:
On Wed, 5 Jul 2023 at 09:54, Christophe Leroy christophe.leroy@csgroup.eu wrote:
Uncompressing a 1.7Mbytes FIT image on U-boot 2023.04 takes approx 7s on a powerpc 8xx. The same on U-boot 2023.07-rc6 takes approx 28s unless watchdog is disabled.
During that decompression, LzmaDec_DecodeReal() calls schedule 1.6 million times, that is every 4µs in average.
In the past it used to be a call to WATCHDOG_RESET() which was just calling hw_watchdog_reset().
But the combination of commit 29caf9305b6 ("cyclic: Use schedule() instead of WATCHDOG_RESET()") and commit 26e8ebcd7cb ("watchdog: mpc8xxx: Make it generic") results in an heavier processing.
However, there is absolutely no point in calling schedule() that often.
By moving and keeping only one call to schedule() in the main loop the number of calls is reduced to 1.2 million which is still too much. So add logic to only call schedule every 1024 times. That leads to a call to schedule approx every 6ms which is still far enough to entertain the watchdog which has a 1s timeout on powerpc 8xx.
powerpc 8xx being one of the slowest targets we have today in U-boot, and most other watchdogs having a timeout of one minutes instead of one second like the 8xx, this fix should not have negative impact on other targets.
Fixes: 29caf9305b6 ("cyclic: Use schedule() instead of WATCHDOG_RESET()") Signed-off-by: Christophe Leroy christophe.leroy@csgroup.eu
lib/lzma/LzmaDec.c | 18 ++++-------------- 1 file changed, 4 insertions(+), 14 deletions(-)
Reviewed-by: Simon Glass sjg@chromium.org
+Stefan Roese I wonder if we need a more general fix?
IIRC, we already have some code in the watchdog IF making sure, that the WDT reset is not called too often. I need to re-check here. And yes, we should probably either move this code to the common schedule() function or add it here. Best by adding a new Kconfig option, configuring the max schedule frequency, e.g. 1000 Hz.
Yes we may add a ratelimitation inside schedule() but if we only do that it will only be a plaster. It will still use a huge amount of CPU just for reading the time and checking the limit.
A caller shouldn't call schedule() million of times per second, I think the root cause need to be fixed anyway. LZMA has this problem, ZLIB has not. I didn't check other decompressors.
Agreed. We should do both, change schedule() and also fix the high frequent callers.
The best would be that schedule() throw a WARN_ONCE() when that happens so that we can identify frantic callers.
Good idea. Let me see if I can come up with such a patch. Will take a few days though as I'm traveling the next few days.
Thanks, Stefan
On Wed, Jul 05, 2023 at 10:34:26AM +0200, Christophe Leroy wrote:
Uncompressing a 1.7Mbytes FIT image on U-boot 2023.04 takes approx 7s on a powerpc 8xx. The same on U-boot 2023.07-rc6 takes approx 28s unless watchdog is disabled.
During that decompression, LzmaDec_DecodeReal() calls schedule 1.6 million times, that is every 4µs in average.
In the past it used to be a call to WATCHDOG_RESET() which was just calling hw_watchdog_reset().
But the combination of commit 29caf9305b6 ("cyclic: Use schedule() instead of WATCHDOG_RESET()") and commit 26e8ebcd7cb ("watchdog: mpc8xxx: Make it generic") results in an heavier processing.
However, there is absolutely no point in calling schedule() that often.
By moving and keeping only one call to schedule() in the main loop the number of calls is reduced to 1.2 million which is still too much. So add logic to only call schedule every 1024 times. That leads to a call to schedule approx every 6ms which is still far enough to entertain the watchdog which has a 1s timeout on powerpc 8xx.
powerpc 8xx being one of the slowest targets we have today in U-boot, and most other watchdogs having a timeout of one minutes instead of one second like the 8xx, this fix should not have negative impact on other targets.
Fixes: 29caf9305b6 ("cyclic: Use schedule() instead of WATCHDOG_RESET()") Signed-off-by: Christophe Leroy christophe.leroy@csgroup.eu Reviewed-by: Simon Glass sjg@chromium.org
Applied to u-boot/master, thanks!
participants (4)
-
Christophe Leroy -
Simon Glass -
Stefan Roese -
Tom Rini