Initial cleanup of arm port timer subsystem
Subsystems (2)
cover Initial cleanup of arm port timer subsystem
From: Andrew Ruder <[email protected]>
Hey all,
Was tracking down an issue on a PXA270 based board where a long y-modem transfer would frequently deadlock the board in the middle of the transfer. Unfortunately I felt like I had opened pandora's box by looking around at other ports.
Very few ports were 32-bit rollover safe! Most looked like they would just skip a __udelay on roll-over. These are fixes for the couple I noticed that would potentially deadlock if __udelay crossed the 32-bit boundary. And some are so complicated that I didn't waste time looking at them.
get_ticks() which returns a unsigned long long had a variety of implementations. Some returned a number that wrapped at 32 bits which worked with some implementations of __udelay. Others returned a number that wraps at a fraction of 32-bits which would hit the roll-over issues at an unexpected place and fail even more often.
There is a lot of opportunity here for tons of code removal by moving more ports to the common code. Here are some notes that I took while going through some that might save someone some time. This was a very cursory look so my apologies if I have missed something.
** u-boot/arch/arm/cpu/arm1136/mx31/timer.c:98:20:unsigned long long get_ticks(void) get_ticks() returns 32-bit number__udelay() depends on 64-bit value - HANGS
** u-boot/arch/arm/cpu/arm1136/mx35/timer.c:75:20:unsigned long long get_ticks(void) get_ticks() returns 32-bit number__udelay() depends on 64-bit value - HANGS
** u-boot/arch/arm/cpu/arm1176/bcm2835/timer.c:37:20:unsigned long long get_ticks(void) get_ticks() returns 32-bit number__udelay() is safe
** u-boot/arch/arm/cpu/arm1176/tnetv107x/timer.c:67:20:unsigned long long get_ticks(void) get_ticks() isn't even a 32-bit number__udelay() is safe
** u-boot/arch/arm/cpu/arm920t/a320/timer.c:74:20:unsigned long long get_ticks(void) get_ticks() returns 64-bit number__udelay() is safe
** u-boot/arch/arm/cpu/arm920t/at91/timer.c:115:20:unsigned long long get_ticks(void) get_ticks() isn't even a 32-bit number__udelay() is safe
** u-boot/arch/arm/cpu/arm920t/ep93xx/timer.c:58:20:unsigned long long get_ticks(void) get_ticks() returns 64-bit number__udelay() is safe until 64-bit rollover
** u-boot/arch/arm/cpu/arm920t/imx/timer.c:70:20:unsigned long long get_ticks(void) get_ticks() returns a 32-bit number__udelay is safe
** u-boot/arch/arm/cpu/arm920t/s3c24x0/timer.c:110:20:unsigned long long get_ticks(void) get_ticks() is complicated. tbu abused for something else__udelay might be safe? It handles 32-bit rollover correctly if get_ticks() is really a 32-bit number
** u-boot/arch/arm/cpu/arm926ejs/armada100/timer.c:182:20:unsigned long long get_ticks(void) get_ticks() isn't a 32-bit number__udelay completely doesn't handle roll-over
** u-boot/arch/arm/cpu/arm926ejs/at91/timer.c:75:20:unsigned long long get_ticks(void) get_ticks() returns 64-bit number__udelay handles roll-over
looks perfect
** u-boot/arch/arm/cpu/arm926ejs/davinci/timer.c:56:20:unsigned long long get_ticks(void) get_ticks() returns 64-bit number
chooses to ignore roll-over
** u-boot/arch/arm/cpu/arm926ejs/kirkwood/timer.c:145:20:unsigned long long get_ticks(void) get_ticks() returns 32-bit number
no idea, too complicated
** u-boot/arch/arm/cpu/arm926ejs/mb86r0x/timer.c:65:20:unsigned long long get_ticks(void) get_ticks() returns 32-bit number__udelay doesn't handle roll-over
** u-boot/arch/arm/cpu/arm926ejs/lpc32xx/timer.c:74:20:unsigned long long get_ticks(void) get_ticks() returns 32-bit number__udelay is safe (doesn't handle roll-over - doesn't need to)
** u-boot/arch/arm/cpu/arm926ejs/mx27/timer.c:111:20:unsigned long long get_ticks(void) get_ticks() returns 32-bit number__udelay broken
** u-boot/arch/arm/cpu/arm926ejs/mxs/timer.c:81:20:unsigned long long get_ticks(void) get_ticks() returns 32-bit number__udelay complicated
** u-boot/arch/arm/cpu/arm926ejs/nomadik/timer.c:63:20:unsigned long long get_ticks(void) get_ticks() isn't a 32-bit number__udelay complicated
** u-boot/arch/arm/cpu/arm926ejs/omap/timer.c:140:20:unsigned long long get_ticks(void) get_ticks() isn't a 32-bit number__udelay looks good
** u-boot/arch/arm/cpu/arm926ejs/orion5x/timer.c:159:20:unsigned long long get_ticks(void) get_ticks() returns 32-bit number__udelay complicated
** u-boot/arch/arm/cpu/arm926ejs/pantheon/timer.c:189:20:unsigned long long get_ticks(void) get_ticks() doesn't return a 32-bit number__udelay doesn't handle roll-over
** u-boot/arch/arm/cpu/arm926ejs/spear/timer.c:111:20:unsigned long long get_ticks(void) get_ticks() doesn't return a 32-bit number__udelay looks good
** u-boot/arch/arm/cpu/armv7/arch_timer.c:24:20:unsigned long long get_ticks(void) get_ticks() returns 64-bit number__udelay doesn't handle 64-bit roll-over
** u-boot/arch/arm/cpu/armv7/at91/timer.c:78:20:unsigned long long get_ticks(void) get_ticks() returns 64-bit number__udelay looks good
looks perfect
** u-boot/arch/arm/cpu/armv7/omap-common/timer.c:96:20:unsigned long long get_ticks(void) get_ticks() returns 32-bit number__udelay complicated but looks good
** u-boot/arch/arm/cpu/armv7/rmobile/timer.c:77:20:unsigned long long get_ticks(void) get_ticks() returns a 64-bit number__udelay looks good
** u-boot/arch/arm/cpu/armv7/s5p-common/timer.c:120:20:unsigned long long get_ticks(void) get_ticks() doesn't return a 32-bit number__udelay looks good
** u-boot/arch/arm/cpu/armv7/sunxi/timer.c:101:20:unsigned long long get_ticks(void) get_ticks() returns 32-bit number__udelay off by 1 on roll-over
** u-boot/arch/arm/cpu/armv7/vf610/timer.c:49:20:unsigned long long get_ticks(void) get_ticks() returns 64-bit number__udelay looks good
** u-boot/arch/arm/cpu/armv7/u8500/timer.c:123:20:unsigned long long get_ticks(void) get_ticks() returns 32-bit number__udelay off by 1 on rollover
** u-boot/arch/arm/cpu/armv7/zynq/timer.c:154:20:unsigned long long get_ticks(void) get_ticks() returns 32-bit number__udelay complicated
** u-boot/arch/arm/cpu/pxa/timer.c:45:20:unsigned long long get_ticks(void) get_ticks() returns a 32-bit number__udelay depends on 64-bit - HANGS
** u-boot/arch/arm/cpu/sa1100/timer.c:58:20:unsigned long long get_ticks(void) get_ticks() returns a 32-bit number__udelay looks good
** u-boot/arch/arm/imx-common/timer.c:74:20:unsigned long long get_ticks(void) get_ticks() returns 64-bit number__udelay doesn't handle 64-bit roll-over
** u-boot/arch/avr32/cpu/interrupts.c:36:20:unsigned long long get_ticks(void) get_ticks() returns 64-bit number__udelay() correctly handles overflow
** u-boot/arch/blackfin/cpu/interrupts.c:39:20:unsigned long long get_ticks(void) get_ticks() returns a 32-bit number__udelay() looks ok i think (pretty complicated)
** u-boot/arch/m68k/lib/time.c:179:20:unsigned long long get_ticks(void) get_ticks() returns a 32-bit number__udelay() looks ok
** u-boot/arch/microblaze/cpu/timer.c:80:20:unsigned long long get_ticks(void) get_ticks() returns a 32-bit number__udelay() looks ok.
** u-boot/arch/mips/cpu/mips32/time.c:58:20:unsigned long long get_ticks(void) get_ticks() returns a 32-bit number__udelay() complicated
** u-boot/arch/mips/cpu/mips64/time.c:58:20:unsigned long long get_ticks(void) get_ticks() might return 64-bit__udelay() complicated
** u-boot/arch/nds32/cpu/n1213/ag101/timer.c:173:20:unsigned long long get_ticks(void) get_ticks() returns 32-bit number__udelay() off-by-1 on roll-over
** u-boot/arch/nds32/cpu/n1213/ag102/timer.c:173:20:unsigned long long get_ticks(void) see above - same code
Andrew Ruder (3):
arm: pxa: use common timer functions
arm: mx31: use common timer functions
arm: mx35: use common timer functions
arch/arm/cpu/arm1136/mx31/timer.c | 104 +----------------------------- arch/arm/cpu/arm1136/mx35/timer.c | 83 ------------------------ arch/arm/cpu/pxa/timer.c | 69 +------------------- arch/arm/include/asm/arch-mx31/imx-regs.h | 10 +++ arch/arm/include/asm/arch-mx35/imx-regs.h | 12 ++++ include/configs/pxa-common.h | 13 ++++ 6 files changed, 38 insertions(+), 253 deletions(-)
Signature · 1 line
2.0.1
PATCH 1/3 arm: pxa: use common timer functions
From: Andrew Ruder <[email protected]>
This patch moves pxa to the common timer functions added in commit
8dfafdd - Introduce common timer functions <Rob Herring>
The (removed) pxa timer code (specifically __udelay()) could deadlock at the 32-bit boundary of get_ticks(). get_ticks() returned a 32-bit value cast up to a 64-bit value. If get_ticks() + tmo in __udelay() crossed the 32-bit boundary, the while condition became unconditionally true and locked the processor. Rather than patch the specific pxa issues, simply move everything over to the common code.
---
32-bit rollover occurs every 22 minutes so even a long y-modem transfer was enough to hit this issue fairly regularly. This has been tested.
arch/arm/cpu/pxa/timer.c | 69 +------------------------------------------- include/configs/pxa-common.h | 13 +++++++++ 2 files changed, 14 insertions(+), 68 deletions(-)
arch/arm/cpu/pxa/timer.c +1 -68
@@ -6,80 +6,13 @@* SPDX-License-Identifier: GPL-2.0+ */ -#include <asm/arch/pxa-regs.h> #include <asm/io.h> #include <common.h> -#include <div64.h> DECLARE_GLOBAL_DATA_PTR ; -#define TIMER_LOAD_VAL 0xffffffff - -#define timestamp (gd->arch.tbl) -#define lastinc (gd->arch.lastinc) - -#if defined(CONFIG_CPU_PXA27X ) || defined(CONFIG_CPU_MONAHANS ) -#define TIMER_FREQ_HZ 3250000 -#elif defined(CONFIG_CPU_PXA25X ) -#define TIMER_FREQ_HZ 3686400 -#else -#error "Timer frequency unknown - please config PXA CPU type" -#endif - -static unsigned long long tick_to_time (unsigned long long tick) -{ -return lldiv (tick *CONFIG_SYS_HZ ,TIMER_FREQ_HZ ); -} - -static unsigned long long us_to_tick (unsigned long long us) -{ -return lldiv (us *TIMER_FREQ_HZ ,1000000 ); -} -int timer_init (void ) { -writel (0 ,OSCR ); +writel (0 ,CONFIG_SYS_TIMER_COUNTER );return 0 ; } - -unsigned long long get_ticks (void ) -{ -/* Current tick value */ -uint32_t now =readl (OSCR ); - -if (now >= lastinc) { -/* -* Normal mode (non roll) -* Move stamp forward with absolute diff ticks -*/ - timestamp += (now - lastinc); - }else { -/* We have rollover of incrementer */ - timestamp += (TIMER_LOAD_VAL - lastinc) + now; - } - - lastinc = now; -return timestamp; -} - -ulong get_timer (ulong base) -{ -return tick_to_time (get_ticks ()) - base; -} - -void __udelay (unsigned long usec) -{ -unsigned long long tmp; -ulong tmo; - - tmo =us_to_tick (usec); - tmp =get_ticks () + tmo;/* get current timestamp */ - -while (get_ticks () < tmp)/* loop till event */ -/*NOP*/ ; -} - -ulong get_tbclk (void ) -{ -return TIMER_FREQ_HZ ; -}
include/configs/pxa-common.h +13 -0
@@ -43,4 +43,17 @@#define CONFIG_USB_STORAGE #endif +/* +* Generic timer support +*/ +#if defined(CONFIG_CPU_PXA27X ) || defined(CONFIG_CPU_MONAHANS ) +#define CONFIG_SYS_TIMER_RATE 3250000 +#elif defined(CONFIG_CPU_PXA25X ) +#define CONFIG_SYS_TIMER_RATE 3686400 +#else +#error "Timer frequency unknown - please config PXA CPU type" +#endif + +#define CONFIG_SYS_TIMER_COUNTER 0x40A00010/* OSCR */ +#endif /* __CONFIG_PXA_COMMON_H__ */
Signature · 1 line
2.0.1
PATCH 1/3 arm: pxa: use common timer functions
Quoted 17 lines · andrew.ruder at elecsyscorp.com
From: Andrew Ruder <[email protected]>
This patch moves pxa to the common timer functions added in commit
8dfafdd - Introduce common timer functions <Rob Herring>
The (removed) pxa timer code (specifically
__udelay()) could deadlock at the 32-bit boundary ofget_ticks().get_ticks()returned a 32-bit value cast up to a 64-bit value. Ifget_ticks()+ tmo in__udelay()crossed the 32-bit boundary, the while condition became unconditionally true and locked the processor. Rather than patch the specific pxa issues, simply move everything over to the common code.Signed-off-by Andrew Ruder <[email protected]>Cc Marek Vasut <[email protected]>---
Quoted 8 lines
32-bit rollover occurs every 22 minutes so even a long y-modem transfer was enough to hit this issue fairly regularly. This has been tested.
arch/arm/cpu/pxa/timer.c | 69 +------------------------------------------- include/configs/pxa-common.h | 13 +++++++++ 2 files changed, 14 insertions(+), 68 deletions(-)
+CC Albert.
Albert , can you please pick this one up ?
Best regards,
Marek Vasut
PATCH 1/3 arm: pxa: use common timer functions
In reply to andrew.ruder at elecsyscorp.comQuoted 13 lines · Andrew Ruder
From: Andrew Ruder <[email protected]>
This patch moves pxa to the common timer functions added in commit
8dfafdd - Introduce common timer functions <Rob Herring>
The (removed) pxa timer code (specifically
__udelay()) could deadlock at the 32-bit boundary ofget_ticks().get_ticks()returned a 32-bit value cast up to a 64-bit value. Ifget_ticks()+ tmo in__udelay()crossed the 32-bit boundary, the while condition became unconditionally true and locked the processor. Rather than patch the specific pxa issues, simply move everything over to the common code.
Quoted 3 lines
Signed-off-by Andrew Ruder <[email protected]>Cc Marek Vasut <[email protected]>Acked-by Marek Vasut <[email protected]>
As it stands today this breaks building a number of PXA boards so it needs to be picked up and re-worked / build tested, thanks!
Signature · 8 lines
Tom -------------- next part -------------- A non-text attachment was scrubbed... Name: not available Type: application/pgp-signature Size: 836 bytes Desc: Digital signature URL: <http://lists.denx.de/pipermail/u-boot/attachments/20140830/7d7877e9/attachment.pgp>
PATCH 1/3 arm: pxa: use common timer functions
Quoted 2 lines · Tom Rini
As it stands today this breaks building a number of PXA boards so it needs to be picked up and re-worked / build tested, thanks!
I'm still planning on getting back to this and taking a look. At the time I was on a slightly older version of U-Boot (2014.07-rc3 I think) and hadn't yet figured out what was going on with all of the Kconfig changes so mostly just made sure it applied cleanly on HEAD. My apologies!
- Andy
PATCH 2/3 arm: mx31: use common timer functions
In reply to andrew.ruder at elecsyscorp.comFrom: Andrew Ruder <[email protected]>
This patch moves mx31 to the common timer functions added in commit
8dfafdd - Introduce common timer functions <Rob Herring>
The (removed) mx31 timer code (specifically __udelay()) could deadlock at the 32-bit boundary of get_ticks(). get_ticks() returned a 32-bit value cast up to a 64-bit value. If get_ticks() + tmo in __udelay() crossed the 32-bit boundary, the while condition became unconditionally true and locks the processor. Rather than patch the specific mx31 issues, simply move everything over to the common code.
---
This patch has been COMPILE tested only. The situation isn't quite as bad on mx31 as 32-bit rollover occurs in about a day and a half.
Before patch:
Configuring for qong board...
456 -rw-r--r-- 1 andy andy 463236 Aug 12 08:36 u-boot.bin
After patch:
Configuring for qong board...
456 -rw-r--r-- 1 andy andy 463248 Aug 12 08:38 u-boot.bin
arch/arm/cpu/arm1136/mx31/timer.c | 104 +----------------------------- arch/arm/include/asm/arch-mx31/imx-regs.h | 10 +++ 2 files changed, 12 insertions(+), 102 deletions(-)
arch/arm/cpu/arm1136/mx31/timer.c +2 -102
@@ -7,9 +7,6 @@#include <common.h> #include <asm/arch/imx-regs.h> -#include <asm/arch/clock.h> -#include <div64.h> -#include <watchdog.h> #include <asm/io.h> #define TIMER_BASE 0x53f90000/* General purpose timer 1 */
@@ -28,57 +25,6 @@DECLARE_GLOBAL_DATA_PTR ; -/* -* "time" is measured in 1 / CONFIG_SYS_HZ seconds, -* "tick" is internal timer period -*/ - -#ifdef CONFIG_MX31_TIMER_HIGH_PRECISION -/* ~0.4% error - measured with stop-watch on 100s boot-delay */ -static inline unsigned long long tick_to_time (unsigned long long tick) -{ - tick *=CONFIG_SYS_HZ ; -do_div (tick,MXC_CLK32 ); -return tick; -} - -static inline unsigned long long time_to_tick (unsigned long long time) -{ - time *=MXC_CLK32 ; -do_div (time,CONFIG_SYS_HZ ); -return time; -} - -static inline unsigned long long us_to_tick (unsigned long long us) -{ - us = us *MXC_CLK32 +999999 ; -do_div (us,1000000 ); -return us; -} -#else -/* ~2% error */ -#define TICK_PER_TIME ((MXC_CLK32 + CONFIG_SYS_HZ / 2) / CONFIG_SYS_HZ) -#define US_PER_TICK (1000000 / MXC_CLK32) - -static inline unsigned long long tick_to_time (unsigned long long tick) -{ -do_div (tick,TICK_PER_TIME ); -return tick; -} - -static inline unsigned long long time_to_tick (unsigned long long time) -{ -return time *TICK_PER_TIME ; -} - -static inline unsigned long long us_to_tick (unsigned long long us) -{ - us +=US_PER_TICK -1 ; -do_div (us,US_PER_TICK ); -return us; -} -#endif -/* The 32768Hz 32-bit timer overruns in 131072 seconds */ int timer_init (void ) {
@@ -95,53 +41,7 @@ int timer_init(void)return 0 ; } -unsigned long long get_ticks (void ) -{ -ulong now =GPTCNT ;/* current tick value */ - -if (now >= gd->arch.lastinc)/* normal mode (non roll) */ -/* move stamp forward with absolut diff ticks */ - gd->arch.tbl += (now - gd->arch.lastinc); -else /* we have rollover of incrementer */ - gd->arch.tbl += (0xFFFFFFFF - gd->arch.lastinc) + now; - gd->arch.lastinc = now; -return gd->arch.tbl; -} - -ulong get_timer_masked (void ) -{ -/* -* get_ticks() returns a long long (64 bit), it wraps in -* 2^64 / MXC_CLK32 = 2^64 / 2^15 = 2^49 ~ 5 * 10^14 (s) ~ -* 5 * 10^9 days... and get_ticks() * CONFIG_SYS_HZ wraps in -* 5 * 10^6 days - long enough. -*/ -return tick_to_time (get_ticks ()); -} - -ulong get_timer (ulong base) -{ -return get_timer_masked () - base; -} - -/* delay x useconds AND preserve advance timestamp value */ -void __udelay (unsigned long usec) -{ -unsigned long long tmp; -ulong tmo; - - tmo =us_to_tick (usec); - tmp =get_ticks () + tmo;/* get current timestamp */ - -while (get_ticks () < tmp)/* loop till event */ -/*NOP*/ ; -} - -/* -* This function is derived from PowerPC code (timebase clock frequency). -* On ARM it returns the number of timer ticks per second. -*/ -ulong get_tbclk (void ) +unsigned long timer_read_counter (void ) { -return MXC_CLK32 ; +return GPTCNT ; }
arch/arm/include/asm/arch-mx31/imx-regs.h +10 -0
@@ -909,9 +909,19 @@ struct esdc_regs {#define MXC_CSPIPERIOD_32KHZ (1 << 15)#define MAX_SPI_BYTES 4 +#define MXC_SPI_BASE_ADDRESSES \ 0x43fa4000, \ 0x50010000, \ 0x53f84000, +/* +* Generic timer support +*/ +#ifdef CONFIG_MX31_CLK32 +#define CONFIG_SYS_TIMER_RATE CONFIG_MX31_CLK32 +#else +#define CONFIG_SYS_TIMER_RATE 32768 +#endif +#endif /* __ASM_ARCH_MX31_IMX_REGS_H */
Signature · 1 line
2.0.1
PATCH 2/3 arm: mx31: use common timer functions
Quoted 13 lines · andrew.ruder at elecsyscorp.com
From: Andrew Ruder <[email protected]>
This patch moves mx31 to the common timer functions added in commit
8dfafdd - Introduce common timer functions <Rob Herring>
The (removed) mx31 timer code (specifically
__udelay()) could deadlock at the 32-bit boundary ofget_ticks().get_ticks()returned a 32-bit value cast up to a 64-bit value. Ifget_ticks()+ tmo in__udelay()crossed the 32-bit boundary, the while condition became unconditionally true and locks the processor. Rather than patch the specific mx31 issues, simply move everything over to the common code.
Quoted 6 lines
Signed-off-by Andrew Ruder <[email protected]>Cc Marek Vasut <[email protected]>Cc Linus Walleij <[email protected]>Cc Wolfgang Denk <[email protected]>Cc Fabio Estevam <[email protected]>Cc Helmut Raiger <[email protected]>
+CC Stefano
Quoted 170 lines
---
This patch has been COMPILE tested only. The situation isn't quite as bad on mx31 as 32-bit rollover occurs in about a day and a half.
Before patch:
Configuring for qong board...
456 -rw-r--r-- 1 andy andy 463236 Aug 12 08:36 u-boot.binAfter patch:
Configuring for qong board...
456 -rw-r--r-- 1 andy andy 463248 Aug 12 08:38 u-boot.binarch/arm/cpu/arm1136/mx31/timer.c | 104 +----------------------------- arch/arm/include/asm/arch-mx31/imx-regs.h | 10 +++ 2 files changed, 12 insertions(+), 102 deletions(-)(quoted diff) +0
-0
b/arch/arm/cpu/arm1136/mx31/timer.cindex f111242..3a81ce4 100644arch/arm/cpu/arm1136/mx31/timer.c +2
-102@@ -7,9 +7,6 @@#include <common.h> #include <asm/arch/imx-regs.h> -#include <asm/arch/clock.h> -#include <div64.h> -#include <watchdog.h> #include <asm/io.h> #define TIMER_BASE 0x53f90000/* General purpose timer 1 */ @@ -28,57 +25,6 @@DECLARE_GLOBAL_DATA_PTR ; -/* -* "time" is measured in 1 / CONFIG_SYS_HZ seconds, -* "tick" is internal timer period -*/ - -#ifdef CONFIG_MX31_TIMER_HIGH_PRECISION -/* ~0.4% error - measured with stop-watch on 100s boot-delay */ -static inline unsigned long long tick_to_time (unsigned long long tick) -{ - tick *=CONFIG_SYS_HZ ; -do_div (tick,MXC_CLK32 ); -return tick; -} - -static inline unsigned long long time_to_tick (unsigned long long time) -{ - time *=MXC_CLK32 ; -do_div (time,CONFIG_SYS_HZ ); -return time; -} - -static inline unsigned long long us_to_tick (unsigned long long us) -{ - us = us *MXC_CLK32 +999999 ; -do_div (us,1000000 ); -return us; -} -#else -/* ~2% error */ -#define TICK_PER_TIME ((MXC_CLK32 + CONFIG_SYS_HZ / 2) / CONFIG_SYS_HZ) -#define US_PER_TICK (1000000 / MXC_CLK32) - -static inline unsigned long long tick_to_time (unsigned long long tick) -{ -do_div (tick,TICK_PER_TIME ); -return tick; -} - -static inline unsigned long long time_to_tick (unsigned long long time) -{ -return time *TICK_PER_TIME ; -} - -static inline unsigned long long us_to_tick (unsigned long long us) -{ - us +=US_PER_TICK -1 ; -do_div (us,US_PER_TICK ); -return us; -} -#endif -/* The 32768Hz 32-bit timer overruns in 131072 seconds */ int timer_init (void ) {@@ -95,53 +41,7 @@ int timer_init(void)return 0 ; } -unsigned long long get_ticks (void ) -{ -ulong now =GPTCNT ;/* current tick value */ - -if (now >= gd->arch.lastinc)/* normal mode (non roll) */ -/* move stamp forward with absolut diff ticks */ - gd->arch.tbl += (now - gd->arch.lastinc); -else /* we have rollover of incrementer */ - gd->arch.tbl += (0xFFFFFFFF - gd->arch.lastinc) + now; - gd->arch.lastinc = now; -return gd->arch.tbl; -} - -ulong get_timer_masked (void ) -{ -/* -* get_ticks() returns a long long (64 bit), it wraps in -* 2^64 / MXC_CLK32 = 2^64 / 2^15 = 2^49 ~ 5 * 10^14 (s) ~ -* 5 * 10^9 days... and get_ticks() * CONFIG_SYS_HZ wraps in -* 5 * 10^6 days - long enough. -*/ -return tick_to_time (get_ticks ()); -} - -ulong get_timer (ulong base) -{ -return get_timer_masked () - base; -} - -/* delay x useconds AND preserve advance timestamp value */ -void __udelay (unsigned long usec) -{ -unsigned long long tmp; -ulong tmo; - - tmo =us_to_tick (usec); - tmp =get_ticks () + tmo;/* get current timestamp */ - -while (get_ticks () < tmp)/* loop till event */ -/*NOP*/ ; -} - -/* -* This function is derived from PowerPC code (timebase clock frequency). -* On ARM it returns the number of timer ticks per second. -*/ -ulong get_tbclk (void ) +unsigned long timer_read_counter (void ) { -return MXC_CLK32 ; +return GPTCNT ; }(quoted diff) +0
-0
b/arch/arm/include/asm/arch-mx31/imx-regs.hindex f23350e..71ebd24 100644arch/arm/include/asm/arch-mx31/imx-regs.h +10
-0@@ -909,9 +909,19 @@ struct esdc_regs {#define MXC_CSPIPERIOD_32KHZ (1 << 15)#define MAX_SPI_BYTES 4 +#define MXC_SPI_BASE_ADDRESSES \ 0x43fa4000, \ 0x50010000, \ 0x53f84000, +/* +* Generic timer support +*/ +#ifdef CONFIG_MX31_CLK32 +#define CONFIG_SYS_TIMER_RATE CONFIG_MX31_CLK32 +#else +#define CONFIG_SYS_TIMER_RATE 32768 +#endif +#endif /* __ASM_ARCH_MX31_IMX_REGS_H */
PATCH 3/3 arm: mx35: use common timer functions
In reply to andrew.ruder at elecsyscorp.comFrom: Andrew Ruder <[email protected]>
This patch moves mx35 to the common timer functions added in commit
8dfafdd - Introduce common timer functions <Rob Herring>
The (removed) mx35 timer code (specifically __udelay()) could deadlock at the 32-bit boundary of get_ticks(). get_ticks() returned a 32-bit value cast up to a 64-bit value. If get_ticks() + tmo in __udelay() crossed the 32-bit boundary, the while condition became unconditionally true and locks the processor. Rather than patch the specific mx35 issues, simply move everything over to the common code.
---
This patch has been COMPILE tested only. The situation isn't quite as bad on mx35 as 32-bit rollover occurs in about a day and a half.
Before patch:
Configuring for woodburn board...
308 -rw-r--r-- 1 andy andy 314232 Aug 12 08:36 u-boot.bin
After patch:
Configuring for woodburn board...
308 -rw-r--r-- 1 andy andy 314208 Aug 12 08:37 u-boot.bin
arch/arm/cpu/arm1136/mx35/timer.c | 83 ------------------------------- arch/arm/include/asm/arch-mx35/imx-regs.h | 12 +++++ 2 files changed, 12 insertions(+), 83 deletions(-)
arch/arm/cpu/arm1136/mx35/timer.c +0 -83
@@ -9,16 +9,11 @@#include <common.h> #include <asm/io.h> -#include <div64.h> #include <asm/arch/imx-regs.h> #include <asm/arch/crm_regs.h> -#include <asm/arch/clock.h> DECLARE_GLOBAL_DATA_PTR ; -#define timestamp (gd->arch.tbl) -#define lastinc (gd->arch.lastinc) -/* General purpose timers bitfields */ #define GPTCR_SWR (1<<15)/* Software reset */ #define GPTCR_FRR (1<<9)/* Freerun / restart */
@@ -26,27 +21,6 @@ DECLARE_GLOBAL_DATA_PTR;#define GPTCR_TEN (1)/* Timer enable */ /* -* "time" is measured in 1 / CONFIG_SYS_HZ seconds, -* "tick" is internal timer period -*/ -/* ~0.4% error - measured with stop-watch on 100s boot-delay */ -static inline unsigned long long tick_to_time (unsigned long long tick) -{ - tick *=CONFIG_SYS_HZ ; -do_div (tick,MXC_CLK32 ); - -return tick; -} - -static inline unsigned long long us_to_tick (unsigned long long us) -{ - us = us *MXC_CLK32 +999999 ; -do_div (us,1000000 ); - -return us; -} - -/* * nothing really to do with interrupts, just starts up a counter. * The 32KHz 32-bit timer overruns in 134217 seconds */
@@ -71,60 +45,3 @@ int timer_init(void)return 0 ; } - -unsigned long long get_ticks (void ) -{ -struct gpt_regs *gpt = (struct gpt_regs *)GPT1_BASE_ADDR ; -ulong now =readl (&gpt->counter);/* current tick value */ - -if (now >= lastinc) { -/* -* normal mode (non roll) -* move stamp forward with absolut diff ticks -*/ - timestamp += (now - lastinc); - }else { -/* we have rollover of incrementer */ - timestamp += (0xFFFFFFFF - lastinc) + now; - } - lastinc = now; -return timestamp; -} - -ulong get_timer_masked (void ) -{ -/* -* get_ticks() returns a long long (64 bit), it wraps in -* 2^64 / MXC_CLK32 = 2^64 / 2^15 = 2^49 ~ 5 * 10^14 (s) ~ -* 5 * 10^9 days... and get_ticks() * CONFIG_SYS_HZ wraps in -* 5 * 10^6 days - long enough. -*/ -return tick_to_time (get_ticks ()); -} - -ulong get_timer (ulong base) -{ -return get_timer_masked () - base; -} - -/* delay x useconds AND preserve advance timstamp value */ -void __udelay (unsigned long usec) -{ -unsigned long long tmp; -ulong tmo; - - tmo =us_to_tick (usec); - tmp =get_ticks () + tmo;/* get current timestamp */ - -while (get_ticks () < tmp)/* loop till event */ -/*NOP*/ ; -} - -/* -* This function is derived from PowerPC code (timebase clock frequency). -* On ARM it returns the number of timer ticks per second. -*/ -ulong get_tbclk (void ) -{ -return MXC_CLK32 ; -}
arch/arm/include/asm/arch-mx35/imx-regs.h +12 -0
@@ -372,4 +372,16 @@ struct aips_regs {#define CCM_RCSR_NF_16BIT_SEL (1 << 14)#endif + +/* +* Generic timer support +*/ +#ifdef CONFIG_MX35_CLK32 +#define CONFIG_SYS_TIMER_RATE CONFIG_MX35_CLK32 +#else +#define CONFIG_SYS_TIMER_RATE 32768 +#endif + +#define CONFIG_SYS_TIMER_COUNTER (GPT1_BASE_ADDR+36) +#endif /* __ASM_ARCH_MX35_H */
Signature · 1 line
2.0.1
PATCH 3/3 arm: mx35: use common timer functions
Hi Andrew,
Quoted 13 lines · andrew.ruder at elecsyscorp.com
From: Andrew Ruder <[email protected]>
This patch moves mx35 to the common timer functions added in commit
8dfafdd - Introduce common timer functions <Rob Herring>
The (removed) mx35 timer code (specifically
__udelay()) could deadlock at the 32-bit boundary ofget_ticks().get_ticks()returned a 32-bit value cast up to a 64-bit value. Ifget_ticks()+ tmo in__udelay()crossed the 32-bit boundary, the while condition became unconditionally true and locks the processor. Rather than patch the specific mx35 issues, simply move everything over to the common code.
Quoted 8 lines
Signed-off-by Andrew Ruder <[email protected]>Cc Marek Vasut <[email protected]>Cc Stefano Babic <[email protected]>---
This patch has been COMPILE tested only. The situation isn't quite as bad on mx35 as 32-bit rollover occurs in about a day and a half.
I have applied both patches for MX31 and MX35, even if I had no time to make myself a test. I hope in this way there is more testers before release.
Applied to u-boot-imx, thanks !
Best regards,
Stefano Babic
Signature · 5 lines
===================================================================== DENX Software Engineering GmbH, MD: Wolfgang Denk & Detlev Zundel HRB 165235 Munich, Office: Kirchenstr.5, D-82194 Groebenzell, Germany Phone: +49-8142-66989-53 Fax: +49-8142-66989-80 Email: sbabic at denx.de =====================================================================