From c0053bd50af57c4ebf032a9de1b07ca78c982452 Mon Sep 17 00:00:00 2001 From: Nishanth Menon Date: Thu, 6 Aug 2015 10:54:24 -0500 Subject: ARM: OMAP5 / DRA7: Introduce workaround for 801819 MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit Add workaround for Cortex-A15 ARM erratum 801819 which says in summary that "A livelock can occur in the L2 cache arbitration that might prevent a snoop from completing. Under certain conditions this can cause the system to deadlock. " Recommended workaround is as follows: Do both of the following: 1) Do not use the write-back no-allocate memory type. 2) Do not issue write-back cacheable stores at any time when the cache is disabled (SCTLR.C=0) and the MMU is enabled (SCTLR.M=1). Because it is implementation defined whether cacheable stores update the cache when the cache is disabled it is not expected that any portable code will execute cacheable stores when the cache is disabled. For implementations of Cortex-A15 configured without the “L2 arbitration register slice” option (typically one or two core systems), you must also do the following: 3) Disable write-streaming in each CPU by setting ACTLR[28:25] = 0b1111 So, we provide an option to disable write streaming on OMAP5 and DRA7. It is a rare condition to occur and may be enabled selectively based on platform acceptance of risk. Applies to: A15 revisions r2p0, r2p1, r2p2, r2p3 or r2p4 and REVIDR[3] is set to 0. Based on ARM errata Document revision 18.0 (22 Nov 2013) Note: the configuration for the workaround needs to be done with each CPU bringup, since CPU0 bringup is done by bootloader, it is recommended to have the workaround in the bootloader, kernel also does ensure that CPU0 has the workaround and makes the workaround active when CPU1 gets active. With CONFIG_SMP disabled, it is expected to be done by the bootloader. This does show significant degradation in synthetic tests such as mbw (https://packages.qa.debian.org/m/mbw.html) mbw -n 100 100|grep AVG (on a test platform) Without enabling the erratum: AVG Method: MEMCPY Elapsed: 0.13406 MiB: 100.00000 Copy: 745.913 MiB/s AVG Method: DUMB Elapsed: 0.06746 MiB: 100.00000 Copy: 1482.357 MiB/s AVG Method: MCBLOCK Elapsed: 0.03058 MiB: 100.00000 Copy: 3270.569 MiB/s After enabling the erratum: AVG Method: MEMCPY Elapsed: 0.13757 MiB: 100.00000 Copy: 726.913 MiB/s AVG Method: DUMB Elapsed: 0.12024 MiB: 100.00000 Copy: 831.668 MiB/s AVG Method: MCBLOCK Elapsed: 0.09243 MiB: 100.00000 Copy: 1081.942 MiB/s Most benchmarks are designed for specific performance analysis, so overall usecase must be considered before making a decision to enable/disable the erratum workaround. Pending internal investigation, the erratum is kept disabled by default. Cc: Russell King Cc: Catalin Marinas Cc: Tony Lindgren Suggested-by: Richard Woodruff Suggested-by: Brad Griffis Signed-off-by: Nishanth Menon Signed-off-by: Tony Lindgren --- arch/arm/mach-omap2/Kconfig | 8 ++++++++ 1 file changed, 8 insertions(+) (limited to 'arch/arm/mach-omap2/Kconfig') diff --git a/arch/arm/mach-omap2/Kconfig b/arch/arm/mach-omap2/Kconfig index 0517f0c1581a..a63d3fe2ca46 100644 --- a/arch/arm/mach-omap2/Kconfig +++ b/arch/arm/mach-omap2/Kconfig @@ -240,4 +240,12 @@ endmenu endif +config OMAP5_ERRATA_801819 + bool "Errata 801819: An eviction from L1 data cache might stall indefinitely" + depends on SOC_OMAP5 || SOC_DRA7XX + help + A livelock can occur in the L2 cache arbitration that might prevent + a snoop from completing. Under certain conditions this can cause the + system to deadlock. + endmenu -- cgit