| Message ID | 20190422161333.5022-2-vigneshr@ti.com |
|---|---|
| State | Accepted |
| Commit | add4967124f6b212fef4fa8e1e68143b1400c994 |
| Delegated to: | Tom Rini |
| Headers | show |
| Series | Add Kconfig to disable cache ops | expand |
On Mon, Apr 22, 2019 at 09:43:32PM +0530, Vignesh Raghavendra wrote: > On AM654 SoC(arm64) which is IO coherent and has L3 Cache, cache > maintenance operations being done to support non-coherent platforms > causes issues. > > For example, here is how U-Boot prepares/handles a buffer to receive > data from a device (DMA Write). This may vary slightly depending on the > driver framework: > > Start DMA to write to destination buffer > Wait for DMA to be done (dma_receive()/dma_memcpy()) > Invalidate destination buffer (invalidate_dcache_range()) > Read from destination buffer > > The invalidate after the DMA is needed in order to read latest data from > memory that’s updated by DMA write. Also, in case random prefetch has > pulled in buffer data during the “wait for DMA” before the DMA has > written to it. This works well for non-coherent architectures. > > In case of coherent architecture with L3 cache, DMA write would directly > update L3 cache contents (assuming cacheline is present in L3) without > updating the DDR memory. So invalidate after “wait for DMA” in above > sequence would discard latest data and read will cause stale data to be > fetched from DDR. Therefore invalidate after “wait for DMA” is not > always correct on coherent architecture. > > Therefore, provide a Kconfig option to disable cache maintenance ops on > coherent architectures. This has added benefit of improving the > performance of DMA transfers as we no longer need to invalidate/flush > individual cache lines(especially for buffer thats several KBs in size). > > In order to facilitate use of same Kconfig across different > architecture, I have added the symbol to top level arch/Kconfig file. > Patch currently disables cache maintenance ops for arm64 only. > flush_dcache_all() and invalidate_dcache_all() are exclusively used > during enabling/disabling dcache and hence are not disabled. > > Signed-off-by: Vignesh Raghavendra <vigneshr@ti.com> Applied to u-boot/master, thanks!
diff --git a/arch/Kconfig b/arch/Kconfig index 2f3d07c13a18..760023b19a81 100644 --- a/arch/Kconfig +++ b/arch/Kconfig @@ -227,6 +227,15 @@ config SYS_CONFIG_NAME The header file include/configs/<CONFIG_SYS_CONFIG_NAME>.h should be included from include/config.h. +config SYS_DISABLE_DCACHE_OPS + bool + help + This option disables dcache flush and dcache invalidation + operations. For example, on coherent systems where cache + operatios are not required, enable this option to avoid them. + Note that, its up to the individual architectures to implement + this functionality. + source "arch/arc/Kconfig" source "arch/arm/Kconfig" source "arch/m68k/Kconfig" diff --git a/arch/arm/cpu/armv8/cache_v8.c b/arch/arm/cpu/armv8/cache_v8.c index 038405173eb1..9ca397e73c9b 100644 --- a/arch/arm/cpu/armv8/cache_v8.c +++ b/arch/arm/cpu/armv8/cache_v8.c @@ -443,6 +443,7 @@ inline void flush_dcache_all(void) debug("flushing dcache successfully.\n"); } +#ifndef CONFIG_SYS_DISABLE_DCACHE_OPS /* * Invalidates range in all levels of D-cache/unified cache */ @@ -458,6 +459,15 @@ void flush_dcache_range(unsigned long start, unsigned long stop) { __asm_flush_dcache_range(start, stop); } +#else +void invalidate_dcache_range(unsigned long start, unsigned long stop) +{ +} + +void flush_dcache_range(unsigned long start, unsigned long stop) +{ +} +#endif /* CONFIG_SYS_DISABLE_DCACHE_OPS */ void dcache_enable(void) {
On AM654 SoC(arm64) which is IO coherent and has L3 Cache, cache maintenance operations being done to support non-coherent platforms causes issues. For example, here is how U-Boot prepares/handles a buffer to receive data from a device (DMA Write). This may vary slightly depending on the driver framework: Start DMA to write to destination buffer Wait for DMA to be done (dma_receive()/dma_memcpy()) Invalidate destination buffer (invalidate_dcache_range()) Read from destination buffer The invalidate after the DMA is needed in order to read latest data from memory that’s updated by DMA write. Also, in case random prefetch has pulled in buffer data during the “wait for DMA” before the DMA has written to it. This works well for non-coherent architectures. In case of coherent architecture with L3 cache, DMA write would directly update L3 cache contents (assuming cacheline is present in L3) without updating the DDR memory. So invalidate after “wait for DMA” in above sequence would discard latest data and read will cause stale data to be fetched from DDR. Therefore invalidate after “wait for DMA” is not always correct on coherent architecture. Therefore, provide a Kconfig option to disable cache maintenance ops on coherent architectures. This has added benefit of improving the performance of DMA transfers as we no longer need to invalidate/flush individual cache lines(especially for buffer thats several KBs in size). In order to facilitate use of same Kconfig across different architecture, I have added the symbol to top level arch/Kconfig file. Patch currently disables cache maintenance ops for arm64 only. flush_dcache_all() and invalidate_dcache_all() are exclusively used during enabling/disabling dcache and hence are not disabled. Signed-off-by: Vignesh Raghavendra <vigneshr@ti.com> --- arch/Kconfig | 9 +++++++++ arch/arm/cpu/armv8/cache_v8.c | 10 ++++++++++ 2 files changed, 19 insertions(+)