Hi Maro,
+More people. There maybe more ideas about this.
On Mon, Mar 21, 2016 at 10:32:46AM +0100, mario.six@gdsys.cc wrote:
Hi Peng,
Quoting Peng Fan van.freenix@gmail.com:
Hi Mario,
On Fri, Mar 18, 2016 at 09:16:48AM +0100, mario.six@gdsys.cc wrote:
Hello,
I've been working on a QorIQ P1022 board (controlcenterd) to run the newest U-Boot on it, and I encountered some strange behavior.
During boot, we get these error messages
" ERROR: Cannot import environment: errno = 12
at common/env_common.c:221/env_import() *** Warning - import failed, using default environment
ERROR: Environment import failed: errno = 12
at common/env_common.c:123/set_default_env() ## Can't malloc 9 bytes "
After bisecting, I found out that the commit 4683b22 (mmc:fsl_esdhc invalidate dcache before read) apparently causes this. With the additional d-cache invalidations in place, malloc and free calls fail.
Now, after some experimenting, I found out that deactivating the d-cache invalidation when reading the first sector is enough; that is, if one replaces the first invalidation
" if (data->flags & MMC_DATA_READ) check_and_invalidate_dcache_range(cmd, data); "
with
" if (data->flags & MMC_DATA_READ && (cmd->cmdidx != MMC_CMD_READ_SINGLE_BLOCK || cmd->cmdarg != 0)) check_and_invalidate_dcache_range(cmd, data); "
in fsl_esdhc.c, it seems to work, but I have no idea why that is the case.
Any ideas would be greatly appreciated.
Before you chaning: " if (data->flags & MMC_DATA_READ) check_and_invalidate_dcache_range(cmd, data); "
can you please try this, https://patchwork.ozlabs.org/patch/511889/ and kindly give some feedback whether the patch can fix you issue or not?
I strongly agree that U-Boot should flow the Linux DMA flow to avoid potential write back and speculative read which may cause DMA data be polluted.
Anyway please first try first invalidate L2, then invalidate L1.
I plan to pick the upper patch or rework for DMA usage, but have not began the work (:
Thanks for the reply!
I tried to add the L2 invalidation from release.S into the invalidate_dcache_range function from ppccache.S:
Oh. I missed you are using ppc. Sorry. Actually the flow I suggested is for ARM.
"
Subject: [PATCH] Invalidate L2 cache before L1 cache
arch/powerpc/lib/ppccache.S | 9 +++++++++ 1 file changed, 9 insertions(+)
diff --git a/arch/powerpc/lib/ppccache.S b/arch/powerpc/lib/ppccache.S index b96dbc6..3f68be2 100644 --- a/arch/powerpc/lib/ppccache.S +++ b/arch/powerpc/lib/ppccache.S @@ -87,6 +87,15 @@ _GLOBAL(flush_dcache_range)
- invalidate_dcache_range(unsigned long start, unsigned long stop)
*/ _GLOBAL(invalidate_dcache_range)
- msync
- lis r2,(L2CSR0_L2FI|L2CSR0_L2LFC)@h
- ori r2,r2,(L2CSR0_L2FI|L2CSR0_L2LFC)@l
- mtspr SPRN_L2CSR0,r2
+2:
- mfspr r5,SPRN_L2CSR0
- and. r1,r5,r2
- bne 2b
- li r5,L1_CACHE_BYTES-1 andc r3,r3,r5 subf r4,r3,r4
-- 1.8.3
"
That had the effect of crashing the board on the first MMC access, sadly.
But, this had me thinking. Correct me if I'm wrong, since I'm in *no way* a DMA expert, but the logic currently works as follows:
- Before the DMA transfer, invalidate the d-cache, so no stale data in the
cache can pollute the result of the DMA transfer. 2. Do the DMA transfer. 3. Invalidate the d-cache again, so the CPU doesn't read stale data from the cache.
Shouldn't the cache be *flushed* in the first step? The DMA transfer didn't
To ARM, invalidate is enough, no need to flush and the flow is 1. Invalidate L1, then invalidate L2 to fix potential write back. 2. DMA read 3. Invalidate L2, then invalidate L1 to fix potential speculative read.
But in uboot, we do not have dma api, only use invalidate_cache_range to do the invalidation and the flow does not follow ARM recommendations after DMA read.
I have no knowledge of PPC (:
I think you need to check the flow of dma_map_area and dmp_unmap_area, then you may get the idea of how ppc handles DMA, then use it in this fsl_esdhc driver. Hope this is useful.
Regards, Peng.
happen yet, so the data in the cache is "legitimate," in the sense that it originated from regular usage of the cache systen, and writing the data from the cache back to the memory doesn't overwrite something we don't want it to.
I did just that, and replaced the invalidation with a flush, and it seems to fix my problem, at least :) (I included the patch at the end of the message).
Thanks again for the help!
Best regards, Mario
-- >8 --
Subject: [PATCH] Replace cache invalidation with flush before DMA transfer
drivers/mmc/fsl_esdhc.c | 25 ++++++++++++++++++++++++- 1 file changed, 24 insertions(+), 1 deletion(-)
diff --git a/drivers/mmc/fsl_esdhc.c b/drivers/mmc/fsl_esdhc.c index ea5f4bf..aa5b29e 100644 --- a/drivers/mmc/fsl_esdhc.c +++ b/drivers/mmc/fsl_esdhc.c @@ -302,6 +302,29 @@ static void check_and_invalidate_dcache_range invalidate_dcache_range(start, end); }
+static void check_and_flush_dcache_range
- (struct mmc_cmd *cmd,
struct mmc_data *data) {+#ifdef CONFIG_FSL_LAYERSCAPE
- unsigned start = 0;
+#else
- unsigned start = (unsigned)data->dest ;
+#endif
- unsigned size = roundup(ARCH_DMA_MINALIGN,
data->blocks*data->blocksize);- unsigned end = start+size ;
+#ifdef CONFIG_FSL_LAYERSCAPE
- dma_addr_t addr;
- addr = virt_to_phys((void *)(data->dest));
- if (upper_32_bits(addr))
printf("Error found for upper 32 bits\n");- else
start = lower_32_bits(addr);+#endif
- flush_dcache_range(start, end);
+}
/*
- Sends a command out on the bus. Takes the mmc pointer,
- a command pointer, and an optional data pointer.
@@ -346,7 +369,7 @@ esdhc_send_cmd(struct mmc *mmc, struct mmc_cmd *cmd, struct mmc_data *data) return err;
if (data->flags & MMC_DATA_READ)
check_and_invalidate_dcache_range(cmd, data);
check_and_flush_dcache_range(cmd, data);}
/* Figure out the transfer arguments */
-- 1.8.3
Thanks, Peng.
Best regards, Mario