问题
I'm implementing cache maintenance functions for ARMv8 (Cortex-A53) running in 32 bit mode.
There is a problems when I try to flush memory region by using virtual addresses (VA). DCacheFlushByRange
looks like this
// some init.
// kDCacheL1 = 0; kDCacheL2 = 2;
while (alignedVirtAddr < endAddr)
{
// Flushing L1
asm volatile("mcr p15, 2, %0, c0, c0, 0" : : "r"(kDCacheL1) :); // select cache
isb();
asm volatile("mcr p15, 0, %0, c7, c14, 1" : : "r"(alignedVirtAddr) :); // clean & invalidate
dsb();
// Flushing L2
asm volatile("mcr p15, 2, %0, c0, c0, 0" : : "r"(kDCacheL2) :); // select cache
isb();
asm volatile("mcr p15, 0, %0, c7, c14, 1" : : "r"(alignedVirtAddr) :); // clean & invalidate
dsb();
alignedVirtAddr += lineSize;
}
DMA is used to validate the functions. DMA copies one buffer into another. Source buffer is flushed before DMA, destination buffer is invalidated after DMA completion. Buffers are 64 bytes aligned. Test
for (uint32_t i = 0; i < kBufSize; i++)
buf1[i] = 0;
for (uint32_t i = 0; i < kBufSize; i++)
buf0[i] = kRefValue;
DCacheFlushByRange(buf0, sizeof(buf0));
// run DMA
while (1) // wait DMA completion;
DCacheInvalidateByRange(buf1, sizeof(buf1));
compare(buf0, buf1);
In dump I could see that buf1
still contains only zeroes. When caches are turned off, result is correct so DMA itself works correctly.
Other point is when whole D-cache is flushed/invalidated by set/way result is correct.
// loops th/ way & set for L1 & L2
asm volatile("mcr p15, 0, %0, c7, c14, 2" : : "r"(setway) :)
So shortly flush/invalidate by set/way work correctly. The same by flashing/invalidating using VA doesn't. What could be a problem?
PS: kBufSize=4096;
, total buffer size is 4096 * sizeof(uint32_t) == 16KB
回答1:
There is no a problems w/ the function itself rather than Cortex-A53 cache implementation features.
From Cortex-A53 TRM
DCIMVAC operations in AArch32 and DC IVAC instructions in AArch64 perform an invalidate of the target address. If the data is dirty within the cluster then a clean is performed before the invalidate.
So there is no actual invalidate, there's clean and invalidate
Normal (at least for me) sequence is
flush(src);
dma(); // copy src -> dst
invalidate(dst);
But due to invalidate() does flush, old data from cache (dst region) is written on top of data in memory after DMA transfer.
Solution/workaround is
flush(src);
invalidate(dst);
dma(); // copy src -> dst
invalidate(dst); // again, that's right*.
* Data from 'dst' memory region could be fetched into a cache in advance. If that happens before DMA put data in memory, an old data from cache would be used. Second invalidate is fine, since data is not marked as 'dirty', it would be performed as 'pure invalidate'. No clean/flush in this case.
来源:https://stackoverflow.com/questions/42081435/flush-invalidate-range-by-virtual-address-armv8-cache