From: Simon Glass Date: Tue, 4 Apr 2017 19:00:19 +0000 (-0600) Subject: dm: core: Ensure DMA regions start up with the cache clean X-Git-Tag: v2017.05-rc2~8^2~24 X-Git-Url: https://git.librecmc.org/?a=commitdiff_plain;h=5a8a8045a928702d9b3fb6f73a4e9717040e09a9;p=oweals%2Fu-boot.git dm: core: Ensure DMA regions start up with the cache clean There is a strange interaction with drivers which use DMA if the cache starts off in a dirty state. Buffer space which the driver reads (but has not previously written) can contain zero bytes from alloc_priv(). This can cause corruption of the memory used by DMA for incoming data. Fix this and add a comment to explain the problem. This allows the dwc2 driver to work correctly with driver model, for example. Signed-off-by: Simon Glass --- diff --git a/drivers/core/device.c b/drivers/core/device.c index e1b0ebffc5..09a115f753 100644 --- a/drivers/core/device.c +++ b/drivers/core/device.c @@ -255,8 +255,36 @@ static void *alloc_priv(int size, uint flags) if (flags & DM_FLAG_ALLOC_PRIV_DMA) { priv = memalign(ARCH_DMA_MINALIGN, size); - if (priv) + if (priv) { memset(priv, '\0', size); + + /* + * Ensure that the zero bytes are flushed to memory. + * This prevents problems if the driver uses this as + * both an input and an output buffer: + * + * 1. Zeroes written to buffer (here) and sit in the + * cache + * 2. Driver issues a read command to DMA + * 3. CPU runs out of cache space and evicts some cache + * data in the buffer, writing zeroes to RAM from + * the memset() above + * 4. DMA completes + * 5. Buffer now has some DMA data and some zeroes + * 6. Data being read is now incorrect + * + * To prevent this, ensure that the cache is clean + * within this range at the start. The driver can then + * use normal flush-after-write, invalidate-before-read + * procedures. + * + * TODO(sjg@chromium.org): Drop this microblaze + * exception. + */ +#ifndef CONFIG_MICROBLAZE + flush_dcache_range((ulong)priv, (ulong)priv + size); +#endif + } } else { priv = calloc(1, size); }