diff options
| author | Chris Metcalf <cmetcalf@tilera.com> | 2012-06-13 14:46:40 -0400 |
|---|---|---|
| committer | Chris Metcalf <cmetcalf@tilera.com> | 2012-07-18 16:40:05 -0400 |
| commit | bbaa22c3a0d0be4406d26e5a73d1e8e504787986 (patch) | |
| tree | 4d00f1bda85d9735c60d7db1cdbdd215d5317ae4 /arch/tile/include/asm/cache.h | |
| parent | 3e219b91533058e242b78ac08aaa91024dd6f369 (diff) | |
| download | blackbird-op-linux-bbaa22c3a0d0be4406d26e5a73d1e8e504787986.tar.gz blackbird-op-linux-bbaa22c3a0d0be4406d26e5a73d1e8e504787986.zip | |
tilegx pci: support I/O to arbitrarily-cached pages
The tilegx PCI root complex support (currently only in linux-next)
is limited to pages that are homed on cached in the default manner,
i.e. "hash-for-home". This change supports delivery of I/O data to
pages that are cached in other ways (locally on a particular core,
uncached, user-managed incoherent, etc.).
A large part of the change is supporting flushing pages from cache
on particular homes so that we can transition the data that we are
delivering to or from the device appropriately. The new homecache_finv*
routines handle this.
Some changes to page_table_range_init() were also required to make
the fixmap code work correctly on tilegx; it hadn't been used there
before.
We also remove some stub mark_caches_evicted_*() routines that
were just no-ops anyway.
Signed-off-by: Chris Metcalf <cmetcalf@tilera.com>
Diffstat (limited to 'arch/tile/include/asm/cache.h')
| -rw-r--r-- | arch/tile/include/asm/cache.h | 12 |
1 files changed, 9 insertions, 3 deletions
diff --git a/arch/tile/include/asm/cache.h b/arch/tile/include/asm/cache.h index 392e5333dd8b..a9a529964e07 100644 --- a/arch/tile/include/asm/cache.h +++ b/arch/tile/include/asm/cache.h @@ -27,11 +27,17 @@ #define L2_CACHE_ALIGN(x) (((x)+(L2_CACHE_BYTES-1)) & -L2_CACHE_BYTES) /* - * TILE-Gx is fully coherent so we don't need to define ARCH_DMA_MINALIGN. + * TILEPro I/O is not always coherent (networking typically uses coherent + * I/O, but PCI traffic does not) and setting ARCH_DMA_MINALIGN to the + * L2 cacheline size helps ensure that kernel heap allocations are aligned. + * TILE-Gx I/O is always coherent when used on hash-for-home pages. + * + * However, it's possible at runtime to request not to use hash-for-home + * for the kernel heap, in which case the kernel will use flush-and-inval + * to manage coherence. As a result, we use L2_CACHE_BYTES for the + * DMA minimum alignment to avoid false sharing in the kernel heap. */ -#ifndef __tilegx__ #define ARCH_DMA_MINALIGN L2_CACHE_BYTES -#endif /* use the cache line size for the L2, which is where it counts */ #define SMP_CACHE_BYTES_SHIFT L2_CACHE_SHIFT |
