svn commit: r343962 - head/sys/arm/arm
mmel at FreeBSD.org
Sun Feb 10 14:25:30 UTC 2019
Date: Sun Feb 10 14:25:29 2019
New Revision: 343962
Properly handle alignment requests bigger that page size.
- for now, alignments bigger that page size is allowed only for buffers
allocated by bus_dmamem_alloc(), cover this fact by KASSERT.
- never bounce buffers allocated by bus_dmamem_alloc(), these always comply
with the required rules (alignment, boundary, address range).
MFC after: 1 week
Reviewed by: jah
--- head/sys/arm/arm/busdma_machdep-v6.c Sun Feb 10 14:02:14 2019 (r343961)
+++ head/sys/arm/arm/busdma_machdep-v6.c Sun Feb 10 14:25:29 2019 (r343962)
@@ -339,16 +339,27 @@ cacheline_bounce(bus_dmamap_t map, bus_addr_t addr, bu
* Note that the addr argument might be either virtual or physical. It doesn't
* matter because we only look at the low-order bits, which are the same in both
- * address spaces.
+ * address spaces and maximum alignment of generic buffer is limited up to page
+ * size.
+ * Bouncing of buffers allocated by bus_dmamem_alloc()is not necessary, these
+ * always comply with the required rules (alignment, boundary, and address
+ * range).
static __inline int
might_bounce(bus_dma_tag_t dmat, bus_dmamap_t map, bus_addr_t addr,
- return ((dmat->flags & BUS_DMA_EXCL_BOUNCE) ||
+ KASSERT(dmat->flags & DMAMAP_DMAMEM_ALLOC ||
+ dmat->alignment <= PAGE_SIZE,
+ ("%s: unsupported alignment (0x%08lx) for buffer not "
+ "allocated by bus_dmamem_alloc()",
+ __func__, dmat->alignment));
+ return (!(dmat->flags & DMAMAP_DMAMEM_ALLOC) &&
+ ((dmat->flags & BUS_DMA_EXCL_BOUNCE) ||
alignment_bounce(dmat, addr) ||
- cacheline_bounce(map, addr, size));
+ cacheline_bounce(map, addr, size)));
More information about the svn-src-all