Skip to contentSkip to conversation

sound: memalloc.h - fix wrong truncation of dma_addr_t

PATCH In review Stefano Panella 2 participants 2 messages Dec 10, 2013 Sound

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 sound: memalloc.h - fix wrong truncation of dma_addr_t

When running a 32bit kernel the hda_intel driver is still reporting a 64bit dma_mask if the HW supports it.

Quoted 1 line

From sound/pci/hda/hda_intel.c:

        /* allow 64bit DMA address if supported by H/W */
        if ((gcap & ICH6_GCAP_64OK) && !pci_set_dma_mask(pci, DMA_BIT_MASK(64)))
                pci_set_consistent_dma_mask(pci, DMA_BIT_MASK(64));
        else {
                pci_set_dma_mask(pci, DMA_BIT_MASK(32));
                pci_set_consistent_dma_mask(pci, DMA_BIT_MASK(32));
        }

which means when there is a call to dma_alloc_coherent from snd_malloc_dev_pages a machine address bigger than 32bit can be returned. This can be true in particular if running the 32bit kernel as a pv dom0 under the Xen Hypervisor or PAE on bare metal.

The problem is that when calling setup_bdle to program the BLE the dma_addr_t returned from the dma_alloc_coherent is wrongly truncated from snd_sgbuf_get_addr if running a 32bit kernel:

static inline dma_addr_t snd_sgbuf_get_addr(struct snd_dma_buffer *dmab,
size_t offset)

{
        struct snd_sg_buf *sgbuf = dmab->private_data;
        dma_addr_t addr = sgbuf->table[offset >> PAGE_SHIFT].addr;
        addr &= PAGE_MASK;
        return addr + offset % PAGE_SIZE;
}

where PAGE_MASK in a 32bit kernel is zeroing the upper 32bit af addr.

Without this patch the HW will fetch the 32bit truncated address, which is not the one obtained from dma_alloc_coherent and will result to a non working audio but can corrupt host memory at a random location.

The current patch apply to v3.13-rc3-74-g6c843f5

---
 include/sound/memalloc.h |    2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
include/sound/memalloc.h +1 -1
index af99839..5f73785 100644
@@ -108,7 +108,7 @@ static inline dma_addr_t snd_sgbuf_get_addr(struct snd_dma_buffer *dmab,
 {
 	struct snd_sg_buf *sgbuf = dmab->private_data;
 	dma_addr_t addr = sgbuf->table[offset >> PAGE_SHIFT].addr;
-	addr &= PAGE_MASK;
+	addr &= ~((dma_addr_t)PAGE_SIZE - 1);
 	return addr + offset % PAGE_SIZE;
 }
 
Signature · 1 line
1.7.9.5
#1 lore raw

At Tue, 10 Dec 2013 14:20:28 +0000,

Quoted 34 lines · Stefano Panella

When running a 32bit kernel the hda_intel driver is still reporting a 64bit dma_mask if the HW supports it.

Quoted 1 line

From sound/pci/hda/hda_intel.c:

        /* allow 64bit DMA address if supported by H/W */
        if ((gcap & ICH6_GCAP_64OK) && !pci_set_dma_mask(pci, DMA_BIT_MASK(64)))
                pci_set_consistent_dma_mask(pci, DMA_BIT_MASK(64));
        else {
                pci_set_dma_mask(pci, DMA_BIT_MASK(32));
                pci_set_consistent_dma_mask(pci, DMA_BIT_MASK(32));
        }

which means when there is a call to dma_alloc_coherent from snd_malloc_dev_pages a machine address bigger than 32bit can be returned. This can be true in particular if running the 32bit kernel as a pv dom0 under the Xen Hypervisor or PAE on bare metal.

The problem is that when calling setup_bdle to program the BLE the dma_addr_t returned from the dma_alloc_coherent is wrongly truncated from snd_sgbuf_get_addr if running a 32bit kernel:

static inline dma_addr_t snd_sgbuf_get_addr(struct snd_dma_buffer *dmab,
size_t offset)

{
        struct snd_sg_buf *sgbuf = dmab->private_data;
        dma_addr_t addr = sgbuf->table[offset >> PAGE_SHIFT].addr;
        addr &= PAGE_MASK;
        return addr + offset % PAGE_SIZE;
}

where PAGE_MASK in a 32bit kernel is zeroing the upper 32bit af addr.

Quoted 8 lines

Without this patch the HW will fetch the 32bit truncated address, which is not the one obtained from dma_alloc_coherent and will result to a non working audio but can corrupt host memory at a random location.

The current patch apply to v3.13-rc3-74-g6c843f5

Thanks for spotting out. Applied now (I added Cc to stable).

It might be better to have a common PAGE_SIZE definition aligned to dma_addr_t, but this patch should be enough for now.

Takashi

Quoted 20 lines
---
 include/sound/memalloc.h |    2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
include/sound/memalloc.h +1 -1
index af99839..5f73785 100644
@@ -108,7 +108,7 @@ static inline dma_addr_t snd_sgbuf_get_addr(struct snd_dma_buffer *dmab,
 {
 	struct snd_sg_buf *sgbuf = dmab->private_data;
 	dma_addr_t addr = sgbuf->table[offset >> PAGE_SHIFT].addr;
-	addr &= PAGE_MASK;
+	addr &= ~((dma_addr_t)PAGE_SIZE - 1);
 	return addr + offset % PAGE_SIZE;
 }
 
Signature · 1 line
1.7.9.5