[PATCH] Fix swiotlb pci_map_sg error handling
authorAndi Kleen <ak@muc.de>
Tue, 20 Dec 2005 13:45:19 +0000 (14:45 +0100)
committerLinus Torvalds <torvalds@g5.osdl.org>
Tue, 20 Dec 2005 18:13:54 +0000 (10:13 -0800)
The overflow checking condition in lib/swiotlb.c was wrong.
It would first run a NULL pointer through virt_to_phys before
testing it. Since pci_map_sg overflow is not that uncommon
and causes data corruption (including broken file systems) when not
properly detected I think it's better to fix it in 2.6.15.

This affects x86-64 and IA64.

Signed-off-by: Andi Kleen <ak@suse.de>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
lib/swiotlb.c

index 57216f3544cad435e2986df9d83b9dbf4ddf37ad..1ff8dcebf7c6edf33abb7eb2583d5b6eab37ce29 100644 (file)
@@ -704,8 +704,9 @@ swiotlb_map_sg(struct device *hwdev, struct scatterlist *sg, int nelems,
                addr = SG_ENT_VIRT_ADDRESS(sg);
                dev_addr = virt_to_phys(addr);
                if (swiotlb_force || address_needs_mapping(hwdev, dev_addr)) {
-                       sg->dma_address = (dma_addr_t) virt_to_phys(map_single(hwdev, addr, sg->length, dir));
-                       if (!sg->dma_address) {
+                       void *map = map_single(hwdev, addr, sg->length, dir);
+                       sg->dma_address = virt_to_bus(map);
+                       if (!map) {
                                /* Don't panic here, we expect map_sg users
                                   to do proper error handling. */
                                swiotlb_full(hwdev, sg->length, dir, 0);