dma-debug: change hash_bucket_find from first-fit to best-fit
Joerg Roedel [Fri, 5 Jun 2009 10:01:35 +0000 (12:01 +0200)]
Some device drivers map the same physical address multiple times to a
dma address. Without an IOMMU this results in the same dma address being
put into the dma-debug hash multiple times. With a first-fit match in
hash_bucket_find() this function may return the wrong dma_debug_entry.

This can result in false positive warnings. This patch fixes it by
changing the first-fit behavior of hash_bucket_find() into a best-fit
algorithm.

Reported-by: Torsten Kaiser <just.for.lkml@googlemail.com>
Reported-by: FUJITA Tomonori <fujita.tomonori@lab.ntt.co.jp>
Signed-off-by: Joerg Roedel <joerg.roedel@amd.com>
Cc: lethal@linux-sh.org
Cc: just.for.lkml@googlemail.com
Cc: hancockrwd@gmail.com
Cc: jens.axboe@oracle.com
Cc: bharrosh@panasas.com
Cc: FUJITA Tomonori <fujita.tomonori@lab.ntt.co.jp>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: <stable@kernel.org>
LKML-Reference: <20090605104132.GE24836@amd.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>

lib/dma-debug.c

index cdd205d..8fcc09c 100644 (file)
@@ -186,15 +186,50 @@ static void put_hash_bucket(struct hash_bucket *bucket,
 static struct dma_debug_entry *hash_bucket_find(struct hash_bucket *bucket,
                                                struct dma_debug_entry *ref)
 {
-       struct dma_debug_entry *entry;
+       struct dma_debug_entry *entry, *ret = NULL;
+       int matches = 0, match_lvl, last_lvl = 0;
 
        list_for_each_entry(entry, &bucket->list, list) {
-               if ((entry->dev_addr == ref->dev_addr) &&
-                   (entry->dev == ref->dev))
+               if ((entry->dev_addr != ref->dev_addr) ||
+                   (entry->dev != ref->dev))
+                       continue;
+
+               /*
+                * Some drivers map the same physical address multiple
+                * times. Without a hardware IOMMU this results in the
+                * same device addresses being put into the dma-debug
+                * hash multiple times too. This can result in false
+                * positives being reported. Therfore we implement a
+                * best-fit algorithm here which returns the entry from
+                * the hash which fits best to the reference value
+                * instead of the first-fit.
+                */
+               matches += 1;
+               match_lvl = 0;
+               entry->size      == ref->size      ? ++match_lvl : match_lvl;
+               entry->type      == ref->type      ? ++match_lvl : match_lvl;
+               entry->direction == ref->direction ? ++match_lvl : match_lvl;
+
+               if (match_lvl == 3) {
+                       /* perfect-fit - return the result */
                        return entry;
+               } else if (match_lvl > last_lvl) {
+                       /*
+                        * We found an entry that fits better then the
+                        * previous one
+                        */
+                       last_lvl = match_lvl;
+                       ret      = entry;
+               }
        }
 
-       return NULL;
+       /*
+        * If we have multiple matches but no perfect-fit, just return
+        * NULL.
+        */
+       ret = (matches == 1) ? ret : NULL;
+
+       return ret;
 }
 
 /*