Skip to contentSkip to conversation

dma-debug: Fix overflow issue in bucket_find_contain

PATCH In review [email protected] 3 participants 3 messages Jul 30, 2022 – Aug 1, 2022

Subsystems (2)
lore
Share & tools
lore
#0 lore raw

PATCH 1/1 dma-debug: Fix overflow issue in bucket_find_contain

From: Yunfei Wang <[email protected]>

There are two issue:
1. If max_rang is set to 0xFFFF_FFFF, and __hash_bucket_find always returns NULL, the rang will be accumulated. When rang is accumulated to 0xFFFF_E000, after executing rang += (1 << HASH_FN_SHIFT) again, rang will overflow to 0, making it impossible to exit the while loop.
2. dev_addr reduce maybe overflow.

So, add range and dev_addr check to avoid overflow.

---
 kernel/dma/debug.c | 8 ++++++--
 1 file changed, 6 insertions(+), 2 deletions(-)
kernel/dma/debug.c +6 -2
index ad731f7858c9..9d7d54cd4c63 100644
@@ -352,6 +352,7 @@ static struct dma_debug_entry *bucket_find_contain(struct hash_bucket **bucket,
 
 	unsigned int max_range = dma_get_max_seg_size(ref->dev);
 	struct dma_debug_entry *entry, index = *ref;
+	unsigned int shift = (1 << HASH_FN_SHIFT);
 	unsigned int range = 0;
 
 	while (range <= max_range) {
@@ -360,12 +361,15 @@ static struct dma_debug_entry *bucket_find_contain(struct hash_bucket **bucket,
 		if (entry)
 			return entry;
 
+		if (max_range - range < shift || index.dev_addr < shift)
+			return NULL;
+
 		/*
 		 * Nothing found, go back a hash bucket
 		 */
 		put_hash_bucket(*bucket, *flags);
-		range          += (1 << HASH_FN_SHIFT);
-		index.dev_addr -= (1 << HASH_FN_SHIFT);
+		range          += shift;
+		index.dev_addr -= shift;
 		*bucket = get_hash_bucket(&index, flags);
 	}
 
Signature · 1 line
2.18.0
#1 lore raw

kernel/dma/debug.c static struct dma_debug_entry *bucket_find_contain(struct hash_bucket **bucket, · line 355

Quoted 1 line
kernel/dma/debug.c +1 -0
+	unsigned int shift = (1 << HASH_FN_SHIFT);

Please just add a constant for that right nxt to the HASH_FN_SHIFT definition.

reviewer
#2 lore raw
In reply to [email protected]
Quoted 30 lines · On 2022-07-30 12:41, [email protected] wrote:

From: Yunfei Wang <[email protected]>

There are two issue:
1. If max_rang is set to 0xFFFF_FFFF, and __hash_bucket_find always returns NULL, the rang will be accumulated. When rang is accumulated to 0xFFFF_E000, after executing rang += (1 << HASH_FN_SHIFT) again, rang will overflow to 0, making it impossible to exit the while loop.
2. dev_addr reduce maybe overflow.

So, add range and dev_addr check to avoid overflow.

---
  kernel/dma/debug.c | 8 ++++++--
  1 file changed, 6 insertions(+), 2 deletions(-)
kernel/dma/debug.c +1 -0
index ad731f7858c9..9d7d54cd4c63 100644
@@ -352,6 +352,7 @@ static struct dma_debug_entry *bucket_find_contain(struct hash_bucket **bucket,
 
  	unsigned int max_range = dma_get_max_seg_size(ref->dev);
  	struct dma_debug_entry *entry, index = *ref;
+	unsigned int shift = (1 << HASH_FN_SHIFT);
  	unsigned int range = 0;
 
  	while (range <= max_range) {

kernel/dma/debug.c static struct dma_debug_entry *bucket_find_contain(struct hash_bucket **bucket, · line 365

Quoted 5 lines
kernel/dma/debug.c +2 -0
@@ -360,12 +361,15 @@ static struct dma_debug_entry *bucket_find_contain(struct hash_bucket **bucket,
  		if (entry)
  			return entry;
 
+		if (max_range - range < shift || index.dev_addr < shift)
+			return NULL;

This seems a bit clunky since the first condition here effectively makes the loop condition redundant.

FWIW I found the whole "range" business here rather hard to make sense of - personally I'd calculate a lower bound for the address then just iterate down to that, but maybe that's just me :/

Otherwise, at the very least we should be capping max_range so that the loop doesn't go beyond HASH_SIZE iterations and pointlessly search the same buckets more than once - it's stupid to even *get* to the point of having to worry about that overflowing. Whether we really care about dev_addr underflow is then another matter.

Really it would seem even more logical to make this a lower-level function that can walk round the dma_entry_hash array directly and not have to monkey about with the fake "index" entry at all, but cleaning up the almost-unnecessary amount of internal abstractions here is maybe more work than it's worth at this point.

Robin.

Quoted 12 lines
kernel/dma/debug.c +3 -2
+
  		/*
  		 * Nothing found, go back a hash bucket
  		 */
  		put_hash_bucket(*bucket, *flags);
-		range          += (1 << HASH_FN_SHIFT);
-		index.dev_addr -= (1 << HASH_FN_SHIFT);
+		range          += shift;
+		index.dev_addr -= shift;
  		*bucket = get_hash_bucket(&index, flags);
  	}