* [PATCH] dma-debug: Avoid uninitialized hash traversal in mapping dumps
@ 2026-11-15 8:25 Avi Weiss
2026-11-16 0:09 ` Anthony Iliopoulos
0 siblings, 1 reply; 3+ messages in thread
From: Avi Weiss @ 2026-11-15 8:25 UTC (permalink / raw)
To: Marek Szyprowski; +Cc: Robin Murphy, Anthony Iliopoulos, iommu, linux-kernel
Booting with dma_debug=off causes dma_debug_init() to return before
initializing the hash buckets. The debugfs dump file is registered
independently, so reading it traverses an uninitialized list and
triggers a NULL pointer dereference in dump_show(). The
debug_dma_dump_mappings() helper has the same unchecked traversal.
Skip both dump paths until dma_debug_initialized is set. Check the
initialization state rather than global_disable so that mappings
recorded before DMA debugging disables itself at runtime remain
available.
With CONFIG_DMA_API_DEBUG=y and CONFIG_DEBUG_FS=y, reading the dump
with dma_debug=off triggers an oops in the unpatched kernel. The
patched kernel returns an empty dump, while dumping with DMA debugging
enabled still returns mapping information.
Fixes: 173735c346c4 ("dma-debug: fix debugfs initialization order")
Signed-off-by: Avi Weiss <thnkslprpt@gmail.com>
---
kernel/dma/debug.c | 6 ++++++
1 file changed, 6 insertions(+)
diff --git a/kernel/dma/debug.c b/kernel/dma/debug.c
index 36c1f7cbb2ae..2f487246d8ce 100644
--- a/kernel/dma/debug.c
+++ b/kernel/dma/debug.c
@@ -543,6 +543,9 @@ void debug_dma_dump_mappings(struct device *dev)
int idx;
phys_addr_t cln;
+ if (!dma_debug_initialized)
+ return;
+
for (idx = 0; idx < HASH_SIZE; idx++) {
struct hash_bucket *bucket = &dma_entry_hash[idx];
struct dma_debug_entry *entry;
@@ -576,6 +579,9 @@ static int dump_show(struct seq_file *seq, void *v)
int idx;
phys_addr_t cln;
+ if (!dma_debug_initialized)
+ return 0;
+
for (idx = 0; idx < HASH_SIZE; idx++) {
struct hash_bucket *bucket = &dma_entry_hash[idx];
struct dma_debug_entry *entry;
base-commit: 97d73c8251226cf27cc6656b4c542869bb8c69fa
--
2.43.0
^ permalink raw reply [flat|nested] 3+ messages in thread* Re: [PATCH] dma-debug: Avoid uninitialized hash traversal in mapping dumps
2026-11-15 8:25 [PATCH] dma-debug: Avoid uninitialized hash traversal in mapping dumps Avi Weiss
@ 2026-11-16 0:09 ` Anthony Iliopoulos
2026-11-16 13:45 ` Robin Murphy
0 siblings, 1 reply; 3+ messages in thread
From: Anthony Iliopoulos @ 2026-11-16 0:09 UTC (permalink / raw)
To: Avi Weiss; +Cc: Marek Szyprowski, Robin Murphy, iommu, linux-kernel
On Sun, Nov 15, 2026 at 10:25:06AM +0200, Avi Weiss wrote:
> Booting with dma_debug=off causes dma_debug_init() to return before
> initializing the hash buckets. The debugfs dump file is registered
> independently, so reading it traverses an uninitialized list and
> triggers a NULL pointer dereference in dump_show(). The
> debug_dma_dump_mappings() helper has the same unchecked traversal.
So the debug_dma_dump_mappings() issue seems preexisting (the risk was
there since its introduction in commit ac26c18bd35d, given that it
doesn't check if dma_debug has been initialized), and it is not related
to decoupling the debugfs init.
> Skip both dump paths until dma_debug_initialized is set. Check the
> initialization state rather than global_disable so that mappings
> recorded before DMA debugging disables itself at runtime remain
> available.
I suppose this makes sense for debug_dma_dump_mappings, but I am not
sure we need this for dump_show. See my next comment.
> With CONFIG_DMA_API_DEBUG=y and CONFIG_DEBUG_FS=y, reading the dump
> with dma_debug=off triggers an oops in the unpatched kernel. The
> patched kernel returns an empty dump, while dumping with DMA debugging
> enabled still returns mapping information.
>
> Fixes: 173735c346c4 ("dma-debug: fix debugfs initialization order")
> Signed-off-by: Avi Weiss <thnkslprpt@gmail.com>
> ---
> kernel/dma/debug.c | 6 ++++++
> 1 file changed, 6 insertions(+)
>
> diff --git a/kernel/dma/debug.c b/kernel/dma/debug.c
> index 36c1f7cbb2ae..2f487246d8ce 100644
> --- a/kernel/dma/debug.c
> +++ b/kernel/dma/debug.c
> @@ -543,6 +543,9 @@ void debug_dma_dump_mappings(struct device *dev)
> int idx;
> phys_addr_t cln;
>
> + if (!dma_debug_initialized)
> + return;
> +
> for (idx = 0; idx < HASH_SIZE; idx++) {
> struct hash_bucket *bucket = &dma_entry_hash[idx];
> struct dma_debug_entry *entry;
> @@ -576,6 +579,9 @@ static int dump_show(struct seq_file *seq, void *v)
> int idx;
> phys_addr_t cln;
>
> + if (!dma_debug_initialized)
> + return 0;
> +
I think it makes better sense to skip creating debugfs files altogether
when debugging is disabled, which was also the case before 15b28bbcd567.
> for (idx = 0; idx < HASH_SIZE; idx++) {
> struct hash_bucket *bucket = &dma_entry_hash[idx];
> struct dma_debug_entry *entry;
>
> base-commit: 97d73c8251226cf27cc6656b4c542869bb8c69fa
> --
> 2.43.0
>
Regards,
Anthony
^ permalink raw reply [flat|nested] 3+ messages in thread* Re: [PATCH] dma-debug: Avoid uninitialized hash traversal in mapping dumps
2026-11-16 0:09 ` Anthony Iliopoulos
@ 2026-11-16 13:45 ` Robin Murphy
0 siblings, 0 replies; 3+ messages in thread
From: Robin Murphy @ 2026-11-16 13:45 UTC (permalink / raw)
To: Anthony Iliopoulos, Avi Weiss; +Cc: Marek Szyprowski, iommu, linux-kernel
On 16/11/2026 12:09 am, Anthony Iliopoulos wrote:
> On Sun, Nov 15, 2026 at 10:25:06AM +0200, Avi Weiss wrote:
>> Booting with dma_debug=off causes dma_debug_init() to return before
>> initializing the hash buckets. The debugfs dump file is registered
>> independently, so reading it traverses an uninitialized list and
>> triggers a NULL pointer dereference in dump_show(). The
>> debug_dma_dump_mappings() helper has the same unchecked traversal.
>
> So the debug_dma_dump_mappings() issue seems preexisting (the risk was
> there since its introduction in commit ac26c18bd35d, given that it
> doesn't check if dma_debug has been initialized), and it is not related
> to decoupling the debugfs init.
>
>> Skip both dump paths until dma_debug_initialized is set. Check the
>> initialization state rather than global_disable so that mappings
>> recorded before DMA debugging disables itself at runtime remain
>> available.
>
> I suppose this makes sense for debug_dma_dump_mappings, but I am not
> sure we need this for dump_show. See my next comment.
>
>> With CONFIG_DMA_API_DEBUG=y and CONFIG_DEBUG_FS=y, reading the dump
>> with dma_debug=off triggers an oops in the unpatched kernel. The
>> patched kernel returns an empty dump, while dumping with DMA debugging
>> enabled still returns mapping information.
>>
>> Fixes: 173735c346c4 ("dma-debug: fix debugfs initialization order")
>> Signed-off-by: Avi Weiss <thnkslprpt@gmail.com>
>> ---
>> kernel/dma/debug.c | 6 ++++++
>> 1 file changed, 6 insertions(+)
>>
>> diff --git a/kernel/dma/debug.c b/kernel/dma/debug.c
>> index 36c1f7cbb2ae..2f487246d8ce 100644
>> --- a/kernel/dma/debug.c
>> +++ b/kernel/dma/debug.c
>> @@ -543,6 +543,9 @@ void debug_dma_dump_mappings(struct device *dev)
>> int idx;
>> phys_addr_t cln;
>>
>> + if (!dma_debug_initialized)
>> + return;
>> +
>> for (idx = 0; idx < HASH_SIZE; idx++) {
>> struct hash_bucket *bucket = &dma_entry_hash[idx];
>> struct dma_debug_entry *entry;
>> @@ -576,6 +579,9 @@ static int dump_show(struct seq_file *seq, void *v)
>> int idx;
>> phys_addr_t cln;
>>
>> + if (!dma_debug_initialized)
>> + return 0;
>> +
>
> I think it makes better sense to skip creating debugfs files altogether
> when debugging is disabled, which was also the case before 15b28bbcd567.
TBH there's probably an argument for just getting rid of
debug_dma_dump_mappings() now too... Hands up anyone who's using a
20-year-old AMD GART as any kind of development machine at all, let
alone for DMA API debugging of new drivers?
Anyone?
Cheers,
Robin.
>
>> for (idx = 0; idx < HASH_SIZE; idx++) {
>> struct hash_bucket *bucket = &dma_entry_hash[idx];
>> struct dma_debug_entry *entry;
>>
>> base-commit: 97d73c8251226cf27cc6656b4c542869bb8c69fa
>> --
>> 2.43.0
>>
>
> Regards,
> Anthony
^ permalink raw reply [flat|nested] 3+ messages in thread
end of thread, other threads:[~2026-11-16 13:45 UTC | newest]
Thread overview: 3+ messages (download: mbox.gz / follow: Atom feed)
-- links below jump to the message on this page --
2026-11-15 8:25 [PATCH] dma-debug: Avoid uninitialized hash traversal in mapping dumps Avi Weiss
2026-11-16 0:09 ` Anthony Iliopoulos
2026-11-16 13:45 ` Robin Murphy
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox
all inboxes | Powered by JetHome®