Skip to contentSkip to conversation

cxl/region: Validate partition index before array access

PATCH Superseded KobaK 3 participants 5 messages Apr 9 – Apr 14 CXL

v1 of 2 Superseded Latest is v2 →

lore
Share & tools
lore

KobaK

author
#0 lore raw

PATCH 1/1 cxl/region: Validate partition index before array access

From: Koba Ko <[email protected]>

Check partition index bounds before accessing cxlds->part[] to prevent out-of-bounds access when part is -1 or invalid.

The partition index is read from cxled->part without validation. If it's negative or exceeds nr_partitions, accessing cxlds->part[part].mode will cause out-of-bounds array access.

Fixes 5ec67596e368 ("cxl/region: Drop goto pattern of construct_region()")
---
 drivers/cxl/core/region.c | 8 ++++++++
 1 file changed, 8 insertions(+)
drivers/cxl/core/region.c +8 -0
index edc267c6cf77a..6be46636db7ee 100644
@@ -3712,6 +3712,14 @@ static struct cxl_region *construct_region(struct cxl_root_decoder *cxlrd,
 	int rc, part = READ_ONCE(cxled->part);
 	struct cxl_region *cxlr;
 
+	if (part < 0 || part >= cxlds->nr_partitions) {
+		dev_err(cxlmd->dev.parent,
+			"%s:%s: invalid partition index %d (max %u)\n",
+			dev_name(&cxlmd->dev), dev_name(&cxled->cxld.dev),
+			part, cxlds->nr_partitions);
+		return ERR_PTR(-ENXIO);
+	}
+
 	do {
 		cxlr = __create_region(cxlrd, cxlds->part[part].mode,
 				       atomic_read(&cxlrd->region_id),
Signature · 1 line
2.43.0
maintainer
#1 lore raw
Quoted 11 lines · KobaK

From: Koba Ko <[email protected]>

Check partition index bounds before accessing cxlds->part[] to prevent out-of-bounds access when part is -1 or invalid.

The partition index is read from cxled->part without validation. If it's negative or exceeds nr_partitions, accessing cxlds->part[part].mode will cause out-of-bounds array access.

Fixes 5ec67596e368 ("cxl/region: Drop goto pattern of construct_region()")

Was this issue encountered during testing or just by inspection (or AI analysis)? I'm just curious on how this condition is triggered and if a regression unit test needs to be added.

DJ

Quoted 23 lines
---
 drivers/cxl/core/region.c | 8 ++++++++
 1 file changed, 8 insertions(+)
drivers/cxl/core/region.c +8 -0
index edc267c6cf77a..6be46636db7ee 100644
@@ -3712,6 +3712,14 @@ static struct cxl_region *construct_region(struct cxl_root_decoder *cxlrd,
 	int rc, part = READ_ONCE(cxled->part);
 	struct cxl_region *cxlr;
 
+	if (part < 0 || part >= cxlds->nr_partitions) {
+		dev_err(cxlmd->dev.parent,
+			"%s:%s: invalid partition index %d (max %u)\n",
+			dev_name(&cxlmd->dev), dev_name(&cxled->cxld.dev),
+			part, cxlds->nr_partitions);
+		return ERR_PTR(-ENXIO);
+	}
+
 	do {
 		cxlr = __create_region(cxlrd, cxlds->part[part].mode,
 				       atomic_read(&cxlrd->region_id),
author
#2 lore raw
In reply to KobaK
Quoted 3 lines · Dave Jiang

Was this issue encountered during testing or just by inspection (or AI analysis)? I'm just curious on how this condition is triggered and if a regression unit test needs to be added.

Caught by AI-assisted code analysis.

Thanks,
Koba

maintainer
#3 lore raw
In reply to KobaK
Quoted 11 lines · KobaK

From: Koba Ko <[email protected]>

Check partition index bounds before accessing cxlds->part[] to prevent out-of-bounds access when part is -1 or invalid.

The partition index is read from cxled->part without validation. If it's negative or exceeds nr_partitions, accessing cxlds->part[part].mode will cause out-of-bounds array access.

Fixes 5ec67596e368 ("cxl/region: Drop goto pattern of construct_region()")

This bounds check need is overstated. All writers of cxled->part either initialize it to -1 or assign it from a bounded walk of cxlds->nr_partitions. There is no path in that code that produces a positive out-of-range index, and the only invalid state is part == -1. So, let's focus on the -1 case.

(If we did want to defensively guard against out-of-range indices, that would need to be applied consistently across all accesses of cxlds->part[part], not just in construct_region(). I don't think that is needed.)

The Fixes: Tag was a bit perplexing. The cited commit does touch any lines related to partition indexing, at least based on "git show 5ec67596e368". However, a close look suggests something else is going on.

Commit be5cbd084027 ("cxl: Kill enum cxl_decoder_mode") which previously reworked this code included the needed check:

+ int rc, part = READ_ONCE(cxled->part);+
+ if (part < 0)
+ return ERR_PTR(-EBUSY);

And that check is indeed preserved in the final lore posting of the patch referenced in the Fixes: tag:
https://lore.kernel.org/all/[email protected]/

The version that landed in the tree however is different:
5ec67596e368 ("cxl/region: Drop goto pattern of construct_region()")

This suggests the check was not removed by that commit directly but instead lost during a merge.

At this point, I suggest a v2 of 'this' patch that simply restores the missing check, no added range check, no debug messaging, simply:

+ if (part < 0)
+ return ERR_PTR(-EBUSY);

It would be useful for Ming and DaveJ to take a close look at the merge to confirm if anything else is off.

Quoted 26 lines
---
 drivers/cxl/core/region.c | 8 ++++++++
 1 file changed, 8 insertions(+)
drivers/cxl/core/region.c +8 -0
index edc267c6cf77a..6be46636db7ee 100644
@@ -3712,6 +3712,14 @@ static struct cxl_region *construct_region(struct cxl_root_decoder *cxlrd,
 	int rc, part = READ_ONCE(cxled->part);
 	struct cxl_region *cxlr;
 
+	if (part < 0 || part >= cxlds->nr_partitions) {
+		dev_err(cxlmd->dev.parent,
+			"%s:%s: invalid partition index %d (max %u)\n",
+			dev_name(&cxlmd->dev), dev_name(&cxled->cxld.dev),
+			part, cxlds->nr_partitions);
+		return ERR_PTR(-ENXIO);
+	}
+
 	do {
 		cxlr = __create_region(cxlrd, cxlds->part[part].mode,
 				       atomic_read(&cxlrd->region_id),
Signature · 1 line
2.43.0
author
#4 lore raw

Thanks for the review. Will address the feedback in v2.

- Koba