Skip to contentSkip to conversation

cxl/region: Reject delete of a provider-locked region

PATCH In review Richard Cheng 2 participants 3 messages Jul 21 CXL

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 cxl/region: Reject delete of a provider-locked region

delete_region_store() destroys a region object without checking CXL_REGION_F_LOCK, unlike commit_store() which refuses locked regions with -EPERM. A region owned by a memdev-attach provider carries F_LOCK, and cxl_region_decode_reset() deliberately skips the HDM decoder reset for F_LOCK regions. As a result, a userspace write to delete_region tears down the region object but leaves the endpoint HDM decoder committed with no owning region, which is an orphan, enabled decoder.

Gate delete_region_store() on CXL_REGION_F_LOCK.

---
 drivers/cxl/core/region.c | 3 +++
 1 file changed, 3 insertions(+)
drivers/cxl/core/region.c +3 -0
index 578622240401..cfc9622c9309 100644
@@ -2894,6 +2894,9 @@ static ssize_t delete_region_store(struct device *dev,
 	if (!cxlr || !sysfs_streq(buf, dev_name(&cxlr->dev)))
 		return -ENODEV;
 
+	if (test_bit(CXL_REGION_F_LOCK, &cxlr->flags))
+		return -EPERM;
+
 	unregister_region(cxlr);
 
 	return len;

base-commit: 5ca04f3ba91f1773bbd5da6d9c654ccc1ba7831d

Signature · 1 line
2.43.0
#1 lore raw
Thank you for your contribution! Sashiko AI review found 2 potential issue(s) to consider:

Thank you for your contribution! Sashiko AI review found 2 potential issue(s) to consider:

New issues:
- [High] A Time-Of-Check to Time-Of-Use (TOCTOU) race condition in delete_region_store() allows userspace to bypass the CXL_REGION_F_LOCK protection, leading to orphaned hardware decoders.

Pre-existing issues:
- [High] Userspace can bypass CXL_REGION_F_LOCK restrictions and tear down region software state by manually detaching targets via targetN sysfs attributes, leading to orphaned hardware decoders. --

--- Patch [1]: [PATCH] cxl/region: Reject delete of a provider-locked region --- commit 447746e2d3844401b6bdd5604c498e319f1e7719
Author: Richard Cheng <[email protected]>

cxl/region: Reject delete of a provider-locked region

delete_region_store() destroys a region object without checking CXL_REGION_F_LOCK, unlike commit_store() which refuses locked regions with -EPERM. This patch gates delete_region_store() on CXL_REGION_F_LOCK.

drivers/cxl/core/region.c static ssize_t delete_region_store(struct device *dev, · line 2899

Quoted 11 lines
drivers/cxl/core/region.c +3 -0
index 578622240401d..cfc9622c93095 100644
@@ -2894,6 +2894,9 @@ static ssize_t delete_region_store(struct device *dev,
 	if (!cxlr || !sysfs_streq(buf, dev_name(&cxlr->dev)))
 		return -ENODEV;
 
+	if (test_bit(CXL_REGION_F_LOCK, &cxlr->flags))
+		return -EPERM;
+

[Severity: High]
Is there a Time-Of-Check to Time-Of-Use (TOCTOU) race condition here?

The CXL_REGION_F_LOCK check is performed without holding synchronization like the region rwsem, leaving a window open before unregister_region() completes.

If userspace writes to delete_region while concurrently writing to targetN to attach a target, delete_region_store() could check the lock flag before it is set.

Then unregister_region() calling device_del() can block in kernfs_drain until the concurrent target attachment finishes and sets CXL_REGION_F_LOCK.

Once resumed, the software region would be destroyed with the lock flag set, skipping the hardware reset. Could this sequence recreate the exact orphaned hardware decoder state this patch intended to prevent?

[Severity: High]
This is a pre-existing issue, but does a similar bypass exist in store_targetN() when userspace manually detaches targets?

If a root user writes \n to a targetN sysfs attribute of a locked region, store_targetN() calls detach_target() which eventually reaches __cxl_decoder_detach(). This software teardown path lacks an F_LOCK check.

Because cxl_region_decode_reset() deliberately skips the hardware reset for F_LOCK regions, the software state is dropped while the hardware decoders remain committed.

The user could then write 0 to the size attribute, freeing the physical memory while it is still actively decoded by hardware.

Should store_targetN() or __cxl_decoder_detach() also be gated by CXL_REGION_F_LOCK to prevent userspace from tearing down the software state of locked regions?

Quoted 3 lines
drivers/cxl/core/region.c +0 -0
 	unregister_region(cxlr);
 
 	return len;
Signature · 1 line
Sashiko AI review · https://sashiko.dev/#/patchset/[email protected]?part=1
#2 lore raw
Quoted 42 lines · On Tue, Jul 21, 2026 at 06:04:02AM +0800, [email protected] wrote:

Thank you for your contribution! Sashiko AI review found 2 potential issue(s) to consider:

New issues:
- [High] A Time-Of-Check to Time-Of-Use (TOCTOU) race condition in delete_region_store() allows userspace to bypass the CXL_REGION_F_LOCK protection, leading to orphaned hardware decoders.

Pre-existing issues:
- [High] Userspace can bypass CXL_REGION_F_LOCK restrictions and tear down region software state by manually detaching targets via targetN sysfs attributes, leading to orphaned hardware decoders. --

--- Patch [1]: [PATCH] cxl/region: Reject delete of a provider-locked region --- commit 447746e2d3844401b6bdd5604c498e319f1e7719
Author: Richard Cheng <[email protected]>

cxl/region: Reject delete of a provider-locked region

delete_region_store() destroys a region object without checking CXL_REGION_F_LOCK, unlike commit_store() which refuses locked regions with -EPERM. This patch gates delete_region_store() on CXL_REGION_F_LOCK.

drivers/cxl/core/region.c static ssize_t delete_region_store(struct device *dev, · line 2899

Quoted 11 lines
drivers/cxl/core/region.c +3 -0
index 578622240401d..cfc9622c93095 100644
@@ -2894,6 +2894,9 @@ static ssize_t delete_region_store(struct device *dev,
 	if (!cxlr || !sysfs_streq(buf, dev_name(&cxlr->dev)))
 		return -ENODEV;
 
+	if (test_bit(CXL_REGION_F_LOCK, &cxlr->flags))
+		return -EPERM;
+

[Severity: High]
Is there a Time-Of-Check to Time-Of-Use (TOCTOU) race condition here?

The CXL_REGION_F_LOCK check is performed without holding synchronization like the region rwsem, leaving a window open before unregister_region() completes.

If userspace writes to delete_region while concurrently writing to targetN to attach a target, delete_region_store() could check the lock flag before it is set.

Quoted 7 lines

Then unregister_region() calling device_del() can block in kernfs_drain until the concurrent target attachment finishes and sets CXL_REGION_F_LOCK.

Once resumed, the software region would be destroyed with the lock flag set, skipping the hardware reset. Could this sequence recreate the exact orphaned hardware decoder state this patch intended to prevent?

I'd push back on this.

I think check-then-act observation is fair, but it doesn't affect the patch's target, which is a provider-locked region that is F_LOCK'd during assembly before any userspace delete is possible, and F_LOCK is never cleared, so the flag is always already set when delete_region_store() runs for it.

The window described by the sashiko-bot requires userspace to concurrently attach an F_LOCK-setting target to the same region.
I don't see a reachable path for that. And it can't be closed with a lock in delete_region_store() because unregister_region() calls device_del() before reset.

Quoted 13 lines

[Severity: High]
This is a pre-existing issue, but does a similar bypass exist in store_targetN() when userspace manually detaches targets?

If a root user writes \n to a targetN sysfs attribute of a locked region, store_targetN() calls detach_target() which eventually reaches __cxl_decoder_detach(). This software teardown path lacks an F_LOCK check.

Because cxl_region_decode_reset() deliberately skips the hardware reset for F_LOCK regions, the software state is dropped while the hardware decoders remain committed.

Quoted 11 lines

The user could then write 0 to the size attribute, freeing the physical memory while it is still actively decoded by hardware.

Should store_targetN() or __cxl_decoder_detach() also be gated by CXL_REGION_F_LOCK to prevent userspace from tearing down the software state of locked regions?

Quoted 3 lines
drivers/cxl/core/region.c +0 -0
 	unregister_region(cxlr);
 
 	return len;

Will send another patch fix for this one.

Best regards,
Richard Cheng.

Quoted 2 lines
Signature · 1 line
Sashiko AI review · https://sashiko.dev/#/patchset/[email protected]?part=1