Skip to contentSkip to conversation

cxl: Fix uninitialized access coordinates

PATCH Superseded Guixin Liu 4 participants 9 messages Aug 12 – Aug 31 CXL

v3 of 4 Superseded Latest is v4 →

lore
Share & tools
lore

Guixin Liu

author
#0 lore raw

v3 cover cxl: Fix uninitialized access coordinates

Two related fixes for stack access_coordinate arrays that are read before anything writes them, both reachable on real topologies and both ending up in userspace through the access coordinate sysfs attributes.

Patch 1 is the v2 cdat fix [1] with one more site covered: the coords array in cxl_switch_gather_bandwidth() has the same defect as the three arrays in cxl_endpoint_gather_bandwidth().

Patch 2 is new. cxl_endpoint_get_perf_coordinates() returns 0 for an RCD without writing the caller's array at all, so cxl_port_perf_data_calculate() combines stack residue with the values parsed from the RCD's CDAT.

Both were pointed out by the Sashiko review bot on the v2 posting of patch 1. The third finding in that review - cxl_coordinates_combine() aliasing its output onto an input and leaving a member alone when an input bandwidth is zero - is the intended "not reported" behaviour and is not addressed here.

[1] https://lore.kernel.org/linux-cxl/[email protected]/

v2->v3:
- also zero initialize coords in cxl_switch_gather_bandwidth(), same defect in the sibling gather path (Sashiko bot)
- new patch 2: clear the output coordinates on the RCD path of cxl_endpoint_get_perf_coordinates(), which currently returns success without writing them (Sashiko bot)
- drop "endpoint" from the patch 1 subject, it now covers both gather paths

v1->v2:
- rebase onto cxl/next
- rewrite the commit message to describe the behaviour rather than narrate the code change (Alison Schofield)

Guixin Liu (2):
cxl/cdat: Fix uninitialized stack use in bandwidth gathering cxl/port: Fix uninitialized coordinates reported for RCDs

 drivers/cxl/core/cdat.c | 8 ++++----
 drivers/cxl/core/port.c | 4 +++-
 2 files changed, 7 insertions(+), 5 deletions(-)

base-commit: 7098e9cd98a05c0c5de2fae0c2465f9d966fdd07

Signature · 1 line
2.43.7
author
#1 lore raw

PATCH v3 1/2 cxl/cdat: Fix uninitialized stack use in bandwidth gathering

cxl_endpoint_gather_bandwidth() declares three access_coordinate arrays on the stack - pci_coord, sw_coord and ep_coord - without initializing them, and relies on its helpers to fill every member. None of them does. cxl_pci_get_bandwidth() and cxl_port_get_switch_dport_bandwidth() assign only read_bandwidth and write_bandwidth, and __cxl_coordinates_combine() assigns an output bandwidth only when both input bandwidths are non-zero. Every member a producer declines to set is read back as whatever was on the stack. cxl_switch_gather_bandwidth() has the same defect in its coords array, which it hands to cxl_pci_get_bandwidth() and then to cxl_coordinates_combine().

Both cases occur on real topologies. The latency members are never written by any of those producers, yet __cxl_coordinates_combine() sums them unconditionally, so the latency reads are undefined on every call. And a device whose CDAT DSLBIS reports no bandwidth for an access class leaves perf->cdat_coord zero for that class, which is exactly the condition that makes __cxl_coordinates_combine() skip the bandwidth assignment and leave the ep_coord entry untouched.

The ep_coord case escapes the function: cxl_bandwidth_add() accumulates it into the per-upstream-port aggregate that is published through the region's access coordinate sysfs attributes, so stack contents are reported to userspace as a bandwidth figure. The latency sums are discarded by cxl_bandwidth_add() rather than published, but they are still computed from uninitialized memory.

Zero initialize the four arrays. Zero is already the value this code uses for "not reported" - both the __cxl_coordinates_combine() guard and coordinates_valid() test for it - so a member no producer sets now reads back as unknown rather than as a plausible number.

Fixes a5ab0de0ebaa ("cxl: Calculate region bandwidth of targets with shared upstream link")
---
 drivers/cxl/core/cdat.c | 8 ++++----
 1 file changed, 4 insertions(+), 4 deletions(-)
drivers/cxl/core/cdat.c +4 -4
index 5c9f07262513..df9261ba84e5 100644
@@ -633,9 +633,9 @@ static int cxl_endpoint_gather_bandwidth(struct cxl_region *cxlr,
 	struct cxl_port *endpoint = to_cxl_port(cxled->cxld.dev.parent);
 	struct cxl_port *parent_port = to_cxl_port(endpoint->dev.parent);
 	struct cxl_port *gp_port = to_cxl_port(parent_port->dev.parent);
-	struct access_coordinate pci_coord[ACCESS_COORDINATE_MAX];
-	struct access_coordinate sw_coord[ACCESS_COORDINATE_MAX];
-	struct access_coordinate ep_coord[ACCESS_COORDINATE_MAX];
+	struct access_coordinate pci_coord[ACCESS_COORDINATE_MAX] = { };
+	struct access_coordinate sw_coord[ACCESS_COORDINATE_MAX] = { };
+	struct access_coordinate ep_coord[ACCESS_COORDINATE_MAX] = { };
 	struct cxl_memdev *cxlmd = cxled_to_memdev(cxled);
 	struct cxl_dev_state *cxlds = cxlmd->cxlds;
 	struct pci_dev *pdev = to_pci_dev(cxlds->dev);
@@ -757,7 +757,7 @@ static struct xarray *cxl_switch_gather_bandwidth(struct cxl_region *cxlr,
 {
 	struct xarray *res_xa __free(free_perf_xa) =
 		kzalloc_obj(*res_xa);
-	struct access_coordinate coords[ACCESS_COORDINATE_MAX];
+	struct access_coordinate coords[ACCESS_COORDINATE_MAX] = { };
 	struct cxl_perf_ctx *ctx, *us_ctx;
 	unsigned long index, us_index;
 	int dev_count = 0;
Signature · 1 line
2.43.7
#2 lore raw

Re: [PATCH v3 1/2] cxl/cdat: Fix uninitialized stack use in bandwidth gathering

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:

Pre-existing issues:
- [High] Zero-initializing ep_coord codifies a logic error where valid PCI bandwidth is discarded if CDAT bandwidth is unreported (0), contradicting subsystem conventions.
- [High] The cxl_port_perf_data_calculate() function declares an uninitialized access_coordinate array on the stack, which is leaked to userspace via sysfs for Restricted CXL Devices (RCDs). --

--- Patch [1]: [PATCH v3 1/2] cxl/cdat: Fix uninitialized stack use in bandwidth gathering --- commit ecbce27f92f7e67349081889c3ca38eb48761104
Author: Guixin Liu <[email protected]>
Subject: cxl/cdat: Fix uninitialized stack use in bandwidth gathering

This commit fixes uninitialized stack usage in cxl_endpoint_gather_bandwidth() and cxl_switch_gather_bandwidth() by zero-initializing the access_coordinate arrays. It prevents uninitialized kernel stack data from being exposed as bandwidth and latency figures.

Quoted 5 lines
drivers/cxl/core/cdat.c +0 -0
index 5c9f072625135..df9261ba84e5a 100644

drivers/cxl/core/cdat.c static int cxl_endpoint_gather_bandwidth(struct cxl_region *cxlr, · line 638

Quoted 9 lines
drivers/cxl/core/cdat.c +3 -3
@@ -633,9 +633,9 @@ static int cxl_endpoint_gather_bandwidth(struct cxl_region *cxlr,
 	struct cxl_port *endpoint = to_cxl_port(cxled->cxld.dev.parent);
 	struct cxl_port *parent_port = to_cxl_port(endpoint->dev.parent);
 	struct cxl_port *gp_port = to_cxl_port(parent_port->dev.parent);
-	struct access_coordinate pci_coord[ACCESS_COORDINATE_MAX];
-	struct access_coordinate sw_coord[ACCESS_COORDINATE_MAX];
-	struct access_coordinate ep_coord[ACCESS_COORDINATE_MAX];
+	struct access_coordinate pci_coord[ACCESS_COORDINATE_MAX] = { };
+	struct access_coordinate sw_coord[ACCESS_COORDINATE_MAX] = { };
+	struct access_coordinate ep_coord[ACCESS_COORDINATE_MAX] = { };

[Severity: High]
This isn't a bug introduced by this patch, but does zero-initializing ep_coord codify a logic error where valid PCI bandwidth is discarded?

If an endpoint CDAT does not report bandwidth, the bandwidth fields in perf->cdat_coord will be 0. When cxl_coordinates_combine() is called later in cxl_endpoint_gather_bandwidth():

	rc = cxl_pci_get_bandwidth(pdev, pci_coord);
	if (rc < 0)
		return rc;

	/*
	 * Min of upstream link bandwidth and Endpoint CDAT bandwidth from
	 * DSLBIS.
	 */
	cxl_coordinates_combine(ep_coord, pci_coord, perf->cdat_coord);

The __cxl_coordinates_combine() helper requires both inputs to be non-zero to assign the output:

	if (c1->write_bandwidth && c2->write_bandwidth)
		out->write_bandwidth = min(c1->write_bandwidth,
					   c2->write_bandwidth);

Because the bandwidth in perf->cdat_coord is 0, ep_coord remains untouched (which is now 0). This completely discards the valid upstream link bandwidth in pci_coord, resulting in an erroneous region aggregate bandwidth of 0.

[Severity: High]
This is a pre-existing issue, but while fixing uninitialized stack variables, did you notice a similar leak in cxl_port_perf_data_calculate() in drivers/cxl/core/cdat.c?

It declares an uninitialized array:

	struct access_coordinate ep_c[ACCESS_COORDINATE_MAX];
	...
	rc = cxl_endpoint_get_perf_coordinates(port, ep_c);

If the device is a Restricted CXL Device (RCD), cxl_endpoint_get_perf_coordinates() in drivers/cxl/core/port.c returns 0 early without initializing the array:

	/*
	 * Skip calculation for RCD. Expectation is HMAT already covers RCD case
	 * since RCH does not support hotplug.
	 */
	if (cxlmd->cxlds->rcd)
		return 0;

cxl_port_perf_data_calculate() then passes this uninitialized array to cxl_coordinates_combine(), which can leak kernel stack memory to userspace via world-readable sysfs region attributes.

Signature · 1 line
Sashiko AI review · https://sashiko.dev/#/patchset/[email protected]?part=1
author
#3 lore raw

Re: [PATCH v3 1/2] cxl/cdat: Fix uninitialized stack use in bandwidth gathering

Quoted 83 lines · 在 2026/8/12 16:44, [email protected] 写道:

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

Pre-existing issues:
- [High] Zero-initializing ep_coord codifies a logic error where valid PCI bandwidth is discarded if CDAT bandwidth is unreported (0), contradicting subsystem conventions.
- [High] The cxl_port_perf_data_calculate() function declares an uninitialized access_coordinate array on the stack, which is leaked to userspace via sysfs for Restricted CXL Devices (RCDs). --

--- Patch [1]: [PATCH v3 1/2] cxl/cdat: Fix uninitialized stack use in bandwidth gathering --- commit ecbce27f92f7e67349081889c3ca38eb48761104
Author: Guixin Liu <[email protected]>
Subject: cxl/cdat: Fix uninitialized stack use in bandwidth gathering

This commit fixes uninitialized stack usage in cxl_endpoint_gather_bandwidth() and cxl_switch_gather_bandwidth() by zero-initializing the access_coordinate arrays. It prevents uninitialized kernel stack data from being exposed as bandwidth and latency figures.

Quoted 5 lines
drivers/cxl/core/cdat.c +0 -0
index 5c9f072625135..df9261ba84e5a 100644

drivers/cxl/core/cdat.c static int cxl_endpoint_gather_bandwidth(struct cxl_region *cxlr, · line 638

Quoted 9 lines
drivers/cxl/core/cdat.c +3 -3
@@ -633,9 +633,9 @@ static int cxl_endpoint_gather_bandwidth(struct cxl_region *cxlr,
  	struct cxl_port *endpoint = to_cxl_port(cxled->cxld.dev.parent);
  	struct cxl_port *parent_port = to_cxl_port(endpoint->dev.parent);
  	struct cxl_port *gp_port = to_cxl_port(parent_port->dev.parent);
-	struct access_coordinate pci_coord[ACCESS_COORDINATE_MAX];
-	struct access_coordinate sw_coord[ACCESS_COORDINATE_MAX];
-	struct access_coordinate ep_coord[ACCESS_COORDINATE_MAX];
+	struct access_coordinate pci_coord[ACCESS_COORDINATE_MAX] = { };
+	struct access_coordinate sw_coord[ACCESS_COORDINATE_MAX] = { };
+	struct access_coordinate ep_coord[ACCESS_COORDINATE_MAX] = { };

[Severity: High]
This isn't a bug introduced by this patch, but does zero-initializing ep_coord codify a logic error where valid PCI bandwidth is discarded?

If an endpoint CDAT does not report bandwidth, the bandwidth fields in perf->cdat_coord will be 0. When cxl_coordinates_combine() is called later in cxl_endpoint_gather_bandwidth():

	rc = cxl_pci_get_bandwidth(pdev, pci_coord);
	if (rc < 0)
		return rc;

	/*
	 * Min of upstream link bandwidth and Endpoint CDAT bandwidth from
	 * DSLBIS.
	 */
	cxl_coordinates_combine(ep_coord, pci_coord, perf->cdat_coord);

The __cxl_coordinates_combine() helper requires both inputs to be non-zero to assign the output:

	if (c1->write_bandwidth && c2->write_bandwidth)
		out->write_bandwidth = min(c1->write_bandwidth,
					   c2->write_bandwidth);

Because the bandwidth in perf->cdat_coord is 0, ep_coord remains untouched (which is now 0). This completely discards the valid upstream link bandwidth in pci_coord, resulting in an erroneous region aggregate bandwidth of 0.

[Severity: High]
This is a pre-existing issue, but while fixing uninitialized stack variables, did you notice a similar leak in cxl_port_perf_data_calculate() in drivers/cxl/core/cdat.c?

It declares an uninitialized array:

	struct access_coordinate ep_c[ACCESS_COORDINATE_MAX];
	...
	rc = cxl_endpoint_get_perf_coordinates(port, ep_c);

If the device is a Restricted CXL Device (RCD), cxl_endpoint_get_perf_coordinates() in drivers/cxl/core/port.c returns 0 early without initializing the array:

	/*
	 * Skip calculation for RCD. Expectation is HMAT already covers RCD case
	 * since RCH does not support hotplug.
	 */
	if (cxlmd->cxlds->rcd)
		return 0;
Quoted 3 lines

cxl_port_perf_data_calculate() then passes this uninitialized array to cxl_coordinates_combine(), which can leak kernel stack memory to userspace via world-readable sysfs region attributes.

On the RCD leak in cxl_port_perf_data_calculate(): that is what patch 2 of this series fixes. cxl_endpoint_get_perf_coordinates() now zeroes the caller's array before the early return for RCDs, so ep_c[] is no longer read before anything writes it. The fix is in the exported helper rather than in the caller because the helper documents @coord as an output parameter, and cxl_test discards its return value.

On ep_coord being discarded when the CDAT reports no bandwidth: the guard that skips the assignment, "if (c1->write_bandwidth && c2->write_bandwidth)" in __cxl_coordinates_combine(), is pre-existing and untouched by this patch. What the patch changes is only what remains in the output when that guard
skips: stack residue before, zero now. A zero bandwidth is how that helper already spells "not reported" - that is why the guard exists, since min() against an unknown value is meaningless. Making the combine step fall back to one input instead would change the meaning of the helper for all of its callers, which is a separate discussion from not reading uninitialized stack. The cover letter says the same about the aliasing question raised on v2.

Best Regards,
Guixin Liu

maintainer
#4 lore raw

Re: [PATCH v3 1/2] cxl/cdat: Fix uninitialized stack use in bandwidth gathering

In reply to Guixin Liu
Quoted 26 lines · Guixin Liu

cxl_endpoint_gather_bandwidth() declares three access_coordinate arrays on the stack - pci_coord, sw_coord and ep_coord - without initializing them, and relies on its helpers to fill every member. None of them does. cxl_pci_get_bandwidth() and cxl_port_get_switch_dport_bandwidth() assign only read_bandwidth and write_bandwidth, and __cxl_coordinates_combine() assigns an output bandwidth only when both input bandwidths are non-zero. Every member a producer declines to set is read back as whatever was on the stack. cxl_switch_gather_bandwidth() has the same defect in its coords array, which it hands to cxl_pci_get_bandwidth() and then to cxl_coordinates_combine().

Both cases occur on real topologies. The latency members are never written by any of those producers, yet __cxl_coordinates_combine() sums them unconditionally, so the latency reads are undefined on every call. And a device whose CDAT DSLBIS reports no bandwidth for an access class leaves perf->cdat_coord zero for that class, which is exactly the condition that makes __cxl_coordinates_combine() skip the bandwidth assignment and leave the ep_coord entry untouched.

The ep_coord case escapes the function: cxl_bandwidth_add() accumulates it into the per-upstream-port aggregate that is published through the region's access coordinate sysfs attributes, so stack contents are reported to userspace as a bandwidth figure. The latency sums are discarded by cxl_bandwidth_add() rather than published, but they are still computed from uninitialized memory.

Quoted 4 lines

Zero initialize the four arrays. Zero is already the value this code uses for "not reported" - both the __cxl_coordinates_combine() guard and coordinates_valid() test for it - so a member no producer sets now reads back as unknown rather than as a plausible number.

Hi Guixin Liu,

Please see my comments on the einj_inject patch about commit message clarity and level of detail. The same issue applies here.
https://lore.kernel.org/linux-cxl/[email protected]/

For this patch, I think the essential story is:

Background:
cxl_endpoint_get_perf_coordinates() returns success for RCDs without calculating coordinates.

Issue:
The output array is left uninitialized on that path.

Impact:
Callers can consume and expose garbage access-coordinate values.

Resolution:
Initialize the coordinates to zero before returning for an RCD.

That is enough to explain why the patch is needed. I do not need the full path through CDAT parsing, __cxl_coordinates_combine(), QoS class selection, cxl_dpa_perf, and sysfs to understand the bug.

Also, as with the other patch, please say how this was found and how it was tested.

-- Alison

Quoted 36 lines
Fixes a5ab0de0ebaa ("cxl: Calculate region bandwidth of targets with shared upstream link")
---
 drivers/cxl/core/cdat.c | 8 ++++----
 1 file changed, 4 insertions(+), 4 deletions(-)
drivers/cxl/core/cdat.c +4 -4
index 5c9f07262513..df9261ba84e5 100644
@@ -633,9 +633,9 @@ static int cxl_endpoint_gather_bandwidth(struct cxl_region *cxlr,
 	struct cxl_port *endpoint = to_cxl_port(cxled->cxld.dev.parent);
 	struct cxl_port *parent_port = to_cxl_port(endpoint->dev.parent);
 	struct cxl_port *gp_port = to_cxl_port(parent_port->dev.parent);
-	struct access_coordinate pci_coord[ACCESS_COORDINATE_MAX];
-	struct access_coordinate sw_coord[ACCESS_COORDINATE_MAX];
-	struct access_coordinate ep_coord[ACCESS_COORDINATE_MAX];
+	struct access_coordinate pci_coord[ACCESS_COORDINATE_MAX] = { };
+	struct access_coordinate sw_coord[ACCESS_COORDINATE_MAX] = { };
+	struct access_coordinate ep_coord[ACCESS_COORDINATE_MAX] = { };
 	struct cxl_memdev *cxlmd = cxled_to_memdev(cxled);
 	struct cxl_dev_state *cxlds = cxlmd->cxlds;
 	struct pci_dev *pdev = to_pci_dev(cxlds->dev);
@@ -757,7 +757,7 @@ static struct xarray *cxl_switch_gather_bandwidth(struct cxl_region *cxlr,
 {
 	struct xarray *res_xa __free(free_perf_xa) =
 		kzalloc_obj(*res_xa);
-	struct access_coordinate coords[ACCESS_COORDINATE_MAX];
+	struct access_coordinate coords[ACCESS_COORDINATE_MAX] = { };
 	struct cxl_perf_ctx *ctx, *us_ctx;
 	unsigned long index, us_index;
 	int dev_count = 0;
Signature · 1 line
2.43.7
maintainer
#5 lore raw

Re: [PATCH v3 1/2] cxl/cdat: Fix uninitialized stack use in bandwidth gathering

Quoted 33 lines · Alison Schofield
Quoted 26 lines · Guixin Liu

cxl_endpoint_gather_bandwidth() declares three access_coordinate arrays on the stack - pci_coord, sw_coord and ep_coord - without initializing them, and relies on its helpers to fill every member. None of them does. cxl_pci_get_bandwidth() and cxl_port_get_switch_dport_bandwidth() assign only read_bandwidth and write_bandwidth, and __cxl_coordinates_combine() assigns an output bandwidth only when both input bandwidths are non-zero. Every member a producer declines to set is read back as whatever was on the stack. cxl_switch_gather_bandwidth() has the same defect in its coords array, which it hands to cxl_pci_get_bandwidth() and then to cxl_coordinates_combine().

Both cases occur on real topologies. The latency members are never written by any of those producers, yet __cxl_coordinates_combine() sums them unconditionally, so the latency reads are undefined on every call. And a device whose CDAT DSLBIS reports no bandwidth for an access class leaves perf->cdat_coord zero for that class, which is exactly the condition that makes __cxl_coordinates_combine() skip the bandwidth assignment and leave the ep_coord entry untouched.

The ep_coord case escapes the function: cxl_bandwidth_add() accumulates it into the per-upstream-port aggregate that is published through the region's access coordinate sysfs attributes, so stack contents are reported to userspace as a bandwidth figure. The latency sums are discarded by cxl_bandwidth_add() rather than published, but they are still computed from uninitialized memory.

Quoted 4 lines

Zero initialize the four arrays. Zero is already the value this code uses for "not reported" - both the __cxl_coordinates_combine() guard and coordinates_valid() test for it - so a member no producer sets now reads back as unknown rather than as a plausible number.

Quoted 7 lines

Hi Guixin Liu,

Please see my comments on the einj_inject patch about commit message clarity and level of detail. The same issue applies here.
https://lore.kernel.org/linux-cxl/[email protected]/

For this patch, I think the essential story is:

Oops, I just criss-crossed your 2 patches. This goes with Patch 2/2. Similar feedback on both patches.
Thanks,

-- Alison

Quoted 61 lines

Background:
cxl_endpoint_get_perf_coordinates() returns success for RCDs without calculating coordinates.

Issue:
The output array is left uninitialized on that path.

Impact:
Callers can consume and expose garbage access-coordinate values.

Resolution:
Initialize the coordinates to zero before returning for an RCD.

That is enough to explain why the patch is needed. I do not need the full path through CDAT parsing, __cxl_coordinates_combine(), QoS class selection, cxl_dpa_perf, and sysfs to understand the bug.

Also, as with the other patch, please say how this was found and how it was tested.

-- Alison

Quoted 36 lines
Fixes a5ab0de0ebaa ("cxl: Calculate region bandwidth of targets with shared upstream link")
---
 drivers/cxl/core/cdat.c | 8 ++++----
 1 file changed, 4 insertions(+), 4 deletions(-)
drivers/cxl/core/cdat.c +4 -4
index 5c9f07262513..df9261ba84e5 100644
@@ -633,9 +633,9 @@ static int cxl_endpoint_gather_bandwidth(struct cxl_region *cxlr,
 	struct cxl_port *endpoint = to_cxl_port(cxled->cxld.dev.parent);
 	struct cxl_port *parent_port = to_cxl_port(endpoint->dev.parent);
 	struct cxl_port *gp_port = to_cxl_port(parent_port->dev.parent);
-	struct access_coordinate pci_coord[ACCESS_COORDINATE_MAX];
-	struct access_coordinate sw_coord[ACCESS_COORDINATE_MAX];
-	struct access_coordinate ep_coord[ACCESS_COORDINATE_MAX];
+	struct access_coordinate pci_coord[ACCESS_COORDINATE_MAX] = { };
+	struct access_coordinate sw_coord[ACCESS_COORDINATE_MAX] = { };
+	struct access_coordinate ep_coord[ACCESS_COORDINATE_MAX] = { };
 	struct cxl_memdev *cxlmd = cxled_to_memdev(cxled);
 	struct cxl_dev_state *cxlds = cxlmd->cxlds;
 	struct pci_dev *pdev = to_pci_dev(cxlds->dev);
@@ -757,7 +757,7 @@ static struct xarray *cxl_switch_gather_bandwidth(struct cxl_region *cxlr,
 {
 	struct xarray *res_xa __free(free_perf_xa) =
 		kzalloc_obj(*res_xa);
-	struct access_coordinate coords[ACCESS_COORDINATE_MAX];
+	struct access_coordinate coords[ACCESS_COORDINATE_MAX] = { };
 	struct cxl_perf_ctx *ctx, *us_ctx;
 	unsigned long index, us_index;
 	int dev_count = 0;
Signature · 1 line
2.43.7
author
#6 lore raw

Re: [PATCH v3 1/2] cxl/cdat: Fix uninitialized stack use in bandwidth gathering

In reply to Alison Schofield
Quoted 53 lines · Alison Schofield
Quoted 26 lines · Guixin Liu

cxl_endpoint_gather_bandwidth() declares three access_coordinate arrays on the stack - pci_coord, sw_coord and ep_coord - without initializing them, and relies on its helpers to fill every member. None of them does. cxl_pci_get_bandwidth() and cxl_port_get_switch_dport_bandwidth() assign only read_bandwidth and write_bandwidth, and __cxl_coordinates_combine() assigns an output bandwidth only when both input bandwidths are non-zero. Every member a producer declines to set is read back as whatever was on the stack. cxl_switch_gather_bandwidth() has the same defect in its coords array, which it hands to cxl_pci_get_bandwidth() and then to cxl_coordinates_combine().

Both cases occur on real topologies. The latency members are never written by any of those producers, yet __cxl_coordinates_combine() sums them unconditionally, so the latency reads are undefined on every call. And a device whose CDAT DSLBIS reports no bandwidth for an access class leaves perf->cdat_coord zero for that class, which is exactly the condition that makes __cxl_coordinates_combine() skip the bandwidth assignment and leave the ep_coord entry untouched.

The ep_coord case escapes the function: cxl_bandwidth_add() accumulates it into the per-upstream-port aggregate that is published through the region's access coordinate sysfs attributes, so stack contents are reported to userspace as a bandwidth figure. The latency sums are discarded by cxl_bandwidth_add() rather than published, but they are still computed from uninitialized memory.

Quoted 4 lines

Zero initialize the four arrays. Zero is already the value this code uses for "not reported" - both the __cxl_coordinates_combine() guard and coordinates_valid() test for it - so a member no producer sets now reads back as unknown rather than as a plausible number.

Hi Guixin Liu,

Please see my comments on the einj_inject patch about commit message clarity and level of detail. The same issue applies here.
https://lore.kernel.org/linux-cxl/[email protected]/

For this patch, I think the essential story is:

Background:
cxl_endpoint_get_perf_coordinates() returns success for RCDs without calculating coordinates.

Issue:
The output array is left uninitialized on that path.

Impact:
Callers can consume and expose garbage access-coordinate values.

Resolution:
Initialize the coordinates to zero before returning for an RCD.

Quoted 8 lines

That is enough to explain why the patch is needed. I do not need the full path through CDAT parsing, __cxl_coordinates_combine(), QoS class selection, cxl_dpa_perf, and sysfs to understand the bug.

Also, as with the other patch, please say how this was found and how it was tested.

-- Alison

Sure, I will cut the commit and explain how I found it and how it was tested in v4.

Best Regards,
Guixin Liu

Quoted 37 lines
Quoted 37 lines
Fixes a5ab0de0ebaa ("cxl: Calculate region bandwidth of targets with shared upstream link")
---
  drivers/cxl/core/cdat.c | 8 ++++----
  1 file changed, 4 insertions(+), 4 deletions(-)
drivers/cxl/core/cdat.c +4 -4
index 5c9f07262513..df9261ba84e5 100644
@@ -633,9 +633,9 @@ static int cxl_endpoint_gather_bandwidth(struct cxl_region *cxlr,
  	struct cxl_port *endpoint = to_cxl_port(cxled->cxld.dev.parent);
  	struct cxl_port *parent_port = to_cxl_port(endpoint->dev.parent);
  	struct cxl_port *gp_port = to_cxl_port(parent_port->dev.parent);
-	struct access_coordinate pci_coord[ACCESS_COORDINATE_MAX];
-	struct access_coordinate sw_coord[ACCESS_COORDINATE_MAX];
-	struct access_coordinate ep_coord[ACCESS_COORDINATE_MAX];
+	struct access_coordinate pci_coord[ACCESS_COORDINATE_MAX] = { };
+	struct access_coordinate sw_coord[ACCESS_COORDINATE_MAX] = { };
+	struct access_coordinate ep_coord[ACCESS_COORDINATE_MAX] = { };
  	struct cxl_memdev *cxlmd = cxled_to_memdev(cxled);
  	struct cxl_dev_state *cxlds = cxlmd->cxlds;
  	struct pci_dev *pdev = to_pci_dev(cxlds->dev);
@@ -757,7 +757,7 @@ static struct xarray *cxl_switch_gather_bandwidth(struct cxl_region *cxlr,
  {
  	struct xarray *res_xa __free(free_perf_xa) =
  		kzalloc_obj(*res_xa);
-	struct access_coordinate coords[ACCESS_COORDINATE_MAX];
+	struct access_coordinate coords[ACCESS_COORDINATE_MAX] = { };
  	struct cxl_perf_ctx *ctx, *us_ctx;
  	unsigned long index, us_index;
  	int dev_count = 0;
Signature · 1 line
2.43.7
author
#7 lore raw

PATCH v3 2/2 cxl/port: Fix uninitialized coordinates reported for RCDs

In reply to Guixin Liu

cxl_endpoint_get_perf_coordinates() walks the CXL path and copies the accumulated numbers into the caller's array with set_access_coordinates() just before it returns. For a Restricted CXL Device it never gets that far:
RCH topologies are covered by HMAT and cannot be hotplugged, so the walk is skipped and the function returns 0 with the output array untouched. Every caller is told the coordinates are valid while nothing was written to them.

Its only consumer, cxl_port_perf_data_calculate(), keeps the array on the stack and does not pre-zero it, and endpoint CDAT parsing is not gated on the topology: cxl_endpoint_port_probe() calls read_cdat_data() and cxl_endpoint_parse_cdat() for every endpoint port, RCD included. So an RCD that exposes a CDAT combines stack residue with the values from its DSMAS entries.

The residue does not stay in the driver. __cxl_coordinates_combine() sums the latencies unconditionally and takes the minimum of the bandwidths when both are non-zero, so it lands in dent->coord, which selects a QoS class through cxl_root->ops.qos_class() and is copied into the memdev's cxl_dpa_perf, where cxl_memdev_update_perf() publishes it through the memdev's access coordinate sysfs attributes.

Clear the output on the RCD path so the function honours its contract for every caller. Zeroing at the single caller instead would leave the exported helper free to report success without writing anything, which the next caller would have to know about; and zero is the value this code already uses for "not reported", so an RCD now reads back as unknown rather than as plausible numbers.

Fixes 5d211c709059 ("cxl: Fix cxl_endpoint_get_perf_coordinate() support for RCH")
---
 drivers/cxl/core/port.c | 4 +++-
 1 file changed, 3 insertions(+), 1 deletion(-)
drivers/cxl/core/port.c +3 -1
index 625e4aa427db..fea43a92744c 100644
@@ -2394,8 +2394,10 @@ int cxl_endpoint_get_perf_coordinates(struct cxl_port *port,
 	 * Skip calculation for RCD. Expectation is HMAT already covers RCD case
 	 * since RCH does not support hotplug.
 	 */
-	if (cxlmd->cxlds->rcd)
+	if (cxlmd->cxlds->rcd) {
+		memset(coord, 0, sizeof(*coord) * ACCESS_COORDINATE_MAX);
 		return 0;
+	}
 
 	/*
 	 * Exit the loop when the parent port of the current iter port is cxl
Signature · 1 line
2.43.7
Reviewed-by on the v3 series
#8 lore raw
In reply to Guixin Liu
Quoted 8 lines · Guixin Liu

Two related fixes for stack access_coordinate arrays that are read before anything writes them, both reachable on real topologies and both ending up in userspace through the access coordinate sysfs attributes.

Patch 1 is the v2 cdat fix [1] with one more site covered: the coords array in cxl_switch_gather_bandwidth() has the same defect as the three arrays in cxl_endpoint_gather_bandwidth().

Quoted 9 lines

Patch 2 is new. cxl_endpoint_get_perf_coordinates() returns 0 for an RCD without writing the caller's array at all, so cxl_port_perf_data_calculate() combines stack residue with the values parsed from the RCD's CDAT.

Both were pointed out by the Sashiko review bot on the v2 posting of patch 1. The third finding in that review - cxl_coordinates_combine() aliasing its output onto an input and leaving a member alone when an input bandwidth is zero - is the intended "not reported" behaviour and is not addressed here.

Hi Guixin,

Thanks for this, both changes LGTM.

Best regards,
Richard Cheng.

Quoted 29 lines

[1] https://lore.kernel.org/linux-cxl/[email protected]/

v2->v3:
- also zero initialize coords in cxl_switch_gather_bandwidth(), same defect in the sibling gather path (Sashiko bot)
- new patch 2: clear the output coordinates on the RCD path of cxl_endpoint_get_perf_coordinates(), which currently returns success without writing them (Sashiko bot)
- drop "endpoint" from the patch 1 subject, it now covers both gather paths

v1->v2:
- rebase onto cxl/next
- rewrite the commit message to describe the behaviour rather than narrate the code change (Alison Schofield)

Guixin Liu (2):
cxl/cdat: Fix uninitialized stack use in bandwidth gathering cxl/port: Fix uninitialized coordinates reported for RCDs

 drivers/cxl/core/cdat.c | 8 ++++----
 drivers/cxl/core/port.c | 4 +++-
 2 files changed, 7 insertions(+), 5 deletions(-)

base-commit: 7098e9cd98a05c0c5de2fae0c2465f9d966fdd07

Signature · 1 line
2.43.7