Skip to contentSkip to conversation

EDAC/versalnet: Fix teardown ordering in mc_remove()

PATCH Applied Prasanna Kumar T S M 3 participants 12 messages Apr 1 – Jul 9

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

PATCH v2 1/3 EDAC/versalnet: Fix teardown ordering in mc_remove()

The teardown sequence in mc_remove() does not mirror the reverse of the initialization order in mc_probe(). In particular, unregister_rpmsg_driver() is called before remove_versalnet(), and cdx_mcdi_finish() is called after rproc_shutdown().

Reorder mc_remove() to reverse the probe initialization sequence, consistent with the probe error-unwind paths.

The rproc reference acquired via rproc_get_by_phandle() during probe is not released in mc_remove(), causing a reference count leak. Add the missing rproc_put() call.

Fixes d5fe2fec6c40 ("EDAC: Add a driver for the AMD Versal NET DDR controller")
---
 drivers/edac/versalnet_edac.c | 5 +++--
 1 file changed, 3 insertions(+), 2 deletions(-)
drivers/edac/versalnet_edac.c +3 -2
index b87fe57aa842..acd51b492772 100644
@@ -955,10 +955,11 @@ static void mc_remove(struct platform_device *pdev)
 {
 	struct mc_priv *priv = platform_get_drvdata(pdev);
 
-	unregister_rpmsg_driver(&amd_rpmsg_driver);
 	remove_versalnet(priv);
-	rproc_shutdown(priv->mcdi->r5_rproc);
 	cdx_mcdi_finish(priv->mcdi);
+	unregister_rpmsg_driver(&amd_rpmsg_driver);
+	rproc_shutdown(priv->mcdi->r5_rproc);
+	rproc_put(priv->mcdi->r5_rproc);
 	kfree(priv->mcdi);
 }
 
Signature · 1 line
2.49.0
#1 lore raw

PATCH v2 2/3 EDAC/versalnet: Fix device name memory leak

The device name allocated via kzalloc() in init_one_mc() is assigned to dev->init_name but never freed on the normal removal path. device_register() copies init_name and then sets dev->init_name to NULL, so the name pointer becomes unreachable from the device. Thus leaking memory.

Use a stack-local char array instead of using kzalloc() for name.

Fixes d5fe2fec6c40 ("EDAC: Add a driver for the AMD Versal NET DDR controller")
---
 drivers/edac/versalnet_edac.c | 10 ++--------
 1 file changed, 2 insertions(+), 8 deletions(-)
drivers/edac/versalnet_edac.c +2 -8
index acd51b492772..012c4f40994d 100644
@@ -779,7 +779,7 @@ static int init_one_mc(struct mc_priv *priv, struct platform_device *pdev, int i
 	struct mem_ctl_info *mci;
 	struct device *dev;
 	enum dev_type dt;
-	char *name;
+	char name[MC_NAME_LEN];
 	int rc;
 
 	config = priv->adec[CONF + i * ADEC_NUM];
@@ -813,13 +813,9 @@ static int init_one_mc(struct mc_priv *priv, struct platform_device *pdev, int i
 	layers[1].is_virt_csrow = false;
 
 	rc = -ENOMEM;
-	name = kzalloc(MC_NAME_LEN, GFP_KERNEL);
-	if (!name)
-		return rc;
-
 	dev = kzalloc(sizeof(*dev), GFP_KERNEL);
 	if (!dev)
-		goto err_name_free;
+		return rc;
 
 	mci = edac_mc_alloc(i, ARRAY_SIZE(layers), layers, sizeof(struct mc_priv));
 	if (!mci) {
@@ -858,8 +854,6 @@ static int init_one_mc(struct mc_priv *priv, struct platform_device *pdev, int i
 	edac_mc_free(mci);
 err_dev_free:
 	kfree(dev);
-err_name_free:
-	kfree(name);
 
 	return rc;
 }
Signature · 1 line
2.49.0
maintainer Applied
#2 lore raw

Re: [PATCH v2 2/3] EDAC/versalnet: Fix device name memory leak

Quoted 6 lines · Prasanna Kumar T S M

The device name allocated via kzalloc() in init_one_mc() is assigned to dev->init_name but never freed on the normal removal path. device_register() copies init_name and then sets dev->init_name to NULL, so the name pointer becomes unreachable from the device. Thus leaking memory.

Quoted 8 lines

Use a stack-local char array instead of using kzalloc() for name.

Fixes d5fe2fec6c40 ("EDAC: Add a driver for the AMD Versal NET DDR controller")
---
 drivers/edac/versalnet_edac.c | 10 ++--------
 1 file changed, 2 insertions(+), 8 deletions(-)

Applied, thanks.

Signature · 4 lines
Regards/Gruss,
    Boris.

https://people.kernel.org/tglx/notes-about-netiquette
#3 lore raw

PATCH v2 3/3 EDAC/versalnet: Fix device_register() error handling in init_one_mc()

In reply to Prasanna Kumar T S M

When device_register() fails, it must be followed by put_device() rather than kfree(), because device_register() calls device_initialize() which sets up the device refcount. The matching release function versal_edac_release() handles the actual kfree().

To simplify error handling and avoid complex unwinding, split device_register() into device_initialize() and device_add(). Initialize the device early so put_device() can be used in all error paths.

Fixes d5fe2fec6c40 ("EDAC: Add a driver for the AMD Versal NET DDR controller")
---
 drivers/edac/versalnet_edac.c | 20 ++++++++++----------
 1 file changed, 10 insertions(+), 10 deletions(-)
drivers/edac/versalnet_edac.c +10 -10
index 012c4f40994d..94580d3c3170 100644
@@ -780,7 +780,7 @@ static int init_one_mc(struct mc_priv *priv, struct platform_device *pdev, int i
 	struct device *dev;
 	enum dev_type dt;
 	char name[MC_NAME_LEN];
-	int rc;
+	int rc = -ENOMEM;
 
 	config = priv->adec[CONF + i * ADEC_NUM];
 	num_chans = FIELD_GET(MC5_NUM_CHANS_MASK, config);
@@ -812,23 +812,23 @@ static int init_one_mc(struct mc_priv *priv, struct platform_device *pdev, int i
 	layers[1].size = num_chans;
 	layers[1].is_virt_csrow = false;
 
-	rc = -ENOMEM;
 	dev = kzalloc(sizeof(*dev), GFP_KERNEL);
 	if (!dev)
 		return rc;
 
+	sprintf(name, "versal-net-ddrmc5-edac-%d", i);
+
+	dev->init_name = name;
+	dev->release = versal_edac_release;
+	device_initialize(dev);
+
 	mci = edac_mc_alloc(i, ARRAY_SIZE(layers), layers, sizeof(struct mc_priv));
 	if (!mci) {
 		edac_printk(KERN_ERR, EDAC_MC, "Failed memory allocation for MC%d\n", i);
 		goto err_dev_free;
 	}
 
-	sprintf(name, "versal-net-ddrmc5-edac-%d", i);
-
-	dev->init_name = name;
-	dev->release = versal_edac_release;
-
-	rc = device_register(dev);
+	rc = device_add(dev);
 	if (rc)
 		goto err_mc_free;
 
@@ -849,11 +849,11 @@ static int init_one_mc(struct mc_priv *priv, struct platform_device *pdev, int i
 	return 0;
 
 err_unreg:
-	device_unregister(mci->pdev);
+	device_del(dev);
 err_mc_free:
 	edac_mc_free(mci);
 err_dev_free:
-	kfree(dev);
+	put_device(dev);
 
 	return rc;
 }
Signature · 1 line
2.49.0
#4 lore raw

Re: [PATCH v2 3/3] EDAC/versalnet: Fix device_register() error handling in init_one_mc()

Quoted 33 lines · Prasanna Kumar T S M

When device_register() fails, it must be followed by put_device() rather than kfree(), because device_register() calls device_initialize() which sets up the device refcount. The matching release function versal_edac_release() handles the actual kfree().

To simplify error handling and avoid complex unwinding, split device_register() into device_initialize() and device_add(). Initialize the device early so put_device() can be used in all error paths.

Fixes d5fe2fec6c40 ("EDAC: Add a driver for the AMD Versal NET DDR controller")
---
  drivers/edac/versalnet_edac.c | 20 ++++++++++----------
  1 file changed, 10 insertions(+), 10 deletions(-)
drivers/edac/versalnet_edac.c +1 -1
index 012c4f40994d..94580d3c3170 100644
@@ -780,7 +780,7 @@ static int init_one_mc(struct mc_priv *priv, struct platform_device *pdev, int i
  	struct device *dev;
  	enum dev_type dt;
  	char name[MC_NAME_LEN];
-	int rc;
+	int rc = -ENOMEM;
 
  	config = priv->adec[CONF + i * ADEC_NUM];
  	num_chans = FIELD_GET(MC5_NUM_CHANS_MASK, config);
@@ -812,23 +812,23 @@ static int init_one_mc(struct mc_priv *priv, struct platform_device *pdev, int i
  	layers[1].size = num_chans;

drivers/edac/versalnet_edac.c static int init_one_mc(struct mc_priv *priv, struct platform_device *pdev, int i · line 821

Quoted 10 lines
drivers/edac/versalnet_edac.c +3 -1
  	layers[1].is_virt_csrow = false;
 
-	rc = -ENOMEM;
  	dev = kzalloc(sizeof(*dev), GFP_KERNEL);
  	if (!dev)
  		return rc;
 
+	sprintf(name, "versal-net-ddrmc5-edac-%d", i);
+
+	dev->init_name = name;

Sashiko's review is valid
-------------------------
If edac_mc_alloc() fails below, we jump to err_dev_free, call put_device(dev), and return from init_one_mc(). Since kobject cleanup can be deferred asynchronously (such as when CONFIG_DEBUG_KOBJECT_RELEASE is enabled), and name is a local stack array, could this result in a use-after-free if dev_name(dev) is accessed during cleanup?
Would it be safer to use dev_set_name(dev, "versal-net-ddrmc5-edac-%d", i) instead of assigning dev->init_name directly?
-------------------------

This is a valid issue, I will fix it in the next iteration.

Quoted 33 lines
drivers/edac/versalnet_edac.c +6 -8
+	dev->release = versal_edac_release;
+	device_initialize(dev);
+
  	mci = edac_mc_alloc(i, ARRAY_SIZE(layers), layers, sizeof(struct mc_priv));
  	if (!mci) {
  		edac_printk(KERN_ERR, EDAC_MC, "Failed memory allocation for MC%d\n", i);
  		goto err_dev_free;
  	}
 
-	sprintf(name, "versal-net-ddrmc5-edac-%d", i);
-
-	dev->init_name = name;
-	dev->release = versal_edac_release;
-
-	rc = device_register(dev);
+	rc = device_add(dev);
  	if (rc)
  		goto err_mc_free;
 
@@ -849,11 +849,11 @@ static int init_one_mc(struct mc_priv *priv, struct platform_device *pdev, int i
  	return 0;
 
  err_unreg:
-	device_unregister(mci->pdev);
+	device_del(dev);
  err_mc_free:
  	edac_mc_free(mci);
  err_dev_free:
-	kfree(dev);
+	put_device(dev);
 
  	return rc;
  }
maintainer
#5 lore raw
In reply to Prasanna Kumar T S M
Quoted 12 lines · Prasanna Kumar T S M

The teardown sequence in mc_remove() does not mirror the reverse of the initialization order in mc_probe(). In particular, unregister_rpmsg_driver() is called before remove_versalnet(), and cdx_mcdi_finish() is called after rproc_shutdown().

Reorder mc_remove() to reverse the probe initialization sequence, consistent with the probe error-unwind paths.

The rproc reference acquired via rproc_get_by_phandle() during probe is not released in mc_remove(), causing a reference count leak. Add the missing rproc_put() call.

Quoted 5 lines
Fixes d5fe2fec6c40 ("EDAC: Add a driver for the AMD Versal NET DDR controller")
---
 drivers/edac/versalnet_edac.c | 5 +++--
 1 file changed, 3 insertions(+), 2 deletions(-)

Sashiko has found things, pls addres them:

https://sashiko.dev/#/patchset/20260401111836.2342918-1-ptsm%40linux.microsoft.com

Signature · 4 lines
Regards/Gruss,
    Boris.

https://people.kernel.org/tglx/notes-about-netiquette
#6 lore raw
Quoted 19 lines · Borislav Petkov
Quoted 12 lines · Prasanna Kumar T S M

The teardown sequence in mc_remove() does not mirror the reverse of the initialization order in mc_probe(). In particular, unregister_rpmsg_driver() is called before remove_versalnet(), and cdx_mcdi_finish() is called after rproc_shutdown().

Reorder mc_remove() to reverse the probe initialization sequence, consistent with the probe error-unwind paths.

The rproc reference acquired via rproc_get_by_phandle() during probe is not released in mc_remove(), causing a reference count leak. Add the missing rproc_put() call.

Quoted 5 lines
Fixes d5fe2fec6c40 ("EDAC: Add a driver for the AMD Versal NET DDR controller")
---
  drivers/edac/versalnet_edac.c | 5 +++--
  1 file changed, 3 insertions(+), 2 deletions(-)
Quoted 4 lines

Sashiko has found things, pls addres them:

https://sashiko.dev/#/patchset/20260401111836.2342918-1-ptsm%40linux.microsoft.com

I asked AI to validate Sashiko's comment. This is its output
------------------------------------------------------------

Analysis

The review comment is not valid (false positive). Here's the detailed
reasoning:

Data structure hierarchy

mc_priv

     └── mcdi: struct cdx_mcdi *            ← kfree'd at step 6 (LAST)
           ├── mcdi: struct cdx_mcdi_data *  ← kfree'd by 
cdx_mcdi_finish() at step 2
           │     └── iface: struct cdx_mcdi_iface (mutex, cmd list, etc.)
           ├── r5_rproc
           └── ...

New mc_remove() order (after commit)

(1) remove_versalnet(priv); // tear down EDAC
(2) cdx_mcdi_finish(priv->mcdi); // wait_for_cleanup →
kfree(cdx->mcdi) → cdx->mcdi = NULL
(3) unregister_rpmsg_driver(&amd_rpmsg_driver); // destroys rpmsg
endpoint
(4) rproc_shutdown(priv->mcdi->r5_rproc);
(5) rproc_put(priv->mcdi->r5_rproc);
(6) kfree(priv->mcdi); // frees outer struct

Why the concern doesn't hold

The reviewer's premise is wrong. The comment says "after priv->mcdi
is freed" — but priv->mcdi (the struct cdx_mcdi * pointer passed to cdx_mcdi_process_cmd) is not freed
until step 6, well after unregister_rpmsg_driver() at step 3.

What IS freed at step 2 is the inner cdx->mcdi (struct cdx_mcdi_data
*). But cdx_mcdi_finish() also sets cdx->mcdi = NULL. If rpmsg_cb() fires between steps 2 and 3:

1. mc_priv->mcdi → still points to valid struct cdx_mcdi (not freed
until step 6)
2. cdx_mcdi_process_cmd(mc_priv->mcdi, ...) is called with a valid
pointer
3. Inside, cdx_mcdi_if(cdx) checks cdx->mcdi → NULL → returns NULL
4. if (!mcdi) return; → exits early, safe

Additionally, cdx_mcdi_wait_for_cleanup() inside cdx_mcdi_finish()
ensures all pending MCDI commands have completed before freeing, so no legitimate MCDI responses should
arrive afterward.

This ordering exactly matches the probe error unwind path (err_init:
does cdx_mcdi_finish → kfree → unregister_rpmsg_driver → rproc_shutdown → rproc_put), which the
commit explicitly states it's aligning with.

Verdict: False positive. The NULL guard in cdx_mcdi_if() /
cdx_mcdi_process_cmd() protects the window, and the outer struct priv->mcdi remains valid throughout.

------------------------------------------------------------

So the comment looks like a false positive. It will be great if someone from AMD verifies this.

Can you please look at the other 2 patches in this series? They are independent of this change.

Thanks,
Prasanna Kumar

maintainer
#7 lore raw
Quoted 6 lines · Prasanna Kumar T S M
Quoted 4 lines

Sashiko has found things, pls addres them:

https://sashiko.dev/#/patchset/20260401111836.2342918-1-ptsm%40linux.microsoft.com

I asked AI to validate Sashiko's comment.

So I asked *you* to address the Sashiko review. You go and ask another AI to validate the review.

Now, if I go and paste the whole conversation to a third AI, the f*ckup is complete.

Tell me: is that the goal of this exercise? Let AIs do the thinking for us and we can all go shopping?

Or maybe *you* should go review Sashiko's review and say, I agree because <proper, comprehensible explanation> or I don't agree because <also proper, comprehensible explanation>.

Then *I* go and verify that.

Hmm...

Signature · 4 lines
Regards/Gruss,
    Boris.

https://people.kernel.org/tglx/notes-about-netiquette
#8 lore raw
Quoted 17 lines · Borislav Petkov
Quoted 6 lines · Prasanna Kumar T S M
Quoted 4 lines

Sashiko has found things, pls addres them:

https://sashiko.dev/#/patchset/20260401111836.2342918-1-ptsm%40linux.microsoft.com

I asked AI to validate Sashiko's comment.

So I asked *you* to address the Sashiko review. You go and ask another AI to validate the review.

Now, if I go and paste the whole conversation to a third AI, the f*ckup is complete.

Tell me: is that the goal of this exercise? Let AIs do the thinking for us and we can all go shopping?

Quoted 8 lines

Or maybe *you* should go review Sashiko's review and say, I agree because <proper, comprehensible explanation> or I don't agree because <also proper, comprehensible explanation>.

Then *I* go and verify that.

Hmm...

Hi,

Sashiko review says
-------------------
If an MCDI response message arrives after priv->mcdi is freed but before the RPMSG endpoint is destroyed, could the still-active rpmsg_cb() pass the dangling priv->mcdi pointer to cdx_mcdi_process_cmd(), leading to a use-after-free?
-------------------

The review is based on the assumption that priv->mcdi is freed before unregister_rpmsg_driver(). But priv->mcdi is valid till the end of the function.

The cdx_mcdi_finish() waits for the mcdi->workqueue to finish processing and the mcdi->workqueue is destroyed. Any subsequent rpmsg_cb() calls cdx_mcdi_process_cmd() which safely without doing anything if mcdi (priv->mcdi->mcdi) is NULL. The review can be safely ignored. Other path in rpmsg_cb() accesses priv->adec and priv->regs, both of them are valid and doesn't cause any use-after-free issue.

The review is a false positive and can be ignored.

Thanks,
Prasanna Kumar

maintainer
#9 lore raw
Quoted 7 lines · Prasanna Kumar T S M

Sashiko review says
-------------------
If an MCDI response message arrives after priv->mcdi is freed but before the RPMSG endpoint is destroyed, could the still-active rpmsg_cb() pass the dangling priv->mcdi pointer to cdx_mcdi_process_cmd(), leading to a use-after-free?
-------------------

Sashiko says:

"If an error message or MCDI response arrives from the remote processor during this race window, rpmsg_cb() will execute concurrently. When processing an error, it calls handle_error(), which dereferences the memory controllers already freed by edac_mc_free() in remove_versalnet()."

IOW, handle_error() does:

	mci = priv->mci[ctl_num];

however, remove_versalnet() has *done* *already*:

        mci = priv->mci[i];
        device_unregister(mci->pdev);
        edac_mc_del_mc(mci->pdev);
        edac_mc_free(mci);

How is that a false positive if you change the teardown sequence to:

remove_versalnet
cdx_mcdi_finish
unregister_rpmsg_driver
rproc_shutdown
rproc_put

?

IOW, the proper fix in all this gunk of unwinding sequences is to *prevent* the hw from sending *any* more messages before you start tearing down things.

I'd leave that to Shubhrajyoti to figure out how to do properly.

Thx.

Signature · 4 lines
Regards/Gruss,
    Boris.

https://people.kernel.org/tglx/notes-about-netiquette
#10 lore raw
In reply to Prasanna Kumar T S M
Quoted 8 lines · Prasanna Kumar T S M

The teardown sequence in mc_remove() does not mirror the reverse of the initialization order in mc_probe(). In particular, unregister_rpmsg_driver() is called before remove_versalnet(), and cdx_mcdi_finish() is called after rproc_shutdown().

Reorder mc_remove() to reverse the probe initialization sequence, consistent with the probe error-unwind paths.

I think that the remote proc should be quiescence first so that no more messages will be
queued. and then the edac should be removed. See below.

Quoted 16 lines

The rproc reference acquired via rproc_get_by_phandle() during probe is not released in mc_remove(), causing a reference count leak. Add the missing rproc_put() call.

Fixes d5fe2fec6c40 ("EDAC: Add a driver for the AMD Versal NET DDR controller")
---
 drivers/edac/versalnet_edac.c | 5 +++--
 1 file changed, 3 insertions(+), 2 deletions(-)
drivers/edac/versalnet_edac.c +0 -0
index b87fe57aa842..acd51b492772 100644

drivers/edac/versalnet_edac.c static void mc_remove(struct platform_device *pdev) · line 958

Quoted 5 lines
drivers/edac/versalnet_edac.c +0 -1
@@ -955,10 +955,11 @@ static void mc_remove(struct platform_device *pdev)
 {
        struct mc_priv *priv = platform_get_drvdata(pdev);
 
-       unregister_rpmsg_driver(&amd_rpmsg_driver);
        remove_versalnet(priv);

Here we are removing the edac but the remoteproc can be triggered. calling the remote callback.

drivers/edac/versalnet_edac.c static void mc_remove(struct platform_device *pdev) · line 962

Quoted 5 lines
drivers/edac/versalnet_edac.c +3 -1
-       rproc_shutdown(priv->mcdi->r5_rproc);
        cdx_mcdi_finish(priv->mcdi);
+       unregister_rpmsg_driver(&amd_rpmsg_driver);
+       rproc_shutdown(priv->mcdi->r5_rproc);
+       rproc_put(priv->mcdi->r5_rproc);

The put is a valid fix.

Quoted 7 lines
drivers/edac/versalnet_edac.c +0 -1
        kfree(priv->mcdi);
 }
 
--

2.49.0

#11 lore raw

Hi Shubrajyoti,

Quoted 34 lines · Shubhrajyoti Datta
Quoted 8 lines · Prasanna Kumar T S M

The teardown sequence in mc_remove() does not mirror the reverse of the initialization order in mc_probe(). In particular, unregister_rpmsg_driver() is called before remove_versalnet(), and cdx_mcdi_finish() is called after rproc_shutdown().

Reorder mc_remove() to reverse the probe initialization sequence, consistent with the probe error-unwind paths.

I think that the remote proc should be quiescence first so that no more messages will be
queued. and then the edac should be removed. See below.

Quoted 19 lines

The rproc reference acquired via rproc_get_by_phandle() during probe is not released in mc_remove(), causing a reference count leak. Add the missing rproc_put() call.

Fixes d5fe2fec6c40 ("EDAC: Add a driver for the AMD Versal NET DDR controller")
---
  drivers/edac/versalnet_edac.c | 5 +++--
  1 file changed, 3 insertions(+), 2 deletions(-)
drivers/edac/versalnet_edac.c +0 -0
index b87fe57aa842..acd51b492772 100644
@@ -955,10 +955,11 @@ static void mc_remove(struct platform_device *pdev)
  {
         struct mc_priv *priv = platform_get_drvdata(pdev);
 
Quoted 5 lines

drivers/edac/versalnet_edac.c static void mc_remove(struct platform_device *pdev) · line 958

Quoted 2 lines
drivers/edac/versalnet_edac.c +0 -1
-       unregister_rpmsg_driver(&amd_rpmsg_driver);
         remove_versalnet(priv);

Here we are removing the edac but the remoteproc can be triggered. calling the remote callback.

Thanks for the review.

This suggests that my change is not complete. I am not sufficiently familiar with this area to provide a correct fix, so I will leave it to someone more knowledgeable.

Quoted 15 lines

drivers/edac/versalnet_edac.c static void mc_remove(struct platform_device *pdev) · line 962

Quoted 6 lines
drivers/edac/versalnet_edac.c +3 -1
-       rproc_shutdown(priv->mcdi->r5_rproc);
         cdx_mcdi_finish(priv->mcdi);
+       unregister_rpmsg_driver(&amd_rpmsg_driver);
+       rproc_shutdown(priv->mcdi->r5_rproc);
+       rproc_put(priv->mcdi->r5_rproc);

The put is a valid fix.

Quoted 7 lines
drivers/edac/versalnet_edac.c +0 -1
         kfree(priv->mcdi);
  }
 
--

2.49.0

Thanks,
Prasanna