[PATCH] dmaengine: idxd: Fix use-after-free in idxd_remove

[PATCH] dmaengine: idxd: Fix use-after-free in idxd_remove

[Ziyou Wang (Lenovo)]

When unbinding the idxd driver, the following call trace is observed:

  general protection fault, probably for non-canonical address 0x18fc1d6937256657
  CPU: 117 UID: 0 PID: 3396 Comm: python3 Tainted: G E 7.2.0-rc5-latest+ #11 PREEMPT(lazy)
  Hardware name: Lenovo WenTian WR5220 G5, BIOS speb50m-2.54
  RIP: 0010:__refill_objects_node+0x2da/0x5e0
  Call Trace:
   <TASK>
   refill_objects+0x1ee/0x2e0
   __pcs_replace_empty_main+0x1e1/0x330
   __kmalloc_noprof+0x503/0x550
   ext4_htree_store_dirent+0x34/0x110 [ext4]
   htree_dirblock_to_tree+0x1a9/0x2c0 [ext4]
   ext4_htree_fill_tree+0x246/0x3d0 [ext4]
   ext4_readdir+0x863/0x9d0 [ext4]
   iterate_dir+0xa6/0x260
   __x64_sys_getdents64+0x76/0x130
   do_syscall_64+0x98/0x620
   entry_SYSCALL_64_after_hwframe+0x76/0x7e
   </TASK>

With the KASAN-enabled kernel, the kernel reports a use-after-free bug:

idxd_unregister_devices() calls device_unregister(wq), which triggers
the wq release callback and frees wq memory via kfree(). However,
idxd->wqs[] array still holds the freed pointer. Later, when
device_unregister(idxd) calls idxd_device_drv_remove(), it accesses
wq->state through idxd->wqs[], resulting in use-after-free.

The same issue exists for idxd->engines[] and idxd->groups[] arrays.

KASAN report:

  BUG: KASAN: slab-use-after-free in idxd_device_drv_remove+0xc1/0x100 [idxd]
  Read of size 4 at addr ff1100038d296414 by task python3/3354
  CPU: 20 UID: 0 PID: 3354 Comm: python3 Tainted: G E 7.2.0-rc5-debug+
  Hardware name: Lenovo WenTian WR5220 G5, BIOS speb50m-2.54

  Call Trace:
    <TASK>
    dump_stack_lvl+0x5b/0x80
    print_report+0x153/0x4b5
    kasan_report+0xbc/0xf0
    idxd_device_drv_remove+0xc1/0x100 [idxd]
    device_release_driver_internal+0x244/0x2e0
    bus_remove_device+0x195/0x2b0
    device_del+0x24b/0x540
    device_unregister+0x17/0x80
    idxd_remove+0x3f/0x120 [idxd]
    pci_device_remove+0x6d/0xf0
    device_release_driver_internal+0x244/0x2e0
    unbind_store+0xae/0xb0
    kernfs_fop_write_iter+0x205/0x2d0
    vfs_write+0x3a9/0x6b0
    ksys_write+0xc7/0x160
    do_syscall_64+0x9c/0x620
    entry_SYSCALL_64_after_hwframe+0x76/0x7e
    </TASK>

  Allocated by task 1219:
    kasan_save_stack+0x20/0x40
    kasan_save_track+0x14/0x30
    __kasan_kmalloc+0x9a/0xb0
    __kmalloc_cache_node_noprof+0x1c6/0x470
    idxd_pci_probe_alloc+0x103f/0x2430 [idxd]
    local_pci_probe+0x71/0xd0
    local_pci_probe_callback+0x20/0x40
    process_one_work+0x389/0x6a0
    worker_thread+0x321+0x590
    kthread+0x1b4/0x200
    ret_from_fork+0x3bd/0x4d0
    ret_from_fork_asm+0x1a/0x30

  Freed by task 3354:
    kasan_save_stack+0x20/0x40
    kasan_save_track+0x14/0x30
    kasan_save_free_info+0x3b/0x70
    __kasan_slab_free+0x6b/0x90
    kfree+0x1d4/0x520
    device_release+0x77/0x120
    kobject_put+0xdb/0x2a0
    idxd_unregister_devices+0x5b/0x130 [idxd]
    idxd_remove+0x2f/0x120 [idxd]
    pci_device_remove+0x6d/0xf0
    device_release_driver_internal+0x244/0x2e0
    unbind_store+0xae/0xb0
    kernfs_fop_write_iter+0x205/0x2d0
    vfs_write+0x3a9/0x6b0
    ksys_write+0xc7/0x160
    do_syscall_64+0x9c/0x620
    entry_SYSCALL_64_after_hwframe+0x76/0x7e

Fix by taking refs on wq/engine/group conf_dev in idxd_unregister_devices()
before calling device_unregister(). This prevents their memory from being
freed until after idxd_device_drv_remove() completes. Add idxd_put_devices()
to release the refs after all cleanup activities are done.

To reproduce:
  cd dsa-perf-micros
  ./scripts/setup_dsa.sh -d dsa0 -w 1 -m s -e 4
  echo "0000:00:01.0" > /sys/bus/pci/drivers/idxd/unbind

Fixes: 47c16ac27d4c ("dmaengine: idxd: fix idxd conf_dev 'struct device' lifetime")
Suggested-by: Adrian Huang (Lenovo) <adrianhuang0701@gmail.com>
Signed-off-by: Ziyou Wang (Lenovo) <ziyou.dev@gmail.com>
---
Added a Suggested-by tag for Adrian, as he came up with the idea for this patch.

 drivers/dma/idxd/init.c  | 15 +++++++++++++++
 drivers/dma/idxd/sysfs.c |  8 ++++++++
 2 files changed, 23 insertions(+)

diff --git a/drivers/dma/idxd/init.c b/drivers/dma/idxd/init.c
index 4b827a329756..212efbfebacc 100644
--- a/drivers/dma/idxd/init.c
+++ b/drivers/dma/idxd/init.c
@@ -1263,6 +1263,20 @@ static void idxd_shutdown(struct pci_dev *pdev)
 	flush_workqueue(idxd->wq);
 }
 
+static void idxd_put_devices(struct idxd_device *idxd)
+{
+	int i;
+
+	for (i = 0; i < idxd->max_wqs; i++)
+		put_device(wq_confdev(idxd->wqs[i]));
+
+	for (i = 0; i < idxd->max_engines; i++)
+		put_device(engine_confdev(idxd->engines[i]));
+
+	for (i = 0; i < idxd->max_groups; i++)
+		put_device(group_confdev(idxd->groups[i]));
+}
+
 static void idxd_remove(struct pci_dev *pdev)
 {
 	struct idxd_device *idxd = pci_get_drvdata(pdev);
@@ -1284,6 +1298,7 @@ static void idxd_remove(struct pci_dev *pdev)
 	if (device_pasid_enabled(idxd))
 		idxd_disable_system_pasid(idxd);
 	pci_iounmap(pdev, idxd->reg_base);
+	idxd_put_devices(idxd);
 	put_device(idxd_confdev(idxd));
 	pci_disable_device(pdev);
 }
diff --git a/drivers/dma/idxd/sysfs.c b/drivers/dma/idxd/sysfs.c
index 6d251095c350..eca96a4dcecd 100644
--- a/drivers/dma/idxd/sysfs.c
+++ b/drivers/dma/idxd/sysfs.c
@@ -1988,21 +1988,29 @@ void idxd_unregister_devices(struct idxd_device *idxd)
 {
 	int i;
 
+	/*
+	 * Take a reference count to the idxd device to prevent the
+	 * corresponding idxd context from being freed, as the driver
+	 * still needs it for the remaining cleanup operations.
+	 */
 	for (i = 0; i < idxd->max_wqs; i++) {
 		struct idxd_wq *wq = idxd->wqs[i];
 
+		get_device(wq_confdev(wq));
 		device_unregister(wq_confdev(wq));
 	}
 
 	for (i = 0; i < idxd->max_engines; i++) {
 		struct idxd_engine *engine = idxd->engines[i];
 
+		get_device(engine_confdev(engine));
 		device_unregister(engine_confdev(engine));
 	}
 
 	for (i = 0; i < idxd->max_groups; i++) {
 		struct idxd_group *group = idxd->groups[i];
 
+		get_device(group_confdev(group));
 		device_unregister(group_confdev(group));
 	}
 }
-- 
2.53.0

Re: [PATCH] dmaengine: idxd: Fix use-after-free in idxd_remove

[sashiko-bot]

Thank you for your contribution! Sashiko AI review found 4 potential issue(s) to consider:

Pre-existing issues:
- [High] Pre-existing Use-After-Free and Double Free in the error path of `idxd_register_devices()` and `idxd_pci_probe_alloc()`.
- [High] Pre-existing double free and UAF of the IDXD device and its resources in `idxd_free()`.
- [High] Flawed `while (j--)` loop condition in device registration cleanup paths causes an out-of-bounds array read on the first iteration failure, and skips unregistering elements on subsequent iteration failures.
- [Medium] Pre-existing memory leak of the `saved_engines` array in the error paths of `idxd_device_config_save`.
--

commit cf8581894d15c4c8e08a44d0ae7b7e0a1a4eb54c
Author: Ziyou Wang (Lenovo) <ziyou.dev@gmail.com>

dmaengine: idxd: Fix use-after-free in idxd_remove

This commit fixes a use-after-free bug in the idxd driver unbind path
by taking reference counts on the wq, engine, and group devices before
unregistering them. It adds idxd_put_devices() to release these references
after all cleanup operations are complete.

> diff --git a/drivers/dma/idxd/init.c b/drivers/dma/idxd/init.c
> index 4b827a3297564..212efbfebacce 100644
> --- a/drivers/dma/idxd/init.c
> +++ b/drivers/dma/idxd/init.c

[ ... ]

[Severity: Medium]
This is a pre-existing issue, but does idxd_device_config_save() leak the
saved_engines array on error?

Looking at idxd_device_config_save() in drivers/dma/idxd/init.c:

	struct idxd_engine **saved_engines =
			kcalloc_node(idxd->max_engines,
				     sizeof(struct idxd_engine *),
				     GFP_KERNEL, dev_to_node(dev));

The saved_engines array is allocated without the __free(kfree) cleanup
attribute. If a memory allocation fails later in the function (for example,
when the device undergoes a reset and memory allocation fails), it jumps to
an error path that returns -ENOMEM.

The idxd_free_saved() function frees the individual structures via
idxd_free_saved_configs(), but assumes the arrays themselves are freed by the
__free cleanup mechanism. Since saved_engines lacks this attribute, is its
memory permanently leaked here?


[Severity: High]
This is a pre-existing issue, but can idxd_free() trigger a double-free
and use-after-free?

Looking at idxd_free() in drivers/dma/idxd/init.c:

static void idxd_free(struct idxd_device *idxd)
{
	if (!idxd)
		return;

	put_device(idxd_confdev(idxd));
	bitmap_free(idxd->opcap_bmap);
	ida_free(&idxd_ida, idxd->id);
	kfree(idxd);
}

When idxd_free() is called on early probe error paths, it first calls
put_device(idxd_confdev(idxd)). This drops the device refcount to 0 and
triggers the idxd_conf_device_release() callback, which explicitly frees
internal structures, calls bitmap_free(idxd->opcap_bmap),
ida_free(&idxd_ida, idxd->id), and then kfree(idxd).

After put_device() returns, idxd_free() immediately repeats these exact
same cleanup steps on the now-freed idxd structure. Does this result in a
direct use-after-free and double-free?

> diff --git a/drivers/dma/idxd/sysfs.c b/drivers/dma/idxd/sysfs.c
> index 6d251095c350f..eca96a4dcecd8 100644
> --- a/drivers/dma/idxd/sysfs.c
> +++ b/drivers/dma/idxd/sysfs.c

[ ... ]

[Severity: High]
This is a pre-existing issue, but does idxd_register_devices() have a similar
use-after-free and double-free pattern in its error path?

Looking at idxd_register_devices() in drivers/dma/idxd/sysfs.c:

 err_group:
	for (i = 0; i < idxd->max_engines; i++)
		device_unregister(engine_confdev(idxd->engines[i]));
 err_engine:
	for (i = 0; i < idxd->max_wqs; i++)
		device_unregister(wq_confdev(idxd->wqs[i]));
 err_wq:
	device_del(idxd_confdev(idxd));

If device_add() fails in idxd_register_wq_devices() or subsequent steps,
this cleanup path calls device_unregister() on previously added engines and
WQs, freeing them. It then calls device_del(idxd_confdev(idxd)), which
unbinds the driver. This invokes idxd_device_drv_remove():

drivers/dma/idxd/device.c:idxd_device_drv_remove() {
	...
	for (i = 0; i < idxd->max_wqs; i++) {
		struct idxd_wq *wq = idxd->wqs[i];
		struct device *wq_dev = wq_confdev(wq);

		if (wq->state == IDXD_WQ_DISABLED)
			continue;
	...
}

This loop blindly accesses the WQ states which have already been freed by
the sysfs registration error handler, causing a direct use-after-free.

Furthermore, after idxd_register_devices() returns an error,
idxd_pci_probe_alloc() in drivers/dma/idxd/init.c jumps to err_dev_register:

drivers/dma/idxd/init.c:idxd_pci_probe_alloc() {
	...
 err_dev_register:
	idxd_cleanup(idxd);
	...
}

Which eventually calls idxd_clean_wqs():

drivers/dma/idxd/init.c:idxd_clean_wqs() {
	...
	for (i = 0; i < idxd->max_wqs; i++) {
		conf_dev = wq_confdev(idxd->wqs[i]);
		put_device(conf_dev);
	}
	...
}

This blindly drops references on objects that were already freed in the
idxd_register_devices() rollback, triggering a double free.


[Severity: High]
This is a pre-existing issue, but do the registration error paths have an
out-of-bounds array read?

Looking at idxd_register_wq_devices() in drivers/dma/idxd/sysfs.c:

	while (j--) {
		wq = idxd->wqs[j];
		device_unregister(wq_confdev(wq));
	}

In idxd_register_wq_devices(), idxd_register_engine_devices(), and
idxd_register_group_devices(), the cleanup paths use j = i - 1 followed by
while (j--).

If device_add() fails on the very first element (i = 0), j becomes -1. The
while (j--) condition evaluates the current value -1 as true (non-zero) and
decrements j to -2. The loop body then accesses index -2 (for example,
idxd->wqs[-2]) and passes the out-of-bounds pointer to device_unregister().

Conversely, if it fails on the second element (i = 1), j becomes 0. The
loop evaluates 0 as false and never executes. Does this skip unregistering
and leak the successfully added element at index 0?

-- 
Sashiko AI review · https://sashiko.dev/#/patchset/20260911065547.161522-1-ziyou.dev@gmail.com?part=1