Re: [PATCH v2] usb: xhci: Check endpoint is valid before dereferencing it

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

 



On 22.12.2022 15.18, Mathias Nyman wrote:
On 22.12.2022 13.52, Ladislav Michl wrote:
On Thu, Dec 22, 2022 at 01:08:47PM +0200, Mathias Nyman wrote:
On 22.12.2022 11.01, Ladislav Michl wrote:
On Thu, Dec 22, 2022 at 07:29:12AM +0000, Jimmy Hu wrote:
When the host controller is not responding, all URBs queued to all
endpoints need to be killed. This can cause a kernel panic if we
dereference an invalid endpoint.

Fix this by using xhci_get_virt_ep() helper to find the endpoint and
checking if the endpoint is valid before dereferencing it.

[233311.853271] xhci-hcd xhci-hcd.1.auto: xHCI host controller not responding, assume dead
[233311.853393] Unable to handle kernel NULL pointer dereference at virtual address 00000000000000e8

[233311.853964] pc : xhci_hc_died+0x10c/0x270
[233311.853971] lr : xhci_hc_died+0x1ac/0x270

[233311.854077] Call trace:
[233311.854085]  xhci_hc_died+0x10c/0x270
[233311.854093]  xhci_stop_endpoint_command_watchdog+0x100/0x1a4
[233311.854105]  call_timer_fn+0x50/0x2d4
[233311.854112]  expire_timers+0xac/0x2e4
[233311.854118]  run_timer_softirq+0x300/0xabc
[233311.854127]  __do_softirq+0x148/0x528
[233311.854135]  irq_exit+0x194/0x1a8
[233311.854143]  __handle_domain_irq+0x164/0x1d0
[233311.854149]  gic_handle_irq.22273+0x10c/0x188
[233311.854156]  el1_irq+0xfc/0x1a8
[233311.854175]  lpm_cpuidle_enter+0x25c/0x418 [msm_pm]
[233311.854185]  cpuidle_enter_state+0x1f0/0x764
[233311.854194]  do_idle+0x594/0x6ac
[233311.854201]  cpu_startup_entry+0x7c/0x80
[233311.854209]  secondary_start_kernel+0x170/0x198

Fixes: 50e8725e7c42 ("xhci: Refactor command watchdog and fix split string.")
Cc: stable@xxxxxxxxxxxxxxx
Signed-off-by: Jimmy Hu <hhhuuu@xxxxxxxxxx>
---
   drivers/usb/host/xhci-ring.c | 5 ++++-
   1 file changed, 4 insertions(+), 1 deletion(-)

diff --git a/drivers/usb/host/xhci-ring.c b/drivers/usb/host/xhci-ring.c
index ddc30037f9ce..f5b0e1ce22af 100644
--- a/drivers/usb/host/xhci-ring.c
+++ b/drivers/usb/host/xhci-ring.c
@@ -1169,7 +1169,10 @@ static void xhci_kill_endpoint_urbs(struct xhci_hcd *xhci,
       struct xhci_virt_ep *ep;
       struct xhci_ring *ring;
-    ep = &xhci->devs[slot_id]->eps[ep_index];
+    ep = xhci_get_virt_ep(xhci, slot_id, ep_index);
+    if (!ep)
+        return;
+

xhci_get_virt_ep also adds check for slot_id == 0. It changes behaviour,
do we really want to skip that slot? Original code went from 0 to
MAX_HC_SLOTS-1.

It seems to be off by one to me. Am I missing anything?

slot_id 0 is always invalid, so this is a good change.

I see. Now reading more carefully:
#define HCS_MAX_SLOTS(p)    (((p) >> 0) & 0xff)
#define MAX_HC_SLOTS        256
So the loop should go:
    for (i = 1; i <= HCS_MAX_SLOTS(xhci->hcs_params1); i++)

yes


Also, what about passing ep directly to xhci_kill_endpoint_urbs
and do the check in xhci_hc_died? Not even compile tested:

passing ep to a function named kill_endpoint_urbs() sound like the
right thing to do, but as a generic change.

I think its a good idea to first do a targeted fix for this null pointer
issue that we can send to stable fist.

Agree. But I still do not understand the root cause. There is a check
for NULL xhci->devs[i] already, so patch does not add much more, except
for test for slot_id == 0. And the eps array is just array of
struct xhci_virt_ep, not a pointers to them, so &xhci->devs[i]->eps[j]
should be always valid pointer. However struct xhci_ring in each eps
is allocated and not protected by any lock here. Is that correct?

I think root cause is that freeing xhci->devs[i] and including rings isn't
protected by the lock, this happens in xhci_free_virt_device() called by
xhci_free_dev(), which in turn may be called by usbcore at any time

So xhci->devs[i] might just suddenly disappear

Patch just checks more often if xhci->devs[i] is valid, between every endpoint.
So the race between xhci_free_virt_device() and xhci_kill_endpoint_urbs()
doesn't trigger null pointer deref as easily.

Jimmy Hu,

Any chance you could try if the change below works for you instead of
using xhci_get_virt_ep().
I don't have a easy way to trigger the issue-


diff --git a/drivers/usb/host/xhci.c b/drivers/usb/host/xhci.c
index 79d7931c048a..50b41213e827 100644
--- a/drivers/usb/host/xhci.c
+++ b/drivers/usb/host/xhci.c
@@ -3974,6 +3974,7 @@ static void xhci_free_dev(struct usb_hcd *hcd, struct usb_device *udev)
 	struct xhci_hcd *xhci = hcd_to_xhci(hcd);
 	struct xhci_virt_device *virt_dev;
 	struct xhci_slot_ctx *slot_ctx;
+	unsigned long flags;
 	int i, ret;
/*
@@ -4000,7 +4001,11 @@ static void xhci_free_dev(struct usb_hcd *hcd, struct usb_device *udev)
 		virt_dev->eps[i].ep_state &= ~EP_STOP_CMD_PENDING;
 	virt_dev->udev = NULL;
 	xhci_disable_slot(xhci, udev->slot_id);
+
+	spin_lock_irqsave(&xhci->lock, flags);
 	xhci_free_virt_device(xhci, udev->slot_id);
+	spin_unlock_irqrestore(&xhci->lock, flags);
+
 }
Thanks
Mathias



[Index of Archives]     [Linux Media]     [Linux Input]     [Linux Audio Users]     [Yosemite News]     [Linux Kernel]     [Linux SCSI]     [Old Linux USB Devel Archive]

  Powered by Linux