On Thu, 29 Jan 2015 11:42:18 -0500
Alan Stern <stern@xxxxxxxxxxxxxxxxxxx> wrote:
> On Wed, 28 Jan 2015, Joe Lawrence wrote:
>
> > This one should have gone over to linux-usb.
> >
> > -- Joe
> >
> > On 01/28/2015 05:04 PM, Joe Lawrence wrote:
> > > Hello linux-usb,
> > >
> > > We've hit a USB use-after-free on Stratus HW during device removal tests.
> > > We're running fio disk I/O to a scsi disk hanging off USB when the USB
> > > controller is hotplug removed. This crash is very consistent (usually the
> > > first device pull during testing). Without I/O, it may take days to
> > > encounter.
> > >
> > > general protection fault: 0000 [#1] SMP
> > > ...
> > > CPU: 35 PID: 19626 Comm: kworker/u97:0 Tainted: PF W O-------------- 3.10.0-210.el7.dev02.x86_64 #1
> > > Hardware name: Stratus ftServer 6800/G7LYY, BIOS BIOS Version 8.0:30 11/12/2014
> > > Workqueue: scsi_tmf_872 scmd_eh_abort_handler
> > > task: ffff88031bd91960 ti: ffff880981318000 task.ti: ffff880981318000
> > > RIP: 0010:[<ffffffff812d5e2d>] [<ffffffff812d5e2d>] kobject_put+0xd/0x60
> > > RSP: 0018:ffff88098131bd28 EFLAGS: 00010002
> > > RAX: 0000000000000000 RBX: 6b6b6b6b6b6b6c0b RCX: 00000001002f0009
> > > RDX: 000000000000002f RSI: ffffea0015719800 RDI: 6b6b6b6b6b6b6c0b
> > > RBP: ffff88098131bd30 R08: ffff88055c6622f0 R09: 00000001002f0008
> > > R10: ffff88085f407a80 R11: ffffffff81450503 R12: ffff8804bab6d248
> > > R13: 00000000ffffff98 R14: 0000000000000086 R15: 0000000000000c20
> > > FS: 0000000000000000(0000) GS:ffff88085fce0000(0000) knlGS:0000000000000000
> > > CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033
> > > CR2: 00007f2ebb5d6008 CR3: 0000001038dc5000 CR4: 00000000001407e0
> > > DR0: 0000000000000000 DR1: 0000000000000000 DR2: 0000000000000000
> > > DR3: 0000000000000000 DR6: 00000000ffff0ff0 DR7: 0000000000000400
> > > Stack:
> > > 0000000000000000 ffff88098131bd40 ffffffff813cd327 ffff88098131bd50
> > > ffffffff8140a65a ffff88098131bd78 ffffffff81416795 0000000000000000
> > > ffff880414791968 ffff880414791978 ffff88098131bd88 ffffffff814175f6
> > > Call Trace:
> > > [<ffffffff813cd327>] put_device+0x17/0x20
> > > [<ffffffff8140a65a>] usb_put_dev+0x1a/0x20
> > > [<ffffffff81416795>] usb_hcd_unlink_urb+0x65/0xe0
> > > [<ffffffff814175f6>] usb_unlink_urb+0x26/0x50
> > > [<ffffffff81418e92>] usb_sg_cancel+0x82/0xe0
> > > [<ffffffffa0074e71>] usb_stor_stop_transport+0x31/0x50 [usb_storage]
> > > [<ffffffffa0073b8d>] command_abort+0xcd/0xe0 [usb_storage]
> > > [<ffffffff813f51ef>] scmd_eh_abort_handler+0xbf/0x480
> > > [<ffffffff8108f06b>] process_one_work+0x17b/0x470
> > > [<ffffffff8108fe4b>] worker_thread+0x11b/0x400
> > > [<ffffffff8108fd30>] ? rescuer_thread+0x400/0x400
> > > [<ffffffff8109722f>] kthread+0xcf/0xe0
> > > [<ffffffff81097160>] ? kthread_create_on_node+0x140/0x140
> > > [<ffffffff8161387c>] ret_from_fork+0x7c/0xb0
> > > [<ffffffff81097160>] ? kthread_create_on_node+0x140/0x140
> > >
> > > from another crash, we know that the URB itself has been freed:
>
> > > but the underlying usb_device->device is still present:
>
> > > It looks like usb_hcd_unlink_urb takes a reference out on the underlying
> > > device but not the URB, which in our test case consistently gets freed
> > > just before usb_hcd_unlink_urb tries to return a reference on urb->device.
>
> The analysis looks correct.
>
> > > Maybe the URB is a reference count short when usb_hcd_unlink_urb calls
> > > unlink1? Somewhere usb-storage missed taking out a ref?
>
> No; it's a simple use-after-free error.
>
> > > A tourniquet hack-patch follows (probably inside out, as the URB should stay
> > > intact while usb_hcd_unlink_urb does its thing) and has been running a
> > > little over 30 surprise device removals under I/O without incident.
> > >
> > > Any better suggestions?
>
> Please try this instead.
>
> Alan Stern
>
>
>
> Index: usb-3.19/drivers/usb/core/hcd.c
> ===================================================================
> --- usb-3.19.orig/drivers/usb/core/hcd.c
> +++ usb-3.19/drivers/usb/core/hcd.c
> @@ -1618,6 +1618,7 @@ static int unlink1(struct usb_hcd *hcd,
> int usb_hcd_unlink_urb (struct urb *urb, int status)
> {
> struct usb_hcd *hcd;
> + struct usb_device *udev = urb->dev;
> int retval = -EIDRM;
> unsigned long flags;
>
> @@ -1629,20 +1630,19 @@ int usb_hcd_unlink_urb (struct urb *urb,
> spin_lock_irqsave(&hcd_urb_unlink_lock, flags);
> if (atomic_read(&urb->use_count) > 0) {
> retval = 0;
> - usb_get_dev(urb->dev);
> + usb_get_dev(udev);
> }
> spin_unlock_irqrestore(&hcd_urb_unlink_lock, flags);
> if (retval == 0) {
> hcd = bus_to_hcd(urb->dev->bus);
> retval = unlink1(hcd, urb, status);
> - usb_put_dev(urb->dev);
> + if (retval == 0)
> + retval = -EINPROGRESS;
> + else if (retval != -EIDRM && retval != -EBUSY)
> + dev_dbg(&udev->dev, "hcd_unlink_urb %p fail %d\n",
> + urb, retval);
> + usb_put_dev(udev);
> }
> -
> - if (retval == 0)
> - retval = -EINPROGRESS;
> - else if (retval != -EIDRM && retval != -EBUSY)
> - dev_dbg(&urb->dev->dev, "hcd_unlink_urb %p fail %d\n",
> - urb, retval);
> return retval;
> }
Hi Alan,
Thanks for the confirmation and refined patch. Caching the usb_device
pointer and moving the unlink1 return checking up looks good to me.
Your patch ran overnight tests with no problems.
Tested-off-by: Joe Lawrence <joe.lawrence@xxxxxxxxxxx>
-- Joe
--
To unsubscribe from this list: send the line "unsubscribe linux-scsi" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
