Hi Sebastian,
On Fri, Jan 15, 2016 at 07:36:25PM +0100, Sebastian Andrzej Siewior wrote:
> I start a binary which should flash the FPGA and re-enumare the PCI-BUS
> and find a new device. It works most of the time. With SLUB debug it
> crashes on each iteration with something like this (compressed output):
>
> | pcieport 0000:00:00.0: AER: Multiple Corrected error received: id=0000
> | Unable to handle kernel paging request for data at address 0x27ef9e3e
> | Faulting instruction address: 0x602f5328
> | Oops: Kernel access of bad area, sig: 11 [#1]
> | Workqueue: events aer_isr
> | GPR24: dd6aa000 6b6b6b6b 605f8378 605f8360 d99b12c0 604fc674 606b1704 d99b12c0
> | NIP [602f5328] pci_walk_bus+0xd4/0x104
>
> Register 25 has the user-after magic. As it turns out, the old PCIe
> device is leaving, generates an error before it left, aer_irq() is fired,
> it schedules a work item. What happens now is that free_irq() is
> invoked, all resources are gone *before* the aes_isr() work item is
> completed.
> So to fix this, I flush the workqueue to ensure that there is no more
> work pending.
> The wait_event() on wait_release should actually synchronized against
> removal. However the condition (->prod_idx == ->cons_idx) is made true
> before the function completes (aer_isr_one_error() is invoked right
> after that) so it does not fulfill its purpose. Therefore I remove it.
>
> Signed-off-by: Sebastian Andrzej Siewior <bigeasy@xxxxxxxxxxxxx>
I propose to merge this patch unchanged, but with the following
changelog. I want to add a bit more detail about the concurrency
problem and remove a bit of the specific detail about your FPGA:
commit 9963c9487f733ef8fe3a06ce3398072a40f955bf
Author: Sebastian Andrzej Siewior <bigeasy@xxxxxxxxxxxxx>
Date: Fri Jan 15 19:36:25 2016 +0100
PCI/AER: Flush workqueue on device remove to avoid use-after-free
A Root Port's AER structure (rpc) contains a queue of events. aer_irq()
enqueues AER status information and schedules aer_isr() to dequeue and
process it. When we remove a device, aer_remove() waits for the queue to
be empty, then frees the rpc struct.
But aer_isr() references the rpc struct after dequeueing and possibly
emptying the queue, which can cause a use-after-free error as in the
following scenario with two threads, aer_isr() on the left and a
concurrent aer_remove() on the right:
Thread A Thread B
-------- --------
aer_irq():
rpc->prod_idx++
aer_remove():
wait_event(rpc->prod_idx == rpc->cons_idx)
# now blocked until queue becomes empty
aer_isr(): # ...
rpc->cons_idx++ # unblocked because queue is now empty
... kfree(rpc)
mutex_unlock(&rpc->rpc_mutex)
Wait until the last scheduled instance of aer_isr() has completed before
freeing the rpc struct by using flush_work() in aer_remove().
I reproduced this use-after-free by flashing a device FPGA and
re-enumerating the bus to find the new device. With SLUB debug, this
crashes with 0x6b bytes (POISON_FREE, the use-after-free magic number) in
GPR25:
pcieport 0000:00:00.0: AER: Multiple Corrected error received: id=0000
Unable to handle kernel paging request for data at address 0x27ef9e3e
Workqueue: events aer_isr
GPR24: dd6aa000 6b6b6b6b 605f8378 605f8360 d99b12c0 604fc674 606b1704 d99b12c0
NIP [602f5328] pci_walk_bus+0xd4/0x104
[bhelgaas: changelog]
Signed-off-by: Sebastian Andrzej Siewior <bigeasy@xxxxxxxxxxxxx>
Signed-off-by: Bjorn Helgaas <bhelgaas@xxxxxxxxxx>
--
To unsubscribe from this list: send the line "unsubscribe linux-pci" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
