On 01/21/2016 01:12 AM, Bjorn Helgaas wrote:
On Wed, Jan 20, 2016 at 12:21:24PM +0800, Chen Fan wrote:
On 01/20/2016 08:24 AM, Bjorn Helgaas wrote:
[+cc Jiang]
Hi Chen,
On Tue, Jan 19, 2016 at 02:43:30PM +0100, Rafael J. Wysocki wrote:
On Tuesday, January 19, 2016 09:45:13 AM Chen Fan wrote:
In our environment, when enable Secure boot, we found an abnormal
This has more information than necessary. I don't think Secure Boot is
really relevant, and nor are the timestamps and stack addresses below.
I just think enable the Secure Boot, probably the firmware assigned
a 0xff interrupt to the device which unauthenticated.
The important thing is that you're changing the way we handle
Interrupt Line being 0xff. That affects more than just Secure Boot
users. It's fine to mention Secure Boot later, as one example of an
affected scenario.
I don't know anything about Secure Boot, but setting Interrupt Line to
0xff would obviously not be a robust way of hiding an unauthenticated
device. But it sounds like you're just speculating about that anyway.
phenomenon as following call trace shows. after investigation, we
found the firmware assigned an irq number 255 which means unknown
or no connection in PCI local spec for i801_smbus, meanwhile the
ACPI didn't configure the pci irq routing. and the 255 irq number
was assigned for megasa msix without IRQF_SHARED. then in this case
during i801_smbus probe, the i801_smbus driver would request irq with
bad irq number 255. but the 255 irq number was assigned for memgasa
with MSIX enable. which will cause request_irq fails, and call trace
shows, actually, we should expose the error early, rather than in request
irq, here we simply fix the problem by return err when find the irq is
255.
See the call trace:
[ 32.459195] ipmi device interface
[ 32.612907] shpchp: Standard Hot Plug PCI Controller Driver version: 0.4
[ 32.800459] ixgbe: Intel(R) 10 Gigabit PCI Express Network Driver - version 4.0.1-k-rh
[ 32.818319] ixgbe: Copyright (c) 1999-2014 Intel Corporation.
[ 32.844009] lpc_ich 0001:80:1f.0: I/O space for ACPI uninitialized
[ 32.850093] i801_smbus 0000:00:1f.3: enabling device (0140 -> 0143)
[ 32.851134] i801_smbus 0000:00:1f.3: can't derive routing for PCI INT C
[ 32.851136] i801_smbus 0000:00:1f.3: PCI INT C: no GSI
[ 32.851164] genirq: Flags mismatch irq 255. 00000080 (i801_smbus) vs. 00000000 (megasa
[ 32.851168] CPU: 0 PID: 2487 Comm: kworker/0:1 Not tainted 3.10.0-229.el7.x86_64 #1
[ 32.851170] Hardware name: FUJITSU PRIMEQUEST 2800E2/D3736, BIOS PRIMEQUEST 2000 Serie5
[ 32.851178] Workqueue: events work_for_cpu_fn
[ 32.851208] ffff88086c330b00 00000000e233a9df ffff88086d57bca0 ffffffff81603f36
[ 32.851227] ffff88086d57bcf8 ffffffff8110d23a ffff88686fe02000 0000000000000246
[ 32.851246] ffff88086a9a8c00 00000000e233a9df ffffffffa00ad220 0000000000000080
[ 32.851247] Call Trace:
[ 32.851261] [<ffffffff81603f36>] dump_stack+0x19/0x1b
[ 32.851271] [<ffffffff8110d23a>] __setup_irq+0x54a/0x570
[ 32.851282] [<ffffffffa00ad220>] ? i801_check_pre.isra.5+0xe0/0xe0 [i2c_i801]
[ 32.851289] [<ffffffff8110d3bc>] request_threaded_irq+0xcc/0x170
[ 32.851298] [<ffffffffa00ae87f>] i801_probe+0x32f/0x508 [i2c_i801]
[ 32.851308] [<ffffffff81308385>] local_pci_probe+0x45/0xa0
[ 32.851315] [<ffffffff8108bfd4>] work_for_cpu_fn+0x14/0x20
[ 32.851323] [<ffffffff8108f0ab>] process_one_work+0x17b/0x470
[ 32.851330] [<ffffffff81090003>] worker_thread+0x293/0x400
[ 32.851338] [<ffffffff8108fd70>] ? rescuer_thread+0x400/0x400
[ 32.851346] [<ffffffff8109726f>] kthread+0xcf/0xe0
[ 32.851353] [<ffffffff810971a0>] ? kthread_create_on_node+0x140/0x140
[ 32.851362] [<ffffffff81613cfc>] ret_from_fork+0x7c/0xb0
[ 32.851369] [<ffffffff810971a0>] ? kthread_create_on_node+0x140/0x140
[ 32.851373] i801_smbus 0000:00:1f.3: Failed to allocate irq 255: -16
[ 32.851435] i801_smbus: probe of 0000:00:1f.3 failed with error -16
Since the Interrupt Line register is writable and might contain any
value, it would be nice if Linux could at least tolerate anything
firmware might leave there without a backtrace, even if we end up not
being able to use the device.
Your patch changes the acpi_pci_irq_enable() return value from 0 to
-EINVAL for this case. You're running v3.10, and this change probably
makes pci_enable_device() fail. I suppose the user-visible effect is
that with your patch,
- there's no backtrace,
- i801_smbus fails with "Failed to enable SMBus PCI device" instead
of with "Failed to allocate irq 255", and
- i801_smbus fails even if no other device is using IRQ 255, instead
of "succeeding" and using an IRQ 255 that probably doesn't work
(this seems like maybe the most important difference)
Jiang has changed this path with 890e4847587f ("PCI: Add
pcibios_alloc_irq() and pcibios_free_irq()"), so I think on newer
kernels, we'll never even call the i801_smbus probe function.
no, on newer kernels, this phenomenon also probably appearance,
with this patch 890e4847587f change, it didn't change the
acpi_pci_irq_enable() return value, with the problem it also return 0,
and then still call __pci_device_probe() to do i801_smbus probe
function in pci_device_probe() function.
I meant that *with your patch*, newer kernels won't call the
i801_smbus probe function.
What behavior are you looking for from i801_smbus? Decline to claim
the device? Try to use the device without interrupts? Try to figure
out an interrupt in some other way?
I think if BIOS assigned 0xff interrupt line to device, and kernel
can't look
up a valid interrupt for the device, we should not allow to use the device.
I'm not 100% sure that 890e4847587f does the right thing by preventing
a driver from claiming a device where we can't set up an IRQ. It's
conceivable that a driver could still operate a device even without an
IRQ.
I don't understanding, does without IRQ for device still work?
Polling drivers do not need IRQs. The PCI core has no idea whether a
driver is interrupt-driven or polling, so we can't assume that a
device with no IRQ is useless.
Got it, I observed the smbus driver has changed to polling when request_irq
failed on new kernel.
can we use a broken_irq flag in pci_dev to mark the device irq if invalid ?
then if a device broken_irq set, we don't need to call request_irq and
directly return failure. of course, maybe we need to check this for all
devices.
BTW, can we skip the 0xff irq number when allocating irq in x86 arch ?
Thanks,
Chen
Bjorn
.
--
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
