A SURPRISE removal of a hotplug PCIe device, caused by a Link Down
event will execute an orderly removal of the driver, which normally
includes releasing the IRQs with pci_free_irq(_vectors):
* SURPRISE removal event causes Link Down
* pciehp_disable_slot()
* pci_device_remove()
* driver->remove()
* pci_free_irq(_vectors)()
* irq_chip->irq_mask()
* pci_msi_mask_irq()
Eventually, msi_set_mask_bit() will attempt to do MMIO over the dead
link, usually resulting in an Unsupported Request error. This can
confuse the firmware on FFS machines, and lead to a system crash.
Since the channel will have been marked "pci_channel_io_perm_failure"
by the hotplug thread, we know we should avoid sending blind IO to a
dead link.
When the device is disconnected, bail out of MSI teardown.
If device removal and Link Down are independent events, there exists a
race condition when the Link Down event occurs right after the
pci_dev_is_disconnected() check. This is outside the scope of this patch.
Signed-off-by: Alexandru Gagniuc <[email protected]>
---
Changes since v2:
* Updated commit message
drivers/pci/msi.c | 3 +++
1 file changed, 3 insertions(+)
diff --git a/drivers/pci/msi.c b/drivers/pci/msi.c
index 4c0b47867258..6b6541ab264f 100644
--- a/drivers/pci/msi.c
+++ b/drivers/pci/msi.c
@@ -227,6 +227,9 @@ static void msi_set_mask_bit(struct irq_data *data, u32 flag)
{
struct msi_desc *desc = irq_data_get_msi_desc(data);
+ if (pci_dev_is_disconnected(msi_desc_to_pci_dev(desc)))
+ return;
+
if (desc->msi_attrib.is_msix) {
msix_mask_irq(desc, flag);
readl(desc->mask_base); /* Flush write to device */
--
2.19.2
[+cc Jon, Jens, Christoph, Sagi, Linus, linux-nvme from related discussion]
[+cc Greg, Oliver, who responded to v2 of this patch]
On Fri, Feb 22, 2019 at 01:48:06PM -0600, Alexandru Gagniuc wrote:
> A SURPRISE removal of a hotplug PCIe device, caused by a Link Down
> event will execute an orderly removal of the driver, which normally
> includes releasing the IRQs with pci_free_irq(_vectors):
>
> * SURPRISE removal event causes Link Down
> * pciehp_disable_slot()
> * pci_device_remove()
> * driver->remove()
> * pci_free_irq(_vectors)()
> * irq_chip->irq_mask()
> * pci_msi_mask_irq()
>
> Eventually, msi_set_mask_bit() will attempt to do MMIO over the dead
> link, usually resulting in an Unsupported Request error. This can
> confuse the firmware on FFS machines, and lead to a system crash.
>
> Since the channel will have been marked "pci_channel_io_perm_failure"
> by the hotplug thread, we know we should avoid sending blind IO to a
> dead link.
> When the device is disconnected, bail out of MSI teardown.
>
> If device removal and Link Down are independent events, there exists a
> race condition when the Link Down event occurs right after the
> pci_dev_is_disconnected() check. This is outside the scope of this patch.
>
> Signed-off-by: Alexandru Gagniuc <[email protected]>
I had actually applied this to pci/msi with the intent of merging it
for v5.1, but by coincidence I noticed [1], where Jon was basically
solving another piece of the same problem, this time in nvme-pci.
AFAICT, the consensus there was that it would be better to find some
sort of platform solution instead of dealing with it in individual
drivers. The PCI core isn't really a driver, but I think the same
argument applies to it: if we had a better way to recover from readl()
errors, that way would work equally well in nvme-pci and the PCI core.
It sounds like the problem has two parts: the PCI core part and the
individual driver part. Solving only the first (eg, with this patch)
isn't enough by itself, and solving the second via some platform
solution would also solve the first. If that's the case, I don't
think it's worth applying this one, but please correct me if I'm
wrong.
Bjorn
[1] https://lore.kernel.org/lkml/[email protected]/T/#u
> ---
> Changes since v2:
> * Updated commit message
>
> drivers/pci/msi.c | 3 +++
> 1 file changed, 3 insertions(+)
>
> diff --git a/drivers/pci/msi.c b/drivers/pci/msi.c
> index 4c0b47867258..6b6541ab264f 100644
> --- a/drivers/pci/msi.c
> +++ b/drivers/pci/msi.c
> @@ -227,6 +227,9 @@ static void msi_set_mask_bit(struct irq_data *data, u32 flag)
> {
> struct msi_desc *desc = irq_data_get_msi_desc(data);
>
> + if (pci_dev_is_disconnected(msi_desc_to_pci_dev(desc)))
> + return;
> +
> if (desc->msi_attrib.is_msix) {
> msix_mask_irq(desc, flag);
> readl(desc->mask_base); /* Flush write to device */
> --
> 2.19.2
>
On Wed, Mar 20, 2019 at 1:52 PM Bjorn Helgaas <[email protected]> wrote:
>
> AFAICT, the consensus there was that it would be better to find some
> sort of platform solution instead of dealing with it in individual
> drivers. The PCI core isn't really a driver, but I think the same
> argument applies to it: if we had a better way to recover from readl()
> errors, that way would work equally well in nvme-pci and the PCI core.
I think that patches with the pattern "if (disconnected) don't do IO"
are fundamentally broken and we should look for alternatives in all
cases.
They are fundamentally broken because they are racy: if it's an actual
sudden disconnect in the middle of IO, there's no guarantee that we'll
even be notified in time.
They are fundamentally broken because they add new magic special cases
that very few people will ever test, and the people who do test them
tend to do so with old irrelevant kernels.
Finally, they are fundamentally broken because they always end up
being just special cases. One or two special case accesses in a
driver, or perhaps all accesses of a particular type in just _one_
special driver.
Yes, yes, I realize that people want error reporting, and that
hot-removal can cause various error conditions (perhaps just parity
errors for the IO, but also perhaps other random errors caused by
firmware perhaps doing special HW setup).
But the "you get a fatal interrupt, so avoid the IO" kind of model is
completely broken, and needs to just be fixed differently. See above
why it's so completely broken.
So if the hw is set up to send some kinf of synchronous interrupt or
machine check that cannot sanely be handled (perhaps because it will
just repeat forever), we should try to just disable said thing.
PCIe allows for just polling for errors on the bridges, afaik. It's
been years since I looked at it, and maybe I'm wrong. And I bet there
are various "platform-specific value add" registers etc that may need
tweaking outside of any standard spec for PCIe error reporting. But
let's do that in a platform driver, to set up the platform to not do
the silly "I'm just going to die if I see an error" thing.
It's way better to have a model where you poll each bridge once a
minute (or one an hour) and let people know "guys, your hardware
reports errors", than make random crappy changes to random drivers
because the hardware was set up to die on said errors.
And if some MIS person wants the "hardware will die" setting, then
they can damn well have that, and then it's not out problem, but it
also means that we don't start changing random drivers for that insane
setting. It's dead, Jim, and it was the users choice.
Notice how in neither case does it make sense to try to do some "if
(disconnected) dont_do_io()" model for the drivers.
Linus
On 3/20/19 4:44 PM, Linus Torvalds wrote:
> On Wed, Mar 20, 2019 at 1:52 PM Bjorn Helgaas <[email protected]> wrote:
>>
>> AFAICT, the consensus there was that it would be better to find some
>> sort of platform solution instead of dealing with it in individual
>> drivers. The PCI core isn't really a driver, but I think the same
>> argument applies to it: if we had a better way to recover from readl()
>> errors, that way would work equally well in nvme-pci and the PCI core.
>
> I think that patches with the pattern "if (disconnected) don't do IO"
> are fundamentally broken and we should look for alternatives in all
> cases.
>
> They are fundamentally broken because they are racy: if it's an actual
> sudden disconnect in the middle of IO, there's no guarantee that we'll
> even be notified in time.
>
> They are fundamentally broken because they add new magic special cases
> that very few people will ever test, and the people who do test them
> tend to do so with old irrelevant kernels.
>
> Finally, they are fundamentally broken because they always end up
> being just special cases. One or two special case accesses in a
> driver, or perhaps all accesses of a particular type in just _one_
> special driver.
>
> Yes, yes, I realize that people want error reporting, and that
> hot-removal can cause various error conditions (perhaps just parity
> errors for the IO, but also perhaps other random errors caused by
> firmware perhaps doing special HW setup).
>
> But the "you get a fatal interrupt, so avoid the IO" kind of model is
> completely broken, and needs to just be fixed differently. See above
> why it's so completely broken.
>
> So if the hw is set up to send some kinf of synchronous interrupt or
> machine check that cannot sanely be handled (perhaps because it will
> just repeat forever), we should try to just disable said thing.
>
> PCIe allows for just polling for errors on the bridges, afaik. It's
> been years since I looked at it, and maybe I'm wrong. And I bet there
> are various "platform-specific value add" registers etc that may need
> tweaking outside of any standard spec for PCIe error reporting. But
> let's do that in a platform driver, to set up the platform to not do
> the silly "I'm just going to die if I see an error" thing.
>
> It's way better to have a model where you poll each bridge once a
> minute (or one an hour) and let people know "guys, your hardware
> reports errors", than make random crappy changes to random drivers
> because the hardware was set up to die on said errors.
>
> And if some MIS person wants the "hardware will die" setting, then
> they can damn well have that, and then it's not out problem, but it
> also means that we don't start changing random drivers for that insane
> setting. It's dead, Jim, and it was the users choice.
>
> Notice how in neither case does it make sense to try to do some "if
> (disconnected) dont_do_io()" model for the drivers.
I disagree with the idea of doing something you know can cause an error
to propagate. That being said, in this particular case we have come to
rely a little too much on the if (disconnected) model.
You mentioned in the other thread that fixing the GHES driver will pay
much higher dividends. I'm working on reviving a couple of changes to do
just that. Some industry folk were very concerned about the "don't
panic()" approach, and I want to make sure I fairly present their
arguments in the cover letter.
I'm hoping one day we'll have the ability to use page tables to prevent
the situations that we're trying to fix today in less than ideal ways.
Although there are other places in msi.c that use if (disconnected), I'm
okay with dropping this change -- provided we come up with an equivalent
fix.
But even if FFS doesn't crash, do we really want to lose hundreds of
milliseconds to SMM --on all cores-- when all it takes is a couple of
cycles to check a flag?
Alex
On Thu, Mar 21, 2019 at 12:27 PM Alex G <[email protected]> wrote:
>
> On 3/20/19 4:44 PM, Linus Torvalds wrote:
> > On Wed, Mar 20, 2019 at 1:52 PM Bjorn Helgaas <[email protected]> wrote:
> >>
> >> AFAICT, the consensus there was that it would be better to find some
> >> sort of platform solution instead of dealing with it in individual
> >> drivers. The PCI core isn't really a driver, but I think the same
> >> argument applies to it: if we had a better way to recover from readl()
> >> errors, that way would work equally well in nvme-pci and the PCI core.
> >
> > I think that patches with the pattern "if (disconnected) don't do IO"
> > are fundamentally broken and we should look for alternatives in all
> > cases.
> >
> > They are fundamentally broken because they are racy: if it's an actual
> > sudden disconnect in the middle of IO, there's no guarantee that we'll
> > even be notified in time.
> >
> > They are fundamentally broken because they add new magic special cases
> > that very few people will ever test, and the people who do test them
> > tend to do so with old irrelevant kernels.
> >
> > Finally, they are fundamentally broken because they always end up
> > being just special cases. One or two special case accesses in a
> > driver, or perhaps all accesses of a particular type in just _one_
> > special driver.
> >
> > Yes, yes, I realize that people want error reporting, and that
> > hot-removal can cause various error conditions (perhaps just parity
> > errors for the IO, but also perhaps other random errors caused by
> > firmware perhaps doing special HW setup).
> >
> > But the "you get a fatal interrupt, so avoid the IO" kind of model is
> > completely broken, and needs to just be fixed differently. See above
> > why it's so completely broken.
> >
> > So if the hw is set up to send some kinf of synchronous interrupt or
> > machine check that cannot sanely be handled (perhaps because it will
> > just repeat forever), we should try to just disable said thing.
> >
> > PCIe allows for just polling for errors on the bridges, afaik. It's
> > been years since I looked at it, and maybe I'm wrong. And I bet there
> > are various "platform-specific value add" registers etc that may need
> > tweaking outside of any standard spec for PCIe error reporting. But
> > let's do that in a platform driver, to set up the platform to not do
> > the silly "I'm just going to die if I see an error" thing.
> >
> > It's way better to have a model where you poll each bridge once a
> > minute (or one an hour) and let people know "guys, your hardware
> > reports errors", than make random crappy changes to random drivers
> > because the hardware was set up to die on said errors.
> >
> > And if some MIS person wants the "hardware will die" setting, then
> > they can damn well have that, and then it's not out problem, but it
> > also means that we don't start changing random drivers for that insane
> > setting. It's dead, Jim, and it was the users choice.
> >
> > Notice how in neither case does it make sense to try to do some "if
> > (disconnected) dont_do_io()" model for the drivers.
>
> I disagree with the idea of doing something you know can cause an error
> to propagate. That being said, in this particular case we have come to
> rely a little too much on the if (disconnected) model.
My main gripe with the if (disconnected) model is that it's only
really good for inactive devices. If a device is being used then odds
are the driver will do an MMIO before the pci core has had a chance to
mark the device as broken so you crash anyway.
> You mentioned in the other thread that fixing the GHES driver will pay
> much higher dividends. I'm working on reviving a couple of changes to do
> just that. Some industry folk were very concerned about the "don't
> panic()" approach, and I want to make sure I fairly present their
> arguments in the cover letter.
>
> I'm hoping one day we'll have the ability to use page tables to prevent
> the situations that we're trying to fix today in less than ideal ways.
> Although there are other places in msi.c that use if (disconnected), I'm
> okay with dropping this change -- provided we come up with an equivalent
> fix.
What's the idea there? Scan the ioremap space for mappings over the
device BARs and swap them with a normal memory page?
> But even if FFS doesn't crash, do we really want to lose hundreds of
> milliseconds to SMM --on all cores-- when all it takes is a couple of
> cycles to check a flag?
Using pci_dev_is_disconnected() to opportunistically avoid waiting for
MMIO timeouts is fair enough IMO, even if it's a bit ugly. It would
help your case if you did some measurements to show the improvement
and look for other cases it might help. It might also be a good idea
to document when it is appropriate to use pci_is_dev_disconnected() so
we aren't stuck having the same argument again and again, but that's
probably a job for Bjorn though.
Oliver