There is a race between releasing an irq on one cpu and fetching it
from XIVE on another cpu as there does not seem to be any locking between
these, probably because xive_irq_chip::irq_shutdown() is supposed to
remove the irq from all queues in the system which it does not do.
As a result, when such released irq appears in a queue, we take it
from the queue but we do not change the current priority on that cpu and
since there is no handler for the irq, EOI is never called and the cpu
current priority remains elevated (7 vs. 0xff==unmasked). If another irq
is assigned to the same cpu, then that device stops working until irq
is moved to another cpu or the device is reset.
This checks if irq is still registered, if not, it assumes no valid irq
was fetched from the loop and if there is none left, it continues to
the irq==0 case (not visible in this patch) and sets priority to 0xff
which is basically unmasking. This calls irq_to_desc() on a hot path now
which is a radix tree lookup; hopefully this won't be noticeable as
that tree is quite small.
Signed-off-by: Alexey Kardashevskiy <redacted>
---
Found it on P9 system with:
- a host with 8 cpus online
- a boot disk on ahci with its msix on cpu#0
- a guest with 2xGPUs + 6xNVLink + 4 cpus
- GPU#0 from the guest is bound to the same cpu#0.
Killing a guest killed ahci and therefore the host because of the race.
Note that VFIO masks interrupts first and only then resets the device.
Alternatives:
1. Fix xive_irq_chip::irq_shutdown() to walk through all cpu queues and
drop deregistered irqs.
2. Exploit chip->irq_get_irqchip_state function from
62e0468650c30f0298 "genirq: Add optional hardware synchronization for shutdown".
Both require deep XIVE knowledge which I do not have.
---
arch/powerpc/sysdev/xive/common.c | 8 ++++++--
1 file changed, 6 insertions(+), 2 deletions(-)
@@ -148,8 +148,12 @@ static u32 xive_scan_interrupts(struct xive_cpu *xc, bool just_peek)irq=xive_read_eq(&xc->queue[prio],just_peek);/* Found something ? That's it */-if(irq)-break;+if(irq){+/* Another CPU may have shut this irq down, check it */+if(irq_to_desc(irq))+break;+irq=0;+}/* Clear pending bits */xc->pending_prio&=~(1<<prio);
@@ -148,8 +148,12 @@ static u32 xive_scan_interrupts(struct xive_cpu *xc, bool just_peek)irq=xive_read_eq(&xc->queue[prio],just_peek);/* Found something ? That's it */-if(irq)-break;+if(irq){+/* Another CPU may have shut this irq down, check it */+if(irq_to_desc(irq))
@@ -148,8 +148,12 @@ static u32 xive_scan_interrupts(struct xive_cpu *xc, bool just_peek)irq=xive_read_eq(&xc->queue[prio],just_peek);/* Found something ? That's it */-if(irq)-break;+if(irq){+/* Another CPU may have shut this irq down, check it */+if(irq_to_desc(irq))
If so, then in order to do EOI, I'll need the desc which is gone,
or
I am missing the point?
All you need is drop the local CPU priority.
I know that, the question was how. I cannot use irq_chip in
arch/powerpc/kernel/irq.c and I do not want to add
ppc_md.enable_irqs().
Well, best is probably to do just that though, but call it something
like ppc_md.orphan_irq() or something like that instead. Another option
as you mention is to try to scrub queues, but that's trickier to do due
to the lockless nature of the queue handling.
Cheers,
Ben.
If so, then in order to do EOI, I'll need the desc which is gone,
or
I am missing the point?
All you need is drop the local CPU priority.
I know that, the question was how. I cannot use irq_chip in
arch/powerpc/kernel/irq.c and I do not want to add
ppc_md.enable_irqs().
Well, best is probably to do just that though, but call it something
like ppc_md.orphan_irq() or something like that instead. Another option
as you mention is to try to scrub queues, but that's trickier to do due
to the lockless nature of the queue handling.
When the IRQ is shutdown, couldn't we cleanup the CPU EQ by filtering
all the dangling entries, and replacing them with zeroes ? That would
be alternative 1, but I don't think we need to scan all cpus. The last
target should be enough.
C.
From: Daniel Henrique Barboza <hidden> Date: 2019-07-14 21:59:55
This patch fixed an issue I was experiencing with virsh start/destroy
of guests with mlx5 and GPU passthrough in a Power 9 server. I
believe it's a similar situation which Alexey described in the post
commit msg.
Tested-by: Daniel Henrique Barboza <redacted>
On 7/12/19 5:20 AM, Alexey Kardashevskiy wrote:
quoted hunk
There is a race between releasing an irq on one cpu and fetching it
from XIVE on another cpu as there does not seem to be any locking between
these, probably because xive_irq_chip::irq_shutdown() is supposed to
remove the irq from all queues in the system which it does not do.
As a result, when such released irq appears in a queue, we take it
from the queue but we do not change the current priority on that cpu and
since there is no handler for the irq, EOI is never called and the cpu
current priority remains elevated (7 vs. 0xff==unmasked). If another irq
is assigned to the same cpu, then that device stops working until irq
is moved to another cpu or the device is reset.
This checks if irq is still registered, if not, it assumes no valid irq
was fetched from the loop and if there is none left, it continues to
the irq==0 case (not visible in this patch) and sets priority to 0xff
which is basically unmasking. This calls irq_to_desc() on a hot path now
which is a radix tree lookup; hopefully this won't be noticeable as
that tree is quite small.
Signed-off-by: Alexey Kardashevskiy <redacted>
---
Found it on P9 system with:
- a host with 8 cpus online
- a boot disk on ahci with its msix on cpu#0
- a guest with 2xGPUs + 6xNVLink + 4 cpus
- GPU#0 from the guest is bound to the same cpu#0.
Killing a guest killed ahci and therefore the host because of the race.
Note that VFIO masks interrupts first and only then resets the device.
Alternatives:
1. Fix xive_irq_chip::irq_shutdown() to walk through all cpu queues and
drop deregistered irqs.
2. Exploit chip->irq_get_irqchip_state function from
62e0468650c30f0298 "genirq: Add optional hardware synchronization for shutdown".
Both require deep XIVE knowledge which I do not have.
---
arch/powerpc/sysdev/xive/common.c | 8 ++++++--
1 file changed, 6 insertions(+), 2 deletions(-)
@@ -148,8 +148,12 @@ static u32 xive_scan_interrupts(struct xive_cpu *xc, bool just_peek)irq=xive_read_eq(&xc->queue[prio],just_peek);/* Found something ? That's it */-if(irq)-break;+if(irq){+/* Another CPU may have shut this irq down, check it */+if(irq_to_desc(irq))+break;+irq=0;+}/* Clear pending bits */xc->pending_prio&=~(1<<prio);
From: Benjamin Herrenschmidt <benh@kernel.crashing.org> Date: 2019-07-14 22:50:14
On Sun, 2019-07-14 at 21:44 +0200, Cédric Le Goater wrote:
quoted
Well, best is probably to do just that though, but call it something
like ppc_md.orphan_irq() or something like that instead. Another option
as you mention is to try to scrub queues, but that's trickier to do due
to the lockless nature of the queue handling.
When the IRQ is shutdown, couldn't we cleanup the CPU EQ by filtering
all the dangling entries, and replacing them with zeroes ? That would
be alternative 1, but I don't think we need to scan all cpus. The last
target should be enough.
It's a bit tricky due to the lockless nature of the queues...
Cheers,
Ben.
On 15/07/2019 05:14, Daniel Henrique Barboza wrote:
This patch fixed an issue I was experiencing with virsh start/destroy
of guests with mlx5 and GPU passthrough in a Power 9 server. I
believe it's a similar situation which Alexey described in the post
commit msg.
Tested-by: Daniel Henrique Barboza <redacted>
Thanks for testing! Unfortunately, this is not the right fix as it only reduces the race window but does not eliminate it, we need something better than this, i.e. do not backport this anywhere please :) Thanks,
On 7/12/19 5:20 AM, Alexey Kardashevskiy wrote:
quoted
There is a race between releasing an irq on one cpu and fetching it
from XIVE on another cpu as there does not seem to be any locking between
these, probably because xive_irq_chip::irq_shutdown() is supposed to
remove the irq from all queues in the system which it does not do.
As a result, when such released irq appears in a queue, we take it
from the queue but we do not change the current priority on that cpu and
since there is no handler for the irq, EOI is never called and the cpu
current priority remains elevated (7 vs. 0xff==unmasked). If another irq
is assigned to the same cpu, then that device stops working until irq
is moved to another cpu or the device is reset.
This checks if irq is still registered, if not, it assumes no valid irq
was fetched from the loop and if there is none left, it continues to
the irq==0 case (not visible in this patch) and sets priority to 0xff
which is basically unmasking. This calls irq_to_desc() on a hot path now
which is a radix tree lookup; hopefully this won't be noticeable as
that tree is quite small.
Signed-off-by: Alexey Kardashevskiy<redacted>
---
Found it on P9 system with:
- a host with 8 cpus online
- a boot disk on ahci with its msix on cpu#0
- a guest with 2xGPUs + 6xNVLink + 4 cpus
- GPU#0 from the guest is bound to the same cpu#0.
Killing a guest killed ahci and therefore the host because of the race.
Note that VFIO masks interrupts first and only then resets the device.
Alternatives:
1. Fix xive_irq_chip::irq_shutdown() to walk through all cpu queues and
drop deregistered irqs.
2. Exploit chip->irq_get_irqchip_state function from
62e0468650c30f0298 "genirq: Add optional hardware synchronization for shutdown".
Both require deep XIVE knowledge which I do not have.
---
arch/powerpc/sysdev/xive/common.c | 8 ++++++--
1 file changed, 6 insertions(+), 2 deletions(-)
@@ -148,8 +148,12 @@ static u32 xive_scan_interrupts(struct xive_cpu *xc, bool just_peek)irq=xive_read_eq(&xc->queue[prio],just_peek);/* Found something ? That's it */-if(irq)-break;+if(irq){+/* Another CPU may have shut this irq down, check it */+if(irq_to_desc(irq))+break;+irq=0;+}/* Clear pending bits */xc->pending_prio&=~(1<<prio);