drivers/iommu/intel/iommu.c | 3 ++- 1 file changed, 2 insertions(+), 1 deletion(-)
When entering intel_iommu_shutdown, system interrupts are disabled,
and the reboot process might be scheduled out by down_write(). If the
scheduled process does not yield (e.g., while(1)), the system will hang.
Signed-off-by: Yunhui Cui <cuiyunhui@bytedance.com>
---
drivers/iommu/intel/iommu.c | 3 ++-
1 file changed, 2 insertions(+), 1 deletion(-)
diff --git a/drivers/iommu/intel/iommu.c b/drivers/iommu/intel/iommu.c
index cc46098f875b..76a1d83b46bf 100644
--- a/drivers/iommu/intel/iommu.c
+++ b/drivers/iommu/intel/iommu.c
@@ -2871,7 +2871,8 @@ void intel_iommu_shutdown(void)
if (no_iommu || dmar_disabled)
return;
- down_write(&dmar_global_lock);
+ if (!down_write_trylock(&dmar_global_lock))
+ return;
/* Disable PMRs explicitly here. */
for_each_iommu(iommu, drhd)
--
2.39.2On 2/20/25 18:15, Yunhui Cui wrote:
> When entering intel_iommu_shutdown, system interrupts are disabled,
> and the reboot process might be scheduled out by down_write(). If the
> scheduled process does not yield (e.g., while(1)), the system will hang.
>
> Signed-off-by: Yunhui Cui<cuiyunhui@bytedance.com>
> ---
> drivers/iommu/intel/iommu.c | 3 ++-
> 1 file changed, 2 insertions(+), 1 deletion(-)
>
> diff --git a/drivers/iommu/intel/iommu.c b/drivers/iommu/intel/iommu.c
> index cc46098f875b..76a1d83b46bf 100644
> --- a/drivers/iommu/intel/iommu.c
> +++ b/drivers/iommu/intel/iommu.c
> @@ -2871,7 +2871,8 @@ void intel_iommu_shutdown(void)
> if (no_iommu || dmar_disabled)
> return;
>
> - down_write(&dmar_global_lock);
> + if (!down_write_trylock(&dmar_global_lock))
> + return;
If system interrupts are disabled here, locking is unnecessary. Hotplug
operations depend on interrupt events, so it's better to remove the
lock. The shutdown helper then appears like this:
void intel_iommu_shutdown(void)
{
struct dmar_drhd_unit *drhd;
struct intel_iommu *iommu = NULL;
if (no_iommu || dmar_disabled)
return;
/*
* System interrupts are disabled when it reaches here. Locking
* is unnecessary when iterating the IOMMU list.
*/
list_for_each_entry(drhd, &dmar_drhd_units, list) {
if (drhd->ignored)
continue;
iommu = drhd->iommu;
/* Disable PMRs explicitly here. */
iommu_disable_protect_mem_regions(iommu);
iommu_disable_translation(iommu);
}
}
Does it work for you?
Thanks,
baolu
在 2025/2/24 9:02, Baolu Lu 写道:
> On 2/20/25 18:15, Yunhui Cui wrote:
>> When entering intel_iommu_shutdown, system interrupts are disabled,
>> and the reboot process might be scheduled out by down_write(). If the
>> scheduled process does not yield (e.g., while(1)), the system will hang.
>>
>> Signed-off-by: Yunhui Cui<cuiyunhui@bytedance.com>
>> ---
>> drivers/iommu/intel/iommu.c | 3 ++-
>> 1 file changed, 2 insertions(+), 1 deletion(-)
>>
>> diff --git a/drivers/iommu/intel/iommu.c b/drivers/iommu/intel/iommu.c
>> index cc46098f875b..76a1d83b46bf 100644
>> --- a/drivers/iommu/intel/iommu.c
>> +++ b/drivers/iommu/intel/iommu.c
>> @@ -2871,7 +2871,8 @@ void intel_iommu_shutdown(void)
>> if (no_iommu || dmar_disabled)
>> return;
>> - down_write(&dmar_global_lock);
Only BSP is running at this point, no DMAR concurrency access protection is needed
anyore, even there is interrupt (only legacy & NMI) coming in, it is impossible to for
scheduler to run any other iommu access code.**
Thanks,
Ethan**
>> + if (!down_write_trylock(&dmar_global_lock))
>> + return;
>
> If system interrupts are disabled here, locking is unnecessary. Hotplug
> operations depend on interrupt events, so it's better to remove the
> lock. The shutdown helper then appears like this:
>
> void intel_iommu_shutdown(void)
> {
> struct dmar_drhd_unit *drhd;
> struct intel_iommu *iommu = NULL;
>
> if (no_iommu || dmar_disabled)
> return;
>
> /*
> * System interrupts are disabled when it reaches here. Locking
> * is unnecessary when iterating the IOMMU list.
> */
> list_for_each_entry(drhd, &dmar_drhd_units, list) {
> if (drhd->ignored)
> continue;
>
> iommu = drhd->iommu;
> /* Disable PMRs explicitly here. */
> iommu_disable_protect_mem_regions(iommu);
> iommu_disable_translation(iommu);
> }
> }
>
> Does it work for you?
>
> Thanks,
> baolu
>
--
"firm, enduring, strong, and long-lived"
Hi Baolu,
On Mon, Feb 24, 2025 at 9:06 AM Baolu Lu <baolu.lu@linux.intel.com> wrote:
>
> On 2/20/25 18:15, Yunhui Cui wrote:
> > When entering intel_iommu_shutdown, system interrupts are disabled,
> > and the reboot process might be scheduled out by down_write(). If the
> > scheduled process does not yield (e.g., while(1)), the system will hang.
> >
> > Signed-off-by: Yunhui Cui<cuiyunhui@bytedance.com>
> > ---
> > drivers/iommu/intel/iommu.c | 3 ++-
> > 1 file changed, 2 insertions(+), 1 deletion(-)
> >
> > diff --git a/drivers/iommu/intel/iommu.c b/drivers/iommu/intel/iommu.c
> > index cc46098f875b..76a1d83b46bf 100644
> > --- a/drivers/iommu/intel/iommu.c
> > +++ b/drivers/iommu/intel/iommu.c
> > @@ -2871,7 +2871,8 @@ void intel_iommu_shutdown(void)
> > if (no_iommu || dmar_disabled)
> > return;
> >
> > - down_write(&dmar_global_lock);
> > + if (!down_write_trylock(&dmar_global_lock))
> > + return;
>
> If system interrupts are disabled here, locking is unnecessary. Hotplug
> operations depend on interrupt events, so it's better to remove the
> lock. The shutdown helper then appears like this:
Currently, intel_iommu_shutdown() is only called by
native_machine_shutdown(). The down_write/up operations can be
removed. Even if there's a hardware access error, IOMMU_WAIT_OP() will
trigger a panic().
>
> void intel_iommu_shutdown(void)
> {
> struct dmar_drhd_unit *drhd;
> struct intel_iommu *iommu = NULL;
>
> if (no_iommu || dmar_disabled)
> return;
>
> /*
> * System interrupts are disabled when it reaches here. Locking
> * is unnecessary when iterating the IOMMU list.
> */
> list_for_each_entry(drhd, &dmar_drhd_units, list) {
> if (drhd->ignored)
> continue;
>
> iommu = drhd->iommu;
> /* Disable PMRs explicitly here. */
> iommu_disable_protect_mem_regions(iommu);
> iommu_disable_translation(iommu);
> }
> }
>
> Does it work for you?
Yes.
>
> Thanks,
> baolu
Thanks,
Yunhui
在 2025/2/20 18:15, Yunhui Cui 写道: > When entering intel_iommu_shutdown, system interrupts are disabled, System interrupts were disabled ? you mean all interrupts were disabled when entering intel_iommu_shutdown(), perhaps it is not true, at least for upstream latest code. > and the reboot process might be scheduled out by down_write(). If the > scheduled process does not yield (e.g., while(1)), the system will hang. No NMI lockup watchdog jumping out here ? Thanks, Ethan > > Signed-off-by: Yunhui Cui <cuiyunhui@bytedance.com> > --- > drivers/iommu/intel/iommu.c | 3 ++- > 1 file changed, 2 insertions(+), 1 deletion(-) > > diff --git a/drivers/iommu/intel/iommu.c b/drivers/iommu/intel/iommu.c > index cc46098f875b..76a1d83b46bf 100644 > --- a/drivers/iommu/intel/iommu.c > +++ b/drivers/iommu/intel/iommu.c > @@ -2871,7 +2871,8 @@ void intel_iommu_shutdown(void) > if (no_iommu || dmar_disabled) > return; > > - down_write(&dmar_global_lock); > + if (!down_write_trylock(&dmar_global_lock)) > + return; > > /* Disable PMRs explicitly here. */ > for_each_iommu(iommu, drhd) -- "firm, enduring, strong, and long-lived"
Hi Ethan,
On Fri, Feb 21, 2025 at 4:40 PM Ethan Zhao <haifeng.zhao@linux.intel.com> wrote:
>
>
> 在 2025/2/20 18:15, Yunhui Cui 写道:
> > When entering intel_iommu_shutdown, system interrupts are disabled,
>
> System interrupts were disabled ? you mean all interrupts were disabled
> when entering intel_iommu_shutdown(), perhaps it is not true, at least
> for upstream latest code.
>
> > and the reboot process might be scheduled out by down_write(). If the
> > scheduled process does not yield (e.g., while(1)), the system will hang.
>
> No NMI lockup watchdog jumping out here ?
Steps to reproduce:
1. Avoid return in:
if (no_iommu || dmar_disabled)
return;
2. Write a.out with while(1).
3. ./a.out &; reboot -f.
4. Observe. Send NMI via BIOS to check system response.
5. Add console=ttyS0,115200 to cmdline to increase reproduction chance.
Let's continue discussing based on the above.
>
> Thanks,
> Ethan
>
> >
> > Signed-off-by: Yunhui Cui <cuiyunhui@bytedance.com>
> > ---
> > drivers/iommu/intel/iommu.c | 3 ++-
> > 1 file changed, 2 insertions(+), 1 deletion(-)
> >
> > diff --git a/drivers/iommu/intel/iommu.c b/drivers/iommu/intel/iommu.c
> > index cc46098f875b..76a1d83b46bf 100644
> > --- a/drivers/iommu/intel/iommu.c
> > +++ b/drivers/iommu/intel/iommu.c
> > @@ -2871,7 +2871,8 @@ void intel_iommu_shutdown(void)
> > if (no_iommu || dmar_disabled)
> > return;
> >
> > - down_write(&dmar_global_lock);
> > + if (!down_write_trylock(&dmar_global_lock))
> > + return;
> >
> > /* Disable PMRs explicitly here. */
> > for_each_iommu(iommu, drhd)
>
> --
> "firm, enduring, strong, and long-lived"
>
Thanks,
Yunhui
在 2025/2/21 17:46, yunhui cui 写道: > Hi Ethan, > > On Fri, Feb 21, 2025 at 4:40 PM Ethan Zhao <haifeng.zhao@linux.intel.com> wrote: >> >> 在 2025/2/20 18:15, Yunhui Cui 写道: >>> When entering intel_iommu_shutdown, system interrupts are disabled, >> System interrupts were disabled ? you mean all interrupts were disabled >> when entering intel_iommu_shutdown(), perhaps it is not true, at least >> for upstream latest code. >> >>> and the reboot process might be scheduled out by down_write(). If the >>> scheduled process does not yield (e.g., while(1)), the system will hang. >> No NMI lockup watchdog jumping out here ? > Steps to reproduce: > > 1. Avoid return in: > if (no_iommu || dmar_disabled) > return; > > 2. Write a.out with while(1). > > 3. ./a.out &; reboot -f. > > 4. Observe. Send NMI via BIOS to check system response. Via BMC ? There is 'NMI' hardware physical button on some machines to trigger NMI to OS for diagnostic purpose, you could check your box for that. but no luck, there is no NMI trigger in my GNR BMC. Thanks, Ethan > > 5. Add console=ttyS0,115200 to cmdline to increase reproduction chance. > > Let's continue discussing based on the above. > >> Thanks, >> Ethan >> >>> Signed-off-by: Yunhui Cui <cuiyunhui@bytedance.com> >>> --- >>> drivers/iommu/intel/iommu.c | 3 ++- >>> 1 file changed, 2 insertions(+), 1 deletion(-) >>> >>> diff --git a/drivers/iommu/intel/iommu.c b/drivers/iommu/intel/iommu.c >>> index cc46098f875b..76a1d83b46bf 100644 >>> --- a/drivers/iommu/intel/iommu.c >>> +++ b/drivers/iommu/intel/iommu.c >>> @@ -2871,7 +2871,8 @@ void intel_iommu_shutdown(void) >>> if (no_iommu || dmar_disabled) >>> return; >>> >>> - down_write(&dmar_global_lock); >>> + if (!down_write_trylock(&dmar_global_lock)) >>> + return; >>> >>> /* Disable PMRs explicitly here. */ >>> for_each_iommu(iommu, drhd) >> -- >> "firm, enduring, strong, and long-lived" >> > Thanks, > Yunhui -- "firm, enduring, strong, and long-lived"
在 2025/2/21 17:46, yunhui cui 写道: > Hi Ethan, > > On Fri, Feb 21, 2025 at 4:40 PM Ethan Zhao <haifeng.zhao@linux.intel.com> wrote: >> >> 在 2025/2/20 18:15, Yunhui Cui 写道: >>> When entering intel_iommu_shutdown, system interrupts are disabled, >> System interrupts were disabled ? you mean all interrupts were disabled >> when entering intel_iommu_shutdown(), perhaps it is not true, at least >> for upstream latest code. >> >>> and the reboot process might be scheduled out by down_write(). If the >>> scheduled process does not yield (e.g., while(1)), the system will hang. >> No NMI lockup watchdog jumping out here ? > Steps to reproduce: > > 1. Avoid return in: > if (no_iommu || dmar_disabled) > return; > > 2. Write a.out with while(1). > > 3. ./a.out &; reboot -f. > > 4. Observe. Send NMI via BIOS to check system response. > > 5. Add console=ttyS0,115200 to cmdline to increase reproduction chance. > > Let's continue discussing based on the above. I will try these steps to reproduce. Per the lastest upstream code, the local processor's interrupt mask is cleaned. so the processor could accept interrupts and handle them. and lagacy interrupt should be restored for later boot if there is lagacy device and as to NMI, no one could stop it. In a short, perhaps it is fact under your hardware configureation that no interrupt event come in to kick the scheduler to run when the a.out (while(1)) got scheduled in, but not because all system interrupts are disabled. Thanks, Ethan >> Thanks, >> Ethan >> >>> Signed-off-by: Yunhui Cui <cuiyunhui@bytedance.com> >>> --- >>> drivers/iommu/intel/iommu.c | 3 ++- >>> 1 file changed, 2 insertions(+), 1 deletion(-) >>> >>> diff --git a/drivers/iommu/intel/iommu.c b/drivers/iommu/intel/iommu.c >>> index cc46098f875b..76a1d83b46bf 100644 >>> --- a/drivers/iommu/intel/iommu.c >>> +++ b/drivers/iommu/intel/iommu.c >>> @@ -2871,7 +2871,8 @@ void intel_iommu_shutdown(void) >>> if (no_iommu || dmar_disabled) >>> return; >>> >>> - down_write(&dmar_global_lock); >>> + if (!down_write_trylock(&dmar_global_lock)) >>> + return; >>> >>> /* Disable PMRs explicitly here. */ >>> for_each_iommu(iommu, drhd) >> -- >> "firm, enduring, strong, and long-lived" >> > Thanks, > Yunhui > -- "firm, enduring, strong, and long-lived"
© 2016 - 2025 Red Hat, Inc.