From: Tvrtko Ursulin tvrtko.ursulin@intel.com
Catch and log any garbage in the register, including no tiles marked, or multiple tiles marked.
Signed-off-by: Tvrtko Ursulin tvrtko.ursulin@intel.com Cc: Matt Roper matthew.d.roper@intel.com --- We caught garbage in DG1_MSTR_TILE_INTR with DG2 (actual value 0xF9D2C008) during glmark and more badness. So I thought lets log all possible failure modes from here and also use per device logging. --- drivers/gpu/drm/i915/i915_irq.c | 33 ++++++++++++++++++++++----------- drivers/gpu/drm/i915/i915_reg.h | 1 + 2 files changed, 23 insertions(+), 11 deletions(-)
diff --git a/drivers/gpu/drm/i915/i915_irq.c b/drivers/gpu/drm/i915/i915_irq.c index 73cebc6aa650..79853d3fc1ed 100644 --- a/drivers/gpu/drm/i915/i915_irq.c +++ b/drivers/gpu/drm/i915/i915_irq.c @@ -2778,24 +2778,30 @@ static irqreturn_t dg1_irq_handler(int irq, void *arg) u32 gu_misc_iir;
if (!intel_irqs_enabled(i915)) - return IRQ_NONE; + goto none;
master_tile_ctl = dg1_master_intr_disable(regs); - if (!master_tile_ctl) { - dg1_master_intr_enable(regs); - return IRQ_NONE; + if (!master_tile_ctl) + goto enable_none; + + if (master_tile_ctl & ~(DG1_MSTR_IRQ | DG1_MSTR_TILE_MASK)) { + drm_warn(&i915->drm, "Garbage in master_tile_ctl: 0x%08x!\n", + master_tile_ctl); + goto enable_none; }
/* FIXME: we only support tile 0 for now. */ - if (master_tile_ctl & DG1_MSTR_TILE(0)) { - master_ctl = raw_reg_read(regs, GEN11_GFX_MSTR_IRQ); - raw_reg_write(regs, GEN11_GFX_MSTR_IRQ, master_ctl); - } else { - DRM_ERROR("Tile not supported: 0x%08x\n", master_tile_ctl); - dg1_master_intr_enable(regs); - return IRQ_NONE; + if (REG_FIELD_GET(DG1_MSTR_TILE_MASK, master_tile_ctl) != + DG1_MSTR_TILE(0)) { + drm_warn(&i915->drm, "Unexpected irq from tile %u!\n", + ilog2(REG_FIELD_GET(DG1_MSTR_TILE_MASK, + master_tile_ctl))); + goto enable_none; }
+ master_ctl = raw_reg_read(regs, GEN11_GFX_MSTR_IRQ); + raw_reg_write(regs, GEN11_GFX_MSTR_IRQ, master_ctl); + gen11_gt_irq_handler(gt, master_ctl);
if (master_ctl & GEN11_DISPLAY_IRQ) @@ -2810,6 +2816,11 @@ static irqreturn_t dg1_irq_handler(int irq, void *arg) pmu_irq_stats(i915, IRQ_HANDLED);
return IRQ_HANDLED; + +enable_none: + dg1_master_intr_enable(regs); +none: + return IRQ_NONE; }
/* Called from drm generic code, passed 'crtc' which diff --git a/drivers/gpu/drm/i915/i915_reg.h b/drivers/gpu/drm/i915/i915_reg.h index d8579ab9384c..eefa301c6430 100644 --- a/drivers/gpu/drm/i915/i915_reg.h +++ b/drivers/gpu/drm/i915/i915_reg.h @@ -5774,6 +5774,7 @@
#define DG1_MSTR_TILE_INTR _MMIO(0x190008) #define DG1_MSTR_IRQ REG_BIT(31) +#define DG1_MSTR_TILE_MASK REG_GENMASK(3, 0) #define DG1_MSTR_TILE(t) REG_BIT(t)
#define GEN11_DISPLAY_INT_CTL _MMIO(0x44200)
On Tue, May 24, 2022 at 10:43:39AM +0100, Tvrtko Ursulin wrote:
From: Tvrtko Ursulin tvrtko.ursulin@intel.com
Catch and log any garbage in the register, including no tiles marked, or multiple tiles marked.
Signed-off-by: Tvrtko Ursulin tvrtko.ursulin@intel.com Cc: Matt Roper matthew.d.roper@intel.com
We caught garbage in DG1_MSTR_TILE_INTR with DG2 (actual value 0xF9D2C008) during glmark and more badness. So I thought lets log all possible failure modes from here and also use per device logging.
drivers/gpu/drm/i915/i915_irq.c | 33 ++++++++++++++++++++++----------- drivers/gpu/drm/i915/i915_reg.h | 1 + 2 files changed, 23 insertions(+), 11 deletions(-)
diff --git a/drivers/gpu/drm/i915/i915_irq.c b/drivers/gpu/drm/i915/i915_irq.c index 73cebc6aa650..79853d3fc1ed 100644 --- a/drivers/gpu/drm/i915/i915_irq.c +++ b/drivers/gpu/drm/i915/i915_irq.c @@ -2778,24 +2778,30 @@ static irqreturn_t dg1_irq_handler(int irq, void *arg) u32 gu_misc_iir;
if (!intel_irqs_enabled(i915))
return IRQ_NONE;
goto none;master_tile_ctl = dg1_master_intr_disable(regs);
- if (!master_tile_ctl) {
dg1_master_intr_enable(regs);return IRQ_NONE;
- if (!master_tile_ctl)
goto enable_none;- if (master_tile_ctl & ~(DG1_MSTR_IRQ | DG1_MSTR_TILE_MASK)) {
drm_warn(&i915->drm, "Garbage in master_tile_ctl: 0x%08x!\n",master_tile_ctl);
I know we have a bunch of them already, but shouldn't we be avoiding printk-based stuff like this inside interrupt handlers? Should we be migrating all these error messages over to trace_printk or something similar that's safer to use?
Matt
goto enable_none;}
/* FIXME: we only support tile 0 for now. */
- if (master_tile_ctl & DG1_MSTR_TILE(0)) {
master_ctl = raw_reg_read(regs, GEN11_GFX_MSTR_IRQ);raw_reg_write(regs, GEN11_GFX_MSTR_IRQ, master_ctl);- } else {
DRM_ERROR("Tile not supported: 0x%08x\n", master_tile_ctl);dg1_master_intr_enable(regs);return IRQ_NONE;
if (REG_FIELD_GET(DG1_MSTR_TILE_MASK, master_tile_ctl) !=
DG1_MSTR_TILE(0)) {drm_warn(&i915->drm, "Unexpected irq from tile %u!\n",ilog2(REG_FIELD_GET(DG1_MSTR_TILE_MASK,master_tile_ctl)));goto enable_none;}
master_ctl = raw_reg_read(regs, GEN11_GFX_MSTR_IRQ);
raw_reg_write(regs, GEN11_GFX_MSTR_IRQ, master_ctl);
gen11_gt_irq_handler(gt, master_ctl);
if (master_ctl & GEN11_DISPLAY_IRQ)
@@ -2810,6 +2816,11 @@ static irqreturn_t dg1_irq_handler(int irq, void *arg) pmu_irq_stats(i915, IRQ_HANDLED);
return IRQ_HANDLED;
+enable_none:
- dg1_master_intr_enable(regs);
+none:
- return IRQ_NONE;
}
/* Called from drm generic code, passed 'crtc' which diff --git a/drivers/gpu/drm/i915/i915_reg.h b/drivers/gpu/drm/i915/i915_reg.h index d8579ab9384c..eefa301c6430 100644 --- a/drivers/gpu/drm/i915/i915_reg.h +++ b/drivers/gpu/drm/i915/i915_reg.h @@ -5774,6 +5774,7 @@
#define DG1_MSTR_TILE_INTR _MMIO(0x190008) #define DG1_MSTR_IRQ REG_BIT(31) +#define DG1_MSTR_TILE_MASK REG_GENMASK(3, 0) #define DG1_MSTR_TILE(t) REG_BIT(t)
#define GEN11_DISPLAY_INT_CTL _MMIO(0x44200)
2.32.0
On 24/05/2022 18:51, Matt Roper wrote:
On Tue, May 24, 2022 at 10:43:39AM +0100, Tvrtko Ursulin wrote:
From: Tvrtko Ursulin tvrtko.ursulin@intel.com
Catch and log any garbage in the register, including no tiles marked, or multiple tiles marked.
Signed-off-by: Tvrtko Ursulin tvrtko.ursulin@intel.com Cc: Matt Roper matthew.d.roper@intel.com
We caught garbage in DG1_MSTR_TILE_INTR with DG2 (actual value 0xF9D2C008) during glmark and more badness. So I thought lets log all possible failure modes from here and also use per device logging.
drivers/gpu/drm/i915/i915_irq.c | 33 ++++++++++++++++++++++----------- drivers/gpu/drm/i915/i915_reg.h | 1 + 2 files changed, 23 insertions(+), 11 deletions(-)
diff --git a/drivers/gpu/drm/i915/i915_irq.c b/drivers/gpu/drm/i915/i915_irq.c index 73cebc6aa650..79853d3fc1ed 100644 --- a/drivers/gpu/drm/i915/i915_irq.c +++ b/drivers/gpu/drm/i915/i915_irq.c @@ -2778,24 +2778,30 @@ static irqreturn_t dg1_irq_handler(int irq, void *arg) u32 gu_misc_iir;
if (!intel_irqs_enabled(i915))
return IRQ_NONE;
goto none;master_tile_ctl = dg1_master_intr_disable(regs);
- if (!master_tile_ctl) {
dg1_master_intr_enable(regs);return IRQ_NONE;
- if (!master_tile_ctl)
goto enable_none;- if (master_tile_ctl & ~(DG1_MSTR_IRQ | DG1_MSTR_TILE_MASK)) {
drm_warn(&i915->drm, "Garbage in master_tile_ctl: 0x%08x!\n",master_tile_ctl);I know we have a bunch of them already, but shouldn't we be avoiding printk-based stuff like this inside interrupt handlers? Should we be migrating all these error messages over to trace_printk or something similar that's safer to use?
Not sure - I kind of think some really unexpected and worrying situations should be loud and on by default. Risk is then spam if not ratelimited. Maybe we should instead ratelimit most errors/warnings coming for irq handlers?
In this particular case at least DRM_ERROR with no device info is the odd one out in the entire file so I'd suggest changing at least that, if the rest of my changes is of questionable benefit.
Regards,
Tvrtko
Matt
goto enable_none;}
/* FIXME: we only support tile 0 for now. */
- if (master_tile_ctl & DG1_MSTR_TILE(0)) {
master_ctl = raw_reg_read(regs, GEN11_GFX_MSTR_IRQ);raw_reg_write(regs, GEN11_GFX_MSTR_IRQ, master_ctl);- } else {
DRM_ERROR("Tile not supported: 0x%08x\n", master_tile_ctl);dg1_master_intr_enable(regs);return IRQ_NONE;
if (REG_FIELD_GET(DG1_MSTR_TILE_MASK, master_tile_ctl) !=
DG1_MSTR_TILE(0)) {drm_warn(&i915->drm, "Unexpected irq from tile %u!\n",ilog2(REG_FIELD_GET(DG1_MSTR_TILE_MASK,master_tile_ctl)));goto enable_none;}
master_ctl = raw_reg_read(regs, GEN11_GFX_MSTR_IRQ);
raw_reg_write(regs, GEN11_GFX_MSTR_IRQ, master_ctl);
gen11_gt_irq_handler(gt, master_ctl);
if (master_ctl & GEN11_DISPLAY_IRQ)
@@ -2810,6 +2816,11 @@ static irqreturn_t dg1_irq_handler(int irq, void *arg) pmu_irq_stats(i915, IRQ_HANDLED);
return IRQ_HANDLED;
+enable_none:
- dg1_master_intr_enable(regs);
+none:
return IRQ_NONE; }
/* Called from drm generic code, passed 'crtc' which
diff --git a/drivers/gpu/drm/i915/i915_reg.h b/drivers/gpu/drm/i915/i915_reg.h index d8579ab9384c..eefa301c6430 100644 --- a/drivers/gpu/drm/i915/i915_reg.h +++ b/drivers/gpu/drm/i915/i915_reg.h @@ -5774,6 +5774,7 @@
#define DG1_MSTR_TILE_INTR _MMIO(0x190008) #define DG1_MSTR_IRQ REG_BIT(31) +#define DG1_MSTR_TILE_MASK REG_GENMASK(3, 0) #define DG1_MSTR_TILE(t) REG_BIT(t)
#define GEN11_DISPLAY_INT_CTL _MMIO(0x44200)
2.32.0
On Wed, May 25, 2022 at 05:03:13PM +0100, Tvrtko Ursulin wrote:
On 24/05/2022 18:51, Matt Roper wrote:
On Tue, May 24, 2022 at 10:43:39AM +0100, Tvrtko Ursulin wrote:
From: Tvrtko Ursulin tvrtko.ursulin@intel.com
Catch and log any garbage in the register, including no tiles marked, or multiple tiles marked.
Signed-off-by: Tvrtko Ursulin tvrtko.ursulin@intel.com Cc: Matt Roper matthew.d.roper@intel.com
We caught garbage in DG1_MSTR_TILE_INTR with DG2 (actual value 0xF9D2C008) during glmark and more badness. So I thought lets log all possible failure modes from here and also use per device logging.
drivers/gpu/drm/i915/i915_irq.c | 33 ++++++++++++++++++++++----------- drivers/gpu/drm/i915/i915_reg.h | 1 + 2 files changed, 23 insertions(+), 11 deletions(-)
diff --git a/drivers/gpu/drm/i915/i915_irq.c b/drivers/gpu/drm/i915/i915_irq.c index 73cebc6aa650..79853d3fc1ed 100644 --- a/drivers/gpu/drm/i915/i915_irq.c +++ b/drivers/gpu/drm/i915/i915_irq.c @@ -2778,24 +2778,30 @@ static irqreturn_t dg1_irq_handler(int irq, void *arg) u32 gu_misc_iir; if (!intel_irqs_enabled(i915))
return IRQ_NONE;
goto none;
- if (!master_tile_ctl) {
dg1_master_intr_enable(regs);return IRQ_NONE;
- if (!master_tile_ctl)
goto enable_none;- if (master_tile_ctl & ~(DG1_MSTR_IRQ | DG1_MSTR_TILE_MASK)) {
drm_warn(&i915->drm, "Garbage in master_tile_ctl: 0x%08x!\n",master_tile_ctl);I know we have a bunch of them already, but shouldn't we be avoiding printk-based stuff like this inside interrupt handlers? Should we be migrating all these error messages over to trace_printk or something similar that's safer to use?
Not sure - I kind of think some really unexpected and worrying situations should be loud and on by default. Risk is then spam if not ratelimited. Maybe we should instead ratelimit most errors/warnings coming for irq handlers?
It's not the risk of spam that's the problem, but rather that printk-based stuff eventually calls into the console code to flush its buffers. That's way more overhead than you want in an interrupt handler so it's bad on its own, but if you're using something slow like a serial console, it becomes even more of a problem.
While the unexpected bits in the master tile register are strange and may point to a bigger problem somewhere else, they're also harmless on their own since we should just ignore those bits and only process the valid tiles.
In this particular case at least DRM_ERROR with no device info is the odd one out in the entire file so I'd suggest changing at least that, if the rest of my changes is of questionable benefit.
Changing DRM_ERROR -> drm_err would probably be fine in the short term since it doesn't really make us any worse off. Changing to drm_warn might not be great since we're generating a lot more lines of output and probably multiplying the already bad overhead that shouldn't be happening in an interrupt handler. But if we could update the interrupt handler to just save away the details and do the actual drm_warn later, outside the interrupt handler code, that would be okay. We should probably work toward something like that for all of our interrupt handler warning/error messages.
Matt
Regards,
Tvrtko
Matt
goto enable_none;
- if (master_tile_ctl & DG1_MSTR_TILE(0)) {
master_ctl = raw_reg_read(regs, GEN11_GFX_MSTR_IRQ);raw_reg_write(regs, GEN11_GFX_MSTR_IRQ, master_ctl);- } else {
DRM_ERROR("Tile not supported: 0x%08x\n", master_tile_ctl);dg1_master_intr_enable(regs);return IRQ_NONE;
- if (REG_FIELD_GET(DG1_MSTR_TILE_MASK, master_tile_ctl) !=
DG1_MSTR_TILE(0)) {drm_warn(&i915->drm, "Unexpected irq from tile %u!\n",ilog2(REG_FIELD_GET(DG1_MSTR_TILE_MASK,master_tile_ctl)));goto enable_none;- master_ctl = raw_reg_read(regs, GEN11_GFX_MSTR_IRQ);
- raw_reg_write(regs, GEN11_GFX_MSTR_IRQ, master_ctl);
- gen11_gt_irq_handler(gt, master_ctl); if (master_ctl & GEN11_DISPLAY_IRQ)
@@ -2810,6 +2816,11 @@ static irqreturn_t dg1_irq_handler(int irq, void *arg) pmu_irq_stats(i915, IRQ_HANDLED); return IRQ_HANDLED;
+enable_none:
- dg1_master_intr_enable(regs);
+none:
- return IRQ_NONE; } /* Called from drm generic code, passed 'crtc' which
diff --git a/drivers/gpu/drm/i915/i915_reg.h b/drivers/gpu/drm/i915/i915_reg.h index d8579ab9384c..eefa301c6430 100644 --- a/drivers/gpu/drm/i915/i915_reg.h +++ b/drivers/gpu/drm/i915/i915_reg.h @@ -5774,6 +5774,7 @@ #define DG1_MSTR_TILE_INTR _MMIO(0x190008) #define DG1_MSTR_IRQ REG_BIT(31) +#define DG1_MSTR_TILE_MASK REG_GENMASK(3, 0) #define DG1_MSTR_TILE(t) REG_BIT(t)
#define GEN11_DISPLAY_INT_CTL _MMIO(0x44200)
2.32.0
On 25/05/2022 19:05, Matt Roper wrote:
On Wed, May 25, 2022 at 05:03:13PM +0100, Tvrtko Ursulin wrote:
On 24/05/2022 18:51, Matt Roper wrote:
On Tue, May 24, 2022 at 10:43:39AM +0100, Tvrtko Ursulin wrote:
From: Tvrtko Ursulin tvrtko.ursulin@intel.com
Catch and log any garbage in the register, including no tiles marked, or multiple tiles marked.
Signed-off-by: Tvrtko Ursulin tvrtko.ursulin@intel.com Cc: Matt Roper matthew.d.roper@intel.com
We caught garbage in DG1_MSTR_TILE_INTR with DG2 (actual value 0xF9D2C008) during glmark and more badness. So I thought lets log all possible failure modes from here and also use per device logging.
drivers/gpu/drm/i915/i915_irq.c | 33 ++++++++++++++++++++++----------- drivers/gpu/drm/i915/i915_reg.h | 1 + 2 files changed, 23 insertions(+), 11 deletions(-)
diff --git a/drivers/gpu/drm/i915/i915_irq.c b/drivers/gpu/drm/i915/i915_irq.c index 73cebc6aa650..79853d3fc1ed 100644 --- a/drivers/gpu/drm/i915/i915_irq.c +++ b/drivers/gpu/drm/i915/i915_irq.c @@ -2778,24 +2778,30 @@ static irqreturn_t dg1_irq_handler(int irq, void *arg) u32 gu_misc_iir; if (!intel_irqs_enabled(i915))
return IRQ_NONE;
goto none;
- if (!master_tile_ctl) {
dg1_master_intr_enable(regs);return IRQ_NONE;
- if (!master_tile_ctl)
goto enable_none;- if (master_tile_ctl & ~(DG1_MSTR_IRQ | DG1_MSTR_TILE_MASK)) {
drm_warn(&i915->drm, "Garbage in master_tile_ctl: 0x%08x!\n",master_tile_ctl);I know we have a bunch of them already, but shouldn't we be avoiding printk-based stuff like this inside interrupt handlers? Should we be migrating all these error messages over to trace_printk or something similar that's safer to use?
Not sure - I kind of think some really unexpected and worrying situations should be loud and on by default. Risk is then spam if not ratelimited. Maybe we should instead ratelimit most errors/warnings coming for irq handlers?
It's not the risk of spam that's the problem, but rather that printk-based stuff eventually calls into the console code to flush its buffers. That's way more overhead than you want in an interrupt handler so it's bad on its own, but if you're using something slow like a serial console, it becomes even more of a problem.
Is it a problem for messages which we never expect to see?
While the unexpected bits in the master tile register are strange and may point to a bigger problem somewhere else, they're also harmless on their own since we should just ignore those bits and only process the valid tiles.
Yes, I was expecting that a patch belonging to multi-tile enablement would be incoming soon, which would be changing:
+ if (REG_FIELD_GET(DG1_MSTR_TILE_MASK, master_tile_ctl) != + DG1_MSTR_TILE(0)) { + drm_warn(&i915->drm, "Unexpected irq from tile %u!\n", + ilog2(REG_FIELD_GET(DG1_MSTR_TILE_MASK, + master_tile_ctl))); + goto enable_none; }
From this patch, into something completely different like walking bit by bit, handling the present tiles, and warning on unexpected ones. What should remain though is warning on no tiles signaled (which what we saw, together with garbage in reserved bits).
In this particular case at least DRM_ERROR with no device info is the odd one out in the entire file so I'd suggest changing at least that, if the rest of my changes is of questionable benefit.
Changing DRM_ERROR -> drm_err would probably be fine in the short term since it doesn't really make us any worse off. Changing to drm_warn might not be great since we're generating a lot more lines of output and
Sorry I don't follow - why does replacing drm_err with drm_warn generate (a lot) more lines of output?
But it can be drm_err for all I care, I don't think we really have consistent story between errors and warnings in this area.
probably multiplying the already bad overhead that shouldn't be happening in an interrupt handler. But if we could update the interrupt handler to just save away the details and do the actual drm_warn later, outside the interrupt handler code, that would be okay. We should probably work toward something like that for all of our interrupt handler warning/error messages.
Not sure I agree - for messages which we don't expect to see it doesn't really matter that there will be overhead when they are hit. Presumably bad things are already happening there so spending effort to optimise those path is questionable.
Regards,
Tvrtko
On Thu, May 26, 2022 at 11:18:17AM +0100, Tvrtko Ursulin wrote:
On 25/05/2022 19:05, Matt Roper wrote:
On Wed, May 25, 2022 at 05:03:13PM +0100, Tvrtko Ursulin wrote:
On 24/05/2022 18:51, Matt Roper wrote:
On Tue, May 24, 2022 at 10:43:39AM +0100, Tvrtko Ursulin wrote:
From: Tvrtko Ursulin tvrtko.ursulin@intel.com
Catch and log any garbage in the register, including no tiles marked, or multiple tiles marked.
Signed-off-by: Tvrtko Ursulin tvrtko.ursulin@intel.com Cc: Matt Roper matthew.d.roper@intel.com
We caught garbage in DG1_MSTR_TILE_INTR with DG2 (actual value 0xF9D2C008) during glmark and more badness. So I thought lets log all possible failure modes from here and also use per device logging.
drivers/gpu/drm/i915/i915_irq.c | 33 ++++++++++++++++++++++----------- drivers/gpu/drm/i915/i915_reg.h | 1 + 2 files changed, 23 insertions(+), 11 deletions(-)
diff --git a/drivers/gpu/drm/i915/i915_irq.c b/drivers/gpu/drm/i915/i915_irq.c index 73cebc6aa650..79853d3fc1ed 100644 --- a/drivers/gpu/drm/i915/i915_irq.c +++ b/drivers/gpu/drm/i915/i915_irq.c @@ -2778,24 +2778,30 @@ static irqreturn_t dg1_irq_handler(int irq, void *arg) u32 gu_misc_iir; if (!intel_irqs_enabled(i915))
return IRQ_NONE;
goto none;
- if (!master_tile_ctl) {
dg1_master_intr_enable(regs);return IRQ_NONE;
- if (!master_tile_ctl)
goto enable_none;- if (master_tile_ctl & ~(DG1_MSTR_IRQ | DG1_MSTR_TILE_MASK)) {
drm_warn(&i915->drm, "Garbage in master_tile_ctl: 0x%08x!\n",master_tile_ctl);I know we have a bunch of them already, but shouldn't we be avoiding printk-based stuff like this inside interrupt handlers? Should we be migrating all these error messages over to trace_printk or something similar that's safer to use?
Not sure - I kind of think some really unexpected and worrying situations should be loud and on by default. Risk is then spam if not ratelimited. Maybe we should instead ratelimit most errors/warnings coming for irq handlers?
It's not the risk of spam that's the problem, but rather that printk-based stuff eventually calls into the console code to flush its buffers. That's way more overhead than you want in an interrupt handler so it's bad on its own, but if you're using something slow like a serial console, it becomes even more of a problem.
Is it a problem for messages which we never expect to see?
Kind of. While not as catastrophic, it's the same argument for why we don't use BUG() anymore...when the impossible does manage to happen there's unnecessary collateral damage on things outside of graphics. If we're adding huge delays inside an interrupt handler (while other interrupts are disabled) that impacts the system-wide usability, not just our own driver.
I'd also argue that these messages actually are semi-expected. Random bits being set shouldn't happen, but in the world of dgpu's, we do occasionally see cases where the PCI link itself goes down for reasons outside our control and then all registers read back as 0xFFFFFFFF, which will probably trigger error messages here (as well as a bunch of other places).
While the unexpected bits in the master tile register are strange and may point to a bigger problem somewhere else, they're also harmless on their own since we should just ignore those bits and only process the valid tiles.
Yes, I was expecting that a patch belonging to multi-tile enablement would be incoming soon, which would be changing:
- if (REG_FIELD_GET(DG1_MSTR_TILE_MASK, master_tile_ctl) !=
DG1_MSTR_TILE(0)) {drm_warn(&i915->drm, "Unexpected irq from tile %u!\n",ilog2(REG_FIELD_GET(DG1_MSTR_TILE_MASK,master_tile_ctl)));goto enable_none;From this patch, into something completely different like walking bit by bit, handling the present tiles, and warning on unexpected ones. What should remain though is warning on no tiles signaled (which what we saw, together with garbage in reserved bits).
Yeah. Although I still feel the interrupt handler should really just be flagging the errors so that the actual prints themselves can happen outside the interrupt.
In this particular case at least DRM_ERROR with no device info is the odd one out in the entire file so I'd suggest changing at least that, if the rest of my changes is of questionable benefit.
Changing DRM_ERROR -> drm_err would probably be fine in the short term since it doesn't really make us any worse off. Changing to drm_warn might not be great since we're generating a lot more lines of output and
Sorry I don't follow - why does replacing drm_err with drm_warn generate (a lot) more lines of output?
Sorry, my mistake; I had it in my mind that we were talking about a drm_WARN_ON rather than just drm_warn (i.e., including a big stacktrace and such). DRM_ERROR -> drm_warn alone shouldn't have any extra negative impact.
But it can be drm_err for all I care, I don't think we really have consistent story between errors and warnings in this area.
probably multiplying the already bad overhead that shouldn't be happening in an interrupt handler. But if we could update the interrupt handler to just save away the details and do the actual drm_warn later, outside the interrupt handler code, that would be okay. We should probably work toward something like that for all of our interrupt handler warning/error messages.
Not sure I agree - for messages which we don't expect to see it doesn't really matter that there will be overhead when they are hit. Presumably bad things are already happening there so spending effort to optimise those path is questionable.
Something bad is happening to graphics is we hit one of these cases. But if we start doing prints while interrupts are disabled, we start having more of a negative impact on the rest of the system too.
Matt
Regards,
Tvrtko
On 27/05/2022 19:42, Matt Roper wrote:
On Thu, May 26, 2022 at 11:18:17AM +0100, Tvrtko Ursulin wrote:
On 25/05/2022 19:05, Matt Roper wrote:
On Wed, May 25, 2022 at 05:03:13PM +0100, Tvrtko Ursulin wrote:
On 24/05/2022 18:51, Matt Roper wrote:
On Tue, May 24, 2022 at 10:43:39AM +0100, Tvrtko Ursulin wrote:
From: Tvrtko Ursulin tvrtko.ursulin@intel.com
Catch and log any garbage in the register, including no tiles marked, or multiple tiles marked.
Signed-off-by: Tvrtko Ursulin tvrtko.ursulin@intel.com Cc: Matt Roper matthew.d.roper@intel.com
We caught garbage in DG1_MSTR_TILE_INTR with DG2 (actual value 0xF9D2C008) during glmark and more badness. So I thought lets log all possible failure modes from here and also use per device logging.
drivers/gpu/drm/i915/i915_irq.c | 33 ++++++++++++++++++++++----------- drivers/gpu/drm/i915/i915_reg.h | 1 + 2 files changed, 23 insertions(+), 11 deletions(-)diff --git a/drivers/gpu/drm/i915/i915_irq.c b/drivers/gpu/drm/i915/i915_irq.c index 73cebc6aa650..79853d3fc1ed 100644 --- a/drivers/gpu/drm/i915/i915_irq.c +++ b/drivers/gpu/drm/i915/i915_irq.c @@ -2778,24 +2778,30 @@ static irqreturn_t dg1_irq_handler(int irq, void *arg) u32 gu_misc_iir; if (!intel_irqs_enabled(i915))
return IRQ_NONE;
goto none; master_tile_ctl = dg1_master_intr_disable(regs);
- if (!master_tile_ctl) {
dg1_master_intr_enable(regs);return IRQ_NONE;
- if (!master_tile_ctl)
goto enable_none;- if (master_tile_ctl & ~(DG1_MSTR_IRQ | DG1_MSTR_TILE_MASK)) {
drm_warn(&i915->drm, "Garbage in master_tile_ctl: 0x%08x!\n",master_tile_ctl);I know we have a bunch of them already, but shouldn't we be avoiding printk-based stuff like this inside interrupt handlers? Should we be migrating all these error messages over to trace_printk or something similar that's safer to use?
Not sure - I kind of think some really unexpected and worrying situations should be loud and on by default. Risk is then spam if not ratelimited. Maybe we should instead ratelimit most errors/warnings coming for irq handlers?
It's not the risk of spam that's the problem, but rather that printk-based stuff eventually calls into the console code to flush its buffers. That's way more overhead than you want in an interrupt handler so it's bad on its own, but if you're using something slow like a serial console, it becomes even more of a problem.
Is it a problem for messages which we never expect to see?
Kind of. While not as catastrophic, it's the same argument for why we don't use BUG() anymore...when the impossible does manage to happen there's unnecessary collateral damage on things outside of graphics. If we're adding huge delays inside an interrupt handler (while other interrupts are disabled) that impacts the system-wide usability, not just our own driver.
I'd also argue that these messages actually are semi-expected. Random bits being set shouldn't happen, but in the world of dgpu's, we do occasionally see cases where the PCI link itself goes down for reasons outside our control and then all registers read back as 0xFFFFFFFF, which will probably trigger error messages here (as well as a bunch of other places).
Could you expand a bit on what is semi-expected and when? I mean the circumstances of PCI link going down. We certainly don't have any code to survive that.
While the unexpected bits in the master tile register are strange and may point to a bigger problem somewhere else, they're also harmless on their own since we should just ignore those bits and only process the valid tiles.
Yes, I was expecting that a patch belonging to multi-tile enablement would be incoming soon, which would be changing:
- if (REG_FIELD_GET(DG1_MSTR_TILE_MASK, master_tile_ctl) !=
DG1_MSTR_TILE(0)) {drm_warn(&i915->drm, "Unexpected irq from tile %u!\n",ilog2(REG_FIELD_GET(DG1_MSTR_TILE_MASK,master_tile_ctl)));goto enable_none;From this patch, into something completely different like walking bit by bit, handling the present tiles, and warning on unexpected ones. What should remain though is warning on no tiles signaled (which what we saw, together with garbage in reserved bits).
Yeah. Although I still feel the interrupt handler should really just be flagging the errors so that the actual prints themselves can happen outside the interrupt.
In this particular case at least DRM_ERROR with no device info is the odd one out in the entire file so I'd suggest changing at least that, if the rest of my changes is of questionable benefit.
Changing DRM_ERROR -> drm_err would probably be fine in the short term since it doesn't really make us any worse off. Changing to drm_warn might not be great since we're generating a lot more lines of output and
Sorry I don't follow - why does replacing drm_err with drm_warn generate (a lot) more lines of output?
Sorry, my mistake; I had it in my mind that we were talking about a drm_WARN_ON rather than just drm_warn (i.e., including a big stacktrace and such). DRM_ERROR -> drm_warn alone shouldn't have any extra negative impact.
But it can be drm_err for all I care, I don't think we really have consistent story between errors and warnings in this area.
probably multiplying the already bad overhead that shouldn't be happening in an interrupt handler. But if we could update the interrupt handler to just save away the details and do the actual drm_warn later, outside the interrupt handler code, that would be okay. We should probably work toward something like that for all of our interrupt handler warning/error messages.
Not sure I agree - for messages which we don't expect to see it doesn't really matter that there will be overhead when they are hit. Presumably bad things are already happening there so spending effort to optimise those path is questionable.
Something bad is happening to graphics is we hit one of these cases. But if we start doing prints while interrupts are disabled, we start having more of a negative impact on the rest of the system too.
Truly for the case of this particular patch I don't think we should care. Rate limiting should be all that is needed in the short term to strike a balance between effort and benefit. But lets first clarify the PCI link going down problem.
Regards,
Tvrtko
On Mon, Jun 06, 2022 at 12:55:20PM +0100, Tvrtko Ursulin wrote:
On 27/05/2022 19:42, Matt Roper wrote:
On Thu, May 26, 2022 at 11:18:17AM +0100, Tvrtko Ursulin wrote:
On 25/05/2022 19:05, Matt Roper wrote:
On Wed, May 25, 2022 at 05:03:13PM +0100, Tvrtko Ursulin wrote:
On 24/05/2022 18:51, Matt Roper wrote:
On Tue, May 24, 2022 at 10:43:39AM +0100, Tvrtko Ursulin wrote: > From: Tvrtko Ursulin tvrtko.ursulin@intel.com > > Catch and log any garbage in the register, including no tiles marked, or > multiple tiles marked. > > Signed-off-by: Tvrtko Ursulin tvrtko.ursulin@intel.com > Cc: Matt Roper matthew.d.roper@intel.com > --- > We caught garbage in DG1_MSTR_TILE_INTR with DG2 (actual value 0xF9D2C008) > during glmark and more badness. So I thought lets log all possible failure > modes from here and also use per device logging. > --- > drivers/gpu/drm/i915/i915_irq.c | 33 ++++++++++++++++++++++----------- > drivers/gpu/drm/i915/i915_reg.h | 1 + > 2 files changed, 23 insertions(+), 11 deletions(-) > > diff --git a/drivers/gpu/drm/i915/i915_irq.c b/drivers/gpu/drm/i915/i915_irq.c > index 73cebc6aa650..79853d3fc1ed 100644 > --- a/drivers/gpu/drm/i915/i915_irq.c > +++ b/drivers/gpu/drm/i915/i915_irq.c > @@ -2778,24 +2778,30 @@ static irqreturn_t dg1_irq_handler(int irq, void *arg) > u32 gu_misc_iir; > if (!intel_irqs_enabled(i915)) > - return IRQ_NONE; > + goto none; > master_tile_ctl = dg1_master_intr_disable(regs); > - if (!master_tile_ctl) { > - dg1_master_intr_enable(regs); > - return IRQ_NONE; > + if (!master_tile_ctl) > + goto enable_none; > + > + if (master_tile_ctl & ~(DG1_MSTR_IRQ | DG1_MSTR_TILE_MASK)) { > + drm_warn(&i915->drm, "Garbage in master_tile_ctl: 0x%08x!\n", > + master_tile_ctl);
I know we have a bunch of them already, but shouldn't we be avoiding printk-based stuff like this inside interrupt handlers? Should we be migrating all these error messages over to trace_printk or something similar that's safer to use?
Not sure - I kind of think some really unexpected and worrying situations should be loud and on by default. Risk is then spam if not ratelimited. Maybe we should instead ratelimit most errors/warnings coming for irq handlers?
It's not the risk of spam that's the problem, but rather that printk-based stuff eventually calls into the console code to flush its buffers. That's way more overhead than you want in an interrupt handler so it's bad on its own, but if you're using something slow like a serial console, it becomes even more of a problem.
Is it a problem for messages which we never expect to see?
Kind of. While not as catastrophic, it's the same argument for why we don't use BUG() anymore...when the impossible does manage to happen there's unnecessary collateral damage on things outside of graphics. If we're adding huge delays inside an interrupt handler (while other interrupts are disabled) that impacts the system-wide usability, not just our own driver.
I'd also argue that these messages actually are semi-expected. Random bits being set shouldn't happen, but in the world of dgpu's, we do occasionally see cases where the PCI link itself goes down for reasons outside our control and then all registers read back as 0xFFFFFFFF, which will probably trigger error messages here (as well as a bunch of other places).
Could you expand a bit on what is semi-expected and when? I mean the circumstances of PCI link going down. We certainly don't have any code to survive that.
Yeah, I'm referring to the "Lost access to MMIO BAR" errors; in the past most of them have ultimately been tracked down to bugs in early firmware, so flashing an updated IFWI/BIOS onto the device usually solved the problems. Generally those buggy firmwares are an internal problem that never make it into the wild, but I think we have also seen cases where they get triggered by physical/electrical problems on a specific part; that can potentially happen to anyone who's unlucky enough to get a defective/damaged unit.
Basically "hardware returns all F's" happens because the CPU initiates an MMIO transaction with the hardware, the hardware fails to produce any response (possibly due to failing hardware, possibly due to firmware/BIOS bugs), so 0xFFFFFFFF gets returned as an autocompletion to prevent the CPU core from hanging.
It looks like we still have a few open here: https://gitlab.freedesktop.org/search?search=%22Lost+access+to+MMIO+BAR%22&a...
and there are some features on specific platforms we haven't turned on yet because they also trigger these failures (which is still under debug).
We don't/can't really do much to handle these problems in i915 today except printing the 'lost access' error so that we know to ignore whatever kinds of bogus errors we get after that point (usually lots of messages about forcewake failing to clear, engine/GuC reset failing to complete, etc.). But aside from i915 being broken, the rest of the platform should generally continue to work, so you can still access the machine over the network, save logs to disk, etc.
Matt
While the unexpected bits in the master tile register are strange and may point to a bigger problem somewhere else, they're also harmless on their own since we should just ignore those bits and only process the valid tiles.
Yes, I was expecting that a patch belonging to multi-tile enablement would be incoming soon, which would be changing:
- if (REG_FIELD_GET(DG1_MSTR_TILE_MASK, master_tile_ctl) !=
DG1_MSTR_TILE(0)) {drm_warn(&i915->drm, "Unexpected irq from tile %u!\n",ilog2(REG_FIELD_GET(DG1_MSTR_TILE_MASK,master_tile_ctl)));goto enable_none;From this patch, into something completely different like walking bit by bit, handling the present tiles, and warning on unexpected ones. What should remain though is warning on no tiles signaled (which what we saw, together with garbage in reserved bits).
Yeah. Although I still feel the interrupt handler should really just be flagging the errors so that the actual prints themselves can happen outside the interrupt.
In this particular case at least DRM_ERROR with no device info is the odd one out in the entire file so I'd suggest changing at least that, if the rest of my changes is of questionable benefit.
Changing DRM_ERROR -> drm_err would probably be fine in the short term since it doesn't really make us any worse off. Changing to drm_warn might not be great since we're generating a lot more lines of output and
Sorry I don't follow - why does replacing drm_err with drm_warn generate (a lot) more lines of output?
Sorry, my mistake; I had it in my mind that we were talking about a drm_WARN_ON rather than just drm_warn (i.e., including a big stacktrace and such). DRM_ERROR -> drm_warn alone shouldn't have any extra negative impact.
But it can be drm_err for all I care, I don't think we really have consistent story between errors and warnings in this area.
probably multiplying the already bad overhead that shouldn't be happening in an interrupt handler. But if we could update the interrupt handler to just save away the details and do the actual drm_warn later, outside the interrupt handler code, that would be okay. We should probably work toward something like that for all of our interrupt handler warning/error messages.
Not sure I agree - for messages which we don't expect to see it doesn't really matter that there will be overhead when they are hit. Presumably bad things are already happening there so spending effort to optimise those path is questionable.
Something bad is happening to graphics is we hit one of these cases. But if we start doing prints while interrupts are disabled, we start having more of a negative impact on the rest of the system too.
Truly for the case of this particular patch I don't think we should care. Rate limiting should be all that is needed in the short term to strike a balance between effort and benefit. But lets first clarify the PCI link going down problem.
Regards,
Tvrtko
On 06/06/2022 16:21, Matt Roper wrote:
On Mon, Jun 06, 2022 at 12:55:20PM +0100, Tvrtko Ursulin wrote:
On 27/05/2022 19:42, Matt Roper wrote:
On Thu, May 26, 2022 at 11:18:17AM +0100, Tvrtko Ursulin wrote:
On 25/05/2022 19:05, Matt Roper wrote:
On Wed, May 25, 2022 at 05:03:13PM +0100, Tvrtko Ursulin wrote:
On 24/05/2022 18:51, Matt Roper wrote: > On Tue, May 24, 2022 at 10:43:39AM +0100, Tvrtko Ursulin wrote: >> From: Tvrtko Ursulin tvrtko.ursulin@intel.com >> >> Catch and log any garbage in the register, including no tiles marked, or >> multiple tiles marked. >> >> Signed-off-by: Tvrtko Ursulin tvrtko.ursulin@intel.com >> Cc: Matt Roper matthew.d.roper@intel.com >> --- >> We caught garbage in DG1_MSTR_TILE_INTR with DG2 (actual value 0xF9D2C008) >> during glmark and more badness. So I thought lets log all possible failure >> modes from here and also use per device logging. >> --- >> drivers/gpu/drm/i915/i915_irq.c | 33 ++++++++++++++++++++++----------- >> drivers/gpu/drm/i915/i915_reg.h | 1 + >> 2 files changed, 23 insertions(+), 11 deletions(-) >> >> diff --git a/drivers/gpu/drm/i915/i915_irq.c b/drivers/gpu/drm/i915/i915_irq.c >> index 73cebc6aa650..79853d3fc1ed 100644 >> --- a/drivers/gpu/drm/i915/i915_irq.c >> +++ b/drivers/gpu/drm/i915/i915_irq.c >> @@ -2778,24 +2778,30 @@ static irqreturn_t dg1_irq_handler(int irq, void *arg) >> u32 gu_misc_iir; >> if (!intel_irqs_enabled(i915)) >> - return IRQ_NONE; >> + goto none; >> master_tile_ctl = dg1_master_intr_disable(regs); >> - if (!master_tile_ctl) { >> - dg1_master_intr_enable(regs); >> - return IRQ_NONE; >> + if (!master_tile_ctl) >> + goto enable_none; >> + >> + if (master_tile_ctl & ~(DG1_MSTR_IRQ | DG1_MSTR_TILE_MASK)) { >> + drm_warn(&i915->drm, "Garbage in master_tile_ctl: 0x%08x!\n", >> + master_tile_ctl); > > I know we have a bunch of them already, but shouldn't we be avoiding > printk-based stuff like this inside interrupt handlers? Should we be > migrating all these error messages over to trace_printk or something > similar that's safer to use?
Not sure - I kind of think some really unexpected and worrying situations should be loud and on by default. Risk is then spam if not ratelimited. Maybe we should instead ratelimit most errors/warnings coming for irq handlers?
It's not the risk of spam that's the problem, but rather that printk-based stuff eventually calls into the console code to flush its buffers. That's way more overhead than you want in an interrupt handler so it's bad on its own, but if you're using something slow like a serial console, it becomes even more of a problem.
Is it a problem for messages which we never expect to see?
Kind of. While not as catastrophic, it's the same argument for why we don't use BUG() anymore...when the impossible does manage to happen there's unnecessary collateral damage on things outside of graphics. If we're adding huge delays inside an interrupt handler (while other interrupts are disabled) that impacts the system-wide usability, not just our own driver.
I'd also argue that these messages actually are semi-expected. Random bits being set shouldn't happen, but in the world of dgpu's, we do occasionally see cases where the PCI link itself goes down for reasons outside our control and then all registers read back as 0xFFFFFFFF, which will probably trigger error messages here (as well as a bunch of other places).
Could you expand a bit on what is semi-expected and when? I mean the circumstances of PCI link going down. We certainly don't have any code to survive that.
Yeah, I'm referring to the "Lost access to MMIO BAR" errors; in the past most of them have ultimately been tracked down to bugs in early firmware, so flashing an updated IFWI/BIOS onto the device usually solved the problems. Generally those buggy firmwares are an internal problem that never make it into the wild, but I think we have also seen cases where they get triggered by physical/electrical problems on a specific part; that can potentially happen to anyone who's unlucky enough to get a defective/damaged unit.
Basically "hardware returns all F's" happens because the CPU initiates an MMIO transaction with the hardware, the hardware fails to produce any response (possibly due to failing hardware, possibly due to firmware/BIOS bugs), so 0xFFFFFFFF gets returned as an autocompletion to prevent the CPU core from hanging.
It looks like we still have a few open here: https://gitlab.freedesktop.org/search?search=%22Lost+access+to+MMIO+BAR%22&a...
and there are some features on specific platforms we haven't turned on yet because they also trigger these failures (which is still under debug).
We don't/can't really do much to handle these problems in i915 today except printing the 'lost access' error so that we know to ignore whatever kinds of bogus errors we get after that point (usually lots of messages about forcewake failing to clear, engine/GuC reset failing to complete, etc.). But aside from i915 being broken, the rest of the platform should generally continue to work, so you can still access the machine over the network, save logs to disk, etc.
Interesting, I missed the addition of 29b6f88d60dd ("drm/i915: Try to detect sudden loss of MMIO access"), thanks!
In case of my "Garbage in master_tile_ctl" or "Unexpected irq from tile.." messages, in case of lost PCI link they should happen only once. I don't think hardware will keep raising interrupts if driver cannot talk to it. But it does seem prudent to go with the rate-limiting flavour just in case.
Regards,
Tvrtko
On Wed, May 25, 2022 at 05:03:13PM +0100, Tvrtko Ursulin wrote:
On 24/05/2022 18:51, Matt Roper wrote:
On Tue, May 24, 2022 at 10:43:39AM +0100, Tvrtko Ursulin wrote:
From: Tvrtko Ursulin tvrtko.ursulin@intel.com
Catch and log any garbage in the register, including no tiles marked, or multiple tiles marked.
Signed-off-by: Tvrtko Ursulin tvrtko.ursulin@intel.com Cc: Matt Roper matthew.d.roper@intel.com
We caught garbage in DG1_MSTR_TILE_INTR with DG2 (actual value 0xF9D2C008) during glmark and more badness. So I thought lets log all possible failure modes from here and also use per device logging.
drivers/gpu/drm/i915/i915_irq.c | 33 ++++++++++++++++++++++----------- drivers/gpu/drm/i915/i915_reg.h | 1 + 2 files changed, 23 insertions(+), 11 deletions(-)
diff --git a/drivers/gpu/drm/i915/i915_irq.c b/drivers/gpu/drm/i915/i915_irq.c index 73cebc6aa650..79853d3fc1ed 100644 --- a/drivers/gpu/drm/i915/i915_irq.c +++ b/drivers/gpu/drm/i915/i915_irq.c @@ -2778,24 +2778,30 @@ static irqreturn_t dg1_irq_handler(int irq, void *arg) u32 gu_misc_iir; if (!intel_irqs_enabled(i915))
return IRQ_NONE;
goto none;
- if (!master_tile_ctl) {
dg1_master_intr_enable(regs);return IRQ_NONE;
- if (!master_tile_ctl)
goto enable_none;- if (master_tile_ctl & ~(DG1_MSTR_IRQ | DG1_MSTR_TILE_MASK)) {
drm_warn(&i915->drm, "Garbage in master_tile_ctl: 0x%08x!\n",master_tile_ctl);I know we have a bunch of them already, but shouldn't we be avoiding printk-based stuff like this inside interrupt handlers? Should we be migrating all these error messages over to trace_printk or something similar that's safer to use?
Not sure - I kind of think some really unexpected and worrying situations should be loud and on by default. Risk is then spam if not ratelimited. Maybe we should instead ratelimit most errors/warnings coming for irq handlers?
In this particular case at least DRM_ERROR with no device info is the odd one out in the entire file so I'd suggest changing at least that, if the rest of my changes is of questionable benefit.
I'd rather remove the printk's from irq rather than adding more. At the very least, they should be the _once variant or ratelimited. One of the few cases to even deserve a unlikely(), even to document this shouldn't really be happening.
Our irq handlers (particularly on dgfx and multi-gt) are already too long running... I don't like making them any onger or slower.
Lucas De Marchi
On 25/05/2022 19:14, Lucas De Marchi wrote:
On Wed, May 25, 2022 at 05:03:13PM +0100, Tvrtko Ursulin wrote:
On 24/05/2022 18:51, Matt Roper wrote:
On Tue, May 24, 2022 at 10:43:39AM +0100, Tvrtko Ursulin wrote:
From: Tvrtko Ursulin tvrtko.ursulin@intel.com
Catch and log any garbage in the register, including no tiles marked, or multiple tiles marked.
Signed-off-by: Tvrtko Ursulin tvrtko.ursulin@intel.com Cc: Matt Roper matthew.d.roper@intel.com
We caught garbage in DG1_MSTR_TILE_INTR with DG2 (actual value 0xF9D2C008) during glmark and more badness. So I thought lets log all possible failure modes from here and also use per device logging.
drivers/gpu/drm/i915/i915_irq.c | 33 ++++++++++++++++++++++----------- drivers/gpu/drm/i915/i915_reg.h | 1 + 2 files changed, 23 insertions(+), 11 deletions(-)
diff --git a/drivers/gpu/drm/i915/i915_irq.c b/drivers/gpu/drm/i915/i915_irq.c index 73cebc6aa650..79853d3fc1ed 100644 --- a/drivers/gpu/drm/i915/i915_irq.c +++ b/drivers/gpu/drm/i915/i915_irq.c @@ -2778,24 +2778,30 @@ static irqreturn_t dg1_irq_handler(int irq, void *arg) u32 gu_misc_iir; if (!intel_irqs_enabled(i915)) - return IRQ_NONE; + goto none; master_tile_ctl = dg1_master_intr_disable(regs); - if (!master_tile_ctl) { - dg1_master_intr_enable(regs); - return IRQ_NONE; + if (!master_tile_ctl) + goto enable_none;
+ if (master_tile_ctl & ~(DG1_MSTR_IRQ | DG1_MSTR_TILE_MASK)) { + drm_warn(&i915->drm, "Garbage in master_tile_ctl: 0x%08x!\n", + master_tile_ctl);
I know we have a bunch of them already, but shouldn't we be avoiding printk-based stuff like this inside interrupt handlers? Should we be migrating all these error messages over to trace_printk or something similar that's safer to use?
Not sure - I kind of think some really unexpected and worrying situations should be loud and on by default. Risk is then spam if not ratelimited. Maybe we should instead ratelimit most errors/warnings coming for irq handlers?
In this particular case at least DRM_ERROR with no device info is the odd one out in the entire file so I'd suggest changing at least that, if the rest of my changes is of questionable benefit.
I'd rather remove the printk's from irq rather than adding more. At the very least, they should be the _once variant or ratelimited. One of the few cases to even deserve a unlikely(), even to document this shouldn't really be happening.
I would support ratelimited for all the unexpected bits set, no bits set, or similar conditions. I wouldn't remove such printks to micro-optimize things. That would potentially lose important feedback in cases when we hit truly unexpected situations.
But annotating them as unlikely would be a good thing.
Our irq handlers (particularly on dgfx and multi-gt) are already too long running... I don't like making them any onger or slower.
How come? I mean which interrupts are a problem there? Isn't GuC supposed to be taking on that load on itself, isn't that one of the main selling points?
Regards,
Tvrtko
dri-devel@lists.freedesktop.org
-
Lucas De Marchi -
Matt Roper -
Tvrtko Ursulin