dmaengine: idxd: Set workqueue state to disabled before trying to re-enable
Subsystems (2)
PATCH 1/1 dmaengine: idxd: Set workqueue state to disabled before trying to re-enable
For a software reset idxd_device_reinit() is called, which will walk the device workqueues to see which ones were enabled, and try to re-enable them. It keys off wq->state being iDXD_WQ_ENABLED, but the first thing idxd_enable_wq() will do is see that the state of the workqueue is enabled, and return 0 instead of attempting to issue a command to enable the workqueue.
So once a workqueue is found that needs to be re-enabled, set the state to disabled prior to calling idxd_enable_wq(). This would accurately reflect the state if the enable fails as well.
--- drivers/dma/idxd/irq.c | 1 + 1 file changed, 1 insertion(+)
drivers/dma/idxd/irq.c +1 -0
@@ -52,6 +52,7 @@ static void idxd_device_reinit(struct work_struct *work)struct idxd_wq *wq = idxd->wqs[i];if (wq->state ==IDXD_WQ_ENABLED ) { + wq->state =IDXD_WQ_DISABLED ; rc =idxd_wq_enable (wq);if (rc <0 ) { dev_warn(dev, "Unable to re-enable wq %s\n",
Signature · 1 line
2.37.2
Dave Jiang
reviewerQuoted 27 lines · Jerry Snitselaar
For a software reset
idxd_device_reinit()is called, which will walk the device workqueues to see which ones were enabled, and try to re-enable them. It keys offwq->statebeingiDXD_WQ_ENABLED, but the first thingidxd_enable_wq()will do is see that the state of the workqueue is enabled, and return 0 instead of attempting to issue a command to enable the workqueue.So once a workqueue is found that needs to be re-enabled, set the state to disabled prior to calling
idxd_enable_wq(). This would accurately reflect the state if the enable fails as well.Cc Fenghua Yu <[email protected]>Cc Dave Jiang <[email protected]>Cc Vinod Koul <[email protected]>Fixes bfe1d56091c1 ("dmaengine: idxd: Init and probe for Intel data accelerators")Signed-off-by Jerry Snitselaar <[email protected]>--- drivers/dma/idxd/irq.c | 1 + 1 file changed, 1 insertion(+)drivers/dma/idxd/irq.c +0
-0
drivers/dma/idxd/irq.c static void idxd_device_reinit(struct work_struct *work) · line 55
Quoted 4 lines
drivers/dma/idxd/irq.c +1
-0@@ -52,6 +52,7 @@ static void idxd_device_reinit(struct work_struct *work)struct idxd_wq *wq = idxd->wqs[i];if (wq->state ==IDXD_WQ_ENABLED ) { + wq->state =IDXD_WQ_DISABLED ;
Might be better off to insert this line in idxd_wq_disable_cleanup(). I think that should put it in sane state.
Quoted 3 lines
drivers/dma/idxd/irq.c +0
-0rc =idxd_wq_enable (wq);if (rc <0 ) { dev_warn(dev, "Unable to re-enable wq %s\n",
Quoted 24 lines · Dave Jiang
Quoted 22 lines · Jerry Snitselaar
For a software reset
idxd_device_reinit()is called, which will walk the device workqueues to see which ones were enabled, and try to re-enable them. It keys offwq->statebeingiDXD_WQ_ENABLED, but the first thingidxd_enable_wq()will do is see that the state of the workqueue is enabled, and return 0 instead of attempting to issue a command to enable the workqueue.So once a workqueue is found that needs to be re-enabled, set the state to disabled prior to calling
idxd_enable_wq(). This would accurately reflect the state if the enable fails as well.Cc Fenghua Yu <[email protected]>Cc Dave Jiang <[email protected]>Cc Vinod Koul <[email protected]>Fixes bfe1d56091c1 ("dmaengine: idxd: Init and probe for Intel data accelerators")Signed-off-by Jerry Snitselaar <[email protected]>--- drivers/dma/idxd/irq.c | 1 + 1 file changed, 1 insertion(+)
Quoted 10 lines
drivers/dma/idxd/irq.c static void idxd_device_reinit(struct work_struct *work) · line 54
Quoted 8 lines
drivers/dma/idxd/irq.c +1
-0@@ -52,6 +52,7 @@ static void idxd_device_reinit(struct work_struct *work)struct idxd_wq *wq = idxd->wqs[i];if (wq->state ==IDXD_WQ_ENABLED ) { + wq->state =IDXD_WQ_DISABLED ;Might be better off to insert this line in
idxd_wq_disable_cleanup(). I think that should put it in sane state.
I don't think that is called in the code path that I was lookng at. I've been looking at this bit of process_misc_interrupts():
halt:
gensts.bits = ioread32(idxd->reg_base + IDXD_GENSTATS_OFFSET);
if (gensts.state == IDXD_DEVICE_STATE_HALT) {
idxd->state = IDXD_DEV_HALTED;
if (gensts.reset_type == IDXD_DEVICE_RESET_SOFTWARE) {
/*
* If we need a software reset, we will throw the work
* on a system workqueue in order to allow interrupts
* for the device command completions.
*/
INIT_WORK(&idxd->work, idxd_device_reinit);
queue_work(idxd->wq, &idxd->work);
} else {
idxd->state = IDXD_DEV_HALTED;
idxd_wqs_quiesce(idxd);
idxd_wqs_unmap_portal(idxd);
spin_lock(&idxd->dev_lock);
idxd_device_clear_state(idxd);
dev_err(&idxd->pdev->dev,
"idxd halted, need %s.\n",
gensts.reset_type == IDXD_DEVICE_RESET_FLR ?
"FLR" : "system reset");
spin_unlock(&idxd->dev_lock);
return -ENXIO;
}
}
return 0;
}So it sees that the device is halted, and sticks idxd_device_reinint() on that workqueue. The idxd_device_reinit() has this loop to re-enable the idxd wqs:
for (i = 0; i < idxd->max_wqs; i++) {
struct idxd_wq *wq = idxd->wqs[i];
if (wq->state == IDXD_WQ_ENABLED) {
wq->state = IDXD_WQ_DISABLED;
rc = idxd_wq_enable(wq);
if (rc < 0) {
dev_warn(dev, "Unable to re-enable wq %s\n",
dev_name(wq_confdev(wq)));
}
}
}Once you go into idxd_wq_enable() though you get this check at the beginning:
if (wq->state == IDXD_WQ_ENABLED) {
dev_dbg(dev, "WQ %d already enabled\n", wq->id);
return 0;
}So IIUC it sees the device is halted, goes to reset it, figures out a wq should be re-enabled, calls idxd_wq_enable() which hits the check, returns 0 and the wq is never really re-enabled, though it will still have wq state set to IDXD_WQ_ENABLED.
Or am I missing something?
Regards,
Jerry
Quoted 3 lines
Quoted 3 lines
drivers/dma/idxd/irq.c +0
-0rc =idxd_wq_enable (wq);if (rc <0 ) { dev_warn(dev, "Unable to re-enable wq %s\n",
Dave Jiang
reviewerQuoted 67 lines · Jerry Snitselaar
Quoted 23 lines · Dave Jiang
Quoted 22 lines · Jerry Snitselaar
For a software reset
idxd_device_reinit()is called, which will walk the device workqueues to see which ones were enabled, and try to re-enable them. It keys offwq->statebeingiDXD_WQ_ENABLED, but the first thingidxd_enable_wq()will do is see that the state of the workqueue is enabled, and return 0 instead of attempting to issue a command to enable the workqueue.So once a workqueue is found that needs to be re-enabled, set the state to disabled prior to calling
idxd_enable_wq(). This would accurately reflect the state if the enable fails as well.Cc Fenghua Yu <[email protected]>Cc Dave Jiang <[email protected]>Cc Vinod Koul <[email protected]>Fixes bfe1d56091c1 ("dmaengine: idxd: Init and probe for Intel data accelerators")Signed-off-by Jerry Snitselaar <[email protected]>--- drivers/dma/idxd/irq.c | 1 + 1 file changed, 1 insertion(+)Quoted 10 lines
drivers/dma/idxd/irq.c static void idxd_device_reinit(struct work_struct *work) · line 54
Quoted 8 lines
drivers/dma/idxd/irq.c +1
-0@@ -52,6 +52,7 @@ static void idxd_device_reinit(struct work_struct *work)struct idxd_wq *wq = idxd->wqs[i];if (wq->state ==IDXD_WQ_ENABLED ) { + wq->state =IDXD_WQ_DISABLED ;Might be better off to insert this line in
idxd_wq_disable_cleanup(). I think that should put it in sane state.I don't think that is called in the code path that I was lookng at. I've been looking at this bit of
process_misc_interrupts():halt:
gensts.bits = ioread32(idxd->reg_base + IDXD_GENSTATS_OFFSET); if (gensts.state == IDXD_DEVICE_STATE_HALT) { idxd->state = IDXD_DEV_HALTED; if (gensts.reset_type == IDXD_DEVICE_RESET_SOFTWARE) { /* * If we need a software reset, we will throw the work * on a system workqueue in order to allow interrupts * for the device command completions. */ INIT_WORK(&idxd->work, idxd_device_reinit); queue_work(idxd->wq, &idxd->work); } else { idxd->state = IDXD_DEV_HALTED; idxd_wqs_quiesce(idxd); idxd_wqs_unmap_portal(idxd); spin_lock(&idxd->dev_lock); idxd_device_clear_state(idxd); dev_err(&idxd->pdev->dev, "idxd halted, need %s.\n", gensts.reset_type == IDXD_DEVICE_RESET_FLR ? "FLR" : "system reset"); spin_unlock(&idxd->dev_lock); return -ENXIO; } } return 0; }
Quoted 2 lines
So it sees that the device is halted, and sticks
idxd_device_reinint()on that workqueue. Theidxd_device_reinit()has this loop to re-enable the idxd wqs:
idxd_device_reinit() should called idxd_device_reset() first. And that should at some point call idxd_wq_disable_cleanup() and clean up the states.
https://elixir.bootlin.com/linux/v6.0-rc2/source/drivers/dma/idxd/irq.c#L42
https://elixir.bootlin.com/linux/v6.0-rc2/source/drivers/dma/idxd/device.c#L725
https://elixir.bootlin.com/linux/v6.0-rc2/source/drivers/dma/idxd/device.c#L711
https://elixir.bootlin.com/linux/v6.0-rc2/source/drivers/dma/idxd/device.c#L376
So if we stick the wq state reset in there, it should show up as "disabled" by the time we try to enable the WQs again. Does that look reasonable?
Quoted 34 lines
for (i = 0; i < idxd->max_wqs; i++) { struct idxd_wq *wq = idxd->wqs[i]; if (wq->state == IDXD_WQ_ENABLED) { wq->state = IDXD_WQ_DISABLED; rc = idxd_wq_enable(wq); if (rc < 0) { dev_warn(dev, "Unable to re-enable wq %s\n", dev_name(wq_confdev(wq))); } } }Once you go into
idxd_wq_enable()though you get this check at the beginning:if (wq->state == IDXD_WQ_ENABLED) { dev_dbg(dev, "WQ %d already enabled\n", wq->id); return 0; }So IIUC it sees the device is halted, goes to reset it, figures out a wq should be re-enabled, calls
idxd_wq_enable()which hits the check, returns 0 and the wq is never really re-enabled, though it will still have wq state set toIDXD_WQ_ENABLED.Or am I missing something?
Regards,
JerryQuoted 3 lines
Quoted 3 lines
drivers/dma/idxd/irq.c +0
-0rc =idxd_wq_enable (wq);if (rc <0 ) { dev_warn(dev, "Unable to re-enable wq %s\n",
Quoted 97 lines · Dave Jiang
Quoted 78 lines · Jerry Snitselaar
Quoted 29 lines · Dave Jiang
Quoted 28 lines · Jerry Snitselaar
For a software reset
idxd_device_reinit()is called, which will walk
the device workqueues to see which ones were enabled, and try to
re-enable them. It keys offwq->statebeingiDXD_WQ_ENABLED, but the
first thingidxd_enable_wq()will do is see that the state of the
workqueue is enabled, and return 0 instead of attempting to issue
a command to enable the workqueue.So once a workqueue is found that needs to be re-enabled, set the state to disabled prior to calling
idxd_enable_wq(). This would accurately reflect the state if the enable fails as well.Cc Fenghua Yu <[email protected]>Cc Dave Jiang <[email protected]>Cc Vinod Koul <[email protected]>Fixes bfe1d56091c1 ("dmaengine: idxd: Init and probe for Inteldata accelerators")
Signed-off-by Jerry Snitselaar <[email protected]>---
drivers/dma/idxd/irq.c| 1 +1 file changed, 1 insertion(+)Quoted 12 lines
drivers/dma/idxd/irq.c static void idxd_device_reinit(struct · line 55
Quoted 9 lines
drivers/dma/idxd/irq.c +1
-0@@ -52,6 +52,7 @@ static void idxd_device_reinit(struct work_struct *work) struct idxd_wq *wq = idxd->wqs[i]; if (wq->state == IDXD_WQ_ENABLED) { + wq->state = IDXD_WQ_DISABLED;Might be better off to insert this line in
idxd_wq_disable_cleanup(). I
think that should put it in sane state.I don't think that is called in the code path that I was lookng at. I've been
looking at this bit ofprocess_misc_interrupts():halt:
gensts.bits =ioread32(idxd->reg_base + IDXD_GENSTATS_OFFSET);
if (gensts.state ==IDXD_DEVICE_STATE_HALT) {idxd->state=IDXD_DEV_HALTED;
if (gensts.reset_type ==IDXD_DEVICE_RESET_SOFTWARE) {
/*
* If we need a software reset, we will throw the work
* on a system workqueue in order to allow interrupts
* for the device command completions. */
INIT_WORK(&idxd->work,idxd_device_reinit);queue_work(idxd->wq, &idxd->work); } else {
idxd->state =IDXD_DEV_HALTED;idxd_wqs_quiesce(idxd);idxd_wqs_unmap_portal(idxd);spin_lock(&idxd->dev_lock);idxd_device_clear_state(idxd); dev_err(&idxd->pdev->dev,
"idxd halted, need %s.\n", gensts.reset_type ==IDXD_DEVICE_RESET_FLR?
"FLR" : "system reset");spin_unlock(&idxd->dev_lock); return-ENXIO;
}
}Quoted 7 lines
return 0;
}So it sees that the device is halted, and sticks
idxd_device_reinint()on that
workqueue. Theidxd_device_reinit()has this loop to re-enable the idxd wqs:
idxd_device_reinit()should calledidxd_device_reset()first. And that
should at some point callidxd_wq_disable_cleanup()and clean up the states.https://elixir.bootlin.com/linux/v6.0-rc2/source/drivers/dma/idxd/irq.c#L42
https://elixir.bootlin.com/linux/v6.0-rc2/source/drivers/dma/idxd/device.c#L725
Quoted 8 lines
https://elixir.bootlin.com/linux/v6.0-rc2/source/drivers/dma/idxd/device.c#L711
https://elixir.bootlin.com/linux/v6.0-rc2/source/drivers/dma/idxd/device.c#L376
So if we stick the wq state reset in there, it should show up as "disabled" by the time we try to enable the WQs again. Does that look reasonable?
Ah, yeah I see that now. So, if it does set the state to disabled in idxd_wq_disable_cleanup(), does it have another means to track which wqs need to be re-enabled for that loop that happens after the idxd_device_reset() call?
Quoted 42 lines
Quoted 41 lines
for (i = 0; i <
idxd->max_wqs; i++) { structidxd_wq*wq = idxd->wqs[i];if (
wq->state==IDXD_WQ_ENABLED) {wq->state=IDXD_WQ_DISABLED;
rc =idxd_wq_enable(wq);
if (rc < 0) {
dev_warn(dev, "Unable to re-enable wq %s\n",
dev_name(wq_confdev(wq))); }
}
}Once you go into
idxd_wq_enable()though you get this check at the beginning:if (
wq->state==IDXD_WQ_ENABLED) {dev_dbg(dev, "WQ %d already enabled\n", wq->id); return 0;
}So IIUC it sees the device is halted, goes to reset it, figures out a wq
should be re-enabled, callsidxd_wq_enable()which hits the check, returns
0 and the wq is never really re-enabled, though it will still have wq state
set toIDXD_WQ_ENABLED.Or am I missing something?
Regards,
JerryQuoted 4 lines
Quoted 4 lines
drivers/dma/idxd/irq.c +0
-0rc = idxd_wq_enable(wq); if (rc < 0) {dev_warn(dev, "Unable to re- enable wq %s\n",
Dave Jiang
reviewerQuoted 100 lines · Jerry Snitselaar
Quoted 99 lines · Dave Jiang
Quoted 78 lines · Jerry Snitselaar
Quoted 29 lines · Dave Jiang
Quoted 28 lines · Jerry Snitselaar
For a software reset
idxd_device_reinit()is called, which will walk
the device workqueues to see which ones were enabled, and try to
re-enable them. It keys offwq->statebeingiDXD_WQ_ENABLED, but the
first thingidxd_enable_wq()will do is see that the state of the
workqueue is enabled, and return 0 instead of attempting to issue
a command to enable the workqueue.So once a workqueue is found that needs to be re-enabled, set the state to disabled prior to calling
idxd_enable_wq(). This would accurately reflect the state if the enable fails as well.Cc Fenghua Yu <[email protected]>Cc Dave Jiang <[email protected]>Cc Vinod Koul <[email protected]>Fixes bfe1d56091c1 ("dmaengine: idxd: Init and probe for Inteldata accelerators")
Signed-off-by Jerry Snitselaar <[email protected]>--- drivers/dma/idxd/irq.c | 1 + 1 file changed, 1 insertion(+)Quoted 12 lines
drivers/dma/idxd/irq.c static void idxd_device_reinit(struct · line 55
Quoted 9 lines
drivers/dma/idxd/irq.c +1
-0@@ -52,6 +52,7 @@ static void idxd_device_reinit(struct work_struct *work) struct idxd_wq *wq = idxd->wqs[i]; if (wq->state == IDXD_WQ_ENABLED) { + wq->state = IDXD_WQ_DISABLED;Might be better off to insert this line in
idxd_wq_disable_cleanup(). I
think that should put it in sane state.I don't think that is called in the code path that I was lookng at. I've been
looking at this bit ofprocess_misc_interrupts():halt:
gensts.bits = ioread32(idxd->reg_base +IDXD_GENSTATS_OFFSET);
if (gensts.state ==IDXD_DEVICE_STATE_HALT) {idxd->state=IDXD_DEV_HALTED;
if (gensts.reset_type ==IDXD_DEVICE_RESET_SOFTWARE) {
/*
* If we need a software reset, we will
throw the work
* on a system workqueue in order to allow
interrupts
* for the device command completions. */
INIT_WORK(&idxd->work,idxd_device_reinit);queue_work(idxd->wq, &idxd->work); } else {
idxd->state =IDXD_DEV_HALTED;idxd_wqs_quiesce(idxd);idxd_wqs_unmap_portal(idxd);spin_lock(&idxd->dev_lock);idxd_device_clear_state(idxd); dev_err(&idxd->pdev->dev,
"idxd halted, need %s.\n", gensts.reset_type ==IDXD_DEVICE_RESET_FLR?
"FLR" : "system reset");spin_unlock(&idxd->dev_lock); return-ENXIO;
}
}Quoted 7 lines
return 0;
}So it sees that the device is halted, and sticks
idxd_device_reinint()on that
workqueue. Theidxd_device_reinit()has this loop to re-enable the idxd wqs:
idxd_device_reinit()should calledidxd_device_reset()first. And that
should at some point callidxd_wq_disable_cleanup()and clean up the states.https://elixir.bootlin.com/linux/v6.0-rc2/source/drivers/dma/idxd/irq.c#L42
https://elixir.bootlin.com/linux/v6.0-rc2/source/drivers/dma/idxd/device.c#L725
https://elixir.bootlin.com/linux/v6.0-rc2/source/drivers/dma/idxd/device.c#L711
https://elixir.bootlin.com/linux/v6.0-rc2/source/drivers/dma/idxd/device.c#L376
Quoted 8 lines
Quoted 4 lines
So if we stick the wq state reset in there, it should show up as "disabled" by the time we try to enable the WQs again. Does that look reasonable?
Ah, yeah I see that now. So, if it does set the state to disabled in
idxd_wq_disable_cleanup(), does it have another means to track which wqs need to be re-enabled for that loop that happens after theidxd_device_reset()call?
Oh I see what you mean... So we can either do what you did or create a mask and mark the WQ that are "enabled" before reset. Maybe that's cleaner rather than relying on the side effect of the WQ state isn't cleared? Thoughts?
Quoted 40 lines
Quoted 40 lines
Quoted 40 lines
for (i = 0; i <
idxd->max_wqs; i++) { structidxd_wq*wq = idxd->wqs[i];if (
wq->state==IDXD_WQ_ENABLED) {wq->state=IDXD_WQ_DISABLED;
rc =idxd_wq_enable(wq);
if (rc < 0) {
dev_warn(dev, "Unable to re-enable
wq %s\n",
dev_name(wq_confdev(wq))); }
}
}Once you go into
idxd_wq_enable()though you get this check at the beginning:if (
wq->state==IDXD_WQ_ENABLED) {dev_dbg(dev, "WQ %d already enabled\n", wq->id); return 0;
}So IIUC it sees the device is halted, goes to reset it, figures out a wq
should be re-enabled, callsidxd_wq_enable()which hits the check, returns
0 and the wq is never really re-enabled, though it will still have wq state
set toIDXD_WQ_ENABLED.Or am I missing something?
Regards,
JerryQuoted 4 lines
Quoted 4 lines
drivers/dma/idxd/irq.c +0
-0rc = idxd_wq_enable(wq); if (rc < 0) { dev_warn(dev, "Unable to re-enable wq %s\n",
Quoted 127 lines · Dave Jiang
Quoted 118 lines · Jerry Snitselaar
Quoted 108 lines · Dave Jiang
Quoted 89 lines · Jerry Snitselaar
Quoted 35 lines · Dave Jiang
Quoted 34 lines · Jerry Snitselaar
For a software reset
idxd_device_reinit()is called, which will
walk
the device workqueues to see which ones were enabled, and try
to
re-enable them. It keys offwq->statebeingiDXD_WQ_ENABLED,
but the
first thingidxd_enable_wq()will do is see that the state of
the
workqueue is enabled, and return 0 instead of attempting to issue
a command to enable the workqueue.So once a workqueue is found that needs to be re-enabled, set the state to disabled prior to calling
idxd_enable_wq().
This would accurately reflect the state if the enable fails as well.Cc Fenghua Yu <[email protected]>Cc Dave Jiang <[email protected]>Cc Vinod Koul <[email protected]>Fixes bfe1d56091c1 ("dmaengine: idxd: Init and probe forIntel
data accelerators")Signed-off-by Jerry Snitselaar <[email protected]>---
drivers/dma/idxd/irq.c| 1 +1 file changed, 1 insertion(+)Quoted 13 lines
drivers/dma/idxd/irq.c static void idxd_device_reinit(struct · line 55
Quoted 10 lines
drivers/dma/idxd/irq.c +1
-0@@ -52,6 +52,7 @@ static void idxd_device_reinit(struct work_struct *work) struct idxd_wq *wq = idxd->wqs[i]; if (wq->state == IDXD_WQ_ENABLED) { + wq->state = IDXD_WQ_DISABLED;Might be better off to insert this line in
idxd_wq_disable_cleanup(). I
think that should put it in sane state.I don't think that is called in the code path that I was lookng at.
I've been
looking at this bit ofprocess_misc_interrupts():halt:
gensts.bits =ioread32(idxd->reg_base + IDXD_GENSTATS_OFFSET);
if (gensts.state ==IDXD_DEVICE_STATE_HALT) {idxd->state=IDXD_DEV_HALTED;
if (gensts.reset_type ==IDXD_DEVICE_RESET_SOFTWARE) {
/*
* If we need a software reset, we will
throw the work
* on a system workqueue in order to allow
interrupts
* for the device command completions. */
INIT_WORK(&idxd->work,idxd_device_reinit);
queue_work(idxd->wq,&idxd->work); } else {
idxd->state =IDXD_DEV_HALTED;idxd_wqs_quiesce(idxd);idxd_wqs_unmap_portal(idxd);spin_lock(&idxd->dev_lock);idxd_device_clear_state(idxd); dev_err(&idxd->pdev->dev, "idxd halted, need %s.\n", gensts.reset_type ==IDXD_DEVICE_RESET_FLR?
"FLR" : "system reset");spin_unlock(&idxd->dev_lock); return-ENXIO;
}
}Quoted 8 lines
return 0;
}So it sees that the device is halted, and sticks
idxd_device_reinint()on that
workqueue. Theidxd_device_reinit()has this loop to re-enable the
idxd wqs:
idxd_device_reinit()should calledidxd_device_reset()first. And that
should at some point callidxd_wq_disable_cleanup()and clean up the
states.https://elixir.bootlin.com/linux/v6.0-rc2/source/drivers/dma/idxd/irq.c#L42
https://elixir.bootlin.com/linux/v6.0-rc2/source/drivers/dma/idxd/device.c#L725
Quoted 9 lines
https://elixir.bootlin.com/linux/v6.0-rc2/source/drivers/dma/idxd/device.c#L711
https://elixir.bootlin.com/linux/v6.0-rc2/source/drivers/dma/idxd/device.c#L376
So if we stick the wq state reset in there, it should show up as "disabled" by the time we try to enable the WQs again. Does that look
reasonable?Quoted 6 lines
Ah, yeah I see that now. So, if it does set the state to disabled in
idxd_wq_disable_cleanup(), does it have another means to track which
wqs need to be re-enabled for that loop that happens after theidxd_device_reset()call?
Quoted 6 lines
Oh I see what you mean... So we can either do what you did or create a
mask and mark the WQ that are "enabled" before reset. Maybe that's cleaner rather than relying on the side effect of the WQ state isn't cleared? Thoughts?
That would be better I think.
Quoted 50 lines
Quoted 49 lines
Quoted 49 lines
Quoted 49 lines
for (i = 0; i <
idxd->max_wqs; i++) { structidxd_wq*wq = idxd->wqs[i];if (
wq->state==IDXD_WQ_ENABLED) {wq->state=IDXD_WQ_DISABLED; rc =idxd_wq_enable(wq); if (rc < 0) { dev_warn(dev, "Unable to re- enable
wq %s\n",
dev_name(wq_confdev(wq)));
}
}
}Once you go into
idxd_wq_enable()though you get this check at the
beginning:if (
wq->state==IDXD_WQ_ENABLED) { dev_dbg(dev, "WQ %d already enabled\n", wq-Quoted 1 line
id);
return 0;
}So IIUC it sees the device is halted, goes to reset it, figures out
a wq
should be re-enabled, callsidxd_wq_enable()which hits the check,
returns
0 and the wq is never really re-enabled, though it will still have
wq state
set toIDXD_WQ_ENABLED.Or am I missing something?
Regards,
JerryQuoted 5 lines
Quoted 5 lines
drivers/dma/idxd/irq.c +0
-0rc = idxd_wq_enable(wq); if (rc < 0) {dev_warn(dev, "Unable to re-
enable wq %s\n",
Quoted 106 lines · Dave Jiang
Quoted 104 lines · Jerry Snitselaar
Quoted 103 lines · Dave Jiang
Quoted 78 lines · Jerry Snitselaar
Quoted 29 lines · Dave Jiang
Quoted 28 lines · Jerry Snitselaar
For a software reset
idxd_device_reinit()is called, which will walk
the device workqueues to see which ones were enabled, and try to
re-enable them. It keys offwq->statebeingiDXD_WQ_ENABLED, but the
first thingidxd_enable_wq()will do is see that the state of the
workqueue is enabled, and return 0 instead of attempting to issue
a command to enable the workqueue.So once a workqueue is found that needs to be re-enabled, set the state to disabled prior to calling
idxd_enable_wq(). This would accurately reflect the state if the enable fails as well.Cc Fenghua Yu <[email protected]>Cc Dave Jiang <[email protected]>Cc Vinod Koul <[email protected]>Fixes bfe1d56091c1 ("dmaengine: idxd: Init and probe for Inteldata accelerators")
Signed-off-by Jerry Snitselaar <[email protected]>--- drivers/dma/idxd/irq.c | 1 + 1 file changed, 1 insertion(+)Quoted 12 lines
drivers/dma/idxd/irq.c static void idxd_device_reinit(struct · line 55
Quoted 9 lines
drivers/dma/idxd/irq.c +1
-0@@ -52,6 +52,7 @@ static void idxd_device_reinit(struct work_struct *work) struct idxd_wq *wq = idxd->wqs[i]; if (wq->state == IDXD_WQ_ENABLED) { + wq->state = IDXD_WQ_DISABLED;Might be better off to insert this line in
idxd_wq_disable_cleanup(). I
think that should put it in sane state.I don't think that is called in the code path that I was lookng at. I've been
looking at this bit ofprocess_misc_interrupts():halt:
gensts.bits = ioread32(idxd->reg_base +IDXD_GENSTATS_OFFSET);
if (gensts.state ==IDXD_DEVICE_STATE_HALT) {idxd->state=IDXD_DEV_HALTED;
if (gensts.reset_type ==IDXD_DEVICE_RESET_SOFTWARE) {
/*
* If we need a software reset, we will
throw the work
* on a system workqueue in order to allow
interrupts
* for the device command completions. */
INIT_WORK(&idxd->work,idxd_device_reinit);queue_work(idxd->wq, &idxd->work); } else {
idxd->state =IDXD_DEV_HALTED;idxd_wqs_quiesce(idxd);idxd_wqs_unmap_portal(idxd);spin_lock(&idxd->dev_lock);idxd_device_clear_state(idxd); dev_err(&idxd->pdev->dev,
"idxd halted, need %s.\n", gensts.reset_type ==IDXD_DEVICE_RESET_FLR?
"FLR" : "system reset");spin_unlock(&idxd->dev_lock); return-ENXIO;
}
}Quoted 7 lines
return 0;
}So it sees that the device is halted, and sticks
idxd_device_reinint()on that
workqueue. Theidxd_device_reinit()has this loop to re-enable the idxd wqs:
idxd_device_reinit()should calledidxd_device_reset()first. And that
should at some point callidxd_wq_disable_cleanup()and clean up the states.https://elixir.bootlin.com/linux/v6.0-rc2/source/drivers/dma/idxd/irq.c#L42
https://elixir.bootlin.com/linux/v6.0-rc2/source/drivers/dma/idxd/device.c#L725
https://elixir.bootlin.com/linux/v6.0-rc2/source/drivers/dma/idxd/device.c#L711
https://elixir.bootlin.com/linux/v6.0-rc2/source/drivers/dma/idxd/device.c#L376
So if we stick the wq state reset in there, it should show up as "disabled" by the time we try to enable the WQs again. Does that look reasonable?
Quoted 9 lines
Quoted 4 lines
Ah, yeah I see that now. So, if it does set the state to disabled in
idxd_wq_disable_cleanup(), does it have another means to track which wqs need to be re-enabled for that loop that happens after theidxd_device_reset()call?Oh I see what you mean... So we can either do what you did or create a mask and mark the WQ that are "enabled" before reset. Maybe that's cleaner rather than relying on the side effect of the WQ state isn't cleared? Thoughts?
Circling back to this. Since max_wqs could theoretically go up to 2^8, I guess this would need to be done with the bitmap_* functions?
Regards,
Jerry
Quoted 41 lines
Quoted 41 lines
Quoted 41 lines
Quoted 41 lines
for (i = 0; i <
idxd->max_wqs; i++) { structidxd_wq*wq = idxd->wqs[i];if (
wq->state==IDXD_WQ_ENABLED) {wq->state=IDXD_WQ_DISABLED;
rc =idxd_wq_enable(wq);
if (rc < 0) {
dev_warn(dev, "Unable to re-enable
wq %s\n",
dev_name(wq_confdev(wq))); }
}
}Once you go into
idxd_wq_enable()though you get this check at the beginning:if (
wq->state==IDXD_WQ_ENABLED) {dev_dbg(dev, "WQ %d already enabled\n", wq->id); return 0;
}So IIUC it sees the device is halted, goes to reset it, figures out a wq
should be re-enabled, callsidxd_wq_enable()which hits the check, returns
0 and the wq is never really re-enabled, though it will still have wq state
set toIDXD_WQ_ENABLED.Or am I missing something?
Regards,
JerryQuoted 4 lines
Quoted 4 lines
drivers/dma/idxd/irq.c +0
-0rc = idxd_wq_enable(wq); if (rc < 0) { dev_warn(dev, "Unable to re-enable wq %s\n",
Dave Jiang
reviewerQuoted 110 lines · Jerry Snitselaar
Quoted 109 lines · Dave Jiang
Quoted 108 lines · Jerry Snitselaar
Quoted 95 lines · Dave Jiang
Quoted 78 lines · Jerry Snitselaar
Quoted 29 lines · Dave Jiang
Quoted 28 lines · Jerry Snitselaar
For a software reset
idxd_device_reinit()is called, which will walk
the device workqueues to see which ones were enabled, and try to
re-enable them. It keys offwq->statebeingiDXD_WQ_ENABLED, but the
first thingidxd_enable_wq()will do is see that the state of the
workqueue is enabled, and return 0 instead of attempting to issue
a command to enable the workqueue.So once a workqueue is found that needs to be re-enabled, set the state to disabled prior to calling
idxd_enable_wq(). This would accurately reflect the state if the enable fails as well.Cc Fenghua Yu <[email protected]>Cc Dave Jiang <[email protected]>Cc Vinod Koul <[email protected]>Fixes bfe1d56091c1 ("dmaengine: idxd: Init and probe for Inteldata accelerators")
Signed-off-by Jerry Snitselaar <[email protected]>--- drivers/dma/idxd/irq.c | 1 + 1 file changed, 1 insertion(+)Quoted 12 lines
drivers/dma/idxd/irq.c static void idxd_device_reinit(struct · line 55
Quoted 9 lines
drivers/dma/idxd/irq.c +1
-0@@ -52,6 +52,7 @@ static void idxd_device_reinit(struct work_struct *work) struct idxd_wq *wq = idxd->wqs[i]; if (wq->state == IDXD_WQ_ENABLED) { + wq->state = IDXD_WQ_DISABLED;Might be better off to insert this line in
idxd_wq_disable_cleanup(). I
think that should put it in sane state.I don't think that is called in the code path that I was lookng at. I've been
looking at this bit ofprocess_misc_interrupts():halt:
gensts.bits = ioread32(idxd->reg_base +IDXD_GENSTATS_OFFSET);
if (gensts.state ==IDXD_DEVICE_STATE_HALT) {idxd->state=IDXD_DEV_HALTED;
if (gensts.reset_type ==IDXD_DEVICE_RESET_SOFTWARE) {
/*
* If we need a software reset, we will
throw the work
* on a system workqueue in order to allow
interrupts
* for the device command completions. */
INIT_WORK(&idxd->work,idxd_device_reinit);queue_work(idxd->wq, &idxd->work); } else {
idxd->state =IDXD_DEV_HALTED;idxd_wqs_quiesce(idxd);idxd_wqs_unmap_portal(idxd);spin_lock(&idxd->dev_lock);idxd_device_clear_state(idxd); dev_err(&idxd->pdev->dev,
"idxd halted, need %s.\n", gensts.reset_type ==IDXD_DEVICE_RESET_FLR?
"FLR" : "system reset");spin_unlock(&idxd->dev_lock); return-ENXIO;
}
}Quoted 7 lines
return 0;
}So it sees that the device is halted, and sticks
idxd_device_reinint()on that
workqueue. Theidxd_device_reinit()has this loop to re-enable the idxd wqs:
idxd_device_reinit()should calledidxd_device_reset()first. And that
should at some point callidxd_wq_disable_cleanup()and clean up the states.https://elixir.bootlin.com/linux/v6.0-rc2/source/drivers/dma/idxd/irq.c#L42
https://elixir.bootlin.com/linux/v6.0-rc2/source/drivers/dma/idxd/device.c#L725
Quoted 8 lines
https://elixir.bootlin.com/linux/v6.0-rc2/source/drivers/dma/idxd/device.c#L711
https://elixir.bootlin.com/linux/v6.0-rc2/source/drivers/dma/idxd/device.c#L376
So if we stick the wq state reset in there, it should show up as "disabled" by the time we try to enable the WQs again. Does that look reasonable?
Ah, yeah I see that now. So, if it does set the state to disabled in
idxd_wq_disable_cleanup(), does it have another means to track which wqs need to be re-enabled for that loop that happens after theidxd_device_reset()call?
Quoted 6 lines
Quoted 4 lines
Oh I see what you mean... So we can either do what you did or create a mask and mark the WQ that are "enabled" before reset. Maybe that's cleaner rather than relying on the side effect of the WQ state isn't cleared? Thoughts?
Circling back to this. Since
max_wqscould theoretically go up to 2^8, I guess this would need to be done with the bitmap_* functions?
Hi Jerry,
I wouldn't say never but I doubt any time soon for 2^8. DSA 1.0 has 8 WQs, and 2.0 (spec just went public) has 16. But yes we can use bitmap to be future proof. Are you currently working on a fix for this? Just don't want to duplicate effort if you already have something going. Thank you!
Quoted 43 lines
Regards,
JerryQuoted 39 lines
Quoted 39 lines
Quoted 39 lines
Quoted 39 lines
for (i = 0; i < idxd->max_wqs; i++) { struct idxd_wq *wq = idxd->wqs[i];if (
wq->state==IDXD_WQ_ENABLED) {wq->state=IDXD_WQ_DISABLED;
rc =idxd_wq_enable(wq);
if (rc < 0) {
dev_warn(dev, "Unable to re-enable
wq %s\n",
dev_name(wq_confdev(wq))); }
}
}Once you go into
idxd_wq_enable()though you get this check at the beginning:if (wq->state == IDXD_WQ_ENABLED) { dev_dbg(dev, "WQ %d already enabled\n", wq->id); return 0; }So IIUC it sees the device is halted, goes to reset it, figures out a wq
should be re-enabled, callsidxd_wq_enable()which hits the check, returns
0 and the wq is never really re-enabled, though it will still have wq state
set toIDXD_WQ_ENABLED.Or am I missing something?
Regards,
JerryQuoted 4 lines
Quoted 4 lines
drivers/dma/idxd/irq.c +0
-0rc = idxd_wq_enable(wq); if (rc < 0) { dev_warn(dev, "Unable to re-enable wq %s\n",
Quoted 119 lines · Dave Jiang
Quoted 110 lines · Jerry Snitselaar
Quoted 109 lines · Dave Jiang
Quoted 108 lines · Jerry Snitselaar
Quoted 95 lines · Dave Jiang
Quoted 78 lines · Jerry Snitselaar
Quoted 29 lines · Dave Jiang
Quoted 28 lines · Jerry Snitselaar
For a software reset
idxd_device_reinit()is called, which will walk
the device workqueues to see which ones were enabled, and try to
re-enable them. It keys offwq->statebeingiDXD_WQ_ENABLED, but the
first thingidxd_enable_wq()will do is see that the state of the
workqueue is enabled, and return 0 instead of attempting to issue
a command to enable the workqueue.So once a workqueue is found that needs to be re-enabled, set the state to disabled prior to calling
idxd_enable_wq(). This would accurately reflect the state if the enable fails as well.Cc Fenghua Yu <[email protected]>Cc Dave Jiang <[email protected]>Cc Vinod Koul <[email protected]>Fixes bfe1d56091c1 ("dmaengine: idxd: Init and probe for Inteldata accelerators")
Signed-off-by Jerry Snitselaar <[email protected]>--- drivers/dma/idxd/irq.c | 1 + 1 file changed, 1 insertion(+)Quoted 12 lines
drivers/dma/idxd/irq.c static void idxd_device_reinit(struct · line 55
Quoted 9 lines
drivers/dma/idxd/irq.c +1
-0@@ -52,6 +52,7 @@ static void idxd_device_reinit(struct work_struct *work) struct idxd_wq *wq = idxd->wqs[i]; if (wq->state == IDXD_WQ_ENABLED) { + wq->state = IDXD_WQ_DISABLED;Might be better off to insert this line in
idxd_wq_disable_cleanup(). I
think that should put it in sane state.I don't think that is called in the code path that I was lookng at. I've been
looking at this bit ofprocess_misc_interrupts():halt:
gensts.bits = ioread32(idxd->reg_base +IDXD_GENSTATS_OFFSET);
if (gensts.state ==IDXD_DEVICE_STATE_HALT) {idxd->state=IDXD_DEV_HALTED;
if (gensts.reset_type ==IDXD_DEVICE_RESET_SOFTWARE) {
/*
* If we need a software reset, we will
throw the work
* on a system workqueue in order to allow
interrupts
* for the device command completions. */
INIT_WORK(&idxd->work,idxd_device_reinit);queue_work(idxd->wq, &idxd->work); } else {
idxd->state =IDXD_DEV_HALTED;idxd_wqs_quiesce(idxd);idxd_wqs_unmap_portal(idxd);spin_lock(&idxd->dev_lock);idxd_device_clear_state(idxd); dev_err(&idxd->pdev->dev,
"idxd halted, need %s.\n", gensts.reset_type ==IDXD_DEVICE_RESET_FLR?
"FLR" : "system reset");spin_unlock(&idxd->dev_lock); return-ENXIO;
}
}Quoted 7 lines
return 0;
}So it sees that the device is halted, and sticks
idxd_device_reinint()on that
workqueue. Theidxd_device_reinit()has this loop to re-enable the idxd wqs:
idxd_device_reinit()should calledidxd_device_reset()first. And that
should at some point callidxd_wq_disable_cleanup()and clean up the states.https://elixir.bootlin.com/linux/v6.0-rc2/source/drivers/dma/idxd/irq.c#L42
https://elixir.bootlin.com/linux/v6.0-rc2/source/drivers/dma/idxd/device.c#L725
Quoted 8 lines
https://elixir.bootlin.com/linux/v6.0-rc2/source/drivers/dma/idxd/device.c#L711
https://elixir.bootlin.com/linux/v6.0-rc2/source/drivers/dma/idxd/device.c#L376
So if we stick the wq state reset in there, it should show up as "disabled" by the time we try to enable the WQs again. Does that look reasonable?
Ah, yeah I see that now. So, if it does set the state to disabled in
idxd_wq_disable_cleanup(), does it have another means to track which wqs need to be re-enabled for that loop that happens after theidxd_device_reset()call?Quoted 6 lines
Quoted 4 lines
Oh I see what you mean... So we can either do what you did or create a mask and mark the WQ that are "enabled" before reset. Maybe that's cleaner rather than relying on the side effect of the WQ state isn't cleared? Thoughts?
Circling back to this. Since
max_wqscould theoretically go up to 2^8, I guess this would need to be done with the bitmap_* functions?
Quoted 7 lines
Hi Jerry,
I wouldn't say never but I doubt any time soon for 2^8. DSA 1.0 has 8 WQs, and 2.0 (spec just went public) has 16. But yes we can use bitmap to be future proof. Are you currently working on a fix for this? Just don't want to duplicate effort if you already have something going. Thank you!
Hi Dave,
Yes, I will try to send it this afternoon.
Regards,
Jerry
Quoted 44 lines
Quoted 44 lines
Regards,
JerryQuoted 39 lines
Quoted 39 lines
Quoted 39 lines
Quoted 39 lines
for (i = 0; i < idxd->max_wqs; i++) { struct idxd_wq *wq = idxd->wqs[i];if (
wq->state==IDXD_WQ_ENABLED) {wq->state=IDXD_WQ_DISABLED;
rc =idxd_wq_enable(wq);
if (rc < 0) {
dev_warn(dev, "Unable to re-enable
wq %s\n",
dev_name(wq_confdev(wq))); }
}
}Once you go into
idxd_wq_enable()though you get this check at the beginning:if (wq->state == IDXD_WQ_ENABLED) { dev_dbg(dev, "WQ %d already enabled\n", wq->id); return 0; }So IIUC it sees the device is halted, goes to reset it, figures out a wq
should be re-enabled, callsidxd_wq_enable()which hits the check, returns
0 and the wq is never really re-enabled, though it will still have wq state
set toIDXD_WQ_ENABLED.Or am I missing something?
Regards,
JerryQuoted 4 lines
Quoted 4 lines
drivers/dma/idxd/irq.c +0
-0rc = idxd_wq_enable(wq); if (rc < 0) { dev_warn(dev, "Unable to re-enable wq %s\n",