Skip to contentSkip to conversation

PCI: Fix faulty logic in pci_reset_bus()

PATCH In review Sinan Kaya 6 participants 13 messages Sep 5, 2018 – Sep 12, 2018

Subsystems (6)
lore
Share & tools
lore

Sinan Kaya

author
#0 lore raw

PATCH v2 1/3 PCI: Fix faulty logic in pci_reset_bus()

From: Dennis Dalessandro <[email protected]>

The pci_rest_bus() function calls into pci_probe_reset_slot() to determine whether to call the slot or bus reset. The check has faulty logic in that it does not account for pci_probe_reset_slot() being able to return an errno. Fix by only calling the slot reset when the function returns 0. Treat < 1 and > 1 the same.

Cc Sinan Kaya <[email protected]>
Fixes 811c5cb37df4 ("PCI: Unify try slot and bus reset API")
---
 drivers/pci/pci.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
drivers/pci/pci.c +1 -1
index 29ff9619b5fa..30b260332a10 100644
@@ -5200,7 +5200,7 @@ static int __pci_reset_bus(struct pci_bus *bus)
  */
 int pci_reset_bus(struct pci_dev *pdev)
 {
-	return pci_probe_reset_slot(pdev->slot) ?
+	return (!pci_probe_reset_slot(pdev->slot)) ?
 	    __pci_reset_slot(pdev->slot) : __pci_reset_bus(pdev->bus);
 }
 EXPORT_SYMBOL_GPL(pci_reset_bus);
Signature · 1 line
2.18.0
author
#1 lore raw

PATCH v2 2/3 PCI: Request reset type as part of function reset

We need a contract between the reset API users and the PCI core about the types of reset that a user needs vs. what PCI core can do internally. If a platform supports hotplug, we need to do hotplug reset as an example.

Expose the reset types to the drivers and try different reset types based on the new reset_type parameter.

Most users are expected to use PCI_RESET_ANY, PCI_RESET_FUNC or PCI_RESET_LINK parameters.

Fixes 811c5cb37df4 ("PCI: Unify try slot and bus reset API")
---
 .../net/ethernet/cavium/liquidio/lio_main.c   |  2 +-
 drivers/pci/pci.c                             | 59 ++++++++++++-------
 drivers/xen/xen-pciback/pci_stub.c            |  6 +-
 include/linux/pci.h                           | 13 +++-
 4 files changed, 55 insertions(+), 25 deletions(-)
drivers/net/ethernet/cavium/liquidio/lio_main.c +1 -1
index 6fb13fa73b27..0ff76722734d 100644
@@ -989,7 +989,7 @@ static void octeon_pci_flr(struct octeon_device *oct)
 	pci_write_config_word(oct->pci_dev, PCI_COMMAND,
 			      PCI_COMMAND_INTX_DISABLE);
 
-	rc = __pci_reset_function_locked(oct->pci_dev);
+	rc = __pci_reset_function_locked(oct->pci_dev, PCI_RESET_ANY);
 
 	if (rc != 0)
 		dev_err(&oct->pci_dev->dev, "Error %d resetting PCI function %d\n",
drivers/pci/pci.c +39 -20

99 lines not on this page · raw message

drivers/xen/xen-pciback/pci_stub.c +3 -3
index 59661db144e5..6dfb805bcb19 100644
@@ -105,7 +105,7 @@ static void pcistub_device_release(struct kref *kref)
 	/* Call the reset function which does not take lock as this
 	 * is called from "unbind" which takes a device_lock mutex.
 	 */
-	__pci_reset_function_locked(dev);
+	__pci_reset_function_locked(dev, PCI_RESET_ANY);
 	if (pci_load_and_free_saved_state(dev, &dev_data->pci_saved_state))
 		dev_info(&dev->dev, "Could not reload PCI state\n");
 	else
@@ -283,7 +283,7 @@ void pcistub_put_pci_dev(struct pci_dev *dev)
 	 * (so it's ready for the next domain)
 	 */
 	device_lock_assert(&dev->dev);
-	__pci_reset_function_locked(dev);
+	__pci_reset_function_locked(dev, PCI_RESET_ANY);
 
 	dev_data = pci_get_drvdata(dev);
 	ret = pci_load_saved_state(dev, dev_data->pci_saved_state);
@@ -417,7 +417,7 @@ static int pcistub_init_device(struct pci_dev *dev)
 		dev_err(&dev->dev, "Could not store PCI conf saved state!\n");
 	else {
 		dev_dbg(&dev->dev, "resetting (FLR, D3, etc) the device\n");
-		__pci_reset_function_locked(dev);
+		__pci_reset_function_locked(dev, PCI_RESET_ANY);
 		pci_restore_state(dev);
 	}
 	/* Now disable the device (this also ensures some private device
include/linux/pci.h +12 -1
index e72ca8dd6241..6924d837c959 100644
@@ -849,6 +849,17 @@ enum {
 	PCI_SCAN_ALL_PCIE_DEVS	= 0x00000040,	/* Scan all, not just dev 0 */
 };
 
+#define PCI_RESET_DEV_SPECIFIC	(1 << 0)
+#define PCI_RESET_FLR		(1 << 1)
+#define PCI_RESET_PM		(1 << 2)
+#define PCI_RESET_SLOT		(1 << 3)
+#define PCI_RESET_BUS		(1 << 4)
+
+#define PCI_RESET_ANY		(~0)
+#define PCI_RESET_FUNC		(PCI_RESET_DEV_SPECIFIC | \
+				 PCI_RESET_FLR | PCI_RESET_PM)
+#define PCI_RESET_LINK		(PCI_RESET_SLOT | PCI_RESET_BUS)
+
 /* These external functions are only available when PCI support is enabled */
 #ifdef CONFIG_PCI
 
@@ -1111,7 +1122,7 @@ u32 pcie_bandwidth_available(struct pci_dev *dev, struct pci_dev **limiting_dev,
 void pcie_print_link_status(struct pci_dev *dev);
 bool pcie_has_flr(struct pci_dev *dev);
 int pcie_flr(struct pci_dev *dev);
-int __pci_reset_function_locked(struct pci_dev *dev);
+int __pci_reset_function_locked(struct pci_dev *dev, u32 reset_type);
 int pci_reset_function(struct pci_dev *dev);
 int pci_reset_function_locked(struct pci_dev *dev);
 int pci_try_reset_function(struct pci_dev *dev);
Signature · 1 line
2.18.0
author
#3 lore raw

PATCH v2 3/3 IB/hfi1: Prefer new __pci_reset_function_locked() API with reset type

In reply to Sinan Kaya

pci_reset_bus() is being called from the probe context and causes a deadlock since pci_reset_bus() also tries to obtain kernel lock.

Use __pci_reset_function_locked() that provides locking guarantees.

Fixes 811c5cb37df4 ("PCI: Unify try slot and bus reset API")
---
 drivers/infiniband/hw/hfi1/pcie.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
drivers/infiniband/hw/hfi1/pcie.c +1 -1
index eec83757d55f..13162289b748 100644
@@ -900,7 +900,7 @@ static int trigger_sbr(struct hfi1_devdata *dd)
 	 * delay after a reset is required.  Per spec requirements,
 	 * the link is either working or not after that point.
 	 */
-	return pci_reset_bus(dev);
+	return __pci_reset_function_locked(dev, PCI_RESET_LINK);
 }
 
 /*
Signature · 1 line
2.18.0
Acked-by on v2 3/3
#4 lore raw

Re: [PATCH for-rc v2 3/3] IB/hfi1: Prefer new __pci_reset_function_locked() API with reset type

Quoted 8 lines · Sinan Kaya

pci_reset_bus() is being called from the probe context and causes a deadlock since pci_reset_bus() also tries to obtain kernel lock.

Use __pci_reset_function_locked() that provides locking guarantees.

Fixes 811c5cb37df4 ("PCI: Unify try slot and bus reset API")
Quoted 17 lines
---
 drivers/infiniband/hw/hfi1/pcie.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
drivers/infiniband/hw/hfi1/pcie.c +1 -1
index eec83757d55f..13162289b748 100644
@@ -900,7 +900,7 @@ static int trigger_sbr(struct hfi1_devdata *dd)
 	 * delay after a reset is required.  Per spec requirements,
 	 * the link is either working or not after that point.
 	 */
-	return pci_reset_bus(dev);
+	return __pci_reset_function_locked(dev, PCI_RESET_LINK);
 }
 
 /*
Signature · 3 lines
Doug Ledford <[email protected]>
    GPG KeyID: B826A3330E572FDD
    Key fingerprint = AE6B 1BDA 122B 23B4 265B  1274 B826 A333 0E57 2FDD
Tested-by on v2 3/3
#5 lore raw

Re: [PATCH for-rc v2 3/3] IB/hfi1: Prefer new __pci_reset_function_locked() API with reset type

In reply to Sinan Kaya
Quoted 8 lines · Sinan Kaya

pci_reset_bus() is being called from the probe context and causes a deadlock since pci_reset_bus() also tries to obtain kernel lock.

Use __pci_reset_function_locked() that provides locking guarantees.

Fixes 811c5cb37df4 ("PCI: Unify try slot and bus reset API")
Reviewed-by on v2 2/3
#2 lore raw

Re: [PATCH for-rc v2 2/3] PCI: Request reset type as part of function reset

In reply to Sinan Kaya
Quoted 7 lines · Sinan Kaya

We need a contract between the reset API users and the PCI core about the types of reset that a user needs vs. what PCI core can do internally. If a platform supports hotplug, we need to do hotplug reset as an example.

Expose the reset types to the drivers and try different reset types based on the new reset_type parameter.

Quoted 7 lines

Most users are expected to use PCI_RESET_ANY, PCI_RESET_FUNC or PCI_RESET_LINK parameters.

Fixes 811c5cb37df4 ("PCI: Unify try slot and bus reset API")
author
#7 lore raw

Re: [PATCH for-rc v2 3/3] IB/hfi1: Prefer new __pci_reset_function_locked() API with reset type

In reply to Bjorn Helgaas
Quoted 5 lines · Bjorn Helgaas
Quoted 2 lines · Sinan Kaya

pci_reset_bus() is being called from the probe context and causes a deadlock since pci_reset_bus() also tries to obtain kernel lock.

This doesn't tell me what the deadlock is.

Sorry, let me try to set the context.

Driver probe routine is being called with pci_dev_lock() held.

pci_reset_bus() tries to obtain the pci_dev_lock() via this path one more time while holding the lock.

pci_dev_lock()

	enter driver probe()
	pci_reset_bus()
		 __pci_reset_bus()
			pci_bus_save_and_disable()
				pci_dev_lock()

	exit driver probe()

pci_dev_unlock()

Quoted 5 lines
Quoted 2 lines

Use __pci_reset_function_locked() that provides locking guarantees.

And the previous patch that adds the "reset_type" parameter doesn't tell me anything about what locking guarantees it provides.

This is sort of implied.

As the name of the function is __pci_reset_function_locked(), it is assumed that the caller is responsible for obtaining the necessary locks before calling this function.

Quoted 6 lines

I want to merge minimal changes for v4.19 to fix the known problems.

It's not clear to me what actually broke hfi1. You mention a deadlock above, so I assume some locking change broke it, but 811c5cb37df4 isn't obviously related to locking.

Previous code was calling secondary bus reset function directly. With the generalization of the code, we are going through device save and restore as follows:

1. obtain device lock
2. save device state
3. release device lock
4. perform secondary bus reset
5. obtain device lock
6. restore device state
7. release device lock

We have introduced the locks in step #1, #3, #5 and #7 into the existing code path.

Quoted 4 lines

It's obvious that 811c5cb37df4 tested the return value of pci_probe_reset_slot() incorrectly, so there's no problem with patch 1/1.

yup.

Quoted 6 lines

But patches 2 & 3:

PCI: Request reset type as part of function reset IB/hfi1: Prefer new __pci_reset_function_locked() API with reset type

do not connect the dots for me.

I hope the above explanation helps. The consensus after conversation with Alex was to reuse the locked API but add masks so that requester can selectively request a slot reset without any locks.

Quoted 25 lines
Quoted 24 lines
Fixes 811c5cb37df4 ("PCI: Unify try slot and bus reset API")
---
  drivers/infiniband/hw/hfi1/pcie.c | 2 +-
  1 file changed, 1 insertion(+), 1 deletion(-)
drivers/infiniband/hw/hfi1/pcie.c +1 -1
index eec83757d55f..13162289b748 100644
@@ -900,7 +900,7 @@ static int trigger_sbr(struct hfi1_devdata *dd)
  	 * delay after a reset is required.  Per spec requirements,
  	 * the link is either working or not after that point.
  	 */
-	return pci_reset_bus(dev);
+	return __pci_reset_function_locked(dev, PCI_RESET_LINK);
  }
 
  /*
Signature · 1 line
2.18.0
author
#9 lore raw

Re: [PATCH for-rc v2 3/3] IB/hfi1: Prefer new __pci_reset_function_locked() API with reset type

In reply to Bjorn Helgaas
Quoted 4 lines · Bjorn Helgaas

Can we just move the pci_bridge_secondary_bus_reset() decl back to include/linux/pci.h temporarily and have trigger_sbr() call that again? I'd rather do that than try to squeeze some API rework into an -rc.

yes, that was the initial proposal from Dennis. If you prefer to have a simpler/temporary change instead for the moment, correct patches are:

https://patchwork.kernel.org/patch/10584283/
https://patchwork.kernel.org/patch/10584277/

#10 lore raw

Re: [PATCH for-rc v2 3/3] IB/hfi1: Prefer new __pci_reset_function_locked() API with reset type

In reply to Bjorn Helgaas
Quoted 5 lines · Bjorn Helgaas

Can we just move the pci_bridge_secondary_bus_reset() decl back to include/linux/pci.h temporarily and have trigger_sbr() call that again? I'd rather do that than try to squeeze some API rework into an -rc.

That was what I was initially hoping for [1]. Just get things back to working for now, especially since we are approaching the later -rc's. That being said, I have reviewed and tested Sinan's patch. So far so good, but I do see the concern.

[1] https://marc.info/?l=linux-pci&m=153573732824788&w=2

-Denny

#11 lore raw

Re: [PATCH for-rc v2 3/3] IB/hfi1: Prefer new __pci_reset_function_locked() API with reset type

In reply to Bjorn Helgaas
Quoted 94 lines · On Tue, 11 Sep 2018 20:59:27 -0500 Bjorn Helgaas <[email protected]> wrote:

Thanks for the quick response (and sorry for my late one).

Quoted 26 lines · Sinan Kaya
Quoted 5 lines · Bjorn Helgaas
Quoted 2 lines · Sinan Kaya

pci_reset_bus() is being called from the probe context and causes a deadlock since pci_reset_bus() also tries to obtain kernel lock.

This doesn't tell me what the deadlock is.

Sorry, let me try to set the context.

Driver probe routine is being called with pci_dev_lock() held.

pci_reset_bus() tries to obtain the pci_dev_lock() via this path one more time while holding the lock.

pci_dev_lock()

	enter driver probe()
	pci_reset_bus()
		 __pci_reset_bus()
			pci_bus_save_and_disable()
				pci_dev_lock()

	exit driver probe()

pci_dev_unlock()

Quoted 1 line
Quoted 6 lines
Quoted 4 lines
Quoted 1 line

Use __pci_reset_function_locked() that provides locking guarantees.

And the previous patch that adds the "reset_type" parameter doesn't tell me anything about what locking guarantees it provides.

This is sort of implied.

Uhh, maybe, but the implication is too subtle for me. It adds a "reset_type" that can be FLR, PM, SLOT, BUS, etc. Those are different types of resets, but those types say nothing about locking.

Quoted 5 lines

As the name of the function is __pci_reset_function_locked(), it is assumed that the caller is responsible for obtaining the necessary locks before calling this function.

Quoted 1 line
Quoted 7 lines
Quoted 5 lines

I want to merge minimal changes for v4.19 to fix the known problems.

It's not clear to me what actually broke hfi1. You mention a deadlock above, so I assume some locking change broke it, but 811c5cb37df4 isn't obviously related to locking.

Previous code was calling secondary bus reset function directly. With the

Ah, so it looks like 409888e0966e ("IB/hfi1: Use pci_try_reset_bus() for initiating PCI Secondary Bus Reset") is an essential part of the picture here. That should probably be part of one of these changelogs.

409888e0966e is where the locking change happened. trigger_sbr() previously called pci_reset_bridge_secondary_bus(), but after 409888e0966e, it calls pci_try_reset_bus(), which at the time called pci_bus_trylock() *before* calling pci_reset_bridge_secondary_bus().

Can we just move the pci_bridge_secondary_bus_reset() decl back to include/linux/pci.h temporarily and have trigger_sbr() call that again? I'd rather do that than try to squeeze some API rework into an -rc.

Quoted 26 lines

generalization of the code, we are going through device save and restore as follows:

1. obtain device lock
2. save device state
3. release device lock
4. perform secondary bus reset
5. obtain device lock
6. restore device state
7. release device lock

We have introduced the locks in step #1, #3, #5 and #7 into the existing code path.

Quoted 4 lines

It's obvious that 811c5cb37df4 tested the return value of pci_probe_reset_slot() incorrectly, so there's no problem with patch 1/1.

yup.

Quoted 5 lines

But patches 2 & 3:

PCI: Request reset type as part of function reset IB/hfi1: Prefer new __pci_reset_function_locked() API with reset type

Quoted 8 lines
Quoted 5 lines
Quoted 1 line

do not connect the dots for me.

I hope the above explanation helps. The consensus after conversation with Alex was to reuse the locked API but add masks so that requester can selectively request a slot reset without any locks.

I'm not buying this yet. There's nothing in that API that says to me "requesting certain types of resets doesn't acquire locks".

The type of reset doesn't determine the locking, it's that the _locked() version of pci_reset_function() requires the caller to provide the locking and the "__" prefix version requires the caller to handle the device save and restore. So the hfi1 requirement was nearly satisfied by the existing exported __pci_reset_function_locked(), but they want a link level reset, not an FLR, but the hotplug or sbr resets are the last options of all the pci_reset_function() choices currently. Therefore adding a reset type as a parameter fixes both the locking issue and the device save/restore, which the hfi1 driver duplicates, and still makes the hfi1 driver work better with hotplug slots, where their previous sbr might trigger surprise hotplugs.

Additionally, the original series unified pci_reset_bus() such that the caller no longer has control of whether a sbr or hotplug reset is used, which is undesirable for vfio-pci (811c5cb37df4 PCI: Unify try slot and bus reset API) which really wants to specify since we do a fair bit or work making sure the user owns the right set of devices for the available type of reset. Abstracting this allows PCI core and vfio to get out of sync. The idea of allowing the caller to specify the type of reset is applicable here as well, unify the function, but still allow the caller to specify.

For rc, I'd certainly agree that a simple revert might be a better option and this API can play out for v4.20, but I definitely want to get back an API that allows vfio to specify the bus reset type before implementations diverge. Thanks,

Alex

Quoted 24 lines
Quoted 24 lines
Quoted 23 lines
Quoted 22 lines
Fixes 811c5cb37df4 ("PCI: Unify try slot and bus reset API")
---
  drivers/infiniband/hw/hfi1/pcie.c | 2 +-
  1 file changed, 1 insertion(+), 1 deletion(-)
drivers/infiniband/hw/hfi1/pcie.c +1 -1
index eec83757d55f..13162289b748 100644
@@ -900,7 +900,7 @@ static int trigger_sbr(struct hfi1_devdata *dd)
  	 * delay after a reset is required.  Per spec requirements,
  	 * the link is either working or not after that point.
  	 */
-	return pci_reset_bus(dev);
+	return __pci_reset_function_locked(dev, PCI_RESET_LINK);
  }
  /*
Signature · 1 line
2.18.0
author
#12 lore raw

Re: [PATCH for-rc v2 3/3] IB/hfi1: Prefer new __pci_reset_function_locked() API with reset type

Quoted 4 lines · Alex Williamson

For rc, I'd certainly agree that a simple revert might be a better option and this API can play out for v4.20, but I definitely want to get back an API that allows vfio to specify the bus reset type before implementations diverge. Thanks,

I posted these per Alex's feedback:

https://patchwork.ozlabs.org/patch/967557/
https://patchwork.ozlabs.org/patch/967558/
https://patchwork.ozlabs.org/patch/967559/
https://patchwork.ozlabs.org/patch/967560/

Please feel free to comment on them targeting v4.20.

#6 lore raw

Re: [PATCH for-rc v2 3/3] IB/hfi1: Prefer new __pci_reset_function_locked() API with reset type

In reply to Sinan Kaya
Quoted 2 lines · Sinan Kaya

pci_reset_bus() is being called from the probe context and causes a deadlock since pci_reset_bus() also tries to obtain kernel lock.

This doesn't tell me what the deadlock is.

Quoted 1 line

Use __pci_reset_function_locked() that provides locking guarantees.

And the previous patch that adds the "reset_type" parameter doesn't tell me anything about what locking guarantees it provides.

I want to merge minimal changes for v4.19 to fix the known problems.

It's not clear to me what actually broke hfi1. You mention a deadlock above, so I assume some locking change broke it, but 811c5cb37df4 isn't obviously related to locking.

It's obvious that 811c5cb37df4 tested the return value of pci_probe_reset_slot() incorrectly, so there's no problem with patch 1/1.

But patches 2 & 3:

PCI: Request reset type as part of function reset
IB/hfi1: Prefer new __pci_reset_function_locked() API with reset type

do not connect the dots for me.

Quoted 23 lines
Fixes 811c5cb37df4 ("PCI: Unify try slot and bus reset API")
---
 drivers/infiniband/hw/hfi1/pcie.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
drivers/infiniband/hw/hfi1/pcie.c +1 -1
index eec83757d55f..13162289b748 100644
@@ -900,7 +900,7 @@ static int trigger_sbr(struct hfi1_devdata *dd)
 	 * delay after a reset is required.  Per spec requirements,
 	 * the link is either working or not after that point.
 	 */
-	return pci_reset_bus(dev);
+	return __pci_reset_function_locked(dev, PCI_RESET_LINK);
 }
 
 /*
Signature · 1 line
2.18.0
#8 lore raw

Re: [PATCH for-rc v2 3/3] IB/hfi1: Prefer new __pci_reset_function_locked() API with reset type

In reply to Sinan Kaya

Thanks for the quick response (and sorry for my late one).

Quoted 26 lines · Sinan Kaya
Quoted 5 lines · Bjorn Helgaas
Quoted 2 lines · Sinan Kaya

pci_reset_bus() is being called from the probe context and causes a deadlock since pci_reset_bus() also tries to obtain kernel lock.

This doesn't tell me what the deadlock is.

Sorry, let me try to set the context.

Driver probe routine is being called with pci_dev_lock() held.

pci_reset_bus() tries to obtain the pci_dev_lock() via this path one more time while holding the lock.

pci_dev_lock()

	enter driver probe()
	pci_reset_bus()
		 __pci_reset_bus()
			pci_bus_save_and_disable()
				pci_dev_lock()

	exit driver probe()

pci_dev_unlock()

Quoted 1 line
Quoted 6 lines
Quoted 4 lines
Quoted 1 line

Use __pci_reset_function_locked() that provides locking guarantees.

And the previous patch that adds the "reset_type" parameter doesn't tell me anything about what locking guarantees it provides.

This is sort of implied.

Uhh, maybe, but the implication is too subtle for me. It adds a "reset_type" that can be FLR, PM, SLOT, BUS, etc. Those are different types of resets, but those types say nothing about locking.

Quoted 5 lines

As the name of the function is __pci_reset_function_locked(), it is assumed that the caller is responsible for obtaining the necessary locks before calling this function.

Quoted 1 line
Quoted 7 lines
Quoted 5 lines

I want to merge minimal changes for v4.19 to fix the known problems.

It's not clear to me what actually broke hfi1. You mention a deadlock above, so I assume some locking change broke it, but 811c5cb37df4 isn't obviously related to locking.

Previous code was calling secondary bus reset function directly. With the

Ah, so it looks like 409888e0966e ("IB/hfi1: Use pci_try_reset_bus() for initiating PCI Secondary Bus Reset") is an essential part of the picture here. That should probably be part of one of these changelogs.

409888e0966e is where the locking change happened. trigger_sbr() previously called pci_reset_bridge_secondary_bus(), but after 409888e0966e, it calls pci_try_reset_bus(), which at the time called pci_bus_trylock() *before* calling pci_reset_bridge_secondary_bus().

Can we just move the pci_bridge_secondary_bus_reset() decl back to include/linux/pci.h temporarily and have trigger_sbr() call that again? I'd rather do that than try to squeeze some API rework into an -rc.

Quoted 23 lines

generalization of the code, we are going through device save and restore as follows:

1. obtain device lock
2. save device state
3. release device lock
4. perform secondary bus reset
5. obtain device lock
6. restore device state
7. release device lock

We have introduced the locks in step #1, #3, #5 and #7 into the existing code path.

Quoted 4 lines

It's obvious that 811c5cb37df4 tested the return value of pci_probe_reset_slot() incorrectly, so there's no problem with patch 1/1.

yup.

Quoted 2 lines

But patches 2 & 3:

Quoted 8 lines
Quoted 4 lines

PCI: Request reset type as part of function reset IB/hfi1: Prefer new __pci_reset_function_locked() API with reset type

do not connect the dots for me.

I hope the above explanation helps. The consensus after conversation with Alex was to reuse the locked API but add masks so that requester can selectively request a slot reset without any locks.

I'm not buying this yet. There's nothing in that API that says to me "requesting certain types of resets doesn't acquire locks".

Quoted 24 lines
Quoted 23 lines
Quoted 22 lines
Fixes 811c5cb37df4 ("PCI: Unify try slot and bus reset API")
---
  drivers/infiniband/hw/hfi1/pcie.c | 2 +-
  1 file changed, 1 insertion(+), 1 deletion(-)
drivers/infiniband/hw/hfi1/pcie.c +1 -1
index eec83757d55f..13162289b748 100644
@@ -900,7 +900,7 @@ static int trigger_sbr(struct hfi1_devdata *dd)
  	 * delay after a reset is required.  Per spec requirements,
  	 * the link is either working or not after that point.
  	 */
-	return pci_reset_bus(dev);
+	return __pci_reset_function_locked(dev, PCI_RESET_LINK);
  }
  /*
Signature · 1 line
2.18.0