Skip to contentSkip to conversation

PCI: Minimizing resource assignment algorithm

PATCH In review Jon Derrick 2 participants 8 messages Sep 28, 2020 – Sep 29, 2020

Subsystems (3)
lore
Share & tools
lore
#0 lore raw

cover PCI: Minimizing resource assignment algorithm

This set adds a minimizing resource assignment algorithm. VMD domains frequently have issues with default hotplug settings and large arrays of drives such as those in JBOFs. This algorithm uses the default or user-specified hotplug resource settings, then tries with minimal settings using 256 for IO, and 1MB for MMIO and Prefetch, and finally tries without additional hotplug resources as if the bridge were not hotplug capable.

This set allows a resource constrained domain to at the very least enumerate and attach drivers to devices, though may not result in supportable hotplug slot if a device is not already occupied in the slot.

Jon Derrick (3):
PCI: Create helper to release/restore bridge resources
PCI: Introduce a minimizing assignment algorithm
PCI: vmd: Wire up VMD for fallback resource assignment

 drivers/pci/controller/vmd.c |   2 +-
 drivers/pci/setup-bus.c      | 147 ++++++++++++++++++++++++++++-------
 include/linux/pci.h          |   2 +
 3 files changed, 124 insertions(+), 27 deletions(-)
Signature · 1 line
2.18.1
#1 lore raw

PATCH 1/3 PCI: Create helper to release/restore bridge resources

Moves bridge release and restore code into a common helper. No functional changes.

---
 drivers/pci/setup-bus.c | 49 +++++++++++++++++++++++------------------
 1 file changed, 28 insertions(+), 21 deletions(-)
drivers/pci/setup-bus.c +28 -21
index 3951e02b7ded..f22502e8e6e6 100644
@@ -2047,6 +2047,33 @@ static void pci_bridge_distribute_available_resources(struct pci_dev *bridge,
 					       available_mmio_pref);
 }
 
+static void release_and_restore_resources(struct list_head *head)
+{
+	struct pci_dev_resource *dev_res;
+
+	list_for_each_entry(dev_res, head, list)
+		pci_bus_release_bridge_resources(dev_res->dev->bus,
+						 dev_res->flags & PCI_RES_TYPE_MASK,
+						 whole_subtree);
+
+	/* Restore size and flags */
+	list_for_each_entry(dev_res, head, list) {
+		struct resource *res = dev_res->res;
+		int idx;
+
+		res->start = dev_res->start;
+		res->end = dev_res->end;
+		res->flags = dev_res->flags;
+
+		if (pci_is_bridge(dev_res->dev)) {
+			idx = res - &dev_res->dev->resource[0];
+			if (idx >= PCI_BRIDGE_RESOURCES &&
+			    idx <= PCI_BRIDGE_RESOURCE_END)
+				res->flags = 0;
+		}
+	}
+}
+
 void pci_assign_unassigned_bridge_resources(struct pci_dev *bridge)
 {
 	struct pci_bus *parent = bridge->subordinate;
@@ -2088,27 +2115,7 @@ void pci_assign_unassigned_bridge_resources(struct pci_dev *bridge)
 	 * Try to release leaf bridge's resources that aren't big enough
 	 * to contain child device resources.
 	 */
-	list_for_each_entry(fail_res, &fail_head, list)
-		pci_bus_release_bridge_resources(fail_res->dev->bus,
-						 fail_res->flags & PCI_RES_TYPE_MASK,
-						 whole_subtree);
-
-	/* Restore size and flags */
-	list_for_each_entry(fail_res, &fail_head, list) {
-		struct resource *res = fail_res->res;
-		int idx;
-
-		res->start = fail_res->start;
-		res->end = fail_res->end;
-		res->flags = fail_res->flags;
-
-		if (pci_is_bridge(fail_res->dev)) {
-			idx = res - &fail_res->dev->resource[0];
-			if (idx >= PCI_BRIDGE_RESOURCES &&
-			    idx <= PCI_BRIDGE_RESOURCE_END)
-				res->flags = 0;
-		}
-	}
+	release_and_restore_resources(&fail_head);
 	free_list(&fail_head);
 
 	goto again;
Signature · 1 line
2.18.1
#2 lore raw

PATCH 2/3 PCI: Introduce a minimizing assignment algorithm

In reply to Jon Derrick

Some PCI domains have limited resources that get exhausted by hotplug resource domains. VMD subdevice domains, for example, tend to support only 32MB MMIO, of which the decoable address space is split between prefetchable and non-prefetchable windows using existing resource assignment algorithms. In addition to these limitations, hotplug bridges require additional resource reservations as specified by default or module parameters "pci=hp{io,mmio,mmiopref}size, further exhausting the domain resources prior to full domain assignment.

Introduce a minimizing assignment algorithm which starts with the default or user-requested hotplug resource values, tries with minimal hotplug resource values, and lastly tries no hotplug resource values.

---
 drivers/pci/setup-bus.c | 98 ++++++++++++++++++++++++++++++++++++++---
 include/linux/pci.h     |  2 +
 2 files changed, 95 insertions(+), 5 deletions(-)
drivers/pci/setup-bus.c +93 -5
index f22502e8e6e6..7beb4f37660b 100644
@@ -1200,6 +1200,35 @@ static void pci_bus_size_cardbus(struct pci_bus *bus,
 	;
 }
 
+enum {
+	PCI_SIZING_VARIANT_DEFAULT,
+	PCI_SIZING_VARIANT_NOHOTPLUG,
+	PCI_SIZING_VARIANT_MINIMUM,
+	PCI_NUM_SIZING_VARIANTS,
+};
+
+static void hotplug_sizes(int sizing_variant, resource_size_t *io,
+			  resource_size_t *mmio, resource_size_t *pref)
+{
+	switch (sizing_variant) {
+	case PCI_SIZING_VARIANT_MINIMUM:
+		*io = 0;
+		*mmio = 0;
+		*pref = 0;
+		break;
+	case PCI_SIZING_VARIANT_NOHOTPLUG:
+		*io = 256;
+		*mmio = 1 << 20;
+		*pref = 1 << 20;
+		break;
+	case PCI_SIZING_VARIANT_DEFAULT:
+	default:
+		*io = pci_hotplug_io_size;
+		*mmio = pci_hotplug_mmio_size;
+		*pref = pci_hotplug_mmio_pref_size;
+	}
+}
+
 void __pci_bus_size_bridges(struct pci_bus *bus, struct list_head *realloc_head)
 {
 	struct pci_dev *dev;
@@ -1248,11 +1277,11 @@ void __pci_bus_size_bridges(struct pci_bus *bus, struct list_head *realloc_head)
 
 	case PCI_HEADER_TYPE_BRIDGE:
 		pci_bridge_check_ranges(bus);
-		if (bus->self->is_hotplug_bridge) {
-			additional_io_size  = pci_hotplug_io_size;
-			additional_mmio_size = pci_hotplug_mmio_size;
-			additional_mmio_pref_size = pci_hotplug_mmio_pref_size;
-		}
+		if (bus->self->is_hotplug_bridge)
+			hotplug_sizes(bus->self->sizing_variant,
+				      &additional_io_size,
+				      &additional_mmio_size,
+				      &additional_mmio_pref_size);
 		/* Fall through */
 	default:
 		pbus_size_io(bus, realloc_head ? 0 : additional_io_size,
@@ -2247,3 +2276,62 @@ void pci_assign_unassigned_bus_resources(struct pci_bus *bus)
 	BUG_ON(!list_empty(&add_list));
 }
 EXPORT_SYMBOL_GPL(pci_assign_unassigned_bus_resources);
+
+static int __set_sizing_variant(struct pci_dev *dev, void *data)
+{
+	if (dev->is_hotplug_bridge)
+		dev->sizing_variant = *((int *) data);
+
+	return 0;
+}
+
+static void release_bridge_resources(struct pci_bus *bus)
+{
+	struct resource *res;
+	struct pci_dev *dev;
+	int i;
+
+	list_for_each_entry(dev, &bus->devices, bus_list) {
+		if (dev->subordinate) {
+			for (i = PCI_BRIDGE_RESOURCES; i < PCI_BRIDGE_RESOURCE_END; i++)
+				reset_resource(&dev->resource[i]);
+
+			release_bridge_resources(dev->subordinate);
+		}
+
+		if (pci_is_root_bus(bus))
+			continue;
+
+		pci_bus_for_each_resource(bus, res, i)
+			reset_resource(res);
+	}
+}
+
+void pci_bus_assign_resources_fallback_sizing(struct pci_bus *bus)
+{
+	LIST_HEAD(fail_head);
+	int i = 0;
+
+	pci_walk_bus(bus, __set_sizing_variant, &i);
+	__pci_bus_assign_resources(bus, NULL, &fail_head);
+
+	if (list_empty(&fail_head))
+		return;
+
+	for (i = 0; i < PCI_NUM_SIZING_VARIANTS; i++) {
+		pci_walk_bus(bus, __set_sizing_variant, &i);
+
+		down_read(&pci_bus_sem);
+		__pci_bus_size_bridges(bus, NULL);
+		up_read(&pci_bus_sem);
+
+		__pci_bus_assign_resources(bus, NULL, &fail_head);
+		if (list_empty(&fail_head))
+			return;
+
+		release_and_restore_resources(&fail_head);
+		release_bridge_resources(bus);
+		free_list(&fail_head);
+	}
+}
+EXPORT_SYMBOL(pci_bus_assign_resources_fallback_sizing);
include/linux/pci.h +2 -0
index 801e9ad0d57e..72ae11d3b5ea 100644
@@ -424,6 +424,7 @@ struct pci_dev {
 	unsigned int	is_hotplug_bridge:1;
 	unsigned int	shpc_managed:1;		/* SHPC owned by shpchp */
 	unsigned int	is_thunderbolt:1;	/* Thunderbolt controller */
+	unsigned int	sizing_variant:2;	/* normal, minimum, no hotplug */
 	/*
 	 * Devices marked being untrusted are the ones that can potentially
 	 * execute DMA attacks and similar. They are typically connected
@@ -1299,6 +1300,7 @@ void pci_assign_unassigned_resources(void);
 void pci_assign_unassigned_bridge_resources(struct pci_dev *bridge);
 void pci_assign_unassigned_bus_resources(struct pci_bus *bus);
 void pci_assign_unassigned_root_bus_resources(struct pci_bus *bus);
+void pci_bus_assign_resources_fallback_sizing(struct pci_bus *bus);
 int pci_reassign_bridge_resources(struct pci_dev *bridge, unsigned long type);
 void pdev_enable_device(struct pci_dev *);
 int pci_enable_resources(struct pci_dev *, int mask);
Signature · 1 line
2.18.1
#6 lore raw

PATCH 3/3 PCI: vmd: Wire up VMD for fallback resource assignment

In reply to Jon Derrick

The VMD subdevice domain would prefer all devices be assigned resources and working rather than a few or none assigned, but with valid hotplug bridge resources. The resource assignment fallback algorithm works best for these requirements.

---
 drivers/pci/controller/vmd.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
drivers/pci/controller/vmd.c +1 -1
index fdc1a206f73e..4debc547c813 100644
@@ -777,7 +777,7 @@ static int vmd_enable_domain(struct vmd_dev *vmd, unsigned long features)
 		dev_set_msi_domain(&vmd->bus->dev, vmd->irq_domain);
 
 	pci_scan_child_bus(vmd->bus);
-	pci_assign_unassigned_bus_resources(vmd->bus);
+	pci_bus_assign_resources_fallback_sizing(vmd->bus);
 
 	/*
 	 * VMD root buses are virtual and don't return true on pci_is_pcie()
Signature · 1 line
2.18.1
#3 lore raw

Re: [PATCH 2/3] PCI: Introduce a minimizing assignment algorithm

In reply to Jon Derrick

Please keep this code in VMD if we really have to do it (although I'd be perfectly fine to let people dumb enough to enable VMD devices to live with the problems). You are adding lots of code that gets copiled into every Linux kernel that supports PCI jut to work around a copletely idiotic invention from Intel that makes life painful for us for no good reason.

#4 lore raw

Re: [PATCH 2/3] PCI: Introduce a minimizing assignment algorithm

Hi Christoph,

Thanks for your valuable feedback as always

Quoted 3 lines · Christoph Hellwig

Please keep this code in VMD if we really have to do it (although I'd be perfectly fine to let people dumb enough to enable VMD devices to live with the problems).

Great! Sounds like you're more open to us working openly within vmd.c then?

Quoted 4 lines

You are adding lots of code that gets
copiled into every Linux kernel that supports PCI jut to work around a copletely idiotic invention from Intel that makes life painful for us for no good reason.

Well this fix in particular may not be needed once the dynamic hotplug resource resizing set is in and build on that. But frankly the generic resource assignment code itself is very difficult to work within and has been discussed at several LPC over the years. I don't see a problem with another algorithm which could be relied upon by other host bridge controller drivers if they want it.

I also spent a good deal of time trying to get the minimizing algorithm into pci_assign_unassigned_root_bus_resources, where the only instance of pci=realloc detection takes place (who knew there were so many originating different paths for resource assignment?). I couldn't make headway there so started fresh. Maybe someone talented could refactor mine into it and save a few instruction bytes.

#5 lore raw

Re: [PATCH 2/3] PCI: Introduce a minimizing assignment algorithm

Quoted 7 lines · Jonathan

Well this fix in particular may not be needed once the dynamic hotplug resource resizing set is in and build on that. But frankly the generic resource assignment code itself is very difficult to work within and has been discussed at several LPC over the years. I don't see a problem with another algorithm which could be relied upon by other host bridge controller drivers if they want it.

Quoted 6 lines

I also spent a good deal of time trying to get the minimizing algorithm into pci_assign_unassigned_root_bus_resources, where the only instance of pci=realloc detection takes place (who knew there were so many originating different paths for resource assignment?). I couldn't make headway there so started fresh. Maybe someone talented could refactor mine into it and save a few instruction bytes.

If the maintainers think there might be other use cases we could also just make it conditional and let VMD select it. I'm just a little worried but all kinds of cruft slipping into core code to work around the various problems vmd creates.