Skip to contentSkip to conversation

PCI: Add mechanism to find topologically near cores

PATCH In review Daniel J Blueman 3 participants 5 messages Dec 23, 2015

Subsystems (4)
lore
Share & tools
lore
#0 lore raw

PATCH 1/2 PCI: Add mechanism to find topologically near cores

Some devices (eg ixgbe) make assumptions about device to core locality when specifying interrupts locality hints and allocate starting from core 0. Moreover, interrupts may not be routable to distant NUMA nodes due to the 8-bit APIC ID space limitations.

Provide a mechanism drivers can use to find cores with reasonable locality to a device; use the existing precendent of RECLAIM_DISTANCE (30), wrapping the offset.

---
 drivers/pci/pci.c   | 15 +++++++++++++++
 include/linux/pci.h |  1 +
 2 files changed, 16 insertions(+)
drivers/pci/pci.c +16 -0
index 314db8c..d5535d1 100644
@@ -4833,6 +4833,22 @@ void __weak pci_fixup_cardbus(struct pci_bus *bus)
 }
 EXPORT_SYMBOL(pci_fixup_cardbus);
 
+int cpu_near_dev(const struct pci_dev *pdev, unsigned offset)
+{
+	/* Start search from node device is on for optimal locality */
+	int localnode = pcibus_to_node(pdev->bus);
+	int cpu = cpumask_first(cpumask_of_node(localnode));
+
+	while (offset--) {
+		do {
+			cpu = (cpu + 1) % nr_cpu_ids;
+		} while (!cpu_online(cpu) || node_distance(cpu_to_node(cpu),
+			localnode) > RECLAIM_DISTANCE);
+	}
+
+	return cpu;
+}
+
 static int __init pci_setup(char *str)
 {
 	while (str) {
include/linux/pci.h +1 -0
index 6ae25aa..f7491bd 100644
@@ -842,6 +842,7 @@ void pci_stop_root_bus(struct pci_bus *bus);
 void pci_remove_root_bus(struct pci_bus *bus);
 void pci_setup_cardbus(struct pci_bus *bus);
 void pci_sort_breadthfirst(void);
+int cpu_near_dev(const struct pci_dev *pdev, unsigned offset);
 #define dev_is_pci(d) ((d)->bus == &pci_bus_type)
 #define dev_is_pf(d) ((dev_is_pci(d) ? to_pci_dev(d)->is_physfn : false))
 #define dev_num_vf(d) ((dev_is_pci(d) ? pci_num_vf(to_pci_dev(d)) : 0))

--
2.5.0

#1 lore raw

PATCH 2/2 ixgbe: Use core to device locality interface

Rather than assuming cores starting from 0 are local to the ethernet device, use the introduced interface to find near cores.

Not only does this improve performance due to spreading interrupts via near NUMA nodes, it prevents assigning cores on distant NUMA nodes, which aren't reachable by device interrupts due to the 8-bit APIC ID limitation.

With Numascale NumaConnect2 systems with Intel ixgbe cards on non-primary PCI domains, all ixgbe NICs would previously revector interrupts to cores 0 to 63 (cores 0 to 47 would be considered near the primary PCI domain). Now, cores 48 to 95 are used, increasing performance and addressing interrupt delivery issues:

do_IRQ: 79.180 No irq handler for vector (irq -1)
do_IRQ: 78.42 No irq handler for vector (irq -1)
do_IRQ: 71.172 No irq handler for vector (irq -1)
do_IRQ: 70.236 No irq handler for vector (irq -1)
do_IRQ: 69.109 No irq handler for vector (irq -1)
do_IRQ: 68.189 No irq handler for vector (irq -1)
do_IRQ: 72.92 No irq handler for vector (irq -1)
do_IRQ: 73.235 No irq handler for vector (irq -1)
do_IRQ: 66.185 No irq handler for vector (irq -1)
do_IRQ: 67.62 No irq handler for vector (irq -1)
do_IRQ: 197 callbacks suppressed

---
 drivers/net/ethernet/intel/ixgbe/ixgbe_lib.c | 6 ++----
 1 file changed, 2 insertions(+), 4 deletions(-)
drivers/net/ethernet/intel/ixgbe/ixgbe_lib.c +2 -4
index f3168bc..12c4ce1 100644
@@ -817,10 +817,8 @@ static int ixgbe_alloc_q_vector(struct ixgbe_adapter *adapter,
 	if ((tcs <= 1) && !(adapter->flags & IXGBE_FLAG_SRIOV_ENABLED)) {
 		u16 rss_i = adapter->ring_feature[RING_F_RSS].indices;
 		if (rss_i > 1 && adapter->atr_sample_rate) {
-			if (cpu_online(v_idx)) {
-				cpu = v_idx;
-				node = cpu_to_node(cpu);
-			}
+			cpu = cpu_near_dev(adapter->pdev, v_idx);
+			node = cpu_to_node(cpu);
 		}
 	}
 

--
2.5.0

#2 lore raw

Re: [PATCH 2/2] ixgbe: Use core to device locality interface

reproduce:

Hi Daniel,

[auto build test ERROR on pci/next]
[also build test ERROR on v4.4-rc6 next-20151223]

url:    https://github.com/0day-ci/linux/commits/Daniel-J-Blueman/PCI-Add-mechanism-to-find-topologically-near-cores/20151223-181947
base:   https://git.kernel.org/pub/scm/linux/kernel/git/helgaas/pci.git next
config: i386-randconfig-a0-12222034 (attached as .config)
reproduce:
        # save the attached .config to linux build tree
        make ARCH=i386 

All errors (new ones prefixed by >>):

>> ERROR: "cpu_near_dev" [drivers/net/ethernet/intel/ixgbe/ixgbe.ko] undefined!

---
0-DAY kernel test infrastructure Open Source Technology Center
https://lists.01.org/pipermail/kbuild-all Intel Corporation
-------------- next part --------------
A non-text attachment was scrubbed...
Name: .config.gz
Type: application/octet-stream
Size: 25593 bytes
Desc: not available
URL: <http://lists.osuosl.org/pipermail/intel-wired-lan/attachments/20151223/47038d47/attachment-0001.obj>

#3 lore raw
In reply to Daniel J Blueman
reproduce: wget https://git.kernel.org/cgit/linux/kernel/git/wfg/lkp-tests.git/plain/sbin/make.cross -O ~/bin/make.cross chmod +x ~/bin/make.cross # save the attached .config to linux build tree…

Hi Daniel,

[auto build test WARNING on pci/next]
[also build test WARNING on v4.4-rc6 next-20151223]

url:    https://github.com/0day-ci/linux/commits/Daniel-J-Blueman/PCI-Add-mechanism-to-find-topologically-near-cores/20151223-181947
base:   https://git.kernel.org/pub/scm/linux/kernel/git/helgaas/pci.git next
config: s390-allyesconfig (attached as .config)

reproduce:
wget https://git.kernel.org/cgit/linux/kernel/git/wfg/lkp-tests.git/plain/sbin/make.cross -O ~/bin/make.cross chmod +x ~/bin/make.cross
# save the attached .config to linux build tree
make.cross ARCH=s390

Note: it may well be a FALSE warning. FWIW you are at least aware of it now.

All warnings (new ones prefixed by >>):

In file included from include/linux/smp.h:12:0,

                    from arch/s390/include/asm/spinlock.h:12,
                    from include/linux/spinlock.h:87,
                    from include/linux/rcupdate.h:38,
                    from include/linux/idr.h:18,
                    from include/linux/kernfs.h:14,
                    from include/linux/sysfs.h:15,
                    from include/linux/kobject.h:21,
                    from include/linux/of.h:21,
                    from drivers/pci/pci.c:13:
   drivers/pci/pci.c: In function 'cpu_near_dev':
>> include/linux/cpumask.h:174:9: warning: array subscript is below array bounds [-Warray-bounds]
     return find_first_bit(cpumask_bits(srcp), nr_cpumask_bits);
            ^

vim +174 include/linux/cpumask.h

da91309e Amir Vadai       2014-06-09  158  
2d3854a3 Rusty Russell    2008-11-05  159  #define for_each_cpu(cpu, mask)			\
2d3854a3 Rusty Russell    2008-11-05  160  	for ((cpu) = 0; (cpu) < 1; (cpu)++, (void)mask)
8bd93a2c Paul E. McKenney 2010-02-22  161  #define for_each_cpu_not(cpu, mask)		\
8bd93a2c Paul E. McKenney 2010-02-22  162  	for ((cpu) = 0; (cpu) < 1; (cpu)++, (void)mask)
2d3854a3 Rusty Russell    2008-11-05  163  #define for_each_cpu_and(cpu, mask, and)	\
2d3854a3 Rusty Russell    2008-11-05  164  	for ((cpu) = 0; (cpu) < 1; (cpu)++, (void)mask, (void)and)
2d3854a3 Rusty Russell    2008-11-05  165  #else
2d3854a3 Rusty Russell    2008-11-05  166  /**
2d3854a3 Rusty Russell    2008-11-05  167   * cpumask_first - get the first cpu in a cpumask
2d3854a3 Rusty Russell    2008-11-05  168   * @srcp: the cpumask pointer
2d3854a3 Rusty Russell    2008-11-05  169   *
2d3854a3 Rusty Russell    2008-11-05  170   * Returns >= nr_cpu_ids if no cpus set.
2d3854a3 Rusty Russell    2008-11-05  171   */
2d3854a3 Rusty Russell    2008-11-05  172  static inline unsigned int cpumask_first(const struct cpumask *srcp)
2d3854a3 Rusty Russell    2008-11-05  173  {
2d3854a3 Rusty Russell    2008-11-05 @174  	return find_first_bit(cpumask_bits(srcp), nr_cpumask_bits);
2d3854a3 Rusty Russell    2008-11-05  175  }
2d3854a3 Rusty Russell    2008-11-05  176  
2d3854a3 Rusty Russell    2008-11-05  177  /**
2d3854a3 Rusty Russell    2008-11-05  178   * cpumask_next - get the next cpu in a cpumask
2d3854a3 Rusty Russell    2008-11-05  179   * @n: the cpu prior to the place to search (ie. return will be > @n)
2d3854a3 Rusty Russell    2008-11-05  180   * @srcp: the cpumask pointer
2d3854a3 Rusty Russell    2008-11-05  181   *
2d3854a3 Rusty Russell    2008-11-05  182   * Returns >= nr_cpu_ids if no further cpus set.

:::::: The code at line 174 was first introduced by commit
:::::: 2d3854a37e8b767a51aba38ed6d22817b0631e33 cpumask: introduce new API, without changing anything

:::::: TO: Rusty Russell <[email protected]>
:::::: CC: Ingo Molnar <[email protected]>

---
0-DAY kernel test infrastructure Open Source Technology Center
https://lists.01.org/pipermail/kbuild-all Intel Corporation
-------------- next part --------------
A non-text attachment was scrubbed...
Name: .config.gz
Type: application/octet-stream
Size: 39398 bytes
Desc: not available
URL: <http://lists.osuosl.org/pipermail/intel-wired-lan/attachments/20151223/58a0bb3e/attachment-0001.obj>

#4 lore raw
In reply to Daniel J Blueman

Hi Daniel,

Quoted 4 lines · Daniel J Blueman

Some devices (eg ixgbe) make assumptions about device to core locality when specifying interrupts locality hints and allocate starting from core 0. Moreover, interrupts may not be routable to distant NUMA nodes due to the 8-bit APIC ID space limitations.

The APIC ID issue is the primary problem you're trying to solve, but this patch doesn't solve it directly because it doesn't look at anything related to the APIC ID domain. Anything you do here is a guess that might work better, but it still won't necessarily work in all cases.

Also, can you add a note about how this relates to the "call driver probe function on node where device is attached" functionality in pci_call_probe()? It seems like that should be enough to keep us from always allocating interrupts on core 0. Maybe that's broken, or maybe ixgbe isn't taking advantage of that?

Quoted 3 lines

Provide a mechanism drivers can use to find cores with reasonable locality to a device; use the existing precendent of RECLAIM_DISTANCE (30), wrapping the offset.

I don't think it's a benefit to reuse RECLAIM_DISTANCE, because that is a different concept that doesn't seem directly related to what you're doing. The name and the existing uses are related to memory zone reclaiming, so I know I would be confused to see it also used for IRQ assignment.

Quoted 11 lines
---
 drivers/pci/pci.c   | 15 +++++++++++++++
 include/linux/pci.h |  1 +
 2 files changed, 16 insertions(+)
drivers/pci/pci.c +0 -0
index 314db8c..d5535d1 100644

drivers/pci/pci.c void __weak pci_fixup_cardbus(struct pci_bus *bus) · line 4836

Quoted 4 lines
drivers/pci/pci.c +1 -0
@@ -4833,6 +4833,22 @@ void __weak pci_fixup_cardbus(struct pci_bus *bus)
 }
 EXPORT_SYMBOL(pci_fixup_cardbus);
 
+int cpu_near_dev(const struct pci_dev *pdev, unsigned offset)

If this becomes a PCI interface, please add a "pci_" prefix to the name.

The distance concept also applies to non-PCI devices, so ideally I think a "pci_nearby_cpu()" or similar interface would be a wrapper around a more generic interface that takes a "struct device *".

I don't really understand what the "offset" parameter is for. It looks like you're using it to spread across CPUs on a node, but that seems like something that should be done internally to this interface.

Quoted 37 lines
drivers/pci/pci.c +15 -0
+{
+	/* Start search from node device is on for optimal locality */
+	int localnode = pcibus_to_node(pdev->bus);
+	int cpu = cpumask_first(cpumask_of_node(localnode));
+
+	while (offset--) {
+		do {
+			cpu = (cpu + 1) % nr_cpu_ids;
+		} while (!cpu_online(cpu) || node_distance(cpu_to_node(cpu),
+			localnode) > RECLAIM_DISTANCE);
+	}
+
+	return cpu;
+}
+
 static int __init pci_setup(char *str)
 {
 	while (str) {
include/linux/pci.h +1 -1
index 6ae25aa..f7491bd 100644
@@ -842,6 +842,7 @@ void pci_stop_root_bus(struct pci_bus *bus);
 void pci_remove_root_bus(struct pci_bus *bus);
 void pci_setup_cardbus(struct pci_bus *bus);
 void pci_sort_breadthfirst(void);
+int cpu_near_dev(const struct pci_dev *pdev, unsigned offset);
 #define dev_is_pci(d) ((d)->bus == &pci_bus_type)
 #define dev_is_pf(d) ((dev_is_pci(d) ? to_pci_dev(d)->is_physfn : false))
 #define dev_num_vf(d) ((dev_is_pci(d) ? pci_num_vf(to_pci_dev(d)) : 0))
--

2.5.0

--
To unsubscribe from this list: send the line "unsubscribe linux-pci" in the body of a message to majordomo at vger.kernel.org
More majordomo info at http://vger.kernel.org/majordomo-info.html