Skip to contentSkip to conversation

x86/smpboot: Cleanup logical package ID

PATCH Superseded Prarit Bhargava 1 participant 4 messages Sep 14, 2017

Subsystems (2)

v2 of 6 Superseded Latest is v6 →

lore
Share & tools
lore
#0 lore raw

v2 cover x86/smpboot: Cleanup logical package ID

Cleanup the logical package ID code by storing the logical package ID in the cpuinfo_x86 struct and calculating the maximum logical package ID after all the CPUs have been enumerated.

[v2]: Decrease logical_packages when the last thread in a socket is removed.

Andi Kleen (2):
perf/x86/intel/uncore: Cache logical pkg id in uncore driver x86/topology: Avoid wasting 128k for package id array

Prarit Bhargava (1):
x86/smpboot: Fix __max_logical_packages estimate

 arch/x86/events/intel/uncore.c       |   1 +
 arch/x86/events/intel/uncore.h       |   1 +
 arch/x86/events/intel/uncore_snbep.c |   2 +-
 arch/x86/include/asm/processor.h     |   6 +-
 arch/x86/kernel/smpboot.c            | 113 +++++++++++------------------------
 5 files changed, 42 insertions(+), 81 deletions(-)
Signature · 1 line
1.8.5.5
#1 lore raw

PATCH v2 1/3 perf/x86/intel/uncore: Cache logical pkg id in uncore driver

From: Andi Kleen <[email protected]>

The SNB-EP uncore driver is the only user of topology_phys_to_logical_pkg in a performance critical path. Change it query the logical pkg ID only once at initialization time and then cache it in box structure.

---
 arch/x86/events/intel/uncore.c       | 1 +
 arch/x86/events/intel/uncore.h       | 1 +
 arch/x86/events/intel/uncore_snbep.c | 2 +-
 3 files changed, 3 insertions(+), 1 deletion(-)
arch/x86/events/intel/uncore.c +1 -0
index 1c5390f1cf09..d13ac98e922b 100644
@@ -939,6 +939,7 @@ static int uncore_pci_probe(struct pci_dev *pdev, const struct pci_device_id *id
 
 	atomic_inc(&box->refcnt);
 	box->pci_phys_id = phys_id;
+	box->logical_pkg_id = topology_phys_to_logical_pkg(box->pci_phys_id);
 	box->pkgid = pkg;
 	box->pci_dev = pdev;
 	box->pmu = pmu;
arch/x86/events/intel/uncore.h +1 -0
index df5989f27b1b..6214cd33a9fb 100644
@@ -99,6 +99,7 @@ struct intel_uncore_extra_reg {
 
 struct intel_uncore_box {
 	int pci_phys_id;
+	int logical_pkg_id;
 	int pkgid;
 	int n_active;	/* number of active events */
 	int n_events;
arch/x86/events/intel/uncore_snbep.c +1 -1
index db1fe377e6dd..11dec9f94cb0 100644
@@ -1056,7 +1056,7 @@ static void snbep_qpi_enable_event(struct intel_uncore_box *box, struct perf_eve
 
 	if (reg1->idx != EXTRA_REG_NONE) {
 		int idx = box->pmu->pmu_idx + SNBEP_PCI_QPI_PORT0_FILTER;
-		int pkg = topology_phys_to_logical_pkg(box->pci_phys_id);
+		int pkg = box->logical_pkg_id;
 		struct pci_dev *filter_pdev = uncore_extra_pci_dev[pkg].dev[idx];
 
 		if (filter_pdev) {
Signature · 1 line
1.8.5.5
#2 lore raw

PATCH v2 2/3 x86/topology: Avoid wasting 128k for package id array

In reply to Prarit Bhargava

From: Andi Kleen <[email protected]>

I was looking at large early boot allocations and noticed that since (1f12e32f x86/topology: Create logical package id) every 64bit system allocates a 128k array to convert logical package ids.

This happens because the array is sized for MAX_LOCAL_APIC and that is always 32k on 64bit systems, and it needs 4 bytes for each entry.

This is fairly wasteful, especially for the common case of having only one socket, where we need 128K just to store a single 4 byte value.

The max logical APIC value is not known at this point, so it's hard to size it correctly.

The previous patch converted the only performance critical user to cache the value, and all others are fairly slow path, so we can just convert the O(1) array lookup to a linear search in cpu_data()

This can also avoid the need for an extra bitmap structure to know if the logical package ID is already allocated. We can also save this information in cpu_data and look it up during the search.

This patch removes the explicit arrays and replaces the lookups with explicit searches.

Overall the new code is somewhat simpler, and needs a lot less run time memory.

The naming of the variables in cpu_data is still not great (_proc sometimes means package and sometimes means logical processor), but I followed the existing (messy) conventions when possible. At some point would be probably good to clean this up too.

Tested on a 2S system, but it would be good
to test on more obscure systems which may have problems with package IDs. I'm copying Prarit who had problematic systems before.

[v2]: Decrease logical_packages when the last thread in a socket is removed.

---
 arch/x86/include/asm/processor.h |  6 +++-
 arch/x86/kernel/smpboot.c        | 62 +++++++++++++++++-----------------------
 2 files changed, 32 insertions(+), 36 deletions(-)
arch/x86/include/asm/processor.h +5 -1
index 3fa26a61eabc..d369d2a82d8f 100644
@@ -124,13 +124,17 @@ struct cpuinfo_x86 {
 	u16			booted_cores;
 	/* Physical processor id: */
 	u16			phys_proc_id;
-	/* Logical processor id: */
+	/* Logical processor (package) id: */
 	u16			logical_proc_id;
+	/* Physical package ID */
+	u16			phys_pkg_id;
 	/* Core id: */
 	u16			cpu_core_id;
 	/* Index into per_cpu list: */
 	u16			cpu_index;
 	u32			microcode;
+	/* Flags */
+	unsigned		logical_proc_set : 1;
 } __randomize_layout;
 
 struct cpuid_regs {
arch/x86/kernel/smpboot.c +27 -35

117 lines not on this page · raw message

Signature · 1 line
1.8.5.5
#3 lore raw

PATCH v2 3/3 x86/smpboot: Fix __max_logical_packages estimate

In reply to Prarit Bhargava

A system booted with a small number of cores enabled per package panics because the estimate of __max_logical_packages is too low. This occurs when the total number of active cores across all packages is less than the maximum core count for a single package.

ie) On a 4 package system with 20 cores/package where only 4 cores are enabled on each package, the value of __max_logical_packages is calculated as DIV_ROUND_UP(16 / 20) = 1 and not 4.

Calculate __max_logical_packages after the cpu enumeration has completed. Use the boot cpu's data to extrapolate the number of packages.

---
 arch/x86/kernel/smpboot.c | 57 +++++++++--------------------------------------
 1 file changed, 10 insertions(+), 47 deletions(-)
arch/x86/kernel/smpboot.c +10 -47
index a6e9bd971b91..1b86ba57ab1a 100644
@@ -288,12 +288,6 @@ int topology_update_package_map(unsigned int pkg, unsigned int cpu)
 	if (new >= 0)
 		goto found;
 
-	if (logical_packages >= __max_logical_packages) {
-		pr_warn("Package %u of CPU %u exceeds BIOS package data %u.\n",
-			logical_packages, cpu, __max_logical_packages);
-		return -ENOSPC;
-	}
-
 	new = logical_packages++;
 	if (new != pkg)
 		pr_info("CPU %u Converting physical %u to logical package %u\n",
@@ -325,46 +319,6 @@ int topology_phys_to_logical_pkg(unsigned int phys_pkg)
 }
 EXPORT_SYMBOL(topology_phys_to_logical_pkg);
 
-static void __init smp_init_package_map(struct cpuinfo_x86 *c, unsigned int cpu)
-{
-	unsigned int ncpus;
-
-	/*
-	 * Today neither Intel nor AMD support heterogenous systems. That
-	 * might change in the future....
-	 *
-	 * While ideally we'd want '* smp_num_siblings' in the below @ncpus
-	 * computation, this won't actually work since some Intel BIOSes
-	 * report inconsistent HT data when they disable HT.
-	 *
-	 * In particular, they reduce the APIC-IDs to only include the cores,
-	 * but leave the CPUID topology to say there are (2) siblings.
-	 * This means we don't know how many threads there will be until
-	 * after the APIC enumeration.
-	 *
-	 * By not including this we'll sometimes over-estimate the number of
-	 * logical packages by the amount of !present siblings, but this is
-	 * still better than MAX_LOCAL_APIC.
-	 *
-	 * We use total_cpus not nr_cpu_ids because nr_cpu_ids can be limited
-	 * on the command line leading to a similar issue as the HT disable
-	 * problem because the hyperthreads are usually enumerated after the
-	 * primary cores.
-	 */
-	ncpus = boot_cpu_data.x86_max_cores;
-	if (!ncpus) {
-		pr_warn("x86_max_cores == zero !?!?");
-		ncpus = 1;
-	}
-
-	__max_logical_packages = DIV_ROUND_UP(total_cpus, ncpus);
-	pr_info("Max logical packages: %u\n", __max_logical_packages);
-
-	logical_packages = 0;
-
-	topology_update_package_map(c->phys_proc_id, cpu);
-}
-
 void __init smp_store_boot_cpu_info(void)
 {
 	int id = 0; /* CPU 0 */
@@ -372,7 +326,7 @@ void __init smp_store_boot_cpu_info(void)
 
 	*c = boot_cpu_data;
 	c->cpu_index = id;
-	smp_init_package_map(c, id);
+	topology_update_package_map(c->phys_proc_id, id);
 }
 
 /*
@@ -1372,7 +1326,16 @@ void __init native_smp_prepare_boot_cpu(void)
 
 void __init native_smp_cpus_done(unsigned int max_cpus)
 {
+	int ncpus;
+
 	pr_debug("Boot done\n");
+	/*
+	 * Today neither Intel nor AMD support heterogenous systems so
+	 * extrapolate the boot cpu's data to all packages.
+	 */
+	ncpus = cpu_data(0).booted_cores * smp_num_siblings;
+	__max_logical_packages = DIV_ROUND_UP(nr_cpu_ids, ncpus);
+	pr_info("Max logical packages: %u\n", __max_logical_packages);
 
 	if (x86_has_numa_in_package)
 		set_sched_topology(x86_numa_in_package_topology);
Signature · 1 line
1.8.5.5