x86/smpboot: Fix panic in __max_logical_packages estimate
Subsystems (2)
v3 cover x86/smpboot: Fix panic in __max_logical_packages estimate
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.
Andi noted that the logical_packages calculation wastes 128k on 64-bit x86 systems and posted patches to fix the allocation. I have put my fix on top of those patches.
Here's an example of the panic:
smpboot: Booting Node 1, Processors #1 OK
smpboot: Package 1 of CPU 1 exceeds BIOS package data 1.
------------[ cut here ]------------
kernel BUG at arch/x86/kernel/cpu/common.c:1087!
invalid opcode: 0000 [#1] SMP
Modules linked in:
CPU: 1 PID: 0 Comm: swapper/1 Not tainted 4.14.0-rc2+ #4This was tested by dropping the number of active cores across all packages to reproduce the bug above. Additional testing included 2 socket and 4 socket systems and hotplugging entire sockets in different order.
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 | 124 +++++++++++++---------------------- 5 files changed, 53 insertions(+), 81 deletions(-)
Signature · 1 line
2.15.0.rc0.39.g2f0e14e64
PATCH v3 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
@@ -947,6 +947,7 @@ static int uncore_pci_probe(struct pci_dev *pdev, const struct pci_device_id *idatomic_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
@@ -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
@@ -1056,7 +1056,7 @@ static void snbep_qpi_enable_event(struct intel_uncore_box *box, struct perf_eveif (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
2.15.0.rc0.39.g2f0e14e64
PATCH v3 2/3 x86/topology: Avoid wasting 128k for package id array
In reply to Prarit BhargavaFrom: 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.
[v3]: Add more logic to keep logical and physical package IDs in synch.
--- arch/x86/include/asm/processor.h | 6 +++- arch/x86/kernel/smpboot.c | 77 +++++++++++++++++++++------------------- 2 files changed, 45 insertions(+), 38 deletions(-)
arch/x86/include/asm/processor.h +5 -1
@@ -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 +40 -37
131 lines not on this page · raw message
Signature · 1 line
2.15.0.rc0.39.g2f0e14e64
Re: [PATCH 2/3 v3] x86/topology: Avoid wasting 128k for package id array
Quoted 8 lines · Prarit Bhargava
static void
remove_siblinginfo(int cpu)
{arch/x86/kernel/smpboot.c +1
-1-int sibling; +int phys_pkg_id, sibling;struct cpuinfo_x86 *c = &cpu_data (cpu);for_each_cpu (sibling,topology_core_cpumask (cpu)) {
arch/x86/kernel/smpboot.c static void remove_siblinginfo(int cpu) · line 1531
Quoted 6 lines
arch/x86/kernel/smpboot.c +3
-0@@ -1529,6 +1526,12 @@ static void remove_siblinginfo(int cpu)cpumask_clear (topology_core_cpumask (cpu)); c->phys_proc_id =0 ; c->cpu_core_id =0 ; + + phys_pkg_id = c->phys_pkg_id; + c->phys_pkg_id =U16_MAX ;
This leaves c->logical_proc_set = 1, which is inconsistent at best. I have no idea why we need this logical_proc_set flag at all.
arch/x86/kernel/smpboot.c static void remove_siblinginfo(int cpu) · line 1533
Quoted 2 lines
arch/x86/kernel/smpboot.c +2
-0+if (topology_phys_to_logical_pkg (phys_pkg_id) <0 ) + logical_packages--;
Now this has another issue. Depending on hotplug ordering the logical package association can change across hotplug operations. I don't know it that's an issue, but this needs to be analyzed before we merge that.
Thanks,
tglx
Re: [PATCH 2/3 v3] x86/topology: Avoid wasting 128k for package id array
Quoted 19 lines · Thomas Gleixner
Quoted 8 lines · Prarit Bhargava
static void
remove_siblinginfo(int cpu)
{arch/x86/kernel/smpboot.c +1
-1-int sibling; +int phys_pkg_id, sibling;struct cpuinfo_x86 *c = &cpu_data (cpu);for_each_cpu (sibling,topology_core_cpumask (cpu)) {arch/x86/kernel/smpboot.c static void remove_siblinginfo(int cpu) · line 1531
Quoted 6 lines
arch/x86/kernel/smpboot.c +3
-0@@ -1529,6 +1526,12 @@ static void remove_siblinginfo(int cpu)cpumask_clear (topology_core_cpumask (cpu)); c->phys_proc_id =0 ; c->cpu_core_id =0 ; + + phys_pkg_id = c->phys_pkg_id; + c->phys_pkg_id =U16_MAX ;This leaves
c->logical_proc_set= 1, which is inconsistent at best. I have no idea why we need thislogical_proc_setflag at all.
Quoted 7 lines
arch/x86/kernel/smpboot.c static void remove_siblinginfo(int cpu) · line 1533
Quoted 2 lines
arch/x86/kernel/smpboot.c +2
-0+if (topology_phys_to_logical_pkg (phys_pkg_id) <0 ) + logical_packages--;Now this has another issue. Depending on hotplug ordering the logical package association can change across hotplug operations. I don't know it that's an issue, but this needs to be analyzed before we merge that.
Thanks for making me look at this Thomas.
Andi, it looks like this is unfortunately an issue. I have a reworked patchset that fixes by dynamically allocating a u16 logical_to_physical_package array that
a) maps the logical to physical package IDs,
b) is dynamically sized to the value of logical_packages, and,
c) is of size logical_packages (not MAX_APICS)
This will resolve the problem that Thomas has pointed out, and it would address the issue of wasting memory.
I'm waiting for a 4S & 8S system to test the LTP hotplug tests on and then I'll post. Andi, I'll send it to you in private email for a quick review.
P.
Quoted 5 lines
Thanks,
tglx
PATCH v3 3/3 x86/smpboot: Fix __max_logical_packages estimate
In reply to Prarit BhargavaA 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.
Here's an example of the panic:
smpboot: Booting Node 1, Processors #1 OK
smpboot: Package 1 of CPU 1 exceeds BIOS package data 1.
------------[ cut here ]------------
kernel BUG at arch/x86/kernel/cpu/common.c:1087!
invalid opcode: 0000 [#1] SMP
Modules linked in:
CPU: 1 PID: 0 Comm: swapper/1 Not tainted 4.14.0-rc2+ #4Calculate __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
@@ -289,12 +289,6 @@ 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 ; - } - logical_packages++;/* Try to keep 1:1 physical to logical mapping */
@@ -338,46 +332,6 @@ 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 */
@@ -385,7 +339,7 @@ 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); }/*
@@ -1385,7 +1339,16 @@ 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
2.15.0.rc0.39.g2f0e14e64