Messages in this thread | | | Date | Mon, 02 Jun 2014 13:20:09 -0400 | From | Prarit Bhargava <> | Subject | Re: [PATCH] x86, Clean up smp_num_siblings calculation |
| |
On 06/02/2014 12:30 PM, Paul Gortmaker wrote: > On 14-06-02 07:51 AM, Prarit Bhargava wrote: >> I have a system on which I have disabled threading in the BIOS, and I am booting >> the kernel with the option "idle=poll". >> >> The kernel displays >> >> process: WARNING: polling idle and HT enabled, performance may degrade >> >> which is incorrect -- I've already disabled HT. >> >> This warning is issued here: >> >> void select_idle_routine(const struct cpuinfo_x86 *c) >> { >> if (boot_option_idle_override == IDLE_POLL && smp_num_siblings > 1) >> pr_warn_once("WARNING: polling idle and HT enabled, performance may degrade\n"); >> >> From my understanding of the other areas of kernel that use smp_num_siblings, >> the value is supposed to be the actual number of threads per core, and >> this value of smp_num_siblings is incorrect. In theory, it should be 1 but it >> is reported as 2. When I looked into how smp_num_siblings is calculated I >> found the following call sequence in the kernel: >> >> start_kernel -> >> check_bugs -> >> identify_boot_cpu -> >> identify_cpu -> >> c_init = init_intel >> init_intel -> >> detect_extended_topology >> (sets value) >> >> OR >> >> c_init = init_amd >> init_amd -> amd_detect_cmp >> -> amd_get_topology >> (sets value) >> -> detect_ht() >> ... (sets value) >> detect_ht() >> (also sets value) >> >> ie) it is set three times in some cases and overwritten in other cases. >> >> It should be noted that nothing in the identify_cpu() path or the cpu_up() >> path requires smp_num_siblings to be set, prior to the final call to >> detect_ht(). >> >> For x86 boxes without X86_FEATURE_XTOPOLOGY, smp_num_siblings is set to a >> value read in a CPUID call in detect_ht(). This value is the *factory >> defined* value in all cases; even if HT is disabled in BIOS the value >> still returns 2 if the CPU supports HT. AMD also reports the factory >> defined value in all cases. >> >> For Intel x86 boxes with X86_FEATURE_XTOPOLOGY, smp_num_siblings is set to a >> value read from the 0xb leaf of CPUID. This value is also the *factory >> defined* value in all cases. >> >> For new-ish AMD x86 boxes, smp_num_siblings is also set to the *factory* >> defined value. >> >> That is, even with threading disabled in BIOSes on these systems, >> >> crash> p smp_num_siblings >> smp_num_siblings = $1 = 0x2 >> >> smp_num_siblings should be calculated a single time on cpu 0 to determine >> whether or not the system is multi-threaded or not. We can easily do >> this by examining the boot cpu's cpu_sibling_mask after the mask has been >> setup in the boot up code path. >> >> After the patch, on a system with HT enabled, >> >> crash> p smp_num_siblings >> smp_num_siblings = $1 = 0x2 >> >> On a system with HT disabled, >> >> crash> p smp_num_siblings >> smp_num_siblings = $1 = 0x1 >> >> Other uses of smp_num_siblings involve oprofile (used after boot), and >> the perf code which is done well after the initial cpus are brought online. >> >> [v2]: After comment from Oren Twaig, rework to single patch. >> Unfortunately there was no easy way to take into account the various >> settings of smp_num_siblings and fix it in two patches. >> >> Cc: Oren Twaig <oren@scalemp.com> >> Cc: Thomas Gleixner <tglx@linutronix.de> >> Cc: Ingo Molnar <mingo@redhat.com> >> Cc: "H. Peter Anvin" <hpa@zytor.com> >> Cc: x86@kernel.org >> Cc: Borislav Petkov <bp@suse.de> >> Cc: Paul Gortmaker <paul.gortmaker@windriver.com> >> Cc: Andrew Morton <akpm@linux-foundation.org> >> Cc: Andi Kleen <ak@linux.intel.com> >> Cc: Dave Jones <davej@redhat.com> >> Cc: Torsten Kaiser <just.for.lkml@googlemail.com> >> Cc: Jan Beulich <JBeulich@suse.com> >> Cc: Jan Kiszka <jan.kiszka@siemens.com> >> Cc: Toshi Kani <toshi.kani@hp.com> >> Cc: Andrew Jones <drjones@redhat.com> >> --- >> arch/x86/kernel/cpu/amd.c | 1 - >> arch/x86/kernel/cpu/common.c | 23 +++++++++++------------ >> arch/x86/kernel/cpu/topology.c | 2 +- >> arch/x86/kernel/smpboot.c | 10 +++++++--- >> 4 files changed, 19 insertions(+), 17 deletions(-) >> >> diff --git a/arch/x86/kernel/cpu/amd.c b/arch/x86/kernel/cpu/amd.c >> index ce8b8ff..6aca2b6 100644 >> --- a/arch/x86/kernel/cpu/amd.c >> +++ b/arch/x86/kernel/cpu/amd.c >> @@ -304,7 +304,6 @@ static void amd_get_topology(struct cpuinfo_x86 *c) >> node_id = ecx & 7; >> >> /* get compute unit information */ >> - smp_num_siblings = ((ebx >> 8) & 3) + 1; >> c->compute_unit_id = ebx & 0xff; >> cores_per_cu += ((ebx >> 8) & 3); >> } else if (cpu_has(c, X86_FEATURE_NODEID_MSR)) { >> diff --git a/arch/x86/kernel/cpu/common.c b/arch/x86/kernel/cpu/common.c >> index a135239..81a5aac 100644 >> --- a/arch/x86/kernel/cpu/common.c >> +++ b/arch/x86/kernel/cpu/common.c >> @@ -507,42 +507,41 @@ void detect_ht(struct cpuinfo_x86 *c) >> u32 eax, ebx, ecx, edx; >> int index_msb, core_bits; >> static bool printed; >> + int threads_per_core; >> >> if (!cpu_has(c, X86_FEATURE_HT)) >> return; >> >> - if (cpu_has(c, X86_FEATURE_CMP_LEGACY)) >> + if (cpu_has(c, X86_FEATURE_CMP_LEGACY)) { >> + threads_per_core = 1; >> goto out; >> + } >> >> if (cpu_has(c, X86_FEATURE_XTOPOLOGY)) >> return; >> >> cpuid(1, &eax, &ebx, &ecx, &edx); >> >> - smp_num_siblings = (ebx & 0xff0000) >> 16; >> + threads_per_core = (ebx & 0xff0000) >> 16; > > I wonder if this code is in need of an update? I recall reading > this thread: > > http://forum.osdev.org/viewtopic.php?f=1&t=23445 > > which suggests that we try CPUID with 0xb, and then 0x4 _before_ > relying on the EBX[23:16] of the older CPUID 0x1. > > AFAICT, the 0xb and 0x4 didn't exist when AP-485 was written ~2002.
I think the first case (0xb leaf) is done when cpu_has(c, X86_FEATURE_XTOPOLOGY) is true. I don't think we've been doing the latter though and it could be something introduced in a new patch?
> > http://datasheets.chipdb.org/Intel/x86/CPUID/24161821.pdf > > Also, there was a discussion of masking the "ht" flag in /proc/cpuinfo > for when it is "off" -- since the common sense interpretation of it > doesn't match the implementation in the specification: > > http://codemonkey.org.uk/2009/11/10/common-hyperthreading-misconception/ > https://lkml.org/lkml/2009/11/13/33
Yeah -- I was actually debating about masking it off when smp_num_siblings == 1, but wasn't sure how we felt about that these days. The problem with mucking around with /proc/cpuinfo is that we're never clear if the values are the values read from the hardware, or the interpreted software values.
I can certainly retest the ht flag masking if one of the x86@kernel.org people give me an ack to do so.
hpa? tglx? Ingo?
P.
| |