linux/arch/x86/kernel/cpu/hygon.c
Borislav Petkov (AMD) 04c3024560 x86/barrier: Do not serialize MSR accesses on AMD
AMD does not have the requirement for a synchronization barrier when
acccessing a certain group of MSRs. Do not incur that unnecessary
penalty there.

There will be a CPUID bit which explicitly states that a MFENCE is not
needed. Once that bit is added to the APM, this will be extended with
it.

While at it, move to processor.h to avoid include hell. Untangling that
file properly is a matter for another day.

Some notes on the performance aspect of why this is relevant, courtesy
of Kishon VijayAbraham <Kishon.VijayAbraham@amd.com>:

On a AMD Zen4 system with 96 cores, a modified ipi-bench[1] on a VM
shows x2AVIC IPI rate is 3% to 4% lower than AVIC IPI rate. The
ipi-bench is modified so that the IPIs are sent between two vCPUs in the
same CCX. This also requires to pin the vCPU to a physical core to
prevent any latencies. This simulates the use case of pinning vCPUs to
the thread of a single CCX to avoid interrupt IPI latency.

In order to avoid run-to-run variance (for both x2AVIC and AVIC), the
below configurations are done:

  1) Disable Power States in BIOS (to prevent the system from going to
     lower power state)

  2) Run the system at fixed frequency 2500MHz (to prevent the system
     from increasing the frequency when the load is more)

With the above configuration:

*) Performance measured using ipi-bench for AVIC:
  Average Latency:  1124.98ns [Time to send IPI from one vCPU to another vCPU]

  Cumulative throughput: 42.6759M/s [Total number of IPIs sent in a second from
  				     48 vCPUs simultaneously]

*) Performance measured using ipi-bench for x2AVIC:
  Average Latency:  1172.42ns [Time to send IPI from one vCPU to another vCPU]

  Cumulative throughput: 40.9432M/s [Total number of IPIs sent in a second from
  				     48 vCPUs simultaneously]

From above, x2AVIC latency is ~4% more than AVIC. However, the expectation is
x2AVIC performance to be better or equivalent to AVIC. Upon analyzing
the perf captures, it is observed significant time is spent in
weak_wrmsr_fence() invoked by x2apic_send_IPI().

With the fix to skip weak_wrmsr_fence()

*) Performance measured using ipi-bench for x2AVIC:
  Average Latency:  1117.44ns [Time to send IPI from one vCPU to another vCPU]

  Cumulative throughput: 42.9608M/s [Total number of IPIs sent in a second from
  				     48 vCPUs simultaneously]

Comparing the performance of x2AVIC with and without the fix, it can be seen
the performance improves by ~4%.

Performance captured using an unmodified ipi-bench using the 'mesh-ipi' option
with and without weak_wrmsr_fence() on a Zen4 system also showed significant
performance improvement without weak_wrmsr_fence(). The 'mesh-ipi' option ignores
CCX or CCD and just picks random vCPU.

  Average throughput (10 iterations) with weak_wrmsr_fence(),
        Cumulative throughput: 4933374 IPI/s

  Average throughput (10 iterations) without weak_wrmsr_fence(),
        Cumulative throughput: 6355156 IPI/s

[1] https://github.com/bytedance/kvm-utils/tree/master/microbenchmark/ipi-bench

Signed-off-by: Borislav Petkov (AMD) <bp@alien8.de>
Link: https://lore.kernel.org/r/20230622095212.20940-1-bp@alien8.de
2023-11-13 10:09:45 +01:00

405 lines
10 KiB
C

// SPDX-License-Identifier: GPL-2.0+
/*
* Hygon Processor Support for Linux
*
* Copyright (C) 2018 Chengdu Haiguang IC Design Co., Ltd.
*
* Author: Pu Wen <puwen@hygon.cn>
*/
#include <linux/io.h>
#include <asm/apic.h>
#include <asm/cpu.h>
#include <asm/smp.h>
#include <asm/numa.h>
#include <asm/cacheinfo.h>
#include <asm/spec-ctrl.h>
#include <asm/delay.h>
#include "cpu.h"
#define APICID_SOCKET_ID_BIT 6
/*
* nodes_per_socket: Stores the number of nodes per socket.
* Refer to CPUID Fn8000_001E_ECX Node Identifiers[10:8]
*/
static u32 nodes_per_socket = 1;
#ifdef CONFIG_NUMA
/*
* To workaround broken NUMA config. Read the comment in
* srat_detect_node().
*/
static int nearby_node(int apicid)
{
int i, node;
for (i = apicid - 1; i >= 0; i--) {
node = __apicid_to_node[i];
if (node != NUMA_NO_NODE && node_online(node))
return node;
}
for (i = apicid + 1; i < MAX_LOCAL_APIC; i++) {
node = __apicid_to_node[i];
if (node != NUMA_NO_NODE && node_online(node))
return node;
}
return first_node(node_online_map); /* Shouldn't happen */
}
#endif
static void hygon_get_topology_early(struct cpuinfo_x86 *c)
{
if (cpu_has(c, X86_FEATURE_TOPOEXT))
smp_num_siblings = ((cpuid_ebx(0x8000001e) >> 8) & 0xff) + 1;
}
/*
* Fixup core topology information for
* (1) Hygon multi-node processors
* Assumption: Number of cores in each internal node is the same.
* (2) Hygon processors supporting compute units
*/
static void hygon_get_topology(struct cpuinfo_x86 *c)
{
/* get information required for multi-node processors */
if (boot_cpu_has(X86_FEATURE_TOPOEXT)) {
int err;
u32 eax, ebx, ecx, edx;
cpuid(0x8000001e, &eax, &ebx, &ecx, &edx);
c->topo.die_id = ecx & 0xff;
c->topo.core_id = ebx & 0xff;
if (smp_num_siblings > 1)
c->x86_max_cores /= smp_num_siblings;
/*
* In case leaf B is available, use it to derive
* topology information.
*/
err = detect_extended_topology(c);
if (!err)
c->x86_coreid_bits = get_count_order(c->x86_max_cores);
/*
* Socket ID is ApicId[6] for the processors with model <= 0x3
* when running on host.
*/
if (!boot_cpu_has(X86_FEATURE_HYPERVISOR) && c->x86_model <= 0x3)
c->topo.pkg_id = c->topo.apicid >> APICID_SOCKET_ID_BIT;
cacheinfo_hygon_init_llc_id(c);
} else if (cpu_has(c, X86_FEATURE_NODEID_MSR)) {
u64 value;
rdmsrl(MSR_FAM10H_NODE_ID, value);
c->topo.die_id = value & 7;
c->topo.llc_id = c->topo.die_id;
} else
return;
if (nodes_per_socket > 1)
set_cpu_cap(c, X86_FEATURE_AMD_DCM);
}
/*
* On Hygon setup the lower bits of the APIC id distinguish the cores.
* Assumes number of cores is a power of two.
*/
static void hygon_detect_cmp(struct cpuinfo_x86 *c)
{
unsigned int bits;
bits = c->x86_coreid_bits;
/* Low order bits define the core id (index of core in socket) */
c->topo.core_id = c->topo.initial_apicid & ((1 << bits)-1);
/* Convert the initial APIC ID into the socket ID */
c->topo.pkg_id = c->topo.initial_apicid >> bits;
/* Use package ID also for last level cache */
c->topo.llc_id = c->topo.die_id = c->topo.pkg_id;
}
static void srat_detect_node(struct cpuinfo_x86 *c)
{
#ifdef CONFIG_NUMA
int cpu = smp_processor_id();
int node;
unsigned int apicid = c->topo.apicid;
node = numa_cpu_node(cpu);
if (node == NUMA_NO_NODE)
node = c->topo.llc_id;
/*
* On multi-fabric platform (e.g. Numascale NumaChip) a
* platform-specific handler needs to be called to fixup some
* IDs of the CPU.
*/
if (x86_cpuinit.fixup_cpu_id)
x86_cpuinit.fixup_cpu_id(c, node);
if (!node_online(node)) {
/*
* Two possibilities here:
*
* - The CPU is missing memory and no node was created. In
* that case try picking one from a nearby CPU.
*
* - The APIC IDs differ from the HyperTransport node IDs.
* Assume they are all increased by a constant offset, but
* in the same order as the HT nodeids. If that doesn't
* result in a usable node fall back to the path for the
* previous case.
*
* This workaround operates directly on the mapping between
* APIC ID and NUMA node, assuming certain relationship
* between APIC ID, HT node ID and NUMA topology. As going
* through CPU mapping may alter the outcome, directly
* access __apicid_to_node[].
*/
int ht_nodeid = c->topo.initial_apicid;
if (__apicid_to_node[ht_nodeid] != NUMA_NO_NODE)
node = __apicid_to_node[ht_nodeid];
/* Pick a nearby node */
if (!node_online(node))
node = nearby_node(apicid);
}
numa_set_node(cpu, node);
#endif
}
static void early_init_hygon_mc(struct cpuinfo_x86 *c)
{
#ifdef CONFIG_SMP
unsigned int bits, ecx;
/* Multi core CPU? */
if (c->extended_cpuid_level < 0x80000008)
return;
ecx = cpuid_ecx(0x80000008);
c->x86_max_cores = (ecx & 0xff) + 1;
/* CPU telling us the core id bits shift? */
bits = (ecx >> 12) & 0xF;
/* Otherwise recompute */
if (bits == 0) {
while ((1 << bits) < c->x86_max_cores)
bits++;
}
c->x86_coreid_bits = bits;
#endif
}
static void bsp_init_hygon(struct cpuinfo_x86 *c)
{
if (cpu_has(c, X86_FEATURE_CONSTANT_TSC)) {
u64 val;
rdmsrl(MSR_K7_HWCR, val);
if (!(val & BIT(24)))
pr_warn(FW_BUG "TSC doesn't count with P0 frequency!\n");
}
if (cpu_has(c, X86_FEATURE_MWAITX))
use_mwaitx_delay();
if (boot_cpu_has(X86_FEATURE_TOPOEXT)) {
u32 ecx;
ecx = cpuid_ecx(0x8000001e);
__max_die_per_package = nodes_per_socket = ((ecx >> 8) & 7) + 1;
} else if (boot_cpu_has(X86_FEATURE_NODEID_MSR)) {
u64 value;
rdmsrl(MSR_FAM10H_NODE_ID, value);
__max_die_per_package = nodes_per_socket = ((value >> 3) & 7) + 1;
}
if (!boot_cpu_has(X86_FEATURE_AMD_SSBD) &&
!boot_cpu_has(X86_FEATURE_VIRT_SSBD)) {
/*
* Try to cache the base value so further operations can
* avoid RMW. If that faults, do not enable SSBD.
*/
if (!rdmsrl_safe(MSR_AMD64_LS_CFG, &x86_amd_ls_cfg_base)) {
setup_force_cpu_cap(X86_FEATURE_LS_CFG_SSBD);
setup_force_cpu_cap(X86_FEATURE_SSBD);
x86_amd_ls_cfg_ssbd_mask = 1ULL << 10;
}
}
}
static void early_init_hygon(struct cpuinfo_x86 *c)
{
u32 dummy;
early_init_hygon_mc(c);
set_cpu_cap(c, X86_FEATURE_K8);
rdmsr_safe(MSR_AMD64_PATCH_LEVEL, &c->microcode, &dummy);
/*
* c->x86_power is 8000_0007 edx. Bit 8 is TSC runs at constant rate
* with P/T states and does not stop in deep C-states
*/
if (c->x86_power & (1 << 8)) {
set_cpu_cap(c, X86_FEATURE_CONSTANT_TSC);
set_cpu_cap(c, X86_FEATURE_NONSTOP_TSC);
}
/* Bit 12 of 8000_0007 edx is accumulated power mechanism. */
if (c->x86_power & BIT(12))
set_cpu_cap(c, X86_FEATURE_ACC_POWER);
/* Bit 14 indicates the Runtime Average Power Limit interface. */
if (c->x86_power & BIT(14))
set_cpu_cap(c, X86_FEATURE_RAPL);
#ifdef CONFIG_X86_64
set_cpu_cap(c, X86_FEATURE_SYSCALL32);
#endif
#if defined(CONFIG_X86_LOCAL_APIC) && defined(CONFIG_PCI)
/*
* ApicID can always be treated as an 8-bit value for Hygon APIC So, we
* can safely set X86_FEATURE_EXTD_APICID unconditionally.
*/
if (boot_cpu_has(X86_FEATURE_APIC))
set_cpu_cap(c, X86_FEATURE_EXTD_APICID);
#endif
/*
* This is only needed to tell the kernel whether to use VMCALL
* and VMMCALL. VMMCALL is never executed except under virt, so
* we can set it unconditionally.
*/
set_cpu_cap(c, X86_FEATURE_VMMCALL);
hygon_get_topology_early(c);
}
static void init_hygon(struct cpuinfo_x86 *c)
{
u64 vm_cr;
early_init_hygon(c);
/*
* Bit 31 in normal CPUID used for nonstandard 3DNow ID;
* 3DNow is IDd by bit 31 in extended CPUID (1*32+31) anyway
*/
clear_cpu_cap(c, 0*32+31);
set_cpu_cap(c, X86_FEATURE_REP_GOOD);
/* get apicid instead of initial apic id from cpuid */
c->topo.apicid = read_apic_id();
/*
* XXX someone from Hygon needs to confirm this DTRT
*
init_spectral_chicken(c);
*/
set_cpu_cap(c, X86_FEATURE_ZEN);
set_cpu_cap(c, X86_FEATURE_CPB);
cpu_detect_cache_sizes(c);
hygon_detect_cmp(c);
hygon_get_topology(c);
srat_detect_node(c);
init_hygon_cacheinfo(c);
if (cpu_has(c, X86_FEATURE_SVM)) {
rdmsrl(MSR_VM_CR, vm_cr);
if (vm_cr & SVM_VM_CR_SVM_DIS_MASK) {
pr_notice_once("SVM disabled (by BIOS) in MSR_VM_CR\n");
clear_cpu_cap(c, X86_FEATURE_SVM);
}
}
if (cpu_has(c, X86_FEATURE_XMM2)) {
/*
* Use LFENCE for execution serialization. On families which
* don't have that MSR, LFENCE is already serializing.
* msr_set_bit() uses the safe accessors, too, even if the MSR
* is not present.
*/
msr_set_bit(MSR_AMD64_DE_CFG,
MSR_AMD64_DE_CFG_LFENCE_SERIALIZE_BIT);
/* A serializing LFENCE stops RDTSC speculation */
set_cpu_cap(c, X86_FEATURE_LFENCE_RDTSC);
}
/*
* Hygon processors have APIC timer running in deep C states.
*/
set_cpu_cap(c, X86_FEATURE_ARAT);
/* Hygon CPUs don't reset SS attributes on SYSRET, Xen does. */
if (!cpu_feature_enabled(X86_FEATURE_XENPV))
set_cpu_bug(c, X86_BUG_SYSRET_SS_ATTRS);
check_null_seg_clears_base(c);
/* Hygon CPUs don't need fencing after x2APIC/TSC_DEADLINE MSR writes. */
clear_cpu_cap(c, X86_FEATURE_APIC_MSRS_FENCE);
}
static void cpu_detect_tlb_hygon(struct cpuinfo_x86 *c)
{
u32 ebx, eax, ecx, edx;
u16 mask = 0xfff;
if (c->extended_cpuid_level < 0x80000006)
return;
cpuid(0x80000006, &eax, &ebx, &ecx, &edx);
tlb_lld_4k[ENTRIES] = (ebx >> 16) & mask;
tlb_lli_4k[ENTRIES] = ebx & mask;
/* Handle DTLB 2M and 4M sizes, fall back to L1 if L2 is disabled */
if (!((eax >> 16) & mask))
tlb_lld_2m[ENTRIES] = (cpuid_eax(0x80000005) >> 16) & 0xff;
else
tlb_lld_2m[ENTRIES] = (eax >> 16) & mask;
/* a 4M entry uses two 2M entries */
tlb_lld_4m[ENTRIES] = tlb_lld_2m[ENTRIES] >> 1;
/* Handle ITLB 2M and 4M sizes, fall back to L1 if L2 is disabled */
if (!(eax & mask)) {
cpuid(0x80000005, &eax, &ebx, &ecx, &edx);
tlb_lli_2m[ENTRIES] = eax & 0xff;
} else
tlb_lli_2m[ENTRIES] = eax & mask;
tlb_lli_4m[ENTRIES] = tlb_lli_2m[ENTRIES] >> 1;
}
static const struct cpu_dev hygon_cpu_dev = {
.c_vendor = "Hygon",
.c_ident = { "HygonGenuine" },
.c_early_init = early_init_hygon,
.c_detect_tlb = cpu_detect_tlb_hygon,
.c_bsp_init = bsp_init_hygon,
.c_init = init_hygon,
.c_x86_vendor = X86_VENDOR_HYGON,
};
cpu_dev_register(hygon_cpu_dev);