2 * Copyright (c) 1996, by Steve Passe
3 * Copyright (c) 2003, by Peter Wemm
6 * Redistribution and use in source and binary forms, with or without
7 * modification, are permitted provided that the following conditions
9 * 1. Redistributions of source code must retain the above copyright
10 * notice, this list of conditions and the following disclaimer.
11 * 2. The name of the developer may NOT be used to endorse or promote products
12 * derived from this software without specific prior written permission.
14 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
15 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
16 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
17 * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
18 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
19 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
20 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
21 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
22 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
23 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
27 #include <sys/cdefs.h>
28 __FBSDID("$FreeBSD$");
34 #include "opt_kstack_pages.h"
36 #include "opt_sched.h"
39 #include <sys/param.h>
40 #include <sys/systm.h>
42 #include <sys/cons.h> /* cngetc() */
43 #include <sys/cpuset.h>
48 #include <sys/kernel.h>
51 #include <sys/malloc.h>
52 #include <sys/memrange.h>
53 #include <sys/mutex.h>
56 #include <sys/sched.h>
58 #include <sys/sysctl.h>
61 #include <vm/vm_param.h>
63 #include <vm/vm_kern.h>
64 #include <vm/vm_extern.h>
65 #include <vm/vm_map.h>
67 #include <x86/apicreg.h>
68 #include <machine/clock.h>
69 #include <machine/cpu.h>
70 #include <machine/cputypes.h>
72 #include <machine/md_var.h>
73 #include <machine/pcb.h>
74 #include <machine/psl.h>
75 #include <machine/smp.h>
76 #include <machine/specialreg.h>
77 #include <x86/ucode.h>
79 static MALLOC_DEFINE(M_CPUS, "cpus", "CPU items");
81 /* lock region used by kernel profiling */
84 int mp_naps; /* # of Applications processors */
85 int boot_cpu_id = -1; /* designated BSP */
87 /* AP uses this during bootstrap. Do not staticize. */
91 /* Free these after use */
92 void *bootstacks[MAXCPU];
95 struct pcb stoppcbs[MAXCPU];
96 struct susppcb **susppcbs;
99 /* Interrupt counts. */
100 static u_long *ipi_preempt_counts[MAXCPU];
101 static u_long *ipi_ast_counts[MAXCPU];
102 u_long *ipi_invltlb_counts[MAXCPU];
103 u_long *ipi_invlrng_counts[MAXCPU];
104 u_long *ipi_invlpg_counts[MAXCPU];
105 u_long *ipi_invlcache_counts[MAXCPU];
106 u_long *ipi_rendezvous_counts[MAXCPU];
107 static u_long *ipi_hardclock_counts[MAXCPU];
110 /* Default cpu_ops implementation. */
111 struct cpu_ops cpu_ops;
114 * Local data and functions.
117 static volatile cpuset_t ipi_stop_nmi_pending;
119 volatile cpuset_t resuming_cpus;
120 volatile cpuset_t toresume_cpus;
122 /* used to hold the AP's until we are ready to release them */
123 struct mtx ap_boot_mtx;
125 /* Set to 1 once we're ready to let the APs out of the pen. */
126 volatile int aps_ready = 0;
129 * Store data from cpu_add() until later in the boot when we actually setup
132 struct cpu_info *cpu_info;
134 int cpu_apic_ids[MAXCPU];
135 _Static_assert(MAXCPU <= MAX_APIC_ID,
136 "MAXCPU cannot be larger that MAX_APIC_ID");
137 _Static_assert(xAPIC_MAX_APIC_ID <= MAX_APIC_ID,
138 "xAPIC_MAX_APIC_ID cannot be larger that MAX_APIC_ID");
140 static void release_aps(void *dummy);
141 static void cpustop_handler_post(u_int cpu);
143 static int hyperthreading_allowed = 1;
144 SYSCTL_INT(_machdep, OID_AUTO, hyperthreading_allowed, CTLFLAG_RDTUN,
145 &hyperthreading_allowed, 0, "Use Intel HTT logical CPUs");
147 static int hyperthreading_intr_allowed = 0;
148 SYSCTL_INT(_machdep, OID_AUTO, hyperthreading_intr_allowed, CTLFLAG_RDTUN,
149 &hyperthreading_intr_allowed, 0,
150 "Allow interrupts on HTT logical CPUs");
152 static struct topo_node topo_root;
154 static int pkg_id_shift;
155 static int node_id_shift;
156 static int core_id_shift;
157 static int disabled_cpus;
162 } static caches[MAX_CACHE_LEVELS];
164 unsigned int boot_address;
166 static bool stop_mwait = false;
167 SYSCTL_BOOL(_machdep, OID_AUTO, stop_mwait, CTLFLAG_RWTUN, &stop_mwait, 0,
168 "Use MONITOR/MWAIT when stopping CPU, if available");
170 #define MiB(v) (v ## ULL << 20)
173 mem_range_AP_init(void)
176 if (mem_range_softc.mr_op && mem_range_softc.mr_op->initAP)
177 mem_range_softc.mr_op->initAP(&mem_range_softc);
181 * Round up to the next power of two, if necessary, and then
183 * Returns -1 if argument is zero.
189 return (fls(x << (1 - powerof2(x))) - 1);
193 * Add a cache level to the cache topology description.
196 add_deterministic_cache(int type, int level, int share_count)
202 printf("unexpected cache type %d\n", type);
205 if (type == 2) /* ignore instruction cache */
207 if (level == 0 || level > MAX_CACHE_LEVELS) {
208 printf("unexpected cache level %d\n", type);
212 if (caches[level - 1].present) {
213 printf("WARNING: multiple entries for L%u data cache\n", level);
214 printf("%u => %u\n", caches[level - 1].id_shift,
215 mask_width(share_count));
217 caches[level - 1].id_shift = mask_width(share_count);
218 caches[level - 1].present = 1;
220 if (caches[level - 1].id_shift > pkg_id_shift) {
221 printf("WARNING: L%u data cache covers more "
222 "APIC IDs than a package (%u > %u)\n", level,
223 caches[level - 1].id_shift, pkg_id_shift);
224 caches[level - 1].id_shift = pkg_id_shift;
226 if (caches[level - 1].id_shift < core_id_shift) {
227 printf("WARNING: L%u data cache covers fewer "
228 "APIC IDs than a core (%u < %u)\n", level,
229 caches[level - 1].id_shift, core_id_shift);
230 caches[level - 1].id_shift = core_id_shift;
237 * Determine topology of processing units and caches for AMD CPUs.
239 * - AMD CPUID Specification (Publication # 25481)
240 * - BKDG for AMD NPT Family 0Fh Processors (Publication # 32559)
241 * - BKDG For AMD Family 10h Processors (Publication # 31116)
242 * - BKDG For AMD Family 15h Models 00h-0Fh Processors (Publication # 42301)
243 * - BKDG For AMD Family 16h Models 00h-0Fh Processors (Publication # 48751)
244 * - PPR For AMD Family 17h Models 00h-0Fh Processors (Publication # 54945)
252 int nodes_per_socket;
257 /* No multi-core capability. */
258 if ((amd_feature2 & AMDID2_CMP) == 0)
261 /* For families 10h and newer. */
262 pkg_id_shift = (cpu_procinfo2 & AMDID_COREID_SIZE) >>
263 AMDID_COREID_SIZE_SHIFT;
265 /* For 0Fh family. */
266 if (pkg_id_shift == 0)
268 mask_width((cpu_procinfo2 & AMDID_CMP_CORES) + 1);
271 * Families prior to 16h define the following value as
272 * cores per compute unit and we don't really care about the AMD
273 * compute units at the moment. Perhaps we should treat them as
274 * cores and cores within the compute units as hardware threads,
275 * but that's up for debate.
276 * Later families define the value as threads per compute unit,
277 * so we are following AMD's nomenclature here.
279 if ((amd_feature2 & AMDID2_TOPOLOGY) != 0 &&
280 CPUID_TO_FAMILY(cpu_id) >= 0x16) {
281 cpuid_count(0x8000001e, 0, p);
282 share_count = ((p[1] >> 8) & 0xff) + 1;
283 core_id_shift = mask_width(share_count);
286 * For Zen (17h), gather Nodes per Processor. Each node is a
287 * Zeppelin die; TR and EPYC CPUs will have multiple dies per
288 * package. Communication latency between dies is higher than
291 nodes_per_socket = ((p[2] >> 8) & 0x7) + 1;
292 node_id_shift = pkg_id_shift - mask_width(nodes_per_socket);
295 if ((amd_feature2 & AMDID2_TOPOLOGY) != 0) {
297 cpuid_count(0x8000001d, i, p);
299 level = (p[0] >> 5) & 0x7;
300 share_count = 1 + ((p[0] >> 14) & 0xfff);
302 if (!add_deterministic_cache(type, level, share_count))
306 if (cpu_exthigh >= 0x80000005) {
307 cpuid_count(0x80000005, 0, p);
308 if (((p[2] >> 24) & 0xff) != 0) {
309 caches[0].id_shift = 0;
310 caches[0].present = 1;
313 if (cpu_exthigh >= 0x80000006) {
314 cpuid_count(0x80000006, 0, p);
315 if (((p[2] >> 16) & 0xffff) != 0) {
316 caches[1].id_shift = 0;
317 caches[1].present = 1;
319 if (((p[3] >> 18) & 0x3fff) != 0) {
320 nodes_per_socket = 1;
321 if ((amd_feature2 & AMDID2_NODE_ID) != 0) {
323 * Handle multi-node processors that
324 * have multiple chips, each with its
325 * own L3 cache, on the same die.
327 v = rdmsr(0xc001100c);
328 nodes_per_socket = 1 + ((v >> 3) & 0x7);
331 pkg_id_shift - mask_width(nodes_per_socket);
332 caches[2].present = 1;
339 * Determine topology of processing units for Intel CPUs
340 * using CPUID Leaf 1 and Leaf 4, if supported.
342 * - Intel 64 Architecture Processor Topology Enumeration
343 * - Intel 64 and IA-32 ArchitecturesSoftware Developer’s Manual,
344 * Volume 3A: System Programming Guide, PROGRAMMING CONSIDERATIONS
345 * FOR HARDWARE MULTI-THREADING CAPABLE PROCESSORS
348 topo_probe_intel_0x4(void)
354 /* Both zero and one here mean one logical processor per package. */
355 max_logical = (cpu_feature & CPUID_HTT) != 0 ?
356 (cpu_procinfo & CPUID_HTT_CORES) >> 16 : 1;
357 if (max_logical <= 1)
360 if (cpu_high >= 0x4) {
361 cpuid_count(0x04, 0, p);
362 max_cores = ((p[0] >> 26) & 0x3f) + 1;
366 core_id_shift = mask_width(max_logical/max_cores);
367 KASSERT(core_id_shift >= 0,
368 ("intel topo: max_cores > max_logical\n"));
369 pkg_id_shift = core_id_shift + mask_width(max_cores);
373 * Determine topology of processing units for Intel CPUs
374 * using CPUID Leaf 11, if supported.
376 * - Intel 64 Architecture Processor Topology Enumeration
377 * - Intel 64 and IA-32 ArchitecturesSoftware Developer’s Manual,
378 * Volume 3A: System Programming Guide, PROGRAMMING CONSIDERATIONS
379 * FOR HARDWARE MULTI-THREADING CAPABLE PROCESSORS
382 topo_probe_intel_0xb(void)
389 /* Fall back if CPU leaf 11 doesn't really exist. */
390 cpuid_count(0x0b, 0, p);
392 topo_probe_intel_0x4();
396 /* We only support three levels for now. */
398 cpuid_count(0x0b, i, p);
401 type = (p[2] >> 8) & 0xff;
406 /* TODO: check for duplicate (re-)assignment */
407 if (type == CPUID_TYPE_SMT)
408 core_id_shift = bits;
409 else if (type == CPUID_TYPE_CORE)
412 printf("unknown CPU level type %d\n", type);
415 if (pkg_id_shift < core_id_shift) {
416 printf("WARNING: core covers more APIC IDs than a package\n");
417 core_id_shift = pkg_id_shift;
422 * Determine topology of caches for Intel CPUs.
424 * - Intel 64 Architecture Processor Topology Enumeration
425 * - Intel 64 and IA-32 Architectures Software Developer’s Manual
426 * Volume 2A: Instruction Set Reference, A-M,
430 topo_probe_intel_caches(void)
438 if (cpu_high < 0x4) {
440 * Available cache level and sizes can be determined
441 * via CPUID leaf 2, but that requires a huge table of hardcoded
442 * values, so for now just assume L1 and L2 caches potentially
443 * shared only by HTT processing units, if HTT is present.
445 caches[0].id_shift = pkg_id_shift;
446 caches[0].present = 1;
447 caches[1].id_shift = pkg_id_shift;
448 caches[1].present = 1;
453 cpuid_count(0x4, i, p);
455 level = (p[0] >> 5) & 0x7;
456 share_count = 1 + ((p[0] >> 14) & 0xfff);
458 if (!add_deterministic_cache(type, level, share_count))
464 * Determine topology of processing units and caches for Intel CPUs.
466 * - Intel 64 Architecture Processor Topology Enumeration
469 topo_probe_intel(void)
473 * Note that 0x1 <= cpu_high < 4 case should be
474 * compatible with topo_probe_intel_0x4() logic when
475 * CPUID.1:EBX[23:16] > 0 (cpu_cores will be 1)
476 * or it should trigger the fallback otherwise.
479 topo_probe_intel_0xb();
480 else if (cpu_high >= 0x1)
481 topo_probe_intel_0x4();
483 topo_probe_intel_caches();
487 * Topology information is queried only on BSP, on which this
488 * code runs and for which it can query CPUID information.
489 * Then topology is extrapolated on all packages using an
490 * assumption that APIC ID to hardware component ID mapping is
492 * That doesn't necesserily imply that the topology is uniform.
497 static int cpu_topo_probed = 0;
498 struct x86_topo_layer {
502 } topo_layers[MAX_CACHE_LEVELS + 4];
503 struct topo_node *parent;
504 struct topo_node *node;
513 CPU_ZERO(&logical_cpus_mask);
517 else if (cpu_vendor_id == CPU_VENDOR_AMD)
519 else if (cpu_vendor_id == CPU_VENDOR_INTEL)
522 KASSERT(pkg_id_shift >= core_id_shift,
523 ("bug in APIC topology discovery"));
526 bzero(topo_layers, sizeof(topo_layers));
528 topo_layers[nlayers].type = TOPO_TYPE_PKG;
529 topo_layers[nlayers].id_shift = pkg_id_shift;
531 printf("Package ID shift: %u\n", topo_layers[nlayers].id_shift);
534 if (pkg_id_shift > node_id_shift && node_id_shift != 0) {
535 topo_layers[nlayers].type = TOPO_TYPE_GROUP;
536 topo_layers[nlayers].id_shift = node_id_shift;
538 printf("Node ID shift: %u\n",
539 topo_layers[nlayers].id_shift);
544 * Consider all caches to be within a package/chip
545 * and "in front" of all sub-components like
546 * cores and hardware threads.
548 for (i = MAX_CACHE_LEVELS - 1; i >= 0; --i) {
549 if (caches[i].present) {
550 if (node_id_shift != 0)
551 KASSERT(caches[i].id_shift <= node_id_shift,
552 ("bug in APIC topology discovery"));
553 KASSERT(caches[i].id_shift <= pkg_id_shift,
554 ("bug in APIC topology discovery"));
555 KASSERT(caches[i].id_shift >= core_id_shift,
556 ("bug in APIC topology discovery"));
558 topo_layers[nlayers].type = TOPO_TYPE_CACHE;
559 topo_layers[nlayers].subtype = i + 1;
560 topo_layers[nlayers].id_shift = caches[i].id_shift;
562 printf("L%u cache ID shift: %u\n",
563 topo_layers[nlayers].subtype,
564 topo_layers[nlayers].id_shift);
569 if (pkg_id_shift > core_id_shift) {
570 topo_layers[nlayers].type = TOPO_TYPE_CORE;
571 topo_layers[nlayers].id_shift = core_id_shift;
573 printf("Core ID shift: %u\n",
574 topo_layers[nlayers].id_shift);
578 topo_layers[nlayers].type = TOPO_TYPE_PU;
579 topo_layers[nlayers].id_shift = 0;
582 topo_init_root(&topo_root);
583 for (i = 0; i <= max_apic_id; ++i) {
584 if (!cpu_info[i].cpu_present)
588 for (layer = 0; layer < nlayers; ++layer) {
589 node_id = i >> topo_layers[layer].id_shift;
590 parent = topo_add_node_by_hwid(parent, node_id,
591 topo_layers[layer].type,
592 topo_layers[layer].subtype);
597 for (layer = 0; layer < nlayers; ++layer) {
598 node_id = boot_cpu_id >> topo_layers[layer].id_shift;
599 node = topo_find_node_by_hwid(parent, node_id,
600 topo_layers[layer].type,
601 topo_layers[layer].subtype);
602 topo_promote_child(node);
610 * Assign logical CPU IDs to local APICs.
615 struct topo_node *node;
619 smt_mask = (1u << core_id_shift) - 1;
622 * Assign CPU IDs to local APIC IDs and disable any CPUs
623 * beyond MAXCPU. CPU 0 is always assigned to the BSP.
627 TOPO_FOREACH(node, &topo_root) {
628 if (node->type != TOPO_TYPE_PU)
631 if ((node->hwid & smt_mask) != (boot_cpu_id & smt_mask))
632 cpu_info[node->hwid].cpu_hyperthread = 1;
634 if (resource_disabled("lapic", node->hwid)) {
635 if (node->hwid != boot_cpu_id)
636 cpu_info[node->hwid].cpu_disabled = 1;
638 printf("Cannot disable BSP, APIC ID = %d\n",
642 if (!hyperthreading_allowed &&
643 cpu_info[node->hwid].cpu_hyperthread)
644 cpu_info[node->hwid].cpu_disabled = 1;
646 if (mp_ncpus >= MAXCPU)
647 cpu_info[node->hwid].cpu_disabled = 1;
649 if (cpu_info[node->hwid].cpu_disabled) {
654 if (cpu_info[node->hwid].cpu_hyperthread)
657 cpu_apic_ids[mp_ncpus] = node->hwid;
658 apic_cpuids[node->hwid] = mp_ncpus;
659 topo_set_pu_id(node, mp_ncpus);
663 KASSERT(mp_maxid >= mp_ncpus - 1,
664 ("%s: counters out of sync: max %d, count %d", __func__, mp_maxid,
667 mp_ncores = mp_ncpus - nhyper;
668 smp_threads_per_core = mp_ncpus / mp_ncores;
672 * Print various information about the SMP system hardware and setup.
675 cpu_mp_announce(void)
677 struct topo_node *node;
678 const char *hyperthread;
679 struct topo_analysis topology;
681 printf("FreeBSD/SMP: ");
682 if (topo_analyze(&topo_root, 1, &topology)) {
683 printf("%d package(s)", topology.entities[TOPO_LEVEL_PKG]);
684 if (topology.entities[TOPO_LEVEL_GROUP] > 1)
685 printf(" x %d groups",
686 topology.entities[TOPO_LEVEL_GROUP]);
687 if (topology.entities[TOPO_LEVEL_CACHEGROUP] > 1)
688 printf(" x %d cache groups",
689 topology.entities[TOPO_LEVEL_CACHEGROUP]);
690 if (topology.entities[TOPO_LEVEL_CORE] > 0)
691 printf(" x %d core(s)",
692 topology.entities[TOPO_LEVEL_CORE]);
693 if (topology.entities[TOPO_LEVEL_THREAD] > 1)
694 printf(" x %d hardware threads",
695 topology.entities[TOPO_LEVEL_THREAD]);
697 printf("Non-uniform topology");
702 printf("FreeBSD/SMP Online: ");
703 if (topo_analyze(&topo_root, 0, &topology)) {
704 printf("%d package(s)",
705 topology.entities[TOPO_LEVEL_PKG]);
706 if (topology.entities[TOPO_LEVEL_GROUP] > 1)
707 printf(" x %d groups",
708 topology.entities[TOPO_LEVEL_GROUP]);
709 if (topology.entities[TOPO_LEVEL_CACHEGROUP] > 1)
710 printf(" x %d cache groups",
711 topology.entities[TOPO_LEVEL_CACHEGROUP]);
712 if (topology.entities[TOPO_LEVEL_CORE] > 0)
713 printf(" x %d core(s)",
714 topology.entities[TOPO_LEVEL_CORE]);
715 if (topology.entities[TOPO_LEVEL_THREAD] > 1)
716 printf(" x %d hardware threads",
717 topology.entities[TOPO_LEVEL_THREAD]);
719 printf("Non-uniform topology");
727 TOPO_FOREACH(node, &topo_root) {
728 switch (node->type) {
730 printf("Package HW ID = %u\n", node->hwid);
733 printf("\tCore HW ID = %u\n", node->hwid);
736 if (cpu_info[node->hwid].cpu_hyperthread)
741 if (node->subtype == 0)
742 printf("\t\tCPU (AP%s): APIC ID: %u"
743 "(disabled)\n", hyperthread, node->hwid);
744 else if (node->id == 0)
745 printf("\t\tCPU0 (BSP): APIC ID: %u\n",
748 printf("\t\tCPU%u (AP%s): APIC ID: %u\n",
749 node->id, hyperthread, node->hwid);
759 * Add a scheduling group, a group of logical processors sharing
760 * a particular cache (and, thus having an affinity), to the scheduling
762 * This function recursively works on lower level caches.
765 x86topo_add_sched_group(struct topo_node *root, struct cpu_group *cg_root)
767 struct topo_node *node;
772 KASSERT(root->type == TOPO_TYPE_SYSTEM || root->type == TOPO_TYPE_CACHE ||
773 root->type == TOPO_TYPE_GROUP,
774 ("x86topo_add_sched_group: bad type: %u", root->type));
775 CPU_COPY(&root->cpuset, &cg_root->cg_mask);
776 cg_root->cg_count = root->cpu_count;
777 if (root->type == TOPO_TYPE_SYSTEM)
778 cg_root->cg_level = CG_SHARE_NONE;
780 cg_root->cg_level = root->subtype;
783 * Check how many core nodes we have under the given root node.
784 * If we have multiple logical processors, but not multiple
785 * cores, then those processors must be hardware threads.
789 while (node != NULL) {
790 if (node->type != TOPO_TYPE_CORE) {
791 node = topo_next_node(root, node);
796 node = topo_next_nonchild_node(root, node);
799 if (cg_root->cg_level != CG_SHARE_NONE &&
800 root->cpu_count > 1 && ncores < 2)
801 cg_root->cg_flags = CG_FLAG_SMT;
804 * Find out how many cache nodes we have under the given root node.
805 * We ignore cache nodes that cover all the same processors as the
806 * root node. Also, we do not descend below found cache nodes.
807 * That is, we count top-level "non-redundant" caches under the root
812 while (node != NULL) {
813 if ((node->type != TOPO_TYPE_GROUP &&
814 node->type != TOPO_TYPE_CACHE) ||
815 (root->type != TOPO_TYPE_SYSTEM &&
816 CPU_CMP(&node->cpuset, &root->cpuset) == 0)) {
817 node = topo_next_node(root, node);
821 node = topo_next_nonchild_node(root, node);
824 cg_root->cg_child = smp_topo_alloc(nchildren);
825 cg_root->cg_children = nchildren;
828 * Now find again the same cache nodes as above and recursively
829 * build scheduling topologies for them.
833 while (node != NULL) {
834 if ((node->type != TOPO_TYPE_GROUP &&
835 node->type != TOPO_TYPE_CACHE) ||
836 (root->type != TOPO_TYPE_SYSTEM &&
837 CPU_CMP(&node->cpuset, &root->cpuset) == 0)) {
838 node = topo_next_node(root, node);
841 cg_root->cg_child[i].cg_parent = cg_root;
842 x86topo_add_sched_group(node, &cg_root->cg_child[i]);
844 node = topo_next_nonchild_node(root, node);
849 * Build the MI scheduling topology from the discovered hardware topology.
854 struct cpu_group *cg_root;
857 return (smp_topo_none());
859 cg_root = smp_topo_alloc(1);
860 x86topo_add_sched_group(&topo_root, cg_root);
865 cpu_alloc(void *dummy __unused)
868 * Dynamically allocate the arrays that depend on the
871 cpu_info = malloc(sizeof(*cpu_info) * (max_apic_id + 1), M_CPUS,
873 apic_cpuids = malloc(sizeof(*apic_cpuids) * (max_apic_id + 1), M_CPUS,
876 SYSINIT(cpu_alloc, SI_SUB_CPU, SI_ORDER_FIRST, cpu_alloc, NULL);
879 * Add a logical CPU to the topology.
882 cpu_add(u_int apic_id, char boot_cpu)
885 if (apic_id > max_apic_id) {
886 panic("SMP: APIC ID %d too high", apic_id);
889 KASSERT(cpu_info[apic_id].cpu_present == 0, ("CPU %u added twice",
891 cpu_info[apic_id].cpu_present = 1;
893 KASSERT(boot_cpu_id == -1,
894 ("CPU %u claims to be BSP, but CPU %u already is", apic_id,
896 boot_cpu_id = apic_id;
897 cpu_info[apic_id].cpu_bsp = 1;
900 printf("SMP: Added CPU %u (%s)\n", apic_id, boot_cpu ? "BSP" :
905 cpu_mp_setmaxid(void)
909 * mp_ncpus and mp_maxid should be already set by calls to cpu_add().
910 * If there were no calls to cpu_add() assume this is a UP system.
921 * Always record BSP in CPU map so that the mbuf init code works
924 CPU_SETOF(0, &all_cpus);
925 return (mp_ncpus > 1);
928 /* Allocate memory for the AP trampoline. */
930 alloc_ap_trampoline(vm_paddr_t *physmap, unsigned int *physmap_idx)
936 for (i = *physmap_idx; i <= *physmap_idx; i -= 2) {
938 * Find a memory region big enough and below the 1MB boundary
939 * for the trampoline code.
940 * NB: needs to be page aligned.
942 if (physmap[i] >= MiB(1) ||
943 (trunc_page(physmap[i + 1]) - round_page(physmap[i])) <
944 round_page(bootMP_size))
949 * Try to steal from the end of the region to mimic previous
950 * behaviour, else fallback to steal from the start.
952 if (physmap[i + 1] < MiB(1)) {
953 boot_address = trunc_page(physmap[i + 1]);
954 if ((physmap[i + 1] - boot_address) < bootMP_size)
955 boot_address -= round_page(bootMP_size);
956 physmap[i + 1] = boot_address;
958 boot_address = round_page(physmap[i]);
959 physmap[i] = boot_address + round_page(bootMP_size);
961 if (physmap[i] == physmap[i + 1] && *physmap_idx != 0) {
962 memmove(&physmap[i], &physmap[i + 2],
963 sizeof(*physmap) * (*physmap_idx - i + 2));
970 boot_address = basemem * 1024 - bootMP_size;
973 "Cannot find enough space for the boot trampoline, placing it at %#x",
979 * AP CPU's call this to initialize themselves.
982 init_secondary_tail(void)
986 pmap_activate_boot(vmspace_pmap(proc0.p_vmspace));
989 * On real hardware, switch to x2apic mode if possible. Do it
990 * after aps_ready was signalled, to avoid manipulating the
991 * mode while BSP might still want to send some IPI to us
992 * (second startup IPI is ignored on modern hardware etc).
996 /* Initialize the PAT MSR. */
999 /* set up CPU registers and state */
1005 /* set up FPU state on the AP */
1012 if (cpu_ops.cpu_init)
1015 /* A quick check from sanity claus */
1016 cpuid = PCPU_GET(cpuid);
1017 if (PCPU_GET(apic_id) != lapic_id()) {
1018 printf("SMP: cpuid = %d\n", cpuid);
1019 printf("SMP: actual apic_id = %d\n", lapic_id());
1020 printf("SMP: correct apic_id = %d\n", PCPU_GET(apic_id));
1021 panic("cpuid mismatch! boom!!");
1024 /* Initialize curthread. */
1025 KASSERT(PCPU_GET(idlethread) != NULL, ("no idle thread"));
1026 PCPU_SET(curthread, PCPU_GET(idlethread));
1028 mtx_lock_spin(&ap_boot_mtx);
1032 /* Init local apic for irq's */
1035 /* Set memory range attributes for this CPU to match the BSP */
1036 mem_range_AP_init();
1040 CTR1(KTR_SMP, "SMP: AP CPU #%d Launched", cpuid);
1042 printf("SMP: AP CPU #%d Launched!\n", cpuid);
1044 printf("%s%d%s", smp_cpus == 2 ? "Launching APs: " : "",
1045 cpuid, smp_cpus == mp_ncpus ? "\n" : " ");
1047 /* Determine if we are a logical CPU. */
1048 if (cpu_info[PCPU_GET(apic_id)].cpu_hyperthread)
1049 CPU_SET(cpuid, &logical_cpus_mask);
1054 if (smp_cpus == mp_ncpus) {
1055 /* enable IPI's, tlb shootdown, freezes etc */
1056 atomic_store_rel_int(&smp_started, 1);
1061 * Enable global pages TLB extension
1062 * This also implicitly flushes the TLB
1064 load_cr4(rcr4() | CR4_PGE);
1065 if (pmap_pcid_enabled)
1066 load_cr4(rcr4() | CR4_PCIDE);
1072 mtx_unlock_spin(&ap_boot_mtx);
1074 /* Wait until all the AP's are up. */
1075 while (atomic_load_acq_int(&smp_started) == 0)
1078 #ifndef EARLY_AP_STARTUP
1079 /* Start per-CPU event timers. */
1080 cpu_initclocks_ap();
1085 panic("scheduler returned us to %s", __func__);
1090 smp_after_idle_runnable(void *arg __unused)
1092 struct thread *idle_td;
1095 for (cpu = 1; cpu < mp_ncpus; cpu++) {
1096 idle_td = pcpu_find(cpu)->pc_idlethread;
1097 while (atomic_load_int(&idle_td->td_lastcpu) == NOCPU &&
1098 atomic_load_int(&idle_td->td_oncpu) == NOCPU)
1100 kmem_free((vm_offset_t)bootstacks[cpu], kstack_pages *
1104 SYSINIT(smp_after_idle_runnable, SI_SUB_SMP, SI_ORDER_ANY,
1105 smp_after_idle_runnable, NULL);
1108 * We tell the I/O APIC code about all the CPUs we want to receive
1109 * interrupts. If we don't want certain CPUs to receive IRQs we
1110 * can simply not tell the I/O APIC code about them in this function.
1111 * We also do not tell it about the BSP since it tells itself about
1112 * the BSP internally to work with UP kernels and on UP machines.
1115 set_interrupt_apic_ids(void)
1119 for (i = 0; i < MAXCPU; i++) {
1120 apic_id = cpu_apic_ids[i];
1123 if (cpu_info[apic_id].cpu_bsp)
1125 if (cpu_info[apic_id].cpu_disabled)
1128 /* Don't let hyperthreads service interrupts. */
1129 if (cpu_info[apic_id].cpu_hyperthread &&
1130 !hyperthreading_intr_allowed)
1138 #ifdef COUNT_XINVLTLB_HITS
1139 u_int xhits_gbl[MAXCPU];
1140 u_int xhits_pg[MAXCPU];
1141 u_int xhits_rng[MAXCPU];
1142 static SYSCTL_NODE(_debug, OID_AUTO, xhits, CTLFLAG_RW, 0, "");
1143 SYSCTL_OPAQUE(_debug_xhits, OID_AUTO, global, CTLFLAG_RW, &xhits_gbl,
1144 sizeof(xhits_gbl), "IU", "");
1145 SYSCTL_OPAQUE(_debug_xhits, OID_AUTO, page, CTLFLAG_RW, &xhits_pg,
1146 sizeof(xhits_pg), "IU", "");
1147 SYSCTL_OPAQUE(_debug_xhits, OID_AUTO, range, CTLFLAG_RW, &xhits_rng,
1148 sizeof(xhits_rng), "IU", "");
1153 u_int ipi_range_size;
1154 SYSCTL_INT(_debug_xhits, OID_AUTO, ipi_global, CTLFLAG_RW, &ipi_global, 0, "");
1155 SYSCTL_INT(_debug_xhits, OID_AUTO, ipi_page, CTLFLAG_RW, &ipi_page, 0, "");
1156 SYSCTL_INT(_debug_xhits, OID_AUTO, ipi_range, CTLFLAG_RW, &ipi_range, 0, "");
1157 SYSCTL_INT(_debug_xhits, OID_AUTO, ipi_range_size, CTLFLAG_RW, &ipi_range_size,
1159 #endif /* COUNT_XINVLTLB_HITS */
1162 * Init and startup IPI.
1165 ipi_startup(int apic_id, int vector)
1169 * This attempts to follow the algorithm described in the
1170 * Intel Multiprocessor Specification v1.4 in section B.4.
1171 * For each IPI, we allow the local APIC ~20us to deliver the
1172 * IPI. If that times out, we panic.
1176 * first we do an INIT IPI: this INIT IPI might be run, resetting
1177 * and running the target CPU. OR this INIT IPI might be latched (P5
1178 * bug), CPU waiting for STARTUP IPI. OR this INIT IPI might be
1181 lapic_ipi_raw(APIC_DEST_DESTFLD | APIC_TRIGMOD_LEVEL |
1182 APIC_LEVEL_ASSERT | APIC_DESTMODE_PHY | APIC_DELMODE_INIT, apic_id);
1183 lapic_ipi_wait(100);
1185 /* Explicitly deassert the INIT IPI. */
1186 lapic_ipi_raw(APIC_DEST_DESTFLD | APIC_TRIGMOD_LEVEL |
1187 APIC_LEVEL_DEASSERT | APIC_DESTMODE_PHY | APIC_DELMODE_INIT,
1190 DELAY(10000); /* wait ~10mS */
1193 * next we do a STARTUP IPI: the previous INIT IPI might still be
1194 * latched, (P5 bug) this 1st STARTUP would then terminate
1195 * immediately, and the previously started INIT IPI would continue. OR
1196 * the previous INIT IPI has already run. and this STARTUP IPI will
1197 * run. OR the previous INIT IPI was ignored. and this STARTUP IPI
1200 lapic_ipi_raw(APIC_DEST_DESTFLD | APIC_TRIGMOD_EDGE |
1201 APIC_LEVEL_ASSERT | APIC_DESTMODE_PHY | APIC_DELMODE_STARTUP |
1203 if (!lapic_ipi_wait(100))
1204 panic("Failed to deliver first STARTUP IPI to APIC %d",
1206 DELAY(200); /* wait ~200uS */
1209 * finally we do a 2nd STARTUP IPI: this 2nd STARTUP IPI should run IF
1210 * the previous STARTUP IPI was cancelled by a latched INIT IPI. OR
1211 * this STARTUP IPI will be ignored, as only ONE STARTUP IPI is
1212 * recognized after hardware RESET or INIT IPI.
1214 lapic_ipi_raw(APIC_DEST_DESTFLD | APIC_TRIGMOD_EDGE |
1215 APIC_LEVEL_ASSERT | APIC_DESTMODE_PHY | APIC_DELMODE_STARTUP |
1217 if (!lapic_ipi_wait(100))
1218 panic("Failed to deliver second STARTUP IPI to APIC %d",
1221 DELAY(200); /* wait ~200uS */
1225 * Send an IPI to specified CPU handling the bitmap logic.
1228 ipi_send_cpu(int cpu, u_int ipi)
1230 u_int bitmap, old, new;
1233 KASSERT(cpu_apic_ids[cpu] != -1, ("IPI to non-existent CPU %d", cpu));
1235 if (IPI_IS_BITMAPED(ipi)) {
1237 ipi = IPI_BITMAP_VECTOR;
1238 cpu_bitmap = &cpuid_to_pcpu[cpu]->pc_ipi_bitmap;
1241 if ((old & bitmap) == bitmap)
1244 if (atomic_fcmpset_int(cpu_bitmap, &old, new))
1250 lapic_ipi_vectored(ipi, cpu_apic_ids[cpu]);
1254 ipi_bitmap_handler(struct trapframe frame)
1256 struct trapframe *oldframe;
1258 int cpu = PCPU_GET(cpuid);
1263 td->td_intr_nesting_level++;
1264 oldframe = td->td_intr_frame;
1265 td->td_intr_frame = &frame;
1266 ipi_bitmap = atomic_readandclear_int(&cpuid_to_pcpu[cpu]->pc_ipi_bitmap);
1267 if (ipi_bitmap & (1 << IPI_PREEMPT)) {
1269 (*ipi_preempt_counts[cpu])++;
1273 if (ipi_bitmap & (1 << IPI_AST)) {
1275 (*ipi_ast_counts[cpu])++;
1277 /* Nothing to do for AST */
1279 if (ipi_bitmap & (1 << IPI_HARDCLOCK)) {
1281 (*ipi_hardclock_counts[cpu])++;
1285 td->td_intr_frame = oldframe;
1286 td->td_intr_nesting_level--;
1291 * send an IPI to a set of cpus.
1294 ipi_selected(cpuset_t cpus, u_int ipi)
1299 * IPI_STOP_HARD maps to a NMI and the trap handler needs a bit
1300 * of help in order to understand what is the source.
1301 * Set the mask of receiving CPUs for this purpose.
1303 if (ipi == IPI_STOP_HARD)
1304 CPU_OR_ATOMIC(&ipi_stop_nmi_pending, &cpus);
1306 while ((cpu = CPU_FFS(&cpus)) != 0) {
1308 CPU_CLR(cpu, &cpus);
1309 CTR3(KTR_SMP, "%s: cpu: %d ipi: %x", __func__, cpu, ipi);
1310 ipi_send_cpu(cpu, ipi);
1315 * send an IPI to a specific CPU.
1318 ipi_cpu(int cpu, u_int ipi)
1322 * IPI_STOP_HARD maps to a NMI and the trap handler needs a bit
1323 * of help in order to understand what is the source.
1324 * Set the mask of receiving CPUs for this purpose.
1326 if (ipi == IPI_STOP_HARD)
1327 CPU_SET_ATOMIC(cpu, &ipi_stop_nmi_pending);
1329 CTR3(KTR_SMP, "%s: cpu: %d ipi: %x", __func__, cpu, ipi);
1330 ipi_send_cpu(cpu, ipi);
1334 * send an IPI to all CPUs EXCEPT myself
1337 ipi_all_but_self(u_int ipi)
1339 cpuset_t other_cpus;
1341 other_cpus = all_cpus;
1342 CPU_CLR(PCPU_GET(cpuid), &other_cpus);
1343 if (IPI_IS_BITMAPED(ipi)) {
1344 ipi_selected(other_cpus, ipi);
1349 * IPI_STOP_HARD maps to a NMI and the trap handler needs a bit
1350 * of help in order to understand what is the source.
1351 * Set the mask of receiving CPUs for this purpose.
1353 if (ipi == IPI_STOP_HARD)
1354 CPU_OR_ATOMIC(&ipi_stop_nmi_pending, &other_cpus);
1356 CTR2(KTR_SMP, "%s: ipi: %x", __func__, ipi);
1357 lapic_ipi_vectored(ipi, APIC_IPI_DEST_OTHERS);
1361 ipi_nmi_handler(void)
1366 * As long as there is not a simple way to know about a NMI's
1367 * source, if the bitmask for the current CPU is present in
1368 * the global pending bitword an IPI_STOP_HARD has been issued
1369 * and should be handled.
1371 cpuid = PCPU_GET(cpuid);
1372 if (!CPU_ISSET(cpuid, &ipi_stop_nmi_pending))
1375 CPU_CLR_ATOMIC(cpuid, &ipi_stop_nmi_pending);
1383 nmi_call_kdb_smp(u_int type, struct trapframe *frame)
1388 cpu = PCPU_GET(cpuid);
1389 if (atomic_cmpset_acq_int(&nmi_kdb_lock, 0, 1)) {
1390 nmi_call_kdb(cpu, type, frame);
1393 savectx(&stoppcbs[cpu]);
1394 CPU_SET_ATOMIC(cpu, &stopped_cpus);
1395 while (!atomic_cmpset_acq_int(&nmi_kdb_lock, 0, 1))
1399 atomic_store_rel_int(&nmi_kdb_lock, 0);
1401 cpustop_handler_post(cpu);
1405 * Handle an IPI_STOP by saving our current context and spinning (or mwaiting,
1406 * if available) until we are resumed.
1409 cpustop_handler(void)
1411 struct monitorbuf *mb;
1415 cpu = PCPU_GET(cpuid);
1417 savectx(&stoppcbs[cpu]);
1419 use_mwait = (stop_mwait && (cpu_feature2 & CPUID2_MON) != 0 &&
1420 !mwait_cpustop_broken);
1422 mb = PCPU_PTR(monitorbuf);
1423 atomic_store_int(&mb->stop_state,
1424 MONITOR_STOPSTATE_STOPPED);
1427 /* Indicate that we are stopped */
1428 CPU_SET_ATOMIC(cpu, &stopped_cpus);
1430 /* Wait for restart */
1431 while (!CPU_ISSET(cpu, &started_cpus)) {
1433 cpu_monitor(mb, 0, 0);
1434 if (atomic_load_int(&mb->stop_state) ==
1435 MONITOR_STOPSTATE_STOPPED)
1436 cpu_mwait(0, MWAIT_C1);
1443 * Halt non-BSP CPUs on panic -- we're never going to need them
1444 * again, and might as well save power / release resources
1445 * (e.g., overprovisioned VM infrastructure).
1447 while (__predict_false(!IS_BSP() && panicstr != NULL))
1451 cpustop_handler_post(cpu);
1455 cpustop_handler_post(u_int cpu)
1458 CPU_CLR_ATOMIC(cpu, &started_cpus);
1459 CPU_CLR_ATOMIC(cpu, &stopped_cpus);
1462 * We don't broadcast TLB invalidations to other CPUs when they are
1463 * stopped. Hence, we clear the TLB before resuming.
1467 #if defined(__amd64__) && defined(DDB)
1468 amd64_db_resume_dbreg();
1471 if (cpu == 0 && cpustop_restartfunc != NULL) {
1472 cpustop_restartfunc();
1473 cpustop_restartfunc = NULL;
1478 * Handle an IPI_SUSPEND by saving our current context and spinning until we
1482 cpususpend_handler(void)
1486 mtx_assert(&smp_ipi_mtx, MA_NOTOWNED);
1488 cpu = PCPU_GET(cpuid);
1489 if (savectx(&susppcbs[cpu]->sp_pcb)) {
1491 fpususpend(susppcbs[cpu]->sp_fpususpend);
1493 npxsuspend(susppcbs[cpu]->sp_fpususpend);
1496 * suspended_cpus is cleared shortly after each AP is restarted
1497 * by a Startup IPI, so that the BSP can proceed to restarting
1500 * resuming_cpus gets cleared when the AP completes
1501 * initialization after having been released by the BSP.
1502 * resuming_cpus is probably not the best name for the
1503 * variable, because it is actually a set of processors that
1504 * haven't resumed yet and haven't necessarily started resuming.
1506 * Note that suspended_cpus is meaningful only for ACPI suspend
1507 * as it's not really used for Xen suspend since the APs are
1508 * automatically restored to the running state and the correct
1509 * context. For the same reason resumectx is never called in
1512 CPU_SET_ATOMIC(cpu, &suspended_cpus);
1513 CPU_SET_ATOMIC(cpu, &resuming_cpus);
1516 * Invalidate the cache after setting the global status bits.
1517 * The last AP to set its bit may end up being an Owner of the
1518 * corresponding cache line in MOESI protocol. The AP may be
1519 * stopped before the cache line is written to the main memory.
1524 fpuresume(susppcbs[cpu]->sp_fpususpend);
1526 npxresume(susppcbs[cpu]->sp_fpususpend);
1530 PCPU_SET(switchtime, 0);
1531 PCPU_SET(switchticks, ticks);
1533 /* Indicate that we have restarted and restored the context. */
1534 CPU_CLR_ATOMIC(cpu, &suspended_cpus);
1537 /* Wait for resume directive */
1538 while (!CPU_ISSET(cpu, &toresume_cpus))
1541 /* Re-apply microcode updates. */
1545 /* Finish removing the identity mapping of low memory for this AP. */
1549 if (cpu_ops.cpu_resume)
1550 cpu_ops.cpu_resume();
1556 /* Resume MCA and local APIC */
1561 /* Indicate that we are resumed */
1562 CPU_CLR_ATOMIC(cpu, &resuming_cpus);
1563 CPU_CLR_ATOMIC(cpu, &suspended_cpus);
1564 CPU_CLR_ATOMIC(cpu, &toresume_cpus);
1569 invlcache_handler(void)
1571 uint32_t generation;
1574 (*ipi_invlcache_counts[PCPU_GET(cpuid)])++;
1575 #endif /* COUNT_IPIS */
1578 * Reading the generation here allows greater parallelism
1579 * since wbinvd is a serializing instruction. Without the
1580 * temporary, we'd wait for wbinvd to complete, then the read
1581 * would execute, then the dependent write, which must then
1582 * complete before return from interrupt.
1584 generation = smp_tlb_generation;
1586 PCPU_SET(smp_tlb_done, generation);
1590 * This is called once the rest of the system is up and running and we're
1591 * ready to let the AP's out of the pen.
1594 release_aps(void *dummy __unused)
1599 atomic_store_rel_int(&aps_ready, 1);
1600 while (smp_started == 0)
1603 SYSINIT(start_aps, SI_SUB_SMP, SI_ORDER_FIRST, release_aps, NULL);
1607 * Setup interrupt counters for IPI handlers.
1610 mp_ipi_intrcnt(void *dummy)
1616 snprintf(buf, sizeof(buf), "cpu%d:invltlb", i);
1617 intrcnt_add(buf, &ipi_invltlb_counts[i]);
1618 snprintf(buf, sizeof(buf), "cpu%d:invlrng", i);
1619 intrcnt_add(buf, &ipi_invlrng_counts[i]);
1620 snprintf(buf, sizeof(buf), "cpu%d:invlpg", i);
1621 intrcnt_add(buf, &ipi_invlpg_counts[i]);
1622 snprintf(buf, sizeof(buf), "cpu%d:invlcache", i);
1623 intrcnt_add(buf, &ipi_invlcache_counts[i]);
1624 snprintf(buf, sizeof(buf), "cpu%d:preempt", i);
1625 intrcnt_add(buf, &ipi_preempt_counts[i]);
1626 snprintf(buf, sizeof(buf), "cpu%d:ast", i);
1627 intrcnt_add(buf, &ipi_ast_counts[i]);
1628 snprintf(buf, sizeof(buf), "cpu%d:rendezvous", i);
1629 intrcnt_add(buf, &ipi_rendezvous_counts[i]);
1630 snprintf(buf, sizeof(buf), "cpu%d:hardclock", i);
1631 intrcnt_add(buf, &ipi_hardclock_counts[i]);
1634 SYSINIT(mp_ipi_intrcnt, SI_SUB_INTR, SI_ORDER_MIDDLE, mp_ipi_intrcnt, NULL);
1638 * Flush the TLB on other CPU's
1641 /* Variables needed for SMP tlb shootdown. */
1642 vm_offset_t smp_tlb_addr1, smp_tlb_addr2;
1643 pmap_t smp_tlb_pmap;
1644 volatile uint32_t smp_tlb_generation;
1647 #define read_eflags() read_rflags()
1651 smp_targeted_tlb_shootdown(cpuset_t mask, u_int vector, pmap_t pmap,
1652 vm_offset_t addr1, vm_offset_t addr2)
1654 cpuset_t other_cpus;
1655 volatile uint32_t *p_cpudone;
1656 uint32_t generation;
1659 /* It is not necessary to signal other CPUs while in the debugger. */
1660 if (kdb_active || panicstr != NULL)
1664 * Check for other cpus. Return if none.
1666 if (CPU_ISFULLSET(&mask)) {
1670 CPU_CLR(PCPU_GET(cpuid), &mask);
1671 if (CPU_EMPTY(&mask))
1675 if (!(read_eflags() & PSL_I))
1676 panic("%s: interrupts disabled", __func__);
1677 mtx_lock_spin(&smp_ipi_mtx);
1678 smp_tlb_addr1 = addr1;
1679 smp_tlb_addr2 = addr2;
1680 smp_tlb_pmap = pmap;
1681 generation = ++smp_tlb_generation;
1682 if (CPU_ISFULLSET(&mask)) {
1683 ipi_all_but_self(vector);
1684 other_cpus = all_cpus;
1685 CPU_CLR(PCPU_GET(cpuid), &other_cpus);
1688 while ((cpu = CPU_FFS(&mask)) != 0) {
1690 CPU_CLR(cpu, &mask);
1691 CTR3(KTR_SMP, "%s: cpu: %d ipi: %x", __func__,
1693 ipi_send_cpu(cpu, vector);
1696 while ((cpu = CPU_FFS(&other_cpus)) != 0) {
1698 CPU_CLR(cpu, &other_cpus);
1699 p_cpudone = &cpuid_to_pcpu[cpu]->pc_smp_tlb_done;
1700 while (*p_cpudone != generation)
1703 mtx_unlock_spin(&smp_ipi_mtx);
1707 smp_masked_invltlb(cpuset_t mask, pmap_t pmap)
1711 smp_targeted_tlb_shootdown(mask, IPI_INVLTLB, pmap, 0, 0);
1712 #ifdef COUNT_XINVLTLB_HITS
1719 smp_masked_invlpg(cpuset_t mask, vm_offset_t addr, pmap_t pmap)
1723 smp_targeted_tlb_shootdown(mask, IPI_INVLPG, pmap, addr, 0);
1724 #ifdef COUNT_XINVLTLB_HITS
1731 smp_masked_invlpg_range(cpuset_t mask, vm_offset_t addr1, vm_offset_t addr2,
1736 smp_targeted_tlb_shootdown(mask, IPI_INVLRNG, pmap,
1738 #ifdef COUNT_XINVLTLB_HITS
1740 ipi_range_size += (addr2 - addr1) / PAGE_SIZE;
1746 smp_cache_flush(void)
1750 smp_targeted_tlb_shootdown(all_cpus, IPI_INVLCACHE, NULL,
1756 * Handlers for TLB related IPIs
1759 invltlb_handler(void)
1761 uint32_t generation;
1763 #ifdef COUNT_XINVLTLB_HITS
1764 xhits_gbl[PCPU_GET(cpuid)]++;
1765 #endif /* COUNT_XINVLTLB_HITS */
1767 (*ipi_invltlb_counts[PCPU_GET(cpuid)])++;
1768 #endif /* COUNT_IPIS */
1771 * Reading the generation here allows greater parallelism
1772 * since invalidating the TLB is a serializing operation.
1774 generation = smp_tlb_generation;
1775 if (smp_tlb_pmap == kernel_pmap)
1781 PCPU_SET(smp_tlb_done, generation);
1785 invlpg_handler(void)
1787 uint32_t generation;
1789 #ifdef COUNT_XINVLTLB_HITS
1790 xhits_pg[PCPU_GET(cpuid)]++;
1791 #endif /* COUNT_XINVLTLB_HITS */
1793 (*ipi_invlpg_counts[PCPU_GET(cpuid)])++;
1794 #endif /* COUNT_IPIS */
1796 generation = smp_tlb_generation; /* Overlap with serialization */
1798 if (smp_tlb_pmap == kernel_pmap)
1800 invlpg(smp_tlb_addr1);
1801 PCPU_SET(smp_tlb_done, generation);
1805 invlrng_handler(void)
1807 vm_offset_t addr, addr2;
1808 uint32_t generation;
1810 #ifdef COUNT_XINVLTLB_HITS
1811 xhits_rng[PCPU_GET(cpuid)]++;
1812 #endif /* COUNT_XINVLTLB_HITS */
1814 (*ipi_invlrng_counts[PCPU_GET(cpuid)])++;
1815 #endif /* COUNT_IPIS */
1817 addr = smp_tlb_addr1;
1818 addr2 = smp_tlb_addr2;
1819 generation = smp_tlb_generation; /* Overlap with serialization */
1821 if (smp_tlb_pmap == kernel_pmap)
1826 } while (addr < addr2);
1828 PCPU_SET(smp_tlb_done, generation);