xref: /freebsd/sys/x86/x86/mp_x86.c (revision 0b32ef71f9f154f4da5037bfcbb4916960d38452)
1 /*-
2  * Copyright (c) 1996, by Steve Passe
3  * Copyright (c) 2003, by Peter Wemm
4  * All rights reserved.
5  *
6  * Redistribution and use in source and binary forms, with or without
7  * modification, are permitted provided that the following conditions
8  * are met:
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.
13  *
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
24  * SUCH DAMAGE.
25  */
26 
27 #include <sys/cdefs.h>
28 #include "opt_acpi.h"
29 #ifdef __i386__
30 #include "opt_apic.h"
31 #endif
32 #include "opt_cpu.h"
33 #include "opt_ddb.h"
34 #include "opt_gdb.h"
35 #include "opt_kstack_pages.h"
36 #include "opt_pmap.h"
37 #include "opt_sched.h"
38 #include "opt_smp.h"
39 #include "opt_stack.h"
40 
41 #include <sys/param.h>
42 #include <sys/systm.h>
43 #include <sys/asan.h>
44 #include <sys/bus.h>
45 #include <sys/cons.h>	/* cngetc() */
46 #include <sys/cpuset.h>
47 #include <sys/csan.h>
48 #include <sys/interrupt.h>
49 #include <sys/kdb.h>
50 #include <sys/kernel.h>
51 #include <sys/ktr.h>
52 #include <sys/lock.h>
53 #include <sys/malloc.h>
54 #include <sys/memrange.h>
55 #include <sys/mutex.h>
56 #include <sys/pcpu.h>
57 #include <sys/proc.h>
58 #include <sys/sched.h>
59 #include <sys/smp.h>
60 #include <sys/sysctl.h>
61 
62 #include <vm/vm.h>
63 #include <vm/vm_param.h>
64 #include <vm/pmap.h>
65 #include <vm/vm_kern.h>
66 #include <vm/vm_extern.h>
67 #include <vm/vm_map.h>
68 
69 #include <x86/apicreg.h>
70 #include <machine/clock.h>
71 #include <machine/cpu.h>
72 #include <machine/cputypes.h>
73 #include <x86/mca.h>
74 #include <machine/md_var.h>
75 #include <machine/pcb.h>
76 #include <machine/psl.h>
77 #include <machine/smp.h>
78 #include <machine/specialreg.h>
79 #include <machine/stack.h>
80 #include <x86/ucode.h>
81 
82 #ifdef DEV_ACPI
83 #include <contrib/dev/acpica/include/acpi.h>
84 #include <dev/acpica/acpivar.h>
85 #endif
86 
87 static MALLOC_DEFINE(M_CPUS, "cpus", "CPU items");
88 
89 int	mp_naps;		/* # of Applications processors */
90 int	boot_cpu_id = -1;	/* designated BSP */
91 
92 /* AP uses this during bootstrap.  Do not staticize.  */
93 char *bootSTK;
94 int bootAP;
95 
96 /* Free these after use */
97 void *bootstacks[MAXCPU];
98 void *dpcpu;
99 
100 struct susppcb **susppcbs;
101 
102 #ifdef COUNT_IPIS
103 /* Interrupt counts. */
104 static u_long *ipi_preempt_counts[MAXCPU];
105 static u_long *ipi_ast_counts[MAXCPU];
106 u_long *ipi_invltlb_counts[MAXCPU];
107 u_long *ipi_invlrng_counts[MAXCPU];
108 u_long *ipi_invlpg_counts[MAXCPU];
109 u_long *ipi_invlcache_counts[MAXCPU];
110 u_long *ipi_rendezvous_counts[MAXCPU];
111 static u_long *ipi_hardclock_counts[MAXCPU];
112 #endif
113 
114 /* Default cpu_ops implementation. */
115 struct cpu_ops cpu_ops;
116 
117 /*
118  * Local data and functions.
119  */
120 
121 static volatile cpuset_t ipi_stop_nmi_pending;
122 
123 volatile cpuset_t resuming_cpus;
124 volatile cpuset_t toresume_cpus;
125 
126 /* used to hold the AP's until we are ready to release them */
127 struct mtx ap_boot_mtx;
128 
129 /* Set to 1 once we're ready to let the APs out of the pen. */
130 volatile int aps_ready = 0;
131 
132 /*
133  * Store data from cpu_add() until later in the boot when we actually setup
134  * the APs.
135  */
136 struct cpu_info *cpu_info;
137 int *apic_cpuids;
138 int cpu_apic_ids[MAXCPU];
139 _Static_assert(MAXCPU <= MAX_APIC_ID,
140     "MAXCPU cannot be larger that MAX_APIC_ID");
141 _Static_assert(xAPIC_MAX_APIC_ID <= MAX_APIC_ID,
142     "xAPIC_MAX_APIC_ID cannot be larger that MAX_APIC_ID");
143 
144 static void	release_aps(void *dummy);
145 static void	cpustop_handler_post(u_int cpu);
146 
147 static int	hyperthreading_allowed = 1;
148 SYSCTL_INT(_machdep, OID_AUTO, hyperthreading_allowed, CTLFLAG_RDTUN,
149 	&hyperthreading_allowed, 0, "Use Intel HTT logical CPUs");
150 
151 static int	hyperthreading_intr_allowed = 0;
152 SYSCTL_INT(_machdep, OID_AUTO, hyperthreading_intr_allowed, CTLFLAG_RDTUN,
153 	&hyperthreading_intr_allowed, 0,
154 	"Allow interrupts on HTT logical CPUs");
155 
156 static int	intr_apic_id_limit = -1;
157 SYSCTL_INT(_machdep, OID_AUTO, intr_apic_id_limit, CTLFLAG_RDTUN,
158 	&intr_apic_id_limit, 0,
159 	"Maximum permitted APIC ID for interrupt delivery (-1 is unlimited)");
160 
161 static struct topo_node topo_root;
162 
163 static int pkg_id_shift;
164 static int node_id_shift;
165 static int core_id_shift;
166 static int disabled_cpus;
167 
168 struct cache_info {
169 	int	id_shift;
170 	int	present;
171 } static caches[MAX_CACHE_LEVELS];
172 
173 static bool stop_mwait = false;
174 SYSCTL_BOOL(_machdep, OID_AUTO, stop_mwait, CTLFLAG_RWTUN, &stop_mwait, 0,
175     "Use MONITOR/MWAIT when stopping CPU, if available");
176 
177 void
mem_range_AP_init(void)178 mem_range_AP_init(void)
179 {
180 
181 	if (mem_range_softc.mr_op && mem_range_softc.mr_op->initAP)
182 		mem_range_softc.mr_op->initAP(&mem_range_softc);
183 }
184 
185 /*
186  * Compute ceil(log2(x)).  Returns -1 if x is zero.
187  */
188 static __inline int
mask_width(u_int x)189 mask_width(u_int x)
190 {
191 
192 	return (x == 0 ? -1 : order_base_2(x));
193 }
194 
195 /*
196  * Add a cache level to the cache topology description.
197  */
198 static int
add_deterministic_cache(int type,int level,int share_count)199 add_deterministic_cache(int type, int level, int share_count)
200 {
201 
202 	if (type == 0)
203 		return (0);
204 	if (type > 3) {
205 		printf("unexpected cache type %d\n", type);
206 		return (1);
207 	}
208 	if (type == 2) /* ignore instruction cache */
209 		return (1);
210 	if (level == 0 || level > MAX_CACHE_LEVELS) {
211 		printf("unexpected cache level %d\n", level);
212 		return (1);
213 	}
214 
215 	if (caches[level - 1].present) {
216 		printf("WARNING: multiple entries for L%u data cache\n", level);
217 		printf("%u => %u\n", caches[level - 1].id_shift,
218 		    mask_width(share_count));
219 	}
220 	caches[level - 1].id_shift = mask_width(share_count);
221 	caches[level - 1].present = 1;
222 
223 	if (caches[level - 1].id_shift > pkg_id_shift) {
224 		printf("WARNING: L%u data cache covers more "
225 		    "APIC IDs than a package (%u > %u)\n", level,
226 		    caches[level - 1].id_shift, pkg_id_shift);
227 		caches[level - 1].id_shift = pkg_id_shift;
228 	}
229 	if (caches[level - 1].id_shift < core_id_shift) {
230 		printf("WARNING: L%u data cache covers fewer "
231 		    "APIC IDs than a core (%u < %u)\n", level,
232 		    caches[level - 1].id_shift, core_id_shift);
233 		caches[level - 1].id_shift = core_id_shift;
234 	}
235 
236 	return (1);
237 }
238 
239 /*
240  * Determine topology of processing units and caches for AMD CPUs.
241  * See:
242  *  - AMD CPUID Specification (Publication # 25481)
243  *  - BKDG for AMD NPT Family 0Fh Processors (Publication # 32559)
244  *  - BKDG For AMD Family 10h Processors (Publication # 31116)
245  *  - BKDG For AMD Family 15h Models 00h-0Fh Processors (Publication # 42301)
246  *  - BKDG For AMD Family 16h Models 00h-0Fh Processors (Publication # 48751)
247  *  - PPR For AMD Family 17h Models 00h-0Fh Processors (Publication # 54945)
248  */
249 static void
topo_probe_amd(void)250 topo_probe_amd(void)
251 {
252 	u_int p[4];
253 	uint64_t v;
254 	int level;
255 	int nodes_per_socket;
256 	int share_count;
257 	int type;
258 	int i;
259 
260 	/* No multi-core capability. */
261 	if ((amd_feature2 & AMDID2_CMP) == 0)
262 		return;
263 
264 	/*
265 	 * XXX Lack of an AMD IOMMU driver prevents use of APIC IDs above
266 	 * xAPIC_MAX_APIC_ID.  This is a workaround so we boot and function on
267 	 * AMD systems with high thread counts, albeit with reduced interrupt
268 	 * performance.
269 	 *
270 	 * We should really set the limit to xAPIC_MAX_APIC_ID by default, and
271 	 * have the IOMMU driver increase it.  That way if a driver is present
272 	 * but disabled, or is otherwise not able to route the interrupts, the
273 	 * system can fall back to a functional state.  That will require a more
274 	 * substantial change though, including having the IOMMU initialize
275 	 * earlier.
276 	 */
277 	if (intr_apic_id_limit == -1)
278 		intr_apic_id_limit = xAPIC_MAX_APIC_ID;
279 
280 	/* For families 10h and newer. */
281 	pkg_id_shift = (cpu_procinfo2 & AMDID_COREID_SIZE) >>
282 	    AMDID_COREID_SIZE_SHIFT;
283 
284 	/* For 0Fh family. */
285 	if (pkg_id_shift == 0)
286 		pkg_id_shift =
287 		    mask_width((cpu_procinfo2 & AMDID_CMP_CORES) + 1);
288 
289 	/*
290 	 * Families prior to 16h define the following value as
291 	 * cores per compute unit and we don't really care about the AMD
292 	 * compute units at the moment.  Perhaps we should treat them as
293 	 * cores and cores within the compute units as hardware threads,
294 	 * but that's up for debate.
295 	 * Later families define the value as threads per compute unit,
296 	 * so we are following AMD's nomenclature here.
297 	 */
298 	if ((amd_feature2 & AMDID2_TOPOLOGY) != 0 &&
299 	    CPUID_TO_FAMILY(cpu_id) >= 0x16) {
300 		cpuid_count(0x8000001e, 0, p);
301 		share_count = ((p[1] >> 8) & 0xff) + 1;
302 		core_id_shift = mask_width(share_count);
303 
304 		/*
305 		 * For Zen (17h), gather Nodes per Processor.  Each node is a
306 		 * Zeppelin die; TR and EPYC CPUs will have multiple dies per
307 		 * package.  Communication latency between dies is higher than
308 		 * within them.
309 		 */
310 		nodes_per_socket = ((p[2] >> 8) & 0x7) + 1;
311 		node_id_shift = pkg_id_shift - mask_width(nodes_per_socket);
312 	}
313 
314 	if ((amd_feature2 & AMDID2_TOPOLOGY) != 0) {
315 		for (i = 0; ; i++) {
316 			cpuid_count(0x8000001d, i, p);
317 			type = p[0] & 0x1f;
318 			level = (p[0] >> 5) & 0x7;
319 			share_count = 1 + ((p[0] >> 14) & 0xfff);
320 
321 			if (!add_deterministic_cache(type, level, share_count))
322 				break;
323 		}
324 	} else {
325 		if (cpu_exthigh >= 0x80000005) {
326 			cpuid_count(0x80000005, 0, p);
327 			if (((p[2] >> 24) & 0xff) != 0) {
328 				caches[0].id_shift = 0;
329 				caches[0].present = 1;
330 			}
331 		}
332 		if (cpu_exthigh >= 0x80000006) {
333 			cpuid_count(0x80000006, 0, p);
334 			if (((p[2] >> 16) & 0xffff) != 0) {
335 				caches[1].id_shift = 0;
336 				caches[1].present = 1;
337 			}
338 			if (((p[3] >> 18) & 0x3fff) != 0) {
339 				nodes_per_socket = 1;
340 				if ((amd_feature2 & AMDID2_NODE_ID) != 0) {
341 					/*
342 					 * Handle multi-node processors that
343 					 * have multiple chips, each with its
344 					 * own L3 cache, on the same die.
345 					 */
346 					v = rdmsr(0xc001100c);
347 					nodes_per_socket = 1 + ((v >> 3) & 0x7);
348 				}
349 				caches[2].id_shift =
350 				    pkg_id_shift - mask_width(nodes_per_socket);
351 				caches[2].present = 1;
352 			}
353 		}
354 	}
355 }
356 
357 /*
358  * Determine topology of processing units for Intel CPUs
359  * using CPUID Leaf 1 and Leaf 4, if supported.
360  * See:
361  *  - Intel 64 Architecture Processor Topology Enumeration
362  *  - Intel 64 and IA-32 ArchitecturesSoftware Developer’s Manual,
363  *    Volume 3A: System Programming Guide, PROGRAMMING CONSIDERATIONS
364  *    FOR HARDWARE MULTI-THREADING CAPABLE PROCESSORS
365  */
366 static void
topo_probe_intel_0x4(void)367 topo_probe_intel_0x4(void)
368 {
369 	u_int p[4];
370 	int max_cores;
371 	int max_logical;
372 
373 	/* Both zero and one here mean one logical processor per package. */
374 	max_logical = (cpu_feature & CPUID_HTT) != 0 ?
375 	    (cpu_procinfo & CPUID_HTT_CORES) >> 16 : 1;
376 	if (max_logical <= 1)
377 		return;
378 
379 	if (cpu_high >= 0x4) {
380 		cpuid_count(0x04, 0, p);
381 		max_cores = ((p[0] >> 26) & 0x3f) + 1;
382 	} else
383 		max_cores = 1;
384 
385 	core_id_shift = mask_width(max_logical/max_cores);
386 	KASSERT(core_id_shift >= 0,
387 	    ("intel topo: max_cores > max_logical\n"));
388 	pkg_id_shift = core_id_shift + mask_width(max_cores);
389 }
390 
391 /*
392  * Determine topology of processing units for Intel CPUs
393  * using CPUID Leaf 1Fh or 0Bh, if supported.
394  * See:
395  *  - Intel 64 Architecture Processor Topology Enumeration
396  *  - Intel 64 and IA-32 ArchitecturesSoftware Developer’s Manual,
397  *    Volume 3A: System Programming Guide, PROGRAMMING CONSIDERATIONS
398  *    FOR HARDWARE MULTI-THREADING CAPABLE PROCESSORS
399  */
400 static void
topo_probe_intel_0xb(void)401 topo_probe_intel_0xb(void)
402 {
403 	u_int leaf;
404 	u_int p[4] = { 0 };
405 	int bits;
406 	int type;
407 	int i;
408 
409 	/* Prefer leaf 1Fh (V2 Extended Topology Enumeration). */
410 	if (cpu_high >= 0x1f) {
411 		leaf = 0x1f;
412 		cpuid_count(leaf, 0, p);
413 	}
414 	/* Fall back to leaf 0Bh (Extended Topology Enumeration). */
415 	if (p[1] == 0) {
416 		leaf = 0x0b;
417 		cpuid_count(leaf, 0, p);
418 	}
419 	/* Fall back to leaf 04h (Deterministic Cache Parameters). */
420 	if (p[1] == 0) {
421 		topo_probe_intel_0x4();
422 		return;
423 	}
424 
425 	/* We only support three levels for now. */
426 	for (i = 0; ; i++) {
427 		cpuid_count(leaf, i, p);
428 
429 		bits = p[0] & 0x1f;
430 		type = (p[2] >> 8) & 0xff;
431 
432 		if (type == 0)
433 			break;
434 
435 		if (type == CPUID_TYPE_SMT)
436 			core_id_shift = bits;
437 		else if (type == CPUID_TYPE_CORE)
438 			pkg_id_shift = bits;
439 		else if (bootverbose)
440 			printf("Topology level type %d shift: %d\n", type, bits);
441 	}
442 
443 	if (pkg_id_shift < core_id_shift) {
444 		printf("WARNING: core covers more APIC IDs than a package\n");
445 		core_id_shift = pkg_id_shift;
446 	}
447 }
448 
449 /*
450  * Determine topology of caches for Intel CPUs.
451  * See:
452  *  - Intel 64 Architecture Processor Topology Enumeration
453  *  - Intel 64 and IA-32 Architectures Software Developer’s Manual
454  *    Volume 2A: Instruction Set Reference, A-M,
455  *    CPUID instruction
456  */
457 static void
topo_probe_intel_caches(void)458 topo_probe_intel_caches(void)
459 {
460 	u_int p[4];
461 	int level;
462 	int share_count;
463 	int type;
464 	int i;
465 
466 	if (cpu_high < 0x4) {
467 		/*
468 		 * Available cache level and sizes can be determined
469 		 * via CPUID leaf 2, but that requires a huge table of hardcoded
470 		 * values, so for now just assume L1 and L2 caches potentially
471 		 * shared only by HTT processing units, if HTT is present.
472 		 */
473 		caches[0].id_shift = pkg_id_shift;
474 		caches[0].present = 1;
475 		caches[1].id_shift = pkg_id_shift;
476 		caches[1].present = 1;
477 		return;
478 	}
479 
480 	for (i = 0; ; i++) {
481 		cpuid_count(0x4, i, p);
482 		type = p[0] & 0x1f;
483 		level = (p[0] >> 5) & 0x7;
484 		share_count = 1 + ((p[0] >> 14) & 0xfff);
485 
486 		if (!add_deterministic_cache(type, level, share_count))
487 			break;
488 	}
489 }
490 
491 /*
492  * Determine topology of processing units and caches for Intel CPUs.
493  * See:
494  *  - Intel 64 Architecture Processor Topology Enumeration
495  */
496 static void
topo_probe_intel(void)497 topo_probe_intel(void)
498 {
499 
500 	/*
501 	 * Note that 0x1 <= cpu_high < 4 case should be
502 	 * compatible with topo_probe_intel_0x4() logic when
503 	 * CPUID.1:EBX[23:16] > 0 (cpu_cores will be 1)
504 	 * or it should trigger the fallback otherwise.
505 	 */
506 	if (cpu_high >= 0xb)
507 		topo_probe_intel_0xb();
508 	else if (cpu_high >= 0x1)
509 		topo_probe_intel_0x4();
510 
511 	topo_probe_intel_caches();
512 }
513 
514 /*
515  * Topology information is queried only on BSP, on which this
516  * code runs and for which it can query CPUID information.
517  * Then topology is extrapolated on all packages using an
518  * assumption that APIC ID to hardware component ID mapping is
519  * homogenious.
520  * That doesn't necesserily imply that the topology is uniform.
521  */
522 void
topo_probe(void)523 topo_probe(void)
524 {
525 	static int cpu_topo_probed = 0;
526 	struct x86_topo_layer {
527 		int type;
528 		int subtype;
529 		int id_shift;
530 	} topo_layers[MAX_CACHE_LEVELS + 5];
531 	struct topo_node *parent;
532 	struct topo_node *node;
533 	int layer;
534 	int nlayers;
535 	int node_id;
536 	int i;
537 #if defined(DEV_ACPI) && MAXMEMDOM > 1
538 	int d, domain;
539 #endif
540 
541 	if (cpu_topo_probed)
542 		return;
543 
544 	CPU_ZERO(&logical_cpus_mask);
545 
546 	if (mp_ncpus <= 1)
547 		; /* nothing */
548 	else if (cpu_vendor_id == CPU_VENDOR_AMD ||
549 	    cpu_vendor_id == CPU_VENDOR_HYGON)
550 		topo_probe_amd();
551 	else if (cpu_vendor_id == CPU_VENDOR_INTEL)
552 		topo_probe_intel();
553 
554 	KASSERT(pkg_id_shift >= core_id_shift,
555 	    ("bug in APIC topology discovery"));
556 
557 	nlayers = 0;
558 	bzero(topo_layers, sizeof(topo_layers));
559 
560 	topo_layers[nlayers].type = TOPO_TYPE_PKG;
561 	topo_layers[nlayers].id_shift = pkg_id_shift;
562 	if (bootverbose)
563 		printf("Package ID shift: %u\n", topo_layers[nlayers].id_shift);
564 	nlayers++;
565 
566 	if (pkg_id_shift > node_id_shift && node_id_shift != 0) {
567 		topo_layers[nlayers].type = TOPO_TYPE_GROUP;
568 		topo_layers[nlayers].id_shift = node_id_shift;
569 		if (bootverbose)
570 			printf("Node ID shift: %u\n",
571 			    topo_layers[nlayers].id_shift);
572 		nlayers++;
573 	}
574 
575 	/*
576 	 * Consider all caches to be within a package/chip
577 	 * and "in front" of all sub-components like
578 	 * cores and hardware threads.
579 	 */
580 	for (i = MAX_CACHE_LEVELS - 1; i >= 0; --i) {
581 		if (caches[i].present) {
582 			if (node_id_shift != 0)
583 				KASSERT(caches[i].id_shift <= node_id_shift,
584 					("bug in APIC topology discovery"));
585 			KASSERT(caches[i].id_shift <= pkg_id_shift,
586 				("bug in APIC topology discovery"));
587 			KASSERT(caches[i].id_shift >= core_id_shift,
588 				("bug in APIC topology discovery"));
589 
590 			topo_layers[nlayers].type = TOPO_TYPE_CACHE;
591 			topo_layers[nlayers].subtype = i + 1;
592 			topo_layers[nlayers].id_shift = caches[i].id_shift;
593 			if (bootverbose)
594 				printf("L%u cache ID shift: %u\n",
595 				    topo_layers[nlayers].subtype,
596 				    topo_layers[nlayers].id_shift);
597 			nlayers++;
598 		}
599 	}
600 
601 	if (pkg_id_shift > core_id_shift) {
602 		topo_layers[nlayers].type = TOPO_TYPE_CORE;
603 		topo_layers[nlayers].id_shift = core_id_shift;
604 		if (bootverbose)
605 			printf("Core ID shift: %u\n",
606 			    topo_layers[nlayers].id_shift);
607 		nlayers++;
608 	}
609 
610 	topo_layers[nlayers].type = TOPO_TYPE_PU;
611 	topo_layers[nlayers].id_shift = 0;
612 	nlayers++;
613 
614 #if defined(DEV_ACPI) && MAXMEMDOM > 1
615 	if (vm_ndomains > 1) {
616 		for (layer = 0; layer < nlayers; ++layer) {
617 			for (i = 0; i <= max_apic_id; ++i) {
618 				if ((i & ((1 << topo_layers[layer].id_shift) - 1)) == 0)
619 					domain = -1;
620 				if (!cpu_info[i].cpu_present)
621 					continue;
622 				d = acpi_pxm_get_cpu_locality(i);
623 				if (domain >= 0 && domain != d)
624 					break;
625 				domain = d;
626 			}
627 			if (i > max_apic_id)
628 				break;
629 		}
630 		KASSERT(layer < nlayers, ("NUMA domain smaller than PU"));
631 		memmove(&topo_layers[layer+1], &topo_layers[layer],
632 		    sizeof(*topo_layers) * (nlayers - layer));
633 		topo_layers[layer].type = TOPO_TYPE_NODE;
634 		topo_layers[layer].subtype = CG_SHARE_NONE;
635 		nlayers++;
636 	}
637 #endif
638 
639 	topo_init_root(&topo_root);
640 	for (i = 0; i <= max_apic_id; ++i) {
641 		if (!cpu_info[i].cpu_present)
642 			continue;
643 
644 		parent = &topo_root;
645 		for (layer = 0; layer < nlayers; ++layer) {
646 #if defined(DEV_ACPI) && MAXMEMDOM > 1
647 			if (topo_layers[layer].type == TOPO_TYPE_NODE) {
648 				node_id = acpi_pxm_get_cpu_locality(i);
649 			} else
650 #endif
651 				node_id = i >> topo_layers[layer].id_shift;
652 			parent = topo_add_node_by_hwid(parent, node_id,
653 			    topo_layers[layer].type,
654 			    topo_layers[layer].subtype);
655 		}
656 	}
657 
658 	parent = &topo_root;
659 	for (layer = 0; layer < nlayers; ++layer) {
660 #if defined(DEV_ACPI) && MAXMEMDOM > 1
661 		if (topo_layers[layer].type == TOPO_TYPE_NODE)
662 			node_id = acpi_pxm_get_cpu_locality(boot_cpu_id);
663 		else
664 #endif
665 			node_id = boot_cpu_id >> topo_layers[layer].id_shift;
666 		node = topo_find_node_by_hwid(parent, node_id,
667 		    topo_layers[layer].type,
668 		    topo_layers[layer].subtype);
669 		topo_promote_child(node);
670 		parent = node;
671 	}
672 
673 	cpu_topo_probed = 1;
674 }
675 
676 /*
677  * Assign logical CPU IDs to local APICs.
678  */
679 void
assign_cpu_ids(void)680 assign_cpu_ids(void)
681 {
682 	struct topo_node *node;
683 	u_int smt_mask;
684 	int nhyper;
685 
686 	smt_mask = (1u << core_id_shift) - 1;
687 
688 	/*
689 	 * Assign CPU IDs to local APIC IDs and disable any CPUs
690 	 * beyond MAXCPU.  CPU 0 is always assigned to the BSP.
691 	 */
692 	mp_ncpus = 0;
693 	nhyper = 0;
694 	TOPO_FOREACH(node, &topo_root) {
695 		if (node->type != TOPO_TYPE_PU)
696 			continue;
697 
698 		if ((node->hwid & smt_mask) != (boot_cpu_id & smt_mask))
699 			cpu_info[node->hwid].cpu_hyperthread = 1;
700 
701 		if (resource_disabled("lapic", node->hwid)) {
702 			if (node->hwid != boot_cpu_id)
703 				cpu_info[node->hwid].cpu_disabled = 1;
704 			else
705 				printf("Cannot disable BSP, APIC ID = %d\n",
706 				    node->hwid);
707 		}
708 
709 		if (!hyperthreading_allowed &&
710 		    cpu_info[node->hwid].cpu_hyperthread)
711 			cpu_info[node->hwid].cpu_disabled = 1;
712 
713 		if (mp_ncpus >= MAXCPU)
714 			cpu_info[node->hwid].cpu_disabled = 1;
715 
716 		if (cpu_info[node->hwid].cpu_disabled) {
717 			disabled_cpus++;
718 			continue;
719 		}
720 
721 		if (cpu_info[node->hwid].cpu_hyperthread)
722 			nhyper++;
723 
724 		cpu_apic_ids[mp_ncpus] = node->hwid;
725 		apic_cpuids[node->hwid] = mp_ncpus;
726 		topo_set_pu_id(node, mp_ncpus);
727 		mp_ncpus++;
728 	}
729 
730 	KASSERT(mp_maxid >= mp_ncpus - 1,
731 	    ("%s: counters out of sync: max %d, count %d", __func__, mp_maxid,
732 	    mp_ncpus));
733 
734 	mp_ncores = mp_ncpus - nhyper;
735 	smp_threads_per_core = mp_ncpus / mp_ncores;
736 }
737 
738 /*
739  * Print various information about the SMP system hardware and setup.
740  */
741 void
cpu_mp_announce(void)742 cpu_mp_announce(void)
743 {
744 	struct topo_node *node;
745 	const char *hyperthread;
746 	struct topo_analysis topology;
747 
748 	printf("FreeBSD/SMP: ");
749 	if (topo_analyze(&topo_root, 1, &topology)) {
750 		printf("%d package(s)", topology.entities[TOPO_LEVEL_PKG]);
751 		if (topology.entities[TOPO_LEVEL_GROUP] > 1)
752 			printf(" x %d groups",
753 			    topology.entities[TOPO_LEVEL_GROUP]);
754 		if (topology.entities[TOPO_LEVEL_CACHEGROUP] > 1)
755 			printf(" x %d cache groups",
756 			    topology.entities[TOPO_LEVEL_CACHEGROUP]);
757 		if (topology.entities[TOPO_LEVEL_CORE] > 0)
758 			printf(" x %d core(s)",
759 			    topology.entities[TOPO_LEVEL_CORE]);
760 		if (topology.entities[TOPO_LEVEL_THREAD] > 1)
761 			printf(" x %d hardware threads",
762 			    topology.entities[TOPO_LEVEL_THREAD]);
763 	} else {
764 		printf("Non-uniform topology");
765 	}
766 	printf("\n");
767 
768 	if (disabled_cpus) {
769 		printf("FreeBSD/SMP Online: ");
770 		if (topo_analyze(&topo_root, 0, &topology)) {
771 			printf("%d package(s)",
772 			    topology.entities[TOPO_LEVEL_PKG]);
773 			if (topology.entities[TOPO_LEVEL_GROUP] > 1)
774 				printf(" x %d groups",
775 				    topology.entities[TOPO_LEVEL_GROUP]);
776 			if (topology.entities[TOPO_LEVEL_CACHEGROUP] > 1)
777 				printf(" x %d cache groups",
778 				    topology.entities[TOPO_LEVEL_CACHEGROUP]);
779 			if (topology.entities[TOPO_LEVEL_CORE] > 0)
780 				printf(" x %d core(s)",
781 				    topology.entities[TOPO_LEVEL_CORE]);
782 			if (topology.entities[TOPO_LEVEL_THREAD] > 1)
783 				printf(" x %d hardware threads",
784 				    topology.entities[TOPO_LEVEL_THREAD]);
785 		} else {
786 			printf("Non-uniform topology");
787 		}
788 		printf("\n");
789 	}
790 
791 	if (!bootverbose)
792 		return;
793 
794 	TOPO_FOREACH(node, &topo_root) {
795 		switch (node->type) {
796 		case TOPO_TYPE_PKG:
797 			printf("Package HW ID = %u\n", node->hwid);
798 			break;
799 		case TOPO_TYPE_CORE:
800 			printf("\tCore HW ID = %u\n", node->hwid);
801 			break;
802 		case TOPO_TYPE_PU:
803 			if (cpu_info[node->hwid].cpu_hyperthread)
804 				hyperthread = "/HT";
805 			else
806 				hyperthread = "";
807 
808 			if (node->subtype == 0)
809 				printf("\t\tCPU (AP%s): APIC ID: %u"
810 				    "(disabled)\n", hyperthread, node->hwid);
811 			else if (node->id == 0)
812 				printf("\t\tCPU0 (BSP): APIC ID: %u\n",
813 				    node->hwid);
814 			else
815 				printf("\t\tCPU%u (AP%s): APIC ID: %u\n",
816 				    node->id, hyperthread, node->hwid);
817 			break;
818 		default:
819 			/* ignored */
820 			break;
821 		}
822 	}
823 }
824 
825 /*
826  * Add a scheduling group, a group of logical processors sharing
827  * a particular cache (and, thus having an affinity), to the scheduling
828  * topology.
829  * This function recursively works on lower level caches.
830  */
831 static void
x86topo_add_sched_group(struct topo_node * root,struct cpu_group * cg_root)832 x86topo_add_sched_group(struct topo_node *root, struct cpu_group *cg_root)
833 {
834 	struct topo_node *node;
835 	int nchildren;
836 	int ncores;
837 	int i;
838 
839 	KASSERT(root->type == TOPO_TYPE_SYSTEM || root->type == TOPO_TYPE_CACHE ||
840 	    root->type == TOPO_TYPE_NODE || root->type == TOPO_TYPE_GROUP,
841 	    ("x86topo_add_sched_group: bad type: %u", root->type));
842 	CPU_COPY(&root->cpuset, &cg_root->cg_mask);
843 	cg_root->cg_count = root->cpu_count;
844 	if (root->type == TOPO_TYPE_CACHE)
845 		cg_root->cg_level = root->subtype;
846 	else
847 		cg_root->cg_level = CG_SHARE_NONE;
848 	if (root->type == TOPO_TYPE_NODE)
849 		cg_root->cg_flags = CG_FLAG_NODE;
850 	else
851 		cg_root->cg_flags = 0;
852 
853 	/*
854 	 * Check how many core nodes we have under the given root node.
855 	 * If we have multiple logical processors, but not multiple
856 	 * cores, then those processors must be hardware threads.
857 	 */
858 	ncores = 0;
859 	node = root;
860 	while (node != NULL) {
861 		if (node->type != TOPO_TYPE_CORE) {
862 			node = topo_next_node(root, node);
863 			continue;
864 		}
865 
866 		ncores++;
867 		node = topo_next_nonchild_node(root, node);
868 	}
869 
870 	if (cg_root->cg_level != CG_SHARE_NONE &&
871 	    root->cpu_count > 1 && ncores < 2)
872 		cg_root->cg_flags |= CG_FLAG_SMT;
873 
874 	/*
875 	 * Find out how many cache nodes we have under the given root node.
876 	 * We ignore cache nodes that cover all the same processors as the
877 	 * root node.  Also, we do not descend below found cache nodes.
878 	 * That is, we count top-level "non-redundant" caches under the root
879 	 * node.
880 	 */
881 	nchildren = 0;
882 	node = root;
883 	while (node != NULL) {
884 		/*
885 		 * When some APICs are disabled by tunables, nodes can end up
886 		 * with an empty cpuset. Nodes with an empty cpuset will be
887 		 * translated into cpu groups with empty cpusets. smp_topo_fill
888 		 * will then set cg_first and cg_last to -1. This isn't
889 		 * correctly handled in all functions. E.g. when
890 		 * cpu_search_lowest and cpu_search_highest loop through all
891 		 * cpus, they call CPU_ISSET on cpu -1 which ends up in a
892 		 * general protection fault.
893 		 *
894 		 * We could fix the scheduler to handle empty cpu groups
895 		 * correctly. Nevertheless, empty cpu groups are causing
896 		 * overhead for no value. So, it makes more sense to just don't
897 		 * create them.
898 		 */
899 		if (CPU_EMPTY(&node->cpuset)) {
900 			node = topo_next_node(root, node);
901 			continue;
902 		}
903 		if (CPU_CMP(&node->cpuset, &root->cpuset) == 0) {
904 			if (node->type == TOPO_TYPE_CACHE &&
905 			    cg_root->cg_level < node->subtype)
906 				cg_root->cg_level = node->subtype;
907 			if (node->type == TOPO_TYPE_NODE)
908 				cg_root->cg_flags |= CG_FLAG_NODE;
909 			node = topo_next_node(root, node);
910 			continue;
911 		}
912 		if (node->type != TOPO_TYPE_GROUP &&
913 		    node->type != TOPO_TYPE_NODE &&
914 		    node->type != TOPO_TYPE_CACHE) {
915 			node = topo_next_node(root, node);
916 			continue;
917 		}
918 		nchildren++;
919 		node = topo_next_nonchild_node(root, node);
920 	}
921 
922 	/*
923 	 * We are not interested in nodes including only one CPU each.
924 	 */
925 	if (nchildren == root->cpu_count)
926 		return;
927 
928 	/*
929 	 * We are not interested in nodes without children.
930 	 */
931 	cg_root->cg_children = nchildren;
932 	if (nchildren == 0)
933 		return;
934 
935 	cg_root->cg_child = smp_topo_alloc(nchildren);
936 
937 	/*
938 	 * Now find again the same cache nodes as above and recursively
939 	 * build scheduling topologies for them.
940 	 */
941 	node = root;
942 	i = 0;
943 	while (node != NULL) {
944 		if ((node->type != TOPO_TYPE_GROUP &&
945 		    node->type != TOPO_TYPE_NODE &&
946 		    node->type != TOPO_TYPE_CACHE) ||
947 		    CPU_CMP(&node->cpuset, &root->cpuset) == 0 ||
948 		    CPU_EMPTY(&node->cpuset)) {
949 			node = topo_next_node(root, node);
950 			continue;
951 		}
952 		cg_root->cg_child[i].cg_parent = cg_root;
953 		x86topo_add_sched_group(node, &cg_root->cg_child[i]);
954 		i++;
955 		node = topo_next_nonchild_node(root, node);
956 	}
957 }
958 
959 /*
960  * Build the MI scheduling topology from the discovered hardware topology.
961  */
962 struct cpu_group *
cpu_topo(void)963 cpu_topo(void)
964 {
965 	struct cpu_group *cg_root;
966 
967 	if (mp_ncpus <= 1)
968 		return (smp_topo_none());
969 
970 	cg_root = smp_topo_alloc(1);
971 	x86topo_add_sched_group(&topo_root, cg_root);
972 	return (cg_root);
973 }
974 
975 static void
cpu_alloc(void * dummy __unused)976 cpu_alloc(void *dummy __unused)
977 {
978 	/*
979 	 * Dynamically allocate the arrays that depend on the
980 	 * maximum APIC ID.
981 	 */
982 	cpu_info = malloc(sizeof(*cpu_info) * (max_apic_id + 1), M_CPUS,
983 	    M_WAITOK | M_ZERO);
984 	apic_cpuids = malloc(sizeof(*apic_cpuids) * (max_apic_id + 1), M_CPUS,
985 	    M_WAITOK | M_ZERO);
986 }
987 SYSINIT(cpu_alloc, SI_SUB_CPU, SI_ORDER_FIRST, cpu_alloc, NULL);
988 
989 /*
990  * Add a logical CPU to the topology.
991  */
992 void
cpu_add(u_int apic_id,char boot_cpu)993 cpu_add(u_int apic_id, char boot_cpu)
994 {
995 
996 	if (apic_id > max_apic_id)
997 		panic("SMP: APIC ID %d too high", apic_id);
998 
999 	KASSERT(cpu_info[apic_id].cpu_present == 0, ("CPU %u added twice",
1000 	    apic_id));
1001 	cpu_info[apic_id].cpu_present = 1;
1002 	if (boot_cpu) {
1003 		KASSERT(boot_cpu_id == -1,
1004 		    ("CPU %u claims to be BSP, but CPU %u already is", apic_id,
1005 		    boot_cpu_id));
1006 		boot_cpu_id = apic_id;
1007 		cpu_info[apic_id].cpu_bsp = 1;
1008 	}
1009 	if (bootverbose)
1010 		printf("SMP: Added CPU %u (%s)\n", apic_id, boot_cpu ? "BSP" :
1011 		    "AP");
1012 }
1013 
1014 void
cpu_mp_setmaxid(void)1015 cpu_mp_setmaxid(void)
1016 {
1017 
1018 	/*
1019 	 * mp_ncpus and mp_maxid should be already set by calls to cpu_add().
1020 	 * If there were no calls to cpu_add() assume this is a UP system.
1021 	 */
1022 	if (mp_ncpus == 0)
1023 		mp_ncpus = 1;
1024 }
1025 
1026 int
cpu_mp_probe(void)1027 cpu_mp_probe(void)
1028 {
1029 
1030 	/*
1031 	 * Always record BSP in CPU map so that the mbuf init code works
1032 	 * correctly.
1033 	 */
1034 	CPU_SETOF(0, &all_cpus);
1035 	return (mp_ncpus > 1);
1036 }
1037 
1038 /*
1039  * AP CPU's call this to initialize themselves.
1040  */
1041 void
init_secondary_tail(void)1042 init_secondary_tail(void)
1043 {
1044 	u_int cpuid;
1045 
1046 	pmap_activate_boot(vmspace_pmap(proc0.p_vmspace));
1047 
1048 	/*
1049 	 * On real hardware, switch to x2apic mode if possible.  Do it
1050 	 * after aps_ready was signalled, to avoid manipulating the
1051 	 * mode while BSP might still want to send some IPI to us
1052 	 * (second startup IPI is ignored on modern hardware etc).
1053 	 */
1054 	lapic_xapic_mode();
1055 
1056 	/* Initialize the PAT MSR. */
1057 	pmap_init_pat();
1058 
1059 	/* set up CPU registers and state */
1060 	cpu_setregs();
1061 
1062 	/* set up SSE/NX */
1063 	initializecpu();
1064 
1065 	/* set up FPU state on the AP */
1066 #ifdef __amd64__
1067 	fpuinit();
1068 #else
1069 	npxinit(false);
1070 #endif
1071 
1072 	if (cpu_ops.cpu_init)
1073 		cpu_ops.cpu_init();
1074 
1075 	/* A quick check from sanity claus */
1076 	cpuid = PCPU_GET(cpuid);
1077 	if (PCPU_GET(apic_id) != lapic_id()) {
1078 		printf("SMP: cpuid = %d\n", cpuid);
1079 		printf("SMP: actual apic_id = %d\n", lapic_id());
1080 		printf("SMP: correct apic_id = %d\n", PCPU_GET(apic_id));
1081 		panic("cpuid mismatch! boom!!");
1082 	}
1083 
1084 	/* Initialize curthread. */
1085 	KASSERT(PCPU_GET(idlethread) != NULL, ("no idle thread"));
1086 	PCPU_SET(curthread, PCPU_GET(idlethread));
1087 	schedinit_ap();
1088 
1089 	mtx_lock_spin(&ap_boot_mtx);
1090 
1091 	mca_init();
1092 
1093 	/* Init local apic for irq's */
1094 	lapic_setup(1);
1095 
1096 	/* Set memory range attributes for this CPU to match the BSP */
1097 	mem_range_AP_init();
1098 
1099 	smp_cpus++;
1100 
1101 	CTR1(KTR_SMP, "SMP: AP CPU #%d Launched", cpuid);
1102 	if (bootverbose)
1103 		printf("SMP: AP CPU #%d Launched!\n", cpuid);
1104 	else
1105 		printf("%s%d%s", smp_cpus == 2 ? "Launching APs: " : "",
1106 		    cpuid, smp_cpus == mp_ncpus ? "\n" : " ");
1107 
1108 	/* Determine if we are a logical CPU. */
1109 	if (cpu_info[PCPU_GET(apic_id)].cpu_hyperthread)
1110 		CPU_SET(cpuid, &logical_cpus_mask);
1111 
1112 	if (bootverbose)
1113 		lapic_dump("AP");
1114 
1115 	if (smp_cpus == mp_ncpus) {
1116 		/* enable IPI's, tlb shootdown, freezes etc */
1117 		atomic_store_rel_int(&smp_started, 1);
1118 	}
1119 
1120 #ifdef __amd64__
1121 	if (pmap_pcid_enabled)
1122 		load_cr4(rcr4() | CR4_PCIDE);
1123 	load_ds(_udatasel);
1124 	load_es(_udatasel);
1125 	load_fs(_ufssel);
1126 #endif
1127 
1128 	mtx_unlock_spin(&ap_boot_mtx);
1129 
1130 	/* Wait until all the AP's are up. */
1131 	while (atomic_load_acq_int(&smp_started) == 0)
1132 		ia32_pause();
1133 
1134 	kcsan_cpu_init(cpuid);
1135 
1136 	sched_ap_entry();
1137 
1138 	panic("scheduler returned us to %s", __func__);
1139 	/* NOTREACHED */
1140 }
1141 
1142 static void
smp_after_idle_runnable(void * arg __unused)1143 smp_after_idle_runnable(void *arg __unused)
1144 {
1145 	int cpu;
1146 
1147 	if (mp_ncpus == 1)
1148 		return;
1149 
1150 	KASSERT(smp_started != 0, ("%s: SMP not started yet", __func__));
1151 
1152 	/*
1153 	 * Wait for all APs to handle an interrupt.  After that, we know that
1154 	 * the APs have entered the scheduler at least once, so the boot stacks
1155 	 * are safe to free.
1156 	 */
1157 	smp_rendezvous(smp_no_rendezvous_barrier, NULL,
1158 	    smp_no_rendezvous_barrier, NULL);
1159 
1160 	for (cpu = 1; cpu < mp_ncpus; cpu++) {
1161 		kmem_free(bootstacks[cpu], kstack_pages * PAGE_SIZE);
1162 	}
1163 }
1164 SYSINIT(smp_after_idle_runnable, SI_SUB_SMP, SI_ORDER_ANY,
1165     smp_after_idle_runnable, NULL);
1166 
1167 /*
1168  * We tell the I/O APIC code about all the CPUs we want to receive
1169  * interrupts.  If we don't want certain CPUs to receive IRQs we
1170  * can simply not tell the I/O APIC code about them in this function.
1171  * We also do not tell it about the BSP since it tells itself about
1172  * the BSP internally to work with UP kernels and on UP machines.
1173  */
1174 void
set_interrupt_apic_ids(void)1175 set_interrupt_apic_ids(void)
1176 {
1177 	u_int i, apic_id;
1178 
1179 	for (i = 0; i < MAXCPU; i++) {
1180 		apic_id = cpu_apic_ids[i];
1181 		if (apic_id == -1)
1182 			continue;
1183 		if (cpu_info[apic_id].cpu_bsp)
1184 			continue;
1185 		if (cpu_info[apic_id].cpu_disabled)
1186 			continue;
1187 		if (intr_apic_id_limit >= 0 && apic_id > intr_apic_id_limit)
1188 			continue;
1189 
1190 		/* Don't let hyperthreads service interrupts. */
1191 		if (cpu_info[apic_id].cpu_hyperthread &&
1192 		    !hyperthreading_intr_allowed)
1193 			continue;
1194 
1195 		intr_add_cpu(i);
1196 	}
1197 }
1198 
1199 #ifdef COUNT_XINVLTLB_HITS
1200 u_int xhits_gbl[MAXCPU];
1201 u_int xhits_pg[MAXCPU];
1202 u_int xhits_rng[MAXCPU];
1203 static SYSCTL_NODE(_debug, OID_AUTO, xhits, CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1204     "");
1205 SYSCTL_OPAQUE(_debug_xhits, OID_AUTO, global, CTLFLAG_RW, &xhits_gbl,
1206     sizeof(xhits_gbl), "IU", "");
1207 SYSCTL_OPAQUE(_debug_xhits, OID_AUTO, page, CTLFLAG_RW, &xhits_pg,
1208     sizeof(xhits_pg), "IU", "");
1209 SYSCTL_OPAQUE(_debug_xhits, OID_AUTO, range, CTLFLAG_RW, &xhits_rng,
1210     sizeof(xhits_rng), "IU", "");
1211 
1212 u_int ipi_global;
1213 u_int ipi_page;
1214 u_int ipi_range;
1215 u_int ipi_range_size;
1216 SYSCTL_INT(_debug_xhits, OID_AUTO, ipi_global, CTLFLAG_RW, &ipi_global, 0, "");
1217 SYSCTL_INT(_debug_xhits, OID_AUTO, ipi_page, CTLFLAG_RW, &ipi_page, 0, "");
1218 SYSCTL_INT(_debug_xhits, OID_AUTO, ipi_range, CTLFLAG_RW, &ipi_range, 0, "");
1219 SYSCTL_INT(_debug_xhits, OID_AUTO, ipi_range_size, CTLFLAG_RW, &ipi_range_size,
1220     0, "");
1221 #endif /* COUNT_XINVLTLB_HITS */
1222 
1223 /*
1224  * Init and startup IPI.
1225  */
1226 void
ipi_startup(int apic_id,int vector)1227 ipi_startup(int apic_id, int vector)
1228 {
1229 
1230 	/*
1231 	 * This attempts to follow the algorithm described in the
1232 	 * Intel Multiprocessor Specification v1.4 in section B.4.
1233 	 * For each IPI, we allow the local APIC ~20us to deliver the
1234 	 * IPI.  If that times out, we panic.
1235 	 */
1236 
1237 	/*
1238 	 * first we do an INIT IPI: this INIT IPI might be run, resetting
1239 	 * and running the target CPU. OR this INIT IPI might be latched (P5
1240 	 * bug), CPU waiting for STARTUP IPI. OR this INIT IPI might be
1241 	 * ignored.
1242 	 */
1243 	lapic_ipi_raw(APIC_DEST_DESTFLD | APIC_TRIGMOD_LEVEL |
1244 	    APIC_LEVEL_ASSERT | APIC_DESTMODE_PHY | APIC_DELMODE_INIT, apic_id);
1245 	lapic_ipi_wait(100);
1246 
1247 	/* Explicitly deassert the INIT IPI. */
1248 	lapic_ipi_raw(APIC_DEST_DESTFLD | APIC_TRIGMOD_LEVEL |
1249 	    APIC_LEVEL_DEASSERT | APIC_DESTMODE_PHY | APIC_DELMODE_INIT,
1250 	    apic_id);
1251 
1252 	DELAY(10000);		/* wait ~10mS */
1253 
1254 	/*
1255 	 * next we do a STARTUP IPI: the previous INIT IPI might still be
1256 	 * latched, (P5 bug) this 1st STARTUP would then terminate
1257 	 * immediately, and the previously started INIT IPI would continue. OR
1258 	 * the previous INIT IPI has already run. and this STARTUP IPI will
1259 	 * run. OR the previous INIT IPI was ignored. and this STARTUP IPI
1260 	 * will run.
1261 	 */
1262 	lapic_ipi_raw(APIC_DEST_DESTFLD | APIC_TRIGMOD_EDGE |
1263 	    APIC_LEVEL_ASSERT | APIC_DESTMODE_PHY | APIC_DELMODE_STARTUP |
1264 	    vector, apic_id);
1265 	if (!lapic_ipi_wait(100))
1266 		panic("Failed to deliver first STARTUP IPI to APIC %d",
1267 		    apic_id);
1268 	DELAY(200);		/* wait ~200uS */
1269 
1270 	/*
1271 	 * finally we do a 2nd STARTUP IPI: this 2nd STARTUP IPI should run IF
1272 	 * the previous STARTUP IPI was cancelled by a latched INIT IPI. OR
1273 	 * this STARTUP IPI will be ignored, as only ONE STARTUP IPI is
1274 	 * recognized after hardware RESET or INIT IPI.
1275 	 */
1276 	lapic_ipi_raw(APIC_DEST_DESTFLD | APIC_TRIGMOD_EDGE |
1277 	    APIC_LEVEL_ASSERT | APIC_DESTMODE_PHY | APIC_DELMODE_STARTUP |
1278 	    vector, apic_id);
1279 	if (!lapic_ipi_wait(100))
1280 		panic("Failed to deliver second STARTUP IPI to APIC %d",
1281 		    apic_id);
1282 
1283 	DELAY(200);		/* wait ~200uS */
1284 }
1285 
1286 static bool
ipi_bitmap_set(int cpu,u_int ipi)1287 ipi_bitmap_set(int cpu, u_int ipi)
1288 {
1289 	u_int bitmap, old, new;
1290 	u_int *cpu_bitmap;
1291 
1292 	bitmap = 1 << ipi;
1293 	cpu_bitmap = &cpuid_to_pcpu[cpu]->pc_ipi_bitmap;
1294 	old = *cpu_bitmap;
1295 	for (;;) {
1296 		if ((old & bitmap) != 0)
1297 			break;
1298 		new = old | bitmap;
1299 		if (atomic_fcmpset_int(cpu_bitmap, &old, new))
1300 			break;
1301 	}
1302 	return (old != 0);
1303 }
1304 
1305 /*
1306  * Send an IPI to specified CPU handling the bitmap logic.
1307  */
1308 static void
ipi_send_cpu(int cpu,u_int ipi)1309 ipi_send_cpu(int cpu, u_int ipi)
1310 {
1311 
1312 	KASSERT((u_int)cpu < MAXCPU && cpu_apic_ids[cpu] != -1,
1313 	    ("IPI to non-existent CPU %d", cpu));
1314 
1315 	if (IPI_IS_BITMAPED(ipi)) {
1316 		if (ipi_bitmap_set(cpu, ipi))
1317 			return;
1318 		ipi = IPI_BITMAP_VECTOR;
1319 	}
1320 	lapic_ipi_vectored(ipi, cpu_apic_ids[cpu]);
1321 }
1322 
1323 void
ipi_bitmap_handler(struct trapframe frame)1324 ipi_bitmap_handler(struct trapframe frame)
1325 {
1326 	struct trapframe *oldframe;
1327 	struct thread *td;
1328 	int cpu = PCPU_GET(cpuid);
1329 	u_int ipi_bitmap;
1330 
1331 	kasan_mark(&frame, sizeof(frame), sizeof(frame), 0);
1332 
1333 	td = curthread;
1334 	ipi_bitmap = atomic_readandclear_int(&cpuid_to_pcpu[cpu]->
1335 	    pc_ipi_bitmap);
1336 
1337 	/*
1338 	 * sched_preempt() must be called to clear the pending preempt
1339 	 * IPI to enable delivery of further preempts.  However, the
1340 	 * critical section will cause extra scheduler lock thrashing
1341 	 * when used unconditionally.  Only critical_enter() if
1342 	 * hardclock must also run, which requires the section entry.
1343 	 */
1344 	if (ipi_bitmap & (1 << IPI_HARDCLOCK))
1345 		critical_enter();
1346 
1347 	td->td_intr_nesting_level++;
1348 	oldframe = td->td_intr_frame;
1349 	td->td_intr_frame = &frame;
1350 #if defined(STACK) || defined(DDB)
1351 	if (ipi_bitmap & (1 << IPI_TRACE))
1352 		stack_capture_intr();
1353 #endif
1354 	if (ipi_bitmap & (1 << IPI_PREEMPT)) {
1355 #ifdef COUNT_IPIS
1356 		(*ipi_preempt_counts[cpu])++;
1357 #endif
1358 		sched_preempt(td);
1359 	}
1360 	if (ipi_bitmap & (1 << IPI_AST)) {
1361 #ifdef COUNT_IPIS
1362 		(*ipi_ast_counts[cpu])++;
1363 #endif
1364 		/* Nothing to do for AST */
1365 	}
1366 	if (ipi_bitmap & (1 << IPI_HARDCLOCK)) {
1367 #ifdef COUNT_IPIS
1368 		(*ipi_hardclock_counts[cpu])++;
1369 #endif
1370 		hardclockintr();
1371 	}
1372 	td->td_intr_frame = oldframe;
1373 	td->td_intr_nesting_level--;
1374 	if (ipi_bitmap & (1 << IPI_HARDCLOCK))
1375 		critical_exit();
1376 }
1377 
1378 /*
1379  * send an IPI to a set of cpus.
1380  */
1381 void
ipi_selected(cpuset_t cpus,u_int ipi)1382 ipi_selected(cpuset_t cpus, u_int ipi)
1383 {
1384 	int cpu;
1385 
1386 	/*
1387 	 * IPI_STOP_HARD maps to a NMI and the trap handler needs a bit
1388 	 * of help in order to understand what is the source.
1389 	 * Set the mask of receiving CPUs for this purpose.
1390 	 */
1391 	if (ipi == IPI_STOP_HARD)
1392 		CPU_OR_ATOMIC(&ipi_stop_nmi_pending, &cpus);
1393 
1394 	CPU_FOREACH_ISSET(cpu, &cpus) {
1395 		CTR3(KTR_SMP, "%s: cpu: %d ipi: %x", __func__, cpu, ipi);
1396 		ipi_send_cpu(cpu, ipi);
1397 	}
1398 }
1399 
1400 /*
1401  * send an IPI to a specific CPU.
1402  */
1403 void
ipi_cpu(int cpu,u_int ipi)1404 ipi_cpu(int cpu, u_int ipi)
1405 {
1406 
1407 	/*
1408 	 * IPI_STOP_HARD maps to a NMI and the trap handler needs a bit
1409 	 * of help in order to understand what is the source.
1410 	 * Set the mask of receiving CPUs for this purpose.
1411 	 */
1412 	if (ipi == IPI_STOP_HARD)
1413 		CPU_SET_ATOMIC(cpu, &ipi_stop_nmi_pending);
1414 
1415 	CTR3(KTR_SMP, "%s: cpu: %d ipi: %x", __func__, cpu, ipi);
1416 	ipi_send_cpu(cpu, ipi);
1417 }
1418 
1419 /*
1420  * send an IPI to all CPUs EXCEPT myself
1421  */
1422 void
ipi_all_but_self(u_int ipi)1423 ipi_all_but_self(u_int ipi)
1424 {
1425 	cpuset_t other_cpus;
1426 	int cpu, c;
1427 
1428 	/*
1429 	 * IPI_STOP_HARD maps to a NMI and the trap handler needs a bit
1430 	 * of help in order to understand what is the source.
1431 	 * Set the mask of receiving CPUs for this purpose.
1432 	 */
1433 	if (ipi == IPI_STOP_HARD) {
1434 		other_cpus = all_cpus;
1435 		CPU_CLR(PCPU_GET(cpuid), &other_cpus);
1436 		CPU_OR_ATOMIC(&ipi_stop_nmi_pending, &other_cpus);
1437 	}
1438 
1439 	CTR2(KTR_SMP, "%s: ipi: %x", __func__, ipi);
1440 	if (IPI_IS_BITMAPED(ipi)) {
1441 		cpu = PCPU_GET(cpuid);
1442 		CPU_FOREACH(c) {
1443 			if (c != cpu)
1444 				ipi_bitmap_set(c, ipi);
1445 		}
1446 		ipi = IPI_BITMAP_VECTOR;
1447 	}
1448 	lapic_ipi_vectored(ipi, APIC_IPI_DEST_OTHERS);
1449 }
1450 
1451 void
ipi_self_from_nmi(u_int vector)1452 ipi_self_from_nmi(u_int vector)
1453 {
1454 
1455 	lapic_ipi_vectored(vector, APIC_IPI_DEST_SELF);
1456 
1457 	/* Wait for IPI to finish. */
1458 	if (!lapic_ipi_wait(50000)) {
1459 		if (KERNEL_PANICKED())
1460 			return;
1461 		else
1462 			panic("APIC: IPI is stuck");
1463 	}
1464 }
1465 
1466 int
ipi_nmi_handler(void)1467 ipi_nmi_handler(void)
1468 {
1469 	u_int cpuid;
1470 
1471 	/*
1472 	 * As long as there is not a simple way to know about a NMI's
1473 	 * source, if the bitmask for the current CPU is present in
1474 	 * the global pending bitword an IPI_STOP_HARD has been issued
1475 	 * and should be handled.
1476 	 */
1477 	cpuid = PCPU_GET(cpuid);
1478 	if (!CPU_ISSET(cpuid, &ipi_stop_nmi_pending))
1479 		return (1);
1480 
1481 	CPU_CLR_ATOMIC(cpuid, &ipi_stop_nmi_pending);
1482 	cpustop_handler();
1483 	return (0);
1484 }
1485 
1486 int nmi_kdb_lock;
1487 
1488 void
nmi_call_kdb_smp(u_int type,struct trapframe * frame)1489 nmi_call_kdb_smp(u_int type, struct trapframe *frame)
1490 {
1491 	int cpu;
1492 	bool call_post;
1493 
1494 	cpu = PCPU_GET(cpuid);
1495 	if (atomic_cmpset_acq_int(&nmi_kdb_lock, 0, 1)) {
1496 		nmi_call_kdb(cpu, type, frame);
1497 		call_post = false;
1498 	} else {
1499 		savectx(&stoppcbs[cpu]);
1500 		CPU_SET_ATOMIC(cpu, &stopped_cpus);
1501 		while (!atomic_cmpset_acq_int(&nmi_kdb_lock, 0, 1))
1502 			ia32_pause();
1503 		call_post = true;
1504 	}
1505 	atomic_store_rel_int(&nmi_kdb_lock, 0);
1506 	if (call_post)
1507 		cpustop_handler_post(cpu);
1508 }
1509 
1510 /*
1511  * Handle an IPI_STOP by saving our current context and spinning (or mwaiting,
1512  * if available) until we are resumed.
1513  */
1514 void
cpustop_handler(void)1515 cpustop_handler(void)
1516 {
1517 	struct monitorbuf *mb;
1518 	u_int cpu;
1519 	bool use_mwait;
1520 
1521 	cpu = PCPU_GET(cpuid);
1522 
1523 	savectx(&stoppcbs[cpu]);
1524 
1525 	use_mwait = (stop_mwait && (cpu_feature2 & CPUID2_MON) != 0 &&
1526 	    !mwait_cpustop_broken);
1527 	if (use_mwait) {
1528 		mb = PCPU_PTR(monitorbuf);
1529 		atomic_store_int(&mb->stop_state,
1530 		    MONITOR_STOPSTATE_STOPPED);
1531 	}
1532 
1533 	/* Indicate that we are stopped */
1534 	CPU_SET_ATOMIC(cpu, &stopped_cpus);
1535 
1536 	/* Wait for restart */
1537 	while (!CPU_ISSET(cpu, &started_cpus)) {
1538 		if (use_mwait) {
1539 			cpu_monitor(mb, 0, 0);
1540 			if (atomic_load_int(&mb->stop_state) ==
1541 			    MONITOR_STOPSTATE_STOPPED)
1542 				cpu_mwait(0, MWAIT_C1);
1543 			continue;
1544 		}
1545 
1546 		ia32_pause();
1547 
1548 		/*
1549 		 * Halt non-BSP CPUs on panic -- we're never going to need them
1550 		 * again, and might as well save power / release resources
1551 		 * (e.g., overprovisioned VM infrastructure).
1552 		 */
1553 		while (__predict_false(!IS_BSP() && KERNEL_PANICKED()))
1554 			halt();
1555 	}
1556 
1557 	cpustop_handler_post(cpu);
1558 }
1559 
1560 static void
cpustop_handler_post(u_int cpu)1561 cpustop_handler_post(u_int cpu)
1562 {
1563 
1564 	CPU_CLR_ATOMIC(cpu, &started_cpus);
1565 	CPU_CLR_ATOMIC(cpu, &stopped_cpus);
1566 
1567 	/*
1568 	 * We don't broadcast TLB invalidations to other CPUs when they are
1569 	 * stopped. Hence, we clear the TLB before resuming.
1570 	 */
1571 	invltlb_glob();
1572 
1573 #if defined(__amd64__) && (defined(DDB) || defined(GDB))
1574 	amd64_db_resume_dbreg();
1575 #endif
1576 
1577 	if (cpu == 0 && cpustop_restartfunc != NULL) {
1578 		cpustop_restartfunc();
1579 		cpustop_restartfunc = NULL;
1580 	}
1581 }
1582 
1583 /*
1584  * Handle an IPI_SUSPEND by saving our current context and spinning until we
1585  * are resumed.
1586  */
1587 void
cpususpend_handler(void)1588 cpususpend_handler(void)
1589 {
1590 	u_int cpu;
1591 
1592 	mtx_assert(&smp_ipi_mtx, MA_NOTOWNED);
1593 
1594 #ifdef __amd64__
1595 	if (vmm_suspend_p)
1596 		vmm_suspend_p();
1597 #endif
1598 
1599 	cpu = PCPU_GET(cpuid);
1600 
1601 #ifdef XENHVM
1602 	/*
1603 	 * Some Xen guest types (PVH) expose a very minimal set of ACPI tables,
1604 	 * and for example have no support for SCI.  That leads to the suspend
1605 	 * stacks not being allocated, and hence when attempting to perform a
1606 	 * Xen triggered suspension FreeBSD will hit a #PF.  Avoid saving the
1607 	 * CPU and FPU contexts if the stacks are not allocated, as the
1608 	 * hypervisor will already take care of this.  Note that we could even
1609 	 * do this for Xen triggered suspensions on guests that have full ACPI
1610 	 * support, but doing so would introduce extra complexity.
1611 	 */
1612 	if (susppcbs == NULL) {
1613 		KASSERT(vm_guest == VM_GUEST_XEN, ("Missing suspend stack"));
1614 		CPU_SET_ATOMIC(cpu, &suspended_cpus);
1615 		CPU_SET_ATOMIC(cpu, &resuming_cpus);
1616 	} else
1617 #endif
1618 	if (savectx(&susppcbs[cpu]->sp_pcb)) {
1619 #ifdef __amd64__
1620 		fpususpend(susppcbs[cpu]->sp_fpususpend);
1621 #else
1622 		npxsuspend(susppcbs[cpu]->sp_fpususpend);
1623 #endif
1624 		/*
1625 		 * suspended_cpus is cleared shortly after each AP is restarted
1626 		 * by a Startup IPI, so that the BSP can proceed to restarting
1627 		 * the next AP.
1628 		 *
1629 		 * resuming_cpus gets cleared when the AP completes
1630 		 * initialization after having been released by the BSP.
1631 		 * resuming_cpus is probably not the best name for the
1632 		 * variable, because it is actually a set of processors that
1633 		 * haven't resumed yet and haven't necessarily started resuming.
1634 		 *
1635 		 * Note that suspended_cpus is meaningful only for ACPI suspend
1636 		 * as it's not really used for Xen suspend since the APs are
1637 		 * automatically restored to the running state and the correct
1638 		 * context.  For the same reason resumectx is never called in
1639 		 * that case.
1640 		 */
1641 		CPU_SET_ATOMIC(cpu, &suspended_cpus);
1642 		CPU_SET_ATOMIC(cpu, &resuming_cpus);
1643 
1644 		/*
1645 		 * Invalidate the cache after setting the global status bits.
1646 		 * The last AP to set its bit may end up being an Owner of the
1647 		 * corresponding cache line in MOESI protocol.  The AP may be
1648 		 * stopped before the cache line is written to the main memory.
1649 		 */
1650 		wbinvd();
1651 	} else {
1652 #ifdef __amd64__
1653 		fpuresume(susppcbs[cpu]->sp_fpususpend);
1654 #else
1655 		npxresume(susppcbs[cpu]->sp_fpususpend);
1656 #endif
1657 		pmap_init_pat();
1658 		initializecpu();
1659 		PCPU_SET(switchtime, 0);
1660 		PCPU_SET(switchticks, ticks);
1661 
1662 		/* Indicate that we have restarted and restored the context. */
1663 		CPU_CLR_ATOMIC(cpu, &suspended_cpus);
1664 	}
1665 
1666 	/* Wait for resume directive */
1667 	while (!CPU_ISSET(cpu, &toresume_cpus))
1668 		ia32_pause();
1669 
1670 	/* Re-apply microcode updates. */
1671 	ucode_reload();
1672 
1673 #ifdef __i386__
1674 	/* Finish removing the identity mapping of low memory for this AP. */
1675 	invltlb_glob();
1676 #endif
1677 
1678 	if (cpu_ops.cpu_resume)
1679 		cpu_ops.cpu_resume();
1680 #ifdef __amd64__
1681 	if (vmm_resume_p)
1682 		vmm_resume_p();
1683 #endif
1684 
1685 	/* Resume MCA and local APIC */
1686 	lapic_xapic_mode();
1687 	mca_resume();
1688 	lapic_setup(0);
1689 
1690 	/* Indicate that we are resumed */
1691 	CPU_CLR_ATOMIC(cpu, &resuming_cpus);
1692 	CPU_CLR_ATOMIC(cpu, &suspended_cpus);
1693 	CPU_CLR_ATOMIC(cpu, &toresume_cpus);
1694 }
1695 
1696 /*
1697  * Handle an IPI_SWI by waking delayed SWI thread.
1698  */
1699 void
ipi_swi_handler(struct trapframe frame)1700 ipi_swi_handler(struct trapframe frame)
1701 {
1702 
1703 	intr_event_handle(clk_intr_event, &frame);
1704 }
1705 
1706 /*
1707  * This is called once the rest of the system is up and running and we're
1708  * ready to let the AP's out of the pen.
1709  */
1710 static void
release_aps(void * dummy __unused)1711 release_aps(void *dummy __unused)
1712 {
1713 
1714 	if (mp_ncpus == 1)
1715 		return;
1716 	atomic_store_rel_int(&aps_ready, 1);
1717 	while (smp_started == 0)
1718 		ia32_pause();
1719 }
1720 SYSINIT(start_aps, SI_SUB_SMP, SI_ORDER_FIRST, release_aps, NULL);
1721 
1722 #ifdef COUNT_IPIS
1723 /*
1724  * Setup interrupt counters for IPI handlers.
1725  */
1726 static void
mp_ipi_intrcnt(void * dummy)1727 mp_ipi_intrcnt(void *dummy)
1728 {
1729 	char buf[64];
1730 	int i;
1731 
1732 	CPU_FOREACH(i) {
1733 		snprintf(buf, sizeof(buf), "cpu%d:invltlb", i);
1734 		intrcnt_add(buf, &ipi_invltlb_counts[i]);
1735 		snprintf(buf, sizeof(buf), "cpu%d:invlrng", i);
1736 		intrcnt_add(buf, &ipi_invlrng_counts[i]);
1737 		snprintf(buf, sizeof(buf), "cpu%d:invlpg", i);
1738 		intrcnt_add(buf, &ipi_invlpg_counts[i]);
1739 		snprintf(buf, sizeof(buf), "cpu%d:invlcache", i);
1740 		intrcnt_add(buf, &ipi_invlcache_counts[i]);
1741 		snprintf(buf, sizeof(buf), "cpu%d:preempt", i);
1742 		intrcnt_add(buf, &ipi_preempt_counts[i]);
1743 		snprintf(buf, sizeof(buf), "cpu%d:ast", i);
1744 		intrcnt_add(buf, &ipi_ast_counts[i]);
1745 		snprintf(buf, sizeof(buf), "cpu%d:rendezvous", i);
1746 		intrcnt_add(buf, &ipi_rendezvous_counts[i]);
1747 		snprintf(buf, sizeof(buf), "cpu%d:hardclock", i);
1748 		intrcnt_add(buf, &ipi_hardclock_counts[i]);
1749 	}
1750 }
1751 SYSINIT(mp_ipi_intrcnt, SI_SUB_INTR, SI_ORDER_MIDDLE, mp_ipi_intrcnt, NULL);
1752 #endif
1753