xref: /freebsd/sys/amd64/amd64/mp_machdep.c (revision c2fe7156e992a295756206f59c282b9e971b8c91)
1 /*-
2  * SPDX-License-Identifier: BSD-2-Clause
3  *
4  * Copyright (c) 1996, by Steve Passe
5  * Copyright (c) 2003, by Peter Wemm
6  * All rights reserved.
7  *
8  * Redistribution and use in source and binary forms, with or without
9  * modification, are permitted provided that the following conditions
10  * are met:
11  * 1. Redistributions of source code must retain the above copyright
12  *    notice, this list of conditions and the following disclaimer.
13  * 2. The name of the developer may NOT be used to endorse or promote products
14  *    derived from this software without specific prior written permission.
15  *
16  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
17  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
18  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
19  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
20  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
21  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
22  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
23  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
24  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
25  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
26  * SUCH DAMAGE.
27  */
28 
29 #include <sys/cdefs.h>
30 #include "opt_acpi.h"
31 #include "opt_cpu.h"
32 #include "opt_ddb.h"
33 #include "opt_kstack_pages.h"
34 #include "opt_sched.h"
35 #include "opt_smp.h"
36 
37 #include <sys/param.h>
38 #include <sys/systm.h>
39 #include <sys/bus.h>
40 #include <sys/cpuset.h>
41 #include <sys/domainset.h>
42 #include <sys/kdb.h>
43 #include <sys/kernel.h>
44 #include <sys/ktr.h>
45 #include <sys/lock.h>
46 #include <sys/malloc.h>
47 #include <sys/memrange.h>
48 #include <sys/mutex.h>
49 #include <sys/pcpu.h>
50 #include <sys/proc.h>
51 #include <sys/sched.h>
52 #include <sys/smp.h>
53 #include <sys/sysctl.h>
54 
55 #include <vm/vm.h>
56 #include <vm/vm_param.h>
57 #include <vm/pmap.h>
58 #include <vm/vm_kern.h>
59 #include <vm/vm_extern.h>
60 #include <vm/vm_page.h>
61 #include <vm/vm_phys.h>
62 
63 #include <x86/apicreg.h>
64 #include <machine/clock.h>
65 #include <machine/cputypes.h>
66 #include <machine/cpufunc.h>
67 #include <x86/mca.h>
68 #include <machine/md_var.h>
69 #include <machine/pcb.h>
70 #include <machine/psl.h>
71 #include <machine/smp.h>
72 #include <machine/specialreg.h>
73 #include <machine/tss.h>
74 #include <x86/ucode.h>
75 #include <machine/cpu.h>
76 #include <x86/init.h>
77 
78 #ifdef DEV_ACPI
79 #include <contrib/dev/acpica/include/acpi.h>
80 #include <dev/acpica/acpivar.h>
81 #endif
82 
83 #define WARMBOOT_TARGET		0
84 #define WARMBOOT_OFF		(KERNBASE + 0x0467)
85 #define WARMBOOT_SEG		(KERNBASE + 0x0469)
86 
87 #define CMOS_REG		(0x70)
88 #define CMOS_DATA		(0x71)
89 #define BIOS_RESET		(0x0f)
90 #define BIOS_WARM		(0x0a)
91 
92 #define GiB(v)			(v ## ULL << 30)
93 
94 #define	AP_BOOTPT_SZ		(PAGE_SIZE * 4)
95 
96 /* Temporary variables for init_secondary()  */
97 static char *doublefault_stack;
98 static char *mce_stack;
99 static char *nmi_stack;
100 static char *dbg_stack;
101 void *bootpcpu;
102 
103 extern u_int mptramp_la57;
104 extern u_int mptramp_nx;
105 smp_targeted_tlb_shootdown_t smp_targeted_tlb_shootdown =
106     &smp_targeted_tlb_shootdown_native;
107 
108 /*
109  * Local data and functions.
110  */
111 
112 static int start_ap(int apic_id, vm_paddr_t boot_address);
113 
114 /*
115  * Initialize the IPI handlers and start up the AP's.
116  */
117 void
cpu_mp_start(void)118 cpu_mp_start(void)
119 {
120 	int i;
121 
122 	/* Initialize the logical ID to APIC ID table. */
123 	for (i = 0; i < MAXCPU; i++) {
124 		cpu_apic_ids[i] = -1;
125 	}
126 
127 	/* Install an inter-CPU IPI for cache and TLB invalidations. */
128 	setidt(IPI_INVLOP, pti ? IDTVEC(invlop_pti) : IDTVEC(invlop),
129 	    SDT_SYSIGT, SEL_KPL, 0);
130 
131 	/* Install an inter-CPU IPI for all-CPU rendezvous */
132 	setidt(IPI_RENDEZVOUS, pti ? IDTVEC(rendezvous_pti) :
133 	    IDTVEC(rendezvous), SDT_SYSIGT, SEL_KPL, 0);
134 
135 	/* Install generic inter-CPU IPI handler */
136 	setidt(IPI_BITMAP_VECTOR, pti ? IDTVEC(ipi_intr_bitmap_handler_pti) :
137 	    IDTVEC(ipi_intr_bitmap_handler), SDT_SYSIGT, SEL_KPL, 0);
138 
139 	/* Install an inter-CPU IPI for CPU stop/restart */
140 	setidt(IPI_STOP, pti ? IDTVEC(cpustop_pti) : IDTVEC(cpustop),
141 	    SDT_SYSIGT, SEL_KPL, 0);
142 
143 	/* Install an inter-CPU IPI for CPU suspend/resume */
144 	setidt(IPI_SUSPEND, pti ? IDTVEC(cpususpend_pti) : IDTVEC(cpususpend),
145 	    SDT_SYSIGT, SEL_KPL, 0);
146 
147 	/* Install an IPI for calling delayed SWI */
148 	setidt(IPI_SWI, pti ? IDTVEC(ipi_swi_pti) : IDTVEC(ipi_swi),
149 	    SDT_SYSIGT, SEL_KPL, 0);
150 
151 	/* Set boot_cpu_id if needed. */
152 	if (boot_cpu_id == -1) {
153 		boot_cpu_id = PCPU_GET(apic_id);
154 		cpu_info[boot_cpu_id].cpu_bsp = 1;
155 	} else
156 		KASSERT(boot_cpu_id == PCPU_GET(apic_id),
157 		    ("BSP's APIC ID doesn't match boot_cpu_id"));
158 
159 	/* Probe logical/physical core configuration. */
160 	topo_probe();
161 
162 	assign_cpu_ids();
163 
164 	mptramp_la57 = la57;
165 	mptramp_nx = pg_nx != 0;
166 	MPASS(kernel_pmap->pm_cr3 < (1UL << 32));
167 	mptramp_pagetables = kernel_pmap->pm_cr3;
168 
169 	/* Start each Application Processor */
170 	start_all_aps();
171 
172 	set_interrupt_apic_ids();
173 
174 #if defined(DEV_ACPI) && MAXMEMDOM > 1
175 	acpi_pxm_set_cpu_locality();
176 #endif
177 }
178 
179 /*
180  * AP CPU's call this to initialize themselves.
181  */
182 void
init_secondary(void)183 init_secondary(void)
184 {
185 	struct pcpu *pc;
186 	struct nmi_pcpu *np;
187 	struct user_segment_descriptor *gdt;
188 	struct region_descriptor ap_gdt;
189 	u_int64_t cr0;
190 	int cpu, gsel_tss, x;
191 
192 	/* Set by the startup code for us to use */
193 	cpu = bootAP;
194 
195 	/* Update microcode before doing anything else. */
196 	ucode_load_ap(cpu);
197 
198 	/* Initialize the PCPU area. */
199 	pc = bootpcpu;
200 	pcpu_init(pc, cpu, sizeof(struct pcpu));
201 	dpcpu_init(dpcpu, cpu);
202 	pc->pc_apic_id = cpu_apic_ids[cpu];
203 	pc->pc_prvspace = pc;
204 	pc->pc_curthread = 0;
205 	pc->pc_tssp = &pc->pc_common_tss;
206 	pc->pc_rsp0 = 0;
207 	pc->pc_pti_rsp0 = (((vm_offset_t)&pc->pc_pti_stack +
208 	    PC_PTI_STACK_SZ * sizeof(uint64_t)) & ~0xful);
209 	gdt = pc->pc_gdt;
210 	pc->pc_tss = (struct system_segment_descriptor *)&gdt[GPROC0_SEL];
211 	pc->pc_fs32p = &gdt[GUFS32_SEL];
212 	pc->pc_gs32p = &gdt[GUGS32_SEL];
213 	pc->pc_ldt = (struct system_segment_descriptor *)&gdt[GUSERLDT_SEL];
214 	pc->pc_ucr3_load_mask = PMAP_UCR3_NOMASK;
215 	/* See comment in pmap_bootstrap(). */
216 	pc->pc_pcid_next = PMAP_PCID_KERN + 2;
217 	pc->pc_pcid_gen = 1;
218 	pc->pc_kpmap_store.pm_pcid = PMAP_PCID_KERN;
219 	pc->pc_kpmap_store.pm_gen = 1;
220 
221 	pc->pc_smp_tlb_gen = 1;
222 
223 	/* Init tss */
224 	pc->pc_common_tss = __pcpu[0].pc_common_tss;
225 	pc->pc_common_tss.tss_iobase = sizeof(struct amd64tss) +
226 	    IOPERM_BITMAP_SIZE;
227 	pc->pc_common_tss.tss_rsp0 = 0;
228 
229 	/* The doublefault stack runs on IST1. */
230 	np = ((struct nmi_pcpu *)&doublefault_stack[DBLFAULT_STACK_SIZE]) - 1;
231 	np->np_pcpu = (register_t)pc;
232 	pc->pc_common_tss.tss_ist1 = (long)np;
233 
234 	/* The NMI stack runs on IST2. */
235 	np = ((struct nmi_pcpu *)&nmi_stack[NMI_STACK_SIZE]) - 1;
236 	np->np_pcpu = (register_t)pc;
237 	pc->pc_common_tss.tss_ist2 = (long)np;
238 
239 	/* The MC# stack runs on IST3. */
240 	np = ((struct nmi_pcpu *)&mce_stack[MCE_STACK_SIZE]) - 1;
241 	np->np_pcpu = (register_t)pc;
242 	pc->pc_common_tss.tss_ist3 = (long)np;
243 
244 	/* The DB# stack runs on IST4. */
245 	np = ((struct nmi_pcpu *)&dbg_stack[DBG_STACK_SIZE]) - 1;
246 	np->np_pcpu = (register_t)pc;
247 	pc->pc_common_tss.tss_ist4 = (long)np;
248 
249 	/* Prepare private GDT */
250 	gdt_segs[GPROC0_SEL].ssd_base = (long)&pc->pc_common_tss;
251 	for (x = 0; x < NGDT; x++) {
252 		if (x != GPROC0_SEL && x != GPROC0_SEL + 1 &&
253 		    x != GUSERLDT_SEL && x != GUSERLDT_SEL + 1)
254 			ssdtosd(&gdt_segs[x], &gdt[x]);
255 	}
256 	ssdtosyssd(&gdt_segs[GPROC0_SEL],
257 	    (struct system_segment_descriptor *)&gdt[GPROC0_SEL]);
258 	ap_gdt.rd_limit = NGDT * sizeof(gdt[0]) - 1;
259 	ap_gdt.rd_base = (u_long)gdt;
260 	lgdt(&ap_gdt);			/* does magic intra-segment return */
261 
262 	wrmsr(MSR_FSBASE, 0);		/* User value */
263 	wrmsr(MSR_GSBASE, (uint64_t)pc);
264 	wrmsr(MSR_KGSBASE, 0);		/* User value */
265 	fix_cpuid();
266 
267 	lidt(&r_idt);
268 
269 	gsel_tss = GSEL(GPROC0_SEL, SEL_KPL);
270 	ltr(gsel_tss);
271 
272 	/*
273 	 * Set to a known state:
274 	 * Set by mpboot.s: CR0_PG, CR0_PE
275 	 * Set by cpu_setregs: CR0_NE, CR0_MP, CR0_TS, CR0_WP, CR0_AM
276 	 */
277 	cr0 = rcr0();
278 	cr0 &= ~(CR0_CD | CR0_NW | CR0_EM);
279 	load_cr0(cr0);
280 
281 	amd64_conf_fast_syscall();
282 
283 	/* signal our startup to the BSP. */
284 	mp_naps++;
285 
286 	/* Spin until the BSP releases the AP's. */
287 	while (atomic_load_acq_int(&aps_ready) == 0)
288 		ia32_pause();
289 
290 	init_secondary_tail();
291 }
292 
293 static void
amd64_mp_alloc_pcpu(void)294 amd64_mp_alloc_pcpu(void)
295 {
296 	vm_page_t m;
297 	int cpu;
298 
299 	/* Allocate pcpu areas to the correct domain. */
300 	for (cpu = 1; cpu < mp_ncpus; cpu++) {
301 #ifdef NUMA
302 		m = NULL;
303 		if (vm_ndomains > 1) {
304 			m = vm_page_alloc_noobj_domain(
305 			    acpi_pxm_get_cpu_locality(cpu_apic_ids[cpu]),
306 			    VM_ALLOC_ZERO);
307 		}
308 		if (m == NULL)
309 #endif
310 			m = vm_page_alloc_noobj(VM_ALLOC_ZERO);
311 		if (m == NULL)
312 			panic("cannot alloc pcpu page for cpu %d", cpu);
313 		pmap_qenter((vm_offset_t)&__pcpu[cpu], &m, 1);
314 	}
315 }
316 
317 /*
318  * start each AP in our list
319  */
320 int
start_all_aps(void)321 start_all_aps(void)
322 {
323 	vm_page_t m_boottramp, m_pml4, m_pdp, m_pd[4];
324 	pml5_entry_t old_pml45;
325 	pml4_entry_t *v_pml4;
326 	pdp_entry_t *v_pdp;
327 	pd_entry_t *v_pd;
328 	vm_paddr_t boot_address;
329 	u_int32_t mpbioswarmvec;
330 	int apic_id, cpu, domain, i;
331 	u_char mpbiosreason;
332 
333 	amd64_mp_alloc_pcpu();
334 	mtx_init(&ap_boot_mtx, "ap boot", NULL, MTX_SPIN);
335 
336 	MPASS(bootMP_size <= PAGE_SIZE);
337 	m_boottramp = vm_page_alloc_noobj_contig(0, 1, 0,
338 	    (1ULL << 20), /* Trampoline should be below 1M for real mode */
339 	    PAGE_SIZE, 0, VM_MEMATTR_DEFAULT);
340 	boot_address = VM_PAGE_TO_PHYS(m_boottramp);
341 
342 	/* Create a transient 1:1 mapping of low 4G */
343 	if (la57) {
344 		m_pml4 = pmap_page_alloc_below_4g(true);
345 		v_pml4 = (pml4_entry_t *)PHYS_TO_DMAP(VM_PAGE_TO_PHYS(m_pml4));
346 	} else {
347 		v_pml4 = &kernel_pmap->pm_pmltop[0];
348 	}
349 	m_pdp = pmap_page_alloc_below_4g(true);
350 	v_pdp = (pdp_entry_t *)PHYS_TO_DMAP(VM_PAGE_TO_PHYS(m_pdp));
351 	m_pd[0] = pmap_page_alloc_below_4g(false);
352 	v_pd = (pd_entry_t *)PHYS_TO_DMAP(VM_PAGE_TO_PHYS(m_pd[0]));
353 	for (i = 0; i < NPDEPG; i++)
354 		v_pd[i] = (i << PDRSHIFT) | X86_PG_V | X86_PG_RW | X86_PG_A |
355 		    X86_PG_M | PG_PS;
356 	m_pd[1] = pmap_page_alloc_below_4g(false);
357 	v_pd = (pd_entry_t *)PHYS_TO_DMAP(VM_PAGE_TO_PHYS(m_pd[1]));
358 	for (i = 0; i < NPDEPG; i++)
359 		v_pd[i] = (NBPDP + (i << PDRSHIFT)) | X86_PG_V | X86_PG_RW |
360 		    X86_PG_A | X86_PG_M | PG_PS;
361 	m_pd[2] = pmap_page_alloc_below_4g(false);
362 	v_pd = (pd_entry_t *)PHYS_TO_DMAP(VM_PAGE_TO_PHYS(m_pd[2]));
363 	for (i = 0; i < NPDEPG; i++)
364 		v_pd[i] = (2UL * NBPDP + (i << PDRSHIFT)) | X86_PG_V |
365 		    X86_PG_RW | X86_PG_A | X86_PG_M | PG_PS;
366 	m_pd[3] = pmap_page_alloc_below_4g(false);
367 	v_pd = (pd_entry_t *)PHYS_TO_DMAP(VM_PAGE_TO_PHYS(m_pd[3]));
368 	for (i = 0; i < NPDEPG; i++)
369 		v_pd[i] = (3UL * NBPDP + (i << PDRSHIFT)) | X86_PG_V |
370 		    X86_PG_RW | X86_PG_A | X86_PG_M | PG_PS;
371 	v_pdp[0] = VM_PAGE_TO_PHYS(m_pd[0]) | X86_PG_V |
372 	    X86_PG_RW | X86_PG_A | X86_PG_M;
373 	v_pdp[1] = VM_PAGE_TO_PHYS(m_pd[1]) | X86_PG_V |
374 	    X86_PG_RW | X86_PG_A | X86_PG_M;
375 	v_pdp[2] = VM_PAGE_TO_PHYS(m_pd[2]) | X86_PG_V |
376 	    X86_PG_RW | X86_PG_A | X86_PG_M;
377 	v_pdp[3] = VM_PAGE_TO_PHYS(m_pd[3]) | X86_PG_V |
378 	    X86_PG_RW | X86_PG_A | X86_PG_M;
379 	old_pml45 = kernel_pmap->pm_pmltop[0];
380 	if (la57) {
381 		kernel_pmap->pm_pmltop[0] = VM_PAGE_TO_PHYS(m_pml4) |
382 		    X86_PG_V | X86_PG_RW | X86_PG_A | X86_PG_M;
383 	}
384 	v_pml4[0] = VM_PAGE_TO_PHYS(m_pdp) | X86_PG_V |
385 	    X86_PG_RW | X86_PG_A | X86_PG_M;
386 	pmap_invalidate_all(kernel_pmap);
387 
388 	/* copy the AP 1st level boot code */
389 	bcopy(mptramp_start, (void *)PHYS_TO_DMAP(boot_address), bootMP_size);
390 	if (bootverbose)
391 		printf("AP boot address %#lx\n", boot_address);
392 
393 	/* save the current value of the warm-start vector */
394 	if (!efi_boot)
395 		mpbioswarmvec = *((u_int32_t *) WARMBOOT_OFF);
396 	outb(CMOS_REG, BIOS_RESET);
397 	mpbiosreason = inb(CMOS_DATA);
398 
399 	/* setup a vector to our boot code */
400 	if (!efi_boot) {
401 		*((volatile u_short *)WARMBOOT_OFF) = WARMBOOT_TARGET;
402 		*((volatile u_short *)WARMBOOT_SEG) = (boot_address >> 4);
403 	}
404 	outb(CMOS_REG, BIOS_RESET);
405 	outb(CMOS_DATA, BIOS_WARM);	/* 'warm-start' */
406 
407 	/* start each AP */
408 	domain = 0;
409 	for (cpu = 1; cpu < mp_ncpus; cpu++) {
410 		apic_id = cpu_apic_ids[cpu];
411 #ifdef NUMA
412 		if (vm_ndomains > 1)
413 			domain = acpi_pxm_get_cpu_locality(apic_id);
414 #endif
415 		/* allocate and set up an idle stack data page */
416 		bootstacks[cpu] = kmem_malloc(kstack_pages * PAGE_SIZE,
417 		    M_WAITOK | M_ZERO);
418 		doublefault_stack = kmem_malloc(DBLFAULT_STACK_SIZE,
419 		    M_WAITOK | M_ZERO);
420 		mce_stack = kmem_malloc(MCE_STACK_SIZE,
421 		    M_WAITOK | M_ZERO);
422 		nmi_stack = kmem_malloc_domainset(
423 		    DOMAINSET_PREF(domain), NMI_STACK_SIZE, M_WAITOK | M_ZERO);
424 		dbg_stack = kmem_malloc_domainset(
425 		    DOMAINSET_PREF(domain), DBG_STACK_SIZE, M_WAITOK | M_ZERO);
426 		dpcpu = kmem_malloc_domainset(DOMAINSET_PREF(domain),
427 		    DPCPU_SIZE, M_WAITOK | M_ZERO);
428 
429 		bootpcpu = &__pcpu[cpu];
430 		bootSTK = (char *)bootstacks[cpu] +
431 		    kstack_pages * PAGE_SIZE - 8;
432 		bootAP = cpu;
433 
434 		/* attempt to start the Application Processor */
435 		if (!start_ap(apic_id, boot_address)) {
436 			/* restore the warmstart vector */
437 			if (!efi_boot)
438 				*(u_int32_t *)WARMBOOT_OFF = mpbioswarmvec;
439 			panic("AP #%d (PHY# %d) failed!", cpu, apic_id);
440 		}
441 
442 		CPU_SET(cpu, &all_cpus);	/* record AP in CPU map */
443 	}
444 
445 	/* restore the warmstart vector */
446 	if (!efi_boot)
447 		*(u_int32_t *)WARMBOOT_OFF = mpbioswarmvec;
448 
449 	outb(CMOS_REG, BIOS_RESET);
450 	outb(CMOS_DATA, mpbiosreason);
451 
452 	/* Destroy transient 1:1 mapping */
453 	kernel_pmap->pm_pmltop[0] = old_pml45;
454 	invlpg(0);
455 	if (la57)
456 		vm_page_free(m_pml4);
457 	vm_page_free(m_pd[3]);
458 	vm_page_free(m_pd[2]);
459 	vm_page_free(m_pd[1]);
460 	vm_page_free(m_pd[0]);
461 	vm_page_free(m_pdp);
462 	vm_page_free(m_boottramp);
463 
464 	/* number of APs actually started */
465 	return (mp_naps);
466 }
467 
468 /*
469  * This function starts the AP (application processor) identified
470  * by the APIC ID 'physicalCpu'.  It does quite a "song and dance"
471  * to accomplish this.  This is necessary because of the nuances
472  * of the different hardware we might encounter.  It isn't pretty,
473  * but it seems to work.
474  */
475 static int
start_ap(int apic_id,vm_paddr_t boot_address)476 start_ap(int apic_id, vm_paddr_t boot_address)
477 {
478 	int vector, ms;
479 	int cpus;
480 
481 	/* calculate the vector */
482 	vector = (boot_address >> 12) & 0xff;
483 
484 	/* used as a watchpoint to signal AP startup */
485 	cpus = mp_naps;
486 
487 	ipi_startup(apic_id, vector);
488 
489 	/* Wait up to 5 seconds for it to start. */
490 	for (ms = 0; ms < 5000; ms++) {
491 		if (mp_naps > cpus)
492 			return 1;	/* return SUCCESS */
493 		DELAY(1000);
494 	}
495 	return 0;		/* return FAILURE */
496 }
497 
498 /*
499  * Flush the TLB on other CPU's
500  */
501 
502 /*
503  * These variables are initialized at startup to reflect how each of
504  * the different kinds of invalidations should be performed on the
505  * current machine and environment.
506  */
507 static enum invl_op_codes invl_op_tlb;
508 static enum invl_op_codes invl_op_pgrng;
509 static enum invl_op_codes invl_op_pg;
510 
511 /*
512  * Scoreboard of IPI completion notifications from target to IPI initiator.
513  *
514  * Each CPU can initiate shootdown IPI independently from other CPUs.
515  * Initiator enters critical section, then fills its local PCPU
516  * shootdown info (pc_smp_tlb_ vars), then clears scoreboard generation
517  * at location (cpu, my_cpuid) for each target cpu.  After that IPI is
518  * sent to all targets which scan for zeroed scoreboard generation
519  * words.  Upon finding such word the shootdown data is read from
520  * corresponding cpu's pcpu, and generation is set.  Meantime initiator
521  * loops waiting for all zeroed generations in scoreboard to update.
522  */
523 static uint32_t *invl_scoreboard;
524 
525 static void
invl_scoreboard_init(void * arg __unused)526 invl_scoreboard_init(void *arg __unused)
527 {
528 	u_int i;
529 
530 	invl_scoreboard = malloc(sizeof(uint32_t) * (mp_maxid + 1) *
531 	    (mp_maxid + 1), M_DEVBUF, M_WAITOK);
532 	for (i = 0; i < (mp_maxid + 1) * (mp_maxid + 1); i++)
533 		invl_scoreboard[i] = 1;
534 
535 	if (pmap_pcid_enabled) {
536 		if (invpcid_works) {
537 			if (pti)
538 				invl_op_tlb = INVL_OP_TLB_INVPCID_PTI;
539 			else
540 				invl_op_tlb = INVL_OP_TLB_INVPCID;
541 			invl_op_pgrng = INVL_OP_PGRNG_INVPCID;
542 			invl_op_pg = INVL_OP_PG_INVPCID;
543 		} else {
544 			invl_op_tlb = INVL_OP_TLB_PCID;
545 			invl_op_pgrng = INVL_OP_PGRNG_PCID;
546 			invl_op_pg = INVL_OP_PG_PCID;
547 		}
548 	} else {
549 		invl_op_tlb = INVL_OP_TLB;
550 		invl_op_pgrng = INVL_OP_PGRNG;
551 		invl_op_pg = INVL_OP_PG;
552 	}
553 }
554 SYSINIT(invl_ops, SI_SUB_SMP - 1, SI_ORDER_ANY, invl_scoreboard_init, NULL);
555 
556 static uint32_t *
invl_scoreboard_getcpu(u_int cpu)557 invl_scoreboard_getcpu(u_int cpu)
558 {
559 	return (invl_scoreboard + cpu * (mp_maxid + 1));
560 }
561 
562 static uint32_t *
invl_scoreboard_slot(u_int cpu)563 invl_scoreboard_slot(u_int cpu)
564 {
565 	return (invl_scoreboard_getcpu(cpu) + PCPU_GET(cpuid));
566 }
567 
568 /*
569  * Used by the pmap to request cache or TLB invalidation on local and
570  * remote processors.  Mask provides the set of remote CPUs that are
571  * to be signalled with the invalidation IPI.  As an optimization, the
572  * curcpu_cb callback is invoked on the calling CPU in a critical
573  * section while waiting for the remote CPUs to complete the operation.
574  *
575  * The callback function is called unconditionally on the caller's
576  * underlying processor, even when this processor is not set in the
577  * mask.  So, the callback function must be prepared to handle such
578  * spurious invocations.
579  *
580  * Interrupts must be enabled when calling the function with smp
581  * started, to avoid deadlock with other IPIs that are protected with
582  * smp_ipi_mtx spinlock at the initiator side.
583  *
584  * Function must be called with the thread pinned, and it unpins on
585  * completion.
586  */
587 void
smp_targeted_tlb_shootdown_native(pmap_t pmap,vm_offset_t addr1,vm_offset_t addr2,smp_invl_cb_t curcpu_cb,enum invl_op_codes op)588 smp_targeted_tlb_shootdown_native(pmap_t pmap, vm_offset_t addr1,
589     vm_offset_t addr2, smp_invl_cb_t curcpu_cb, enum invl_op_codes op)
590 {
591 	cpuset_t mask;
592 	uint32_t generation, *p_cpudone;
593 	int cpu;
594 	bool is_all;
595 
596 	/*
597 	 * It is not necessary to signal other CPUs while booting or
598 	 * when in the debugger.
599 	 */
600 	if (__predict_false(kdb_active || KERNEL_PANICKED() || !smp_started))
601 		goto local_cb;
602 
603 	KASSERT(curthread->td_pinned > 0, ("curthread not pinned"));
604 
605 	/*
606 	 * Make a stable copy of the set of CPUs on which the pmap is active.
607 	 * See if we have to interrupt other CPUs.
608 	 */
609 	CPU_COPY(pmap_invalidate_cpu_mask(pmap), &mask);
610 	is_all = CPU_CMP(&mask, &all_cpus) == 0;
611 	CPU_CLR(curcpu, &mask);
612 	if (CPU_EMPTY(&mask))
613 		goto local_cb;
614 
615 	/*
616 	 * Initiator must have interrupts enabled, which prevents
617 	 * non-invalidation IPIs that take smp_ipi_mtx spinlock,
618 	 * from deadlocking with us.  On the other hand, preemption
619 	 * must be disabled to pin initiator to the instance of the
620 	 * pcpu pc_smp_tlb data and scoreboard line.
621 	 */
622 	KASSERT((read_rflags() & PSL_I) != 0,
623 	    ("smp_targeted_tlb_shootdown: interrupts disabled"));
624 	critical_enter();
625 
626 	PCPU_SET(smp_tlb_addr1, addr1);
627 	PCPU_SET(smp_tlb_addr2, addr2);
628 	PCPU_SET(smp_tlb_pmap, pmap);
629 	generation = PCPU_GET(smp_tlb_gen);
630 	if (++generation == 0)
631 		generation = 1;
632 	PCPU_SET(smp_tlb_gen, generation);
633 	PCPU_SET(smp_tlb_op, op);
634 	/* Fence between filling smp_tlb fields and clearing scoreboard. */
635 	atomic_thread_fence_rel();
636 
637 	CPU_FOREACH_ISSET(cpu, &mask) {
638 		KASSERT(*invl_scoreboard_slot(cpu) != 0,
639 		    ("IPI scoreboard is zero, initiator %d target %d",
640 		    curcpu, cpu));
641 		*invl_scoreboard_slot(cpu) = 0;
642 	}
643 
644 	/*
645 	 * IPI acts as a fence between writing to the scoreboard above
646 	 * (zeroing slot) and reading from it below (wait for
647 	 * acknowledgment).
648 	 */
649 	if (is_all) {
650 		ipi_all_but_self(IPI_INVLOP);
651 	} else {
652 		ipi_selected(mask, IPI_INVLOP);
653 	}
654 	curcpu_cb(pmap, addr1, addr2);
655 	CPU_FOREACH_ISSET(cpu, &mask) {
656 		p_cpudone = invl_scoreboard_slot(cpu);
657 		while (atomic_load_int(p_cpudone) != generation)
658 			ia32_pause();
659 	}
660 
661 	/*
662 	 * Unpin before leaving critical section.  If the thread owes
663 	 * preemption, this allows scheduler to select thread on any
664 	 * CPU from its cpuset.
665 	 */
666 	sched_unpin();
667 	critical_exit();
668 
669 	return;
670 
671 local_cb:
672 	critical_enter();
673 	curcpu_cb(pmap, addr1, addr2);
674 	sched_unpin();
675 	critical_exit();
676 }
677 
678 void
smp_masked_invltlb(pmap_t pmap,smp_invl_cb_t curcpu_cb)679 smp_masked_invltlb(pmap_t pmap, smp_invl_cb_t curcpu_cb)
680 {
681 	if (invlpgb_works && pmap == kernel_pmap) {
682 		invlpgb(INVLPGB_GLOB, 0, 0);
683 
684 		/*
685 		 * TLBSYNC syncs only against INVLPGB executed on the
686 		 * same CPU.  Since current thread is pinned by
687 		 * caller, we do not need to enter critical section to
688 		 * prevent migration.
689 		 */
690 		tlbsync();
691 		sched_unpin();
692 		return;
693 	}
694 
695 	smp_targeted_tlb_shootdown(pmap, 0, 0, curcpu_cb, invl_op_tlb);
696 #ifdef COUNT_XINVLTLB_HITS
697 	ipi_global++;
698 #endif
699 }
700 
701 void
smp_masked_invlpg(vm_offset_t addr,pmap_t pmap,smp_invl_cb_t curcpu_cb)702 smp_masked_invlpg(vm_offset_t addr, pmap_t pmap, smp_invl_cb_t curcpu_cb)
703 {
704 	if (invlpgb_works && pmap == kernel_pmap) {
705 		invlpgb(INVLPGB_GLOB | INVLPGB_VA | trunc_page(addr), 0, 0);
706 		tlbsync();
707 		sched_unpin();
708 		return;
709 	}
710 
711 	smp_targeted_tlb_shootdown(pmap, addr, 0, curcpu_cb, invl_op_pg);
712 #ifdef COUNT_XINVLTLB_HITS
713 	ipi_page++;
714 #endif
715 }
716 
717 void
smp_masked_invlpg_range(vm_offset_t addr1,vm_offset_t addr2,pmap_t pmap,smp_invl_cb_t curcpu_cb)718 smp_masked_invlpg_range(vm_offset_t addr1, vm_offset_t addr2, pmap_t pmap,
719     smp_invl_cb_t curcpu_cb)
720 {
721 	if (invlpgb_works && pmap == kernel_pmap) {
722 		vm_offset_t va;
723 		uint64_t cnt, total;
724 
725 		addr1 = trunc_page(addr1);
726 		addr2 = round_page(addr2);
727 		total = atop(addr2 - addr1);
728 		for (va = addr1; total > 0;) {
729 			if ((va & PDRMASK) != 0 || total < NPDEPG) {
730 				cnt = atop(NBPDR - (va & PDRMASK));
731 				if (cnt > total)
732 					cnt = total;
733 				if (cnt > invlpgb_maxcnt + 1)
734 					cnt = invlpgb_maxcnt + 1;
735 				invlpgb(INVLPGB_GLOB | INVLPGB_VA | va, 0,
736 				    cnt - 1);
737 				va += ptoa(cnt);
738 				total -= cnt;
739 			} else {
740 				cnt = total / NPTEPG;
741 				if (cnt > invlpgb_maxcnt + 1)
742 					cnt = invlpgb_maxcnt + 1;
743 				invlpgb(INVLPGB_GLOB | INVLPGB_VA | va, 0,
744 				    INVLPGB_2M_CNT | (cnt - 1));
745 				va += cnt << PDRSHIFT;
746 				total -= cnt * NPTEPG;
747 			}
748 		}
749 		tlbsync();
750 		sched_unpin();
751 		return;
752 	}
753 
754 	smp_targeted_tlb_shootdown(pmap, addr1, addr2, curcpu_cb,
755 	    invl_op_pgrng);
756 #ifdef COUNT_XINVLTLB_HITS
757 	ipi_range++;
758 	ipi_range_size += (addr2 - addr1) / PAGE_SIZE;
759 #endif
760 }
761 
762 void
smp_cache_flush(smp_invl_cb_t curcpu_cb)763 smp_cache_flush(smp_invl_cb_t curcpu_cb)
764 {
765 	smp_targeted_tlb_shootdown(kernel_pmap, 0, 0, curcpu_cb, INVL_OP_CACHE);
766 }
767 
768 /*
769  * Handlers for TLB related IPIs
770  */
771 static void
invltlb_handler(pmap_t smp_tlb_pmap)772 invltlb_handler(pmap_t smp_tlb_pmap)
773 {
774 #ifdef COUNT_XINVLTLB_HITS
775 	xhits_gbl[PCPU_GET(cpuid)]++;
776 #endif /* COUNT_XINVLTLB_HITS */
777 #ifdef COUNT_IPIS
778 	(*ipi_invltlb_counts[PCPU_GET(cpuid)])++;
779 #endif /* COUNT_IPIS */
780 
781 	if (smp_tlb_pmap == kernel_pmap)
782 		invltlb_glob();
783 	else
784 		invltlb();
785 }
786 
787 static void
invltlb_invpcid_handler(pmap_t smp_tlb_pmap)788 invltlb_invpcid_handler(pmap_t smp_tlb_pmap)
789 {
790 	struct invpcid_descr d;
791 
792 #ifdef COUNT_XINVLTLB_HITS
793 	xhits_gbl[PCPU_GET(cpuid)]++;
794 #endif /* COUNT_XINVLTLB_HITS */
795 #ifdef COUNT_IPIS
796 	(*ipi_invltlb_counts[PCPU_GET(cpuid)])++;
797 #endif /* COUNT_IPIS */
798 
799 	d.pcid = pmap_get_pcid(smp_tlb_pmap);
800 	d.pad = 0;
801 	d.addr = 0;
802 	invpcid(&d, smp_tlb_pmap == kernel_pmap ? INVPCID_CTXGLOB :
803 	    INVPCID_CTX);
804 }
805 
806 static void
invltlb_invpcid_pti_handler(pmap_t smp_tlb_pmap)807 invltlb_invpcid_pti_handler(pmap_t smp_tlb_pmap)
808 {
809 	struct invpcid_descr d;
810 
811 #ifdef COUNT_XINVLTLB_HITS
812 	xhits_gbl[PCPU_GET(cpuid)]++;
813 #endif /* COUNT_XINVLTLB_HITS */
814 #ifdef COUNT_IPIS
815 	(*ipi_invltlb_counts[PCPU_GET(cpuid)])++;
816 #endif /* COUNT_IPIS */
817 
818 	d.pcid = pmap_get_pcid(smp_tlb_pmap);
819 	d.pad = 0;
820 	d.addr = 0;
821 	if (smp_tlb_pmap == kernel_pmap) {
822 		/*
823 		 * This invalidation actually needs to clear kernel
824 		 * mappings from the TLB in the current pmap, but
825 		 * since we were asked for the flush in the kernel
826 		 * pmap, achieve it by performing global flush.
827 		 */
828 		invpcid(&d, INVPCID_CTXGLOB);
829 	} else {
830 		invpcid(&d, INVPCID_CTX);
831 		if (smp_tlb_pmap == PCPU_GET(curpmap) &&
832 		    smp_tlb_pmap->pm_ucr3 != PMAP_NO_CR3)
833 			PCPU_SET(ucr3_load_mask, ~CR3_PCID_SAVE);
834 	}
835 }
836 
837 static void
invltlb_pcid_handler(pmap_t smp_tlb_pmap)838 invltlb_pcid_handler(pmap_t smp_tlb_pmap)
839 {
840 #ifdef COUNT_XINVLTLB_HITS
841 	xhits_gbl[PCPU_GET(cpuid)]++;
842 #endif /* COUNT_XINVLTLB_HITS */
843 #ifdef COUNT_IPIS
844 	(*ipi_invltlb_counts[PCPU_GET(cpuid)])++;
845 #endif /* COUNT_IPIS */
846 
847 	if (smp_tlb_pmap == kernel_pmap) {
848 		invltlb_glob();
849 	} else {
850 		/*
851 		 * The current pmap might not be equal to
852 		 * smp_tlb_pmap.  The clearing of the pm_gen in
853 		 * pmap_invalidate_all() takes care of TLB
854 		 * invalidation when switching to the pmap on this
855 		 * CPU.
856 		 */
857 		if (smp_tlb_pmap == PCPU_GET(curpmap)) {
858 			load_cr3(smp_tlb_pmap->pm_cr3 |
859 			    pmap_get_pcid(smp_tlb_pmap));
860 			if (smp_tlb_pmap->pm_ucr3 != PMAP_NO_CR3)
861 				PCPU_SET(ucr3_load_mask, ~CR3_PCID_SAVE);
862 		}
863 	}
864 }
865 
866 static void
invlpg_handler(vm_offset_t smp_tlb_addr1)867 invlpg_handler(vm_offset_t smp_tlb_addr1)
868 {
869 #ifdef COUNT_XINVLTLB_HITS
870 	xhits_pg[PCPU_GET(cpuid)]++;
871 #endif /* COUNT_XINVLTLB_HITS */
872 #ifdef COUNT_IPIS
873 	(*ipi_invlpg_counts[PCPU_GET(cpuid)])++;
874 #endif /* COUNT_IPIS */
875 
876 	invlpg(smp_tlb_addr1);
877 }
878 
879 static void
invlpg_invpcid_handler(pmap_t smp_tlb_pmap,vm_offset_t smp_tlb_addr1)880 invlpg_invpcid_handler(pmap_t smp_tlb_pmap, vm_offset_t smp_tlb_addr1)
881 {
882 	struct invpcid_descr d;
883 
884 #ifdef COUNT_XINVLTLB_HITS
885 	xhits_pg[PCPU_GET(cpuid)]++;
886 #endif /* COUNT_XINVLTLB_HITS */
887 #ifdef COUNT_IPIS
888 	(*ipi_invlpg_counts[PCPU_GET(cpuid)])++;
889 #endif /* COUNT_IPIS */
890 
891 	pmap_invlpg(smp_tlb_pmap, smp_tlb_addr1);
892 	if (smp_tlb_pmap == PCPU_GET(curpmap) &&
893 	    smp_tlb_pmap->pm_ucr3 != PMAP_NO_CR3 &&
894 	    PCPU_GET(ucr3_load_mask) == PMAP_UCR3_NOMASK) {
895 		d.pcid = pmap_get_pcid(smp_tlb_pmap) | PMAP_PCID_USER_PT;
896 		d.pad = 0;
897 		d.addr = smp_tlb_addr1;
898 		invpcid(&d, INVPCID_ADDR);
899 	}
900 }
901 
902 static void
invlpg_pcid_handler(pmap_t smp_tlb_pmap,vm_offset_t smp_tlb_addr1)903 invlpg_pcid_handler(pmap_t smp_tlb_pmap, vm_offset_t smp_tlb_addr1)
904 {
905 	uint64_t kcr3, ucr3;
906 	uint32_t pcid;
907 
908 #ifdef COUNT_XINVLTLB_HITS
909 	xhits_pg[PCPU_GET(cpuid)]++;
910 #endif /* COUNT_XINVLTLB_HITS */
911 #ifdef COUNT_IPIS
912 	(*ipi_invlpg_counts[PCPU_GET(cpuid)])++;
913 #endif /* COUNT_IPIS */
914 
915 	invlpg(smp_tlb_addr1);
916 	if (smp_tlb_pmap == PCPU_GET(curpmap) &&
917 	    (ucr3 = smp_tlb_pmap->pm_ucr3) != PMAP_NO_CR3 &&
918 	    PCPU_GET(ucr3_load_mask) == PMAP_UCR3_NOMASK) {
919 		pcid = pmap_get_pcid(smp_tlb_pmap);
920 		kcr3 = smp_tlb_pmap->pm_cr3 | pcid | CR3_PCID_SAVE;
921 		ucr3 |= pcid | PMAP_PCID_USER_PT | CR3_PCID_SAVE;
922 		pmap_pti_pcid_invlpg(ucr3, kcr3, smp_tlb_addr1);
923 	}
924 }
925 
926 static void
invlrng_handler(vm_offset_t smp_tlb_addr1,vm_offset_t smp_tlb_addr2)927 invlrng_handler(vm_offset_t smp_tlb_addr1, vm_offset_t smp_tlb_addr2)
928 {
929 	vm_offset_t addr;
930 
931 #ifdef COUNT_XINVLTLB_HITS
932 	xhits_rng[PCPU_GET(cpuid)]++;
933 #endif /* COUNT_XINVLTLB_HITS */
934 #ifdef COUNT_IPIS
935 	(*ipi_invlrng_counts[PCPU_GET(cpuid)])++;
936 #endif /* COUNT_IPIS */
937 
938 	addr = smp_tlb_addr1;
939 	do {
940 		invlpg(addr);
941 		addr += PAGE_SIZE;
942 	} while (addr < smp_tlb_addr2);
943 }
944 
945 static void
invlrng_invpcid_handler(pmap_t smp_tlb_pmap,vm_offset_t smp_tlb_addr1,vm_offset_t smp_tlb_addr2)946 invlrng_invpcid_handler(pmap_t smp_tlb_pmap, vm_offset_t smp_tlb_addr1,
947     vm_offset_t smp_tlb_addr2)
948 {
949 	struct invpcid_descr d;
950 	vm_offset_t addr;
951 
952 #ifdef COUNT_XINVLTLB_HITS
953 	xhits_rng[PCPU_GET(cpuid)]++;
954 #endif /* COUNT_XINVLTLB_HITS */
955 #ifdef COUNT_IPIS
956 	(*ipi_invlrng_counts[PCPU_GET(cpuid)])++;
957 #endif /* COUNT_IPIS */
958 
959 	addr = smp_tlb_addr1;
960 	if (smp_tlb_pmap == kernel_pmap && PCPU_GET(pcid_invlpg_workaround)) {
961 		struct invpcid_descr d = { 0 };
962 
963 		invpcid(&d, INVPCID_CTXGLOB);
964 	} else {
965 		do {
966 			invlpg(addr);
967 			addr += PAGE_SIZE;
968 		} while (addr < smp_tlb_addr2);
969 	}
970 	if (smp_tlb_pmap == PCPU_GET(curpmap) &&
971 	    smp_tlb_pmap->pm_ucr3 != PMAP_NO_CR3 &&
972 	    PCPU_GET(ucr3_load_mask) == PMAP_UCR3_NOMASK) {
973 		d.pcid = pmap_get_pcid(smp_tlb_pmap) | PMAP_PCID_USER_PT;
974 		d.pad = 0;
975 		d.addr = smp_tlb_addr1;
976 		do {
977 			invpcid(&d, INVPCID_ADDR);
978 			d.addr += PAGE_SIZE;
979 		} while (d.addr < smp_tlb_addr2);
980 	}
981 }
982 
983 static void
invlrng_pcid_handler(pmap_t smp_tlb_pmap,vm_offset_t smp_tlb_addr1,vm_offset_t smp_tlb_addr2)984 invlrng_pcid_handler(pmap_t smp_tlb_pmap, vm_offset_t smp_tlb_addr1,
985     vm_offset_t smp_tlb_addr2)
986 {
987 	vm_offset_t addr;
988 	uint64_t kcr3, ucr3;
989 	uint32_t pcid;
990 
991 #ifdef COUNT_XINVLTLB_HITS
992 	xhits_rng[PCPU_GET(cpuid)]++;
993 #endif /* COUNT_XINVLTLB_HITS */
994 #ifdef COUNT_IPIS
995 	(*ipi_invlrng_counts[PCPU_GET(cpuid)])++;
996 #endif /* COUNT_IPIS */
997 
998 	addr = smp_tlb_addr1;
999 	do {
1000 		invlpg(addr);
1001 		addr += PAGE_SIZE;
1002 	} while (addr < smp_tlb_addr2);
1003 	if (smp_tlb_pmap == PCPU_GET(curpmap) &&
1004 	    (ucr3 = smp_tlb_pmap->pm_ucr3) != PMAP_NO_CR3 &&
1005 	    PCPU_GET(ucr3_load_mask) == PMAP_UCR3_NOMASK) {
1006 		pcid = pmap_get_pcid(smp_tlb_pmap);
1007 		kcr3 = smp_tlb_pmap->pm_cr3 | pcid | CR3_PCID_SAVE;
1008 		ucr3 |= pcid | PMAP_PCID_USER_PT | CR3_PCID_SAVE;
1009 		pmap_pti_pcid_invlrng(ucr3, kcr3, smp_tlb_addr1, smp_tlb_addr2);
1010 	}
1011 }
1012 
1013 static void
invlcache_handler(void)1014 invlcache_handler(void)
1015 {
1016 #ifdef COUNT_IPIS
1017 	(*ipi_invlcache_counts[PCPU_GET(cpuid)])++;
1018 #endif /* COUNT_IPIS */
1019 	wbinvd();
1020 }
1021 
1022 static void
invlop_handler_one_req(enum invl_op_codes smp_tlb_op,pmap_t smp_tlb_pmap,vm_offset_t smp_tlb_addr1,vm_offset_t smp_tlb_addr2)1023 invlop_handler_one_req(enum invl_op_codes smp_tlb_op, pmap_t smp_tlb_pmap,
1024     vm_offset_t smp_tlb_addr1, vm_offset_t smp_tlb_addr2)
1025 {
1026 	switch (smp_tlb_op) {
1027 	case INVL_OP_TLB:
1028 		invltlb_handler(smp_tlb_pmap);
1029 		break;
1030 	case INVL_OP_TLB_INVPCID:
1031 		invltlb_invpcid_handler(smp_tlb_pmap);
1032 		break;
1033 	case INVL_OP_TLB_INVPCID_PTI:
1034 		invltlb_invpcid_pti_handler(smp_tlb_pmap);
1035 		break;
1036 	case INVL_OP_TLB_PCID:
1037 		invltlb_pcid_handler(smp_tlb_pmap);
1038 		break;
1039 	case INVL_OP_PGRNG:
1040 		invlrng_handler(smp_tlb_addr1, smp_tlb_addr2);
1041 		break;
1042 	case INVL_OP_PGRNG_INVPCID:
1043 		invlrng_invpcid_handler(smp_tlb_pmap, smp_tlb_addr1,
1044 		    smp_tlb_addr2);
1045 		break;
1046 	case INVL_OP_PGRNG_PCID:
1047 		invlrng_pcid_handler(smp_tlb_pmap, smp_tlb_addr1,
1048 		    smp_tlb_addr2);
1049 		break;
1050 	case INVL_OP_PG:
1051 		invlpg_handler(smp_tlb_addr1);
1052 		break;
1053 	case INVL_OP_PG_INVPCID:
1054 		invlpg_invpcid_handler(smp_tlb_pmap, smp_tlb_addr1);
1055 		break;
1056 	case INVL_OP_PG_PCID:
1057 		invlpg_pcid_handler(smp_tlb_pmap, smp_tlb_addr1);
1058 		break;
1059 	case INVL_OP_CACHE:
1060 		invlcache_handler();
1061 		break;
1062 	default:
1063 		__assert_unreachable();
1064 		break;
1065 	}
1066 }
1067 
1068 void
invlop_handler(void)1069 invlop_handler(void)
1070 {
1071 	struct pcpu *initiator_pc;
1072 	pmap_t smp_tlb_pmap;
1073 	vm_offset_t smp_tlb_addr1, smp_tlb_addr2;
1074 	u_int initiator_cpu_id;
1075 	enum invl_op_codes smp_tlb_op;
1076 	uint32_t *scoreboard, smp_tlb_gen;
1077 
1078 	scoreboard = invl_scoreboard_getcpu(PCPU_GET(cpuid));
1079 	for (;;) {
1080 		for (initiator_cpu_id = 0; initiator_cpu_id <= mp_maxid;
1081 		    initiator_cpu_id++) {
1082 			if (atomic_load_int(&scoreboard[initiator_cpu_id]) == 0)
1083 				break;
1084 		}
1085 		if (initiator_cpu_id > mp_maxid)
1086 			break;
1087 		initiator_pc = cpuid_to_pcpu[initiator_cpu_id];
1088 
1089 		/*
1090 		 * This acquire fence and its corresponding release
1091 		 * fence in smp_targeted_tlb_shootdown() is between
1092 		 * reading zero scoreboard slot and accessing PCPU of
1093 		 * initiator for pc_smp_tlb values.
1094 		 */
1095 		atomic_thread_fence_acq();
1096 		smp_tlb_pmap = initiator_pc->pc_smp_tlb_pmap;
1097 		smp_tlb_addr1 = initiator_pc->pc_smp_tlb_addr1;
1098 		smp_tlb_addr2 = initiator_pc->pc_smp_tlb_addr2;
1099 		smp_tlb_op = initiator_pc->pc_smp_tlb_op;
1100 		smp_tlb_gen = initiator_pc->pc_smp_tlb_gen;
1101 
1102 		/*
1103 		 * Ensure that we do not make our scoreboard
1104 		 * notification visible to the initiator until the
1105 		 * pc_smp_tlb values are read.  The corresponding
1106 		 * fence is implicitly provided by the barrier in the
1107 		 * IPI send operation before the APIC ICR register
1108 		 * write.
1109 		 *
1110 		 * As an optimization, the request is acknowledged
1111 		 * before the actual invalidation is performed.  It is
1112 		 * safe because target CPU cannot return to userspace
1113 		 * before handler finishes. Only NMI can preempt the
1114 		 * handler, but NMI would see the kernel handler frame
1115 		 * and not touch not-invalidated user page table.
1116 		 */
1117 		atomic_thread_fence_acq();
1118 		atomic_store_int(&scoreboard[initiator_cpu_id], smp_tlb_gen);
1119 
1120 		invlop_handler_one_req(smp_tlb_op, smp_tlb_pmap, smp_tlb_addr1,
1121 		    smp_tlb_addr2);
1122 	}
1123 }
1124