xref: /netbsd-src/sys/uvm/pmap/pmap_tlb.c (revision 53b02e147d4ed531c0d2a5ca9b3e8026ba3e99b5)
1 /*	$NetBSD: pmap_tlb.c,v 1.49 2021/10/27 06:54:15 simonb Exp $	*/
2 
3 /*-
4  * Copyright (c) 2010 The NetBSD Foundation, Inc.
5  * All rights reserved.
6  *
7  * This code is derived from software contributed to The NetBSD Foundation
8  * by Matt Thomas at 3am Software Foundry.
9  *
10  * Redistribution and use in source and binary forms, with or without
11  * modification, are permitted provided that the following conditions
12  * are met:
13  * 1. Redistributions of source code must retain the above copyright
14  *    notice, this list of conditions and the following disclaimer.
15  * 2. Redistributions in binary form must reproduce the above copyright
16  *    notice, this list of conditions and the following disclaimer in the
17  *    documentation and/or other materials provided with the distribution.
18  *
19  * THIS SOFTWARE IS PROVIDED BY THE NETBSD FOUNDATION, INC. AND CONTRIBUTORS
20  * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED
21  * TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
22  * PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE FOUNDATION OR CONTRIBUTORS
23  * BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
24  * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
25  * SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
26  * INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
27  * CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
28  * ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
29  * POSSIBILITY OF SUCH DAMAGE.
30  */
31 
32 #include <sys/cdefs.h>
33 
34 __KERNEL_RCSID(0, "$NetBSD: pmap_tlb.c,v 1.49 2021/10/27 06:54:15 simonb Exp $");
35 
36 /*
37  * Manages address spaces in a TLB.
38  *
39  * Normally there is a 1:1 mapping between a TLB and a CPU.  However, some
40  * implementations may share a TLB between multiple CPUs (really CPU thread
41  * contexts).  This requires the TLB abstraction to be separated from the
42  * CPU abstraction.  It also requires that the TLB be locked while doing
43  * TLB activities.
44  *
45  * For each TLB, we track the ASIDs in use in a bitmap and a list of pmaps
46  * that have a valid ASID.
47  *
48  * We allocate ASIDs in increasing order until we have exhausted the supply,
49  * then reinitialize the ASID space, and start allocating again at 1.  When
50  * allocating from the ASID bitmap, we skip any ASID who has a corresponding
51  * bit set in the ASID bitmap.  Eventually this causes the ASID bitmap to fill
52  * and, when completely filled, a reinitialization of the ASID space.
53  *
54  * To reinitialize the ASID space, the ASID bitmap is reset and then the ASIDs
55  * of non-kernel TLB entries get recorded in the ASID bitmap.  If the entries
56  * in TLB consume more than half of the ASID space, all ASIDs are invalidated,
57  * the ASID bitmap is recleared, and the list of pmaps is emptied.  Otherwise,
58  * (the normal case), any ASID present in the TLB (even those which are no
59  * longer used by a pmap) will remain active (allocated) and all other ASIDs
60  * will be freed.  If the size of the TLB is much smaller than the ASID space,
61  * this algorithm completely avoids TLB invalidation.
62  *
63  * For multiprocessors, we also have to deal TLB invalidation requests from
64  * other CPUs, some of which are dealt with the reinitialization of the ASID
65  * space.  Whereas above we keep the ASIDs of those pmaps which have active
66  * TLB entries, this type of reinitialization preserves the ASIDs of any
67  * "onproc" user pmap and all other ASIDs will be freed.  We must do this
68  * since we can't change the current ASID.
69  *
70  * Each pmap has two bitmaps: pm_active and pm_onproc.  Each bit in pm_active
71  * indicates whether that pmap has an allocated ASID for a CPU.  Each bit in
72  * pm_onproc indicates that the pmap's ASID is in use, i.e. a CPU has it in its
73  * "current ASID" field, e.g. the ASID field of the COP 0 register EntryHi for
74  * MIPS, or the ASID field of TTBR0 for AA64.  The bit number used in these
75  * bitmaps comes from the CPU's cpu_index().  Even though these bitmaps contain
76  * the bits for all CPUs, the bits that correspond to the bits belonging to
77  * the CPUs sharing a TLB can only be manipulated while holding that TLB's
78  * lock.  Atomic ops must be used to update them since multiple CPUs may be
79  * changing different sets of bits at same time but these sets never overlap.
80  *
81  * When a change to the local TLB may require a change in the TLB's of other
82  * CPUs, we try to avoid sending an IPI if at all possible.  For instance, if
83  * we are updating a PTE and that PTE previously was invalid and therefore
84  * couldn't support an active mapping, there's no need for an IPI since there
85  * can't be a TLB entry to invalidate.  The other case is when we change a PTE
86  * to be modified we just update the local TLB.  If another TLB has a stale
87  * entry, a TLB MOD exception will be raised and that will cause the local TLB
88  * to be updated.
89  *
90  * We never need to update a non-local TLB if the pmap doesn't have a valid
91  * ASID for that TLB.  If it does have a valid ASID but isn't current "onproc"
92  * we simply reset its ASID for that TLB and then when it goes "onproc" it
93  * will allocate a new ASID and any existing TLB entries will be orphaned.
94  * Only in the case that pmap has an "onproc" ASID do we actually have to send
95  * an IPI.
96  *
97  * Once we determined we must send an IPI to shootdown a TLB, we need to send
98  * it to one of CPUs that share that TLB.  We choose the lowest numbered CPU
99  * that has one of the pmap's ASID "onproc".  In reality, any CPU sharing that
100  * TLB would do, but interrupting an active CPU seems best.
101  *
102  * A TLB might have multiple shootdowns active concurrently.  The shootdown
103  * logic compresses these into a few cases:
104  *	0) nobody needs to have its TLB entries invalidated
105  *	1) one ASID needs to have its TLB entries invalidated
106  *	2) more than one ASID needs to have its TLB entries invalidated
107  *	3) the kernel needs to have its TLB entries invalidated
108  *	4) the kernel and one or more ASID need their TLB entries invalidated.
109  *
110  * And for each case we do:
111  *	0) nothing,
112  *	1) if that ASID is still "onproc", we invalidate the TLB entries for
113  *	   that single ASID.  If not, just reset the pmap's ASID to invalidate
114  *	   and let it allocate a new ASID the next time it goes "onproc",
115  *	2) we reinitialize the ASID space (preserving any "onproc" ASIDs) and
116  *	   invalidate all non-wired non-global TLB entries,
117  *	3) we invalidate all of the non-wired global TLB entries,
118  *	4) we reinitialize the ASID space (again preserving any "onproc" ASIDs)
119  *	   invalidate all non-wired TLB entries.
120  *
121  * As you can see, shootdowns are not concerned with addresses, just address
122  * spaces.  Since the number of TLB entries is usually quite small, this avoids
123  * a lot of overhead for not much gain.
124  */
125 
126 #define __PMAP_PRIVATE
127 
128 #include "opt_multiprocessor.h"
129 
130 #include <sys/param.h>
131 
132 #include <sys/atomic.h>
133 #include <sys/cpu.h>
134 #include <sys/kernel.h>			/* for cold */
135 #include <sys/mutex.h>
136 #include <sys/proc.h>
137 #include <sys/systm.h>
138 
139 #include <uvm/uvm.h>
140 
141 static kmutex_t pmap_tlb0_lock __cacheline_aligned;
142 
143 #define	IFCONSTANT(x)	(__builtin_constant_p((x)) ? (x) : 0)
144 
145 #if KERNEL_PID > 31
146 #error "KERNEL_PID expected in range 0-31"
147 #endif
148 
149 #define	TLBINFO_ASID_MARK_UNUSED(ti, asid) \
150 	__BITMAP_CLR((asid), &(ti)->ti_asid_bitmap)
151 #define	TLBINFO_ASID_MARK_USED(ti, asid) \
152 	__BITMAP_SET((asid), &(ti)->ti_asid_bitmap)
153 #define	TLBINFO_ASID_INUSE_P(ti, asid) \
154 	__BITMAP_ISSET((asid), &(ti)->ti_asid_bitmap)
155 #define	TLBINFO_ASID_RESET(ti) \
156 	do {								\
157 		__BITMAP_ZERO(&ti->ti_asid_bitmap);			\
158 		for (tlb_asid_t asid = 0; asid <= KERNEL_PID; asid++) 	\
159 			TLBINFO_ASID_MARK_USED(ti, asid);	 	\
160 	} while (0)
161 #define	TLBINFO_ASID_INITIAL_FREE(asid_max) \
162 	(asid_max + 1 /* 0 */ - (1 + KERNEL_PID))
163 
164 struct pmap_tlb_info pmap_tlb0_info = {
165 	.ti_name = "tlb0",
166 	.ti_asid_hint = KERNEL_PID + 1,
167 #ifdef PMAP_TLB_NUM_PIDS
168 	.ti_asid_max = IFCONSTANT(PMAP_TLB_NUM_PIDS - 1),
169 	.ti_asids_free = IFCONSTANT(
170 		TLBINFO_ASID_INITIAL_FREE(PMAP_TLB_NUM_PIDS - 1)),
171 #endif
172 	.ti_asid_bitmap._b[0] = __BITS(0, KERNEL_PID),
173 #ifdef PMAP_TLB_WIRED_UPAGES
174 	.ti_wired = PMAP_TLB_WIRED_UPAGES,
175 #endif
176 	.ti_lock = &pmap_tlb0_lock,
177 	.ti_pais = LIST_HEAD_INITIALIZER(pmap_tlb0_info.ti_pais),
178 #if defined(MULTIPROCESSOR) && PMAP_TLB_MAX > 1
179 	.ti_tlbinvop = TLBINV_NOBODY,
180 #endif
181 };
182 
183 #undef IFCONSTANT
184 
185 #if defined(MULTIPROCESSOR) && PMAP_TLB_MAX > 1
186 struct pmap_tlb_info *pmap_tlbs[PMAP_TLB_MAX] = {
187 	[0] = &pmap_tlb0_info,
188 };
189 u_int pmap_ntlbs = 1;
190 #endif
191 
192 #ifdef MULTIPROCESSOR
193 __unused static inline bool
194 pmap_tlb_intersecting_active_p(pmap_t pm, struct pmap_tlb_info *ti)
195 {
196 #if PMAP_TLB_MAX == 1
197 	return !kcpuset_iszero(pm->pm_active);
198 #else
199 	return kcpuset_intersecting_p(pm->pm_active, ti->ti_kcpuset);
200 #endif
201 }
202 
203 static inline bool
204 pmap_tlb_intersecting_onproc_p(pmap_t pm, struct pmap_tlb_info *ti)
205 {
206 #if PMAP_TLB_MAX == 1
207 	return !kcpuset_iszero(pm->pm_onproc);
208 #else
209 	return kcpuset_intersecting_p(pm->pm_onproc, ti->ti_kcpuset);
210 #endif
211 }
212 #endif
213 
214 static void
215 pmap_tlb_pai_check(struct pmap_tlb_info *ti, bool locked_p)
216 {
217 	UVMHIST_FUNC(__func__);
218 	UVMHIST_CALLARGS(maphist, "(ti=%#jx)", (uintptr_t)ti, 0, 0, 0);
219 
220 #ifdef DIAGNOSTIC
221 	struct pmap_asid_info *pai;
222 	if (!locked_p)
223 		TLBINFO_LOCK(ti);
224 	LIST_FOREACH(pai, &ti->ti_pais, pai_link) {
225 		KASSERT(pai != NULL);
226 		KASSERT(PAI_PMAP(pai, ti) != pmap_kernel());
227 		KASSERT(pai->pai_asid > KERNEL_PID);
228 		KASSERTMSG(pai->pai_asid <= ti->ti_asid_max,
229 		    "pm %p asid %#x", PAI_PMAP(pai, ti), pai->pai_asid);
230 		KASSERTMSG(TLBINFO_ASID_INUSE_P(ti, pai->pai_asid),
231 		    "pm %p asid %u", PAI_PMAP(pai, ti), pai->pai_asid);
232 #ifdef MULTIPROCESSOR
233 		KASSERT(pmap_tlb_intersecting_active_p(PAI_PMAP(pai, ti), ti));
234 #endif
235 	}
236 	if (!locked_p)
237 		TLBINFO_UNLOCK(ti);
238 #endif
239 	UVMHIST_LOG(maphist, " <-- done", 0, 0, 0, 0);
240 }
241 
242 static void
243 pmap_tlb_pai_reset(struct pmap_tlb_info *ti, struct pmap_asid_info *pai,
244 	struct pmap *pm)
245 {
246 	UVMHIST_FUNC(__func__);
247 	UVMHIST_CALLARGS(maphist, "(ti=%#jx, pai=%#jx, pm=%#jx): asid %u",
248 	    (uintptr_t)ti, (uintptr_t)pai, (uintptr_t)pm, pai->pai_asid);
249 
250 	/*
251 	 * We must have an ASID but it must not be onproc (on a processor).
252 	 */
253 	KASSERT(pai->pai_asid > KERNEL_PID);
254 	KASSERT(pai->pai_asid <= ti->ti_asid_max);
255 #if defined(MULTIPROCESSOR)
256 	KASSERT(pmap_tlb_intersecting_active_p(pm, ti));
257 	KASSERT(!pmap_tlb_intersecting_onproc_p(pm, ti));
258 #endif
259 	LIST_REMOVE(pai, pai_link);
260 #ifdef DIAGNOSTIC
261 	pai->pai_link.le_prev = NULL;	/* tagged as unlinked */
262 #endif
263 	/*
264 	 * If the platform has a cheap way to flush ASIDs then free the ASID
265 	 * back into the pool.  On multiprocessor systems, we will flush the
266 	 * ASID from the TLB when it's allocated.  That way we know the flush
267 	 * was always done in the correct TLB space.  On uniprocessor systems,
268 	 * just do the flush now since we know that it has been used.  This has
269 	 * a bit less overhead.  Either way, this will mean that we will only
270 	 * need to flush all ASIDs if all ASIDs are in use and we need to
271 	 * allocate a new one.
272 	 */
273 	if (PMAP_TLB_FLUSH_ASID_ON_RESET) {
274 #ifndef MULTIPROCESSOR
275 		UVMHIST_LOG(maphist, " ... asid %u flushed", pai->pai_asid, 0,
276 		    0, 0);
277 		tlb_invalidate_asids(pai->pai_asid, pai->pai_asid);
278 #endif
279 		if (TLBINFO_ASID_INUSE_P(ti, pai->pai_asid)) {
280 			UVMHIST_LOG(maphist, " ... asid marked unused",
281 			    pai->pai_asid, 0, 0, 0);
282 			TLBINFO_ASID_MARK_UNUSED(ti, pai->pai_asid);
283 			ti->ti_asids_free++;
284 		}
285 	}
286 	/*
287 	 * Note that we don't mark the ASID as not in use in the TLB's ASID
288 	 * bitmap (thus it can't be allocated until the ASID space is exhausted
289 	 * and therefore reinitialized).  We don't want to flush the TLB for
290 	 * entries belonging to this ASID so we will let natural TLB entry
291 	 * replacement flush them out of the TLB.  Any new entries for this
292 	 * pmap will need a new ASID allocated.
293 	 */
294 	pai->pai_asid = 0;
295 
296 #if defined(MULTIPROCESSOR)
297 	/*
298 	 * The bits in pm_active belonging to this TLB can only be changed
299 	 * while this TLB's lock is held.
300 	 */
301 #if PMAP_TLB_MAX == 1
302 	kcpuset_zero(pm->pm_active);
303 #else
304 	kcpuset_remove(pm->pm_active, ti->ti_kcpuset);
305 #endif
306 	KASSERT(!pmap_tlb_intersecting_active_p(pm, ti));
307 #endif /* MULTIPROCESSOR */
308 
309 	UVMHIST_LOG(maphist, " <-- done", 0, 0, 0, 0);
310 }
311 
312 void
313 pmap_tlb_info_evcnt_attach(struct pmap_tlb_info *ti)
314 {
315 #if defined(MULTIPROCESSOR) && !defined(PMAP_TLB_NO_SYNCI_EVCNT)
316 	evcnt_attach_dynamic_nozero(&ti->ti_evcnt_synci_desired,
317 	    EVCNT_TYPE_MISC, NULL,
318 	    ti->ti_name, "icache syncs desired");
319 	evcnt_attach_dynamic_nozero(&ti->ti_evcnt_synci_asts,
320 	    EVCNT_TYPE_MISC, &ti->ti_evcnt_synci_desired,
321 	    ti->ti_name, "icache sync asts");
322 	evcnt_attach_dynamic_nozero(&ti->ti_evcnt_synci_all,
323 	    EVCNT_TYPE_MISC, &ti->ti_evcnt_synci_asts,
324 	    ti->ti_name, "icache full syncs");
325 	evcnt_attach_dynamic_nozero(&ti->ti_evcnt_synci_pages,
326 	    EVCNT_TYPE_MISC, &ti->ti_evcnt_synci_asts,
327 	    ti->ti_name, "icache pages synced");
328 	evcnt_attach_dynamic_nozero(&ti->ti_evcnt_synci_duplicate,
329 	    EVCNT_TYPE_MISC, &ti->ti_evcnt_synci_desired,
330 	    ti->ti_name, "icache dup pages skipped");
331 	evcnt_attach_dynamic_nozero(&ti->ti_evcnt_synci_deferred,
332 	    EVCNT_TYPE_MISC, &ti->ti_evcnt_synci_desired,
333 	    ti->ti_name, "icache pages deferred");
334 #endif /* MULTIPROCESSOR && !PMAP_TLB_NO_SYNCI_EVCNT */
335 	evcnt_attach_dynamic_nozero(&ti->ti_evcnt_asid_reinits,
336 	    EVCNT_TYPE_MISC, NULL,
337 	    ti->ti_name, "asid pool reinit");
338 }
339 
340 void
341 pmap_tlb_info_init(struct pmap_tlb_info *ti)
342 {
343 #if defined(MULTIPROCESSOR)
344 #if PMAP_TLB_MAX == 1
345 	KASSERT(ti == &pmap_tlb0_info);
346 #else
347 	if (ti != &pmap_tlb0_info) {
348 		KASSERT(pmap_ntlbs < PMAP_TLB_MAX);
349 
350 		KASSERT(pmap_tlbs[pmap_ntlbs] == NULL);
351 
352 		ti->ti_lock = mutex_obj_alloc(MUTEX_DEFAULT, IPL_SCHED);
353 		TLBINFO_ASID_RESET(ti);
354 		ti->ti_asid_hint = KERNEL_PID + 1;
355 		ti->ti_asid_max = pmap_tlbs[0]->ti_asid_max;
356 		ti->ti_asids_free = TLBINFO_ASID_INITIAL_FREE(ti->ti_asid_max);
357 		ti->ti_tlbinvop = TLBINV_NOBODY;
358 		ti->ti_victim = NULL;
359 		kcpuset_create(&ti->ti_kcpuset, true);
360 		ti->ti_index = pmap_ntlbs++;
361 		ti->ti_wired = 0;
362 		pmap_tlbs[ti->ti_index] = ti;
363 		snprintf(ti->ti_name, sizeof(ti->ti_name), "tlb%u",
364 		    ti->ti_index);
365 		pmap_tlb_info_evcnt_attach(ti);
366 
367 		KASSERT(ti->ti_asid_max < PMAP_TLB_BITMAP_LENGTH);
368 		return;
369 	}
370 #endif
371 #endif /* MULTIPROCESSOR */
372 	KASSERT(ti == &pmap_tlb0_info);
373 	KASSERT(ti->ti_lock == &pmap_tlb0_lock);
374 
375 	mutex_init(ti->ti_lock, MUTEX_DEFAULT, IPL_SCHED);
376 #if defined(MULTIPROCESSOR) && PMAP_TLB_MAX > 1
377 	kcpuset_create(&ti->ti_kcpuset, true);
378 	kcpuset_set(ti->ti_kcpuset, cpu_index(curcpu()));
379 #endif
380 
381 	const tlb_asid_t asid_max = pmap_md_tlb_asid_max();
382 	if (ti->ti_asid_max == 0 || asid_max < ti->ti_asid_max) {
383 		ti->ti_asid_max = asid_max;
384 		ti->ti_asids_free = TLBINFO_ASID_INITIAL_FREE(ti->ti_asid_max);
385 	}
386 
387 	KASSERT(ti->ti_asid_max < PMAP_TLB_BITMAP_LENGTH);
388 }
389 
390 #if defined(MULTIPROCESSOR)
391 void
392 pmap_tlb_info_attach(struct pmap_tlb_info *ti, struct cpu_info *ci)
393 {
394 	KASSERT(!CPU_IS_PRIMARY(ci));
395 	KASSERT(ci->ci_data.cpu_idlelwp != NULL);
396 	KASSERT(cold);
397 
398 	TLBINFO_LOCK(ti);
399 #if PMAP_TLB_MAX > 1
400 	kcpuset_set(ti->ti_kcpuset, cpu_index(ci));
401 	cpu_set_tlb_info(ci, ti);
402 #endif
403 
404 	/*
405 	 * Do any MD tlb info init.
406 	 */
407 	pmap_md_tlb_info_attach(ti, ci);
408 
409 	/*
410 	 * The kernel pmap uses the kcpuset_running set so it's always
411 	 * up-to-date.
412 	 */
413 	TLBINFO_UNLOCK(ti);
414 }
415 #endif /* MULTIPROCESSOR */
416 
417 #ifdef DIAGNOSTIC
418 static size_t
419 pmap_tlb_asid_count(struct pmap_tlb_info *ti)
420 {
421 	size_t count = 0;
422 	for (tlb_asid_t asid = 1; asid <= ti->ti_asid_max; asid++) {
423 		if (TLBINFO_ASID_INUSE_P(ti, asid))
424 			count++;
425 	}
426 	return count;
427 }
428 #endif
429 
430 static void
431 pmap_tlb_asid_reinitialize(struct pmap_tlb_info *ti, enum tlb_invalidate_op op)
432 {
433 	UVMHIST_FUNC(__func__);
434 	UVMHIST_CALLARGS(maphist, "(ti=%#jx, op=%ju)", (uintptr_t)ti, op, 0, 0);
435 
436 	pmap_tlb_pai_check(ti, true);
437 
438 	ti->ti_evcnt_asid_reinits.ev_count++;
439 
440 	/*
441 	 * First, clear the ASID bitmap (except for ASID 0 which belongs
442 	 * to the kernel).
443 	 */
444 	ti->ti_asids_free = TLBINFO_ASID_INITIAL_FREE(ti->ti_asid_max);
445 	ti->ti_asid_hint = KERNEL_PID + 1;
446 	TLBINFO_ASID_RESET(ti);
447 
448 	switch (op) {
449 #if defined(MULTIPROCESSOR) && defined(PMAP_TLB_NEED_SHOOTDOWN)
450 	case TLBINV_ALL:
451 		tlb_invalidate_all();
452 		break;
453 	case TLBINV_ALLUSER:
454 		tlb_invalidate_asids(KERNEL_PID + 1, ti->ti_asid_max);
455 		break;
456 #endif /* MULTIPROCESSOR && PMAP_TLB_NEED_SHOOTDOWN */
457 	case TLBINV_NOBODY: {
458 		/*
459 		 * If we are just reclaiming ASIDs in the TLB, let's go find
460 		 * what ASIDs are in use in the TLB.  Since this is a
461 		 * semi-expensive operation, we don't want to do it too often.
462 		 * So if more half of the ASIDs are in use, we don't have
463 		 * enough free ASIDs so invalidate the TLB entries with ASIDs
464 		 * and clear the ASID bitmap.  That will force everyone to
465 		 * allocate a new ASID.
466 		 */
467 #if !defined(MULTIPROCESSOR) || defined(PMAP_TLB_NEED_SHOOTDOWN)
468 		pmap_tlb_asid_check();
469 		const u_int asids_found = tlb_record_asids(
470 		    ti->ti_asid_bitmap._b, ti->ti_asid_max);
471 		pmap_tlb_asid_check();
472 #ifdef DIAGNOSTIC
473 		const u_int asids_count = pmap_tlb_asid_count(ti);
474 #endif
475 		KASSERTMSG(asids_found == asids_count,
476 		    "found %u != count %u", asids_found, asids_count);
477 		if (__predict_false(asids_found >= ti->ti_asid_max / 2)) {
478 			tlb_invalidate_asids(KERNEL_PID + 1, ti->ti_asid_max);
479 #else /* MULTIPROCESSOR && !PMAP_TLB_NEED_SHOOTDOWN */
480 			/*
481 			 * For those systems (PowerPC) that don't require
482 			 * cross cpu TLB shootdowns, we have to invalidate the
483 			 * entire TLB because we can't record the ASIDs in use
484 			 * on the other CPUs.  This is hopefully cheaper than
485 			 * than trying to use an IPI to record all the ASIDs
486 			 * on all the CPUs (which would be a synchronization
487 			 * nightmare).
488 			 */
489 			tlb_invalidate_all();
490 #endif /* MULTIPROCESSOR && !PMAP_TLB_NEED_SHOOTDOWN */
491 			TLBINFO_ASID_RESET(ti);
492 			ti->ti_asids_free = TLBINFO_ASID_INITIAL_FREE(
493 				ti->ti_asid_max);
494 #if !defined(MULTIPROCESSOR) || defined(PMAP_TLB_NEED_SHOOTDOWN)
495 		} else {
496 			ti->ti_asids_free -= asids_found;
497 		}
498 #endif /* !MULTIPROCESSOR || PMAP_TLB_NEED_SHOOTDOWN */
499 		KASSERTMSG(ti->ti_asids_free <= ti->ti_asid_max, "%u",
500 		    ti->ti_asids_free);
501 		break;
502 	}
503 	default:
504 		panic("%s: unexpected op %d", __func__, op);
505 	}
506 
507 	/*
508 	 * Now go through the active ASIDs.  If the ASID is on a processor or
509 	 * we aren't invalidating all ASIDs and the TLB has an entry owned by
510 	 * that ASID, mark it as in use.  Otherwise release the ASID.
511 	 */
512 	struct pmap_asid_info *pai, *next;
513 	for (pai = LIST_FIRST(&ti->ti_pais); pai != NULL; pai = next) {
514 		struct pmap * const pm = PAI_PMAP(pai, ti);
515 		next = LIST_NEXT(pai, pai_link);
516 		KASSERT(pm != pmap_kernel());
517 		KASSERT(pai->pai_asid > KERNEL_PID);
518 #if defined(MULTIPROCESSOR)
519 		if (pmap_tlb_intersecting_onproc_p(pm, ti)) {
520 			if (!TLBINFO_ASID_INUSE_P(ti, pai->pai_asid)) {
521 				TLBINFO_ASID_MARK_USED(ti, pai->pai_asid);
522 				ti->ti_asids_free--;
523 			}
524 			continue;
525 		}
526 #endif /* MULTIPROCESSOR */
527 		if (TLBINFO_ASID_INUSE_P(ti, pai->pai_asid)) {
528 			KASSERT(op == TLBINV_NOBODY);
529 		} else {
530 			pmap_tlb_pai_reset(ti, pai, pm);
531 		}
532 	}
533 #ifdef DIAGNOSTIC
534 	size_t free_count __diagused = ti->ti_asid_max - pmap_tlb_asid_count(ti);
535 	KASSERTMSG(free_count == ti->ti_asids_free,
536 	    "bitmap error: %zu != %u", free_count, ti->ti_asids_free);
537 #endif
538 	UVMHIST_LOG(maphist, " <-- done", 0, 0, 0, 0);
539 }
540 
541 #if defined(MULTIPROCESSOR) && defined(PMAP_TLB_NEED_SHOOTDOWN)
542 #if PMAP_TLB_MAX == 1
543 #error shootdown not required for single TLB systems
544 #endif
545 void
546 pmap_tlb_shootdown_process(void)
547 {
548 	struct cpu_info * const ci = curcpu();
549 	struct pmap_tlb_info * const ti = cpu_tlb_info(ci);
550 #ifdef DIAGNOSTIC
551 	struct pmap * const pm = curlwp->l_proc->p_vmspace->vm_map.pmap;
552 #endif
553 
554 	KASSERT(cpu_intr_p());
555 	KASSERTMSG(ci->ci_cpl >= IPL_SCHED, "%s: cpl (%d) < IPL_SCHED (%d)",
556 	    __func__, ci->ci_cpl, IPL_SCHED);
557 
558 	TLBINFO_LOCK(ti);
559 
560 	switch (ti->ti_tlbinvop) {
561 	case TLBINV_ONE: {
562 		/*
563 		 * We only need to invalidate one user ASID.
564 		 */
565 		struct pmap_asid_info * const pai = PMAP_PAI(ti->ti_victim, ti);
566 		KASSERT(ti->ti_victim != pmap_kernel());
567 		if (pmap_tlb_intersecting_onproc_p(ti->ti_victim, ti)) {
568 			/*
569 			 * The victim is an active pmap so we will just
570 			 * invalidate its TLB entries.
571 			 */
572 			KASSERT(pai->pai_asid > KERNEL_PID);
573 			pmap_tlb_asid_check();
574 			tlb_invalidate_asids(pai->pai_asid, pai->pai_asid);
575 			pmap_tlb_asid_check();
576 		} else if (pai->pai_asid) {
577 			/*
578 			 * The victim is no longer an active pmap for this TLB.
579 			 * So simply clear its ASID and when pmap_activate is
580 			 * next called for this pmap, it will allocate a new
581 			 * ASID.
582 			 */
583 			KASSERT(!pmap_tlb_intersecting_onproc_p(pm, ti));
584 			pmap_tlb_pai_reset(ti, pai, PAI_PMAP(pai, ti));
585 		}
586 		break;
587 	}
588 	case TLBINV_ALLUSER:
589 		/*
590 		 * Flush all user TLB entries.
591 		 */
592 		pmap_tlb_asid_reinitialize(ti, TLBINV_ALLUSER);
593 		break;
594 	case TLBINV_ALLKERNEL:
595 		/*
596 		 * We need to invalidate all global TLB entries.
597 		 */
598 		pmap_tlb_asid_check();
599 		tlb_invalidate_globals();
600 		pmap_tlb_asid_check();
601 		break;
602 	case TLBINV_ALL:
603 		/*
604 		 * Flush all the TLB entries (user and kernel).
605 		 */
606 		pmap_tlb_asid_reinitialize(ti, TLBINV_ALL);
607 		break;
608 	case TLBINV_NOBODY:
609 		/*
610 		 * Might be spurious or another SMT CPU sharing this TLB
611 		 * could have already done the work.
612 		 */
613 		break;
614 	}
615 
616 	/*
617 	 * Indicate we are done with shutdown event.
618 	 */
619 	ti->ti_victim = NULL;
620 	ti->ti_tlbinvop = TLBINV_NOBODY;
621 	TLBINFO_UNLOCK(ti);
622 }
623 
624 /*
625  * This state machine could be encoded into an array of integers but since all
626  * the values fit in 3 bits, the 5 entry "table" fits in a 16 bit value which
627  * can be loaded in a single instruction.
628  */
629 #define	TLBINV_MAP(op, nobody, one, alluser, allkernel, all)	\
630 	((((   (nobody) << 3 * TLBINV_NOBODY)			\
631 	 | (      (one) << 3 * TLBINV_ONE)			\
632 	 | (  (alluser) << 3 * TLBINV_ALLUSER)			\
633 	 | ((allkernel) << 3 * TLBINV_ALLKERNEL)		\
634 	 | (      (all) << 3 * TLBINV_ALL)) >> 3 * (op)) & 7)
635 
636 #define	TLBINV_USER_MAP(op)	\
637 	TLBINV_MAP(op, TLBINV_ONE, TLBINV_ALLUSER, TLBINV_ALLUSER,	\
638 	    TLBINV_ALL, TLBINV_ALL)
639 
640 #define	TLBINV_KERNEL_MAP(op)	\
641 	TLBINV_MAP(op, TLBINV_ALLKERNEL, TLBINV_ALL, TLBINV_ALL,	\
642 	    TLBINV_ALLKERNEL, TLBINV_ALL)
643 
644 bool
645 pmap_tlb_shootdown_bystanders(pmap_t pm)
646 {
647 	/*
648 	 * We don't need to deal with our own TLB.
649 	 */
650 
651 	UVMHIST_FUNC(__func__);
652 	UVMHIST_CALLARGS(maphist, "pm %#jx", (uintptr_t)pm, 0, 0, 0);
653 
654 	const struct cpu_info * const ci = curcpu();
655 	kcpuset_t *pm_active = ci->ci_shootdowncpus;
656 	kcpuset_copy(pm_active, pm->pm_active);
657 	kcpuset_remove(pm_active, cpu_tlb_info(curcpu())->ti_kcpuset);
658 	const bool kernel_p = (pm == pmap_kernel());
659 	bool ipi_sent = false;
660 
661 	/*
662 	 * If pm_active gets more bits set, then it's after all our changes
663 	 * have been made so they will already be cognizant of them.
664 	 */
665 
666 	for (size_t i = 0; !kcpuset_iszero(pm_active); i++) {
667 		KASSERT(i < pmap_ntlbs);
668 		struct pmap_tlb_info * const ti = pmap_tlbs[i];
669 		KASSERT(tlbinfo_index(ti) == i);
670 		/*
671 		 * Skip this TLB if there are no active mappings for it.
672 		 */
673 		if (!kcpuset_intersecting_p(pm_active, ti->ti_kcpuset))
674 			continue;
675 		struct pmap_asid_info * const pai = PMAP_PAI(pm, ti);
676 		kcpuset_remove(pm_active, ti->ti_kcpuset);
677 		TLBINFO_LOCK(ti);
678 		cpuid_t j = kcpuset_ffs_intersecting(pm->pm_onproc,
679 		    ti->ti_kcpuset);
680 		// post decrement since ffs returns bit + 1 or 0 if no bit
681 		if (j-- > 0) {
682 			if (kernel_p) {
683 				ti->ti_tlbinvop =
684 				    TLBINV_KERNEL_MAP(ti->ti_tlbinvop);
685 				ti->ti_victim = NULL;
686 			} else {
687 				KASSERT(pai->pai_asid);
688 				if (__predict_false(ti->ti_victim == pm)) {
689 					KASSERT(ti->ti_tlbinvop == TLBINV_ONE);
690 					/*
691 					 * We still need to invalidate this one
692 					 * ASID so there's nothing to change.
693 					 */
694 				} else {
695 					ti->ti_tlbinvop =
696 					    TLBINV_USER_MAP(ti->ti_tlbinvop);
697 					if (ti->ti_tlbinvop == TLBINV_ONE)
698 						ti->ti_victim = pm;
699 					else
700 						ti->ti_victim = NULL;
701 				}
702 			}
703 			TLBINFO_UNLOCK(ti);
704 			/*
705 			 * Now we can send out the shootdown IPIs to a CPU
706 			 * that shares this TLB and is currently using this
707 			 * pmap.  That CPU will process the IPI and do the
708 			 * all the work.  Any other CPUs sharing that TLB
709 			 * will take advantage of that work.  pm_onproc might
710 			 * change now that we have released the lock but we
711 			 * can tolerate spurious shootdowns.
712 			 */
713 			cpu_send_ipi(cpu_lookup(j), IPI_SHOOTDOWN);
714 			ipi_sent = true;
715 			continue;
716 		}
717 		if (!pmap_tlb_intersecting_active_p(pm, ti)) {
718 			/*
719 			 * If this pmap has an ASID assigned but it's not
720 			 * currently running, nuke its ASID.  Next time the
721 			 * pmap is activated, it will allocate a new ASID.
722 			 * And best of all, we avoid an IPI.
723 			 */
724 			KASSERT(!kernel_p);
725 			pmap_tlb_pai_reset(ti, pai, pm);
726 			//ti->ti_evcnt_lazy_shots.ev_count++;
727 		}
728 		TLBINFO_UNLOCK(ti);
729 	}
730 
731 	UVMHIST_LOG(maphist, " <-- done (ipi_sent=%jd)", ipi_sent, 0, 0, 0);
732 
733 	return ipi_sent;
734 }
735 #endif /* MULTIPROCESSOR && PMAP_TLB_NEED_SHOOTDOWN */
736 
737 int
738 pmap_tlb_update_addr(pmap_t pm, vaddr_t va, pt_entry_t pte, u_int flags)
739 {
740 	struct pmap_tlb_info * const ti = cpu_tlb_info(curcpu());
741 	struct pmap_asid_info * const pai = PMAP_PAI(pm, ti);
742 	int rv = -1;
743 
744 	UVMHIST_FUNC(__func__);
745 	UVMHIST_CALLARGS(maphist, " (pm=%#jx va=%#jx, pte=%#jx flags=%#jx)",
746 	    (uintptr_t)pm, va, pte_value(pte), flags);
747 
748 	KASSERT(kpreempt_disabled());
749 
750 	KASSERTMSG(pte_valid_p(pte), "va %#"PRIxVADDR" %#"PRIxPTE,
751 	    va, pte_value(pte));
752 
753 	TLBINFO_LOCK(ti);
754 	if (pm == pmap_kernel() || PMAP_PAI_ASIDVALID_P(pai, ti)) {
755 		pmap_tlb_asid_check();
756 		rv = tlb_update_addr(va, pai->pai_asid, pte,
757 		    (flags & PMAP_TLB_INSERT) != 0);
758 		pmap_tlb_asid_check();
759 		UVMHIST_LOG(maphist,
760 		    "   %jd <-- tlb_update_addr(%#jx, %#jx, %#jx, ...)",
761 		    rv, va, pai->pai_asid, pte_value(pte));
762 		KASSERTMSG((flags & PMAP_TLB_INSERT) == 0 || rv == 1,
763 		    "pmap %p (asid %u) va %#"PRIxVADDR" pte %#"PRIxPTE" rv %d",
764 		    pm, pai->pai_asid, va, pte_value(pte), rv);
765 	}
766 #if defined(MULTIPROCESSOR) && defined(PMAP_TLB_NEED_SHOOTDOWN)
767 	if (flags & PMAP_TLB_NEED_IPI)
768 		pm->pm_shootdown_pending = 1;
769 #endif
770 	TLBINFO_UNLOCK(ti);
771 
772 	UVMHIST_LOG(maphist, "   <-- done (rv=%jd)", rv, 0, 0, 0);
773 
774 	return rv;
775 }
776 
777 void
778 pmap_tlb_invalidate_addr(pmap_t pm, vaddr_t va)
779 {
780 	struct pmap_tlb_info * const ti = cpu_tlb_info(curcpu());
781 	struct pmap_asid_info * const pai = PMAP_PAI(pm, ti);
782 
783 	UVMHIST_FUNC(__func__);
784 	UVMHIST_CALLARGS(maphist, " (pm=%#jx va=%#jx) ti=%#jx asid=%#jx",
785 	    (uintptr_t)pm, va, (uintptr_t)ti, pai->pai_asid);
786 
787 	KASSERT(kpreempt_disabled());
788 
789 	TLBINFO_LOCK(ti);
790 	if (pm == pmap_kernel() || PMAP_PAI_ASIDVALID_P(pai, ti)) {
791 		pmap_tlb_asid_check();
792 		UVMHIST_LOG(maphist, " invalidating %#jx asid %#jx",
793 		    va, pai->pai_asid, 0, 0);
794 		tlb_invalidate_addr(va, pai->pai_asid);
795 		pmap_tlb_asid_check();
796 	}
797 #if defined(MULTIPROCESSOR) && defined(PMAP_TLB_NEED_SHOOTDOWN)
798 	pm->pm_shootdown_pending = 1;
799 #endif
800 	TLBINFO_UNLOCK(ti);
801 	UVMHIST_LOG(maphist, " <-- done", 0, 0, 0, 0);
802 }
803 
804 static inline void
805 pmap_tlb_asid_alloc(struct pmap_tlb_info *ti, pmap_t pm,
806 	struct pmap_asid_info *pai)
807 {
808 	/*
809 	 * We shouldn't have an ASID assigned, and thusly must not be onproc
810 	 * nor active.
811 	 */
812 	KASSERT(pm != pmap_kernel());
813 	KASSERT(pai->pai_asid == 0);
814 	KASSERT(pai->pai_link.le_prev == NULL);
815 #if defined(MULTIPROCESSOR)
816 	KASSERT(!pmap_tlb_intersecting_onproc_p(pm, ti));
817 	KASSERT(!pmap_tlb_intersecting_active_p(pm, ti));
818 #endif
819 	KASSERT(ti->ti_asids_free > 0);
820 	KASSERT(ti->ti_asid_hint > KERNEL_PID);
821 
822 	/*
823 	 * If the last ASID allocated was the maximum ASID, then the
824 	 * hint will be out of range.  Reset the hint to first
825 	 * available ASID.
826 	 */
827 	if (PMAP_TLB_FLUSH_ASID_ON_RESET
828 	    && ti->ti_asid_hint > ti->ti_asid_max) {
829 		ti->ti_asid_hint = KERNEL_PID + 1;
830 	}
831 	KASSERTMSG(ti->ti_asid_hint <= ti->ti_asid_max, "hint %u",
832 	    ti->ti_asid_hint);
833 
834 	/*
835 	 * Let's see if the hinted ASID is free.  If not search for
836 	 * a new one.
837 	 */
838 	if (__predict_true(TLBINFO_ASID_INUSE_P(ti, ti->ti_asid_hint))) {
839 		const size_t nbpw = NBBY * sizeof(ti->ti_asid_bitmap._b[0]);
840 		size_t i;
841 		u_long bits;
842 		for (i = 0; (bits = ~ti->ti_asid_bitmap._b[i]) == 0; i++) {
843 			KASSERT(i < __arraycount(ti->ti_asid_bitmap._b) - 1);
844 		}
845 		/*
846 		 * ffs wants to find the first bit set while we want
847 		 * to find the first bit cleared.
848 		 */
849 		const u_int n = __builtin_ffsl(bits) - 1;
850 		KASSERTMSG((bits << (nbpw - (n+1))) == (1ul << (nbpw-1)),
851 		    "n %u bits %#lx", n, bits);
852 		KASSERT(n < nbpw);
853 		ti->ti_asid_hint = n + i * nbpw;
854 	}
855 
856 	KASSERT(ti->ti_asid_hint > KERNEL_PID);
857 	KASSERT(ti->ti_asid_hint <= ti->ti_asid_max);
858 	KASSERTMSG(PMAP_TLB_FLUSH_ASID_ON_RESET
859 	    || TLBINFO_ASID_INUSE_P(ti, ti->ti_asid_hint - 1),
860 	    "hint %u bitmap %p", ti->ti_asid_hint, &ti->ti_asid_bitmap);
861 	KASSERTMSG(!TLBINFO_ASID_INUSE_P(ti, ti->ti_asid_hint),
862 	    "hint %u bitmap %p", ti->ti_asid_hint, &ti->ti_asid_bitmap);
863 
864 	/*
865 	 * The hint contains our next ASID so take it and advance the hint.
866 	 * Mark it as used and insert the pai into the list of active asids.
867 	 * There is also one less asid free in this TLB.
868 	 */
869 	KASSERT(ti->ti_asid_hint > KERNEL_PID);
870 	pai->pai_asid = ti->ti_asid_hint++;
871 #ifdef MULTIPROCESSOR
872 	if (PMAP_TLB_FLUSH_ASID_ON_RESET) {
873 		/*
874 		 * Clean the new ASID from the TLB.
875 		 */
876 		tlb_invalidate_asids(pai->pai_asid, pai->pai_asid);
877 	}
878 #endif
879 	TLBINFO_ASID_MARK_USED(ti, pai->pai_asid);
880 	LIST_INSERT_HEAD(&ti->ti_pais, pai, pai_link);
881 	ti->ti_asids_free--;
882 
883 #if defined(MULTIPROCESSOR)
884 	/*
885 	 * Mark that we now have an active ASID for all CPUs sharing this TLB.
886 	 * The bits in pm_active belonging to this TLB can only be changed
887 	 * while this TLBs lock is held.
888 	 */
889 #if PMAP_TLB_MAX == 1
890 	kcpuset_copy(pm->pm_active, kcpuset_running);
891 #else
892 	kcpuset_merge(pm->pm_active, ti->ti_kcpuset);
893 #endif
894 #endif
895 }
896 
897 /*
898  * Acquire a TLB address space tag (called ASID or TLBPID) and return it.
899  * ASID might have already been previously acquired.
900  */
901 void
902 pmap_tlb_asid_acquire(pmap_t pm, struct lwp *l)
903 {
904 	struct cpu_info * const ci = l->l_cpu;
905 	struct pmap_tlb_info * const ti = cpu_tlb_info(ci);
906 	struct pmap_asid_info * const pai = PMAP_PAI(pm, ti);
907 
908 	UVMHIST_FUNC(__func__);
909 	UVMHIST_CALLARGS(maphist, "(pm=%#jx, l=%#jx, ti=%#jx)", (uintptr_t)pm,
910 	    (uintptr_t)l, (uintptr_t)ti, 0);
911 
912 	KASSERT(kpreempt_disabled());
913 
914 	/*
915 	 * Kernels use a fixed ASID and thus doesn't need to acquire one.
916 	 */
917 	if (pm == pmap_kernel()) {
918 		UVMHIST_LOG(maphist, " <-- done (kernel)", 0, 0, 0, 0);
919 		return;
920 	}
921 
922 	TLBINFO_LOCK(ti);
923 	KASSERT(pai->pai_asid <= KERNEL_PID || pai->pai_link.le_prev != NULL);
924 	KASSERT(pai->pai_asid > KERNEL_PID || pai->pai_link.le_prev == NULL);
925 	pmap_tlb_pai_check(ti, true);
926 	if (__predict_false(!PMAP_PAI_ASIDVALID_P(pai, ti))) {
927 		/*
928 		 * If we've run out ASIDs, reinitialize the ASID space.
929 		 */
930 		if (__predict_false(tlbinfo_noasids_p(ti))) {
931 			KASSERT(l == curlwp);
932 			UVMHIST_LOG(maphist, " asid reinit", 0, 0, 0, 0);
933 			pmap_tlb_asid_reinitialize(ti, TLBINV_NOBODY);
934 			KASSERT(!tlbinfo_noasids_p(ti));
935 		}
936 
937 		/*
938 		 * Get an ASID.
939 		 */
940 		pmap_tlb_asid_alloc(ti, pm, pai);
941 		UVMHIST_LOG(maphist, "allocated asid %#jx", pai->pai_asid,
942 		    0, 0, 0);
943 	}
944 	pmap_tlb_pai_check(ti, true);
945 #if defined(MULTIPROCESSOR)
946 	KASSERT(kcpuset_isset(pm->pm_active, cpu_index(ci)));
947 #endif
948 
949 	if (l == curlwp) {
950 #if defined(MULTIPROCESSOR)
951 		/*
952 		 * The bits in pm_onproc belonging to this TLB can only
953 		 * be changed while this TLBs lock is held unless atomic
954 		 * operations are used.
955 		 */
956 		KASSERT(pm != pmap_kernel());
957 		kcpuset_atomic_set(pm->pm_onproc, cpu_index(ci));
958 #endif
959 		ci->ci_pmap_asid_cur = pai->pai_asid;
960 		UVMHIST_LOG(maphist, "setting asid to %#jx", pai->pai_asid,
961 		    0, 0, 0);
962 		tlb_set_asid(pai->pai_asid, pm);
963 		pmap_tlb_asid_check();
964 	} else {
965 		printf("%s: l (%p) != curlwp %p\n", __func__, l, curlwp);
966 	}
967 	TLBINFO_UNLOCK(ti);
968 	UVMHIST_LOG(maphist, " <-- done", 0, 0, 0, 0);
969 }
970 
971 void
972 pmap_tlb_asid_deactivate(pmap_t pm)
973 {
974 	UVMHIST_FUNC(__func__);
975 	UVMHIST_CALLARGS(maphist, "pm %#jx", (uintptr_t)pm, 0, 0, 0);
976 
977 	KASSERT(kpreempt_disabled());
978 #if defined(MULTIPROCESSOR)
979 	/*
980 	 * The kernel pmap is aways onproc and active and must never have
981 	 * those bits cleared.  If pmap_remove_all was called, it has already
982 	 * deactivated the pmap and thusly onproc will be 0 so there's nothing
983 	 * to do.
984 	 */
985 	if (pm != pmap_kernel() && !kcpuset_iszero(pm->pm_onproc)) {
986 		struct cpu_info * const ci = curcpu();
987 		KASSERT(!cpu_intr_p());
988 		KASSERTMSG(kcpuset_isset(pm->pm_onproc, cpu_index(ci)),
989 		    "%s: pmap %p onproc %p doesn't include cpu %d (%p)",
990 		    __func__, pm, pm->pm_onproc, cpu_index(ci), ci);
991 		/*
992 		 * The bits in pm_onproc that belong to this TLB can
993 		 * be changed while this TLBs lock is not held as long
994 		 * as we use atomic ops.
995 		 */
996 		kcpuset_atomic_clear(pm->pm_onproc, cpu_index(ci));
997 	}
998 #endif
999 	curcpu()->ci_pmap_asid_cur = KERNEL_PID;
1000 	tlb_set_asid(KERNEL_PID, pmap_kernel());
1001 
1002 	pmap_tlb_pai_check(cpu_tlb_info(curcpu()), false);
1003 #if defined(DEBUG)
1004 	pmap_tlb_asid_check();
1005 #endif
1006 	UVMHIST_LOG(maphist, " <-- done (pm=%#jx)", (uintptr_t)pm, 0, 0, 0);
1007 }
1008 
1009 void
1010 pmap_tlb_asid_release_all(struct pmap *pm)
1011 {
1012 	UVMHIST_FUNC(__func__);
1013 	UVMHIST_CALLARGS(maphist, "(pm=%#jx)", (uintptr_t)pm, 0, 0, 0);
1014 
1015 	KASSERT(pm != pmap_kernel());
1016 #if defined(MULTIPROCESSOR)
1017 	//KASSERT(!kcpuset_iszero(pm->pm_onproc)); // XXX
1018 	struct cpu_info * const ci __diagused = curcpu();
1019 	KASSERT(!kcpuset_isotherset(pm->pm_onproc, cpu_index(ci)));
1020 #if PMAP_TLB_MAX > 1
1021 	for (u_int i = 0; !kcpuset_iszero(pm->pm_active); i++) {
1022 		KASSERT(i < pmap_ntlbs);
1023 		struct pmap_tlb_info * const ti = pmap_tlbs[i];
1024 #else
1025 		struct pmap_tlb_info * const ti = &pmap_tlb0_info;
1026 #endif
1027 		struct pmap_asid_info * const pai = PMAP_PAI(pm, ti);
1028 		TLBINFO_LOCK(ti);
1029 		if (PMAP_PAI_ASIDVALID_P(pai, ti)) {
1030 			/*
1031 			 * This pmap should not be in use by any other cpu so
1032 			 * we can just reset and be happy.
1033 			 */
1034 			if (ti->ti_victim == pm)
1035 				ti->ti_victim = NULL;
1036 			pmap_tlb_pai_reset(ti, pai, pm);
1037 		}
1038 		KASSERT(pai->pai_link.le_prev == NULL);
1039 		TLBINFO_UNLOCK(ti);
1040 #if PMAP_TLB_MAX > 1
1041 	}
1042 #endif
1043 #ifdef DIAGNOSTIC
1044 	for (size_t i = 0; i < (PMAP_TLB_MAX > 1 ? pmap_ntlbs : 1); i++) {
1045 		KASSERTMSG(pm->pm_pai[i].pai_asid == 0,
1046 		    "pm %p i %zu asid %u",
1047 		    pm, i, pm->pm_pai[i].pai_asid);
1048 	}
1049 #endif
1050 #else
1051 	/*
1052 	 * Handle the case of an UP kernel which only has, at most, one TLB.
1053 	 * If the pmap has an ASID allocated, free it.
1054 	 */
1055 	struct pmap_tlb_info * const ti = &pmap_tlb0_info;
1056 	struct pmap_asid_info * const pai = PMAP_PAI(pm, ti);
1057 	TLBINFO_LOCK(ti);
1058 	if (pai->pai_asid > KERNEL_PID) {
1059 		if (curcpu()->ci_pmap_asid_cur == pai->pai_asid) {
1060 			tlb_invalidate_asids(pai->pai_asid, pai->pai_asid);
1061 		} else {
1062 			pmap_tlb_pai_reset(ti, pai, pm);
1063 		}
1064 	}
1065 	TLBINFO_UNLOCK(ti);
1066 #endif /* MULTIPROCESSOR */
1067 	UVMHIST_LOG(maphist, " <-- done", 0, 0, 0, 0);
1068 }
1069 
1070 void
1071 pmap_tlb_asid_check(void)
1072 {
1073 #ifdef DEBUG
1074 	kpreempt_disable();
1075 	const tlb_asid_t asid __debugused = tlb_get_asid();
1076 	KDASSERTMSG(asid == curcpu()->ci_pmap_asid_cur,
1077 	   "%s: asid (%#x) != current asid (%#x)",
1078 	    __func__, asid, curcpu()->ci_pmap_asid_cur);
1079 	kpreempt_enable();
1080 #endif
1081 }
1082 
1083 #ifdef DEBUG
1084 void
1085 pmap_tlb_check(pmap_t pm, bool (*func)(void *, vaddr_t, tlb_asid_t, pt_entry_t))
1086 {
1087 	struct pmap_tlb_info * const ti = cpu_tlb_info(curcpu());
1088 	struct pmap_asid_info * const pai = PMAP_PAI(pm, ti);
1089 	TLBINFO_LOCK(ti);
1090 	if (pm == pmap_kernel() || pai->pai_asid > KERNEL_PID)
1091 		tlb_walk(pm, func);
1092 	TLBINFO_UNLOCK(ti);
1093 }
1094 #endif /* DEBUG */
1095