1 /*	$NetBSD: pmap_tlb.c,v 1.18 2016/07/23 20:06:25 matt 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 1.18 2016/07/23 20:06:25 matt 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 pmap's ASID is active (equal to the ASID in COP 0
73  * register EntryHi) on a CPU.  The bit number comes from the CPU's cpu_index().
74  * Even though these bitmaps contain the bits for all CPUs, the bits that
75  * correspond to the bits belonging to the CPUs sharing a TLB can only be
76  * manipulated while holding that TLB's lock.  Atomic ops must be used to
77  * update them since multiple CPUs may be changing different sets of bits at
78  * same time but these sets never overlap.
79  *
80  * When a change to the local TLB may require a change in the TLB's of other
81  * CPUs, we try to avoid sending an IPI if at all possible.  For instance, if
82  * we are updating a PTE and that PTE previously was invalid and therefore
83  * couldn't support an active mapping, there's no need for an IPI since there
84  * can't be a TLB entry to invalidate.  The other case is when we change a PTE
85  * to be modified we just update the local TLB.  If another TLB has a stale
86  * entry, a TLB MOD exception will be raised and that will cause the local TLB
87  * to be updated.
88  *
89  * We never need to update a non-local TLB if the pmap doesn't have a valid
90  * ASID for that TLB.  If it does have a valid ASID but isn't current "onproc"
91  * we simply reset its ASID for that TLB and then when it goes "onproc" it
92  * will allocate a new ASID and any existing TLB entries will be orphaned.
93  * Only in the case that pmap has an "onproc" ASID do we actually have to send
94  * an IPI.
95  *
96  * Once we determined we must send an IPI to shootdown a TLB, we need to send
97  * it to one of CPUs that share that TLB.  We choose the lowest numbered CPU
98  * that has one of the pmap's ASID "onproc".  In reality, any CPU sharing that
99  * TLB would do, but interrupting an active CPU seems best.
100  *
101  * A TLB might have multiple shootdowns active concurrently.  The shootdown
102  * logic compresses these into a few cases:
103  *	0) nobody needs to have its TLB entries invalidated
104  *	1) one ASID needs to have its TLB entries invalidated
105  *	2) more than one ASID needs to have its TLB entries invalidated
106  *	3) the kernel needs to have its TLB entries invalidated
107  *	4) the kernel and one or more ASID need their TLB entries invalidated.
108  *
109  * And for each case we do:
110  *	0) nothing,
111  *	1) if that ASID is still "onproc", we invalidate the TLB entries for
112  *	   that single ASID.  If not, just reset the pmap's ASID to invalidate
113  *	   and let it allocate a new ASID the next time it goes "onproc",
114  *	2) we reinitialize the ASID space (preserving any "onproc" ASIDs) and
115  *	   invalidate all non-wired non-global TLB entries,
116  *	3) we invalidate all of the non-wired global TLB entries,
117  *	4) we reinitialize the ASID space (again preserving any "onproc" ASIDs)
118  *	   invalidate all non-wired TLB entries.
119  *
120  * As you can see, shootdowns are not concerned with addresses, just address
121  * spaces.  Since the number of TLB entries is usually quite small, this avoids
122  * a lot of overhead for not much gain.
123  */
124 
125 #define __PMAP_PRIVATE
126 
127 #include "opt_multiprocessor.h"
128 
129 #include <sys/param.h>
130 #include <sys/systm.h>
131 #include <sys/proc.h>
132 #include <sys/mutex.h>
133 #include <sys/atomic.h>
134 #include <sys/kernel.h>			/* for cold */
135 #include <sys/cpu.h>
136 
137 #include <uvm/uvm.h>
138 
139 static kmutex_t pmap_tlb0_lock __cacheline_aligned;
140 
141 #define	IFCONSTANT(x)	(__builtin_constant_p((x)) ? (x) : 0)
142 
143 struct pmap_tlb_info pmap_tlb0_info = {
144 	.ti_name = "tlb0",
145 	.ti_asid_hint = KERNEL_PID + 1,
146 #ifdef PMAP_TLB_NUM_PIDS
147 	.ti_asid_max = IFCONSTANT(PMAP_TLB_NUM_PIDS - 1),
148 	.ti_asids_free = IFCONSTANT(PMAP_TLB_NUM_PIDS - (1 + KERNEL_PID)),
149 #endif
150 	.ti_asid_bitmap[0] = (2 << KERNEL_PID) - 1,
151 #ifdef PMAP_TLB_WIRED_UPAGES
152 	.ti_wired = PMAP_TLB_WIRED_UPAGES,
153 #endif
154 	.ti_lock = &pmap_tlb0_lock,
155 	.ti_pais = LIST_HEAD_INITIALIZER(pmap_tlb0_info.ti_pais),
156 #if defined(MULTIPROCESSOR) && PMAP_TLB_MAX > 1
157 	.ti_tlbinvop = TLBINV_NOBODY,
158 #endif
159 };
160 
161 #undef IFCONSTANT
162 
163 #if defined(MULTIPROCESSOR) && PMAP_TLB_MAX > 1
164 struct pmap_tlb_info *pmap_tlbs[PMAP_TLB_MAX] = {
165 	[0] = &pmap_tlb0_info,
166 };
167 u_int pmap_ntlbs = 1;
168 #endif
169 
170 #define	__BITMAP_SET(bm, n) \
171 	((bm)[(n) / (8*sizeof(bm[0]))] |= 1LU << ((n) % (8*sizeof(bm[0]))))
172 #define	__BITMAP_CLR(bm, n) \
173 	((bm)[(n) / (8*sizeof(bm[0]))] &= ~(1LU << ((n) % (8*sizeof(bm[0])))))
174 #define	__BITMAP_ISSET_P(bm, n) \
175 	(((bm)[(n) / (8*sizeof(bm[0]))] & (1LU << ((n) % (8*sizeof(bm[0]))))) != 0)
176 
177 #define	TLBINFO_ASID_MARK_UNUSED(ti, asid) \
178 	__BITMAP_CLR((ti)->ti_asid_bitmap, (asid))
179 #define	TLBINFO_ASID_MARK_USED(ti, asid) \
180 	__BITMAP_SET((ti)->ti_asid_bitmap, (asid))
181 #define	TLBINFO_ASID_INUSE_P(ti, asid) \
182 	__BITMAP_ISSET_P((ti)->ti_asid_bitmap, (asid))
183 
184 #ifdef MULTIPROCESSOR
185 __unused static inline bool
pmap_tlb_intersecting_active_p(pmap_t pm,struct pmap_tlb_info * ti)186 pmap_tlb_intersecting_active_p(pmap_t pm, struct pmap_tlb_info *ti)
187 {
188 #if PMAP_TLB_MAX == 1
189 	return !kcpuset_iszero(pm->pm_active);
190 #else
191 	return kcpuset_intersecting_p(pm->pm_active, ti->ti_kcpuset);
192 #endif
193 }
194 
195 static inline bool
pmap_tlb_intersecting_onproc_p(pmap_t pm,struct pmap_tlb_info * ti)196 pmap_tlb_intersecting_onproc_p(pmap_t pm, struct pmap_tlb_info *ti)
197 {
198 #if PMAP_TLB_MAX == 1
199 	return !kcpuset_iszero(pm->pm_onproc);
200 #else
201 	return kcpuset_intersecting_p(pm->pm_onproc, ti->ti_kcpuset);
202 #endif
203 }
204 #endif
205 
206 static void
pmap_tlb_pai_check(struct pmap_tlb_info * ti,bool locked_p)207 pmap_tlb_pai_check(struct pmap_tlb_info *ti, bool locked_p)
208 {
209 #ifdef DIAGNOSTIC
210 	struct pmap_asid_info *pai;
211 	if (!locked_p)
212 		TLBINFO_LOCK(ti);
213 	LIST_FOREACH(pai, &ti->ti_pais, pai_link) {
214 		KASSERT(pai != NULL);
215 		KASSERT(PAI_PMAP(pai, ti) != pmap_kernel());
216 		KASSERT(pai->pai_asid > KERNEL_PID);
217 		KASSERTMSG(pai->pai_asid <= ti->ti_asid_max,
218 		    "pm %p asid %#x", PAI_PMAP(pai, ti), pai->pai_asid);
219 		KASSERTMSG(TLBINFO_ASID_INUSE_P(ti, pai->pai_asid),
220 		    "pm %p asid %u", PAI_PMAP(pai, ti), pai->pai_asid);
221 #ifdef MULTIPROCESSOR
222 		KASSERT(pmap_tlb_intersecting_active_p(PAI_PMAP(pai, ti), ti));
223 #endif
224 	}
225 	if (!locked_p)
226 		TLBINFO_UNLOCK(ti);
227 #endif
228 }
229 
230 static void
pmap_tlb_pai_reset(struct pmap_tlb_info * ti,struct pmap_asid_info * pai,struct pmap * pm)231 pmap_tlb_pai_reset(struct pmap_tlb_info *ti, struct pmap_asid_info *pai,
232 	struct pmap *pm)
233 {
234 	UVMHIST_FUNC(__func__); UVMHIST_CALLED(maphist);
235 	UVMHIST_LOG(maphist, "(ti=%p, pai=%p, pm=%p): asid %u",
236 	    ti, pai, pm, pai->pai_asid);
237 
238 	/*
239 	 * We must have an ASID but it must not be onproc (on a processor).
240 	 */
241 	KASSERT(pai->pai_asid > KERNEL_PID);
242 	KASSERT(pai->pai_asid <= ti->ti_asid_max);
243 #if defined(MULTIPROCESSOR)
244 	KASSERT(pmap_tlb_intersecting_active_p(pm, ti));
245 	KASSERT(!pmap_tlb_intersecting_onproc_p(pm, ti));
246 #endif
247 	LIST_REMOVE(pai, pai_link);
248 #ifdef DIAGNOSTIC
249 	pai->pai_link.le_prev = NULL;	/* tagged as unlinked */
250 #endif
251 	/*
252 	 * If the platform has a cheap way to flush ASIDs then free the ASID
253 	 * back into the pool.  On multiprocessor systems, we will flush the
254 	 * ASID from the TLB when it's allocated.  That way we know the flush
255 	 * was always done in the correct TLB space.  On uniprocessor systems,
256 	 * just do the flush now since we know that it has been used.  This has
257 	 * a bit less overhead.  Either way, this will mean that we will only
258 	 * need to flush all ASIDs if all ASIDs are in use and we need to
259 	 * allocate a new one.
260 	 */
261 	if (PMAP_TLB_FLUSH_ASID_ON_RESET) {
262 #ifndef MULTIPROCESSOR
263 		tlb_invalidate_asids(pai->pai_asid, pai->pai_asid);
264 #endif
265 		if (TLBINFO_ASID_INUSE_P(ti, pai->pai_asid)) {
266 			TLBINFO_ASID_MARK_UNUSED(ti, pai->pai_asid);
267 			ti->ti_asids_free++;
268 		}
269 	}
270 	/*
271 	 * Note that we don't mark the ASID as not in use in the TLB's ASID
272 	 * bitmap (thus it can't be allocated until the ASID space is exhausted
273 	 * and therefore reinitialized).  We don't want to flush the TLB for
274 	 * entries belonging to this ASID so we will let natural TLB entry
275 	 * replacement flush them out of the TLB.  Any new entries for this
276 	 * pmap will need a new ASID allocated.
277 	 */
278 	pai->pai_asid = 0;
279 
280 #if defined(MULTIPROCESSOR)
281 	/*
282 	 * The bits in pm_active belonging to this TLB can only be changed
283 	 * while this TLB's lock is held.
284 	 */
285 #if PMAP_TLB_MAX == 1
286 	kcpuset_zero(pm->pm_active);
287 #else
288 	kcpuset_remove(pm->pm_active, ti->ti_kcpuset);
289 #endif
290 	KASSERT(!pmap_tlb_intersecting_active_p(pm, ti));
291 #endif /* MULTIPROCESSOR */
292 
293 	UVMHIST_LOG(maphist, " <-- done", 0, 0, 0, 0);
294 }
295 
296 void
pmap_tlb_info_evcnt_attach(struct pmap_tlb_info * ti)297 pmap_tlb_info_evcnt_attach(struct pmap_tlb_info *ti)
298 {
299 #if defined(MULTIPROCESSOR)
300 	evcnt_attach_dynamic_nozero(&ti->ti_evcnt_synci_desired,
301 	    EVCNT_TYPE_MISC, NULL,
302 	    ti->ti_name, "icache syncs desired");
303 	evcnt_attach_dynamic_nozero(&ti->ti_evcnt_synci_asts,
304 	    EVCNT_TYPE_MISC, &ti->ti_evcnt_synci_desired,
305 	    ti->ti_name, "icache sync asts");
306 	evcnt_attach_dynamic_nozero(&ti->ti_evcnt_synci_all,
307 	    EVCNT_TYPE_MISC, &ti->ti_evcnt_synci_asts,
308 	    ti->ti_name, "icache full syncs");
309 	evcnt_attach_dynamic_nozero(&ti->ti_evcnt_synci_pages,
310 	    EVCNT_TYPE_MISC, &ti->ti_evcnt_synci_asts,
311 	    ti->ti_name, "icache pages synced");
312 	evcnt_attach_dynamic_nozero(&ti->ti_evcnt_synci_duplicate,
313 	    EVCNT_TYPE_MISC, &ti->ti_evcnt_synci_desired,
314 	    ti->ti_name, "icache dup pages skipped");
315 	evcnt_attach_dynamic_nozero(&ti->ti_evcnt_synci_deferred,
316 	    EVCNT_TYPE_MISC, &ti->ti_evcnt_synci_desired,
317 	    ti->ti_name, "icache pages deferred");
318 #endif /* MULTIPROCESSOR */
319 	evcnt_attach_dynamic_nozero(&ti->ti_evcnt_asid_reinits,
320 	    EVCNT_TYPE_MISC, NULL,
321 	    ti->ti_name, "asid pool reinit");
322 }
323 
324 void
pmap_tlb_info_init(struct pmap_tlb_info * ti)325 pmap_tlb_info_init(struct pmap_tlb_info *ti)
326 {
327 #if defined(MULTIPROCESSOR)
328 #if PMAP_TLB_MAX == 1
329 	KASSERT(ti == &pmap_tlb0_info);
330 #else
331 	if (ti != &pmap_tlb0_info) {
332 		KASSERT(pmap_ntlbs < PMAP_TLB_MAX);
333 
334 		KASSERT(pmap_tlbs[pmap_ntlbs] == NULL);
335 
336 		ti->ti_lock = mutex_obj_alloc(MUTEX_DEFAULT, IPL_SCHED);
337 		ti->ti_asid_bitmap[0] = (2 << KERNEL_PID) - 1;
338 		ti->ti_asid_hint = KERNEL_PID + 1;
339 		ti->ti_asid_max = pmap_tlbs[0]->ti_asid_max;
340 		ti->ti_asids_free = ti->ti_asid_max - KERNEL_PID;
341 		ti->ti_tlbinvop = TLBINV_NOBODY,
342 		ti->ti_victim = NULL;
343 		kcpuset_create(&ti->ti_kcpuset, true);
344 		ti->ti_index = pmap_ntlbs++;
345 		ti->ti_wired = 0;
346 		pmap_tlbs[ti->ti_index] = ti;
347 		snprintf(ti->ti_name, sizeof(ti->ti_name), "tlb%u",
348 		    ti->ti_index);
349 		pmap_tlb_info_evcnt_attach(ti);
350 		return;
351 	}
352 #endif
353 #endif /* MULTIPROCESSOR */
354 	KASSERT(ti == &pmap_tlb0_info);
355 	KASSERT(ti->ti_lock == &pmap_tlb0_lock);
356 	//printf("ti_lock %p ", ti->ti_lock);
357 	mutex_init(ti->ti_lock, MUTEX_DEFAULT, IPL_SCHED);
358 #if defined(MULTIPROCESSOR) && PMAP_TLB_MAX > 1
359 	kcpuset_create(&ti->ti_kcpuset, true);
360 	kcpuset_set(ti->ti_kcpuset, cpu_index(curcpu()));
361 #endif
362 	//printf("asid ");
363 	if (ti->ti_asid_max == 0) {
364 		ti->ti_asid_max = pmap_md_tlb_asid_max();
365 		ti->ti_asids_free = ti->ti_asid_max - KERNEL_PID;
366 	}
367 
368 	KASSERT(ti->ti_asid_max < sizeof(ti->ti_asid_bitmap)*8);
369 }
370 
371 #if defined(MULTIPROCESSOR)
372 void
pmap_tlb_info_attach(struct pmap_tlb_info * ti,struct cpu_info * ci)373 pmap_tlb_info_attach(struct pmap_tlb_info *ti, struct cpu_info *ci)
374 {
375 	KASSERT(!CPU_IS_PRIMARY(ci));
376 	KASSERT(ci->ci_data.cpu_idlelwp != NULL);
377 	KASSERT(cold);
378 
379 	TLBINFO_LOCK(ti);
380 #if PMAP_TLB_MAX > 1
381 	kcpuset_set(ti->ti_kcpuset, cpu_index(ci));
382 	cpu_set_tlb_info(ci, ti);
383 #endif
384 
385 	/*
386 	 * Do any MD tlb info init.
387 	 */
388 	pmap_md_tlb_info_attach(ti, ci);
389 
390 	/*
391 	 * The kernel pmap uses the kcpuset_running set so it's always
392 	 * up-to-date.
393 	 */
394 	TLBINFO_UNLOCK(ti);
395 }
396 #endif /* MULTIPROCESSOR */
397 
398 #ifdef DIAGNOSTIC
399 static size_t
pmap_tlb_asid_count(struct pmap_tlb_info * ti)400 pmap_tlb_asid_count(struct pmap_tlb_info *ti)
401 {
402 	size_t count = 0;
403 	for (tlb_asid_t asid = 1; asid <= ti->ti_asid_max; asid++) {
404 		count += TLBINFO_ASID_INUSE_P(ti, asid);
405 	}
406 	return count;
407 }
408 #endif
409 
410 static void
pmap_tlb_asid_reinitialize(struct pmap_tlb_info * ti,enum tlb_invalidate_op op)411 pmap_tlb_asid_reinitialize(struct pmap_tlb_info *ti, enum tlb_invalidate_op op)
412 {
413 	const size_t asid_bitmap_words =
414 	    ti->ti_asid_max / (8 * sizeof(ti->ti_asid_bitmap[0]));
415 
416 	UVMHIST_FUNC(__func__); UVMHIST_CALLED(maphist);
417 	UVMHIST_LOG(maphist, "(ti=%p, op=%u)", ti, op, 0, 0);
418 
419 	pmap_tlb_pai_check(ti, true);
420 
421 	ti->ti_evcnt_asid_reinits.ev_count++;
422 
423 	/*
424 	 * First, clear the ASID bitmap (except for ASID 0 which belongs
425 	 * to the kernel).
426 	 */
427 	ti->ti_asids_free = ti->ti_asid_max - KERNEL_PID;
428 	ti->ti_asid_hint = KERNEL_PID + 1;
429 	ti->ti_asid_bitmap[0] = (2 << KERNEL_PID) - 1;
430 	for (size_t word = 1; word <= asid_bitmap_words; word++) {
431 		ti->ti_asid_bitmap[word] = 0;
432 	}
433 
434 	switch (op) {
435 #if defined(MULTIPROCESSOR) && defined(PMAP_TLB_NEED_SHOOTDOWN)
436 	case TLBINV_ALL:
437 		tlb_invalidate_all();
438 		break;
439 	case TLBINV_ALLUSER:
440 		tlb_invalidate_asids(KERNEL_PID + 1, ti->ti_asid_max);
441 		break;
442 #endif /* MULTIPROCESSOR && PMAP_TLB_NEED_SHOOTDOWN */
443 	case TLBINV_NOBODY: {
444 		/*
445 		 * If we are just reclaiming ASIDs in the TLB, let's go find
446 		 * what ASIDs are in use in the TLB.  Since this is a
447 		 * semi-expensive operation, we don't want to do it too often.
448 		 * So if more half of the ASIDs are in use, we don't have
449 		 * enough free ASIDs so invalidate the TLB entries with ASIDs
450 		 * and clear the ASID bitmap.  That will force everyone to
451 		 * allocate a new ASID.
452 		 */
453 #if !defined(MULTIPROCESSOR) || defined(PMAP_TLB_NEED_SHOOTDOWN)
454 		pmap_tlb_asid_check();
455 		const u_int asids_found = tlb_record_asids(ti->ti_asid_bitmap,
456 		    ti->ti_asid_max);
457 		pmap_tlb_asid_check();
458 		KASSERT(asids_found == pmap_tlb_asid_count(ti));
459 		if (__predict_false(asids_found >= ti->ti_asid_max / 2)) {
460 			tlb_invalidate_asids(KERNEL_PID + 1, ti->ti_asid_max);
461 #else /* MULTIPROCESSOR && !PMAP_TLB_NEED_SHOOTDOWN */
462 			/*
463 			 * For those systems (PowerPC) that don't require
464 			 * cross cpu TLB shootdowns, we have to invalidate the
465 			 * entire TLB because we can't record the ASIDs in use
466 			 * on the other CPUs.  This is hopefully cheaper than
467 			 * than trying to use an IPI to record all the ASIDs
468 			 * on all the CPUs (which would be a synchronization
469 			 * nightmare).
470 			 */
471 			tlb_invalidate_all();
472 #endif /* MULTIPROCESSOR && !PMAP_TLB_NEED_SHOOTDOWN */
473 			ti->ti_asid_bitmap[0] = (2 << KERNEL_PID) - 1;
474 			for (size_t word = 1;
475 			     word <= asid_bitmap_words;
476 			     word++) {
477 				ti->ti_asid_bitmap[word] = 0;
478 			}
479 			ti->ti_asids_free = ti->ti_asid_max - KERNEL_PID;
480 #if !defined(MULTIPROCESSOR) || defined(PMAP_TLB_NEED_SHOOTDOWN)
481 		} else {
482 			ti->ti_asids_free -= asids_found;
483 		}
484 #endif /* !MULTIPROCESSOR || PMAP_TLB_NEED_SHOOTDOWN */
485 		KASSERTMSG(ti->ti_asids_free <= ti->ti_asid_max, "%u",
486 		    ti->ti_asids_free);
487 		break;
488 	}
489 	default:
490 		panic("%s: unexpected op %d", __func__, op);
491 	}
492 
493 	/*
494 	 * Now go through the active ASIDs.  If the ASID is on a processor or
495 	 * we aren't invalidating all ASIDs and the TLB has an entry owned by
496 	 * that ASID, mark it as in use.  Otherwise release the ASID.
497 	 */
498 	struct pmap_asid_info *pai, *next;
499 	for (pai = LIST_FIRST(&ti->ti_pais); pai != NULL; pai = next) {
500 		struct pmap * const pm = PAI_PMAP(pai, ti);
501 		next = LIST_NEXT(pai, pai_link);
502 		KASSERT(pm != pmap_kernel());
503 		KASSERT(pai->pai_asid > KERNEL_PID);
504 #if defined(MULTIPROCESSOR)
505 		if (pmap_tlb_intersecting_onproc_p(pm, ti)) {
506 			if (!TLBINFO_ASID_INUSE_P(ti, pai->pai_asid)) {
507 				TLBINFO_ASID_MARK_USED(ti, pai->pai_asid);
508 				ti->ti_asids_free--;
509 			}
510 			continue;
511 		}
512 #endif /* MULTIPROCESSOR */
513 		if (TLBINFO_ASID_INUSE_P(ti, pai->pai_asid)) {
514 			KASSERT(op == TLBINV_NOBODY);
515 		} else {
516 			pmap_tlb_pai_reset(ti, pai, pm);
517 		}
518 	}
519 #ifdef DIAGNOSTIC
520 	size_t free_count __diagused = ti->ti_asid_max - pmap_tlb_asid_count(ti);
521 	KASSERTMSG(free_count == ti->ti_asids_free,
522 	    "bitmap error: %zu != %u", free_count, ti->ti_asids_free);
523 #endif
524 	UVMHIST_LOG(maphist, " <-- done", 0, 0, 0, 0);
525 }
526 
527 #if defined(MULTIPROCESSOR) && defined(PMAP_TLB_NEED_SHOOTDOWN)
528 #if PMAP_TLB_MAX == 1
529 #error shootdown not required for single TLB systems
530 #endif
531 void
pmap_tlb_shootdown_process(void)532 pmap_tlb_shootdown_process(void)
533 {
534 	struct cpu_info * const ci = curcpu();
535 	struct pmap_tlb_info * const ti = cpu_tlb_info(ci);
536 #ifdef DIAGNOSTIC
537 	struct pmap * const pm = curlwp->l_proc->p_vmspace->vm_map.pmap;
538 #endif
539 
540 	KASSERT(cpu_intr_p());
541 	KASSERTMSG(ci->ci_cpl >= IPL_SCHED,
542 	    "%s: cpl (%d) < IPL_SCHED (%d)",
543 	    __func__, ci->ci_cpl, IPL_SCHED);
544 
545 	TLBINFO_LOCK(ti);
546 
547 	switch (ti->ti_tlbinvop) {
548 	case TLBINV_ONE: {
549 		/*
550 		 * We only need to invalidate one user ASID.
551 		 */
552 		struct pmap_asid_info * const pai = PMAP_PAI(ti->ti_victim, ti);
553 		KASSERT(ti->ti_victim != pmap_kernel());
554 		if (!pmap_tlb_intersecting_onproc_p(ti->ti_victim, ti)) {
555 			/*
556 			 * The victim is an active pmap so we will just
557 			 * invalidate its TLB entries.
558 			 */
559 			KASSERT(pai->pai_asid > KERNEL_PID);
560 			pmap_tlb_asid_check();
561 			tlb_invalidate_asids(pai->pai_asid, pai->pai_asid);
562 			pmap_tlb_asid_check();
563 		} else if (pai->pai_asid) {
564 			/*
565 			 * The victim is no longer an active pmap for this TLB.
566 			 * So simply clear its ASID and when pmap_activate is
567 			 * next called for this pmap, it will allocate a new
568 			 * ASID.
569 			 */
570 			KASSERT(!pmap_tlb_intersecting_onproc_p(pm, ti));
571 			pmap_tlb_pai_reset(ti, pai, PAI_PMAP(pai, ti));
572 		}
573 		break;
574 	}
575 	case TLBINV_ALLUSER:
576 		/*
577 		 * Flush all user TLB entries.
578 		 */
579 		pmap_tlb_asid_reinitialize(ti, TLBINV_ALLUSER);
580 		break;
581 	case TLBINV_ALLKERNEL:
582 		/*
583 		 * We need to invalidate all global TLB entries.
584 		 */
585 		pmap_tlb_asid_check();
586 		tlb_invalidate_globals();
587 		pmap_tlb_asid_check();
588 		break;
589 	case TLBINV_ALL:
590 		/*
591 		 * Flush all the TLB entries (user and kernel).
592 		 */
593 		pmap_tlb_asid_reinitialize(ti, TLBINV_ALL);
594 		break;
595 	case TLBINV_NOBODY:
596 		/*
597 		 * Might be spurious or another SMT CPU sharing this TLB
598 		 * could have already done the work.
599 		 */
600 		break;
601 	}
602 
603 	/*
604 	 * Indicate we are done with shutdown event.
605 	 */
606 	ti->ti_victim = NULL;
607 	ti->ti_tlbinvop = TLBINV_NOBODY;
608 	TLBINFO_UNLOCK(ti);
609 }
610 
611 /*
612  * This state machine could be encoded into an array of integers but since all
613  * the values fit in 3 bits, the 5 entry "table" fits in a 16 bit value which
614  * can be loaded in a single instruction.
615  */
616 #define	TLBINV_MAP(op, nobody, one, alluser, allkernel, all)	\
617 	((((   (nobody) << 3*TLBINV_NOBODY)			\
618 	 | (      (one) << 3*TLBINV_ONE)			\
619 	 | (  (alluser) << 3*TLBINV_ALLUSER)			\
620 	 | ((allkernel) << 3*TLBINV_ALLKERNEL)			\
621 	 | (      (all) << 3*TLBINV_ALL)) >> 3*(op)) & 7)
622 
623 #define	TLBINV_USER_MAP(op)	\
624 	TLBINV_MAP(op, TLBINV_ONE, TLBINV_ALLUSER, TLBINV_ALLUSER,	\
625 	    TLBINV_ALL, TLBINV_ALL)
626 
627 #define	TLBINV_KERNEL_MAP(op)	\
628 	TLBINV_MAP(op, TLBINV_ALLKERNEL, TLBINV_ALL, TLBINV_ALL,	\
629 	    TLBINV_ALLKERNEL, TLBINV_ALL)
630 
631 bool
pmap_tlb_shootdown_bystanders(pmap_t pm)632 pmap_tlb_shootdown_bystanders(pmap_t pm)
633 {
634 	/*
635 	 * We don't need to deal with our own TLB.
636 	 */
637 
638 	UVMHIST_FUNC(__func__); UVMHIST_CALLED(maphist);
639 
640 	kcpuset_t *pm_active;
641 	kcpuset_clone(&pm_active, pm->pm_active);
642 	kcpuset_remove(pm_active, cpu_tlb_info(curcpu())->ti_kcpuset);
643 	const bool kernel_p = (pm == pmap_kernel());
644 	bool ipi_sent = false;
645 
646 	/*
647 	 * If pm_active gets more bits set, then it's after all our changes
648 	 * have been made so they will already be cognizant of them.
649 	 */
650 
651 	for (size_t i = 0; !kcpuset_iszero(pm_active); i++) {
652 		KASSERT(i < pmap_ntlbs);
653 		struct pmap_tlb_info * const ti = pmap_tlbs[i];
654 		KASSERT(tlbinfo_index(ti) == i);
655 		/*
656 		 * Skip this TLB if there are no active mappings for it.
657 		 */
658 		if (!kcpuset_intersecting_p(pm_active, ti->ti_kcpuset))
659 			continue;
660 		struct pmap_asid_info * const pai = PMAP_PAI(pm, ti);
661 		kcpuset_remove(pm_active, ti->ti_kcpuset);
662 		TLBINFO_LOCK(ti);
663 		cpuid_t j = kcpuset_ffs_intersecting(pm->pm_onproc,
664 		    ti->ti_kcpuset);
665 		// post decrement since ffs returns bit + 1 or 0 if no bit
666 		if (j-- > 0) {
667 			if (kernel_p) {
668 				ti->ti_tlbinvop =
669 				    TLBINV_KERNEL_MAP(ti->ti_tlbinvop);
670 				ti->ti_victim = NULL;
671 			} else {
672 				KASSERT(pai->pai_asid);
673 				if (__predict_false(ti->ti_victim == pm)) {
674 					KASSERT(ti->ti_tlbinvop == TLBINV_ONE);
675 					/*
676 					 * We still need to invalidate this one
677 					 * ASID so there's nothing to change.
678 					 */
679 				} else {
680 					ti->ti_tlbinvop =
681 					    TLBINV_USER_MAP(ti->ti_tlbinvop);
682 					if (ti->ti_tlbinvop == TLBINV_ONE)
683 						ti->ti_victim = pm;
684 					else
685 						ti->ti_victim = NULL;
686 				}
687 			}
688 			TLBINFO_UNLOCK(ti);
689 			/*
690 			 * Now we can send out the shootdown IPIs to a CPU
691 			 * that shares this TLB and is currently using this
692 			 * pmap.  That CPU will process the IPI and do the
693 			 * all the work.  Any other CPUs sharing that TLB
694 			 * will take advantage of that work.  pm_onproc might
695 			 * change now that we have released the lock but we
696 			 * can tolerate spurious shootdowns.
697 			 */
698 			cpu_send_ipi(cpu_lookup(j), IPI_SHOOTDOWN);
699 			ipi_sent = true;
700 			continue;
701 		}
702 		if (!pmap_tlb_intersecting_active_p(pm, ti)) {
703 			/*
704 			 * If this pmap has an ASID assigned but it's not
705 			 * currently running, nuke its ASID.  Next time the
706 			 * pmap is activated, it will allocate a new ASID.
707 			 * And best of all, we avoid an IPI.
708 			 */
709 			KASSERT(!kernel_p);
710 			pmap_tlb_pai_reset(ti, pai, pm);
711 			//ti->ti_evcnt_lazy_shots.ev_count++;
712 		}
713 		TLBINFO_UNLOCK(ti);
714 	}
715 
716 	kcpuset_destroy(pm_active);
717 
718 	UVMHIST_LOG(maphist, " <-- done (ipi_sent=%d)", ipi_sent, 0, 0, 0);
719 
720 	return ipi_sent;
721 }
722 #endif /* MULTIPROCESSOR && PMAP_TLB_NEED_SHOOTDOWN */
723 
724 #ifndef PMAP_TLB_HWPAGEWALKER
725 int
pmap_tlb_update_addr(pmap_t pm,vaddr_t va,pt_entry_t pte,u_int flags)726 pmap_tlb_update_addr(pmap_t pm, vaddr_t va, pt_entry_t pte, u_int flags)
727 {
728 	struct pmap_tlb_info * const ti = cpu_tlb_info(curcpu());
729 	struct pmap_asid_info * const pai = PMAP_PAI(pm, ti);
730 	int rv = -1;
731 
732 	UVMHIST_FUNC(__func__); UVMHIST_CALLED(maphist);
733 	UVMHIST_LOG(maphist,
734 	    " (pm=%p va=%#"PRIxVADDR", pte=%#"PRIxPTE" flags=%#x)",
735 	    pm, va, pte_value(pte), flags);
736 
737 	KASSERT(kpreempt_disabled());
738 
739 	KASSERTMSG(pte_valid_p(pte), "va %#"PRIxVADDR" %#"PRIxPTE,
740 	    va, pte_value(pte));
741 
742 	TLBINFO_LOCK(ti);
743 	if (pm == pmap_kernel() || PMAP_PAI_ASIDVALID_P(pai, ti)) {
744 		pmap_tlb_asid_check();
745 		rv = tlb_update_addr(va, pai->pai_asid, pte,
746 		    (flags & PMAP_TLB_INSERT) != 0);
747 		pmap_tlb_asid_check();
748 		UVMHIST_LOG(maphist,
749 		     "   %d <-- tlb_update_addr(%#"PRIxVADDR", %#x, %#"PRIxPTE", ...)",
750 		     rv, va, pai->pai_asid, pte_value(pte));
751 		KASSERTMSG((flags & PMAP_TLB_INSERT) == 0 || rv == 1,
752 		    "pmap %p (asid %u) va %#"PRIxVADDR" pte %#"PRIxPTE" rv %d",
753 		    pm, pai->pai_asid, va, pte_value(pte), rv);
754 	}
755 #if defined(MULTIPROCESSOR) && defined(PMAP_TLB_NEED_SHOOTDOWN)
756 	if (flags & PMAP_TLB_NEED_IPI)
757 		pm->pm_shootdown_pending = 1;
758 #endif
759 	TLBINFO_UNLOCK(ti);
760 
761 	UVMHIST_LOG(maphist, "   <-- done (rv=%d)", rv, 0, 0, 0);
762 
763 	return rv;
764 }
765 #endif /* !PMAP_TLB_HWPAGEWALKER */
766 
767 void
pmap_tlb_invalidate_addr(pmap_t pm,vaddr_t va)768 pmap_tlb_invalidate_addr(pmap_t pm, vaddr_t va)
769 {
770 	struct pmap_tlb_info * const ti = cpu_tlb_info(curcpu());
771 	struct pmap_asid_info * const pai = PMAP_PAI(pm, ti);
772 
773 	UVMHIST_FUNC(__func__); UVMHIST_CALLED(maphist);
774 	UVMHIST_LOG(maphist, " (pm=%p va=%#"PRIxVADDR") ti=%p asid=%#x",
775 	    pm, va, ti, pai->pai_asid);
776 
777 	KASSERT(kpreempt_disabled());
778 
779 	TLBINFO_LOCK(ti);
780 	if (pm == pmap_kernel() || PMAP_PAI_ASIDVALID_P(pai, ti)) {
781 		pmap_tlb_asid_check();
782 		UVMHIST_LOG(maphist, " invalidating %#"PRIxVADDR" asid %#x",
783 		    va, pai->pai_asid, 0, 0);
784 		tlb_invalidate_addr(va, pai->pai_asid);
785 		pmap_tlb_asid_check();
786 	}
787 #if defined(MULTIPROCESSOR) && defined(PMAP_TLB_NEED_SHOOTDOWN)
788 	pm->pm_shootdown_pending = 1;
789 #endif
790 	TLBINFO_UNLOCK(ti);
791 	UVMHIST_LOG(maphist, " <-- done", 0, 0, 0, 0);
792 }
793 
794 static inline void
pmap_tlb_asid_alloc(struct pmap_tlb_info * ti,pmap_t pm,struct pmap_asid_info * pai)795 pmap_tlb_asid_alloc(struct pmap_tlb_info *ti, pmap_t pm,
796 	struct pmap_asid_info *pai)
797 {
798 	/*
799 	 * We shouldn't have an ASID assigned, and thusly must not be onproc
800 	 * nor active.
801 	 */
802 	KASSERT(pm != pmap_kernel());
803 	KASSERT(pai->pai_asid == 0);
804 	KASSERT(pai->pai_link.le_prev == NULL);
805 #if defined(MULTIPROCESSOR)
806 	KASSERT(!pmap_tlb_intersecting_onproc_p(pm, ti));
807 	KASSERT(!pmap_tlb_intersecting_active_p(pm, ti));
808 #endif
809 	KASSERT(ti->ti_asids_free > 0);
810 	KASSERT(ti->ti_asid_hint > KERNEL_PID);
811 
812 	/*
813 	 * If the last ASID allocated was the maximum ASID, then the
814 	 * hint will be out of range.  Reset the hint to first
815 	 * available ASID.
816 	 */
817 	if (PMAP_TLB_FLUSH_ASID_ON_RESET
818 	    && ti->ti_asid_hint > ti->ti_asid_max) {
819 		ti->ti_asid_hint = KERNEL_PID + 1;
820 	}
821 	KASSERTMSG(ti->ti_asid_hint <= ti->ti_asid_max, "hint %u",
822 	    ti->ti_asid_hint);
823 
824 	/*
825 	 * Let's see if the hinted ASID is free.  If not search for
826 	 * a new one.
827 	 */
828 	if (__predict_true(TLBINFO_ASID_INUSE_P(ti, ti->ti_asid_hint))) {
829 		const size_t nbpw __diagused = 8*sizeof(ti->ti_asid_bitmap[0]);
830 		size_t i;
831 		u_long bits;
832 		for (i = 0; (bits = ~ti->ti_asid_bitmap[i]) == 0; i++) {
833 			KASSERT(i < __arraycount(ti->ti_asid_bitmap) - 1);
834 		}
835 		/*
836 		 * ffs wants to find the first bit set while we want
837 		 * to find the first bit cleared.
838 		 */
839 		const u_int n = __builtin_ffsl(bits) - 1;
840 		KASSERTMSG((bits << (nbpw - (n+1))) == (1ul << (nbpw-1)),
841 		    "n %u bits %#lx", n, bits);
842 		KASSERT(n < nbpw);
843 		ti->ti_asid_hint = n + i * nbpw;
844 	}
845 
846 	KASSERT(ti->ti_asid_hint > KERNEL_PID);
847 	KASSERT(ti->ti_asid_hint <= ti->ti_asid_max);
848 	KASSERTMSG(PMAP_TLB_FLUSH_ASID_ON_RESET
849 	    || TLBINFO_ASID_INUSE_P(ti, ti->ti_asid_hint - 1),
850 	    "hint %u bitmap %p", ti->ti_asid_hint, ti->ti_asid_bitmap);
851 	KASSERTMSG(!TLBINFO_ASID_INUSE_P(ti, ti->ti_asid_hint),
852 	    "hint %u bitmap %p", ti->ti_asid_hint, ti->ti_asid_bitmap);
853 
854 	/*
855 	 * The hint contains our next ASID so take it and advance the hint.
856 	 * Mark it as used and insert the pai into the list of active asids.
857 	 * There is also one less asid free in this TLB.
858 	 */
859 	KASSERT(ti->ti_asid_hint > KERNEL_PID);
860 	pai->pai_asid = ti->ti_asid_hint++;
861 #ifdef MULTIPROCESSOR
862 	if (PMAP_TLB_FLUSH_ASID_ON_RESET) {
863 		/*
864 		 * Clean the new ASID from the TLB.
865 		 */
866 		tlb_invalidate_asids(pai->pai_asid, pai->pai_asid);
867 	}
868 #endif
869 	TLBINFO_ASID_MARK_USED(ti, pai->pai_asid);
870 	LIST_INSERT_HEAD(&ti->ti_pais, pai, pai_link);
871 	ti->ti_asids_free--;
872 
873 #if defined(MULTIPROCESSOR)
874 	/*
875 	 * Mark that we now have an active ASID for all CPUs sharing this TLB.
876 	 * The bits in pm_active belonging to this TLB can only be changed
877 	 * while this TLBs lock is held.
878 	 */
879 #if PMAP_TLB_MAX == 1
880 	kcpuset_copy(pm->pm_active, kcpuset_running);
881 #else
882 	kcpuset_merge(pm->pm_active, ti->ti_kcpuset);
883 #endif
884 #endif
885 }
886 
887 /*
888  * Acquire a TLB address space tag (called ASID or TLBPID) and return it.
889  * ASID might have already been previously acquired.
890  */
891 void
pmap_tlb_asid_acquire(pmap_t pm,struct lwp * l)892 pmap_tlb_asid_acquire(pmap_t pm, struct lwp *l)
893 {
894 	struct cpu_info * const ci = l->l_cpu;
895 	struct pmap_tlb_info * const ti = cpu_tlb_info(ci);
896 	struct pmap_asid_info * const pai = PMAP_PAI(pm, ti);
897 
898 	UVMHIST_FUNC(__func__); UVMHIST_CALLED(maphist);
899 	UVMHIST_LOG(maphist, "(pm=%p, l=%p, ti=%p)", pm, l, ti, 0);
900 
901 	KASSERT(kpreempt_disabled());
902 
903 	/*
904 	 * Kernels use a fixed ASID and thus doesn't need to acquire one.
905 	 */
906 	if (pm == pmap_kernel()) {
907 		UVMHIST_LOG(maphist, " <-- done (kernel)", 0, 0, 0, 0);
908 		return;
909 	}
910 
911 	TLBINFO_LOCK(ti);
912 	KASSERT(pai->pai_asid <= KERNEL_PID || pai->pai_link.le_prev != NULL);
913 	KASSERT(pai->pai_asid > KERNEL_PID || pai->pai_link.le_prev == NULL);
914 	pmap_tlb_pai_check(ti, true);
915 	if (__predict_false(!PMAP_PAI_ASIDVALID_P(pai, ti))) {
916 		/*
917 		 * If we've run out ASIDs, reinitialize the ASID space.
918 		 */
919 		if (__predict_false(tlbinfo_noasids_p(ti))) {
920 			KASSERT(l == curlwp);
921 			UVMHIST_LOG(maphist, " asid reinit", 0, 0, 0, 0);
922 			pmap_tlb_asid_reinitialize(ti, TLBINV_NOBODY);
923 			KASSERT(!tlbinfo_noasids_p(ti));
924 		}
925 
926 		/*
927 		 * Get an ASID.
928 		 */
929 		pmap_tlb_asid_alloc(ti, pm, pai);
930 		UVMHIST_LOG(maphist, "allocated asid %#x", pai->pai_asid, 0, 0, 0);
931 	}
932 	pmap_tlb_pai_check(ti, true);
933 #if defined(MULTIPROCESSOR)
934 	KASSERT(kcpuset_isset(pm->pm_active, cpu_index(ci)));
935 #endif
936 
937 	if (l == curlwp) {
938 #if defined(MULTIPROCESSOR)
939 		/*
940 		 * The bits in pm_onproc belonging to this TLB can only
941 		 * be changed while this TLBs lock is held unless atomic
942 		 * operations are used.
943 		 */
944 		KASSERT(pm != pmap_kernel());
945 		kcpuset_atomic_set(pm->pm_onproc, cpu_index(ci));
946 #endif
947 		ci->ci_pmap_asid_cur = pai->pai_asid;
948 		UVMHIST_LOG(maphist, "setting asid to %#x", pai->pai_asid, 0, 0, 0);
949 		tlb_set_asid(pai->pai_asid);
950 		pmap_tlb_asid_check();
951 	} else {
952 		printf("%s: l (%p) != curlwp %p\n", __func__, l, curlwp);
953 	}
954 	TLBINFO_UNLOCK(ti);
955 	UVMHIST_LOG(maphist, " <-- done", 0, 0, 0, 0);
956 }
957 
958 void
pmap_tlb_asid_deactivate(pmap_t pm)959 pmap_tlb_asid_deactivate(pmap_t pm)
960 {
961 	UVMHIST_FUNC(__func__); UVMHIST_CALLED(maphist);
962 
963 	KASSERT(kpreempt_disabled());
964 #if defined(MULTIPROCESSOR)
965 	/*
966 	 * The kernel pmap is aways onproc and active and must never have
967 	 * those bits cleared.  If pmap_remove_all was called, it has already
968 	 * deactivated the pmap and thusly onproc will be 0 so there's nothing
969 	 * to do.
970 	 */
971 	if (pm != pmap_kernel() && !kcpuset_iszero(pm->pm_onproc)) {
972 		struct cpu_info * const ci = curcpu();
973 		KASSERT(!cpu_intr_p());
974 		KASSERTMSG(kcpuset_isset(pm->pm_onproc, cpu_index(ci)),
975 		    "%s: pmap %p onproc %p doesn't include cpu %d (%p)",
976 		    __func__, pm, pm->pm_onproc, cpu_index(ci), ci);
977 		/*
978 		 * The bits in pm_onproc that belong to this TLB can
979 		 * be changed while this TLBs lock is not held as long
980 		 * as we use atomic ops.
981 		 */
982 		kcpuset_atomic_clear(pm->pm_onproc, cpu_index(ci));
983 	}
984 #endif
985 	curcpu()->ci_pmap_asid_cur = KERNEL_PID;
986 	UVMHIST_LOG(maphist, " <-- done (pm=%p)", pm, 0, 0, 0);
987 	tlb_set_asid(KERNEL_PID);
988 	pmap_tlb_pai_check(cpu_tlb_info(curcpu()), false);
989 #if defined(DEBUG)
990 	pmap_tlb_asid_check();
991 #endif
992 }
993 
994 void
pmap_tlb_asid_release_all(struct pmap * pm)995 pmap_tlb_asid_release_all(struct pmap *pm)
996 {
997 	UVMHIST_FUNC(__func__); UVMHIST_CALLED(maphist);
998 	UVMHIST_LOG(maphist, "(pm=%p)", pm, 0, 0, 0);
999 
1000 	KASSERT(pm != pmap_kernel());
1001 #if defined(MULTIPROCESSOR)
1002 	//KASSERT(!kcpuset_iszero(pm->pm_onproc)); // XXX
1003 	struct cpu_info * const ci __diagused = curcpu();
1004 	KASSERT(!kcpuset_isotherset(pm->pm_onproc, cpu_index(ci)));
1005 #if PMAP_TLB_MAX > 1
1006 	for (u_int i = 0; !kcpuset_iszero(pm->pm_active); i++) {
1007 		KASSERT(i < pmap_ntlbs);
1008 		struct pmap_tlb_info * const ti = pmap_tlbs[i];
1009 #else
1010 		struct pmap_tlb_info * const ti = &pmap_tlb0_info;
1011 #endif
1012 		struct pmap_asid_info * const pai = PMAP_PAI(pm, ti);
1013 		TLBINFO_LOCK(ti);
1014 		if (PMAP_PAI_ASIDVALID_P(pai, ti)) {
1015 			/*
1016 			 * This pmap should not be in use by any other cpu so
1017 			 * we can just reset and be happy.
1018 			 */
1019 			if (ti->ti_victim == pm)
1020 				ti->ti_victim = NULL;
1021 			pmap_tlb_pai_reset(ti, pai, pm);
1022 		}
1023 		KASSERT(pai->pai_link.le_prev == NULL);
1024 		TLBINFO_UNLOCK(ti);
1025 #if PMAP_TLB_MAX > 1
1026 	}
1027 #endif
1028 #ifdef DIAGNOSTIC
1029 	for (size_t i = 0; i < (PMAP_TLB_MAX > 1 ? pmap_ntlbs : 1); i++) {
1030 		KASSERTMSG(pm->pm_pai[i].pai_asid == 0,
1031 		    "pm %p i %zu asid %u",
1032 		    pm, i, pm->pm_pai[i].pai_asid);
1033 	}
1034 #endif
1035 #else
1036 	/*
1037 	 * Handle the case of an UP kernel which only has, at most, one TLB.
1038 	 * If the pmap has an ASID allocated, free it.
1039 	 */
1040 	struct pmap_tlb_info * const ti = &pmap_tlb0_info;
1041 	struct pmap_asid_info * const pai = PMAP_PAI(pm, ti);
1042 	TLBINFO_LOCK(ti);
1043 	if (pai->pai_asid > KERNEL_PID) {
1044 		if (curcpu()->ci_pmap_asid_cur == pai->pai_asid) {
1045 			tlb_invalidate_asids(pai->pai_asid, pai->pai_asid);
1046 		} else {
1047 			pmap_tlb_pai_reset(ti, pai, pm);
1048 		}
1049 	}
1050 	TLBINFO_UNLOCK(ti);
1051 #endif /* MULTIPROCESSOR */
1052 	UVMHIST_LOG(maphist, " <-- done", 0, 0, 0, 0);
1053 }
1054 
1055 void
pmap_tlb_asid_check(void)1056 pmap_tlb_asid_check(void)
1057 {
1058 #ifdef DEBUG
1059 	kpreempt_disable();
1060 	const tlb_asid_t asid __debugused = tlb_get_asid();
1061 	KDASSERTMSG(asid == curcpu()->ci_pmap_asid_cur,
1062 	   "%s: asid (%#x) != current asid (%#x)",
1063 	    __func__, asid, curcpu()->ci_pmap_asid_cur);
1064 	kpreempt_enable();
1065 #endif
1066 }
1067 
1068 #ifdef DEBUG
1069 void
pmap_tlb_check(pmap_t pm,bool (* func)(void *,vaddr_t,tlb_asid_t,pt_entry_t))1070 pmap_tlb_check(pmap_t pm, bool (*func)(void *, vaddr_t, tlb_asid_t, pt_entry_t))
1071 {
1072         struct pmap_tlb_info * const ti = cpu_tlb_info(curcpu());
1073         struct pmap_asid_info * const pai = PMAP_PAI(pm, ti);
1074         TLBINFO_LOCK(ti);
1075         if (pm == pmap_kernel() || pai->pai_asid > KERNEL_PID)
1076 		tlb_walk(pm, func);
1077         TLBINFO_UNLOCK(ti);
1078 }
1079 #endif /* DEBUG */
1080