xref: /netbsd-src/sys/uvm/pmap/pmap_tlb.c (revision 6a493d6bc668897c91594964a732d38505b70cbb)
1 /*	$NetBSD: pmap_tlb.c,v 1.3 2013/07/17 23:15:20 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 1.3 2013/07/17 23:15:20 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_mutex __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 - (KERNEL_PID + 1)),
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_mutex,
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_USED(ti, asid) \
178 	__BITMAP_SET((ti)->ti_asid_bitmap, (asid))
179 #define	TLBINFO_ASID_INUSE_P(ti, asid) \
180 	__BITMAP_ISSET_P((ti)->ti_asid_bitmap, (asid))
181 
182 static void
183 pmap_pai_check(struct pmap_tlb_info *ti)
184 {
185 #ifdef DIAGNOSTIC
186 	struct pmap_asid_info *pai;
187 	LIST_FOREACH(pai, &ti->ti_pais, pai_link) {
188 		KASSERT(pai != NULL);
189 		KASSERT(PAI_PMAP(pai, ti) != pmap_kernel());
190 		KASSERT(pai->pai_asid > KERNEL_PID);
191 		KASSERT(TLBINFO_ASID_INUSE_P(ti, pai->pai_asid));
192 	}
193 #endif
194 }
195 
196 #ifdef MULTIPROCESSOR
197 static inline bool
198 pmap_tlb_intersecting_active_p(pmap_t pm, struct pmap_tlb_info *ti)
199 {
200 #if PMAP_TLB_MAX == 1
201 	return !kcpuset_iszero(pm->pm_active);
202 #else
203 	return kcpuset_intersecting_p(pm->pm_active, ti->ti_kcpuset);
204 #endif
205 }
206 
207 static inline bool
208 pmap_tlb_intersecting_onproc_p(pmap_t pm, struct pmap_tlb_info *ti)
209 {
210 #if PMAP_TLB_MAX == 1
211 	return !kcpuset_iszero(pm->pm_onproc);
212 #else
213 	return kcpuset_intersecting_p(pm->pm_onproc, ti->ti_kcpuset);
214 #endif
215 }
216 #endif
217 
218 static inline void
219 pmap_pai_reset(struct pmap_tlb_info *ti, struct pmap_asid_info *pai,
220 	struct pmap *pm)
221 {
222 	/*
223 	 * We must have an ASID but it must not be onproc (on a processor).
224 	 */
225 	KASSERT(pai->pai_asid > KERNEL_PID);
226 #if defined(MULTIPROCESSOR)
227 	KASSERT(!pmap_tlb_intersecting_onproc_p(pm, ti));
228 #endif
229 	LIST_REMOVE(pai, pai_link);
230 #ifdef DIAGNOSTIC
231 	pai->pai_link.le_prev = NULL;	/* tagged as unlinked */
232 #endif
233 	/*
234 	 * Note that we don't mark the ASID as not in use in the TLB's ASID
235 	 * bitmap (thus it can't be allocated until the ASID space is exhausted
236 	 * and therefore reinitialized).  We don't want to flush the TLB for
237 	 * entries belonging to this ASID so we will let natural TLB entry
238 	 * replacement flush them out of the TLB.  Any new entries for this
239 	 * pmap will need a new ASID allocated.
240 	 */
241 	pai->pai_asid = 0;
242 
243 #if defined(MULTIPROCESSOR)
244 	/*
245 	 * The bits in pm_active belonging to this TLB can only be changed
246 	 * while this TLB's lock is held.
247 	 */
248 #if PMAP_TLB_MAX == 1
249 	kcpuset_zero(pm->pm_active);
250 #else
251 	kcpuset_atomicly_remove(pm->pm_active, ti->ti_kcpuset);
252 #endif
253 #endif /* MULTIPROCESSOR */
254 }
255 
256 void
257 pmap_tlb_info_evcnt_attach(struct pmap_tlb_info *ti)
258 {
259 #if defined(MULTIPROCESSOR)
260 	evcnt_attach_dynamic_nozero(&ti->ti_evcnt_synci_desired,
261 	    EVCNT_TYPE_MISC, NULL,
262 	    ti->ti_name, "icache syncs desired");
263 	evcnt_attach_dynamic_nozero(&ti->ti_evcnt_synci_asts,
264 	    EVCNT_TYPE_MISC, &ti->ti_evcnt_synci_desired,
265 	    ti->ti_name, "icache sync asts");
266 	evcnt_attach_dynamic_nozero(&ti->ti_evcnt_synci_all,
267 	    EVCNT_TYPE_MISC, &ti->ti_evcnt_synci_asts,
268 	    ti->ti_name, "icache full syncs");
269 	evcnt_attach_dynamic_nozero(&ti->ti_evcnt_synci_pages,
270 	    EVCNT_TYPE_MISC, &ti->ti_evcnt_synci_asts,
271 	    ti->ti_name, "icache pages synced");
272 	evcnt_attach_dynamic_nozero(&ti->ti_evcnt_synci_duplicate,
273 	    EVCNT_TYPE_MISC, &ti->ti_evcnt_synci_desired,
274 	    ti->ti_name, "icache dup pages skipped");
275 	evcnt_attach_dynamic_nozero(&ti->ti_evcnt_synci_deferred,
276 	    EVCNT_TYPE_MISC, &ti->ti_evcnt_synci_desired,
277 	    ti->ti_name, "icache pages deferred");
278 #endif /* MULTIPROCESSOR */
279 	evcnt_attach_dynamic_nozero(&ti->ti_evcnt_asid_reinits,
280 	    EVCNT_TYPE_MISC, NULL,
281 	    ti->ti_name, "asid pool reinit");
282 }
283 
284 void
285 pmap_tlb_info_init(struct pmap_tlb_info *ti)
286 {
287 #if defined(MULTIPROCESSOR)
288 #if PMAP_TLB_MAX == 1
289 	KASSERT(ti == &pmap_tlb0_info);
290 #else
291 	if (ti != &pmap_tlb0_info) {
292 		KASSERT(pmap_ntlbs < PMAP_TLB_MAX);
293 
294 		KASSERT(pmap_tlbs[pmap_ntlbs] == NULL);
295 
296 		ti->ti_lock = mutex_obj_alloc(MUTEX_DEFAULT, IPL_SCHED);
297 		ti->ti_asid_bitmap[0] = (2 << KERNEL_PID) - 1;
298 		ti->ti_asid_hint = KERNEL_PID + 1;
299 		ti->ti_asid_max = pmap_tlbs[0]->ti_asid_max;
300 		ti->ti_asids_free = ti->ti_asid_max - KERNEL_PID;
301 		ti->ti_tlbinvop = TLBINV_NOBODY,
302 		ti->ti_victim = NULL;
303 		kcpuset_create(&ti->ti_kcpuset, true);
304 		ti->ti_index = pmap_ntlbs++;
305 		ti->ti_wired = 0;
306 		pmap_tlbs[ti->ti_index] = ti;
307 		snprintf(ti->ti_name, sizeof(ti->ti_name), "tlb%u",
308 		    ti->ti_index);
309 		pmap_tlb_info_evcnt_attach(ti);
310 		return;
311 	}
312 #endif
313 #endif /* MULTIPROCESSOR */
314 	KASSERT(ti == &pmap_tlb0_info);
315 	mutex_init(ti->ti_lock, MUTEX_DEFAULT, IPL_SCHED);
316 #if defined(MULTIPROCESSOR) && PMAP_TLB_MAX > 1
317 	kcpuset_create(&ti->ti_kcpuset, true);
318 	kcpuset_set(&ti->ti_kcpuset, cpu_index(curcpu()));
319 #endif
320 	if (ti->ti_asid_max == 0) {
321 		ti->ti_asid_max = pmap_md_tlb_asid_max();
322 		ti->ti_asids_free = ti->ti_asid_max - (KERNEL_PID + 1);
323 	}
324 
325 	KASSERT(ti->ti_asid_max < sizeof(ti->ti_asid_bitmap)*8);
326 }
327 
328 #if defined(MULTIPROCESSOR)
329 void
330 pmap_tlb_info_attach(struct pmap_tlb_info *ti, struct cpu_info *ci)
331 {
332 	KASSERT(!CPU_IS_PRIMARY(ci));
333 	KASSERT(ci->ci_data.cpu_idlelwp != NULL);
334 	KASSERT(cold);
335 
336 	TLBINFO_LOCK(ti);
337 #if PMAP_TLB_MAX > 1
338 	kcpuset_set(ti->ti_kcpuset, cpu_index(ci));
339 #endif
340 	cpu_set_tlb_info(ci, ti);
341 
342 	/*
343 	 * Do any MD tlb info init.
344 	 */
345 	pmap_md_tlb_info_attach(ti, ci);
346 
347 	/*
348 	 * The kernel pmap uses the kcpuset_running set so it's always
349 	 * up-to-date.
350 	 */
351 	TLBINFO_UNLOCK(ti);
352 }
353 #endif /* MULTIPROCESSOR */
354 
355 #ifdef DIAGNOSTIC
356 static size_t
357 pmap_tlb_asid_count(struct pmap_tlb_info *ti)
358 {
359 	size_t count = 0;
360 	for (tlb_asid_t asid = 1; asid <= ti->ti_asid_max; asid++) {
361 		count += TLBINFO_ASID_INUSE_P(ti, asid);
362 	}
363 	return count;
364 }
365 #endif
366 
367 static void
368 pmap_tlb_asid_reinitialize(struct pmap_tlb_info *ti, enum tlb_invalidate_op op)
369 {
370 	const size_t asid_bitmap_words =
371 	    ti->ti_asid_max / (8 * sizeof(ti->ti_asid_bitmap[0]));
372 
373 	pmap_pai_check(ti);
374 
375 	/*
376 	 * First, clear the ASID bitmap (except for ASID 0 which belongs
377 	 * to the kernel).
378 	 */
379 	ti->ti_asids_free = ti->ti_asid_max - KERNEL_PID;
380 	ti->ti_asid_hint = KERNEL_PID + 1;
381 	ti->ti_asid_bitmap[0] = (2 << KERNEL_PID) - 1;
382 	for (size_t word = 1; word <= asid_bitmap_words; word++) {
383 		ti->ti_asid_bitmap[word] = 0;
384 	}
385 
386 	switch (op) {
387 #if defined(MULTIPROCESSOR) && defined(PMAP_NEED_TLB_SHOOTDOWN)
388 	case TLBINV_ALL:
389 		tlb_invalidate_all();
390 		break;
391 	case TLBINV_ALLUSER:
392 		tlb_invalidate_asids(KERNEL_PID + 1, ti->ti_asid_max);
393 		break;
394 #endif /* MULTIPROCESSOR && PMAP_NEED_TLB_SHOOTDOWN */
395 	case TLBINV_NOBODY: {
396 		/*
397 		 * If we are just reclaiming ASIDs in the TLB, let's go find
398 		 * what ASIDs are in use in the TLB.  Since this is a
399 		 * semi-expensive operation, we don't want to do it too often.
400 		 * So if more half of the ASIDs are in use, we don't have
401 		 * enough free ASIDs so invalidate the TLB entries with ASIDs
402 		 * and clear the ASID bitmap.  That will force everyone to
403 		 * allocate a new ASID.
404 		 */
405 #if !defined(MULTIPROCESSOR) || defined(PMAP_NEED_TLB_SHOOTDOWN)
406 		pmap_tlb_asid_check();
407 		const u_int asids_found = tlb_record_asids(ti->ti_asid_bitmap);
408 		pmap_tlb_asid_check();
409 		KASSERT(asids_found == pmap_tlb_asid_count(ti));
410 		if (__predict_false(asids_found >= ti->ti_asid_max / 2)) {
411 			tlb_invalidate_asids(KERNEL_PID + 1, ti->ti_asid_max);
412 #else /* MULTIPROCESSOR && !PMAP_NEED_TLB_SHOOTDOWN */
413 			/*
414 			 * For those systems (PowerPC) that don't need require
415 			 * cross cpu TLB shootdowns, we have to invalidate the
416 			 * entire TLB because we can't record the ASIDs in use
417 			 * on the other CPUs.  This is hopefully cheaper than
418 			 * than trying to use an IPI to record all the ASIDs
419 			 * on all the CPUs (which would be a synchronization
420 			 * nightmare).
421 			 */
422 			tlb_invalidate_all();
423 #endif /* MULTIPROCESSOR && !PMAP_NEED_TLB_SHOOTDOWN */
424 			ti->ti_asid_bitmap[0] = (2 << KERNEL_PID) - 1;
425 			for (size_t word = 1;
426 			     word <= asid_bitmap_words;
427 			     word++) {
428 				ti->ti_asid_bitmap[word] = 0;
429 			}
430 #if !defined(MULTIPROCESSOR) || defined(PMAP_NEED_TLB_SHOOTDOWN)
431 		} else {
432 			ti->ti_asids_free -= asids_found;
433 		}
434 #endif /* !MULTIPROCESSOR || PMAP_NEED_TLB_SHOOTDOWN */
435 		break;
436 	}
437 	default:
438 		panic("%s: unexpected op %d", __func__, op);
439 	}
440 
441 	/*
442 	 * Now go through the active ASIDs.  If the ASID is on a processor or
443 	 * we aren't invalidating all ASIDs and the TLB has an entry owned by
444 	 * that ASID, mark it as in use.  Otherwise release the ASID.
445 	 */
446 	struct pmap_asid_info *pai, *next;
447 	for (pai = LIST_FIRST(&ti->ti_pais); pai != NULL; pai = next) {
448 		struct pmap * const pm = PAI_PMAP(pai, ti);
449 		next = LIST_NEXT(pai, pai_link);
450 		KASSERT(pm != pmap_kernel());
451 		KASSERT(pai->pai_asid > KERNEL_PID);
452 #if defined(MULTIPROCESSOR)
453 		if (pmap_tlb_intersecting_onproc_p(pm, ti)) {
454 			if (!TLBINFO_ASID_INUSE_P(ti, pai->pai_asid)) {
455 				TLBINFO_ASID_MARK_USED(ti, pai->pai_asid);
456 				ti->ti_asids_free--;
457 			}
458 			continue;
459 		}
460 #endif /* MULTIPROCESSOR */
461 		if (TLBINFO_ASID_INUSE_P(ti, pai->pai_asid)) {
462 			KASSERT(op == TLBINV_NOBODY);
463 		} else {
464 			pmap_pai_reset(ti, pai, pm);
465 		}
466 	}
467 #ifdef DIAGNOSTIC
468 	size_t free_count = ti->ti_asid_max - pmap_tlb_asid_count(ti);
469 	if (free_count != ti->ti_asids_free)
470 		panic("%s: bitmap error: %zu != %u",
471 		    __func__, free_count, ti->ti_asids_free);
472 #endif
473 }
474 
475 #if defined(MULTIPROCESSOR) && defined(PMAP_NEED_TLB_SHOOTDOWN)
476 #if PMAP_MAX_TLB == 1
477 #error shootdown not required for single TLB systems
478 #endif
479 void
480 pmap_tlb_shootdown_process(void)
481 {
482 	struct cpu_info * const ci = curcpu();
483 	struct pmap_tlb_info * const ti = cpu_tlb_info(ci);
484 #ifdef DIAGNOSTIC
485 	struct pmap * const pm = curlwp->l_proc->p_vmspace->vm_map.pmap;
486 #endif
487 
488 	KASSERT(cpu_intr_p());
489 	KASSERTMSG(ci->ci_cpl >= IPL_SCHED,
490 	    "%s: cpl (%d) < IPL_SCHED (%d)",
491 	    __func__, ci->ci_cpl, IPL_SCHED);
492 
493 	TLBINFO_LOCK(ti);
494 
495 	switch (ti->ti_tlbinvop) {
496 	case TLBINV_ONE: {
497 		/*
498 		 * We only need to invalidate one user ASID.
499 		 */
500 		struct pmap_asid_info * const pai = PMAP_PAI(ti->ti_victim, ti);
501 		KASSERT(ti->ti_victim != pmap_kernel());
502 		if (!pmap_tlb_intersecting_onproc_p(ti_victim->pm_onproc, ti)) {
503 			/*
504 			 * The victim is an active pmap so we will just
505 			 * invalidate its TLB entries.
506 			 */
507 			KASSERT(pai->pai_asid > KERNEL_PID);
508 			pmap_tlb_asid_check();
509 			tlb_invalidate_asids(pai->pai_asid, pai->pai_asid);
510 			pmap_tlb_asid_check();
511 		} else if (pai->pai_asid) {
512 			/*
513 			 * The victim is no longer an active pmap for this TLB.
514 			 * So simply clear its ASID and when pmap_activate is
515 			 * next called for this pmap, it will allocate a new
516 			 * ASID.
517 			 */
518 			KASSERT(!pmap_tlb_intersecting_onproc_p(pm, ti));
519 			pmap_pai_reset(ti, pai, PAI_PMAP(pai, ti));
520 		}
521 		break;
522 	}
523 	case TLBINV_ALLUSER:
524 		/*
525 		 * Flush all user TLB entries.
526 		 */
527 		pmap_tlb_asid_reinitialize(ti, TLBINV_ALLUSER);
528 		break;
529 	case TLBINV_ALLKERNEL:
530 		/*
531 		 * We need to invalidate all global TLB entries.
532 		 */
533 		pmap_tlb_asid_check();
534 		tlb_invalidate_globals();
535 		pmap_tlb_asid_check();
536 		break;
537 	case TLBINV_ALL:
538 		/*
539 		 * Flush all the TLB entries (user and kernel).
540 		 */
541 		pmap_tlb_asid_reinitialize(ti, TLBINV_ALL);
542 		break;
543 	case TLBINV_NOBODY:
544 		/*
545 		 * Might be spurious or another SMT CPU sharing this TLB
546 		 * could have already done the work.
547 		 */
548 		break;
549 	}
550 
551 	/*
552 	 * Indicate we are done with shutdown event.
553 	 */
554 	ti->ti_victim = NULL;
555 	ti->ti_tlbinvop = TLBINV_NOBODY;
556 	TLBINFO_UNLOCK(ti);
557 }
558 
559 /*
560  * This state machine could be encoded into an array of integers but since all
561  * the values fit in 3 bits, the 5 entry "table" fits in a 16 bit value which
562  * can be loaded in a single instruction.
563  */
564 #define	TLBINV_MAP(op, nobody, one, alluser, allkernel, all)	\
565 	((((   (nobody) << 3*TLBINV_NOBODY)			\
566 	 | (      (one) << 3*TLBINV_ONE)			\
567 	 | (  (alluser) << 3*TLBINV_ALLUSER)			\
568 	 | ((allkernel) << 3*TLBINV_ALLKERNEL)			\
569 	 | (      (all) << 3*TLBINV_ALL)) >> 3*(op)) & 7)
570 
571 #define	TLBINV_USER_MAP(op)	\
572 	TLBINV_MAP(op, TLBINV_ONE, TLBINV_ALLUSER, TLBINV_ALLUSER,	\
573 	    TLBINV_ALL, TLBINV_ALL)
574 
575 #define	TLBINV_KERNEL_MAP(op)	\
576 	TLBINV_MAP(op, TLBINV_ALLKERNEL, TLBINV_ALL, TLBINV_ALL,	\
577 	    TLBINV_ALLKERNEL, TLBINV_ALL)
578 
579 bool
580 pmap_tlb_shootdown_bystanders(pmap_t pm)
581 {
582 	/*
583 	 * We don't need to deal our own TLB.
584 	 */
585 	kcpuset_t *pm_active;
586 
587 	kcpuset_clone(&pm_active, pm->pm_active);
588 	kcpuset_atomicly_remove(pm->pm_active,
589 	    cpu_tlb_info(curcpu())->ti_kcpuset);
590 	const bool kernel_p = (pm == pmap_kernel());
591 	bool ipi_sent = false;
592 
593 	/*
594 	 * If pm_active gets more bits set, then it's after all our changes
595 	 * have been made so they will already be cognizant of them.
596 	 */
597 
598 	for (size_t i = 0; !kcpuset_iszero(pm_active); i++) {
599 		KASSERT(i < pmap_ntlbs);
600 		struct pmap_tlb_info * const ti = pmap_tlbs[i];
601 		KASSERT(tlbinfo_index(ti) == i);
602 		/*
603 		 * Skip this TLB if there are no active mappings for it.
604 		 */
605 		if (!kcpuset_intersecting_p(pm_active, ti->ti_kcpuset))
606 			continue;
607 		struct pmap_asid_info * const pai = PMAP_PAI(pm, ti);
608 		kcpuset_remove(pm_active, ti->ti_kcpuset);
609 		TLBINFO_LOCK(ti);
610 		if (pmap_tlb_intersecting_onproc_p(pm, ti)) {
611 			cpuid_t j = kcpuset_ffs_intersecting(pm->pm_onproc,
612 			    ti->ti_kcpuset);
613 			if (kernel_p) {
614 				ti->ti_tlbinvop =
615 				    TLBINV_KERNEL_MAP(ti->ti_tlbinvop);
616 				ti->ti_victim = NULL;
617 			} else {
618 				KASSERT(pai->pai_asid);
619 				if (__predict_false(ti->ti_victim == pm)) {
620 					KASSERT(ti->ti_tlbinvop == TLBINV_ONE);
621 					/*
622 					 * We still need to invalidate this one
623 					 * ASID so there's nothing to change.
624 					 */
625 				} else {
626 					ti->ti_tlbinvop =
627 					    TLBINV_USER_MAP(ti->ti_tlbinvop);
628 					if (ti->ti_tlbinvop == TLBINV_ONE)
629 						ti->ti_victim = pm;
630 					else
631 						ti->ti_victim = NULL;
632 				}
633 			}
634 			TLBINFO_UNLOCK(ti);
635 			/*
636 			 * Now we can send out the shootdown IPIs to a CPU
637 			 * that shares this TLB and is currently using this
638 			 * pmap.  That CPU will process the IPI and do the
639 			 * all the work.  Any other CPUs sharing that TLB
640 			 * will take advantage of that work.  pm_onproc might
641 			 * change now that we have released the lock but we
642 			 * can tolerate spurious shootdowns.
643 			 */
644 			cpu_send_ipi(cpu_lookup(j), IPI_SHOOTDOWN);
645 			ipi_sent = true;
646 			continue;
647 		}
648 		if (!pmap_tlb_intersecting_active_p(pm, ti)) {
649 			/*
650 			 * If this pmap has an ASID assigned but it's not
651 			 * currently running, nuke its ASID.  Next time the
652 			 * pmap is activated, it will allocate a new ASID.
653 			 * And best of all, we avoid an IPI.
654 			 */
655 			KASSERT(!kernel_p);
656 			pmap_pai_reset(ti, pai, pm);
657 			//ti->ti_evcnt_lazy_shots.ev_count++;
658 		}
659 		TLBINFO_UNLOCK(ti);
660 	}
661 
662 	kcpuset_destroy(pm_active);
663 
664 	return ipi_sent;
665 }
666 #endif /* MULTIPROCESSOR && PMAP_NEED_TLB_SHOOTDOWN */
667 
668 #ifndef PMAP_TLB_HWPAGEWALKER
669 int
670 pmap_tlb_update_addr(pmap_t pm, vaddr_t va, pt_entry_t pt_entry, u_int flags)
671 {
672 	struct pmap_tlb_info * const ti = cpu_tlb_info(curcpu());
673 	struct pmap_asid_info * const pai = PMAP_PAI(pm, ti);
674 	int rv = -1;
675 
676 	KASSERT(kpreempt_disabled());
677 
678 	TLBINFO_LOCK(ti);
679 	if (pm == pmap_kernel() || PMAP_PAI_ASIDVALID_P(pai, ti)) {
680 		pmap_tlb_asid_check();
681 		rv = tlb_update_addr(va, pai->pai_asid, pt_entry,
682 		    (flags & PMAP_TLB_INSERT) != 0);
683 		pmap_tlb_asid_check();
684 	}
685 #if defined(MULTIPROCESSOR) && defined(PMAP_NEED_TLB_SHOOTDOWN)
686 	pm->pm_shootdown_pending = (flags & PMAP_TLB_NEED_IPI) != 0;
687 #endif
688 	TLBINFO_UNLOCK(ti);
689 
690 	return rv;
691 }
692 #endif /* !PMAP_TLB_HWPAGEWALKER */
693 
694 void
695 pmap_tlb_invalidate_addr(pmap_t pm, vaddr_t va)
696 {
697 	struct pmap_tlb_info * const ti = cpu_tlb_info(curcpu());
698 	struct pmap_asid_info * const pai = PMAP_PAI(pm, ti);
699 
700 	KASSERT(kpreempt_disabled());
701 
702 	TLBINFO_LOCK(ti);
703 	if (pm == pmap_kernel() || PMAP_PAI_ASIDVALID_P(pai, ti)) {
704 		pmap_tlb_asid_check();
705 		tlb_invalidate_addr(va, pai->pai_asid);
706 		pmap_tlb_asid_check();
707 	}
708 #if defined(MULTIPROCESSOR) && defined(PMAP_NEED_TLB_SHOOTDOWN)
709 	pm->pm_shootdown_pending = 1;
710 #endif
711 	TLBINFO_UNLOCK(ti);
712 }
713 
714 static inline void
715 pmap_tlb_asid_alloc(struct pmap_tlb_info *ti, pmap_t pm,
716 	struct pmap_asid_info *pai)
717 {
718 	/*
719 	 * We shouldn't have an ASID assigned, and thusly must not be onproc
720 	 * nor active.
721 	 */
722 	KASSERT(pm != pmap_kernel());
723 	KASSERT(pai->pai_asid == 0);
724 	KASSERT(pai->pai_link.le_prev == NULL);
725 #if defined(MULTIPROCESSOR)
726 	KASSERT(pmap_tlb_intersecting_onproc_p(pm, ti));
727 	KASSERT(pmap_tlb_intersecting_active_p(pm, ti));
728 #endif
729 	KASSERT(ti->ti_asids_free > 0);
730 	KASSERT(ti->ti_asid_hint <= ti->ti_asid_max);
731 
732 	/*
733 	 * Let's see if the hinted ASID is free.  If not search for
734 	 * a new one.
735 	 */
736 	if (__predict_false(TLBINFO_ASID_INUSE_P(ti, ti->ti_asid_hint))) {
737 #ifdef DIAGNOSTIC
738 		const size_t words = __arraycount(ti->ti_asid_bitmap);
739 #endif
740 		const size_t nbpw = 8 * sizeof(ti->ti_asid_bitmap[0]);
741 		for (size_t i = 0; i < ti->ti_asid_hint / nbpw; i++) {
742 			KASSERT(~ti->ti_asid_bitmap[i] == 0);
743 		}
744 		for (size_t i = ti->ti_asid_hint / nbpw;; i++) {
745 			KASSERT(i < words);
746 			/*
747 			 * ffs wants to find the first bit set while we want
748 			 * to find the first bit cleared.
749 			 */
750 			u_long bits = ~ti->ti_asid_bitmap[i];
751 			if (__predict_true(bits)) {
752 				u_int n = 0;
753 				if ((bits & 0xffffffff) == 0)  {
754 					bits = (bits >> 31) >> 1;
755 					KASSERT(bits);
756 					n += 32;
757 				}
758 				n += ffs(bits) - 1;
759 				KASSERT(n < nbpw);
760 				ti->ti_asid_hint = n + i * nbpw;
761 				break;
762 			}
763 		}
764 		KASSERT(ti->ti_asid_hint > KERNEL_PID);
765 		KASSERT(TLBINFO_ASID_INUSE_P(ti, ti->ti_asid_hint-1));
766 		KASSERT(!TLBINFO_ASID_INUSE_P(ti, ti->ti_asid_hint));
767 	}
768 
769 	/*
770 	 * The hint contains our next ASID so take it and advance the hint.
771 	 * Mark it as used and insert the pai into the list of active asids.
772 	 * There is also one less asid free in this TLB.
773 	 */
774 	pai->pai_asid = ti->ti_asid_hint++;
775 	TLBINFO_ASID_MARK_USED(ti, pai->pai_asid);
776 	LIST_INSERT_HEAD(&ti->ti_pais, pai, pai_link);
777 	ti->ti_asids_free--;
778 
779 #if defined(MULTIPROCESSOR)
780 	/*
781 	 * Mark that we now have an active ASID for all CPUs sharing this TLB.
782 	 * The bits in pm_active belonging to this TLB can only be changed
783 	 * while this TLBs lock is held.
784 	 */
785 #if PMAP_TLB_MAX == 1
786 	kcpuset_copy(pm->pm_active, kcpuset_running);
787 #else
788 	kcpuset_atomicly_merge(pm->pm_active, ti->ti_kcpuset);
789 #endif
790 #endif
791 }
792 
793 /*
794  * Acquire a TLB address space tag (called ASID or TLBPID) and return it.
795  * ASID might have already been previously acquired.
796  */
797 void
798 pmap_tlb_asid_acquire(pmap_t pm, struct lwp *l)
799 {
800 	struct cpu_info * const ci = l->l_cpu;
801 	struct pmap_tlb_info * const ti = cpu_tlb_info(ci);
802 	struct pmap_asid_info * const pai = PMAP_PAI(pm, ti);
803 
804 	KASSERT(kpreempt_disabled());
805 
806 	/*
807 	 * Kernels use a fixed ASID and thus doesn't need to acquire one.
808 	 */
809 	if (pm == pmap_kernel())
810 		return;
811 
812 	TLBINFO_LOCK(ti);
813 	KASSERT(pai->pai_asid <= KERNEL_PID || pai->pai_link.le_prev != NULL);
814 	KASSERT(pai->pai_asid > KERNEL_PID || pai->pai_link.le_prev == NULL);
815 	pmap_pai_check(ti);
816 	if (__predict_false(!PMAP_PAI_ASIDVALID_P(pai, ti))) {
817 		/*
818 		 * If we've run out ASIDs, reinitialize the ASID space.
819 		 */
820 		if (__predict_false(tlbinfo_noasids_p(ti))) {
821 			KASSERT(l == curlwp);
822 			pmap_tlb_asid_reinitialize(ti, TLBINV_NOBODY);
823 		}
824 
825 		/*
826 		 * Get an ASID.
827 		 */
828 		pmap_tlb_asid_alloc(ti, pm, pai);
829 	}
830 
831 	if (l == curlwp) {
832 #if defined(MULTIPROCESSOR)
833 		/*
834 		 * The bits in pm_onproc belonging to this TLB can only
835 		 * be changed while this TLBs lock is held unless atomic
836 		 * operations are used.
837 		 */
838 		kcpuset_atomic_set(pm->pm_onproc, cpu_index(ci));
839 #endif
840 		ci->ci_pmap_asid_cur = pai->pai_asid;
841 		tlb_set_asid(pai->pai_asid);
842 		pmap_tlb_asid_check();
843 	} else {
844 		printf("%s: l (%p) != curlwp %p\n", __func__, l, curlwp);
845 	}
846 	TLBINFO_UNLOCK(ti);
847 }
848 
849 void
850 pmap_tlb_asid_deactivate(pmap_t pm)
851 {
852 	KASSERT(kpreempt_disabled());
853 #if defined(MULTIPROCESSOR)
854 	/*
855 	 * The kernel pmap is aways onproc and active and must never have
856 	 * those bits cleared.  If pmap_remove_all was called, it has already
857 	 * deactivated the pmap and thusly onproc will be 0 so there's nothing
858 	 * to do.
859 	 */
860 	if (pm != pmap_kernel() && pm->pm_onproc != 0) {
861 		struct cpu_info * const ci = curcpu();
862 		KASSERT(!cpu_intr_p());
863 		KASSERTMSG(kcpuset_isset(pm->pm_onproc, cpu_index(ci)),
864 		    "%s: pmap %p onproc %p doesn't include cpu %d (%p)",
865 		    __func__, pm, pm->pm_onproc, cpu_index(ci), ci);
866 		/*
867 		 * The bits in pm_onproc that belong to this TLB can
868 		 * be changed while this TLBs lock is not held as long
869 		 * as we use atomic ops.
870 		 */
871 		kcpuset_atomic_clear(pm->pm_onproc, cpu_index(ci));
872 	}
873 #elif defined(DEBUG)
874 	curcpu()->ci_pmap_asid_cur = 0;
875 	tlb_set_asid(0);
876 	pmap_tlb_asid_check();
877 #endif
878 }
879 
880 void
881 pmap_tlb_asid_release_all(struct pmap *pm)
882 {
883 	KASSERT(pm != pmap_kernel());
884 #if defined(MULTIPROCESSOR)
885 	KASSERT(kcpuset_iszero(pm->pm_onproc));
886 #if PMAP_TLB_MAX > 1
887 	for (u_int i = 0; !kcpuset_iszero(pm->pm_active); i++) {
888 		KASSERT(i < pmap_ntlbs);
889 		struct pmap_tlb_info * const ti = pmap_tlbs[i];
890 #else
891 		struct pmap_tlb_info * const ti = &pmap_tlb0_info;
892 #endif
893 		if (!pmap_tlb_intersecting_onproc_p(pm, ti)) {
894 			struct pmap_asid_info * const pai = PMAP_PAI(pm, ti);
895 			TLBINFO_LOCK(ti);
896 			KASSERT(ti->ti_victim != pm);
897 			pmap_pai_reset(ti, pai, pm);
898 			TLBINFO_UNLOCK(ti);
899 		}
900 #if PMAP_TLB_MAX > 1
901 	}
902 #endif
903 #else
904 	/*
905 	 * Handle the case of an UP kernel which only has, at most, one ASID.
906 	 * If the pmap has an ASID allocated, free it.
907 	 */
908 	struct pmap_tlb_info * const ti = &pmap_tlb0_info;
909 	struct pmap_asid_info * const pai = PMAP_PAI(pm, ti);
910 	TLBINFO_LOCK(ti);
911 	if (pai->pai_asid > KERNEL_PID) {
912 		pmap_pai_reset(ti, pai, pm);
913 	}
914 	TLBINFO_UNLOCK(ti);
915 #endif /* MULTIPROCESSOR */
916 }
917 
918 void
919 pmap_tlb_asid_check(void)
920 {
921 #ifdef DEBUG
922 	kpreempt_disable();
923 	const tlb_asid_t asid = tlb_get_asid();
924 	KDASSERTMSG(asid == curcpu()->ci_pmap_asid_cur,
925 	   "%s: asid (%#x) != current asid (%#x)",
926 	    __func__, asid, curcpu()->ci_pmap_asid_cur);
927 	kpreempt_enable();
928 #endif
929 }
930 
931 #ifdef DEBUG
932 void
933 pmap_tlb_check(pmap_t pm, bool (*func)(void *, vaddr_t, tlb_asid_t, pt_entry_t))
934 {
935         struct pmap_tlb_info * const ti = cpu_tlb_info(curcpu());
936         struct pmap_asid_info * const pai = PMAP_PAI(pm, ti);
937         TLBINFO_LOCK(ti);
938         if (pm == pmap_kernel() || pai->pai_asid > KERNEL_PID)
939 		tlb_walk(pm, func);
940         TLBINFO_UNLOCK(ti);
941 }
942 #endif /* DEBUG */
943