xref: /netbsd-src/sys/uvm/pmap/pmap_tlb.c (revision fbbd98ed5c5a00f438f37f36c44c3c644ad5bea5)
1 /*	$NetBSD: pmap_tlb.c,v 1.62 2024/01/01 16:56:30 skrll 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 1.62 2024/01/01 16:56:30 skrll 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
pmap_tlb_intersecting_active_p(pmap_t pm,struct pmap_tlb_info * ti)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
pmap_tlb_intersecting_onproc_p(pmap_t pm,struct pmap_tlb_info * ti)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
pmap_tlb_pai_check(struct pmap_tlb_info * ti,bool locked_p)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
pmap_tlb_pai_reset(struct pmap_tlb_info * ti,struct pmap_asid_info * pai,struct pmap * pm)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
pmap_tlb_info_evcnt_attach(struct pmap_tlb_info * ti)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
pmap_tlb_info_init(struct pmap_tlb_info * ti)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(__type_fit(tlb_asid_t, ti->ti_asid_max + 1));
388 	KASSERT(ti->ti_asid_max < PMAP_TLB_BITMAP_LENGTH);
389 }
390 
391 #if defined(MULTIPROCESSOR)
392 void
pmap_tlb_info_attach(struct pmap_tlb_info * ti,struct cpu_info * ci)393 pmap_tlb_info_attach(struct pmap_tlb_info *ti, struct cpu_info *ci)
394 {
395 	KASSERT(!CPU_IS_PRIMARY(ci));
396 	KASSERT(ci->ci_data.cpu_idlelwp != NULL);
397 	KASSERT(cold);
398 
399 	TLBINFO_LOCK(ti);
400 #if PMAP_TLB_MAX > 1
401 	kcpuset_set(ti->ti_kcpuset, cpu_index(ci));
402 	cpu_set_tlb_info(ci, ti);
403 #endif
404 
405 	/*
406 	 * Do any MD tlb info init.
407 	 */
408 	pmap_md_tlb_info_attach(ti, ci);
409 
410 	/*
411 	 * The kernel pmap uses the kcpuset_running set so it's always
412 	 * up-to-date.
413 	 */
414 	TLBINFO_UNLOCK(ti);
415 }
416 #endif /* MULTIPROCESSOR */
417 
418 #ifdef DIAGNOSTIC
419 static size_t
pmap_tlb_asid_count(struct pmap_tlb_info * ti)420 pmap_tlb_asid_count(struct pmap_tlb_info *ti)
421 {
422 	size_t count = 0;
423 	for (tlb_asid_t asid = 1; asid <= ti->ti_asid_max; asid++) {
424 		if (TLBINFO_ASID_INUSE_P(ti, asid))
425 			count++;
426 	}
427 	return count;
428 }
429 #endif
430 
431 static void
pmap_tlb_asid_reinitialize(struct pmap_tlb_info * ti,enum tlb_invalidate_op op)432 pmap_tlb_asid_reinitialize(struct pmap_tlb_info *ti, enum tlb_invalidate_op op)
433 {
434 	UVMHIST_FUNC(__func__);
435 	UVMHIST_CALLARGS(maphist, "(ti=%#jx, op=%ju)", (uintptr_t)ti, op, 0, 0);
436 
437 	pmap_tlb_pai_check(ti, true);
438 
439 	ti->ti_evcnt_asid_reinits.ev_count++;
440 
441 	/*
442 	 * First, clear the ASID bitmap (except for ASID 0 which belongs
443 	 * to the kernel).
444 	 */
445 	ti->ti_asids_free = TLBINFO_ASID_INITIAL_FREE(ti->ti_asid_max);
446 	ti->ti_asid_hint = KERNEL_PID + 1;
447 	TLBINFO_ASID_RESET(ti);
448 
449 	switch (op) {
450 #if defined(MULTIPROCESSOR) && defined(PMAP_TLB_NEED_SHOOTDOWN)
451 	case TLBINV_ALL:
452 		tlb_invalidate_all();
453 		break;
454 	case TLBINV_ALLUSER:
455 		tlb_invalidate_asids(KERNEL_PID + 1, ti->ti_asid_max);
456 		break;
457 #endif /* MULTIPROCESSOR && PMAP_TLB_NEED_SHOOTDOWN */
458 	case TLBINV_NOBODY: {
459 		/*
460 		 * If we are just reclaiming ASIDs in the TLB, let's go find
461 		 * what ASIDs are in use in the TLB.  Since this is a
462 		 * semi-expensive operation, we don't want to do it too often.
463 		 * So if more half of the ASIDs are in use, we don't have
464 		 * enough free ASIDs so invalidate the TLB entries with ASIDs
465 		 * and clear the ASID bitmap.  That will force everyone to
466 		 * allocate a new ASID.
467 		 */
468 #if !defined(MULTIPROCESSOR) || defined(PMAP_TLB_NEED_SHOOTDOWN)
469 		pmap_tlb_asid_check();
470 		const u_int asids_found = tlb_record_asids(
471 		    ti->ti_asid_bitmap._b, ti->ti_asid_max);
472 		pmap_tlb_asid_check();
473 #ifdef DIAGNOSTIC
474 		const u_int asids_count = pmap_tlb_asid_count(ti);
475 		KASSERTMSG(asids_found == asids_count,
476 		    "found %u != count %u", asids_found, asids_count);
477 #endif
478 		if (__predict_false(asids_found >= ti->ti_asid_max / 2)) {
479 			tlb_invalidate_asids(KERNEL_PID + 1, ti->ti_asid_max);
480 #else /* MULTIPROCESSOR && !PMAP_TLB_NEED_SHOOTDOWN */
481 			/*
482 			 * For those systems (PowerPC) that don't require
483 			 * cross cpu TLB shootdowns, we have to invalidate the
484 			 * entire TLB because we can't record the ASIDs in use
485 			 * on the other CPUs.  This is hopefully cheaper than
486 			 * than trying to use an IPI to record all the ASIDs
487 			 * on all the CPUs (which would be a synchronization
488 			 * nightmare).
489 			 */
490 			tlb_invalidate_all();
491 #endif /* MULTIPROCESSOR && !PMAP_TLB_NEED_SHOOTDOWN */
492 			TLBINFO_ASID_RESET(ti);
493 			ti->ti_asids_free = TLBINFO_ASID_INITIAL_FREE(
494 				ti->ti_asid_max);
495 #if !defined(MULTIPROCESSOR) || defined(PMAP_TLB_NEED_SHOOTDOWN)
496 		} else {
497 			ti->ti_asids_free -= asids_found;
498 		}
499 #endif /* !MULTIPROCESSOR || PMAP_TLB_NEED_SHOOTDOWN */
500 		KASSERTMSG(ti->ti_asids_free <= ti->ti_asid_max, "%u",
501 		    ti->ti_asids_free);
502 		break;
503 	}
504 	default:
505 		panic("%s: unexpected op %d", __func__, op);
506 	}
507 
508 	/*
509 	 * Now go through the active ASIDs.  If the ASID is on a processor or
510 	 * we aren't invalidating all ASIDs and the TLB has an entry owned by
511 	 * that ASID, mark it as in use.  Otherwise release the ASID.
512 	 */
513 	struct pmap_asid_info *pai, *next;
514 	for (pai = LIST_FIRST(&ti->ti_pais); pai != NULL; pai = next) {
515 		struct pmap * const pm = PAI_PMAP(pai, ti);
516 		next = LIST_NEXT(pai, pai_link);
517 		KASSERT(pm != pmap_kernel());
518 		KASSERT(pai->pai_asid > KERNEL_PID);
519 #if defined(MULTIPROCESSOR)
520 		if (pmap_tlb_intersecting_onproc_p(pm, ti)) {
521 			if (!TLBINFO_ASID_INUSE_P(ti, pai->pai_asid)) {
522 				TLBINFO_ASID_MARK_USED(ti, pai->pai_asid);
523 				ti->ti_asids_free--;
524 			}
525 			continue;
526 		}
527 #endif /* MULTIPROCESSOR */
528 		if (TLBINFO_ASID_INUSE_P(ti, pai->pai_asid)) {
529 			KASSERT(op == TLBINV_NOBODY);
530 		} else {
531 			pmap_tlb_pai_reset(ti, pai, pm);
532 		}
533 	}
534 #ifdef DIAGNOSTIC
535 	size_t free_count __diagused = ti->ti_asid_max - pmap_tlb_asid_count(ti);
536 	KASSERTMSG(free_count == ti->ti_asids_free,
537 	    "bitmap error: %zu != %u", free_count, ti->ti_asids_free);
538 #endif
539 	UVMHIST_LOG(maphist, " <-- done", 0, 0, 0, 0);
540 }
541 
542 #if defined(MULTIPROCESSOR) && defined(PMAP_TLB_NEED_SHOOTDOWN)
543 #if PMAP_TLB_MAX == 1
544 #error shootdown not required for single TLB systems
545 #endif
546 void
pmap_tlb_shootdown_process(void)547 pmap_tlb_shootdown_process(void)
548 {
549 	struct cpu_info * const ci = curcpu();
550 	struct pmap_tlb_info * const ti = cpu_tlb_info(ci);
551 
552 	UVMHIST_FUNC(__func__);
553 	UVMHIST_CALLED(maphist);
554 
555 	KASSERT(cpu_intr_p());
556 	KASSERTMSG(ci->ci_cpl >= IPL_SCHED, "%s: cpl (%d) < IPL_SCHED (%d)",
557 	    __func__, ci->ci_cpl, IPL_SCHED);
558 
559 	TLBINFO_LOCK(ti);
560 	UVMHIST_LOG(maphist, "ti %#jx", (uintptr_t)ti, 0, 0, 0);
561 
562 	switch (ti->ti_tlbinvop) {
563 	case TLBINV_ONE: {
564 		/*
565 		 * We only need to invalidate one user ASID.
566 		 */
567 		UVMHIST_LOG(maphist, "TLBINV_ONE ti->ti_victim %#jx", (uintptr_t)ti->ti_victim, 0, 0, 0);
568 		struct pmap_asid_info * const pai = PMAP_PAI(ti->ti_victim, ti);
569 		KASSERT(ti->ti_victim != pmap_kernel());
570 		if (pmap_tlb_intersecting_onproc_p(ti->ti_victim, ti)) {
571 			UVMHIST_LOG(maphist, "... onproc asid %jd", pai->pai_asid, 0, 0, 0);
572 			/*
573 			 * The victim is an active pmap so we will just
574 			 * invalidate its TLB entries.
575 			 */
576 			KASSERT(pai->pai_asid > KERNEL_PID);
577 			pmap_tlb_asid_check();
578 			tlb_invalidate_asids(pai->pai_asid, pai->pai_asid);
579 			pmap_tlb_asid_check();
580 		} else if (pai->pai_asid) {
581 			UVMHIST_LOG(maphist, "... not active asid %jd", pai->pai_asid, 0, 0, 0);
582 			/*
583 			 * The victim is no longer an active pmap for this TLB.
584 			 * So simply clear its ASID and when pmap_activate is
585 			 * next called for this pmap, it will allocate a new
586 			 * ASID.
587 			 */
588 			pmap_tlb_pai_reset(ti, pai, PAI_PMAP(pai, ti));
589 		}
590 		break;
591 	}
592 	case TLBINV_ALLUSER:
593 		/*
594 		 * Flush all user TLB entries.
595 		 */
596 		pmap_tlb_asid_reinitialize(ti, TLBINV_ALLUSER);
597 		break;
598 	case TLBINV_ALLKERNEL:
599 		/*
600 		 * We need to invalidate all global TLB entries.
601 		 */
602 		pmap_tlb_asid_check();
603 		tlb_invalidate_globals();
604 		pmap_tlb_asid_check();
605 		break;
606 	case TLBINV_ALL:
607 		/*
608 		 * Flush all the TLB entries (user and kernel).
609 		 */
610 		pmap_tlb_asid_reinitialize(ti, TLBINV_ALL);
611 		break;
612 	case TLBINV_NOBODY:
613 		/*
614 		 * Might be spurious or another SMT CPU sharing this TLB
615 		 * could have already done the work.
616 		 */
617 		break;
618 	}
619 
620 	/*
621 	 * Indicate we are done with shutdown event.
622 	 */
623 	ti->ti_victim = NULL;
624 	ti->ti_tlbinvop = TLBINV_NOBODY;
625 	TLBINFO_UNLOCK(ti);
626 }
627 
628 /*
629  * This state machine could be encoded into an array of integers but since all
630  * the values fit in 3 bits, the 5 entry "table" fits in a 16 bit value which
631  * can be loaded in a single instruction.
632  */
633 #define	TLBINV_MAP(op, nobody, one, alluser, allkernel, all)	\
634 	((((   (nobody) << 3 * TLBINV_NOBODY)			\
635 	 | (      (one) << 3 * TLBINV_ONE)			\
636 	 | (  (alluser) << 3 * TLBINV_ALLUSER)			\
637 	 | ((allkernel) << 3 * TLBINV_ALLKERNEL)		\
638 	 | (      (all) << 3 * TLBINV_ALL)) >> 3 * (op)) & 7)
639 
640 #define	TLBINV_USER_MAP(op)	\
641 	TLBINV_MAP(op, TLBINV_ONE, TLBINV_ALLUSER, TLBINV_ALLUSER,	\
642 	    TLBINV_ALL, TLBINV_ALL)
643 
644 #define	TLBINV_KERNEL_MAP(op)	\
645 	TLBINV_MAP(op, TLBINV_ALLKERNEL, TLBINV_ALL, TLBINV_ALL,	\
646 	    TLBINV_ALLKERNEL, TLBINV_ALL)
647 
648 bool
pmap_tlb_shootdown_bystanders(pmap_t pm)649 pmap_tlb_shootdown_bystanders(pmap_t pm)
650 {
651 	/*
652 	 * We don't need to deal with our own TLB.
653 	 */
654 
655 	UVMHIST_FUNC(__func__);
656 	UVMHIST_CALLARGS(maphist, "pm %#jx", (uintptr_t)pm, 0, 0, 0);
657 
658 	KASSERT(kpreempt_disabled());
659 
660 	const struct cpu_info * const ci = curcpu();
661 
662 	kcpuset_t *pm_active = ci->ci_shootdowncpus;
663 	kcpuset_copy(pm_active, pm->pm_active);
664 	kcpuset_remove(pm_active, cpu_tlb_info(curcpu())->ti_kcpuset);
665 	const bool kernel_p = (pm == pmap_kernel());
666 	bool ipi_sent = false;
667 
668 	/*
669 	 * If pm_active gets more bits set, then it's after all our changes
670 	 * have been made so they will already be cognizant of them.
671 	 */
672 
673 	for (size_t i = 0; !kcpuset_iszero(pm_active); i++) {
674 		KASSERT(i < pmap_ntlbs);
675 		struct pmap_tlb_info * const ti = pmap_tlbs[i];
676 		KASSERT(tlbinfo_index(ti) == i);
677 		UVMHIST_LOG(maphist, "ti %#jx", (uintptr_t)ti, 0, 0, 0);
678 		/*
679 		 * Skip this TLB if there are no active mappings for it.
680 		 */
681 		if (!kcpuset_intersecting_p(pm_active, ti->ti_kcpuset))
682 			continue;
683 		struct pmap_asid_info * const pai = PMAP_PAI(pm, ti);
684 		kcpuset_remove(pm_active, ti->ti_kcpuset);
685 		TLBINFO_LOCK(ti);
686 		cpuid_t j = kcpuset_ffs_intersecting(pm->pm_onproc,
687 		    ti->ti_kcpuset);
688 		// post decrement since ffs returns bit + 1 or 0 if no bit
689 		if (j-- > 0) {
690 			if (kernel_p) {
691 				ti->ti_tlbinvop =
692 				    TLBINV_KERNEL_MAP(ti->ti_tlbinvop);
693 				ti->ti_victim = NULL;
694 			} else {
695 				KASSERT(pai->pai_asid);
696 				if (__predict_false(ti->ti_victim == pm)) {
697 					KASSERT(ti->ti_tlbinvop == TLBINV_ONE);
698 					/*
699 					 * We still need to invalidate this one
700 					 * ASID so there's nothing to change.
701 					 */
702 				} else {
703 					ti->ti_tlbinvop =
704 					    TLBINV_USER_MAP(ti->ti_tlbinvop);
705 					if (ti->ti_tlbinvop == TLBINV_ONE)
706 						ti->ti_victim = pm;
707 					else
708 						ti->ti_victim = NULL;
709 				}
710 			}
711 			UVMHIST_LOG(maphist, "tlbinvop %jx victim %#jx", ti->ti_tlbinvop,
712 			    (uintptr_t)ti->ti_victim, 0, 0);
713 			TLBINFO_UNLOCK(ti);
714 			/*
715 			 * Now we can send out the shootdown IPIs to a CPU
716 			 * that shares this TLB and is currently using this
717 			 * pmap.  That CPU will process the IPI and do the
718 			 * all the work.  Any other CPUs sharing that TLB
719 			 * will take advantage of that work.  pm_onproc might
720 			 * change now that we have released the lock but we
721 			 * can tolerate spurious shootdowns.
722 			 */
723 			cpu_send_ipi(cpu_lookup(j), IPI_SHOOTDOWN);
724 			ipi_sent = true;
725 			continue;
726 		}
727 		if (!pmap_tlb_intersecting_active_p(pm, ti)) {
728 			UVMHIST_LOG(maphist, "pm %#jx not active", (uintptr_t)pm, 0, 0, 0);
729 			/*
730 			 * If this pmap has an ASID assigned but it's not
731 			 * currently running, nuke its ASID.  Next time the
732 			 * pmap is activated, it will allocate a new ASID.
733 			 * And best of all, we avoid an IPI.
734 			 */
735 			KASSERT(!kernel_p);
736 			pmap_tlb_pai_reset(ti, pai, pm);
737 			//ti->ti_evcnt_lazy_shots.ev_count++;
738 		}
739 		TLBINFO_UNLOCK(ti);
740 	}
741 
742 	UVMHIST_LOG(maphist, " <-- done (ipi_sent=%jd)", ipi_sent, 0, 0, 0);
743 
744 	return ipi_sent;
745 }
746 #endif /* MULTIPROCESSOR && PMAP_TLB_NEED_SHOOTDOWN */
747 
748 int
pmap_tlb_update_addr(pmap_t pm,vaddr_t va,pt_entry_t pte,u_int flags)749 pmap_tlb_update_addr(pmap_t pm, vaddr_t va, pt_entry_t pte, u_int flags)
750 {
751 	KASSERT(kpreempt_disabled());
752 
753 	struct pmap_tlb_info * const ti = cpu_tlb_info(curcpu());
754 	struct pmap_asid_info * const pai = PMAP_PAI(pm, ti);
755 	int rv = -1;
756 
757 	UVMHIST_FUNC(__func__);
758 	UVMHIST_CALLARGS(maphist, " (pm=%#jx va=%#jx, pte=%#jx flags=%#jx)",
759 	    (uintptr_t)pm, va, pte_value(pte), flags);
760 
761 	KASSERTMSG(pte_valid_p(pte), "va %#"PRIxVADDR" %#"PRIxPTE,
762 	    va, pte_value(pte));
763 
764 	TLBINFO_LOCK(ti);
765 	if (pm == pmap_kernel() || PMAP_PAI_ASIDVALID_P(pai, ti)) {
766 		pmap_tlb_asid_check();
767 		rv = tlb_update_addr(va, pai->pai_asid, pte,
768 		    (flags & PMAP_TLB_INSERT) != 0);
769 		pmap_tlb_asid_check();
770 		UVMHIST_LOG(maphist,
771 		    "   %jd <-- tlb_update_addr(%#jx, %#jx, %#jx, ...)",
772 		    rv, va, pai->pai_asid, pte_value(pte));
773 		KASSERTMSG((flags & PMAP_TLB_INSERT) == 0 || rv == 1,
774 		    "pmap %p (asid %u) va %#"PRIxVADDR" pte %#"PRIxPTE" rv %d",
775 		    pm, pai->pai_asid, va, pte_value(pte), rv);
776 	}
777 #if defined(MULTIPROCESSOR) && defined(PMAP_TLB_NEED_SHOOTDOWN)
778 	if (flags & PMAP_TLB_NEED_IPI)
779 		pm->pm_shootdown_pending = 1;
780 #endif
781 	TLBINFO_UNLOCK(ti);
782 
783 	UVMHIST_LOG(maphist, "   <-- done (rv=%jd)", rv, 0, 0, 0);
784 
785 	return rv;
786 }
787 
788 void
pmap_tlb_invalidate_addr(pmap_t pm,vaddr_t va)789 pmap_tlb_invalidate_addr(pmap_t pm, vaddr_t va)
790 {
791 	KASSERT(kpreempt_disabled());
792 
793 	struct pmap_tlb_info * const ti = cpu_tlb_info(curcpu());
794 	struct pmap_asid_info * const pai = PMAP_PAI(pm, ti);
795 
796 	UVMHIST_FUNC(__func__);
797 	UVMHIST_CALLARGS(maphist, " (pm=%#jx va=%#jx) ti=%#jx asid=%#jx",
798 	    (uintptr_t)pm, va, (uintptr_t)ti, pai->pai_asid);
799 
800 	TLBINFO_LOCK(ti);
801 	if (pm == pmap_kernel() || PMAP_PAI_ASIDVALID_P(pai, ti)) {
802 		pmap_tlb_asid_check();
803 		UVMHIST_LOG(maphist, " invalidating %#jx asid %#jx",
804 		    va, pai->pai_asid, 0, 0);
805 		tlb_invalidate_addr(va, pai->pai_asid);
806 		pmap_tlb_asid_check();
807 	}
808 #if defined(MULTIPROCESSOR) && defined(PMAP_TLB_NEED_SHOOTDOWN)
809 	pm->pm_shootdown_pending = 1;
810 #endif
811 	TLBINFO_UNLOCK(ti);
812 	UVMHIST_LOG(maphist, " <-- done", 0, 0, 0, 0);
813 }
814 
815 static inline void
pmap_tlb_asid_alloc(struct pmap_tlb_info * ti,pmap_t pm,struct pmap_asid_info * pai)816 pmap_tlb_asid_alloc(struct pmap_tlb_info *ti, pmap_t pm,
817 	struct pmap_asid_info *pai)
818 {
819 	/*
820 	 * We shouldn't have an ASID assigned, and thusly must not be onproc
821 	 * nor active.
822 	 */
823 	KASSERT(pm != pmap_kernel());
824 	KASSERT(pai->pai_asid == 0);
825 	KASSERT(pai->pai_link.le_prev == NULL);
826 #if defined(MULTIPROCESSOR)
827 	KASSERT(!pmap_tlb_intersecting_onproc_p(pm, ti));
828 	KASSERT(!pmap_tlb_intersecting_active_p(pm, ti));
829 #endif
830 	KASSERT(ti->ti_asids_free > 0);
831 	KASSERT(ti->ti_asid_hint > KERNEL_PID);
832 
833 	/*
834 	 * If the last ASID allocated was the maximum ASID, then the
835 	 * hint will be out of range.  Reset the hint to first
836 	 * available ASID.
837 	 */
838 	if (PMAP_TLB_FLUSH_ASID_ON_RESET
839 	    && ti->ti_asid_hint > ti->ti_asid_max) {
840 		ti->ti_asid_hint = KERNEL_PID + 1;
841 	}
842 	KASSERTMSG(ti->ti_asid_hint <= ti->ti_asid_max, "hint %u",
843 	    ti->ti_asid_hint);
844 
845 	/*
846 	 * Let's see if the hinted ASID is free.  If not search for
847 	 * a new one.
848 	 */
849 	if (__predict_true(TLBINFO_ASID_INUSE_P(ti, ti->ti_asid_hint))) {
850 		const size_t nbpw = NBBY * sizeof(ti->ti_asid_bitmap._b[0]);
851 		size_t i;
852 		u_long bits;
853 		for (i = 0; (bits = ~ti->ti_asid_bitmap._b[i]) == 0; i++) {
854 			KASSERT(i < __arraycount(ti->ti_asid_bitmap._b) - 1);
855 		}
856 		/*
857 		 * ffs wants to find the first bit set while we want
858 		 * to find the first bit cleared.
859 		 */
860 		const u_int n = __builtin_ffsl(bits) - 1;
861 		KASSERTMSG((bits << (nbpw - (n+1))) == (1ul << (nbpw-1)),
862 		    "n %u bits %#lx", n, bits);
863 		KASSERT(n < nbpw);
864 		ti->ti_asid_hint = n + i * nbpw;
865 	}
866 
867 	KASSERT(ti->ti_asid_hint > KERNEL_PID);
868 	KASSERT(ti->ti_asid_hint <= ti->ti_asid_max);
869 	KASSERTMSG(PMAP_TLB_FLUSH_ASID_ON_RESET
870 	    || TLBINFO_ASID_INUSE_P(ti, ti->ti_asid_hint - 1),
871 	    "hint %u bitmap %p", ti->ti_asid_hint, &ti->ti_asid_bitmap);
872 	KASSERTMSG(!TLBINFO_ASID_INUSE_P(ti, ti->ti_asid_hint),
873 	    "hint %u bitmap %p", ti->ti_asid_hint, &ti->ti_asid_bitmap);
874 
875 	/*
876 	 * The hint contains our next ASID so take it and advance the hint.
877 	 * Mark it as used and insert the pai into the list of active asids.
878 	 * There is also one less asid free in this TLB.
879 	 */
880 	pai->pai_asid = ti->ti_asid_hint++;
881 #ifdef MULTIPROCESSOR
882 	if (PMAP_TLB_FLUSH_ASID_ON_RESET) {
883 		/*
884 		 * Clean the new ASID from the TLB.
885 		 */
886 		tlb_invalidate_asids(pai->pai_asid, pai->pai_asid);
887 	}
888 #endif
889 	TLBINFO_ASID_MARK_USED(ti, pai->pai_asid);
890 	LIST_INSERT_HEAD(&ti->ti_pais, pai, pai_link);
891 	ti->ti_asids_free--;
892 
893 #if defined(MULTIPROCESSOR)
894 	/*
895 	 * Mark that we now have an active ASID for all CPUs sharing this TLB.
896 	 * The bits in pm_active belonging to this TLB can only be changed
897 	 * while this TLBs lock is held.
898 	 */
899 #if PMAP_TLB_MAX == 1
900 	kcpuset_copy(pm->pm_active, kcpuset_running);
901 #else
902 	kcpuset_merge(pm->pm_active, ti->ti_kcpuset);
903 #endif
904 #endif
905 }
906 
907 /*
908  * Acquire a TLB address space tag (called ASID or TLBPID) and return it.
909  * ASID might have already been previously acquired.
910  */
911 void
pmap_tlb_asid_acquire(pmap_t pm,struct lwp * l)912 pmap_tlb_asid_acquire(pmap_t pm, struct lwp *l)
913 {
914 	KASSERT(kpreempt_disabled());
915 
916 	struct cpu_info * const ci = l->l_cpu;
917 	struct pmap_tlb_info * const ti = cpu_tlb_info(ci);
918 	struct pmap_asid_info * const pai = PMAP_PAI(pm, ti);
919 
920 	UVMHIST_FUNC(__func__);
921 	UVMHIST_CALLARGS(maphist, "(pm=%#jx, l=%#jx, ti=%#jx)", (uintptr_t)pm,
922 	    (uintptr_t)l, (uintptr_t)ti, 0);
923 
924 	/*
925 	 * Kernels use a fixed ASID and thus doesn't need to acquire one.
926 	 */
927 	if (pm == pmap_kernel()) {
928 		UVMHIST_LOG(maphist, " <-- done (kernel)", 0, 0, 0, 0);
929 		return;
930 	}
931 
932 	TLBINFO_LOCK(ti);
933 	KASSERT(pai->pai_asid <= KERNEL_PID || pai->pai_link.le_prev != NULL);
934 	KASSERT(pai->pai_asid > KERNEL_PID || pai->pai_link.le_prev == NULL);
935 	pmap_tlb_pai_check(ti, true);
936 
937 	if (__predict_false(!tlbinfo_asids_p(ti))) {
938 #if defined(MULTIPROCESSOR)
939 		/*
940 		 * Mark that we are active for all CPUs sharing this TLB.
941 		 * The bits in pm_active belonging to this TLB can only
942 		 * be changed  while this TLBs lock is held.
943 		 */
944 #if PMAP_TLB_MAX == 1
945 		kcpuset_copy(pm->pm_active, kcpuset_running);
946 #else
947 		kcpuset_merge(pm->pm_active, ti->ti_kcpuset);
948 #endif
949 #endif
950 	} else if (__predict_false(!PMAP_PAI_ASIDVALID_P(pai, ti))) {
951 		/*
952 		 * If we've run out ASIDs, reinitialize the ASID space.
953 		 */
954 		if (__predict_false(tlbinfo_noasids_p(ti))) {
955 			KASSERT(l == curlwp);
956 			UVMHIST_LOG(maphist, " asid reinit", 0, 0, 0, 0);
957 			pmap_tlb_asid_reinitialize(ti, TLBINV_NOBODY);
958 			KASSERT(!tlbinfo_noasids_p(ti));
959 		}
960 
961 		/*
962 		 * Get an ASID.
963 		 */
964 		pmap_tlb_asid_alloc(ti, pm, pai);
965 		UVMHIST_LOG(maphist, "allocated asid %#jx", pai->pai_asid,
966 		    0, 0, 0);
967 	}
968 	pmap_tlb_pai_check(ti, true);
969 #if defined(MULTIPROCESSOR)
970 	KASSERT(kcpuset_isset(pm->pm_active, cpu_index(ci)));
971 #endif
972 
973 	if (l == curlwp) {
974 #if defined(MULTIPROCESSOR)
975 		/*
976 		 * The bits in pm_onproc belonging to this TLB can only
977 		 * be changed while this TLBs lock is held unless atomic
978 		 * operations are used.
979 		 */
980 		KASSERT(pm != pmap_kernel());
981 		kcpuset_atomic_set(pm->pm_onproc, cpu_index(ci));
982 #endif
983 		ci->ci_pmap_asid_cur = pai->pai_asid;
984 		UVMHIST_LOG(maphist, "setting asid to %#jx", pai->pai_asid,
985 		    0, 0, 0);
986 		tlb_set_asid(pai->pai_asid, pm);
987 		pmap_tlb_asid_check();
988 	} else {
989 		printf("%s: l (%p) != curlwp %p\n", __func__, l, curlwp);
990 	}
991 	TLBINFO_UNLOCK(ti);
992 	UVMHIST_LOG(maphist, " <-- done", 0, 0, 0, 0);
993 }
994 
995 void
pmap_tlb_asid_deactivate(pmap_t pm)996 pmap_tlb_asid_deactivate(pmap_t pm)
997 {
998 	UVMHIST_FUNC(__func__);
999 	UVMHIST_CALLARGS(maphist, "pm %#jx", (uintptr_t)pm, 0, 0, 0);
1000 
1001 	KASSERT(kpreempt_disabled());
1002 #if defined(MULTIPROCESSOR)
1003 	/*
1004 	 * The kernel pmap is aways onproc and active and must never have
1005 	 * those bits cleared.  If pmap_remove_all was called, it has already
1006 	 * deactivated the pmap and thusly onproc will be 0 so there's nothing
1007 	 * to do.
1008 	 */
1009 	if (pm != pmap_kernel() && !kcpuset_iszero(pm->pm_onproc)) {
1010 		struct cpu_info * const ci = curcpu();
1011 		KASSERT(!cpu_intr_p());
1012 		KASSERTMSG(kcpuset_isset(pm->pm_onproc, cpu_index(ci)),
1013 		    "%s: pmap %p onproc %p doesn't include cpu %d (%p)",
1014 		    __func__, pm, pm->pm_onproc, cpu_index(ci), ci);
1015 		/*
1016 		 * The bits in pm_onproc that belong to this TLB can
1017 		 * be changed while this TLBs lock is not held as long
1018 		 * as we use atomic ops.
1019 		 */
1020 		kcpuset_atomic_clear(pm->pm_onproc, cpu_index(ci));
1021 	}
1022 #endif
1023 	curcpu()->ci_pmap_asid_cur = KERNEL_PID;
1024 	tlb_set_asid(KERNEL_PID, pmap_kernel());
1025 
1026 	pmap_tlb_pai_check(cpu_tlb_info(curcpu()), false);
1027 #if defined(DEBUG)
1028 	pmap_tlb_asid_check();
1029 #endif
1030 	UVMHIST_LOG(maphist, " <-- done (pm=%#jx)", (uintptr_t)pm, 0, 0, 0);
1031 }
1032 
1033 void
pmap_tlb_asid_release_all(struct pmap * pm)1034 pmap_tlb_asid_release_all(struct pmap *pm)
1035 {
1036 	UVMHIST_FUNC(__func__);
1037 	UVMHIST_CALLARGS(maphist, "(pm=%#jx)", (uintptr_t)pm, 0, 0, 0);
1038 
1039 	KASSERT(pm != pmap_kernel());
1040 #if defined(MULTIPROCESSOR)
1041 	//KASSERT(!kcpuset_iszero(pm->pm_onproc)); // XXX
1042 	struct cpu_info * const ci __diagused = curcpu();
1043 	KASSERT(!kcpuset_isotherset(pm->pm_onproc, cpu_index(ci)));
1044 #if PMAP_TLB_MAX > 1
1045 	for (u_int i = 0; !kcpuset_iszero(pm->pm_active); i++) {
1046 		KASSERT(i < pmap_ntlbs);
1047 		struct pmap_tlb_info * const ti = pmap_tlbs[i];
1048 #else
1049 		struct pmap_tlb_info * const ti = &pmap_tlb0_info;
1050 #endif
1051 		struct pmap_asid_info * const pai = PMAP_PAI(pm, ti);
1052 		TLBINFO_LOCK(ti);
1053 		if (PMAP_PAI_ASIDVALID_P(pai, ti)) {
1054 			/*
1055 			 * This pmap should not be in use by any other cpu so
1056 			 * we can just reset and be happy.
1057 			 */
1058 			if (ti->ti_victim == pm)
1059 				ti->ti_victim = NULL;
1060 			if (__predict_true(tlbinfo_asids_p(ti)))
1061 				pmap_tlb_pai_reset(ti, pai, pm);
1062 		}
1063 		KASSERT(pai->pai_link.le_prev == NULL);
1064 		TLBINFO_UNLOCK(ti);
1065 #if PMAP_TLB_MAX > 1
1066 	}
1067 #endif
1068 #ifdef DIAGNOSTIC
1069 	for (size_t i = 0; i < (PMAP_TLB_MAX > 1 ? pmap_ntlbs : 1); i++) {
1070 		KASSERTMSG(pm->pm_pai[i].pai_asid == 0,
1071 		    "pm %p i %zu asid %u",
1072 		    pm, i, pm->pm_pai[i].pai_asid);
1073 	}
1074 #endif
1075 #else
1076 	/*
1077 	 * Handle the case of an UP kernel which only has, at most, one TLB.
1078 	 * If the pmap has an ASID allocated, free it.
1079 	 */
1080 	struct pmap_tlb_info * const ti = &pmap_tlb0_info;
1081 	struct pmap_asid_info * const pai = PMAP_PAI(pm, ti);
1082 	TLBINFO_LOCK(ti);
1083 	if (pai->pai_asid > KERNEL_PID) {
1084 		if (curcpu()->ci_pmap_asid_cur == pai->pai_asid) {
1085 			tlb_invalidate_asids(pai->pai_asid, pai->pai_asid);
1086 		} else {
1087 			pmap_tlb_pai_reset(ti, pai, pm);
1088 		}
1089 	}
1090 	TLBINFO_UNLOCK(ti);
1091 #endif /* MULTIPROCESSOR */
1092 	UVMHIST_LOG(maphist, " <-- done", 0, 0, 0, 0);
1093 }
1094 
1095 void
pmap_tlb_asid_check(void)1096 pmap_tlb_asid_check(void)
1097 {
1098 	UVMHIST_FUNC(__func__);
1099 	UVMHIST_CALLED(maphist);
1100 
1101 #ifdef DEBUG
1102 	kpreempt_disable();
1103 	const tlb_asid_t asid __debugused = tlb_get_asid();
1104 	UVMHIST_LOG(maphist, " asid %u vs pmap_cur_asid %u", asid,
1105 	    curcpu()->ci_pmap_asid_cur, 0, 0);
1106 	KDASSERTMSG(asid == curcpu()->ci_pmap_asid_cur,
1107 	   "%s: asid (%#x) != current asid (%#x)",
1108 	    __func__, asid, curcpu()->ci_pmap_asid_cur);
1109 	kpreempt_enable();
1110 #endif
1111 	UVMHIST_LOG(maphist, " <-- done", 0, 0, 0, 0);
1112 }
1113 
1114 #ifdef DEBUG
1115 void
pmap_tlb_check(pmap_t pm,bool (* func)(void *,vaddr_t,tlb_asid_t,pt_entry_t))1116 pmap_tlb_check(pmap_t pm, bool (*func)(void *, vaddr_t, tlb_asid_t, pt_entry_t))
1117 {
1118 	struct pmap_tlb_info * const ti = cpu_tlb_info(curcpu());
1119 	struct pmap_asid_info * const pai = PMAP_PAI(pm, ti);
1120 	TLBINFO_LOCK(ti);
1121 	if (pm == pmap_kernel() || pai->pai_asid > KERNEL_PID)
1122 		tlb_walk(pm, func);
1123 	TLBINFO_UNLOCK(ti);
1124 }
1125 #endif /* DEBUG */
1126 
1127 #ifdef DDB
1128 void
1129 pmap_db_tlb_print(struct pmap *pm,
1130     void (*pr)(const char *, ...) __printflike(1, 2))
1131 {
1132 #if !defined(MULTIPROCESSOR) || PMAP_TLB_MAX == 1
1133 	pr(" asid %5u\n", pm->pm_pai[0].pai_asid);
1134 #else
1135 	for (size_t i = 0; i < (PMAP_TLB_MAX > 1 ? pmap_ntlbs : 1); i++) {
1136 		pr(" tlb %zu  asid %5u\n", i, pm->pm_pai[i].pai_asid);
1137 	}
1138 #endif
1139 }
1140 #endif /* DDB */
1141