xref: /netbsd-src/sys/uvm/pmap/pmap_segtab.c (revision 946379e7b37692fc43f68eb0d1c10daa0a7f3b6c)
1 /*	$NetBSD: pmap_segtab.c,v 1.2 2015/06/11 08:04:44 matt Exp $	*/
2 
3 /*-
4  * Copyright (c) 1998, 2001 The NetBSD Foundation, Inc.
5  * All rights reserved.
6  *
7  * This code is derived from software contributed to The NetBSD Foundation
8  * by Jason R. Thorpe of the Numerical Aerospace Simulation Facility,
9  * NASA Ames Research Center and by Chris G. Demetriou.
10  *
11  * Redistribution and use in source and binary forms, with or without
12  * modification, are permitted provided that the following conditions
13  * are met:
14  * 1. Redistributions of source code must retain the above copyright
15  *    notice, this list of conditions and the following disclaimer.
16  * 2. Redistributions in binary form must reproduce the above copyright
17  *    notice, this list of conditions and the following disclaimer in the
18  *    documentation and/or other materials provided with the distribution.
19  *
20  * THIS SOFTWARE IS PROVIDED BY THE NETBSD FOUNDATION, INC. AND CONTRIBUTORS
21  * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED
22  * TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
23  * PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE FOUNDATION OR CONTRIBUTORS
24  * BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
25  * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
26  * SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
27  * INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
28  * CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
29  * ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
30  * POSSIBILITY OF SUCH DAMAGE.
31  */
32 
33 /*
34  * Copyright (c) 1992, 1993
35  *	The Regents of the University of California.  All rights reserved.
36  *
37  * This code is derived from software contributed to Berkeley by
38  * the Systems Programming Group of the University of Utah Computer
39  * Science Department and Ralph Campbell.
40  *
41  * Redistribution and use in source and binary forms, with or without
42  * modification, are permitted provided that the following conditions
43  * are met:
44  * 1. Redistributions of source code must retain the above copyright
45  *    notice, this list of conditions and the following disclaimer.
46  * 2. Redistributions in binary form must reproduce the above copyright
47  *    notice, this list of conditions and the following disclaimer in the
48  *    documentation and/or other materials provided with the distribution.
49  * 3. Neither the name of the University nor the names of its contributors
50  *    may be used to endorse or promote products derived from this software
51  *    without specific prior written permission.
52  *
53  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
54  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
55  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
56  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
57  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
58  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
59  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
60  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
61  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
62  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
63  * SUCH DAMAGE.
64  *
65  *	@(#)pmap.c	8.4 (Berkeley) 1/26/94
66  */
67 
68 #include <sys/cdefs.h>
69 
70 __KERNEL_RCSID(0, "$NetBSD: pmap_segtab.c,v 1.2 2015/06/11 08:04:44 matt Exp $");
71 
72 /*
73  *	Manages physical address maps.
74  *
75  *	In addition to hardware address maps, this
76  *	module is called upon to provide software-use-only
77  *	maps which may or may not be stored in the same
78  *	form as hardware maps.  These pseudo-maps are
79  *	used to store intermediate results from copy
80  *	operations to and from address spaces.
81  *
82  *	Since the information managed by this module is
83  *	also stored by the logical address mapping module,
84  *	this module may throw away valid virtual-to-physical
85  *	mappings at almost any time.  However, invalidations
86  *	of virtual-to-physical mappings must be done as
87  *	requested.
88  *
89  *	In order to cope with hardware architectures which
90  *	make virtual-to-physical map invalidates expensive,
91  *	this module may delay invalidate or reduced protection
92  *	operations until such time as they are actually
93  *	necessary.  This module is given full information as
94  *	to which processors are currently using which maps,
95  *	and to when physical maps must be made correct.
96  */
97 
98 #define __PMAP_PRIVATE
99 
100 #include "opt_multiprocessor.h"
101 
102 #include <sys/param.h>
103 #include <sys/systm.h>
104 #include <sys/proc.h>
105 #include <sys/mutex.h>
106 #include <sys/atomic.h>
107 
108 #include <uvm/uvm.h>
109 
110 CTASSERT(NBPG >= sizeof(pmap_segtab_t));
111 
112 struct pmap_segtab_info {
113 	pmap_segtab_t *free_segtab;	/* free list kept locally */
114 #ifdef DEBUG
115 	uint32_t nget_segtab;
116 	uint32_t nput_segtab;
117 	uint32_t npage_segtab;
118 #define	SEGTAB_ADD(n, v)	(pmap_segtab_info.n ## _segtab += (v))
119 #else
120 #define	SEGTAB_ADD(n, v)	((void) 0)
121 #endif
122 #ifdef PMAP_PTP_CACHE
123 	struct pgflist ptp_pgflist;	/* Keep a list of idle page tables. */
124 #endif
125 } pmap_segtab_info = {
126 #ifdef PMAP_PTP_CACHE
127 	.ptp_pgflist = LIST_HEAD_INITIALIZER(pmap_segtab_info.ptp_pgflist),
128 #endif
129 };
130 
131 kmutex_t pmap_segtab_lock __cacheline_aligned;
132 
133 static inline struct vm_page *
134 pmap_pte_pagealloc(void)
135 {
136 	struct vm_page *pg;
137 
138 	pg = PMAP_ALLOC_POOLPAGE(UVM_PGA_ZERO|UVM_PGA_USERESERVE);
139 	if (pg) {
140 #ifdef UVM_PAGE_TRKOWN
141 		pg->owner_tag = NULL;
142 #endif
143 		UVM_PAGE_OWN(pg, "pmap-ptp");
144 	}
145 
146 	return pg;
147 }
148 
149 static inline pt_entry_t *
150 pmap_segmap(struct pmap *pmap, vaddr_t va)
151 {
152 	pmap_segtab_t *stp = pmap->pm_segtab;
153 	KASSERT(pmap != pmap_kernel() || !pmap_md_direct_mapped_vaddr_p(va));
154 #ifdef _LP64
155 	stp = stp->seg_seg[(va >> XSEGSHIFT) & (NSEGPG - 1)];
156 	if (stp == NULL)
157 		return NULL;
158 #endif
159 
160 	return stp->seg_tab[(va >> SEGSHIFT) & (PMAP_SEGTABSIZE - 1)];
161 }
162 
163 pt_entry_t *
164 pmap_pte_lookup(pmap_t pmap, vaddr_t va)
165 {
166 	pt_entry_t *pte = pmap_segmap(pmap, va);
167 	if (pte == NULL)
168 		return NULL;
169 
170 	return pte + ((va >> PGSHIFT) & (NPTEPG - 1));
171 }
172 
173 static void
174 pmap_segtab_free(pmap_segtab_t *stp)
175 {
176 	/*
177 	 * Insert the the segtab into the segtab freelist.
178 	 */
179 	mutex_spin_enter(&pmap_segtab_lock);
180 	stp->seg_seg[0] = pmap_segtab_info.free_segtab;
181 	pmap_segtab_info.free_segtab = stp;
182 	SEGTAB_ADD(nput, 1);
183 	mutex_spin_exit(&pmap_segtab_lock);
184 }
185 
186 static void
187 pmap_segtab_release(pmap_t pmap, pmap_segtab_t **stp_p, bool free_stp,
188 	pte_callback_t callback, uintptr_t flags,
189 	vaddr_t va, vsize_t vinc)
190 {
191 	pmap_segtab_t *stp = *stp_p;
192 
193 	for (size_t i = va / vinc; i < PMAP_SEGTABSIZE; i++, va += vinc) {
194 #ifdef _LP64
195 		if (vinc > NBSEG) {
196 			if (stp->seg_seg[i] != NULL) {
197 				pmap_segtab_release(pmap, &stp->seg_seg[i],
198 				    true, callback, flags, va, vinc / NSEGPG);
199 				KASSERT(stp->seg_seg[i] == NULL);
200 			}
201 			continue;
202 		}
203 #endif
204 		KASSERT(vinc == NBSEG);
205 
206 		/* get pointer to segment map */
207 		pt_entry_t *pte = stp->seg_tab[i];
208 		if (pte == NULL)
209 			continue;
210 
211 		/*
212 		 * If our caller want a callback, do so.
213 		 */
214 		if (callback != NULL) {
215 			(*callback)(pmap, va, va + vinc, pte, flags);
216 		}
217 #ifdef DEBUG
218 		for (size_t j = 0; j < NPTEPG; j++) {
219 			if (!pte_zero_p(pte[j]))
220 				panic("%s: pte entry %p not 0 (%#"PRIxPTE")",
221 				    __func__, &pte[j], pte_value(pte[j]));
222 		}
223 #endif
224 		// PMAP_UNMAP_POOLPAGE should handle any VCA issues itself
225 		paddr_t pa = PMAP_UNMAP_POOLPAGE((vaddr_t)pte);
226 		struct vm_page *pg = PHYS_TO_VM_PAGE(pa);
227 #ifdef PMAP_PTP_CACHE
228 		mutex_spin_enter(&pmap_segtab_lock);
229 		LIST_INSERT_HEAD(&pmap_segtab_info.ptp_pgflist, pg, listq.list);
230 		mutex_spin_exit(&pmap_segtab_lock);
231 #else
232 		uvm_pagefree(pg);
233 #endif
234 
235 		stp->seg_tab[i] = NULL;
236 	}
237 
238 	if (free_stp) {
239 		pmap_segtab_free(stp);
240 		*stp_p = NULL;
241 	}
242 }
243 
244 /*
245  *	Create and return a physical map.
246  *
247  *	If the size specified for the map
248  *	is zero, the map is an actual physical
249  *	map, and may be referenced by the
250  *	hardware.
251  *
252  *	If the size specified is non-zero,
253  *	the map will be used in software only, and
254  *	is bounded by that size.
255  */
256 static pmap_segtab_t *
257 pmap_segtab_alloc(void)
258 {
259 	pmap_segtab_t *stp;
260 
261  again:
262 	mutex_spin_enter(&pmap_segtab_lock);
263 	if (__predict_true((stp = pmap_segtab_info.free_segtab) != NULL)) {
264 		pmap_segtab_info.free_segtab = stp->seg_seg[0];
265 		stp->seg_seg[0] = NULL;
266 		SEGTAB_ADD(nget, 1);
267 	}
268 	mutex_spin_exit(&pmap_segtab_lock);
269 
270 	if (__predict_false(stp == NULL)) {
271 		struct vm_page * const stp_pg = pmap_pte_pagealloc();
272 
273 		if (__predict_false(stp_pg == NULL)) {
274 			/*
275 			 * XXX What else can we do?  Could we deadlock here?
276 			 */
277 			uvm_wait("pmap_create");
278 			goto again;
279 		}
280 		SEGTAB_ADD(npage, 1);
281 		const paddr_t stp_pa = VM_PAGE_TO_PHYS(stp_pg);
282 
283 		stp = (pmap_segtab_t *)PMAP_MAP_POOLPAGE(stp_pa);
284 		const size_t n = NBPG / sizeof(*stp);
285 		if (n > 1) {
286 			/*
287 			 * link all the segtabs in this page together
288 			 */
289 			for (size_t i = 1; i < n - 1; i++) {
290 				stp[i].seg_seg[0] = &stp[i+1];
291 			}
292 			/*
293 			 * Now link the new segtabs into the free segtab list.
294 			 */
295 			mutex_spin_enter(&pmap_segtab_lock);
296 			stp[n-1].seg_seg[0] = pmap_segtab_info.free_segtab;
297 			pmap_segtab_info.free_segtab = stp + 1;
298 			SEGTAB_ADD(nput, n - 1);
299 			mutex_spin_exit(&pmap_segtab_lock);
300 		}
301 	}
302 
303 #ifdef DEBUG
304 	for (size_t i = 0; i < PMAP_SEGTABSIZE; i++) {
305 		if (stp->seg_tab[i] != 0)
306 			panic("%s: pm_segtab.seg_tab[%zu] != 0", __func__, i);
307 	}
308 #endif
309 	return stp;
310 }
311 
312 /*
313  * Allocate the top segment table for the pmap.
314  */
315 void
316 pmap_segtab_init(pmap_t pmap)
317 {
318 
319 	pmap->pm_segtab = pmap_segtab_alloc();
320 }
321 
322 /*
323  *	Retire the given physical map from service.
324  *	Should only be called if the map contains
325  *	no valid mappings.
326  */
327 void
328 pmap_segtab_destroy(pmap_t pmap, pte_callback_t func, uintptr_t flags)
329 {
330 	if (pmap->pm_segtab == NULL)
331 		return;
332 
333 #ifdef _LP64
334 	const vsize_t vinc = NBXSEG;
335 #else
336 	const vsize_t vinc = NBSEG;
337 #endif
338 	pmap_segtab_release(pmap, &pmap->pm_segtab,
339 	    func == NULL, func, flags, pmap->pm_minaddr, vinc);
340 }
341 
342 /*
343  *	Make a new pmap (vmspace) active for the given process.
344  */
345 void
346 pmap_segtab_activate(struct pmap *pm, struct lwp *l)
347 {
348 	if (l == curlwp) {
349 		KASSERT(pm == l->l_proc->p_vmspace->vm_map.pmap);
350 		if (pm == pmap_kernel()) {
351 			l->l_cpu->ci_pmap_user_segtab = (void*)0xdeadbabe;
352 #ifdef _LP64
353 			l->l_cpu->ci_pmap_user_seg0tab = (void*)0xdeadbabe;
354 #endif
355 		} else {
356 			l->l_cpu->ci_pmap_user_segtab = pm->pm_segtab;
357 #ifdef _LP64
358 			l->l_cpu->ci_pmap_user_seg0tab = pm->pm_segtab->seg_seg[0];
359 #endif
360 		}
361 	}
362 }
363 
364 /*
365  *	Act on the given range of addresses from the specified map.
366  *
367  *	It is assumed that the start and end are properly rounded to
368  *	the page size.
369  */
370 void
371 pmap_pte_process(pmap_t pmap, vaddr_t sva, vaddr_t eva,
372 	pte_callback_t callback, uintptr_t flags)
373 {
374 #if 0
375 	printf("%s: %p, %"PRIxVADDR", %"PRIxVADDR", %p, %"PRIxPTR"\n",
376 	    __func__, pmap, sva, eva, callback, flags);
377 #endif
378 	while (sva < eva) {
379 		vaddr_t lastseg_va = pmap_trunc_seg(sva) + NBSEG;
380 		if (lastseg_va == 0 || lastseg_va > eva)
381 			lastseg_va = eva;
382 
383 		/*
384 		 * If VA belongs to an unallocated segment,
385 		 * skip to the next segment boundary.
386 		 */
387 		pt_entry_t * const pte = pmap_pte_lookup(pmap, sva);
388 		if (pte != NULL) {
389 			/*
390 			 * Callback to deal with the ptes for this segment.
391 			 */
392 			(*callback)(pmap, sva, lastseg_va, pte, flags);
393 		}
394 		/*
395 		 * In theory we could release pages with no entries,
396 		 * but that takes more effort than we want here.
397 		 */
398 		sva = lastseg_va;
399 	}
400 }
401 
402 /*
403  *	Return a pointer for the pte that corresponds to the specified virtual
404  *	address (va) in the target physical map, allocating if needed.
405  */
406 pt_entry_t *
407 pmap_pte_reserve(pmap_t pmap, vaddr_t va, int flags)
408 {
409 	pmap_segtab_t *stp = pmap->pm_segtab;
410 	pt_entry_t *pte;
411 
412 	pte = pmap_pte_lookup(pmap, va);
413 	if (__predict_false(pte == NULL)) {
414 #ifdef _LP64
415 		pmap_segtab_t ** const stp_p =
416 		    &stp->seg_seg[(va >> XSEGSHIFT) & (NSEGPG - 1)];
417 		if (__predict_false((stp = *stp_p) == NULL)) {
418 			pmap_segtab_t *nstp = pmap_segtab_alloc();
419 #ifdef MULTIPROCESSOR
420 			pmap_segtab_t *ostp = atomic_cas_ptr(stp_p, NULL, nstp);
421 			if (__predict_false(ostp != NULL)) {
422 				pmap_segtab_free(nstp);
423 				nstp = ostp;
424 			}
425 #else
426 			*stp_p = nstp;
427 #endif /* MULTIPROCESSOR */
428 			stp = nstp;
429 		}
430 		KASSERT(stp == pmap->pm_segtab->seg_seg[(va >> XSEGSHIFT) & (NSEGPG - 1)]);
431 #endif /* _LP64 */
432 		struct vm_page *pg = NULL;
433 #ifdef PMAP_PTP_CACHE
434 		mutex_spin_enter(&pmap_segtab_lock);
435 		if ((pg = LIST_FIRST(&pmap_segtab_info.ptp_pgflist)) != NULL) {
436 			LIST_REMOVE(pg, listq.list);
437 			KASSERT(LIST_FIRST(&pmap_segtab_info.ptp_pgflist) != pg);
438 		}
439 		mutex_spin_exit(&pmap_segtab_lock);
440 #endif
441 		if (pg == NULL)
442 			pg = pmap_pte_pagealloc();
443 		if (pg == NULL) {
444 			if (flags & PMAP_CANFAIL)
445 				return NULL;
446 			panic("%s: cannot allocate page table page "
447 			    "for va %" PRIxVADDR, __func__, va);
448 		}
449 
450 		const paddr_t pa = VM_PAGE_TO_PHYS(pg);
451 		pte = (pt_entry_t *)PMAP_MAP_POOLPAGE(pa);
452 		pt_entry_t ** const pte_p =
453 		    &stp->seg_tab[(va >> SEGSHIFT) & (PMAP_SEGTABSIZE - 1)];
454 #ifdef MULTIPROCESSOR
455 		pt_entry_t *opte = atomic_cas_ptr(pte_p, NULL, pte);
456 		/*
457 		 * If another thread allocated the segtab needed for this va
458 		 * free the page we just allocated.
459 		 */
460 		if (__predict_false(opte != NULL)) {
461 #ifdef PMAP_PTP_CACHE
462 			mutex_spin_enter(&pmap_segtab_lock);
463 			LIST_INSERT_HEAD(&pmap_segtab_info.ptp_pgflist,
464 			    pg, listq.list);
465 			mutex_spin_exit(&pmap_segtab_lock);
466 #else
467 			PMAP_UNMAP_POOLPAGE((vaddr_t)pte);
468 			uvm_pagefree(pg);
469 #endif
470 			pte = opte;
471 		}
472 #else
473 		*pte_p = pte;
474 #endif
475 		KASSERT(pte == stp->seg_tab[(va >> SEGSHIFT) & (PMAP_SEGTABSIZE - 1)]);
476 
477 #ifdef DEBUG
478 		for (size_t i = 0; i < NPTEPG; i++) {
479 			if (!pte_zero_p(pte[i]))
480 				panic("%s: new segmap %p not empty @ %zu",
481 				    __func__, pte, i);
482 		}
483 #endif
484 		pte += (va >> PGSHIFT) & (NPTEPG - 1);
485 	}
486 
487 	return pte;
488 }
489