xref: /netbsd-src/sys/uvm/pmap/pmap_segtab.c (revision e89934bbf778a6d6d6894877c4da59d0c7835b0f)
1 /*	$NetBSD: pmap_segtab.c,v 1.4 2016/11/23 03:30:53 mrg 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 1.4 2016/11/23 03:30:53 mrg 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 void
134 pmap_check_stp(pmap_segtab_t *stp, const char *caller, const char *why)
135 {
136 #ifdef DEBUG
137 	for (size_t i = 0; i < PMAP_SEGTABSIZE; i++) {
138 		if (stp->seg_tab[i] != 0) {
139 #ifdef DEBUG_NOISY
140 			for (size_t j = i; j < PMAP_SEGTABSIZE; j++)
141 				printf("%s: pm_segtab.seg_tab[%zu] = 0x%p\n",
142 				       caller, j, stp->seg_tab[j]);
143 #endif
144 			panic("%s: pm_segtab.seg_tab[%zu] != 0 (0x%p): %s",
145 			      caller, i, stp->seg_tab[i], why);
146 		}
147 	}
148 #endif
149 }
150 
151 static inline struct vm_page *
152 pmap_pte_pagealloc(void)
153 {
154 	struct vm_page *pg;
155 
156 	pg = PMAP_ALLOC_POOLPAGE(UVM_PGA_ZERO|UVM_PGA_USERESERVE);
157 	if (pg) {
158 #ifdef UVM_PAGE_TRKOWN
159 		pg->owner_tag = NULL;
160 #endif
161 		UVM_PAGE_OWN(pg, "pmap-ptp");
162 	}
163 
164 	return pg;
165 }
166 
167 static inline pt_entry_t *
168 pmap_segmap(struct pmap *pmap, vaddr_t va)
169 {
170 	pmap_segtab_t *stp = pmap->pm_segtab;
171 	KASSERT(pmap != pmap_kernel() || !pmap_md_direct_mapped_vaddr_p(va));
172 #ifdef _LP64
173 	stp = stp->seg_seg[(va >> XSEGSHIFT) & (NSEGPG - 1)];
174 	if (stp == NULL)
175 		return NULL;
176 #endif
177 
178 	return stp->seg_tab[(va >> SEGSHIFT) & (PMAP_SEGTABSIZE - 1)];
179 }
180 
181 pt_entry_t *
182 pmap_pte_lookup(pmap_t pmap, vaddr_t va)
183 {
184 	pt_entry_t *pte = pmap_segmap(pmap, va);
185 	if (pte == NULL)
186 		return NULL;
187 
188 	return pte + ((va >> PGSHIFT) & (NPTEPG - 1));
189 }
190 
191 static void
192 pmap_segtab_free(pmap_segtab_t *stp)
193 {
194 	/*
195 	 * Insert the the segtab into the segtab freelist.
196 	 */
197 	mutex_spin_enter(&pmap_segtab_lock);
198 	stp->seg_seg[0] = pmap_segtab_info.free_segtab;
199 	pmap_segtab_info.free_segtab = stp;
200 	SEGTAB_ADD(nput, 1);
201 	mutex_spin_exit(&pmap_segtab_lock);
202 }
203 
204 static void
205 pmap_segtab_release(pmap_t pmap, pmap_segtab_t **stp_p, bool free_stp,
206 	pte_callback_t callback, uintptr_t flags,
207 	vaddr_t va, vsize_t vinc)
208 {
209 	pmap_segtab_t *stp = *stp_p;
210 
211 	for (size_t i = (va / vinc) & (PMAP_SEGTABSIZE - 1);
212 	     i < PMAP_SEGTABSIZE;
213 	     i++, va += vinc) {
214 #ifdef _LP64
215 		if (vinc > NBSEG) {
216 			if (stp->seg_seg[i] != NULL) {
217 				pmap_segtab_release(pmap, &stp->seg_seg[i],
218 				    true, callback, flags, va, vinc / NSEGPG);
219 				KASSERT(stp->seg_seg[i] == NULL);
220 			}
221 			continue;
222 		}
223 #endif
224 		KASSERT(vinc == NBSEG);
225 
226 		/* get pointer to segment map */
227 		pt_entry_t *pte = stp->seg_tab[i];
228 		if (pte == NULL)
229 			continue;
230 
231 		/*
232 		 * If our caller want a callback, do so.
233 		 */
234 		if (callback != NULL) {
235 			(*callback)(pmap, va, va + vinc, pte, flags);
236 		}
237 #ifdef DEBUG
238 		for (size_t j = 0; j < NPTEPG; j++) {
239 			if (!pte_zero_p(pte[j]))
240 				panic("%s: pte entry %p not 0 (%#"PRIxPTE")",
241 				    __func__, &pte[j], pte_value(pte[j]));
242 		}
243 #endif
244 		// PMAP_UNMAP_POOLPAGE should handle any VCA issues itself
245 		paddr_t pa = PMAP_UNMAP_POOLPAGE((vaddr_t)pte);
246 		struct vm_page *pg = PHYS_TO_VM_PAGE(pa);
247 #ifdef PMAP_PTP_CACHE
248 		mutex_spin_enter(&pmap_segtab_lock);
249 		LIST_INSERT_HEAD(&pmap_segtab_info.ptp_pgflist, pg, listq.list);
250 		mutex_spin_exit(&pmap_segtab_lock);
251 #else
252 		uvm_pagefree(pg);
253 #endif
254 
255 		stp->seg_tab[i] = NULL;
256 	}
257 
258 	if (free_stp) {
259 		pmap_check_stp(stp, __func__,
260 			       vinc == NBSEG ? "release seg" : "release xseg");
261 		pmap_segtab_free(stp);
262 		*stp_p = NULL;
263 	}
264 }
265 
266 /*
267  *	Create and return a physical map.
268  *
269  *	If the size specified for the map
270  *	is zero, the map is an actual physical
271  *	map, and may be referenced by the
272  *	hardware.
273  *
274  *	If the size specified is non-zero,
275  *	the map will be used in software only, and
276  *	is bounded by that size.
277  */
278 static pmap_segtab_t *
279 pmap_segtab_alloc(void)
280 {
281 	pmap_segtab_t *stp;
282 	bool found_on_freelist = false;
283 
284  again:
285 	mutex_spin_enter(&pmap_segtab_lock);
286 	if (__predict_true((stp = pmap_segtab_info.free_segtab) != NULL)) {
287 		pmap_segtab_info.free_segtab = stp->seg_seg[0];
288 		stp->seg_seg[0] = NULL;
289 		SEGTAB_ADD(nget, 1);
290 		found_on_freelist = true;
291 	}
292 	mutex_spin_exit(&pmap_segtab_lock);
293 
294 	if (__predict_false(stp == NULL)) {
295 		struct vm_page * const stp_pg = pmap_pte_pagealloc();
296 
297 		if (__predict_false(stp_pg == NULL)) {
298 			/*
299 			 * XXX What else can we do?  Could we deadlock here?
300 			 */
301 			uvm_wait("pmap_create");
302 			goto again;
303 		}
304 		SEGTAB_ADD(npage, 1);
305 		const paddr_t stp_pa = VM_PAGE_TO_PHYS(stp_pg);
306 
307 		stp = (pmap_segtab_t *)PMAP_MAP_POOLPAGE(stp_pa);
308 		const size_t n = NBPG / sizeof(*stp);
309 		if (n > 1) {
310 			/*
311 			 * link all the segtabs in this page together
312 			 */
313 			for (size_t i = 1; i < n - 1; i++) {
314 				stp[i].seg_seg[0] = &stp[i+1];
315 			}
316 			/*
317 			 * Now link the new segtabs into the free segtab list.
318 			 */
319 			mutex_spin_enter(&pmap_segtab_lock);
320 			stp[n-1].seg_seg[0] = pmap_segtab_info.free_segtab;
321 			pmap_segtab_info.free_segtab = stp + 1;
322 			SEGTAB_ADD(nput, n - 1);
323 			mutex_spin_exit(&pmap_segtab_lock);
324 		}
325 	}
326 
327 	pmap_check_stp(stp, __func__,
328 		       found_on_freelist ? "from free list" : "allocated");
329 
330 	return stp;
331 }
332 
333 /*
334  * Allocate the top segment table for the pmap.
335  */
336 void
337 pmap_segtab_init(pmap_t pmap)
338 {
339 
340 	pmap->pm_segtab = pmap_segtab_alloc();
341 }
342 
343 /*
344  *	Retire the given physical map from service.
345  *	Should only be called if the map contains
346  *	no valid mappings.
347  */
348 void
349 pmap_segtab_destroy(pmap_t pmap, pte_callback_t func, uintptr_t flags)
350 {
351 	if (pmap->pm_segtab == NULL)
352 		return;
353 
354 #ifdef _LP64
355 	const vsize_t vinc = NBXSEG;
356 #else
357 	const vsize_t vinc = NBSEG;
358 #endif
359 	pmap_segtab_release(pmap, &pmap->pm_segtab,
360 	    func == NULL, func, flags, pmap->pm_minaddr, vinc);
361 }
362 
363 /*
364  *	Make a new pmap (vmspace) active for the given process.
365  */
366 void
367 pmap_segtab_activate(struct pmap *pm, struct lwp *l)
368 {
369 	if (l == curlwp) {
370 		struct cpu_info * const ci = l->l_cpu;
371 		KASSERT(pm == l->l_proc->p_vmspace->vm_map.pmap);
372 		if (pm == pmap_kernel()) {
373 			ci->ci_pmap_user_segtab = PMAP_INVALID_SEGTAB_ADDRESS;
374 #ifdef _LP64
375 			ci->ci_pmap_user_seg0tab = PMAP_INVALID_SEGTAB_ADDRESS;
376 #endif
377 		} else {
378 			ci->ci_pmap_user_segtab = pm->pm_segtab;
379 #ifdef _LP64
380 			ci->ci_pmap_user_seg0tab = pm->pm_segtab->seg_seg[0];
381 #endif
382 		}
383 	}
384 }
385 
386 /*
387  *	Act on the given range of addresses from the specified map.
388  *
389  *	It is assumed that the start and end are properly rounded to
390  *	the page size.
391  */
392 void
393 pmap_pte_process(pmap_t pmap, vaddr_t sva, vaddr_t eva,
394 	pte_callback_t callback, uintptr_t flags)
395 {
396 #if 0
397 	printf("%s: %p, %"PRIxVADDR", %"PRIxVADDR", %p, %"PRIxPTR"\n",
398 	    __func__, pmap, sva, eva, callback, flags);
399 #endif
400 	while (sva < eva) {
401 		vaddr_t lastseg_va = pmap_trunc_seg(sva) + NBSEG;
402 		if (lastseg_va == 0 || lastseg_va > eva)
403 			lastseg_va = eva;
404 
405 		/*
406 		 * If VA belongs to an unallocated segment,
407 		 * skip to the next segment boundary.
408 		 */
409 		pt_entry_t * const pte = pmap_pte_lookup(pmap, sva);
410 		if (pte != NULL) {
411 			/*
412 			 * Callback to deal with the ptes for this segment.
413 			 */
414 			(*callback)(pmap, sva, lastseg_va, pte, flags);
415 		}
416 		/*
417 		 * In theory we could release pages with no entries,
418 		 * but that takes more effort than we want here.
419 		 */
420 		sva = lastseg_va;
421 	}
422 }
423 
424 /*
425  *	Return a pointer for the pte that corresponds to the specified virtual
426  *	address (va) in the target physical map, allocating if needed.
427  */
428 pt_entry_t *
429 pmap_pte_reserve(pmap_t pmap, vaddr_t va, int flags)
430 {
431 	pmap_segtab_t *stp = pmap->pm_segtab;
432 	pt_entry_t *pte;
433 
434 	pte = pmap_pte_lookup(pmap, va);
435 	if (__predict_false(pte == NULL)) {
436 #ifdef _LP64
437 		pmap_segtab_t ** const stp_p =
438 		    &stp->seg_seg[(va >> XSEGSHIFT) & (NSEGPG - 1)];
439 		if (__predict_false((stp = *stp_p) == NULL)) {
440 			pmap_segtab_t *nstp = pmap_segtab_alloc();
441 #ifdef MULTIPROCESSOR
442 			pmap_segtab_t *ostp = atomic_cas_ptr(stp_p, NULL, nstp);
443 			if (__predict_false(ostp != NULL)) {
444 				pmap_check_stp(nstp, __func__, "reserve");
445 				pmap_segtab_free(nstp);
446 				nstp = ostp;
447 			}
448 #else
449 			*stp_p = nstp;
450 #endif /* MULTIPROCESSOR */
451 			stp = nstp;
452 		}
453 		KASSERT(stp == pmap->pm_segtab->seg_seg[(va >> XSEGSHIFT) & (NSEGPG - 1)]);
454 #endif /* _LP64 */
455 		struct vm_page *pg = NULL;
456 #ifdef PMAP_PTP_CACHE
457 		mutex_spin_enter(&pmap_segtab_lock);
458 		if ((pg = LIST_FIRST(&pmap_segtab_info.ptp_pgflist)) != NULL) {
459 			LIST_REMOVE(pg, listq.list);
460 			KASSERT(LIST_FIRST(&pmap_segtab_info.ptp_pgflist) != pg);
461 		}
462 		mutex_spin_exit(&pmap_segtab_lock);
463 #endif
464 		if (pg == NULL)
465 			pg = pmap_pte_pagealloc();
466 		if (pg == NULL) {
467 			if (flags & PMAP_CANFAIL)
468 				return NULL;
469 			panic("%s: cannot allocate page table page "
470 			    "for va %" PRIxVADDR, __func__, va);
471 		}
472 
473 		const paddr_t pa = VM_PAGE_TO_PHYS(pg);
474 		pte = (pt_entry_t *)PMAP_MAP_POOLPAGE(pa);
475 		pt_entry_t ** const pte_p =
476 		    &stp->seg_tab[(va >> SEGSHIFT) & (PMAP_SEGTABSIZE - 1)];
477 #ifdef MULTIPROCESSOR
478 		pt_entry_t *opte = atomic_cas_ptr(pte_p, NULL, pte);
479 		/*
480 		 * If another thread allocated the segtab needed for this va
481 		 * free the page we just allocated.
482 		 */
483 		if (__predict_false(opte != NULL)) {
484 #ifdef PMAP_PTP_CACHE
485 			mutex_spin_enter(&pmap_segtab_lock);
486 			LIST_INSERT_HEAD(&pmap_segtab_info.ptp_pgflist,
487 			    pg, listq.list);
488 			mutex_spin_exit(&pmap_segtab_lock);
489 #else
490 			PMAP_UNMAP_POOLPAGE((vaddr_t)pte);
491 			uvm_pagefree(pg);
492 #endif
493 			pte = opte;
494 		}
495 #else
496 		*pte_p = pte;
497 #endif
498 		KASSERT(pte == stp->seg_tab[(va >> SEGSHIFT) & (PMAP_SEGTABSIZE - 1)]);
499 
500 #ifdef DEBUG
501 		for (size_t i = 0; i < NPTEPG; i++) {
502 			if (!pte_zero_p(pte[i]))
503 				panic("%s: new segmap %p not empty @ %zu",
504 				    __func__, pte, i);
505 		}
506 #endif
507 		pte += (va >> PGSHIFT) & (NPTEPG - 1);
508 	}
509 
510 	return pte;
511 }
512