xref: /netbsd-src/sys/kern/subr_physmap.c (revision deb6f0161a9109e7de9b519dc8dfb9478668dcdd)
1 /*	$NetBSD: subr_physmap.c,v 1.2 2013/01/19 01:04:51 rmind Exp $	*/
2 
3 /*-
4  * Copyright (c) 2013 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 of 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 __KERNEL_RCSID(1, "$NetBSD: subr_physmap.c,v 1.2 2013/01/19 01:04:51 rmind Exp $");
34 
35 #include <sys/param.h>
36 #include <sys/physmap.h>
37 #include <sys/kmem.h>
38 
39 #include <dev/mm.h>
40 
41 /*
42  * This file contain support routines used to create and destroy lists of
43  * physical pages from lists of pages or ranges of virtual address.  By using
44  * these physical maps, the kernel can avoid mapping physical I/O in the
45  * kernel's address space in most cases.
46  */
47 
48 typedef struct {
49 	physmap_t *pc_physmap;
50 	physmap_segment_t *pc_segs;
51 	vsize_t pc_offset;
52 	vsize_t pc_klen;
53 	vaddr_t pc_kva;
54 	u_int pc_nsegs;
55 	vm_prot_t pc_prot;
56 	bool pc_direct_mapped;
57 } physmap_cookie_t;
58 
59 /*
60  * Allocate a physmap structure that requires "maxsegs" segments.
61  */
62 static physmap_t *
63 physmap_alloc(size_t maxsegs)
64 {
65 	const size_t mapsize = offsetof(physmap_t, pm_segs[maxsegs]);
66 
67 	KASSERT(maxsegs > 0);
68 
69 	physmap_t * const map = kmem_zalloc(mapsize, KM_SLEEP);
70 	map->pm_maxsegs = maxsegs;
71 
72 	return map;
73 }
74 
75 static int
76 physmap_fill(physmap_t *map, pmap_t pmap, vaddr_t va, vsize_t len)
77 {
78 	size_t nsegs = map->pm_nsegs;
79 	physmap_segment_t *ps = &map->pm_segs[nsegs];
80 	vsize_t offset = va - trunc_page(va);
81 
82 	if (nsegs == 0) {
83 		if (!pmap_extract(pmap, va, &ps->ps_addr)) {
84 			return EFAULT;
85 		}
86 		ps->ps_len = MIN(len, PAGE_SIZE - offset);
87 		if (ps->ps_len == len) {
88 			map->pm_nsegs = 1;
89 			return 0;
90 		}
91 		offset = 0;
92 	} else {
93 		/*
94 		 * Backup to the last segment since we have to see if we can
95 		 * merge virtual addresses that are physically contiguous into
96 		 * as few segments as possible.
97 		 */
98 		ps--;
99 		nsegs--;
100 	}
101 
102 	paddr_t lastaddr = ps->ps_addr + ps->ps_len;
103 	for (;;) {
104 		paddr_t curaddr;
105 		if (!pmap_extract(pmap, va, &curaddr)) {
106 			return EFAULT;
107 		}
108 		if (curaddr != lastaddr) {
109 			ps++;
110 			nsegs++;
111 			KASSERT(nsegs < map->pm_maxsegs);
112 			ps->ps_addr = curaddr;
113 			lastaddr = curaddr;
114 		}
115 		if (offset + len > PAGE_SIZE) {
116 			ps->ps_len += PAGE_SIZE - offset;
117 			lastaddr = ps->ps_addr + ps->ps_len;
118 			len -= PAGE_SIZE - offset;
119 			lastaddr += PAGE_SIZE - offset;
120 			offset = 0;
121 		} else {
122 			ps->ps_len += len;
123 			map->pm_nsegs = nsegs + 1;
124 			return 0;
125 		}
126 	}
127 }
128 
129 /*
130  * Create a physmap and populate it with the pages that are used to mapped
131  * linear range of virtual addresses.  It is assumed that uvm_vslock has been
132  * called to lock these pages into memory.
133  */
134 int
135 physmap_create_linear(physmap_t **map_p, const struct vmspace *vs, vaddr_t va,
136 	vsize_t len)
137 {
138 	const size_t maxsegs = atop(round_page(va + len) - trunc_page(va));
139 	physmap_t * const map = physmap_alloc(maxsegs);
140 	int error = physmap_fill(map, vs->vm_map.pmap, va, len);
141 	if (error) {
142 		physmap_destroy(map);
143 		*map_p = NULL;
144 		return error;
145 	}
146 	*map_p = map;
147 	return 0;
148 }
149 
150 /*
151  * Create a physmap and populate it with the pages that are contained in an
152  * iovec array.  It is assumed that uvm_vslock has been called to lock these
153  * pages into memory.
154  */
155 int
156 physmap_create_iov(physmap_t **map_p, const struct vmspace *vs,
157 	struct iovec *iov, size_t iovlen)
158 {
159 	size_t maxsegs = 0;
160 	for (size_t i = 0; i < iovlen; i++) {
161 		const vaddr_t start = (vaddr_t) iov[i].iov_base;
162 		const vaddr_t end = start + iov[i].iov_len;
163 		maxsegs += atop(round_page(end) - trunc_page(start));
164 	}
165 	physmap_t * const map = physmap_alloc(maxsegs);
166 
167 	for (size_t i = 0; i < iovlen; i++) {
168 		int error = physmap_fill(map, vs->vm_map.pmap,
169 		    (vaddr_t) iov[i].iov_base, iov[i].iov_len);
170 		if (error) {
171 			physmap_destroy(map);
172 			*map_p = NULL;
173 			return error;
174 		}
175 	}
176 	*map_p = map;
177 	return 0;
178 }
179 
180 /*
181  * This uses a list of vm_page structure to create a physmap.
182  */
183 physmap_t *
184 physmap_create_pagelist(struct vm_page **pgs, size_t npgs)
185 {
186 	physmap_t * const map = physmap_alloc(npgs);
187 
188 	physmap_segment_t *ps = map->pm_segs;
189 
190 	/*
191 	 * Initialize the first segment.
192 	 */
193 	paddr_t lastaddr = VM_PAGE_TO_PHYS(pgs[0]);
194 	ps->ps_addr = lastaddr;
195 	ps->ps_len = PAGE_SIZE;
196 
197 	for (pgs++; npgs-- > 1; pgs++) {
198 		/*
199 		 * lastaddr needs to be increased by a page.
200 		 */
201 		lastaddr += PAGE_SIZE;
202 		paddr_t curaddr = VM_PAGE_TO_PHYS(*pgs);
203 		if (curaddr != lastaddr) {
204 			/*
205 			 * If the addresses are not the same, we need to use
206 			 * a new segemnt.  Set its address and update lastaddr.
207 			 */
208 			ps++;
209 			ps->ps_addr = curaddr;
210 			lastaddr = curaddr;
211 		}
212 		/*
213 		 * Increase this segment's length by a page
214 		 */
215 		ps->ps_len += PAGE_SIZE;
216 	}
217 
218 	map->pm_nsegs = ps + 1 - map->pm_segs;
219 	return map;
220 }
221 
222 void
223 physmap_destroy(physmap_t *map)
224 {
225 	const size_t mapsize = offsetof(physmap_t, pm_segs[map->pm_maxsegs]);
226 
227 	kmem_free(map, mapsize);
228 }
229 
230 void *
231 physmap_map_init(physmap_t *map, size_t offset, vm_prot_t prot)
232 {
233 	physmap_cookie_t * const pc = kmem_zalloc(sizeof(*pc), KM_SLEEP);
234 
235 	KASSERT(prot == VM_PROT_READ || prot == (VM_PROT_READ|VM_PROT_WRITE));
236 
237 	pc->pc_physmap = map;
238 	pc->pc_segs = map->pm_segs;
239 	pc->pc_nsegs = map->pm_nsegs;
240 	pc->pc_prot = prot;
241 	pc->pc_klen = 0;
242 	pc->pc_kva = 0;
243 	pc->pc_direct_mapped = false;
244 
245 	/*
246 	 * Skip to the first segment we are interested in.
247 	 */
248 	while (offset >= pc->pc_segs->ps_len) {
249 		offset -= pc->pc_segs->ps_len;
250 		pc->pc_segs++;
251 		pc->pc_nsegs--;
252 	}
253 
254 	pc->pc_offset = offset;
255 
256 	return pc;
257 }
258 
259 size_t
260 physmap_map(void *cookie, vaddr_t *kvap)
261 {
262 	physmap_cookie_t * const pc = cookie;
263 
264 	/*
265 	 * If there is currently a non-direct mapped KVA region allocated,
266 	 * free it now.
267 	 */
268 	if (pc->pc_kva != 0 && !pc->pc_direct_mapped) {
269 		pmap_kremove(pc->pc_kva, pc->pc_klen);
270 		pmap_update(pmap_kernel());
271 		uvm_km_free(kernel_map, pc->pc_kva, pc->pc_klen,
272 		    UVM_KMF_VAONLY);
273 	}
274 
275 	/*
276 	 * If there are no more segments to process, return 0 indicating
277 	 * we are done.
278 	 */
279 	if (pc->pc_nsegs == 0) {
280 		return 0;
281 	}
282 
283 	/*
284 	 * Get starting physical address of this segment and its length.
285 	 */
286 	paddr_t pa = pc->pc_segs->ps_addr + pc->pc_offset;
287 	const size_t koff = pa & PAGE_MASK;
288 	const size_t len = pc->pc_segs->ps_len - pc->pc_offset;
289 
290 	/*
291 	 * Now that we have the starting offset in the page, reset to the
292 	 * beginning of the page.
293 	 */
294 	pa = trunc_page(pa);
295 
296 	/*
297 	 * We are now done with this segment; advance to the next one.
298 	 */
299 	pc->pc_segs++;
300 	pc->pc_nsegs--;
301 	pc->pc_offset = 0;
302 
303 	/*
304 	 * Find out how many pages we are mapping.
305 	 */
306 	pc->pc_klen = round_page(len);
307 #ifdef __HAVE_MM_MD_DIRECT_MAPPED_PHYS
308 	/*
309 	 * Always try to direct map it since that's nearly zero cost.
310 	 */
311 	pc->pc_direct_mapped = mm_md_direct_mapped_phys(pa, &pc->pc_kva);
312 #endif
313 	if (!pc->pc_direct_mapped) {
314 		/*
315 		 * If we can't direct map it, we have to allocate some KVA
316 		 * so we map it via the kernel_map.
317 		 */
318 		pc->pc_kva = uvm_km_alloc(kernel_map, pc->pc_klen,
319 		    atop(pa) & uvmexp.ncolors,
320 		    UVM_KMF_VAONLY | UVM_KMF_WAITVA | UVM_KMF_COLORMATCH);
321 		KASSERT(pc->pc_kva != 0);
322 
323 		/*
324 		 * Setup mappings for this segment.
325 		 */
326 		for (size_t poff = 0; poff < pc->pc_klen; poff += PAGE_SIZE) {
327 			pmap_kenter_pa(pc->pc_kva + poff, pa + poff,
328 			    pc->pc_prot, 0);
329 		}
330 		/*
331 		 * Make them real.
332 		 */
333 		pmap_update(pmap_kernel());
334 	}
335 	/*
336 	 * Return the starting KVA (including offset into the page) and
337 	 * the length of this segment.
338 	 */
339 	*kvap = pc->pc_kva + koff;
340 	return len;
341 }
342 
343 void
344 physmap_map_fini(void *cookie)
345 {
346 	physmap_cookie_t * const pc = cookie;
347 
348 	/*
349 	 * If there is currently a non-direct mapped KVA region allocated,
350 	 * free it now.
351 	 */
352 	if (pc->pc_kva != 0 && !pc->pc_direct_mapped) {
353 		pmap_kremove(pc->pc_kva, pc->pc_klen);
354 		pmap_update(pmap_kernel());
355 		uvm_km_free(kernel_map, pc->pc_kva, pc->pc_klen,
356 		    UVM_KMF_VAONLY);
357 	}
358 
359 	/*
360 	 * Free the cookie.
361 	 */
362 	kmem_free(pc, sizeof(*pc));
363 }
364 
365 /*
366  * genio needs to zero pages past the EOF or without backing storage (think
367  * sparse files).  But since we are using physmaps, there is no kva to use with
368  * memset so we need a helper to obtain a kva and memset the desired memory.
369  */
370 void
371 physmap_zero(physmap_t *map, size_t offset, size_t len)
372 {
373 	void * const cookie = physmap_map_init(map, offset,
374 	    VM_PROT_READ|VM_PROT_WRITE);
375 
376 	for (;;) {
377 		vaddr_t kva;
378 		size_t seglen = physmap_map(cookie, &kva);
379 		KASSERT(seglen != 0);
380 		if (seglen > len)
381 			seglen = len;
382 		memset((void *)kva, 0, seglen);
383 		if (seglen == len)
384 			break;
385 	}
386 
387 	physmap_map_fini(cookie);
388 }
389