xref: /illumos-gate/usr/src/uts/common/vm/seg_umap.c (revision 8119dad84d6416f13557b0ba8e2aaf9064cbcfd3)
1 /*
2  * This file and its contents are supplied under the terms of the
3  * Common Development and Distribution License ("CDDL"), version 1.0.
4  * You may only use this file in accordance with the terms of version
5  * 1.0 of the CDDL.
6  *
7  * A full copy of the text of the CDDL should have accompanied this
8  * source.  A copy of the CDDL is also available via the Internet at
9  * http://www.illumos.org/license/CDDL.
10  */
11 
12 /*
13  * Copyright 2018 Joyent, Inc.
14  */
15 
16 /*
17  * VM - Kernel-to-user mapping segment
18  *
19  * The umap segment driver was primarily designed to facilitate the comm page:
20  * a portion of kernel memory shared with userspace so that certain (namely
21  * clock-related) actions could operate without making an expensive trip into
22  * the kernel.
23  *
24  * Since the initial requirements for the comm page are slim, advanced features
25  * of the segment driver such as per-page protection have been left
26  * unimplemented at this time.
27  */
28 
29 
30 #include <sys/types.h>
31 #include <sys/param.h>
32 #include <sys/errno.h>
33 #include <sys/cred.h>
34 #include <sys/kmem.h>
35 #include <sys/lgrp.h>
36 #include <sys/mman.h>
37 
38 #include <vm/hat.h>
39 #include <vm/as.h>
40 #include <vm/seg.h>
41 #include <vm/seg_kmem.h>
42 #include <vm/seg_umap.h>
43 
44 
45 static boolean_t segumap_verify_safe(caddr_t, size_t);
46 static int segumap_dup(struct seg *, struct seg *);
47 static int segumap_unmap(struct seg *, caddr_t, size_t);
48 static void segumap_free(struct seg *);
49 static faultcode_t segumap_fault(struct hat *, struct seg *, caddr_t, size_t,
50     enum fault_type, enum seg_rw);
51 static faultcode_t segumap_faulta(struct seg *, caddr_t);
52 static int segumap_setprot(struct seg *, caddr_t, size_t, uint_t);
53 static int segumap_checkprot(struct seg *, caddr_t, size_t, uint_t);
54 static int segumap_sync(struct seg *, caddr_t, size_t, int, uint_t);
55 static size_t segumap_incore(struct seg *, caddr_t, size_t, char *);
56 static int segumap_lockop(struct seg *, caddr_t, size_t, int, int, ulong_t *,
57     size_t);
58 static int segumap_getprot(struct seg *, caddr_t, size_t, uint_t *);
59 static u_offset_t segumap_getoffset(struct seg *, caddr_t);
60 static int segumap_gettype(struct seg *, caddr_t);
61 static int segumap_getvp(struct seg *, caddr_t, struct vnode **);
62 static int segumap_advise(struct seg *, caddr_t, size_t, uint_t);
63 static void segumap_dump(struct seg *);
64 static int segumap_pagelock(struct seg *, caddr_t, size_t, struct page ***,
65     enum lock_type, enum seg_rw);
66 static int segumap_setpagesize(struct seg *, caddr_t, size_t, uint_t);
67 static int segumap_getmemid(struct seg *, caddr_t, memid_t *);
68 static int segumap_capable(struct seg *, segcapability_t);
69 
70 static struct seg_ops segumap_ops = {
71 	segumap_dup,
72 	segumap_unmap,
73 	segumap_free,
74 	segumap_fault,
75 	segumap_faulta,
76 	segumap_setprot,
77 	segumap_checkprot,
78 	NULL,			/* kluster: disabled */
79 	NULL,			/* swapout: disabled */
80 	segumap_sync,
81 	segumap_incore,
82 	segumap_lockop,
83 	segumap_getprot,
84 	segumap_getoffset,
85 	segumap_gettype,
86 	segumap_getvp,
87 	segumap_advise,
88 	segumap_dump,
89 	segumap_pagelock,
90 	segumap_setpagesize,
91 	segumap_getmemid,
92 	NULL,			/* getpolicy: disabled */
93 	segumap_capable,
94 	seg_inherit_notsup
95 };
96 
97 
98 /*
99  * Create a kernel/user-mapped segment.
100  */
101 int
102 segumap_create(struct seg **segpp, void *argsp)
103 {
104 	struct seg *seg = *segpp;
105 	segumap_crargs_t *a = (struct segumap_crargs *)argsp;
106 	segumap_data_t *data;
107 
108 	ASSERT((uintptr_t)a->kaddr > _userlimit);
109 
110 	/*
111 	 * Check several aspects of the mapping request to ensure validity:
112 	 * - kernel pages must reside entirely in kernel space
113 	 * - target protection must be user-accessible
114 	 * - kernel address must be page-aligned
115 	 * - kernel address must reside inside a "safe" segment
116 	 */
117 	if ((uintptr_t)a->kaddr <= _userlimit ||
118 	    ((uintptr_t)a->kaddr + seg->s_size) < (uintptr_t)a->kaddr ||
119 	    (a->prot & PROT_USER) == 0 ||
120 	    ((uintptr_t)a->kaddr & PAGEOFFSET) != 0 ||
121 	    !segumap_verify_safe(a->kaddr, seg->s_size)) {
122 		return (EINVAL);
123 	}
124 
125 	data = kmem_zalloc(sizeof (*data), KM_SLEEP);
126 	rw_init(&data->sud_lock, NULL, RW_DEFAULT, NULL);
127 	data->sud_kaddr = a->kaddr;
128 	data->sud_prot = a->prot;
129 
130 	seg->s_ops = &segumap_ops;
131 	seg->s_data = data;
132 	return (0);
133 }
134 
135 static boolean_t
136 segumap_verify_safe(caddr_t kaddr, size_t len)
137 {
138 	struct seg *seg;
139 
140 	/*
141 	 * Presently, only pages which are backed by segkmem are allowed to be
142 	 * shared with userspace.  This prevents nasty paging behavior with
143 	 * other drivers such as seg_kp.  Furthermore, the backing kernel
144 	 * segment must completely contain the region to be mapped.
145 	 *
146 	 * Failing these checks is fatal for now since such mappings are done
147 	 * in a very limited context from the kernel.
148 	 */
149 	AS_LOCK_ENTER(&kas, RW_READER);
150 	seg = as_segat(&kas, kaddr);
151 	VERIFY(seg != NULL);
152 	VERIFY(seg->s_base + seg->s_size >= kaddr + len);
153 	VERIFY(seg->s_ops == &segkmem_ops);
154 	AS_LOCK_EXIT(&kas);
155 
156 	return (B_TRUE);
157 }
158 
159 static int
160 segumap_dup(struct seg *seg, struct seg *newseg)
161 {
162 	segumap_data_t *sud = (segumap_data_t *)seg->s_data;
163 	segumap_data_t *newsud;
164 
165 	ASSERT(seg->s_as && AS_WRITE_HELD(seg->s_as));
166 
167 	newsud = kmem_zalloc(sizeof (segumap_data_t), KM_SLEEP);
168 	rw_init(&newsud->sud_lock, NULL, RW_DEFAULT, NULL);
169 	newsud->sud_kaddr = sud->sud_kaddr;
170 	newsud->sud_prot = sud->sud_prot;
171 
172 	newseg->s_ops = seg->s_ops;
173 	newseg->s_data = newsud;
174 	return (0);
175 }
176 
177 static int
178 segumap_unmap(struct seg *seg, caddr_t addr, size_t len)
179 {
180 	segumap_data_t *sud = (segumap_data_t *)seg->s_data;
181 
182 	ASSERT(seg->s_as && AS_WRITE_HELD(seg->s_as));
183 
184 	/* Only allow unmap of entire segment */
185 	if (addr != seg->s_base || len != seg->s_size) {
186 		return (EINVAL);
187 	}
188 	if (sud->sud_softlockcnt != 0) {
189 		return (EAGAIN);
190 	}
191 
192 	/*
193 	 * Unconditionally unload the entire segment range.
194 	 */
195 	hat_unload(seg->s_as->a_hat, addr, len, HAT_UNLOAD_UNMAP);
196 
197 	seg_free(seg);
198 	return (0);
199 }
200 
201 static void
202 segumap_free(struct seg *seg)
203 {
204 	segumap_data_t *data = (segumap_data_t *)seg->s_data;
205 
206 	ASSERT(data != NULL);
207 
208 	rw_destroy(&data->sud_lock);
209 	VERIFY(data->sud_softlockcnt == 0);
210 	kmem_free(data, sizeof (*data));
211 	seg->s_data = NULL;
212 }
213 
214 /* ARGSUSED */
215 static faultcode_t
216 segumap_fault(struct hat *hat, struct seg *seg, caddr_t addr, size_t len,
217     enum fault_type type, enum seg_rw tw)
218 {
219 	segumap_data_t *sud = (segumap_data_t *)seg->s_data;
220 
221 	ASSERT(seg->s_as && AS_LOCK_HELD(seg->s_as));
222 
223 	if (type == F_PROT) {
224 		/*
225 		 * Since protection on the segment is fixed, there is nothing
226 		 * to do but report an error for protection faults.
227 		 */
228 		return (FC_PROT);
229 	} else if (type == F_SOFTUNLOCK) {
230 		size_t plen = btop(len);
231 
232 		rw_enter(&sud->sud_lock, RW_WRITER);
233 		VERIFY(sud->sud_softlockcnt >= plen);
234 		sud->sud_softlockcnt -= plen;
235 		rw_exit(&sud->sud_lock);
236 		return (0);
237 	}
238 
239 	ASSERT(type == F_INVAL || type == F_SOFTLOCK);
240 	rw_enter(&sud->sud_lock, RW_WRITER);
241 
242 	if (type == F_INVAL ||
243 	    (type == F_SOFTLOCK && sud->sud_softlockcnt == 0)) {
244 		/*
245 		 * Load the (entire) segment into the HAT.
246 		 *
247 		 * It's possible that threads racing into as_fault will cause
248 		 * seg_umap to load the same range multiple times in quick
249 		 * succession.  Redundant hat_devload operations are safe.
250 		 */
251 		for (uintptr_t i = 0; i < seg->s_size; i += PAGESIZE) {
252 			pfn_t pfn;
253 
254 			pfn = hat_getpfnum(kas.a_hat, sud->sud_kaddr + i);
255 			VERIFY(pfn != PFN_INVALID);
256 			hat_devload(seg->s_as->a_hat, seg->s_base + i,
257 			    PAGESIZE, pfn, sud->sud_prot, HAT_LOAD);
258 		}
259 	}
260 	if (type == F_SOFTLOCK) {
261 		size_t nval = sud->sud_softlockcnt + btop(len);
262 
263 		if (sud->sud_softlockcnt >= nval) {
264 			rw_exit(&sud->sud_lock);
265 			return (FC_MAKE_ERR(EOVERFLOW));
266 		}
267 		sud->sud_softlockcnt = nval;
268 	}
269 
270 	rw_exit(&sud->sud_lock);
271 	return (0);
272 }
273 
274 /* ARGSUSED */
275 static faultcode_t
276 segumap_faulta(struct seg *seg, caddr_t addr)
277 {
278 	/* Do nothing since asynch pagefault should not load translation. */
279 	return (0);
280 }
281 
282 /* ARGSUSED */
283 static int
284 segumap_setprot(struct seg *seg, caddr_t addr, size_t len, uint_t prot)
285 {
286 	ASSERT(seg->s_as && AS_LOCK_HELD(seg->s_as));
287 
288 	/*
289 	 * The seg_umap driver does not yet allow protection to be changed.
290 	 */
291 	return (EACCES);
292 }
293 
294 /* ARGSUSED */
295 static int
296 segumap_checkprot(struct seg *seg, caddr_t addr, size_t len, uint_t prot)
297 {
298 	segumap_data_t *sud = (segumap_data_t *)seg->s_data;
299 	int error = 0;
300 
301 	ASSERT(seg->s_as && AS_LOCK_HELD(seg->s_as));
302 
303 	rw_enter(&sud->sud_lock, RW_READER);
304 	if ((sud->sud_prot & prot) != prot) {
305 		error = EACCES;
306 	}
307 	rw_exit(&sud->sud_lock);
308 	return (error);
309 }
310 
311 /* ARGSUSED */
312 static int
313 segumap_sync(struct seg *seg, caddr_t addr, size_t len, int attr, uint_t flags)
314 {
315 	/* Always succeed since there are no backing store to sync */
316 	return (0);
317 }
318 
319 /* ARGSUSED */
320 static size_t
321 segumap_incore(struct seg *seg, caddr_t addr, size_t len, char *vec)
322 {
323 	size_t sz = 0;
324 
325 	ASSERT(seg->s_as && AS_LOCK_HELD(seg->s_as));
326 
327 	len = (len + PAGEOFFSET) & PAGEMASK;
328 	while (len > 0) {
329 		*vec = 1;
330 		sz += PAGESIZE;
331 		vec++;
332 		len -= PAGESIZE;
333 	}
334 	return (sz);
335 }
336 
337 /* ARGSUSED */
338 static int
339 segumap_lockop(struct seg *seg, caddr_t addr, size_t len, int attr, int op,
340     ulong_t *lockmap, size_t pos)
341 {
342 	/* Report success since kernel pages are always in memory. */
343 	return (0);
344 }
345 
346 static int
347 segumap_getprot(struct seg *seg, caddr_t addr, size_t len, uint_t *protv)
348 {
349 	segumap_data_t *sud = (segumap_data_t *)seg->s_data;
350 	size_t pgno;
351 	uint_t prot;
352 
353 	ASSERT(seg->s_as && AS_LOCK_HELD(seg->s_as));
354 
355 	rw_enter(&sud->sud_lock, RW_READER);
356 	prot = sud->sud_prot;
357 	rw_exit(&sud->sud_lock);
358 
359 	/*
360 	 * Reporting protection is simple since it is not tracked per-page.
361 	 */
362 	pgno = seg_page(seg, addr + len) - seg_page(seg, addr) + 1;
363 	while (pgno > 0) {
364 		protv[--pgno] = prot;
365 	}
366 	return (0);
367 }
368 
369 /* ARGSUSED */
370 static u_offset_t
371 segumap_getoffset(struct seg *seg, caddr_t addr)
372 {
373 	/*
374 	 * To avoid leaking information about the layout of the kernel address
375 	 * space, always report '0' as the offset.
376 	 */
377 	return (0);
378 }
379 
380 /* ARGSUSED */
381 static int
382 segumap_gettype(struct seg *seg, caddr_t addr)
383 {
384 	/*
385 	 * Since already-existing kernel pages are being mapped into userspace,
386 	 * always report the segment type as shared.
387 	 */
388 	return (MAP_SHARED);
389 }
390 
391 /* ARGSUSED */
392 static int
393 segumap_getvp(struct seg *seg, caddr_t addr, struct vnode **vpp)
394 {
395 	ASSERT(seg->s_as && AS_LOCK_HELD(seg->s_as));
396 
397 	*vpp = NULL;
398 	return (0);
399 }
400 
401 /* ARGSUSED */
402 static int
403 segumap_advise(struct seg *seg, caddr_t addr, size_t len, uint_t behav)
404 {
405 	if (behav == MADV_PURGE) {
406 		/* Purge does not make sense for this mapping */
407 		return (EINVAL);
408 	}
409 	/* Indicate success for everything else. */
410 	return (0);
411 }
412 
413 /* ARGSUSED */
414 static void
415 segumap_dump(struct seg *seg)
416 {
417 	/*
418 	 * Since this is a mapping to share kernel data with userspace, nothing
419 	 * additional should be dumped.
420 	 */
421 }
422 
423 /* ARGSUSED */
424 static int
425 segumap_pagelock(struct seg *seg, caddr_t addr, size_t len, struct page ***ppp,
426     enum lock_type type, enum seg_rw rw)
427 {
428 	return (ENOTSUP);
429 }
430 
431 /* ARGSUSED */
432 static int
433 segumap_setpagesize(struct seg *seg, caddr_t addr, size_t len, uint_t szc)
434 {
435 	return (ENOTSUP);
436 }
437 
438 static int
439 segumap_getmemid(struct seg *seg, caddr_t addr, memid_t *memidp)
440 {
441 	segumap_data_t *sud = (segumap_data_t *)seg->s_data;
442 
443 	memidp->val[0] = (uintptr_t)sud->sud_kaddr;
444 	memidp->val[1] = (uintptr_t)(addr - seg->s_base);
445 	return (0);
446 }
447 
448 /* ARGSUSED */
449 static int
450 segumap_capable(struct seg *seg, segcapability_t capability)
451 {
452 	/* no special capablities */
453 	return (0);
454 }
455