xref: /freebsd/sys/powerpc/aim/slb.c (revision 9a14aa017b21c292740c00ee098195cd46642730)
1 /*-
2  * Copyright (c) 2010 Nathan Whitehorn
3  * All rights reserved.
4  *
5  * Redistribution and use in source and binary forms, with or without
6  * modification, are permitted provided that the following conditions
7  * are met:
8  *
9  * 1. Redistributions of source code must retain the above copyright
10  *    notice, this list of conditions and the following disclaimer.
11  * 2. Redistributions in binary form must reproduce the above copyright
12  *    notice, this list of conditions and the following disclaimer in the
13  *    documentation and/or other materials provided with the distribution.
14  *
15  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR
16  * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES
17  * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED.
18  * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT,
19  * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT
20  * NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
21  * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
22  * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
23  * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF
24  * THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
25  *
26  * $FreeBSD$
27  */
28 
29 #include <sys/param.h>
30 #include <sys/kernel.h>
31 #include <sys/lock.h>
32 #include <sys/mutex.h>
33 #include <sys/proc.h>
34 #include <sys/systm.h>
35 
36 #include <vm/vm.h>
37 #include <vm/pmap.h>
38 #include <vm/uma.h>
39 #include <vm/vm.h>
40 #include <vm/vm_map.h>
41 #include <vm/vm_page.h>
42 #include <vm/vm_pageout.h>
43 
44 #include <machine/md_var.h>
45 #include <machine/platform.h>
46 #include <machine/pmap.h>
47 #include <machine/vmparam.h>
48 
49 uintptr_t moea64_get_unique_vsid(void);
50 void moea64_release_vsid(uint64_t vsid);
51 static void slb_zone_init(void *);
52 
53 static uma_zone_t slbt_zone;
54 static uma_zone_t slb_cache_zone;
55 int n_slbs = 64;
56 
57 SYSINIT(slb_zone_init, SI_SUB_KMEM, SI_ORDER_ANY, slb_zone_init, NULL);
58 
59 struct slbtnode {
60 	uint16_t	ua_alloc;
61 	uint8_t		ua_level;
62 	/* Only 36 bits needed for full 64-bit address space. */
63 	uint64_t	ua_base;
64 	union {
65 		struct slbtnode	*ua_child[16];
66 		struct slb	slb_entries[16];
67 	} u;
68 };
69 
70 /*
71  * For a full 64-bit address space, there are 36 bits in play in an
72  * esid, so 8 levels, with the leaf being at level 0.
73  *
74  * |3333|3322|2222|2222|1111|1111|11  |    |    |  esid
75  * |5432|1098|7654|3210|9876|5432|1098|7654|3210|  bits
76  * +----+----+----+----+----+----+----+----+----+--------
77  * | 8  | 7  | 6  | 5  | 4  | 3  | 2  | 1  | 0  | level
78  */
79 #define UAD_ROOT_LEVEL  8
80 #define UAD_LEAF_LEVEL  0
81 
82 static inline int
83 esid2idx(uint64_t esid, int level)
84 {
85 	int shift;
86 
87 	shift = level * 4;
88 	return ((esid >> shift) & 0xF);
89 }
90 
91 /*
92  * The ua_base field should have 0 bits after the first 4*(level+1)
93  * bits; i.e. only
94  */
95 #define uad_baseok(ua)                          \
96 	(esid2base(ua->ua_base, ua->ua_level) == ua->ua_base)
97 
98 
99 static inline uint64_t
100 esid2base(uint64_t esid, int level)
101 {
102 	uint64_t mask;
103 	int shift;
104 
105 	shift = (level + 1) * 4;
106 	mask = ~((1ULL << shift) - 1);
107 	return (esid & mask);
108 }
109 
110 /*
111  * Allocate a new leaf node for the specified esid/vmhandle from the
112  * parent node.
113  */
114 static struct slb *
115 make_new_leaf(uint64_t esid, uint64_t slbv, struct slbtnode *parent)
116 {
117 	struct slbtnode *child;
118 	struct slb *retval;
119 	int idx;
120 
121 	idx = esid2idx(esid, parent->ua_level);
122 	KASSERT(parent->u.ua_child[idx] == NULL, ("Child already exists!"));
123 
124 	/* unlock and M_WAITOK and loop? */
125 	child = uma_zalloc(slbt_zone, M_NOWAIT | M_ZERO);
126 	KASSERT(child != NULL, ("unhandled NULL case"));
127 
128 	child->ua_level = UAD_LEAF_LEVEL;
129 	child->ua_base = esid2base(esid, child->ua_level);
130 	idx = esid2idx(esid, child->ua_level);
131 	child->u.slb_entries[idx].slbv = slbv;
132 	child->u.slb_entries[idx].slbe = (esid << SLBE_ESID_SHIFT) | SLBE_VALID;
133 	setbit(&child->ua_alloc, idx);
134 
135 	retval = &child->u.slb_entries[idx];
136 
137 	/*
138 	 * The above stores must be visible before the next one, so
139 	 * that a lockless searcher always sees a valid path through
140 	 * the tree.
141 	 */
142 	powerpc_sync();
143 
144 	idx = esid2idx(esid, parent->ua_level);
145 	parent->u.ua_child[idx] = child;
146 	setbit(&parent->ua_alloc, idx);
147 
148 	return (retval);
149 }
150 
151 /*
152  * Allocate a new intermediate node to fit between the parent and
153  * esid.
154  */
155 static struct slbtnode*
156 make_intermediate(uint64_t esid, struct slbtnode *parent)
157 {
158 	struct slbtnode *child, *inter;
159 	int idx, level;
160 
161 	idx = esid2idx(esid, parent->ua_level);
162 	child = parent->u.ua_child[idx];
163 	KASSERT(esid2base(esid, child->ua_level) != child->ua_base,
164 	    ("No need for an intermediate node?"));
165 
166 	/*
167 	 * Find the level where the existing child and our new esid
168 	 * meet.  It must be lower than parent->ua_level or we would
169 	 * have chosen a different index in parent.
170 	 */
171 	level = child->ua_level + 1;
172 	while (esid2base(esid, level) !=
173 	    esid2base(child->ua_base, level))
174 		level++;
175 	KASSERT(level < parent->ua_level,
176 	    ("Found splitting level %d for %09jx and %09jx, "
177 	    "but it's the same as %p's",
178 	    level, esid, child->ua_base, parent));
179 
180 	/* unlock and M_WAITOK and loop? */
181 	inter = uma_zalloc(slbt_zone, M_NOWAIT | M_ZERO);
182 	KASSERT(inter != NULL, ("unhandled NULL case"));
183 
184 	/* Set up intermediate node to point to child ... */
185 	inter->ua_level = level;
186 	inter->ua_base = esid2base(esid, inter->ua_level);
187 	idx = esid2idx(child->ua_base, inter->ua_level);
188 	inter->u.ua_child[idx] = child;
189 	setbit(&inter->ua_alloc, idx);
190 	powerpc_sync();
191 
192 	/* Set up parent to point to intermediate node ... */
193 	idx = esid2idx(inter->ua_base, parent->ua_level);
194 	parent->u.ua_child[idx] = inter;
195 	setbit(&parent->ua_alloc, idx);
196 
197 	return (inter);
198 }
199 
200 uint64_t
201 kernel_va_to_slbv(vm_offset_t va)
202 {
203 	uint64_t slbv;
204 
205 	/* Set kernel VSID to deterministic value */
206 	slbv = (KERNEL_VSID((uintptr_t)va >> ADDR_SR_SHFT)) << SLBV_VSID_SHIFT;
207 
208 	/* Figure out if this is a large-page mapping */
209 	if (hw_direct_map && va < VM_MIN_KERNEL_ADDRESS) {
210 		/*
211 		 * XXX: If we have set up a direct map, assumes
212 		 * all physical memory is mapped with large pages.
213 		 */
214 		if (mem_valid(va, 0) == 0)
215 			slbv |= SLBV_L;
216 	}
217 
218 	return (slbv);
219 }
220 
221 struct slb *
222 user_va_to_slb_entry(pmap_t pm, vm_offset_t va)
223 {
224 	uint64_t esid = va >> ADDR_SR_SHFT;
225 	struct slbtnode *ua;
226 	int idx;
227 
228 	ua = pm->pm_slb_tree_root;
229 
230 	for (;;) {
231 		KASSERT(uad_baseok(ua), ("uad base %016jx level %d bad!",
232 		    ua->ua_base, ua->ua_level));
233 		idx = esid2idx(esid, ua->ua_level);
234 
235 		/*
236 		 * This code is specific to ppc64 where a load is
237 		 * atomic, so no need for atomic_load macro.
238 		 */
239 		if (ua->ua_level == UAD_LEAF_LEVEL)
240 			return ((ua->u.slb_entries[idx].slbe & SLBE_VALID) ?
241 			    &ua->u.slb_entries[idx] : NULL);
242 
243 		ua = ua->u.ua_child[idx];
244 		if (ua == NULL ||
245 		    esid2base(esid, ua->ua_level) != ua->ua_base)
246 			return (NULL);
247 	}
248 
249 	return (NULL);
250 }
251 
252 uint64_t
253 va_to_vsid(pmap_t pm, vm_offset_t va)
254 {
255 	struct slb *entry;
256 
257 	/* Shortcut kernel case */
258 	if (pm == kernel_pmap)
259 		return (KERNEL_VSID((uintptr_t)va >> ADDR_SR_SHFT));
260 
261 	/*
262 	 * If there is no vsid for this VA, we need to add a new entry
263 	 * to the PMAP's segment table.
264 	 */
265 
266 	entry = user_va_to_slb_entry(pm, va);
267 
268 	if (entry == NULL)
269 		return (allocate_user_vsid(pm,
270 		    (uintptr_t)va >> ADDR_SR_SHFT, 0));
271 
272 	return ((entry->slbv & SLBV_VSID_MASK) >> SLBV_VSID_SHIFT);
273 }
274 
275 uint64_t
276 allocate_user_vsid(pmap_t pm, uint64_t esid, int large)
277 {
278 	uint64_t vsid, slbv;
279 	struct slbtnode *ua, *next, *inter;
280 	struct slb *slb;
281 	int idx;
282 
283 	KASSERT(pm != kernel_pmap, ("Attempting to allocate a kernel VSID"));
284 
285 	PMAP_LOCK_ASSERT(pm, MA_OWNED);
286 	vsid = moea64_get_unique_vsid();
287 
288 	slbv = vsid << SLBV_VSID_SHIFT;
289 	if (large)
290 		slbv |= SLBV_L;
291 
292 	ua = pm->pm_slb_tree_root;
293 
294 	/* Descend to the correct leaf or NULL pointer. */
295 	for (;;) {
296 		KASSERT(uad_baseok(ua),
297 		   ("uad base %09jx level %d bad!", ua->ua_base, ua->ua_level));
298 		idx = esid2idx(esid, ua->ua_level);
299 
300 		if (ua->ua_level == UAD_LEAF_LEVEL) {
301 			ua->u.slb_entries[idx].slbv = slbv;
302 			eieio();
303 			ua->u.slb_entries[idx].slbe = (esid << SLBE_ESID_SHIFT)
304 			    | SLBE_VALID;
305 			setbit(&ua->ua_alloc, idx);
306 			slb = &ua->u.slb_entries[idx];
307 			break;
308 		}
309 
310 		next = ua->u.ua_child[idx];
311 		if (next == NULL) {
312 			slb = make_new_leaf(esid, slbv, ua);
313 			break;
314                 }
315 
316 		/*
317 		 * Check if the next item down has an okay ua_base.
318 		 * If not, we need to allocate an intermediate node.
319 		 */
320 		if (esid2base(esid, next->ua_level) != next->ua_base) {
321 			inter = make_intermediate(esid, ua);
322 			slb = make_new_leaf(esid, slbv, inter);
323 			break;
324 		}
325 
326 		ua = next;
327 	}
328 
329 	/*
330 	 * Someone probably wants this soon, and it may be a wired
331 	 * SLB mapping, so pre-spill this entry.
332 	 */
333 	eieio();
334 	slb_insert_user(pm, slb);
335 
336 	return (vsid);
337 }
338 
339 void
340 free_vsid(pmap_t pm, uint64_t esid, int large)
341 {
342 	struct slbtnode *ua;
343 	int idx;
344 
345 	PMAP_LOCK_ASSERT(pm, MA_OWNED);
346 
347 	ua = pm->pm_slb_tree_root;
348 	/* Descend to the correct leaf. */
349 	for (;;) {
350 		KASSERT(uad_baseok(ua),
351 		   ("uad base %09jx level %d bad!", ua->ua_base, ua->ua_level));
352 
353 		idx = esid2idx(esid, ua->ua_level);
354 		if (ua->ua_level == UAD_LEAF_LEVEL) {
355 			ua->u.slb_entries[idx].slbv = 0;
356 			eieio();
357 			ua->u.slb_entries[idx].slbe = 0;
358 			clrbit(&ua->ua_alloc, idx);
359 			return;
360 		}
361 
362 		ua = ua->u.ua_child[idx];
363 		if (ua == NULL ||
364 		    esid2base(esid, ua->ua_level) != ua->ua_base) {
365 			/* Perhaps just return instead of assert? */
366 			KASSERT(0,
367 			    ("Asked to remove an entry that was never inserted!"));
368 			return;
369 		}
370 	}
371 }
372 
373 static void
374 free_slb_tree_node(struct slbtnode *ua)
375 {
376 	int idx;
377 
378 	for (idx = 0; idx < 16; idx++) {
379 		if (ua->ua_level != UAD_LEAF_LEVEL) {
380 			if (ua->u.ua_child[idx] != NULL)
381 				free_slb_tree_node(ua->u.ua_child[idx]);
382 		} else {
383 			if (ua->u.slb_entries[idx].slbv != 0)
384 				moea64_release_vsid(ua->u.slb_entries[idx].slbv
385 				    >> SLBV_VSID_SHIFT);
386 		}
387 	}
388 
389 	uma_zfree(slbt_zone, ua);
390 }
391 
392 void
393 slb_free_tree(pmap_t pm)
394 {
395 
396 	free_slb_tree_node(pm->pm_slb_tree_root);
397 }
398 
399 struct slbtnode *
400 slb_alloc_tree(void)
401 {
402 	struct slbtnode *root;
403 
404 	root = uma_zalloc(slbt_zone, M_NOWAIT | M_ZERO);
405 	root->ua_level = UAD_ROOT_LEVEL;
406 
407 	return (root);
408 }
409 
410 /* Lock entries mapping kernel text and stacks */
411 
412 void
413 slb_insert_kernel(uint64_t slbe, uint64_t slbv)
414 {
415 	struct slb *slbcache;
416 	int i;
417 
418 	/* We don't want to be preempted while modifying the kernel map */
419 	critical_enter();
420 
421 	slbcache = PCPU_GET(slb);
422 
423 	/* Check for an unused slot, abusing the user slot as a full flag */
424 	if (slbcache[USER_SLB_SLOT].slbe == 0) {
425 		for (i = 0; i < n_slbs; i++) {
426 			if (i == USER_SLB_SLOT)
427 				continue;
428 			if (!(slbcache[i].slbe & SLBE_VALID))
429 				goto fillkernslb;
430 		}
431 
432 		if (i == n_slbs)
433 			slbcache[USER_SLB_SLOT].slbe = 1;
434 	}
435 
436 	i = mftb() % n_slbs;
437 	if (i == USER_SLB_SLOT)
438 			i = (i+1) % n_slbs;
439 
440 fillkernslb:
441 	KASSERT(i != USER_SLB_SLOT,
442 	    ("Filling user SLB slot with a kernel mapping"));
443 	slbcache[i].slbv = slbv;
444 	slbcache[i].slbe = slbe | (uint64_t)i;
445 
446 	/* If it is for this CPU, put it in the SLB right away */
447 	if (pmap_bootstrapped) {
448 		/* slbie not required */
449 		__asm __volatile ("slbmte %0, %1" ::
450 		    "r"(slbcache[i].slbv), "r"(slbcache[i].slbe));
451 	}
452 
453 	critical_exit();
454 }
455 
456 void
457 slb_insert_user(pmap_t pm, struct slb *slb)
458 {
459 	int i;
460 
461 	PMAP_LOCK_ASSERT(pm, MA_OWNED);
462 
463 	if (pm->pm_slb_len < n_slbs) {
464 		i = pm->pm_slb_len;
465 		pm->pm_slb_len++;
466 	} else {
467 		i = mftb() % n_slbs;
468 	}
469 
470 	/* Note that this replacement is atomic with respect to trap_subr */
471 	pm->pm_slb[i] = slb;
472 }
473 
474 static void *
475 slb_uma_real_alloc(uma_zone_t zone, int bytes, u_int8_t *flags, int wait)
476 {
477 	static vm_offset_t realmax = 0;
478 	void *va;
479 	vm_page_t m;
480 	int pflags;
481 
482 	if (realmax == 0)
483 		realmax = platform_real_maxaddr();
484 
485 	*flags = UMA_SLAB_PRIV;
486 	if ((wait & (M_NOWAIT | M_USE_RESERVE)) == M_NOWAIT)
487 		pflags = VM_ALLOC_INTERRUPT | VM_ALLOC_NOOBJ | VM_ALLOC_WIRED;
488 	else
489 		pflags = VM_ALLOC_SYSTEM | VM_ALLOC_NOOBJ | VM_ALLOC_WIRED;
490 	if (wait & M_ZERO)
491 		pflags |= VM_ALLOC_ZERO;
492 
493 	for (;;) {
494 		m = vm_page_alloc_contig(NULL, 0, pflags, 1, 0, realmax,
495 		    PAGE_SIZE, PAGE_SIZE, VM_MEMATTR_DEFAULT);
496 		if (m == NULL) {
497 			if (wait & M_NOWAIT)
498 				return (NULL);
499 			VM_WAIT;
500 		} else
501                         break;
502         }
503 
504 	va = (void *) VM_PAGE_TO_PHYS(m);
505 
506 	if (!hw_direct_map)
507 		pmap_kenter((vm_offset_t)va, VM_PAGE_TO_PHYS(m));
508 
509 	if ((wait & M_ZERO) && (m->flags & PG_ZERO) == 0)
510 		bzero(va, PAGE_SIZE);
511 
512 	return (va);
513 }
514 
515 static void
516 slb_zone_init(void *dummy)
517 {
518 
519 	slbt_zone = uma_zcreate("SLB tree node", sizeof(struct slbtnode),
520 	    NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_VM);
521 	slb_cache_zone = uma_zcreate("SLB cache",
522 	    (n_slbs + 1)*sizeof(struct slb *), NULL, NULL, NULL, NULL,
523 	    UMA_ALIGN_PTR, UMA_ZONE_VM);
524 
525 	if (platform_real_maxaddr() != VM_MAX_ADDRESS) {
526 		uma_zone_set_allocf(slb_cache_zone, slb_uma_real_alloc);
527 		uma_zone_set_allocf(slbt_zone, slb_uma_real_alloc);
528 	}
529 }
530 
531 struct slb **
532 slb_alloc_user_cache(void)
533 {
534 	return (uma_zalloc(slb_cache_zone, M_ZERO));
535 }
536 
537 void
538 slb_free_user_cache(struct slb **slb)
539 {
540 	uma_zfree(slb_cache_zone, slb);
541 }
542