xref: /titanic_41/usr/src/uts/i86pc/vm/i86_mmu.c (revision 56f33205c9ed776c3c909e07d52e94610a675740)
1 /*
2  * CDDL HEADER START
3  *
4  * The contents of this file are subject to the terms of the
5  * Common Development and Distribution License (the "License").
6  * You may not use this file except in compliance with the License.
7  *
8  * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9  * or http://www.opensolaris.org/os/licensing.
10  * See the License for the specific language governing permissions
11  * and limitations under the License.
12  *
13  * When distributing Covered Code, include this CDDL HEADER in each
14  * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15  * If applicable, add the following below this CDDL HEADER, with the
16  * fields enclosed by brackets "[]" replaced with your own identifying
17  * information: Portions Copyright [yyyy] [name of copyright owner]
18  *
19  * CDDL HEADER END
20  */
21 /*
22  * Copyright 2010 Sun Microsystems, Inc.  All rights reserved.
23  * Use is subject to license terms.
24  */
25 
26 #include <sys/t_lock.h>
27 #include <sys/memlist.h>
28 #include <sys/cpuvar.h>
29 #include <sys/vmem.h>
30 #include <sys/mman.h>
31 #include <sys/vm.h>
32 #include <sys/kmem.h>
33 #include <sys/cmn_err.h>
34 #include <sys/debug.h>
35 #include <sys/vm_machparam.h>
36 #include <sys/tss.h>
37 #include <sys/vnode.h>
38 #include <vm/hat.h>
39 #include <vm/anon.h>
40 #include <vm/as.h>
41 #include <vm/page.h>
42 #include <vm/seg.h>
43 #include <vm/seg_kmem.h>
44 #include <vm/seg_map.h>
45 #include <vm/hat_i86.h>
46 #include <sys/promif.h>
47 #include <sys/x86_archext.h>
48 #include <sys/systm.h>
49 #include <sys/archsystm.h>
50 #include <sys/sunddi.h>
51 #include <sys/ddidmareq.h>
52 #include <sys/controlregs.h>
53 #include <sys/reboot.h>
54 #include <sys/kdi.h>
55 #include <sys/bootconf.h>
56 #include <sys/bootsvcs.h>
57 #include <sys/bootinfo.h>
58 #include <vm/kboot_mmu.h>
59 
60 #ifdef __xpv
61 #include <sys/hypervisor.h>
62 #endif
63 
64 caddr_t
i86devmap(pfn_t pf,pgcnt_t pgcnt,uint_t prot)65 i86devmap(pfn_t pf, pgcnt_t pgcnt, uint_t prot)
66 {
67 	caddr_t addr;
68 	caddr_t addr1;
69 	page_t *pp;
70 
71 	addr1 = addr = vmem_alloc(heap_arena, mmu_ptob(pgcnt), VM_SLEEP);
72 
73 	for (; pgcnt != 0; addr += MMU_PAGESIZE, ++pf, --pgcnt) {
74 		pp = page_numtopp_nolock(pf);
75 		if (pp == NULL) {
76 			hat_devload(kas.a_hat, addr, MMU_PAGESIZE, pf,
77 			    prot | HAT_NOSYNC, HAT_LOAD_LOCK);
78 		} else {
79 			hat_memload(kas.a_hat, addr, pp,
80 			    prot | HAT_NOSYNC, HAT_LOAD_LOCK);
81 		}
82 	}
83 
84 	return (addr1);
85 }
86 
87 /*
88  * This routine is like page_numtopp, but accepts only free pages, which
89  * it allocates (unfrees) and returns with the exclusive lock held.
90  * It is used by machdep.c/dma_init() to find contiguous free pages.
91  *
92  * XXX this and some others should probably be in vm_machdep.c
93  */
94 page_t *
page_numtopp_alloc(pfn_t pfnum)95 page_numtopp_alloc(pfn_t pfnum)
96 {
97 	page_t *pp;
98 
99 retry:
100 	pp = page_numtopp_nolock(pfnum);
101 	if (pp == NULL) {
102 		return (NULL);
103 	}
104 
105 	if (!page_trylock(pp, SE_EXCL)) {
106 		return (NULL);
107 	}
108 
109 	if (page_pptonum(pp) != pfnum) {
110 		page_unlock(pp);
111 		goto retry;
112 	}
113 
114 	if (!PP_ISFREE(pp)) {
115 		page_unlock(pp);
116 		return (NULL);
117 	}
118 	if (pp->p_szc) {
119 		page_demote_free_pages(pp);
120 		page_unlock(pp);
121 		goto retry;
122 	}
123 
124 	/* If associated with a vnode, destroy mappings */
125 
126 	if (pp->p_vnode) {
127 
128 		page_destroy_free(pp);
129 
130 		if (!page_lock(pp, SE_EXCL, (kmutex_t *)NULL, P_NO_RECLAIM)) {
131 			return (NULL);
132 		}
133 
134 		if (page_pptonum(pp) != pfnum) {
135 			page_unlock(pp);
136 			goto retry;
137 		}
138 	}
139 
140 	if (!PP_ISFREE(pp)) {
141 		page_unlock(pp);
142 		return (NULL);
143 	}
144 
145 	if (!page_reclaim(pp, (kmutex_t *)NULL))
146 		return (NULL);
147 
148 	return (pp);
149 }
150 
151 /*
152  * Flag is not set early in boot. Once it is set we are no longer
153  * using boot's page tables.
154  */
155 uint_t khat_running = 0;
156 
157 /*
158  * This procedure is callable only while the boot loader is in charge of the
159  * MMU. It assumes that PA == VA for page table pointers.  It doesn't live in
160  * kboot_mmu.c since it's used from common code.
161  */
162 pfn_t
va_to_pfn(void * vaddr)163 va_to_pfn(void *vaddr)
164 {
165 	uintptr_t	des_va = ALIGN2PAGE(vaddr);
166 	uintptr_t	va = des_va;
167 	size_t		len;
168 	uint_t		prot;
169 	pfn_t		pfn;
170 
171 	if (khat_running)
172 		panic("va_to_pfn(): called too late\n");
173 
174 	if (kbm_probe(&va, &len, &pfn, &prot) == 0)
175 		return (PFN_INVALID);
176 	if (va > des_va)
177 		return (PFN_INVALID);
178 	if (va < des_va)
179 		pfn += mmu_btop(des_va - va);
180 	return (pfn);
181 }
182 
183 /*
184  * Initialize a special area in the kernel that always holds some PTEs for
185  * faster performance. This always holds segmap's PTEs.
186  * In the 32 bit kernel this maps the kernel heap too.
187  */
188 void
hat_kmap_init(uintptr_t base,size_t len)189 hat_kmap_init(uintptr_t base, size_t len)
190 {
191 	uintptr_t map_addr;	/* base rounded down to large page size */
192 	uintptr_t map_eaddr;	/* base + len rounded up */
193 	size_t map_len;
194 	caddr_t ptes;		/* mapping area in kernel for kmap ptes */
195 	size_t window_size;	/* size of mapping area for ptes */
196 	ulong_t htable_cnt;	/* # of page tables to cover map_len */
197 	ulong_t i;
198 	htable_t *ht;
199 	uintptr_t va;
200 
201 	/*
202 	 * We have to map in an area that matches an entire page table.
203 	 * The PTEs are large page aligned to avoid spurious pagefaults
204 	 * on the hypervisor.
205 	 */
206 	map_addr = base & LEVEL_MASK(1);
207 	map_eaddr = (base + len + LEVEL_SIZE(1) - 1) & LEVEL_MASK(1);
208 	map_len = map_eaddr - map_addr;
209 	window_size = mmu_btop(map_len) * mmu.pte_size;
210 	window_size = (window_size + LEVEL_SIZE(1)) & LEVEL_MASK(1);
211 	htable_cnt = map_len >> LEVEL_SHIFT(1);
212 
213 	/*
214 	 * allocate vmem for the kmap_ptes
215 	 */
216 	ptes = vmem_xalloc(heap_arena, window_size, LEVEL_SIZE(1), 0,
217 	    0, NULL, NULL, VM_SLEEP);
218 	mmu.kmap_htables =
219 	    kmem_alloc(htable_cnt * sizeof (htable_t *), KM_SLEEP);
220 
221 	/*
222 	 * Map the page tables that cover kmap into the allocated range.
223 	 * Note we don't ever htable_release() the kmap page tables - they
224 	 * can't ever be stolen, freed, etc.
225 	 */
226 	for (va = map_addr, i = 0; i < htable_cnt; va += LEVEL_SIZE(1), ++i) {
227 		ht = htable_create(kas.a_hat, va, 0, NULL);
228 		if (ht == NULL)
229 			panic("hat_kmap_init: ht == NULL");
230 		mmu.kmap_htables[i] = ht;
231 
232 		hat_devload(kas.a_hat, ptes + i * MMU_PAGESIZE,
233 		    MMU_PAGESIZE, ht->ht_pfn,
234 #ifdef __xpv
235 		    PROT_READ | HAT_NOSYNC | HAT_UNORDERED_OK,
236 #else
237 		    PROT_READ | PROT_WRITE | HAT_NOSYNC | HAT_UNORDERED_OK,
238 #endif
239 		    HAT_LOAD | HAT_LOAD_NOCONSIST);
240 	}
241 
242 	/*
243 	 * set information in mmu to activate handling of kmap
244 	 */
245 	mmu.kmap_addr = map_addr;
246 	mmu.kmap_eaddr = map_eaddr;
247 	mmu.kmap_ptes = (x86pte_t *)ptes;
248 }
249 
250 extern caddr_t	kpm_vbase;
251 extern size_t	kpm_size;
252 
253 #ifdef __xpv
254 /*
255  * Create the initial segkpm mappings for the hypervisor. To avoid having
256  * to deal with page tables being read only, we make all mappings
257  * read only at first.
258  */
259 static void
xen_kpm_create(paddr_t paddr,level_t lvl)260 xen_kpm_create(paddr_t paddr, level_t lvl)
261 {
262 	ulong_t pg_off;
263 
264 	for (pg_off = 0; pg_off < LEVEL_SIZE(lvl); pg_off += MMU_PAGESIZE) {
265 		kbm_map((uintptr_t)kpm_vbase + paddr, (paddr_t)0, 0, 1);
266 		kbm_read_only((uintptr_t)kpm_vbase + paddr + pg_off,
267 		    paddr + pg_off);
268 	}
269 }
270 
271 /*
272  * Try to make all kpm mappings writable. Failures are ok, as those
273  * are just pagetable, GDT, etc. pages.
274  */
275 static void
xen_kpm_finish_init(void)276 xen_kpm_finish_init(void)
277 {
278 	pfn_t gdtpfn = mmu_btop(CPU->cpu_m.mcpu_gdtpa);
279 	pfn_t pfn;
280 	page_t *pp;
281 
282 	for (pfn = 0; pfn < mfn_count; ++pfn) {
283 		/*
284 		 * skip gdt
285 		 */
286 		if (pfn == gdtpfn)
287 			continue;
288 
289 		/*
290 		 * p_index is a hint that this is a pagetable
291 		 */
292 		pp = page_numtopp_nolock(pfn);
293 		if (pp && pp->p_index) {
294 			pp->p_index = 0;
295 			continue;
296 		}
297 		(void) xen_kpm_page(pfn, PT_VALID | PT_WRITABLE);
298 	}
299 }
300 #endif
301 
302 /*
303  * Routine to pre-allocate data structures for hat_kern_setup(). It computes
304  * how many pagetables it needs by walking the boot loader's page tables.
305  */
306 /*ARGSUSED*/
307 void
hat_kern_alloc(caddr_t segmap_base,size_t segmap_size,caddr_t ekernelheap)308 hat_kern_alloc(
309 	caddr_t	segmap_base,
310 	size_t	segmap_size,
311 	caddr_t	ekernelheap)
312 {
313 	uintptr_t	last_va = (uintptr_t)-1;	/* catch 1st time */
314 	uintptr_t	va = 0;
315 	size_t		size;
316 	pfn_t		pfn;
317 	uint_t		prot;
318 	uint_t		table_cnt = 1;
319 	uint_t		mapping_cnt;
320 	level_t		start_level;
321 	level_t		l;
322 	struct memlist	*pmem;
323 	level_t		lpagel = mmu.max_page_level;
324 	uint64_t	paddr;
325 	int64_t		psize;
326 	int		nwindows;
327 
328 	if (kpm_size > 0) {
329 		/*
330 		 * Create the kpm page tables.  When running on the
331 		 * hypervisor these are made read/only at first.
332 		 * Later we'll add write permission where possible.
333 		 */
334 		for (pmem = phys_install; pmem; pmem = pmem->ml_next) {
335 			paddr = pmem->ml_address;
336 			psize = pmem->ml_size;
337 			while (psize >= MMU_PAGESIZE) {
338 				/* find the largest page size */
339 				for (l = lpagel; l > 0; l--) {
340 					if ((paddr & LEVEL_OFFSET(l)) == 0 &&
341 					    psize > LEVEL_SIZE(l))
342 						break;
343 				}
344 
345 #if defined(__xpv)
346 				/*
347 				 * Create read/only mappings to avoid
348 				 * conflicting with pagetable usage
349 				 */
350 				xen_kpm_create(paddr, l);
351 #else
352 				kbm_map((uintptr_t)kpm_vbase + paddr, paddr,
353 				    l, 1);
354 #endif
355 				paddr += LEVEL_SIZE(l);
356 				psize -= LEVEL_SIZE(l);
357 			}
358 		}
359 	}
360 
361 	/*
362 	 * If this machine doesn't have a kpm segment, we need to allocate
363 	 * a small number of 'windows' which can be used to map pagetables.
364 	 */
365 	nwindows = (kpm_size == 0) ? 2 * NCPU : 0;
366 
367 #if defined(__xpv)
368 	/*
369 	 * On a hypervisor, these windows are also used by the xpv_panic
370 	 * code, where we need one window for each level of the pagetable
371 	 * hierarchy.
372 	 */
373 	nwindows = MAX(nwindows, mmu.max_level);
374 #endif
375 
376 	if (nwindows != 0) {
377 		/*
378 		 * Create the page windows and 1 page of VA in
379 		 * which we map the PTEs of those windows.
380 		 */
381 		mmu.pwin_base = vmem_xalloc(heap_arena, nwindows * MMU_PAGESIZE,
382 		    LEVEL_SIZE(1), 0, 0, NULL, NULL, VM_SLEEP);
383 		ASSERT(nwindows <= MMU_PAGESIZE / mmu.pte_size);
384 		mmu.pwin_pte_va = vmem_xalloc(heap_arena, MMU_PAGESIZE,
385 		    MMU_PAGESIZE, 0, 0, NULL, NULL, VM_SLEEP);
386 
387 		/*
388 		 * Find/Create the page table window mappings.
389 		 */
390 		paddr = 0;
391 		(void) find_pte((uintptr_t)mmu.pwin_base, &paddr, 0, 0);
392 		ASSERT(paddr != 0);
393 		ASSERT((paddr & MMU_PAGEOFFSET) == 0);
394 		mmu.pwin_pte_pa = paddr;
395 #ifdef __xpv
396 		(void) find_pte((uintptr_t)mmu.pwin_pte_va, NULL, 0, 0);
397 		kbm_read_only((uintptr_t)mmu.pwin_pte_va, mmu.pwin_pte_pa);
398 #else
399 		kbm_map((uintptr_t)mmu.pwin_pte_va, mmu.pwin_pte_pa, 0, 1);
400 #endif
401 	}
402 
403 	/*
404 	 * Walk the boot loader's page tables and figure out
405 	 * how many tables and page mappings there will be.
406 	 */
407 	while (kbm_probe(&va, &size, &pfn, &prot) != 0) {
408 		/*
409 		 * At each level, if the last_va falls into a new htable,
410 		 * increment table_cnt. We can stop at the 1st level where
411 		 * they are in the same htable.
412 		 */
413 		start_level = 0;
414 		while (start_level <= mmu.max_page_level) {
415 			if (size == LEVEL_SIZE(start_level))
416 				break;
417 			start_level++;
418 		}
419 
420 		for (l = start_level; l < mmu.max_level; ++l) {
421 			if (va >> LEVEL_SHIFT(l + 1) ==
422 			    last_va >> LEVEL_SHIFT(l + 1))
423 				break;
424 			++table_cnt;
425 		}
426 		last_va = va;
427 		l = (start_level == 0) ? 1 : start_level;
428 		va = (va & LEVEL_MASK(l)) + LEVEL_SIZE(l);
429 	}
430 
431 	/*
432 	 * Besides the boot loader mappings, we're going to fill in
433 	 * the entire top level page table for the kernel. Make sure there's
434 	 * enough reserve for that too.
435 	 */
436 	table_cnt += mmu.top_level_count - ((kernelbase >>
437 	    LEVEL_SHIFT(mmu.max_level)) & (mmu.top_level_count - 1));
438 
439 #if defined(__i386)
440 	/*
441 	 * The 32 bit PAE hat allocates tables one level below the top when
442 	 * kernelbase isn't 1 Gig aligned. We'll just be sloppy and allocate
443 	 * a bunch more to the reserve. Any unused will be returned later.
444 	 * Note we've already counted these mappings, just not the extra
445 	 * pagetables.
446 	 */
447 	if (mmu.pae_hat != 0 && (kernelbase & LEVEL_OFFSET(mmu.max_level)) != 0)
448 		table_cnt += mmu.ptes_per_table -
449 		    ((kernelbase & LEVEL_OFFSET(mmu.max_level)) >>
450 		    LEVEL_SHIFT(mmu.max_level - 1));
451 #endif
452 
453 	/*
454 	 * Add 1/4 more into table_cnt for extra slop.  The unused
455 	 * slop is freed back when we htable_adjust_reserve() later.
456 	 */
457 	table_cnt += table_cnt >> 2;
458 
459 	/*
460 	 * We only need mapping entries (hments) for shared pages.
461 	 * This should be far, far fewer than the total possible,
462 	 * We'll allocate enough for 1/16 of all possible PTEs.
463 	 */
464 	mapping_cnt = (table_cnt * mmu.ptes_per_table) >> 4;
465 
466 	/*
467 	 * Now create the initial htable/hment reserves
468 	 */
469 	htable_initial_reserve(table_cnt);
470 	hment_reserve(mapping_cnt);
471 	x86pte_cpu_init(CPU);
472 }
473 
474 
475 /*
476  * This routine handles the work of creating the kernel's initial mappings
477  * by deciphering the mappings in the page tables created by the boot program.
478  *
479  * We maintain large page mappings, but only to a level 1 pagesize.
480  * The boot loader can only add new mappings once this function starts.
481  * In particular it can not change the pagesize used for any existing
482  * mappings or this code breaks!
483  */
484 
485 void
hat_kern_setup(void)486 hat_kern_setup(void)
487 {
488 	/*
489 	 * Attach htables to the existing pagetables
490 	 */
491 	/* BEGIN CSTYLED */
492 	htable_attach(kas.a_hat, 0, mmu.max_level, NULL,
493 #ifdef __xpv
494 	    mmu_btop(xen_info->pt_base - ONE_GIG));
495 #else
496 	    mmu_btop(getcr3()));
497 #endif
498 	/* END CSTYLED */
499 
500 #if defined(__i386) && !defined(__xpv)
501 	CPU->cpu_tss->tss_cr3 = dftss0->tss_cr3 = getcr3();
502 #endif /* __i386 */
503 
504 #if defined(__xpv) && defined(__amd64)
505 	/*
506 	 * Try to make the kpm mappings r/w. Failures here are OK, as
507 	 * it's probably just a pagetable
508 	 */
509 	xen_kpm_finish_init();
510 #endif
511 
512 	/*
513 	 * The kernel HAT is now officially open for business.
514 	 */
515 	khat_running = 1;
516 
517 	CPUSET_ATOMIC_ADD(kas.a_hat->hat_cpus, CPU->cpu_id);
518 	CPU->cpu_current_hat = kas.a_hat;
519 }
520