xref: /netbsd-src/sys/arch/xen/x86/x86_xpmap.c (revision 9fb66d812c00ebfb445c0b47dea128f32aa6fe96)
1 /*	$NetBSD: x86_xpmap.c,v 1.90 2020/09/06 02:18:53 riastradh Exp $	*/
2 
3 /*
4  * Copyright (c) 2017 The NetBSD Foundation, Inc.
5  * All rights reserved.
6  *
7  * This code is derived from software contributed to The NetBSD Foundation
8  * by Maxime Villard.
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 /*
33  * Copyright (c) 2006 Mathieu Ropert <mro@adviseo.fr>
34  *
35  * Permission to use, copy, modify, and distribute this software for any
36  * purpose with or without fee is hereby granted, provided that the above
37  * copyright notice and this permission notice appear in all copies.
38  *
39  * THE SOFTWARE IS PROVIDED "AS IS" AND THE AUTHOR DISCLAIMS ALL WARRANTIES
40  * WITH REGARD TO THIS SOFTWARE INCLUDING ALL IMPLIED WARRANTIES OF
41  * MERCHANTABILITY AND FITNESS. IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR
42  * ANY SPECIAL, DIRECT, INDIRECT, OR CONSEQUENTIAL DAMAGES OR ANY DAMAGES
43  * WHATSOEVER RESULTING FROM LOSS OF USE, DATA OR PROFITS, WHETHER IN AN
44  * ACTION OF CONTRACT, NEGLIGENCE OR OTHER TORTIOUS ACTION, ARISING OUT OF
45  * OR IN CONNECTION WITH THE USE OR PERFORMANCE OF THIS SOFTWARE.
46  */
47 
48 /*
49  * Copyright (c) 2006, 2007 Manuel Bouyer.
50  *
51  * Redistribution and use in source and binary forms, with or without
52  * modification, are permitted provided that the following conditions
53  * are met:
54  * 1. Redistributions of source code must retain the above copyright
55  *    notice, this list of conditions and the following disclaimer.
56  * 2. Redistributions in binary form must reproduce the above copyright
57  *    notice, this list of conditions and the following disclaimer in the
58  *    documentation and/or other materials provided with the distribution.
59  *
60  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR
61  * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES
62  * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED.
63  * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT,
64  * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT
65  * NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
66  * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
67  * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
68  * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF
69  * THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
70  */
71 
72 /*
73  * Copyright (c) 2004 Christian Limpach.
74  * All rights reserved.
75  *
76  * Redistribution and use in source and binary forms, with or without
77  * modification, are permitted provided that the following conditions
78  * are met:
79  * 1. Redistributions of source code must retain the above copyright
80  *    notice, this list of conditions and the following disclaimer.
81  * 2. Redistributions in binary form must reproduce the above copyright
82  *    notice, this list of conditions and the following disclaimer in the
83  *    documentation and/or other materials provided with the distribution.
84  *
85  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR
86  * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES
87  * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED.
88  * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT,
89  * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT
90  * NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
91  * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
92  * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
93  * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF
94  * THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
95  */
96 
97 #include <sys/cdefs.h>
98 __KERNEL_RCSID(0, "$NetBSD: x86_xpmap.c,v 1.90 2020/09/06 02:18:53 riastradh Exp $");
99 
100 #include "opt_xen.h"
101 #include "opt_ddb.h"
102 #include "ksyms.h"
103 
104 #include <sys/param.h>
105 #include <sys/systm.h>
106 #include <sys/mutex.h>
107 #include <sys/cpu.h>
108 #include <sys/kernel.h>
109 
110 #include <uvm/uvm.h>
111 
112 #include <machine/gdt.h>
113 
114 #include <xen/xenfunc.h>
115 
116 #include <dev/isa/isareg.h>
117 #include <machine/isa_machdep.h>
118 
119 #ifdef XENDEBUG
120 #define	__PRINTK(x) printk x
121 #else
122 #define	__PRINTK(x)
123 #endif
124 
125 /* Xen requires the start_info struct to be page aligned */
126 union start_info_union start_info_union __aligned(PAGE_SIZE);
127 
128 volatile shared_info_t *HYPERVISOR_shared_info __read_mostly;
129 unsigned long *xpmap_phys_to_machine_mapping __read_mostly;
130 kmutex_t pte_lock __cacheline_aligned;
131 vaddr_t xen_dummy_page;
132 pt_entry_t xpmap_pg_nx __read_mostly;
133 
134 #define XPQUEUE_SIZE 2048
135 static mmu_update_t xpq_queue_array[MAXCPUS][XPQUEUE_SIZE];
136 
137 void xen_failsafe_handler(void);
138 
139 extern struct xenstore_domain_interface *xenstore_interface; /* XXX */
140 
141 static void xen_bt_set_readonly(vaddr_t);
142 static void xen_bootstrap_tables(vaddr_t, vaddr_t, size_t, size_t, bool);
143 
144 vaddr_t xen_locore(void);
145 
146 /*
147  * kcpuset internally uses an array of uint32_t while xen uses an array of
148  * u_long. As we're little-endian we can cast one to the other.
149  */
150 typedef union {
151 #ifdef _LP64
152 	uint32_t xcpum_km[2];
153 #else
154 	uint32_t xcpum_km[1];
155 #endif
156 	u_long xcpum_xm;
157 } xcpumask_t;
158 
159 void
160 xen_failsafe_handler(void)
161 {
162 
163 	panic("xen_failsafe_handler called!\n");
164 }
165 
166 void
167 xen_set_ldt(vaddr_t base, uint32_t entries)
168 {
169 	vaddr_t va;
170 	vaddr_t end;
171 	pt_entry_t *ptp;
172 	int s;
173 
174 #ifdef __x86_64__
175 	end = base + (entries << 3);
176 #else
177 	end = base + entries * sizeof(union descriptor);
178 #endif
179 
180 	for (va = base; va < end; va += PAGE_SIZE) {
181 		KASSERT(va >= VM_MIN_KERNEL_ADDRESS);
182 		ptp = kvtopte(va);
183 		pmap_pte_clearbits(ptp, PTE_W);
184 	}
185 	s = splvm();
186 	xpq_queue_set_ldt(base, entries);
187 	splx(s);
188 }
189 
190 void
191 xpq_flush_queue(void)
192 {
193 	mmu_update_t *xpq_queue;
194 	int done = 0, ret;
195 	size_t xpq_idx;
196 
197 	KASSERT(curcpu()->ci_ilevel >= IPL_VM || cold);
198 
199 	xpq_idx = curcpu()->ci_xpq_idx;
200 	xpq_queue = xpq_queue_array[curcpu()->ci_cpuid];
201 
202 retry:
203 	ret = HYPERVISOR_mmu_update(xpq_queue, xpq_idx, &done, DOMID_SELF);
204 
205 	if (ret < 0 && xpq_idx != 0) {
206 		printf("xpq_flush_queue: %zu entries (%d successful) on "
207 		    "cpu%d (%ld)\n",
208 		    xpq_idx, done, curcpu()->ci_index, curcpu()->ci_cpuid);
209 
210 		if (done != 0) {
211 			xpq_queue += done;
212 			xpq_idx -= done;
213 			done = 0;
214 			goto retry;
215 		}
216 
217 		panic("HYPERVISOR_mmu_update failed, ret: %d\n", ret);
218 	}
219 	curcpu()->ci_xpq_idx = 0;
220 }
221 
222 static inline void
223 xpq_increment_idx(void)
224 {
225 	KASSERT(curcpu()->ci_ilevel >= IPL_VM || cold);
226 	if (__predict_false(++curcpu()->ci_xpq_idx == XPQUEUE_SIZE))
227 		xpq_flush_queue();
228 }
229 
230 void
231 xpq_queue_machphys_update(paddr_t ma, paddr_t pa)
232 {
233 	mmu_update_t *xpq_queue = xpq_queue_array[curcpu()->ci_cpuid];
234 	size_t xpq_idx = curcpu()->ci_xpq_idx;
235 
236 	xpq_queue[xpq_idx].ptr = ma | MMU_MACHPHYS_UPDATE;
237 	xpq_queue[xpq_idx].val = pa >> PAGE_SHIFT;
238 	xpq_increment_idx();
239 }
240 
241 void
242 xpq_queue_pte_update(paddr_t ptr, pt_entry_t val)
243 {
244 	mmu_update_t *xpq_queue = xpq_queue_array[curcpu()->ci_cpuid];
245 	size_t xpq_idx = curcpu()->ci_xpq_idx;
246 
247 	xpq_queue[xpq_idx].ptr = ptr | MMU_NORMAL_PT_UPDATE;
248 	xpq_queue[xpq_idx].val = val;
249 	xpq_increment_idx();
250 }
251 
252 void
253 xpq_queue_pt_switch(paddr_t pa)
254 {
255 	struct mmuext_op op;
256 
257 	xpq_flush_queue();
258 
259 	op.cmd = MMUEXT_NEW_BASEPTR;
260 	op.arg1.mfn = pa >> PAGE_SHIFT;
261 	if (HYPERVISOR_mmuext_op(&op, 1, NULL, DOMID_SELF) < 0)
262 		panic(__func__);
263 }
264 
265 void
266 xpq_queue_pin_table(paddr_t pa, int lvl)
267 {
268 	struct mmuext_op op;
269 
270 	xpq_flush_queue();
271 
272 	op.cmd = lvl;
273 	op.arg1.mfn = pa >> PAGE_SHIFT;
274 	if (HYPERVISOR_mmuext_op(&op, 1, NULL, DOMID_SELF) < 0)
275 		panic(__func__);
276 }
277 
278 void
279 xpq_queue_unpin_table(paddr_t pa)
280 {
281 	struct mmuext_op op;
282 
283 	xpq_flush_queue();
284 
285 	op.cmd = MMUEXT_UNPIN_TABLE;
286 	op.arg1.mfn = pa >> PAGE_SHIFT;
287 	if (HYPERVISOR_mmuext_op(&op, 1, NULL, DOMID_SELF) < 0)
288 		panic(__func__);
289 }
290 
291 void
292 xpq_queue_set_ldt(vaddr_t va, uint32_t entries)
293 {
294 	struct mmuext_op op;
295 
296 	xpq_flush_queue();
297 
298 	KASSERT(va == (va & ~PAGE_MASK));
299 	op.cmd = MMUEXT_SET_LDT;
300 	op.arg1.linear_addr = va;
301 	op.arg2.nr_ents = entries;
302 	if (HYPERVISOR_mmuext_op(&op, 1, NULL, DOMID_SELF) < 0)
303 		panic(__func__);
304 }
305 
306 void
307 xpq_queue_tlb_flush(void)
308 {
309 	struct mmuext_op op;
310 
311 	xpq_flush_queue();
312 
313 	op.cmd = MMUEXT_TLB_FLUSH_LOCAL;
314 	if (HYPERVISOR_mmuext_op(&op, 1, NULL, DOMID_SELF) < 0)
315 		panic(__func__);
316 }
317 
318 void
319 xpq_flush_cache(void)
320 {
321 	int s = splvm();
322 
323 	xpq_flush_queue();
324 
325 	asm("wbinvd":::"memory");
326 	splx(s);
327 }
328 
329 void
330 xpq_queue_invlpg(vaddr_t va)
331 {
332 	struct mmuext_op op;
333 
334 	xpq_flush_queue();
335 
336 	op.cmd = MMUEXT_INVLPG_LOCAL;
337 	op.arg1.linear_addr = (va & ~PAGE_MASK);
338 	if (HYPERVISOR_mmuext_op(&op, 1, NULL, DOMID_SELF) < 0)
339 		panic(__func__);
340 }
341 
342 void
343 xen_mcast_invlpg(vaddr_t va, kcpuset_t *kc)
344 {
345 	xcpumask_t xcpumask;
346 	mmuext_op_t op;
347 
348 	kcpuset_export_u32(kc, &xcpumask.xcpum_km[0], sizeof(xcpumask));
349 
350 	xpq_flush_queue();
351 
352 	op.cmd = MMUEXT_INVLPG_MULTI;
353 	op.arg1.linear_addr = va;
354 	set_xen_guest_handle(op.arg2.vcpumask, &xcpumask.xcpum_xm);
355 
356 	if (HYPERVISOR_mmuext_op(&op, 1, NULL, DOMID_SELF) < 0)
357 		panic(__func__);
358 }
359 
360 void
361 xen_bcast_invlpg(vaddr_t va)
362 {
363 	mmuext_op_t op;
364 
365 	xpq_flush_queue();
366 
367 	op.cmd = MMUEXT_INVLPG_ALL;
368 	op.arg1.linear_addr = va;
369 
370 	if (HYPERVISOR_mmuext_op(&op, 1, NULL, DOMID_SELF) < 0)
371 		panic(__func__);
372 }
373 
374 /* This is a synchronous call. */
375 void
376 xen_mcast_tlbflush(kcpuset_t *kc)
377 {
378 	xcpumask_t xcpumask;
379 	mmuext_op_t op;
380 
381 	kcpuset_export_u32(kc, &xcpumask.xcpum_km[0], sizeof(xcpumask));
382 
383 	xpq_flush_queue();
384 
385 	op.cmd = MMUEXT_TLB_FLUSH_MULTI;
386 	set_xen_guest_handle(op.arg2.vcpumask, &xcpumask.xcpum_xm);
387 
388 	if (HYPERVISOR_mmuext_op(&op, 1, NULL, DOMID_SELF) < 0)
389 		panic(__func__);
390 }
391 
392 /* This is a synchronous call. */
393 void
394 xen_bcast_tlbflush(void)
395 {
396 	mmuext_op_t op;
397 
398 	xpq_flush_queue();
399 
400 	op.cmd = MMUEXT_TLB_FLUSH_ALL;
401 
402 	if (HYPERVISOR_mmuext_op(&op, 1, NULL, DOMID_SELF) < 0)
403 		panic(__func__);
404 }
405 
406 void
407 xen_copy_page(paddr_t srcpa, paddr_t dstpa)
408 {
409 	mmuext_op_t op;
410 
411 	op.cmd = MMUEXT_COPY_PAGE;
412 	op.arg1.mfn = xpmap_ptom(dstpa) >> PAGE_SHIFT;
413 	op.arg2.src_mfn = xpmap_ptom(srcpa) >> PAGE_SHIFT;
414 
415 	if (HYPERVISOR_mmuext_op(&op, 1, NULL, DOMID_SELF) < 0)
416 		panic(__func__);
417 }
418 
419 void
420 xen_pagezero(paddr_t pa)
421 {
422 	mmuext_op_t op;
423 
424 	op.cmd = MMUEXT_CLEAR_PAGE;
425 	op.arg1.mfn = xpmap_ptom(pa) >> PAGE_SHIFT;
426 
427 	if (HYPERVISOR_mmuext_op(&op, 1, NULL, DOMID_SELF) < 0)
428 		panic(__func__);
429 }
430 
431 int
432 xpq_update_foreign(paddr_t ptr, pt_entry_t val, int dom, u_int flags)
433 {
434 	mmu_update_t op;
435 	int ok;
436 	int err;
437 
438 	xpq_flush_queue();
439 
440 	op.ptr = ptr;
441 	if (flags & PMAP_MD_XEN_NOTR)
442 		op.ptr |= MMU_PT_UPDATE_NO_TRANSLATE;
443 	op.val = val;
444 	/*
445 	 * here we return a negative error number as Xen error to
446 	 * pmap_enter_ma. only calls from privcmd.c should end here, and
447 	 * it can deal with it.
448 	 */
449 	if ((err = HYPERVISOR_mmu_update(&op, 1, &ok, dom)) < 0) {
450 		return err;
451 	}
452 	return 0;
453 }
454 
455 #if L2_SLOT_KERNBASE > 0
456 #define TABLE_L2_ENTRIES (2 * (NKL2_KIMG_ENTRIES + 1))
457 #else
458 #define TABLE_L2_ENTRIES (NKL2_KIMG_ENTRIES + 1)
459 #endif
460 
461 #ifdef __x86_64__
462 #define PDIRSZ	PTP_LEVELS
463 #else
464 /*
465  * For PAE, we need an L3 page, a single contiguous L2 "superpage" of 4 pages
466  * (all of them mapped by the L3 page), and a shadow page for L3[3].
467  */
468 #define PDIRSZ	(1 + 4 + 1)
469 #endif
470 
471 /*
472  * Xen locore: get rid of the Xen bootstrap tables. Build and switch to new page
473  * tables.
474  *
475  * Virtual address space of the kernel when leaving this function:
476  * +--------------+------------------+-------------+------------+---------------
477  * | KERNEL IMAGE | BOOTSTRAP TABLES | PROC0 UAREA | DUMMY PAGE | HYPER. SHARED
478  * +--------------+------------------+-------------+------------+---------------
479  *
480  * ------+-----------------+-------------+
481  *  INFO | EARLY ZERO PAGE | ISA I/O MEM |
482  * ------+-----------------+-------------+
483  *
484  * DUMMY PAGE is either a PDG for amd64 or a GDT for i386.
485  *
486  * (HYPER. SHARED INFO + EARLY ZERO PAGE + ISA I/O MEM) have no physical
487  * addresses preallocated.
488  */
489 vaddr_t
490 xen_locore(void)
491 {
492 	size_t nL2, oldcount, mapsize;
493 	vaddr_t our_tables, xen_tables;
494 	u_int descs[4];
495 
496 	xen_init_features();
497 
498 	xpmap_phys_to_machine_mapping =
499 	    (unsigned long *)xen_start_info.mfn_list;
500 
501 	/* Set the NX/XD bit, if available. descs[3] = %edx. */
502 	x86_cpuid(0x80000001, descs);
503 	xpmap_pg_nx = (descs[3] & CPUID_NOX) ? PTE_NX : 0;
504 
505 	/* Space after Xen boostrap tables should be free */
506 	xen_tables = xen_start_info.pt_base;
507 	our_tables = xen_tables + (xen_start_info.nr_pt_frames * PAGE_SIZE);
508 
509 	/*
510 	 * Calculate how much space we need. First, everything mapped before
511 	 * the Xen bootstrap tables.
512 	 */
513 	mapsize = xen_tables - KERNTEXTOFF;
514 
515 	/* After the tables we'll have:
516 	 *  - UAREA
517 	 *  - dummy user PGD (x86_64)
518 	 *  - HYPERVISOR_shared_info
519 	 *  - early_zerop
520 	 *  - ISA I/O mem (if needed)
521 	 */
522 	mapsize += UPAGES * PAGE_SIZE;
523 #ifdef __x86_64__
524 	mapsize += PAGE_SIZE;
525 #endif
526 	mapsize += PAGE_SIZE;
527 	mapsize += PAGE_SIZE;
528 #ifdef DOM0OPS
529 	if (xendomain_is_dom0()) {
530 		mapsize += IOM_SIZE;
531 	}
532 #endif
533 
534 	/*
535 	 * At this point, mapsize doesn't include the table size.
536 	 */
537 #ifdef __x86_64__
538 	nL2 = TABLE_L2_ENTRIES;
539 #else
540 	nL2 = (mapsize + (NBPD_L2 - 1)) >> L2_SHIFT;
541 #endif
542 
543 	/*
544 	 * Now compute how many L2 pages we need exactly. This is useful only
545 	 * on i386, since the initial count for amd64 is already enough.
546 	 */
547 	while (KERNTEXTOFF + mapsize + (nL2 + PDIRSZ) * PAGE_SIZE >
548 	    KERNBASE + (nL2 << L2_SHIFT)) {
549 		nL2++;
550 	}
551 
552 #ifdef i386
553 	/*
554 	 * One more L2 page: we'll allocate several pages after kva_start
555 	 * in pmap_bootstrap() before pmap_growkernel(), which have not been
556 	 * counted here. It's not a big issue to allocate one more L2 as
557 	 * pmap_growkernel() will be called anyway.
558 	 */
559 	nL2++;
560 	nkptp[1] = nL2;
561 #endif
562 
563 	/*
564 	 * Install bootstrap pages. We may need more L2 pages than will
565 	 * have the final table here, as it's installed after the final table.
566 	 */
567 	oldcount = nL2;
568 
569 bootstrap_again:
570 
571 	/*
572 	 * Xen space we'll reclaim may not be enough for our new page tables,
573 	 * move bootstrap tables if necessary.
574 	 */
575 	if (our_tables < xen_tables + ((nL2 + PDIRSZ) * PAGE_SIZE))
576 		our_tables = xen_tables + ((nL2 + PDIRSZ) * PAGE_SIZE);
577 
578 	/*
579 	 * Make sure the number of L2 pages we have is enough to map everything
580 	 * from KERNBASE to the bootstrap tables themselves.
581 	 */
582 	if (our_tables + ((oldcount + PDIRSZ) * PAGE_SIZE) >
583 	    KERNBASE + (oldcount << L2_SHIFT)) {
584 		oldcount++;
585 		goto bootstrap_again;
586 	}
587 
588 	/* Create temporary tables */
589 	xen_bootstrap_tables(xen_tables, our_tables,
590 	    xen_start_info.nr_pt_frames, oldcount, false);
591 
592 	/* Create final tables */
593 	xen_bootstrap_tables(our_tables, xen_tables,
594 	    oldcount + PDIRSZ, nL2, true);
595 
596 	/* Zero out PROC0 UAREA and DUMMY PAGE. */
597 	memset((void *)(xen_tables + ((nL2 + PDIRSZ) * PAGE_SIZE)), 0,
598 	    (UPAGES + 1) * PAGE_SIZE);
599 
600 	/* Finally, flush TLB. */
601 	xpq_queue_tlb_flush();
602 
603 	return (xen_tables + ((nL2 + PDIRSZ) * PAGE_SIZE));
604 }
605 
606 /*
607  * Build a new table and switch to it.
608  * old_count is # of old tables (including L4, L3 and L2).
609  * new_count is # of new tables (PTE only).
610  * We assume the areas don't overlap.
611  */
612 static void
613 xen_bootstrap_tables(vaddr_t old_pgd, vaddr_t new_pgd, size_t old_count,
614     size_t new_count, bool final)
615 {
616 	pd_entry_t *L4cpu, *L4, *L3, *L2, *pte;
617 	paddr_t addr;
618 	vaddr_t page, avail, map_end;
619 	int i;
620 	extern char __rodata_start;
621 	extern char __data_start;
622 	extern char __kernel_end;
623 	extern char *early_zerop; /* from pmap.c */
624 #ifdef i386
625 	extern union descriptor tmpgdt[];
626 #endif
627 
628 	/*
629 	 * Layout of RW area after the kernel image:
630 	 *     xencons_interface (if present)
631 	 *     xenstore_interface (if present)
632 	 *     table pages (new_count + PDIRSZ entries)
633 	 * Extra mappings (only when final is true):
634 	 *     UAREA
635 	 *     dummy user PGD (x86_64 only) / GDT page (i386 only)
636 	 *     HYPERVISOR_shared_info
637 	 *     early_zerop
638 	 *     ISA I/O mem (if needed)
639 	 */
640 	map_end = new_pgd + ((new_count + PDIRSZ) * PAGE_SIZE);
641 	if (final) {
642 		map_end += UPAGES * PAGE_SIZE;
643 		xen_dummy_page = (vaddr_t)map_end;
644 		map_end += PAGE_SIZE;
645 		HYPERVISOR_shared_info = (shared_info_t *)map_end;
646 		map_end += PAGE_SIZE;
647 		early_zerop = (char *)map_end;
648 		map_end += PAGE_SIZE;
649 	}
650 
651 	/*
652 	 * We always set atdevbase, as it's used by init386 to find the first
653 	 * available VA. map_end is updated only if we are dom0, so
654 	 * atdevbase -> atdevbase + IOM_SIZE will be mapped only in
655 	 * this case.
656 	 */
657 	if (final) {
658 		atdevbase = map_end;
659 #ifdef DOM0OPS
660 		if (xendomain_is_dom0()) {
661 			/* ISA I/O mem */
662 			map_end += IOM_SIZE;
663 		}
664 #endif
665 	}
666 
667 	__PRINTK(("xen_bootstrap_tables map_end 0x%lx\n", map_end));
668 	__PRINTK(("console %#lx ", xen_start_info.console_mfn));
669 	__PRINTK(("xenstore %#" PRIx32 "\n", xen_start_info.store_mfn));
670 
671 	avail = new_pgd;
672 
673 	/*
674 	 * Create our page tables.
675 	 */
676 
677 #ifdef __x86_64__
678 	/* per-cpu L4 */
679 	L4cpu = (pd_entry_t *)avail;
680 	memset(L4cpu, 0, PAGE_SIZE);
681 	avail += PAGE_SIZE;
682 
683 	/* pmap_kernel L4 */
684 	L4 = (pd_entry_t *)avail;
685 	memset(L4, 0, PAGE_SIZE);
686 	avail += PAGE_SIZE;
687 
688 	/* L3 */
689 	L3 = (pd_entry_t *)avail;
690 	memset(L3, 0, PAGE_SIZE);
691 	avail += PAGE_SIZE;
692 
693 	/* link L4->L3 */
694 	addr = ((u_long)L3) - KERNBASE;
695 	L4cpu[pl4_pi(KERNTEXTOFF)] = xpmap_ptom_masked(addr) | PTE_P | PTE_W;
696 	L4[pl4_pi(KERNTEXTOFF)] = xpmap_ptom_masked(addr) | PTE_P | PTE_W;
697 
698 	/* L2 */
699 	L2 = (pd_entry_t *)avail;
700 	memset(L2, 0, PAGE_SIZE);
701 	avail += PAGE_SIZE;
702 
703 	/* link L3->L2 */
704 	addr = ((u_long)L2) - KERNBASE;
705 	L3[pl3_pi(KERNTEXTOFF)] = xpmap_ptom_masked(addr) | PTE_P | PTE_W;
706 #else
707 	/* no L4 on i386PAE */
708 	__USE(L4cpu);
709 	__USE(L4);
710 
711 	/* L3 */
712 	L3 = (pd_entry_t *)avail;
713 	memset(L3, 0, PAGE_SIZE);
714 	avail += PAGE_SIZE;
715 
716 	/*
717 	 * Our PAE-style level 2, 5 contiguous pages (4 L2 + 1 shadow).
718 	 *                  +-----------------+----------------+---------+
719 	 * Physical layout: | 3 * USERLAND L2 | L2 KERN SHADOW | L2 KERN |
720 	 *                  +-----------------+----------------+---------+
721 	 * However, we enter L3[3] into L2 KERN, and not L2 KERN SHADOW.
722 	 * This way, L2[L2_SLOT_KERN] always points to the shadow.
723 	 */
724 	L2 = (pd_entry_t *)avail;
725 	memset(L2, 0, PAGE_SIZE * 5);
726 	avail += PAGE_SIZE * 5;
727 
728 	/*
729 	 * Link L2 pages in L3, with a special case for L2 KERN. Xen doesn't
730 	 * want RW permissions in L3 entries, it'll add them itself.
731 	 */
732 	addr = ((u_long)L2) - KERNBASE;
733 	for (i = 0; i < 3; i++, addr += PAGE_SIZE) {
734 		L3[i] = xpmap_ptom_masked(addr) | PTE_P;
735 	}
736 	addr += PAGE_SIZE;
737 	L3[3] = xpmap_ptom_masked(addr) | PTE_P;
738 #endif
739 
740 	/* Level 1 */
741 	page = KERNTEXTOFF;
742 	for (i = 0; i < new_count; i ++) {
743 		vaddr_t cur_page = page;
744 
745 		pte = (pd_entry_t *)avail;
746 		memset(pte, 0, PAGE_SIZE);
747 		avail += PAGE_SIZE;
748 
749 		while (pl2_pi(page) == pl2_pi(cur_page)) {
750 			if (page >= map_end) {
751 				/* not mapped at all */
752 				pte[pl1_pi(page)] = 0;
753 				page += PAGE_SIZE;
754 				continue;
755 			}
756 			pte[pl1_pi(page)] = xpmap_ptom_masked(page - KERNBASE);
757 			if (page == (vaddr_t)HYPERVISOR_shared_info) {
758 				pte[pl1_pi(page)] = xen_start_info.shared_info;
759 			}
760 			if ((xpmap_ptom_masked(page - KERNBASE) >> PAGE_SHIFT)
761 			    == xen_start_info.console.domU.mfn) {
762 				xencons_interface = (void *)page;
763 				pte[pl1_pi(page)] = xen_start_info.console_mfn;
764 				pte[pl1_pi(page)] <<= PAGE_SHIFT;
765 			}
766 			if ((xpmap_ptom_masked(page - KERNBASE) >> PAGE_SHIFT)
767 			    == xen_start_info.store_mfn) {
768 				xenstore_interface = (void *)page;
769 				pte[pl1_pi(page)] = xen_start_info.store_mfn;
770 				pte[pl1_pi(page)] <<= PAGE_SHIFT;
771 			}
772 #ifdef DOM0OPS
773 			if (page >= (vaddr_t)atdevbase &&
774 			    page < (vaddr_t)atdevbase + IOM_SIZE) {
775 				pte[pl1_pi(page)] =
776 				    IOM_BEGIN + (page - (vaddr_t)atdevbase);
777 				pte[pl1_pi(page)] |= xpmap_pg_nx;
778 			}
779 #endif
780 
781 			pte[pl1_pi(page)] |= PTE_P;
782 			if (page < (vaddr_t)&__rodata_start) {
783 				/* Map the kernel text RX. Nothing to do. */
784 			} else if (page >= (vaddr_t)&__rodata_start &&
785 			    page < (vaddr_t)&__data_start) {
786 				/* Map the kernel rodata R. */
787 				pte[pl1_pi(page)] |= xpmap_pg_nx;
788 			} else if (page >= old_pgd &&
789 			    page < old_pgd + (old_count * PAGE_SIZE)) {
790 				/* Map the old page tables R. */
791 				pte[pl1_pi(page)] |= xpmap_pg_nx;
792 			} else if (page >= new_pgd &&
793 			    page < new_pgd + ((new_count + PDIRSZ) * PAGE_SIZE)) {
794 				/* Map the new page tables R. */
795 				pte[pl1_pi(page)] |= xpmap_pg_nx;
796 #ifdef i386
797 			} else if (page == (vaddr_t)tmpgdt) {
798 				/*
799 				 * Map bootstrap gdt R/O. Later, we will re-add
800 				 * this page to uvm after making it writable.
801 				 */
802 				pte[pl1_pi(page)] = 0;
803 				page += PAGE_SIZE;
804 				continue;
805 #endif
806 			} else if (page >= (vaddr_t)&__data_start &&
807 			    page < (vaddr_t)&__kernel_end) {
808 				/* Map the kernel data+bss RW. */
809 				pte[pl1_pi(page)] |= PTE_W | xpmap_pg_nx;
810 			} else {
811 				/* Map the page RW. */
812 				pte[pl1_pi(page)] |= PTE_W | xpmap_pg_nx;
813 			}
814 
815 			page += PAGE_SIZE;
816 		}
817 
818 		addr = ((u_long)pte) - KERNBASE;
819 		L2[pl2_pi(cur_page)] = xpmap_ptom_masked(addr) | PTE_W | PTE_P;
820 
821 		/* Mark readonly */
822 		xen_bt_set_readonly((vaddr_t)pte);
823 	}
824 
825 	/* Install recursive page tables mapping */
826 #ifdef __x86_64__
827 	/* Recursive entry in pmap_kernel(). */
828 	L4[PDIR_SLOT_PTE] = xpmap_ptom_masked((paddr_t)L4 - KERNBASE)
829 	    | PTE_P | xpmap_pg_nx;
830 	/* Recursive entry in higher-level per-cpu PD. */
831 	L4cpu[PDIR_SLOT_PTE] = xpmap_ptom_masked((paddr_t)L4cpu - KERNBASE)
832 	    | PTE_P | xpmap_pg_nx;
833 
834 	/* Mark tables RO */
835 	xen_bt_set_readonly((vaddr_t)L2);
836 #else
837 	/* Copy L2 KERN into L2 KERN SHADOW, and reference the latter in cpu0. */
838 	memcpy(&L2[L2_SLOT_KERN + NPDPG], &L2[L2_SLOT_KERN], PAGE_SIZE);
839 	cpu_info_primary.ci_kpm_pdir = &L2[L2_SLOT_KERN + NPDPG];
840 	cpu_info_primary.ci_kpm_pdirpa =
841 	    (vaddr_t)cpu_info_primary.ci_kpm_pdir - KERNBASE;
842 
843 	/*
844 	 * We don't enter a recursive entry from the L3 PD. Instead, we enter
845 	 * the first 4 L2 pages, which includes the kernel's L2 shadow. But we
846 	 * have to enter the shadow after switching %cr3, or Xen will refcount
847 	 * some PTEs with the wrong type.
848 	 */
849 	addr = (u_long)L2 - KERNBASE;
850 	for (i = 0; i < 3; i++, addr += PAGE_SIZE) {
851 		L2[PDIR_SLOT_PTE + i] = xpmap_ptom_masked(addr) | PTE_P |
852 		    xpmap_pg_nx;
853 	}
854 
855 	/* Mark tables RO, and pin L2 KERN SHADOW. */
856 	addr = (u_long)L2 - KERNBASE;
857 	for (i = 0; i < 5; i++, addr += PAGE_SIZE) {
858 		xen_bt_set_readonly(((vaddr_t)L2) + PAGE_SIZE * i);
859 	}
860 	if (final) {
861 		addr = (u_long)L2 - KERNBASE + 3 * PAGE_SIZE;
862 		xpq_queue_pin_l2_table(xpmap_ptom_masked(addr));
863 	}
864 #endif
865 
866 	xen_bt_set_readonly((vaddr_t)L3);
867 #ifdef __x86_64__
868 	xen_bt_set_readonly((vaddr_t)L4cpu);
869 #endif
870 
871 	/* Pin the PGD */
872 #ifdef __x86_64__
873 	xpq_queue_pin_l4_table(xpmap_ptom_masked(new_pgd - KERNBASE));
874 #else
875 	xpq_queue_pin_l3_table(xpmap_ptom_masked(new_pgd - KERNBASE));
876 #endif
877 
878 	/* Save phys. addr of PDP, for libkvm. */
879 #ifdef __x86_64__
880 	PDPpaddr = (u_long)L4 - KERNBASE;
881 #else
882 	PDPpaddr = (u_long)L2 - KERNBASE; /* PDP is the L2 with PAE */
883 #endif
884 
885 	/* Switch to new tables */
886 	xpq_queue_pt_switch(xpmap_ptom_masked(new_pgd - KERNBASE));
887 
888 	if (final) {
889 #ifdef __x86_64__
890 		/* Save the address of the real per-cpu L4 page. */
891 		cpu_info_primary.ci_kpm_pdir = L4cpu;
892 		cpu_info_primary.ci_kpm_pdirpa = ((paddr_t)L4cpu - KERNBASE);
893 #else
894 		/* Save the address of the L3 page */
895 		cpu_info_primary.ci_pae_l3_pdir = L3;
896 		cpu_info_primary.ci_pae_l3_pdirpa = (new_pgd - KERNBASE);
897 
898 		/* Now enter the kernel's PTE mappings */
899 		addr = (u_long)L2 - KERNBASE + PAGE_SIZE * 3;
900 		xpq_queue_pte_update(
901 		    xpmap_ptom(((vaddr_t)&L2[PDIR_SLOT_PTE + 3]) - KERNBASE),
902 		    xpmap_ptom_masked(addr) | PTE_P);
903 		xpq_flush_queue();
904 #endif
905 	}
906 
907 	/*
908 	 * Now we can safely reclaim the space taken by the old tables.
909 	 */
910 
911 	/* Unpin old PGD */
912 	xpq_queue_unpin_table(xpmap_ptom_masked(old_pgd - KERNBASE));
913 
914 	/* Mark old tables RW */
915 	page = old_pgd;
916 	addr = xpmap_mtop((paddr_t)L2[pl2_pi(page)] & PTE_4KFRAME);
917 	pte = (pd_entry_t *)((u_long)addr + KERNBASE);
918 	pte += pl1_pi(page);
919 	while (page < old_pgd + (old_count * PAGE_SIZE) && page < map_end) {
920 		addr = xpmap_ptom(((u_long)pte) - KERNBASE);
921 		xpq_queue_pte_update(addr, *pte | PTE_W);
922 		page += PAGE_SIZE;
923 		/*
924 		 * Our PTEs are contiguous so it's safe to just "++" here.
925 		 */
926 		pte++;
927 	}
928 	xpq_flush_queue();
929 }
930 
931 /*
932  * Mark a page read-only, assuming vaddr = paddr + KERNBASE.
933  */
934 static void
935 xen_bt_set_readonly(vaddr_t page)
936 {
937 	pt_entry_t entry;
938 
939 	entry = xpmap_ptom_masked(page - KERNBASE);
940 	entry |= PTE_P | xpmap_pg_nx;
941 
942 	HYPERVISOR_update_va_mapping(page, entry, UVMF_INVLPG);
943 }
944 
945 #ifdef __x86_64__
946 void
947 xen_set_user_pgd(paddr_t page)
948 {
949 	struct mmuext_op op;
950 
951 	int s = splvm();
952 	xpq_flush_queue();
953 	splx(s);
954 	op.cmd = MMUEXT_NEW_USER_BASEPTR;
955 	op.arg1.mfn = xpmap_ptom_masked(page) >> PAGE_SHIFT;
956 	if (HYPERVISOR_mmuext_op(&op, 1, NULL, DOMID_SELF) < 0)
957 		panic("xen_set_user_pgd: failed to install new user page"
958 			" directory %#" PRIxPADDR, page);
959 }
960 #endif /* __x86_64__ */
961