Pull bugzilla-7897 into release branch
[sfrench/cifs-2.6.git] / drivers / kvm / paging_tmpl.h
1 /*
2  * Kernel-based Virtual Machine driver for Linux
3  *
4  * This module enables machines with Intel VT-x extensions to run virtual
5  * machines without emulation or binary translation.
6  *
7  * MMU support
8  *
9  * Copyright (C) 2006 Qumranet, Inc.
10  *
11  * Authors:
12  *   Yaniv Kamay  <yaniv@qumranet.com>
13  *   Avi Kivity   <avi@qumranet.com>
14  *
15  * This work is licensed under the terms of the GNU GPL, version 2.  See
16  * the COPYING file in the top-level directory.
17  *
18  */
19
20 /*
21  * We need the mmu code to access both 32-bit and 64-bit guest ptes,
22  * so the code in this file is compiled twice, once per pte size.
23  */
24
25 #if PTTYPE == 64
26         #define pt_element_t u64
27         #define guest_walker guest_walker64
28         #define FNAME(name) paging##64_##name
29         #define PT_BASE_ADDR_MASK PT64_BASE_ADDR_MASK
30         #define PT_DIR_BASE_ADDR_MASK PT64_DIR_BASE_ADDR_MASK
31         #define PT_INDEX(addr, level) PT64_INDEX(addr, level)
32         #define SHADOW_PT_INDEX(addr, level) PT64_INDEX(addr, level)
33         #define PT_LEVEL_MASK(level) PT64_LEVEL_MASK(level)
34         #define PT_PTE_COPY_MASK PT64_PTE_COPY_MASK
35         #ifdef CONFIG_X86_64
36         #define PT_MAX_FULL_LEVELS 4
37         #else
38         #define PT_MAX_FULL_LEVELS 2
39         #endif
40 #elif PTTYPE == 32
41         #define pt_element_t u32
42         #define guest_walker guest_walker32
43         #define FNAME(name) paging##32_##name
44         #define PT_BASE_ADDR_MASK PT32_BASE_ADDR_MASK
45         #define PT_DIR_BASE_ADDR_MASK PT32_DIR_BASE_ADDR_MASK
46         #define PT_INDEX(addr, level) PT32_INDEX(addr, level)
47         #define SHADOW_PT_INDEX(addr, level) PT64_INDEX(addr, level)
48         #define PT_LEVEL_MASK(level) PT32_LEVEL_MASK(level)
49         #define PT_PTE_COPY_MASK PT32_PTE_COPY_MASK
50         #define PT_MAX_FULL_LEVELS 2
51 #else
52         #error Invalid PTTYPE value
53 #endif
54
55 /*
56  * The guest_walker structure emulates the behavior of the hardware page
57  * table walker.
58  */
59 struct guest_walker {
60         int level;
61         gfn_t table_gfn[PT_MAX_FULL_LEVELS];
62         pt_element_t *table;
63         pt_element_t *ptep;
64         pt_element_t inherited_ar;
65         gfn_t gfn;
66         u32 error_code;
67 };
68
69 /*
70  * Fetch a guest pte for a guest virtual address
71  */
72 static int FNAME(walk_addr)(struct guest_walker *walker,
73                             struct kvm_vcpu *vcpu, gva_t addr,
74                             int write_fault, int user_fault, int fetch_fault)
75 {
76         hpa_t hpa;
77         struct kvm_memory_slot *slot;
78         pt_element_t *ptep;
79         pt_element_t root;
80         gfn_t table_gfn;
81
82         pgprintk("%s: addr %lx\n", __FUNCTION__, addr);
83         walker->level = vcpu->mmu.root_level;
84         walker->table = NULL;
85         root = vcpu->cr3;
86 #if PTTYPE == 64
87         if (!is_long_mode(vcpu)) {
88                 walker->ptep = &vcpu->pdptrs[(addr >> 30) & 3];
89                 root = *walker->ptep;
90                 if (!(root & PT_PRESENT_MASK))
91                         goto not_present;
92                 --walker->level;
93         }
94 #endif
95         table_gfn = (root & PT64_BASE_ADDR_MASK) >> PAGE_SHIFT;
96         walker->table_gfn[walker->level - 1] = table_gfn;
97         pgprintk("%s: table_gfn[%d] %lx\n", __FUNCTION__,
98                  walker->level - 1, table_gfn);
99         slot = gfn_to_memslot(vcpu->kvm, table_gfn);
100         hpa = safe_gpa_to_hpa(vcpu, root & PT64_BASE_ADDR_MASK);
101         walker->table = kmap_atomic(pfn_to_page(hpa >> PAGE_SHIFT), KM_USER0);
102
103         ASSERT((!is_long_mode(vcpu) && is_pae(vcpu)) ||
104                (vcpu->cr3 & ~(PAGE_MASK | CR3_FLAGS_MASK)) == 0);
105
106         walker->inherited_ar = PT_USER_MASK | PT_WRITABLE_MASK;
107
108         for (;;) {
109                 int index = PT_INDEX(addr, walker->level);
110                 hpa_t paddr;
111
112                 ptep = &walker->table[index];
113                 ASSERT(((unsigned long)walker->table & PAGE_MASK) ==
114                        ((unsigned long)ptep & PAGE_MASK));
115
116                 if (!is_present_pte(*ptep))
117                         goto not_present;
118
119                 if (write_fault && !is_writeble_pte(*ptep))
120                         if (user_fault || is_write_protection(vcpu))
121                                 goto access_error;
122
123                 if (user_fault && !(*ptep & PT_USER_MASK))
124                         goto access_error;
125
126 #if PTTYPE == 64
127                 if (fetch_fault && is_nx(vcpu) && (*ptep & PT64_NX_MASK))
128                         goto access_error;
129 #endif
130
131                 if (!(*ptep & PT_ACCESSED_MASK))
132                         *ptep |= PT_ACCESSED_MASK;      /* avoid rmw */
133
134                 if (walker->level == PT_PAGE_TABLE_LEVEL) {
135                         walker->gfn = (*ptep & PT_BASE_ADDR_MASK)
136                                 >> PAGE_SHIFT;
137                         break;
138                 }
139
140                 if (walker->level == PT_DIRECTORY_LEVEL
141                     && (*ptep & PT_PAGE_SIZE_MASK)
142                     && (PTTYPE == 64 || is_pse(vcpu))) {
143                         walker->gfn = (*ptep & PT_DIR_BASE_ADDR_MASK)
144                                 >> PAGE_SHIFT;
145                         walker->gfn += PT_INDEX(addr, PT_PAGE_TABLE_LEVEL);
146                         break;
147                 }
148
149                 if (walker->level != 3 || is_long_mode(vcpu))
150                         walker->inherited_ar &= walker->table[index];
151                 table_gfn = (*ptep & PT_BASE_ADDR_MASK) >> PAGE_SHIFT;
152                 paddr = safe_gpa_to_hpa(vcpu, *ptep & PT_BASE_ADDR_MASK);
153                 kunmap_atomic(walker->table, KM_USER0);
154                 walker->table = kmap_atomic(pfn_to_page(paddr >> PAGE_SHIFT),
155                                             KM_USER0);
156                 --walker->level;
157                 walker->table_gfn[walker->level - 1 ] = table_gfn;
158                 pgprintk("%s: table_gfn[%d] %lx\n", __FUNCTION__,
159                          walker->level - 1, table_gfn);
160         }
161         walker->ptep = ptep;
162         pgprintk("%s: pte %llx\n", __FUNCTION__, (u64)*ptep);
163         return 1;
164
165 not_present:
166         walker->error_code = 0;
167         goto err;
168
169 access_error:
170         walker->error_code = PFERR_PRESENT_MASK;
171
172 err:
173         if (write_fault)
174                 walker->error_code |= PFERR_WRITE_MASK;
175         if (user_fault)
176                 walker->error_code |= PFERR_USER_MASK;
177         if (fetch_fault)
178                 walker->error_code |= PFERR_FETCH_MASK;
179         return 0;
180 }
181
182 static void FNAME(release_walker)(struct guest_walker *walker)
183 {
184         if (walker->table)
185                 kunmap_atomic(walker->table, KM_USER0);
186 }
187
188 static void FNAME(set_pte)(struct kvm_vcpu *vcpu, u64 guest_pte,
189                            u64 *shadow_pte, u64 access_bits, gfn_t gfn)
190 {
191         ASSERT(*shadow_pte == 0);
192         access_bits &= guest_pte;
193         *shadow_pte = (guest_pte & PT_PTE_COPY_MASK);
194         set_pte_common(vcpu, shadow_pte, guest_pte & PT_BASE_ADDR_MASK,
195                        guest_pte & PT_DIRTY_MASK, access_bits, gfn);
196 }
197
198 static void FNAME(set_pde)(struct kvm_vcpu *vcpu, u64 guest_pde,
199                            u64 *shadow_pte, u64 access_bits, gfn_t gfn)
200 {
201         gpa_t gaddr;
202
203         ASSERT(*shadow_pte == 0);
204         access_bits &= guest_pde;
205         gaddr = (gpa_t)gfn << PAGE_SHIFT;
206         if (PTTYPE == 32 && is_cpuid_PSE36())
207                 gaddr |= (guest_pde & PT32_DIR_PSE36_MASK) <<
208                         (32 - PT32_DIR_PSE36_SHIFT);
209         *shadow_pte = guest_pde & PT_PTE_COPY_MASK;
210         set_pte_common(vcpu, shadow_pte, gaddr,
211                        guest_pde & PT_DIRTY_MASK, access_bits, gfn);
212 }
213
214 /*
215  * Fetch a shadow pte for a specific level in the paging hierarchy.
216  */
217 static u64 *FNAME(fetch)(struct kvm_vcpu *vcpu, gva_t addr,
218                               struct guest_walker *walker)
219 {
220         hpa_t shadow_addr;
221         int level;
222         u64 *prev_shadow_ent = NULL;
223         pt_element_t *guest_ent = walker->ptep;
224
225         if (!is_present_pte(*guest_ent))
226                 return NULL;
227
228         shadow_addr = vcpu->mmu.root_hpa;
229         level = vcpu->mmu.shadow_root_level;
230         if (level == PT32E_ROOT_LEVEL) {
231                 shadow_addr = vcpu->mmu.pae_root[(addr >> 30) & 3];
232                 shadow_addr &= PT64_BASE_ADDR_MASK;
233                 --level;
234         }
235
236         for (; ; level--) {
237                 u32 index = SHADOW_PT_INDEX(addr, level);
238                 u64 *shadow_ent = ((u64 *)__va(shadow_addr)) + index;
239                 struct kvm_mmu_page *shadow_page;
240                 u64 shadow_pte;
241                 int metaphysical;
242                 gfn_t table_gfn;
243
244                 if (is_present_pte(*shadow_ent) || is_io_pte(*shadow_ent)) {
245                         if (level == PT_PAGE_TABLE_LEVEL)
246                                 return shadow_ent;
247                         shadow_addr = *shadow_ent & PT64_BASE_ADDR_MASK;
248                         prev_shadow_ent = shadow_ent;
249                         continue;
250                 }
251
252                 if (level == PT_PAGE_TABLE_LEVEL) {
253
254                         if (walker->level == PT_DIRECTORY_LEVEL) {
255                                 if (prev_shadow_ent)
256                                         *prev_shadow_ent |= PT_SHADOW_PS_MARK;
257                                 FNAME(set_pde)(vcpu, *guest_ent, shadow_ent,
258                                                walker->inherited_ar,
259                                                walker->gfn);
260                         } else {
261                                 ASSERT(walker->level == PT_PAGE_TABLE_LEVEL);
262                                 FNAME(set_pte)(vcpu, *guest_ent, shadow_ent,
263                                                walker->inherited_ar,
264                                                walker->gfn);
265                         }
266                         return shadow_ent;
267                 }
268
269                 if (level - 1 == PT_PAGE_TABLE_LEVEL
270                     && walker->level == PT_DIRECTORY_LEVEL) {
271                         metaphysical = 1;
272                         table_gfn = (*guest_ent & PT_BASE_ADDR_MASK)
273                                 >> PAGE_SHIFT;
274                 } else {
275                         metaphysical = 0;
276                         table_gfn = walker->table_gfn[level - 2];
277                 }
278                 shadow_page = kvm_mmu_get_page(vcpu, table_gfn, addr, level-1,
279                                                metaphysical, shadow_ent);
280                 shadow_addr = shadow_page->page_hpa;
281                 shadow_pte = shadow_addr | PT_PRESENT_MASK | PT_ACCESSED_MASK
282                         | PT_WRITABLE_MASK | PT_USER_MASK;
283                 *shadow_ent = shadow_pte;
284                 prev_shadow_ent = shadow_ent;
285         }
286 }
287
288 /*
289  * The guest faulted for write.  We need to
290  *
291  * - check write permissions
292  * - update the guest pte dirty bit
293  * - update our own dirty page tracking structures
294  */
295 static int FNAME(fix_write_pf)(struct kvm_vcpu *vcpu,
296                                u64 *shadow_ent,
297                                struct guest_walker *walker,
298                                gva_t addr,
299                                int user,
300                                int *write_pt)
301 {
302         pt_element_t *guest_ent;
303         int writable_shadow;
304         gfn_t gfn;
305         struct kvm_mmu_page *page;
306
307         if (is_writeble_pte(*shadow_ent))
308                 return !user || (*shadow_ent & PT_USER_MASK);
309
310         writable_shadow = *shadow_ent & PT_SHADOW_WRITABLE_MASK;
311         if (user) {
312                 /*
313                  * User mode access.  Fail if it's a kernel page or a read-only
314                  * page.
315                  */
316                 if (!(*shadow_ent & PT_SHADOW_USER_MASK) || !writable_shadow)
317                         return 0;
318                 ASSERT(*shadow_ent & PT_USER_MASK);
319         } else
320                 /*
321                  * Kernel mode access.  Fail if it's a read-only page and
322                  * supervisor write protection is enabled.
323                  */
324                 if (!writable_shadow) {
325                         if (is_write_protection(vcpu))
326                                 return 0;
327                         *shadow_ent &= ~PT_USER_MASK;
328                 }
329
330         guest_ent = walker->ptep;
331
332         if (!is_present_pte(*guest_ent)) {
333                 *shadow_ent = 0;
334                 return 0;
335         }
336
337         gfn = walker->gfn;
338
339         if (user) {
340                 /*
341                  * Usermode page faults won't be for page table updates.
342                  */
343                 while ((page = kvm_mmu_lookup_page(vcpu, gfn)) != NULL) {
344                         pgprintk("%s: zap %lx %x\n",
345                                  __FUNCTION__, gfn, page->role.word);
346                         kvm_mmu_zap_page(vcpu, page);
347                 }
348         } else if (kvm_mmu_lookup_page(vcpu, gfn)) {
349                 pgprintk("%s: found shadow page for %lx, marking ro\n",
350                          __FUNCTION__, gfn);
351                 *guest_ent |= PT_DIRTY_MASK;
352                 *write_pt = 1;
353                 return 0;
354         }
355         mark_page_dirty(vcpu->kvm, gfn);
356         *shadow_ent |= PT_WRITABLE_MASK;
357         *guest_ent |= PT_DIRTY_MASK;
358         rmap_add(vcpu, shadow_ent);
359
360         return 1;
361 }
362
363 /*
364  * Page fault handler.  There are several causes for a page fault:
365  *   - there is no shadow pte for the guest pte
366  *   - write access through a shadow pte marked read only so that we can set
367  *     the dirty bit
368  *   - write access to a shadow pte marked read only so we can update the page
369  *     dirty bitmap, when userspace requests it
370  *   - mmio access; in this case we will never install a present shadow pte
371  *   - normal guest page fault due to the guest pte marked not present, not
372  *     writable, or not executable
373  *
374  *  Returns: 1 if we need to emulate the instruction, 0 otherwise, or
375  *           a negative value on error.
376  */
377 static int FNAME(page_fault)(struct kvm_vcpu *vcpu, gva_t addr,
378                                u32 error_code)
379 {
380         int write_fault = error_code & PFERR_WRITE_MASK;
381         int user_fault = error_code & PFERR_USER_MASK;
382         int fetch_fault = error_code & PFERR_FETCH_MASK;
383         struct guest_walker walker;
384         u64 *shadow_pte;
385         int fixed;
386         int write_pt = 0;
387         int r;
388
389         pgprintk("%s: addr %lx err %x\n", __FUNCTION__, addr, error_code);
390         kvm_mmu_audit(vcpu, "pre page fault");
391
392         r = mmu_topup_memory_caches(vcpu);
393         if (r)
394                 return r;
395
396         /*
397          * Look up the shadow pte for the faulting address.
398          */
399         r = FNAME(walk_addr)(&walker, vcpu, addr, write_fault, user_fault,
400                              fetch_fault);
401
402         /*
403          * The page is not mapped by the guest.  Let the guest handle it.
404          */
405         if (!r) {
406                 pgprintk("%s: guest page fault\n", __FUNCTION__);
407                 inject_page_fault(vcpu, addr, walker.error_code);
408                 FNAME(release_walker)(&walker);
409                 return 0;
410         }
411
412         shadow_pte = FNAME(fetch)(vcpu, addr, &walker);
413         pgprintk("%s: shadow pte %p %llx\n", __FUNCTION__,
414                  shadow_pte, *shadow_pte);
415
416         /*
417          * Update the shadow pte.
418          */
419         if (write_fault)
420                 fixed = FNAME(fix_write_pf)(vcpu, shadow_pte, &walker, addr,
421                                             user_fault, &write_pt);
422         else
423                 fixed = fix_read_pf(shadow_pte);
424
425         pgprintk("%s: updated shadow pte %p %llx\n", __FUNCTION__,
426                  shadow_pte, *shadow_pte);
427
428         FNAME(release_walker)(&walker);
429
430         /*
431          * mmio: emulate if accessible, otherwise its a guest fault.
432          */
433         if (is_io_pte(*shadow_pte)) {
434                 return 1;
435         }
436
437         ++kvm_stat.pf_fixed;
438         kvm_mmu_audit(vcpu, "post page fault (fixed)");
439
440         return write_pt;
441 }
442
443 static gpa_t FNAME(gva_to_gpa)(struct kvm_vcpu *vcpu, gva_t vaddr)
444 {
445         struct guest_walker walker;
446         gpa_t gpa = UNMAPPED_GVA;
447         int r;
448
449         r = FNAME(walk_addr)(&walker, vcpu, vaddr, 0, 0, 0);
450
451         if (r) {
452                 gpa = (gpa_t)walker.gfn << PAGE_SHIFT;
453                 gpa |= vaddr & ~PAGE_MASK;
454         }
455
456         FNAME(release_walker)(&walker);
457         return gpa;
458 }
459
460 #undef pt_element_t
461 #undef guest_walker
462 #undef FNAME
463 #undef PT_BASE_ADDR_MASK
464 #undef PT_INDEX
465 #undef SHADOW_PT_INDEX
466 #undef PT_LEVEL_MASK
467 #undef PT_PTE_COPY_MASK
468 #undef PT_NON_PTE_COPY_MASK
469 #undef PT_DIR_BASE_ADDR_MASK
470 #undef PT_MAX_FULL_LEVELS