7ed8cc3434c5f91efb35acd4cebf09923849107d
[~shefty/rdma-dev.git] / arch / x86 / xen / p2m.c
1 /*
2  * Xen leaves the responsibility for maintaining p2m mappings to the
3  * guests themselves, but it must also access and update the p2m array
4  * during suspend/resume when all the pages are reallocated.
5  *
6  * The p2m table is logically a flat array, but we implement it as a
7  * three-level tree to allow the address space to be sparse.
8  *
9  *                               Xen
10  *                                |
11  *     p2m_top              p2m_top_mfn
12  *       /  \                   /   \
13  * p2m_mid p2m_mid      p2m_mid_mfn p2m_mid_mfn
14  *    / \      / \         /           /
15  *  p2m p2m p2m p2m p2m p2m p2m ...
16  *
17  * The p2m_mid_mfn pages are mapped by p2m_top_mfn_p.
18  *
19  * The p2m_top and p2m_top_mfn levels are limited to 1 page, so the
20  * maximum representable pseudo-physical address space is:
21  *  P2M_TOP_PER_PAGE * P2M_MID_PER_PAGE * P2M_PER_PAGE pages
22  *
23  * P2M_PER_PAGE depends on the architecture, as a mfn is always
24  * unsigned long (8 bytes on 64-bit, 4 bytes on 32), leading to
25  * 512 and 1024 entries respectively. 
26  *
27  * In short, these structures contain the Machine Frame Number (MFN) of the PFN.
28  *
29  * However not all entries are filled with MFNs. Specifically for all other
30  * leaf entries, or for the top  root, or middle one, for which there is a void
31  * entry, we assume it is  "missing". So (for example)
32  *  pfn_to_mfn(0x90909090)=INVALID_P2M_ENTRY.
33  *
34  * We also have the possibility of setting 1-1 mappings on certain regions, so
35  * that:
36  *  pfn_to_mfn(0xc0000)=0xc0000
37  *
38  * The benefit of this is, that we can assume for non-RAM regions (think
39  * PCI BARs, or ACPI spaces), we can create mappings easily b/c we
40  * get the PFN value to match the MFN.
41  *
42  * For this to work efficiently we have one new page p2m_identity and
43  * allocate (via reserved_brk) any other pages we need to cover the sides
44  * (1GB or 4MB boundary violations). All entries in p2m_identity are set to
45  * INVALID_P2M_ENTRY type (Xen toolstack only recognizes that and MFNs,
46  * no other fancy value).
47  *
48  * On lookup we spot that the entry points to p2m_identity and return the
49  * identity value instead of dereferencing and returning INVALID_P2M_ENTRY.
50  * If the entry points to an allocated page, we just proceed as before and
51  * return the PFN.  If the PFN has IDENTITY_FRAME_BIT set we unmask that in
52  * appropriate functions (pfn_to_mfn).
53  *
54  * The reason for having the IDENTITY_FRAME_BIT instead of just returning the
55  * PFN is that we could find ourselves where pfn_to_mfn(pfn)==pfn for a
56  * non-identity pfn. To protect ourselves against we elect to set (and get) the
57  * IDENTITY_FRAME_BIT on all identity mapped PFNs.
58  *
59  * This simplistic diagram is used to explain the more subtle piece of code.
60  * There is also a digram of the P2M at the end that can help.
61  * Imagine your E820 looking as so:
62  *
63  *                    1GB                                           2GB
64  * /-------------------+---------\/----\         /----------\    /---+-----\
65  * | System RAM        | Sys RAM ||ACPI|         | reserved |    | Sys RAM |
66  * \-------------------+---------/\----/         \----------/    \---+-----/
67  *                               ^- 1029MB                       ^- 2001MB
68  *
69  * [1029MB = 263424 (0x40500), 2001MB = 512256 (0x7D100),
70  *  2048MB = 524288 (0x80000)]
71  *
72  * And dom0_mem=max:3GB,1GB is passed in to the guest, meaning memory past 1GB
73  * is actually not present (would have to kick the balloon driver to put it in).
74  *
75  * When we are told to set the PFNs for identity mapping (see patch: "xen/setup:
76  * Set identity mapping for non-RAM E820 and E820 gaps.") we pass in the start
77  * of the PFN and the end PFN (263424 and 512256 respectively). The first step
78  * is to reserve_brk a top leaf page if the p2m[1] is missing. The top leaf page
79  * covers 512^2 of page estate (1GB) and in case the start or end PFN is not
80  * aligned on 512^2*PAGE_SIZE (1GB) we loop on aligned 1GB PFNs from start pfn
81  * to end pfn.  We reserve_brk top leaf pages if they are missing (means they
82  * point to p2m_mid_missing).
83  *
84  * With the E820 example above, 263424 is not 1GB aligned so we allocate a
85  * reserve_brk page which will cover the PFNs estate from 0x40000 to 0x80000.
86  * Each entry in the allocate page is "missing" (points to p2m_missing).
87  *
88  * Next stage is to determine if we need to do a more granular boundary check
89  * on the 4MB (or 2MB depending on architecture) off the start and end pfn's.
90  * We check if the start pfn and end pfn violate that boundary check, and if
91  * so reserve_brk a middle (p2m[x][y]) leaf page. This way we have a much finer
92  * granularity of setting which PFNs are missing and which ones are identity.
93  * In our example 263424 and 512256 both fail the check so we reserve_brk two
94  * pages. Populate them with INVALID_P2M_ENTRY (so they both have "missing"
95  * values) and assign them to p2m[1][2] and p2m[1][488] respectively.
96  *
97  * At this point we would at minimum reserve_brk one page, but could be up to
98  * three. Each call to set_phys_range_identity has at maximum a three page
99  * cost. If we were to query the P2M at this stage, all those entries from
100  * start PFN through end PFN (so 1029MB -> 2001MB) would return
101  * INVALID_P2M_ENTRY ("missing").
102  *
103  * The next step is to walk from the start pfn to the end pfn setting
104  * the IDENTITY_FRAME_BIT on each PFN. This is done in set_phys_range_identity.
105  * If we find that the middle leaf is pointing to p2m_missing we can swap it
106  * over to p2m_identity - this way covering 4MB (or 2MB) PFN space.  At this
107  * point we do not need to worry about boundary aligment (so no need to
108  * reserve_brk a middle page, figure out which PFNs are "missing" and which
109  * ones are identity), as that has been done earlier.  If we find that the
110  * middle leaf is not occupied by p2m_identity or p2m_missing, we dereference
111  * that page (which covers 512 PFNs) and set the appropriate PFN with
112  * IDENTITY_FRAME_BIT. In our example 263424 and 512256 end up there, and we
113  * set from p2m[1][2][256->511] and p2m[1][488][0->256] with
114  * IDENTITY_FRAME_BIT set.
115  *
116  * All other regions that are void (or not filled) either point to p2m_missing
117  * (considered missing) or have the default value of INVALID_P2M_ENTRY (also
118  * considered missing). In our case, p2m[1][2][0->255] and p2m[1][488][257->511]
119  * contain the INVALID_P2M_ENTRY value and are considered "missing."
120  *
121  * This is what the p2m ends up looking (for the E820 above) with this
122  * fabulous drawing:
123  *
124  *    p2m         /--------------\
125  *  /-----\       | &mfn_list[0],|                           /-----------------\
126  *  |  0  |------>| &mfn_list[1],|    /---------------\      | ~0, ~0, ..      |
127  *  |-----|       |  ..., ~0, ~0 |    | ~0, ~0, [x]---+----->| IDENTITY [@256] |
128  *  |  1  |---\   \--------------/    | [p2m_identity]+\     | IDENTITY [@257] |
129  *  |-----|    \                      | [p2m_identity]+\\    | ....            |
130  *  |  2  |--\  \-------------------->|  ...          | \\   \----------------/
131  *  |-----|   \                       \---------------/  \\
132  *  |  3  |\   \                                          \\  p2m_identity
133  *  |-----| \   \-------------------->/---------------\   /-----------------\
134  *  | ..  +->+                        | [p2m_identity]+-->| ~0, ~0, ~0, ... |
135  *  \-----/ /                         | [p2m_identity]+-->| ..., ~0         |
136  *         / /---------------\        | ....          |   \-----------------/
137  *        /  | IDENTITY[@0]  |      /-+-[x], ~0, ~0.. |
138  *       /   | IDENTITY[@256]|<----/  \---------------/
139  *      /    | ~0, ~0, ....  |
140  *     |     \---------------/
141  *     |
142  *     p2m_missing             p2m_missing
143  * /------------------\     /------------\
144  * | [p2m_mid_missing]+---->| ~0, ~0, ~0 |
145  * | [p2m_mid_missing]+---->| ..., ~0    |
146  * \------------------/     \------------/
147  *
148  * where ~0 is INVALID_P2M_ENTRY. IDENTITY is (PFN | IDENTITY_BIT)
149  */
150
151 #include <linux/init.h>
152 #include <linux/module.h>
153 #include <linux/list.h>
154 #include <linux/hash.h>
155 #include <linux/sched.h>
156 #include <linux/seq_file.h>
157
158 #include <asm/cache.h>
159 #include <asm/setup.h>
160
161 #include <asm/xen/page.h>
162 #include <asm/xen/hypercall.h>
163 #include <asm/xen/hypervisor.h>
164 #include <xen/grant_table.h>
165
166 #include "multicalls.h"
167 #include "xen-ops.h"
168
169 static void __init m2p_override_init(void);
170
171 unsigned long xen_max_p2m_pfn __read_mostly;
172
173 #define P2M_PER_PAGE            (PAGE_SIZE / sizeof(unsigned long))
174 #define P2M_MID_PER_PAGE        (PAGE_SIZE / sizeof(unsigned long *))
175 #define P2M_TOP_PER_PAGE        (PAGE_SIZE / sizeof(unsigned long **))
176
177 #define MAX_P2M_PFN             (P2M_TOP_PER_PAGE * P2M_MID_PER_PAGE * P2M_PER_PAGE)
178
179 /* Placeholders for holes in the address space */
180 static RESERVE_BRK_ARRAY(unsigned long, p2m_missing, P2M_PER_PAGE);
181 static RESERVE_BRK_ARRAY(unsigned long *, p2m_mid_missing, P2M_MID_PER_PAGE);
182 static RESERVE_BRK_ARRAY(unsigned long, p2m_mid_missing_mfn, P2M_MID_PER_PAGE);
183
184 static RESERVE_BRK_ARRAY(unsigned long **, p2m_top, P2M_TOP_PER_PAGE);
185 static RESERVE_BRK_ARRAY(unsigned long, p2m_top_mfn, P2M_TOP_PER_PAGE);
186 static RESERVE_BRK_ARRAY(unsigned long *, p2m_top_mfn_p, P2M_TOP_PER_PAGE);
187
188 static RESERVE_BRK_ARRAY(unsigned long, p2m_identity, P2M_PER_PAGE);
189
190 RESERVE_BRK(p2m_mid, PAGE_SIZE * (MAX_DOMAIN_PAGES / (P2M_PER_PAGE * P2M_MID_PER_PAGE)));
191 RESERVE_BRK(p2m_mid_mfn, PAGE_SIZE * (MAX_DOMAIN_PAGES / (P2M_PER_PAGE * P2M_MID_PER_PAGE)));
192
193 /* We might hit two boundary violations at the start and end, at max each
194  * boundary violation will require three middle nodes. */
195 RESERVE_BRK(p2m_mid_identity, PAGE_SIZE * 2 * 3);
196
197 static inline unsigned p2m_top_index(unsigned long pfn)
198 {
199         BUG_ON(pfn >= MAX_P2M_PFN);
200         return pfn / (P2M_MID_PER_PAGE * P2M_PER_PAGE);
201 }
202
203 static inline unsigned p2m_mid_index(unsigned long pfn)
204 {
205         return (pfn / P2M_PER_PAGE) % P2M_MID_PER_PAGE;
206 }
207
208 static inline unsigned p2m_index(unsigned long pfn)
209 {
210         return pfn % P2M_PER_PAGE;
211 }
212
213 static void p2m_top_init(unsigned long ***top)
214 {
215         unsigned i;
216
217         for (i = 0; i < P2M_TOP_PER_PAGE; i++)
218                 top[i] = p2m_mid_missing;
219 }
220
221 static void p2m_top_mfn_init(unsigned long *top)
222 {
223         unsigned i;
224
225         for (i = 0; i < P2M_TOP_PER_PAGE; i++)
226                 top[i] = virt_to_mfn(p2m_mid_missing_mfn);
227 }
228
229 static void p2m_top_mfn_p_init(unsigned long **top)
230 {
231         unsigned i;
232
233         for (i = 0; i < P2M_TOP_PER_PAGE; i++)
234                 top[i] = p2m_mid_missing_mfn;
235 }
236
237 static void p2m_mid_init(unsigned long **mid)
238 {
239         unsigned i;
240
241         for (i = 0; i < P2M_MID_PER_PAGE; i++)
242                 mid[i] = p2m_missing;
243 }
244
245 static void p2m_mid_mfn_init(unsigned long *mid)
246 {
247         unsigned i;
248
249         for (i = 0; i < P2M_MID_PER_PAGE; i++)
250                 mid[i] = virt_to_mfn(p2m_missing);
251 }
252
253 static void p2m_init(unsigned long *p2m)
254 {
255         unsigned i;
256
257         for (i = 0; i < P2M_MID_PER_PAGE; i++)
258                 p2m[i] = INVALID_P2M_ENTRY;
259 }
260
261 /*
262  * Build the parallel p2m_top_mfn and p2m_mid_mfn structures
263  *
264  * This is called both at boot time, and after resuming from suspend:
265  * - At boot time we're called very early, and must use extend_brk()
266  *   to allocate memory.
267  *
268  * - After resume we're called from within stop_machine, but the mfn
269  *   tree should alreay be completely allocated.
270  */
271 void __ref xen_build_mfn_list_list(void)
272 {
273         unsigned long pfn;
274
275         /* Pre-initialize p2m_top_mfn to be completely missing */
276         if (p2m_top_mfn == NULL) {
277                 p2m_mid_missing_mfn = extend_brk(PAGE_SIZE, PAGE_SIZE);
278                 p2m_mid_mfn_init(p2m_mid_missing_mfn);
279
280                 p2m_top_mfn_p = extend_brk(PAGE_SIZE, PAGE_SIZE);
281                 p2m_top_mfn_p_init(p2m_top_mfn_p);
282
283                 p2m_top_mfn = extend_brk(PAGE_SIZE, PAGE_SIZE);
284                 p2m_top_mfn_init(p2m_top_mfn);
285         } else {
286                 /* Reinitialise, mfn's all change after migration */
287                 p2m_mid_mfn_init(p2m_mid_missing_mfn);
288         }
289
290         for (pfn = 0; pfn < xen_max_p2m_pfn; pfn += P2M_PER_PAGE) {
291                 unsigned topidx = p2m_top_index(pfn);
292                 unsigned mididx = p2m_mid_index(pfn);
293                 unsigned long **mid;
294                 unsigned long *mid_mfn_p;
295
296                 mid = p2m_top[topidx];
297                 mid_mfn_p = p2m_top_mfn_p[topidx];
298
299                 /* Don't bother allocating any mfn mid levels if
300                  * they're just missing, just update the stored mfn,
301                  * since all could have changed over a migrate.
302                  */
303                 if (mid == p2m_mid_missing) {
304                         BUG_ON(mididx);
305                         BUG_ON(mid_mfn_p != p2m_mid_missing_mfn);
306                         p2m_top_mfn[topidx] = virt_to_mfn(p2m_mid_missing_mfn);
307                         pfn += (P2M_MID_PER_PAGE - 1) * P2M_PER_PAGE;
308                         continue;
309                 }
310
311                 if (mid_mfn_p == p2m_mid_missing_mfn) {
312                         /*
313                          * XXX boot-time only!  We should never find
314                          * missing parts of the mfn tree after
315                          * runtime.  extend_brk() will BUG if we call
316                          * it too late.
317                          */
318                         mid_mfn_p = extend_brk(PAGE_SIZE, PAGE_SIZE);
319                         p2m_mid_mfn_init(mid_mfn_p);
320
321                         p2m_top_mfn_p[topidx] = mid_mfn_p;
322                 }
323
324                 p2m_top_mfn[topidx] = virt_to_mfn(mid_mfn_p);
325                 mid_mfn_p[mididx] = virt_to_mfn(mid[mididx]);
326         }
327 }
328
329 void xen_setup_mfn_list_list(void)
330 {
331         BUG_ON(HYPERVISOR_shared_info == &xen_dummy_shared_info);
332
333         HYPERVISOR_shared_info->arch.pfn_to_mfn_frame_list_list =
334                 virt_to_mfn(p2m_top_mfn);
335         HYPERVISOR_shared_info->arch.max_pfn = xen_max_p2m_pfn;
336 }
337
338 /* Set up p2m_top to point to the domain-builder provided p2m pages */
339 void __init xen_build_dynamic_phys_to_machine(void)
340 {
341         unsigned long *mfn_list = (unsigned long *)xen_start_info->mfn_list;
342         unsigned long max_pfn = min(MAX_DOMAIN_PAGES, xen_start_info->nr_pages);
343         unsigned long pfn;
344
345         xen_max_p2m_pfn = max_pfn;
346
347         p2m_missing = extend_brk(PAGE_SIZE, PAGE_SIZE);
348         p2m_init(p2m_missing);
349
350         p2m_mid_missing = extend_brk(PAGE_SIZE, PAGE_SIZE);
351         p2m_mid_init(p2m_mid_missing);
352
353         p2m_top = extend_brk(PAGE_SIZE, PAGE_SIZE);
354         p2m_top_init(p2m_top);
355
356         p2m_identity = extend_brk(PAGE_SIZE, PAGE_SIZE);
357         p2m_init(p2m_identity);
358
359         /*
360          * The domain builder gives us a pre-constructed p2m array in
361          * mfn_list for all the pages initially given to us, so we just
362          * need to graft that into our tree structure.
363          */
364         for (pfn = 0; pfn < max_pfn; pfn += P2M_PER_PAGE) {
365                 unsigned topidx = p2m_top_index(pfn);
366                 unsigned mididx = p2m_mid_index(pfn);
367
368                 if (p2m_top[topidx] == p2m_mid_missing) {
369                         unsigned long **mid = extend_brk(PAGE_SIZE, PAGE_SIZE);
370                         p2m_mid_init(mid);
371
372                         p2m_top[topidx] = mid;
373                 }
374
375                 /*
376                  * As long as the mfn_list has enough entries to completely
377                  * fill a p2m page, pointing into the array is ok. But if
378                  * not the entries beyond the last pfn will be undefined.
379                  */
380                 if (unlikely(pfn + P2M_PER_PAGE > max_pfn)) {
381                         unsigned long p2midx;
382
383                         p2midx = max_pfn % P2M_PER_PAGE;
384                         for ( ; p2midx < P2M_PER_PAGE; p2midx++)
385                                 mfn_list[pfn + p2midx] = INVALID_P2M_ENTRY;
386                 }
387                 p2m_top[topidx][mididx] = &mfn_list[pfn];
388         }
389
390         m2p_override_init();
391 }
392
393 unsigned long get_phys_to_machine(unsigned long pfn)
394 {
395         unsigned topidx, mididx, idx;
396
397         if (unlikely(pfn >= MAX_P2M_PFN))
398                 return INVALID_P2M_ENTRY;
399
400         topidx = p2m_top_index(pfn);
401         mididx = p2m_mid_index(pfn);
402         idx = p2m_index(pfn);
403
404         /*
405          * The INVALID_P2M_ENTRY is filled in both p2m_*identity
406          * and in p2m_*missing, so returning the INVALID_P2M_ENTRY
407          * would be wrong.
408          */
409         if (p2m_top[topidx][mididx] == p2m_identity)
410                 return IDENTITY_FRAME(pfn);
411
412         return p2m_top[topidx][mididx][idx];
413 }
414 EXPORT_SYMBOL_GPL(get_phys_to_machine);
415
416 static void *alloc_p2m_page(void)
417 {
418         return (void *)__get_free_page(GFP_KERNEL | __GFP_REPEAT);
419 }
420
421 static void free_p2m_page(void *p)
422 {
423         free_page((unsigned long)p);
424 }
425
426 /* 
427  * Fully allocate the p2m structure for a given pfn.  We need to check
428  * that both the top and mid levels are allocated, and make sure the
429  * parallel mfn tree is kept in sync.  We may race with other cpus, so
430  * the new pages are installed with cmpxchg; if we lose the race then
431  * simply free the page we allocated and use the one that's there.
432  */
433 static bool alloc_p2m(unsigned long pfn)
434 {
435         unsigned topidx, mididx;
436         unsigned long ***top_p, **mid;
437         unsigned long *top_mfn_p, *mid_mfn;
438
439         topidx = p2m_top_index(pfn);
440         mididx = p2m_mid_index(pfn);
441
442         top_p = &p2m_top[topidx];
443         mid = *top_p;
444
445         if (mid == p2m_mid_missing) {
446                 /* Mid level is missing, allocate a new one */
447                 mid = alloc_p2m_page();
448                 if (!mid)
449                         return false;
450
451                 p2m_mid_init(mid);
452
453                 if (cmpxchg(top_p, p2m_mid_missing, mid) != p2m_mid_missing)
454                         free_p2m_page(mid);
455         }
456
457         top_mfn_p = &p2m_top_mfn[topidx];
458         mid_mfn = p2m_top_mfn_p[topidx];
459
460         BUG_ON(virt_to_mfn(mid_mfn) != *top_mfn_p);
461
462         if (mid_mfn == p2m_mid_missing_mfn) {
463                 /* Separately check the mid mfn level */
464                 unsigned long missing_mfn;
465                 unsigned long mid_mfn_mfn;
466
467                 mid_mfn = alloc_p2m_page();
468                 if (!mid_mfn)
469                         return false;
470
471                 p2m_mid_mfn_init(mid_mfn);
472
473                 missing_mfn = virt_to_mfn(p2m_mid_missing_mfn);
474                 mid_mfn_mfn = virt_to_mfn(mid_mfn);
475                 if (cmpxchg(top_mfn_p, missing_mfn, mid_mfn_mfn) != missing_mfn)
476                         free_p2m_page(mid_mfn);
477                 else
478                         p2m_top_mfn_p[topidx] = mid_mfn;
479         }
480
481         if (p2m_top[topidx][mididx] == p2m_identity ||
482             p2m_top[topidx][mididx] == p2m_missing) {
483                 /* p2m leaf page is missing */
484                 unsigned long *p2m;
485                 unsigned long *p2m_orig = p2m_top[topidx][mididx];
486
487                 p2m = alloc_p2m_page();
488                 if (!p2m)
489                         return false;
490
491                 p2m_init(p2m);
492
493                 if (cmpxchg(&mid[mididx], p2m_orig, p2m) != p2m_orig)
494                         free_p2m_page(p2m);
495                 else
496                         mid_mfn[mididx] = virt_to_mfn(p2m);
497         }
498
499         return true;
500 }
501
502 static bool __init __early_alloc_p2m(unsigned long pfn)
503 {
504         unsigned topidx, mididx, idx;
505
506         topidx = p2m_top_index(pfn);
507         mididx = p2m_mid_index(pfn);
508         idx = p2m_index(pfn);
509
510         /* Pfff.. No boundary cross-over, lets get out. */
511         if (!idx)
512                 return false;
513
514         WARN(p2m_top[topidx][mididx] == p2m_identity,
515                 "P2M[%d][%d] == IDENTITY, should be MISSING (or alloced)!\n",
516                 topidx, mididx);
517
518         /*
519          * Could be done by xen_build_dynamic_phys_to_machine..
520          */
521         if (p2m_top[topidx][mididx] != p2m_missing)
522                 return false;
523
524         /* Boundary cross-over for the edges: */
525         if (idx) {
526                 unsigned long *p2m = extend_brk(PAGE_SIZE, PAGE_SIZE);
527                 unsigned long *mid_mfn_p;
528
529                 p2m_init(p2m);
530
531                 p2m_top[topidx][mididx] = p2m;
532
533                 /* For save/restore we need to MFN of the P2M saved */
534                 
535                 mid_mfn_p = p2m_top_mfn_p[topidx];
536                 WARN(mid_mfn_p[mididx] != virt_to_mfn(p2m_missing),
537                         "P2M_TOP_P[%d][%d] != MFN of p2m_missing!\n",
538                         topidx, mididx);
539                 mid_mfn_p[mididx] = virt_to_mfn(p2m);
540
541         }
542         return idx != 0;
543 }
544 unsigned long __init set_phys_range_identity(unsigned long pfn_s,
545                                       unsigned long pfn_e)
546 {
547         unsigned long pfn;
548
549         if (unlikely(pfn_s >= MAX_P2M_PFN || pfn_e >= MAX_P2M_PFN))
550                 return 0;
551
552         if (unlikely(xen_feature(XENFEAT_auto_translated_physmap)))
553                 return pfn_e - pfn_s;
554
555         if (pfn_s > pfn_e)
556                 return 0;
557
558         for (pfn = (pfn_s & ~(P2M_MID_PER_PAGE * P2M_PER_PAGE - 1));
559                 pfn < ALIGN(pfn_e, (P2M_MID_PER_PAGE * P2M_PER_PAGE));
560                 pfn += P2M_MID_PER_PAGE * P2M_PER_PAGE)
561         {
562                 unsigned topidx = p2m_top_index(pfn);
563                 unsigned long *mid_mfn_p;
564                 unsigned long **mid;
565
566                 mid = p2m_top[topidx];
567                 mid_mfn_p = p2m_top_mfn_p[topidx];
568                 if (mid == p2m_mid_missing) {
569                         mid = extend_brk(PAGE_SIZE, PAGE_SIZE);
570
571                         p2m_mid_init(mid);
572
573                         p2m_top[topidx] = mid;
574
575                         BUG_ON(mid_mfn_p != p2m_mid_missing_mfn);
576                 }
577                 /* And the save/restore P2M tables.. */
578                 if (mid_mfn_p == p2m_mid_missing_mfn) {
579                         mid_mfn_p = extend_brk(PAGE_SIZE, PAGE_SIZE);
580                         p2m_mid_mfn_init(mid_mfn_p);
581
582                         p2m_top_mfn_p[topidx] = mid_mfn_p;
583                         p2m_top_mfn[topidx] = virt_to_mfn(mid_mfn_p);
584                         /* Note: we don't set mid_mfn_p[midix] here,
585                          * look in __early_alloc_p2m */
586                 }
587         }
588
589         __early_alloc_p2m(pfn_s);
590         __early_alloc_p2m(pfn_e);
591
592         for (pfn = pfn_s; pfn < pfn_e; pfn++)
593                 if (!__set_phys_to_machine(pfn, IDENTITY_FRAME(pfn)))
594                         break;
595
596         if (!WARN((pfn - pfn_s) != (pfn_e - pfn_s),
597                 "Identity mapping failed. We are %ld short of 1-1 mappings!\n",
598                 (pfn_e - pfn_s) - (pfn - pfn_s)))
599                 printk(KERN_DEBUG "1-1 mapping on %lx->%lx\n", pfn_s, pfn);
600
601         return pfn - pfn_s;
602 }
603
604 /* Try to install p2m mapping; fail if intermediate bits missing */
605 bool __set_phys_to_machine(unsigned long pfn, unsigned long mfn)
606 {
607         unsigned topidx, mididx, idx;
608
609         if (unlikely(xen_feature(XENFEAT_auto_translated_physmap))) {
610                 BUG_ON(pfn != mfn && mfn != INVALID_P2M_ENTRY);
611                 return true;
612         }
613         if (unlikely(pfn >= MAX_P2M_PFN)) {
614                 BUG_ON(mfn != INVALID_P2M_ENTRY);
615                 return true;
616         }
617
618         topidx = p2m_top_index(pfn);
619         mididx = p2m_mid_index(pfn);
620         idx = p2m_index(pfn);
621
622         /* For sparse holes were the p2m leaf has real PFN along with
623          * PCI holes, stick in the PFN as the MFN value.
624          */
625         if (mfn != INVALID_P2M_ENTRY && (mfn & IDENTITY_FRAME_BIT)) {
626                 if (p2m_top[topidx][mididx] == p2m_identity)
627                         return true;
628
629                 /* Swap over from MISSING to IDENTITY if needed. */
630                 if (p2m_top[topidx][mididx] == p2m_missing) {
631                         WARN_ON(cmpxchg(&p2m_top[topidx][mididx], p2m_missing,
632                                 p2m_identity) != p2m_missing);
633                         return true;
634                 }
635         }
636
637         if (p2m_top[topidx][mididx] == p2m_missing)
638                 return mfn == INVALID_P2M_ENTRY;
639
640         p2m_top[topidx][mididx][idx] = mfn;
641
642         return true;
643 }
644
645 bool set_phys_to_machine(unsigned long pfn, unsigned long mfn)
646 {
647         if (unlikely(!__set_phys_to_machine(pfn, mfn)))  {
648                 if (!alloc_p2m(pfn))
649                         return false;
650
651                 if (!__set_phys_to_machine(pfn, mfn))
652                         return false;
653         }
654
655         return true;
656 }
657
658 #define M2P_OVERRIDE_HASH_SHIFT 10
659 #define M2P_OVERRIDE_HASH       (1 << M2P_OVERRIDE_HASH_SHIFT)
660
661 static RESERVE_BRK_ARRAY(struct list_head, m2p_overrides, M2P_OVERRIDE_HASH);
662 static DEFINE_SPINLOCK(m2p_override_lock);
663
664 static void __init m2p_override_init(void)
665 {
666         unsigned i;
667
668         m2p_overrides = extend_brk(sizeof(*m2p_overrides) * M2P_OVERRIDE_HASH,
669                                    sizeof(unsigned long));
670
671         for (i = 0; i < M2P_OVERRIDE_HASH; i++)
672                 INIT_LIST_HEAD(&m2p_overrides[i]);
673 }
674
675 static unsigned long mfn_hash(unsigned long mfn)
676 {
677         return hash_long(mfn, M2P_OVERRIDE_HASH_SHIFT);
678 }
679
680 /* Add an MFN override for a particular page */
681 int m2p_add_override(unsigned long mfn, struct page *page,
682                 struct gnttab_map_grant_ref *kmap_op)
683 {
684         unsigned long flags;
685         unsigned long pfn;
686         unsigned long uninitialized_var(address);
687         unsigned level;
688         pte_t *ptep = NULL;
689
690         pfn = page_to_pfn(page);
691         if (!PageHighMem(page)) {
692                 address = (unsigned long)__va(pfn << PAGE_SHIFT);
693                 ptep = lookup_address(address, &level);
694                 if (WARN(ptep == NULL || level != PG_LEVEL_4K,
695                                         "m2p_add_override: pfn %lx not mapped", pfn))
696                         return -EINVAL;
697         }
698         WARN_ON(PagePrivate(page));
699         SetPagePrivate(page);
700         set_page_private(page, mfn);
701         page->index = pfn_to_mfn(pfn);
702
703         if (unlikely(!set_phys_to_machine(pfn, FOREIGN_FRAME(mfn))))
704                 return -ENOMEM;
705
706         if (kmap_op != NULL) {
707                 if (!PageHighMem(page)) {
708                         struct multicall_space mcs =
709                                 xen_mc_entry(sizeof(*kmap_op));
710
711                         MULTI_grant_table_op(mcs.mc,
712                                         GNTTABOP_map_grant_ref, kmap_op, 1);
713
714                         xen_mc_issue(PARAVIRT_LAZY_MMU);
715                 }
716                 /* let's use dev_bus_addr to record the old mfn instead */
717                 kmap_op->dev_bus_addr = page->index;
718                 page->index = (unsigned long) kmap_op;
719         }
720         spin_lock_irqsave(&m2p_override_lock, flags);
721         list_add(&page->lru,  &m2p_overrides[mfn_hash(mfn)]);
722         spin_unlock_irqrestore(&m2p_override_lock, flags);
723
724         return 0;
725 }
726 EXPORT_SYMBOL_GPL(m2p_add_override);
727 int m2p_remove_override(struct page *page, bool clear_pte)
728 {
729         unsigned long flags;
730         unsigned long mfn;
731         unsigned long pfn;
732         unsigned long uninitialized_var(address);
733         unsigned level;
734         pte_t *ptep = NULL;
735
736         pfn = page_to_pfn(page);
737         mfn = get_phys_to_machine(pfn);
738         if (mfn == INVALID_P2M_ENTRY || !(mfn & FOREIGN_FRAME_BIT))
739                 return -EINVAL;
740
741         if (!PageHighMem(page)) {
742                 address = (unsigned long)__va(pfn << PAGE_SHIFT);
743                 ptep = lookup_address(address, &level);
744
745                 if (WARN(ptep == NULL || level != PG_LEVEL_4K,
746                                         "m2p_remove_override: pfn %lx not mapped", pfn))
747                         return -EINVAL;
748         }
749
750         spin_lock_irqsave(&m2p_override_lock, flags);
751         list_del(&page->lru);
752         spin_unlock_irqrestore(&m2p_override_lock, flags);
753         WARN_ON(!PagePrivate(page));
754         ClearPagePrivate(page);
755
756         if (clear_pte) {
757                 struct gnttab_map_grant_ref *map_op =
758                         (struct gnttab_map_grant_ref *) page->index;
759                 set_phys_to_machine(pfn, map_op->dev_bus_addr);
760                 if (!PageHighMem(page)) {
761                         struct multicall_space mcs;
762                         struct gnttab_unmap_grant_ref *unmap_op;
763
764                         /*
765                          * It might be that we queued all the m2p grant table
766                          * hypercalls in a multicall, then m2p_remove_override
767                          * get called before the multicall has actually been
768                          * issued. In this case handle is going to -1 because
769                          * it hasn't been modified yet.
770                          */
771                         if (map_op->handle == -1)
772                                 xen_mc_flush();
773                         /*
774                          * Now if map_op->handle is negative it means that the
775                          * hypercall actually returned an error.
776                          */
777                         if (map_op->handle == GNTST_general_error) {
778                                 printk(KERN_WARNING "m2p_remove_override: "
779                                                 "pfn %lx mfn %lx, failed to modify kernel mappings",
780                                                 pfn, mfn);
781                                 return -1;
782                         }
783
784                         mcs = xen_mc_entry(
785                                         sizeof(struct gnttab_unmap_grant_ref));
786                         unmap_op = mcs.args;
787                         unmap_op->host_addr = map_op->host_addr;
788                         unmap_op->handle = map_op->handle;
789                         unmap_op->dev_bus_addr = 0;
790
791                         MULTI_grant_table_op(mcs.mc,
792                                         GNTTABOP_unmap_grant_ref, unmap_op, 1);
793
794                         xen_mc_issue(PARAVIRT_LAZY_MMU);
795
796                         set_pte_at(&init_mm, address, ptep,
797                                         pfn_pte(pfn, PAGE_KERNEL));
798                         __flush_tlb_single(address);
799                         map_op->host_addr = 0;
800                 }
801         } else
802                 set_phys_to_machine(pfn, page->index);
803
804         return 0;
805 }
806 EXPORT_SYMBOL_GPL(m2p_remove_override);
807
808 struct page *m2p_find_override(unsigned long mfn)
809 {
810         unsigned long flags;
811         struct list_head *bucket = &m2p_overrides[mfn_hash(mfn)];
812         struct page *p, *t, *ret;
813
814         ret = NULL;
815
816         list_for_each_entry_safe(p, t, bucket, lru) {
817                 if (page_private(p) == mfn) {
818                         ret = p;
819                         break;
820                 }
821         }
822
823         return ret;
824 }
825
826 unsigned long m2p_find_override_pfn(unsigned long mfn, unsigned long pfn)
827 {
828         struct page *p = m2p_find_override(mfn);
829         unsigned long ret = pfn;
830
831         if (p)
832                 ret = page_to_pfn(p);
833
834         return ret;
835 }
836 EXPORT_SYMBOL_GPL(m2p_find_override_pfn);
837
838 #ifdef CONFIG_XEN_DEBUG_FS
839 #include <linux/debugfs.h>
840 #include "debugfs.h"
841 static int p2m_dump_show(struct seq_file *m, void *v)
842 {
843         static const char * const level_name[] = { "top", "middle",
844                                                 "entry", "abnormal", "error"};
845 #define TYPE_IDENTITY 0
846 #define TYPE_MISSING 1
847 #define TYPE_PFN 2
848 #define TYPE_UNKNOWN 3
849         static const char * const type_name[] = {
850                                 [TYPE_IDENTITY] = "identity",
851                                 [TYPE_MISSING] = "missing",
852                                 [TYPE_PFN] = "pfn",
853                                 [TYPE_UNKNOWN] = "abnormal"};
854         unsigned long pfn, prev_pfn_type = 0, prev_pfn_level = 0;
855         unsigned int uninitialized_var(prev_level);
856         unsigned int uninitialized_var(prev_type);
857
858         if (!p2m_top)
859                 return 0;
860
861         for (pfn = 0; pfn < MAX_DOMAIN_PAGES; pfn++) {
862                 unsigned topidx = p2m_top_index(pfn);
863                 unsigned mididx = p2m_mid_index(pfn);
864                 unsigned idx = p2m_index(pfn);
865                 unsigned lvl, type;
866
867                 lvl = 4;
868                 type = TYPE_UNKNOWN;
869                 if (p2m_top[topidx] == p2m_mid_missing) {
870                         lvl = 0; type = TYPE_MISSING;
871                 } else if (p2m_top[topidx] == NULL) {
872                         lvl = 0; type = TYPE_UNKNOWN;
873                 } else if (p2m_top[topidx][mididx] == NULL) {
874                         lvl = 1; type = TYPE_UNKNOWN;
875                 } else if (p2m_top[topidx][mididx] == p2m_identity) {
876                         lvl = 1; type = TYPE_IDENTITY;
877                 } else if (p2m_top[topidx][mididx] == p2m_missing) {
878                         lvl = 1; type = TYPE_MISSING;
879                 } else if (p2m_top[topidx][mididx][idx] == 0) {
880                         lvl = 2; type = TYPE_UNKNOWN;
881                 } else if (p2m_top[topidx][mididx][idx] == IDENTITY_FRAME(pfn)) {
882                         lvl = 2; type = TYPE_IDENTITY;
883                 } else if (p2m_top[topidx][mididx][idx] == INVALID_P2M_ENTRY) {
884                         lvl = 2; type = TYPE_MISSING;
885                 } else if (p2m_top[topidx][mididx][idx] == pfn) {
886                         lvl = 2; type = TYPE_PFN;
887                 } else if (p2m_top[topidx][mididx][idx] != pfn) {
888                         lvl = 2; type = TYPE_PFN;
889                 }
890                 if (pfn == 0) {
891                         prev_level = lvl;
892                         prev_type = type;
893                 }
894                 if (pfn == MAX_DOMAIN_PAGES-1) {
895                         lvl = 3;
896                         type = TYPE_UNKNOWN;
897                 }
898                 if (prev_type != type) {
899                         seq_printf(m, " [0x%lx->0x%lx] %s\n",
900                                 prev_pfn_type, pfn, type_name[prev_type]);
901                         prev_pfn_type = pfn;
902                         prev_type = type;
903                 }
904                 if (prev_level != lvl) {
905                         seq_printf(m, " [0x%lx->0x%lx] level %s\n",
906                                 prev_pfn_level, pfn, level_name[prev_level]);
907                         prev_pfn_level = pfn;
908                         prev_level = lvl;
909                 }
910         }
911         return 0;
912 #undef TYPE_IDENTITY
913 #undef TYPE_MISSING
914 #undef TYPE_PFN
915 #undef TYPE_UNKNOWN
916 }
917
918 static int p2m_dump_open(struct inode *inode, struct file *filp)
919 {
920         return single_open(filp, p2m_dump_show, NULL);
921 }
922
923 static const struct file_operations p2m_dump_fops = {
924         .open           = p2m_dump_open,
925         .read           = seq_read,
926         .llseek         = seq_lseek,
927         .release        = single_release,
928 };
929
930 static struct dentry *d_mmu_debug;
931
932 static int __init xen_p2m_debugfs(void)
933 {
934         struct dentry *d_xen = xen_init_debugfs();
935
936         if (d_xen == NULL)
937                 return -ENOMEM;
938
939         d_mmu_debug = debugfs_create_dir("mmu", d_xen);
940
941         debugfs_create_file("p2m", 0600, d_mmu_debug, NULL, &p2m_dump_fops);
942         return 0;
943 }
944 fs_initcall(xen_p2m_debugfs);
945 #endif /* CONFIG_XEN_DEBUG_FS */