Pileus Git - ~andy/linux/blob - 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_mid_missing           p2m_missing
 143  * /-----------------\     /------------\
 144  * | [p2m_missing]   +---->| ~0, ~0, ~0 |
 145  * | [p2m_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/balloon.h>
 165 #include <xen/grant_table.h>
 166
 167 #include "multicalls.h"
 168 #include "xen-ops.h"
 169
 170 static void __init m2p_override_init(void);
 171
 172 unsigned long xen_max_p2m_pfn __read_mostly;
 173
 174 #define P2M_PER_PAGE            (PAGE_SIZE / sizeof(unsigned long))
 175 #define P2M_MID_PER_PAGE        (PAGE_SIZE / sizeof(unsigned long *))
 176 #define P2M_TOP_PER_PAGE        (PAGE_SIZE / sizeof(unsigned long **))
 177
 178 #define MAX_P2M_PFN             (P2M_TOP_PER_PAGE * P2M_MID_PER_PAGE * P2M_PER_PAGE)
 179
 180 /* Placeholders for holes in the address space */
 181 static RESERVE_BRK_ARRAY(unsigned long, p2m_missing, P2M_PER_PAGE);
 182 static RESERVE_BRK_ARRAY(unsigned long *, p2m_mid_missing, P2M_MID_PER_PAGE);
 183 static RESERVE_BRK_ARRAY(unsigned long, p2m_mid_missing_mfn, P2M_MID_PER_PAGE);
 184
 185 static RESERVE_BRK_ARRAY(unsigned long **, p2m_top, P2M_TOP_PER_PAGE);
 186 static RESERVE_BRK_ARRAY(unsigned long, p2m_top_mfn, P2M_TOP_PER_PAGE);
 187 static RESERVE_BRK_ARRAY(unsigned long *, p2m_top_mfn_p, P2M_TOP_PER_PAGE);
 188
 189 static RESERVE_BRK_ARRAY(unsigned long, p2m_identity, P2M_PER_PAGE);
 190
 191 RESERVE_BRK(p2m_mid, PAGE_SIZE * (MAX_DOMAIN_PAGES / (P2M_PER_PAGE * P2M_MID_PER_PAGE)));
 192 RESERVE_BRK(p2m_mid_mfn, PAGE_SIZE * (MAX_DOMAIN_PAGES / (P2M_PER_PAGE * P2M_MID_PER_PAGE)));
 193
 194 /* We might hit two boundary violations at the start and end, at max each
 195  * boundary violation will require three middle nodes. */
 196 RESERVE_BRK(p2m_mid_identity, PAGE_SIZE * 2 * 3);
 197
 198 /* When we populate back during bootup, the amount of pages can vary. The
 199  * max we have is seen is 395979, but that does not mean it can't be more.
 200  * Some machines can have 3GB I/O holes even. With early_can_reuse_p2m_middle
 201  * it can re-use Xen provided mfn_list array, so we only need to allocate at
 202  * most three P2M top nodes. */
 203 RESERVE_BRK(p2m_populated, PAGE_SIZE * 3);
 204
 205 static inline unsigned p2m_top_index(unsigned long pfn)
 206 {
 207         BUG_ON(pfn >= MAX_P2M_PFN);
 208         return pfn / (P2M_MID_PER_PAGE * P2M_PER_PAGE);
 209 }
 210
 211 static inline unsigned p2m_mid_index(unsigned long pfn)
 212 {
 213         return (pfn / P2M_PER_PAGE) % P2M_MID_PER_PAGE;
 214 }
 215
 216 static inline unsigned p2m_index(unsigned long pfn)
 217 {
 218         return pfn % P2M_PER_PAGE;
 219 }
 220
 221 static void p2m_top_init(unsigned long ***top)
 222 {
 223         unsigned i;
 224
 225         for (i = 0; i < P2M_TOP_PER_PAGE; i++)
 226                 top[i] = p2m_mid_missing;
 227 }
 228
 229 static void p2m_top_mfn_init(unsigned long *top)
 230 {
 231         unsigned i;
 232
 233         for (i = 0; i < P2M_TOP_PER_PAGE; i++)
 234                 top[i] = virt_to_mfn(p2m_mid_missing_mfn);
 235 }
 236
 237 static void p2m_top_mfn_p_init(unsigned long **top)
 238 {
 239         unsigned i;
 240
 241         for (i = 0; i < P2M_TOP_PER_PAGE; i++)
 242                 top[i] = p2m_mid_missing_mfn;
 243 }
 244
 245 static void p2m_mid_init(unsigned long **mid)
 246 {
 247         unsigned i;
 248
 249         for (i = 0; i < P2M_MID_PER_PAGE; i++)
 250                 mid[i] = p2m_missing;
 251 }
 252
 253 static void p2m_mid_mfn_init(unsigned long *mid)
 254 {
 255         unsigned i;
 256
 257         for (i = 0; i < P2M_MID_PER_PAGE; i++)
 258                 mid[i] = virt_to_mfn(p2m_missing);
 259 }
 260
 261 static void p2m_init(unsigned long *p2m)
 262 {
 263         unsigned i;
 264
 265         for (i = 0; i < P2M_MID_PER_PAGE; i++)
 266                 p2m[i] = INVALID_P2M_ENTRY;
 267 }
 268
 269 /*
 270  * Build the parallel p2m_top_mfn and p2m_mid_mfn structures
 271  *
 272  * This is called both at boot time, and after resuming from suspend:
 273  * - At boot time we're called very early, and must use extend_brk()
 274  *   to allocate memory.
 275  *
 276  * - After resume we're called from within stop_machine, but the mfn
 277  *   tree should alreay be completely allocated.
 278  */
 279 void __ref xen_build_mfn_list_list(void)
 280 {
 281         unsigned long pfn;
 282
 283         if (xen_feature(XENFEAT_auto_translated_physmap))
 284                 return;
 285
 286         /* Pre-initialize p2m_top_mfn to be completely missing */
 287         if (p2m_top_mfn == NULL) {
 288                 p2m_mid_missing_mfn = extend_brk(PAGE_SIZE, PAGE_SIZE);
 289                 p2m_mid_mfn_init(p2m_mid_missing_mfn);
 290
 291                 p2m_top_mfn_p = extend_brk(PAGE_SIZE, PAGE_SIZE);
 292                 p2m_top_mfn_p_init(p2m_top_mfn_p);
 293
 294                 p2m_top_mfn = extend_brk(PAGE_SIZE, PAGE_SIZE);
 295                 p2m_top_mfn_init(p2m_top_mfn);
 296         } else {
 297                 /* Reinitialise, mfn's all change after migration */
 298                 p2m_mid_mfn_init(p2m_mid_missing_mfn);
 299         }
 300
 301         for (pfn = 0; pfn < xen_max_p2m_pfn; pfn += P2M_PER_PAGE) {
 302                 unsigned topidx = p2m_top_index(pfn);
 303                 unsigned mididx = p2m_mid_index(pfn);
 304                 unsigned long **mid;
 305                 unsigned long *mid_mfn_p;
 306
 307                 mid = p2m_top[topidx];
 308                 mid_mfn_p = p2m_top_mfn_p[topidx];
 309
 310                 /* Don't bother allocating any mfn mid levels if
 311                  * they're just missing, just update the stored mfn,
 312                  * since all could have changed over a migrate.
 313                  */
 314                 if (mid == p2m_mid_missing) {
 315                         BUG_ON(mididx);
 316                         BUG_ON(mid_mfn_p != p2m_mid_missing_mfn);
 317                         p2m_top_mfn[topidx] = virt_to_mfn(p2m_mid_missing_mfn);
 318                         pfn += (P2M_MID_PER_PAGE - 1) * P2M_PER_PAGE;
 319                         continue;
 320                 }
 321
 322                 if (mid_mfn_p == p2m_mid_missing_mfn) {
 323                         /*
 324                          * XXX boot-time only!  We should never find
 325                          * missing parts of the mfn tree after
 326                          * runtime.  extend_brk() will BUG if we call
 327                          * it too late.
 328                          */
 329                         mid_mfn_p = extend_brk(PAGE_SIZE, PAGE_SIZE);
 330                         p2m_mid_mfn_init(mid_mfn_p);
 331
 332                         p2m_top_mfn_p[topidx] = mid_mfn_p;
 333                 }
 334
 335                 p2m_top_mfn[topidx] = virt_to_mfn(mid_mfn_p);
 336                 mid_mfn_p[mididx] = virt_to_mfn(mid[mididx]);
 337         }
 338 }
 339
 340 void xen_setup_mfn_list_list(void)
 341 {
 342         if (xen_feature(XENFEAT_auto_translated_physmap))
 343                 return;
 344
 345         BUG_ON(HYPERVISOR_shared_info == &xen_dummy_shared_info);
 346
 347         HYPERVISOR_shared_info->arch.pfn_to_mfn_frame_list_list =
 348                 virt_to_mfn(p2m_top_mfn);
 349         HYPERVISOR_shared_info->arch.max_pfn = xen_max_p2m_pfn;
 350 }
 351
 352 /* Set up p2m_top to point to the domain-builder provided p2m pages */
 353 void __init xen_build_dynamic_phys_to_machine(void)
 354 {
 355         unsigned long *mfn_list;
 356         unsigned long max_pfn;
 357         unsigned long pfn;
 358
 359          if (xen_feature(XENFEAT_auto_translated_physmap))
 360                 return;
 361
 362         mfn_list = (unsigned long *)xen_start_info->mfn_list;
 363         max_pfn = min(MAX_DOMAIN_PAGES, xen_start_info->nr_pages);
 364         xen_max_p2m_pfn = max_pfn;
 365
 366         p2m_missing = extend_brk(PAGE_SIZE, PAGE_SIZE);
 367         p2m_init(p2m_missing);
 368
 369         p2m_mid_missing = extend_brk(PAGE_SIZE, PAGE_SIZE);
 370         p2m_mid_init(p2m_mid_missing);
 371
 372         p2m_top = extend_brk(PAGE_SIZE, PAGE_SIZE);
 373         p2m_top_init(p2m_top);
 374
 375         p2m_identity = extend_brk(PAGE_SIZE, PAGE_SIZE);
 376         p2m_init(p2m_identity);
 377
 378         /*
 379          * The domain builder gives us a pre-constructed p2m array in
 380          * mfn_list for all the pages initially given to us, so we just
 381          * need to graft that into our tree structure.
 382          */
 383         for (pfn = 0; pfn < max_pfn; pfn += P2M_PER_PAGE) {
 384                 unsigned topidx = p2m_top_index(pfn);
 385                 unsigned mididx = p2m_mid_index(pfn);
 386
 387                 if (p2m_top[topidx] == p2m_mid_missing) {
 388                         unsigned long **mid = extend_brk(PAGE_SIZE, PAGE_SIZE);
 389                         p2m_mid_init(mid);
 390
 391                         p2m_top[topidx] = mid;
 392                 }
 393
 394                 /*
 395                  * As long as the mfn_list has enough entries to completely
 396                  * fill a p2m page, pointing into the array is ok. But if
 397                  * not the entries beyond the last pfn will be undefined.
 398                  */
 399                 if (unlikely(pfn + P2M_PER_PAGE > max_pfn)) {
 400                         unsigned long p2midx;
 401
 402                         p2midx = max_pfn % P2M_PER_PAGE;
 403                         for ( ; p2midx < P2M_PER_PAGE; p2midx++)
 404                                 mfn_list[pfn + p2midx] = INVALID_P2M_ENTRY;
 405                 }
 406                 p2m_top[topidx][mididx] = &mfn_list[pfn];
 407         }
 408
 409         m2p_override_init();
 410 }
 411 #ifdef CONFIG_X86_64
 412 #include <linux/bootmem.h>
 413 unsigned long __init xen_revector_p2m_tree(void)
 414 {
 415         unsigned long va_start;
 416         unsigned long va_end;
 417         unsigned long pfn;
 418         unsigned long pfn_free = 0;
 419         unsigned long *mfn_list = NULL;
 420         unsigned long size;
 421
 422         va_start = xen_start_info->mfn_list;
 423         /*We copy in increments of P2M_PER_PAGE * sizeof(unsigned long),
 424          * so make sure it is rounded up to that */
 425         size = PAGE_ALIGN(xen_start_info->nr_pages * sizeof(unsigned long));
 426         va_end = va_start + size;
 427
 428         /* If we were revectored already, don't do it again. */
 429         if (va_start <= __START_KERNEL_map && va_start >= __PAGE_OFFSET)
 430                 return 0;
 431
 432         mfn_list = alloc_bootmem_align(size, PAGE_SIZE);
 433         if (!mfn_list) {
 434                 pr_warn("Could not allocate space for a new P2M tree!\n");
 435                 return xen_start_info->mfn_list;
 436         }
 437         /* Fill it out with INVALID_P2M_ENTRY value */
 438         memset(mfn_list, 0xFF, size);
 439
 440         for (pfn = 0; pfn < ALIGN(MAX_DOMAIN_PAGES, P2M_PER_PAGE); pfn += P2M_PER_PAGE) {
 441                 unsigned topidx = p2m_top_index(pfn);
 442                 unsigned mididx;
 443                 unsigned long *mid_p;
 444
 445                 if (!p2m_top[topidx])
 446                         continue;
 447
 448                 if (p2m_top[topidx] == p2m_mid_missing)
 449                         continue;
 450
 451                 mididx = p2m_mid_index(pfn);
 452                 mid_p = p2m_top[topidx][mididx];
 453                 if (!mid_p)
 454                         continue;
 455                 if ((mid_p == p2m_missing) || (mid_p == p2m_identity))
 456                         continue;
 457
 458                 if ((unsigned long)mid_p == INVALID_P2M_ENTRY)
 459                         continue;
 460
 461                 /* The old va. Rebase it on mfn_list */
 462                 if (mid_p >= (unsigned long *)va_start && mid_p <= (unsigned long *)va_end) {
 463                         unsigned long *new;
 464
 465                         if (pfn_free  > (size / sizeof(unsigned long))) {
 466                                 WARN(1, "Only allocated for %ld pages, but we want %ld!\n",
 467                                      size / sizeof(unsigned long), pfn_free);
 468                                 return 0;
 469                         }
 470                         new = &mfn_list[pfn_free];
 471
 472                         copy_page(new, mid_p);
 473                         p2m_top[topidx][mididx] = &mfn_list[pfn_free];
 474                         p2m_top_mfn_p[topidx][mididx] = virt_to_mfn(&mfn_list[pfn_free]);
 475
 476                         pfn_free += P2M_PER_PAGE;
 477
 478                 }
 479                 /* This should be the leafs allocated for identity from _brk. */
 480         }
 481         return (unsigned long)mfn_list;
 482
 483 }
 484 #else
 485 unsigned long __init xen_revector_p2m_tree(void)
 486 {
 487         return 0;
 488 }
 489 #endif
 490 unsigned long get_phys_to_machine(unsigned long pfn)
 491 {
 492         unsigned topidx, mididx, idx;
 493
 494         if (unlikely(pfn >= MAX_P2M_PFN))
 495                 return INVALID_P2M_ENTRY;
 496
 497         topidx = p2m_top_index(pfn);
 498         mididx = p2m_mid_index(pfn);
 499         idx = p2m_index(pfn);
 500
 501         /*
 502          * The INVALID_P2M_ENTRY is filled in both p2m_*identity
 503          * and in p2m_*missing, so returning the INVALID_P2M_ENTRY
 504          * would be wrong.
 505          */
 506         if (p2m_top[topidx][mididx] == p2m_identity)
 507                 return IDENTITY_FRAME(pfn);
 508
 509         return p2m_top[topidx][mididx][idx];
 510 }
 511 EXPORT_SYMBOL_GPL(get_phys_to_machine);
 512
 513 static void *alloc_p2m_page(void)
 514 {
 515         return (void *)__get_free_page(GFP_KERNEL | __GFP_REPEAT);
 516 }
 517
 518 static void free_p2m_page(void *p)
 519 {
 520         free_page((unsigned long)p);
 521 }
 522
 523 /*
 524  * Fully allocate the p2m structure for a given pfn.  We need to check
 525  * that both the top and mid levels are allocated, and make sure the
 526  * parallel mfn tree is kept in sync.  We may race with other cpus, so
 527  * the new pages are installed with cmpxchg; if we lose the race then
 528  * simply free the page we allocated and use the one that's there.
 529  */
 530 static bool alloc_p2m(unsigned long pfn)
 531 {
 532         unsigned topidx, mididx;
 533         unsigned long ***top_p, **mid;
 534         unsigned long *top_mfn_p, *mid_mfn;
 535
 536         topidx = p2m_top_index(pfn);
 537         mididx = p2m_mid_index(pfn);
 538
 539         top_p = &p2m_top[topidx];
 540         mid = *top_p;
 541
 542         if (mid == p2m_mid_missing) {
 543                 /* Mid level is missing, allocate a new one */
 544                 mid = alloc_p2m_page();
 545                 if (!mid)
 546                         return false;
 547
 548                 p2m_mid_init(mid);
 549
 550                 if (cmpxchg(top_p, p2m_mid_missing, mid) != p2m_mid_missing)
 551                         free_p2m_page(mid);
 552         }
 553
 554         top_mfn_p = &p2m_top_mfn[topidx];
 555         mid_mfn = p2m_top_mfn_p[topidx];
 556
 557         BUG_ON(virt_to_mfn(mid_mfn) != *top_mfn_p);
 558
 559         if (mid_mfn == p2m_mid_missing_mfn) {
 560                 /* Separately check the mid mfn level */
 561                 unsigned long missing_mfn;
 562                 unsigned long mid_mfn_mfn;
 563
 564                 mid_mfn = alloc_p2m_page();
 565                 if (!mid_mfn)
 566                         return false;
 567
 568                 p2m_mid_mfn_init(mid_mfn);
 569
 570                 missing_mfn = virt_to_mfn(p2m_mid_missing_mfn);
 571                 mid_mfn_mfn = virt_to_mfn(mid_mfn);
 572                 if (cmpxchg(top_mfn_p, missing_mfn, mid_mfn_mfn) != missing_mfn)
 573                         free_p2m_page(mid_mfn);
 574                 else
 575                         p2m_top_mfn_p[topidx] = mid_mfn;
 576         }
 577
 578         if (p2m_top[topidx][mididx] == p2m_identity ||
 579             p2m_top[topidx][mididx] == p2m_missing) {
 580                 /* p2m leaf page is missing */
 581                 unsigned long *p2m;
 582                 unsigned long *p2m_orig = p2m_top[topidx][mididx];
 583
 584                 p2m = alloc_p2m_page();
 585                 if (!p2m)
 586                         return false;
 587
 588                 p2m_init(p2m);
 589
 590                 if (cmpxchg(&mid[mididx], p2m_orig, p2m) != p2m_orig)
 591                         free_p2m_page(p2m);
 592                 else
 593                         mid_mfn[mididx] = virt_to_mfn(p2m);
 594         }
 595
 596         return true;
 597 }
 598
 599 static bool __init early_alloc_p2m_middle(unsigned long pfn, bool check_boundary)
 600 {
 601         unsigned topidx, mididx, idx;
 602         unsigned long *p2m;
 603         unsigned long *mid_mfn_p;
 604
 605         topidx = p2m_top_index(pfn);
 606         mididx = p2m_mid_index(pfn);
 607         idx = p2m_index(pfn);
 608
 609         /* Pfff.. No boundary cross-over, lets get out. */
 610         if (!idx && check_boundary)
 611                 return false;
 612
 613         WARN(p2m_top[topidx][mididx] == p2m_identity,
 614                 "P2M[%d][%d] == IDENTITY, should be MISSING (or alloced)!\n",
 615                 topidx, mididx);
 616
 617         /*
 618          * Could be done by xen_build_dynamic_phys_to_machine..
 619          */
 620         if (p2m_top[topidx][mididx] != p2m_missing)
 621                 return false;
 622
 623         /* Boundary cross-over for the edges: */
 624         p2m = extend_brk(PAGE_SIZE, PAGE_SIZE);
 625
 626         p2m_init(p2m);
 627
 628         p2m_top[topidx][mididx] = p2m;
 629
 630         /* For save/restore we need to MFN of the P2M saved */
 631
 632         mid_mfn_p = p2m_top_mfn_p[topidx];
 633         WARN(mid_mfn_p[mididx] != virt_to_mfn(p2m_missing),
 634                 "P2M_TOP_P[%d][%d] != MFN of p2m_missing!\n",
 635                 topidx, mididx);
 636         mid_mfn_p[mididx] = virt_to_mfn(p2m);
 637
 638         return true;
 639 }
 640
 641 static bool __init early_alloc_p2m(unsigned long pfn)
 642 {
 643         unsigned topidx = p2m_top_index(pfn);
 644         unsigned long *mid_mfn_p;
 645         unsigned long **mid;
 646
 647         mid = p2m_top[topidx];
 648         mid_mfn_p = p2m_top_mfn_p[topidx];
 649         if (mid == p2m_mid_missing) {
 650                 mid = extend_brk(PAGE_SIZE, PAGE_SIZE);
 651
 652                 p2m_mid_init(mid);
 653
 654                 p2m_top[topidx] = mid;
 655
 656                 BUG_ON(mid_mfn_p != p2m_mid_missing_mfn);
 657         }
 658         /* And the save/restore P2M tables.. */
 659         if (mid_mfn_p == p2m_mid_missing_mfn) {
 660                 mid_mfn_p = extend_brk(PAGE_SIZE, PAGE_SIZE);
 661                 p2m_mid_mfn_init(mid_mfn_p);
 662
 663                 p2m_top_mfn_p[topidx] = mid_mfn_p;
 664                 p2m_top_mfn[topidx] = virt_to_mfn(mid_mfn_p);
 665                 /* Note: we don't set mid_mfn_p[midix] here,
 666                  * look in early_alloc_p2m_middle */
 667         }
 668         return true;
 669 }
 670
 671 /*
 672  * Skim over the P2M tree looking at pages that are either filled with
 673  * INVALID_P2M_ENTRY or with 1:1 PFNs. If found, re-use that page and
 674  * replace the P2M leaf with a p2m_missing or p2m_identity.
 675  * Stick the old page in the new P2M tree location.
 676  */
 677 bool __init early_can_reuse_p2m_middle(unsigned long set_pfn, unsigned long set_mfn)
 678 {
 679         unsigned topidx;
 680         unsigned mididx;
 681         unsigned ident_pfns;
 682         unsigned inv_pfns;
 683         unsigned long *p2m;
 684         unsigned long *mid_mfn_p;
 685         unsigned idx;
 686         unsigned long pfn;
 687
 688         /* We only look when this entails a P2M middle layer */
 689         if (p2m_index(set_pfn))
 690                 return false;
 691
 692         for (pfn = 0; pfn < MAX_DOMAIN_PAGES; pfn += P2M_PER_PAGE) {
 693                 topidx = p2m_top_index(pfn);
 694
 695                 if (!p2m_top[topidx])
 696                         continue;
 697
 698                 if (p2m_top[topidx] == p2m_mid_missing)
 699                         continue;
 700
 701                 mididx = p2m_mid_index(pfn);
 702                 p2m = p2m_top[topidx][mididx];
 703                 if (!p2m)
 704                         continue;
 705
 706                 if ((p2m == p2m_missing) || (p2m == p2m_identity))
 707                         continue;
 708
 709                 if ((unsigned long)p2m == INVALID_P2M_ENTRY)
 710                         continue;
 711
 712                 ident_pfns = 0;
 713                 inv_pfns = 0;
 714                 for (idx = 0; idx < P2M_PER_PAGE; idx++) {
 715                         /* IDENTITY_PFNs are 1:1 */
 716                         if (p2m[idx] == IDENTITY_FRAME(pfn + idx))
 717                                 ident_pfns++;
 718                         else if (p2m[idx] == INVALID_P2M_ENTRY)
 719                                 inv_pfns++;
 720                         else
 721                                 break;
 722                 }
 723                 if ((ident_pfns == P2M_PER_PAGE) || (inv_pfns == P2M_PER_PAGE))
 724                         goto found;
 725         }
 726         return false;
 727 found:
 728         /* Found one, replace old with p2m_identity or p2m_missing */
 729         p2m_top[topidx][mididx] = (ident_pfns ? p2m_identity : p2m_missing);
 730         /* And the other for save/restore.. */
 731         mid_mfn_p = p2m_top_mfn_p[topidx];
 732         /* NOTE: Even if it is a p2m_identity it should still be point to
 733          * a page filled with INVALID_P2M_ENTRY entries. */
 734         mid_mfn_p[mididx] = virt_to_mfn(p2m_missing);
 735
 736         /* Reset where we want to stick the old page in. */
 737         topidx = p2m_top_index(set_pfn);
 738         mididx = p2m_mid_index(set_pfn);
 739
 740         /* This shouldn't happen */
 741         if (WARN_ON(p2m_top[topidx] == p2m_mid_missing))
 742                 early_alloc_p2m(set_pfn);
 743
 744         if (WARN_ON(p2m_top[topidx][mididx] != p2m_missing))
 745                 return false;
 746
 747         p2m_init(p2m);
 748         p2m_top[topidx][mididx] = p2m;
 749         mid_mfn_p = p2m_top_mfn_p[topidx];
 750         mid_mfn_p[mididx] = virt_to_mfn(p2m);
 751
 752         return true;
 753 }
 754 bool __init early_set_phys_to_machine(unsigned long pfn, unsigned long mfn)
 755 {
 756         if (unlikely(!__set_phys_to_machine(pfn, mfn)))  {
 757                 if (!early_alloc_p2m(pfn))
 758                         return false;
 759
 760                 if (early_can_reuse_p2m_middle(pfn, mfn))
 761                         return __set_phys_to_machine(pfn, mfn);
 762
 763                 if (!early_alloc_p2m_middle(pfn, false /* boundary crossover OK!*/))
 764                         return false;
 765
 766                 if (!__set_phys_to_machine(pfn, mfn))
 767                         return false;
 768         }
 769
 770         return true;
 771 }
 772 unsigned long __init set_phys_range_identity(unsigned long pfn_s,
 773                                       unsigned long pfn_e)
 774 {
 775         unsigned long pfn;
 776
 777         if (unlikely(pfn_s >= MAX_P2M_PFN || pfn_e >= MAX_P2M_PFN))
 778                 return 0;
 779
 780         if (unlikely(xen_feature(XENFEAT_auto_translated_physmap)))
 781                 return pfn_e - pfn_s;
 782
 783         if (pfn_s > pfn_e)
 784                 return 0;
 785
 786         for (pfn = (pfn_s & ~(P2M_MID_PER_PAGE * P2M_PER_PAGE - 1));
 787                 pfn < ALIGN(pfn_e, (P2M_MID_PER_PAGE * P2M_PER_PAGE));
 788                 pfn += P2M_MID_PER_PAGE * P2M_PER_PAGE)
 789         {
 790                 WARN_ON(!early_alloc_p2m(pfn));
 791         }
 792
 793         early_alloc_p2m_middle(pfn_s, true);
 794         early_alloc_p2m_middle(pfn_e, true);
 795
 796         for (pfn = pfn_s; pfn < pfn_e; pfn++)
 797                 if (!__set_phys_to_machine(pfn, IDENTITY_FRAME(pfn)))
 798                         break;
 799
 800         if (!WARN((pfn - pfn_s) != (pfn_e - pfn_s),
 801                 "Identity mapping failed. We are %ld short of 1-1 mappings!\n",
 802                 (pfn_e - pfn_s) - (pfn - pfn_s)))
 803                 printk(KERN_DEBUG "1-1 mapping on %lx->%lx\n", pfn_s, pfn);
 804
 805         return pfn - pfn_s;
 806 }
 807
 808 /* Try to install p2m mapping; fail if intermediate bits missing */
 809 bool __set_phys_to_machine(unsigned long pfn, unsigned long mfn)
 810 {
 811         unsigned topidx, mididx, idx;
 812
 813         /* don't track P2M changes in autotranslate guests */
 814         if (unlikely(xen_feature(XENFEAT_auto_translated_physmap)))
 815                 return true;
 816
 817         if (unlikely(pfn >= MAX_P2M_PFN)) {
 818                 BUG_ON(mfn != INVALID_P2M_ENTRY);
 819                 return true;
 820         }
 821
 822         topidx = p2m_top_index(pfn);
 823         mididx = p2m_mid_index(pfn);
 824         idx = p2m_index(pfn);
 825
 826         /* For sparse holes were the p2m leaf has real PFN along with
 827          * PCI holes, stick in the PFN as the MFN value.
 828          */
 829         if (mfn != INVALID_P2M_ENTRY && (mfn & IDENTITY_FRAME_BIT)) {
 830                 if (p2m_top[topidx][mididx] == p2m_identity)
 831                         return true;
 832
 833                 /* Swap over from MISSING to IDENTITY if needed. */
 834                 if (p2m_top[topidx][mididx] == p2m_missing) {
 835                         WARN_ON(cmpxchg(&p2m_top[topidx][mididx], p2m_missing,
 836                                 p2m_identity) != p2m_missing);
 837                         return true;
 838                 }
 839         }
 840
 841         if (p2m_top[topidx][mididx] == p2m_missing)
 842                 return mfn == INVALID_P2M_ENTRY;
 843
 844         p2m_top[topidx][mididx][idx] = mfn;
 845
 846         return true;
 847 }
 848
 849 bool set_phys_to_machine(unsigned long pfn, unsigned long mfn)
 850 {
 851         if (unlikely(!__set_phys_to_machine(pfn, mfn)))  {
 852                 if (!alloc_p2m(pfn))
 853                         return false;
 854
 855                 if (!__set_phys_to_machine(pfn, mfn))
 856                         return false;
 857         }
 858
 859         return true;
 860 }
 861
 862 #define M2P_OVERRIDE_HASH_SHIFT 10
 863 #define M2P_OVERRIDE_HASH       (1 << M2P_OVERRIDE_HASH_SHIFT)
 864
 865 static RESERVE_BRK_ARRAY(struct list_head, m2p_overrides, M2P_OVERRIDE_HASH);
 866 static DEFINE_SPINLOCK(m2p_override_lock);
 867
 868 static void __init m2p_override_init(void)
 869 {
 870         unsigned i;
 871
 872         m2p_overrides = extend_brk(sizeof(*m2p_overrides) * M2P_OVERRIDE_HASH,
 873                                    sizeof(unsigned long));
 874
 875         for (i = 0; i < M2P_OVERRIDE_HASH; i++)
 876                 INIT_LIST_HEAD(&m2p_overrides[i]);
 877 }
 878
 879 static unsigned long mfn_hash(unsigned long mfn)
 880 {
 881         return hash_long(mfn, M2P_OVERRIDE_HASH_SHIFT);
 882 }
 883
 884 /* Add an MFN override for a particular page */
 885 int m2p_add_override(unsigned long mfn, struct page *page,
 886                 struct gnttab_map_grant_ref *kmap_op)
 887 {
 888         unsigned long flags;
 889         unsigned long pfn;
 890         unsigned long uninitialized_var(address);
 891         unsigned level;
 892         pte_t *ptep = NULL;
 893
 894         pfn = page_to_pfn(page);
 895         if (!PageHighMem(page)) {
 896                 address = (unsigned long)__va(pfn << PAGE_SHIFT);
 897                 ptep = lookup_address(address, &level);
 898                 if (WARN(ptep == NULL || level != PG_LEVEL_4K,
 899                                         "m2p_add_override: pfn %lx not mapped", pfn))
 900                         return -EINVAL;
 901         }
 902         WARN_ON(PagePrivate(page));
 903         SetPagePrivate(page);
 904         set_page_private(page, mfn);
 905         page->index = pfn_to_mfn(pfn);
 906
 907         if (unlikely(!set_phys_to_machine(pfn, FOREIGN_FRAME(mfn))))
 908                 return -ENOMEM;
 909
 910         if (kmap_op != NULL) {
 911                 if (!PageHighMem(page)) {
 912                         struct multicall_space mcs =
 913                                 xen_mc_entry(sizeof(*kmap_op));
 914
 915                         MULTI_grant_table_op(mcs.mc,
 916                                         GNTTABOP_map_grant_ref, kmap_op, 1);
 917
 918                         xen_mc_issue(PARAVIRT_LAZY_MMU);
 919                 }
 920         }
 921         spin_lock_irqsave(&m2p_override_lock, flags);
 922         list_add(&page->lru,  &m2p_overrides[mfn_hash(mfn)]);
 923         spin_unlock_irqrestore(&m2p_override_lock, flags);
 924
 925         /* p2m(m2p(mfn)) == mfn: the mfn is already present somewhere in
 926          * this domain. Set the FOREIGN_FRAME_BIT in the p2m for the other
 927          * pfn so that the following mfn_to_pfn(mfn) calls will return the
 928          * pfn from the m2p_override (the backend pfn) instead.
 929          * We need to do this because the pages shared by the frontend
 930          * (xen-blkfront) can be already locked (lock_page, called by
 931          * do_read_cache_page); when the userspace backend tries to use them
 932          * with direct_IO, mfn_to_pfn returns the pfn of the frontend, so
 933          * do_blockdev_direct_IO is going to try to lock the same pages
 934          * again resulting in a deadlock.
 935          * As a side effect get_user_pages_fast might not be safe on the
 936          * frontend pages while they are being shared with the backend,
 937          * because mfn_to_pfn (that ends up being called by GUPF) will
 938          * return the backend pfn rather than the frontend pfn. */
 939         pfn = mfn_to_pfn_no_overrides(mfn);
 940         if (get_phys_to_machine(pfn) == mfn)
 941                 set_phys_to_machine(pfn, FOREIGN_FRAME(mfn));
 942
 943         return 0;
 944 }
 945 EXPORT_SYMBOL_GPL(m2p_add_override);
 946 int m2p_remove_override(struct page *page,
 947                 struct gnttab_map_grant_ref *kmap_op)
 948 {
 949         unsigned long flags;
 950         unsigned long mfn;
 951         unsigned long pfn;
 952         unsigned long uninitialized_var(address);
 953         unsigned level;
 954         pte_t *ptep = NULL;
 955
 956         pfn = page_to_pfn(page);
 957         mfn = get_phys_to_machine(pfn);
 958         if (mfn == INVALID_P2M_ENTRY || !(mfn & FOREIGN_FRAME_BIT))
 959                 return -EINVAL;
 960
 961         if (!PageHighMem(page)) {
 962                 address = (unsigned long)__va(pfn << PAGE_SHIFT);
 963                 ptep = lookup_address(address, &level);
 964
 965                 if (WARN(ptep == NULL || level != PG_LEVEL_4K,
 966                                         "m2p_remove_override: pfn %lx not mapped", pfn))
 967                         return -EINVAL;
 968         }
 969
 970         spin_lock_irqsave(&m2p_override_lock, flags);
 971         list_del(&page->lru);
 972         spin_unlock_irqrestore(&m2p_override_lock, flags);
 973         WARN_ON(!PagePrivate(page));
 974         ClearPagePrivate(page);
 975
 976         set_phys_to_machine(pfn, page->index);
 977         if (kmap_op != NULL) {
 978                 if (!PageHighMem(page)) {
 979                         struct multicall_space mcs;
 980                         struct gnttab_unmap_and_replace *unmap_op;
 981                         struct page *scratch_page = get_balloon_scratch_page();
 982                         unsigned long scratch_page_address = (unsigned long)
 983                                 __va(page_to_pfn(scratch_page) << PAGE_SHIFT);
 984
 985                         /*
 986                          * It might be that we queued all the m2p grant table
 987                          * hypercalls in a multicall, then m2p_remove_override
 988                          * get called before the multicall has actually been
 989                          * issued. In this case handle is going to -1 because
 990                          * it hasn't been modified yet.
 991                          */
 992                         if (kmap_op->handle == -1)
 993                                 xen_mc_flush();
 994                         /*
 995                          * Now if kmap_op->handle is negative it means that the
 996                          * hypercall actually returned an error.
 997                          */
 998                         if (kmap_op->handle == GNTST_general_error) {
 999                                 printk(KERN_WARNING "m2p_remove_override: "
1000                                                 "pfn %lx mfn %lx, failed to modify kernel mappings",
1001                                                 pfn, mfn);
1002                                 put_balloon_scratch_page();
1003                                 return -1;
1004                         }
1005
1006                         xen_mc_batch();
1007
1008                         mcs = __xen_mc_entry(
1009                                         sizeof(struct gnttab_unmap_and_replace));
1010                         unmap_op = mcs.args;
1011                         unmap_op->host_addr = kmap_op->host_addr;
1012                         unmap_op->new_addr = scratch_page_address;
1013                         unmap_op->handle = kmap_op->handle;
1014
1015                         MULTI_grant_table_op(mcs.mc,
1016                                         GNTTABOP_unmap_and_replace, unmap_op, 1);
1017
1018                         mcs = __xen_mc_entry(0);
1019                         MULTI_update_va_mapping(mcs.mc, scratch_page_address,
1020                                         pfn_pte(page_to_pfn(scratch_page),
1021                                         PAGE_KERNEL_RO), 0);
1022
1023                         xen_mc_issue(PARAVIRT_LAZY_MMU);
1024
1025                         kmap_op->host_addr = 0;
1026                         put_balloon_scratch_page();
1027                 }
1028         }
1029
1030         /* p2m(m2p(mfn)) == FOREIGN_FRAME(mfn): the mfn is already present
1031          * somewhere in this domain, even before being added to the
1032          * m2p_override (see comment above in m2p_add_override).
1033          * If there are no other entries in the m2p_override corresponding
1034          * to this mfn, then remove the FOREIGN_FRAME_BIT from the p2m for
1035          * the original pfn (the one shared by the frontend): the backend
1036          * cannot do any IO on this page anymore because it has been
1037          * unshared. Removing the FOREIGN_FRAME_BIT from the p2m entry of
1038          * the original pfn causes mfn_to_pfn(mfn) to return the frontend
1039          * pfn again. */
1040         mfn &= ~FOREIGN_FRAME_BIT;
1041         pfn = mfn_to_pfn_no_overrides(mfn);
1042         if (get_phys_to_machine(pfn) == FOREIGN_FRAME(mfn) &&
1043                         m2p_find_override(mfn) == NULL)
1044                 set_phys_to_machine(pfn, mfn);
1045
1046         return 0;
1047 }
1048 EXPORT_SYMBOL_GPL(m2p_remove_override);
1049
1050 struct page *m2p_find_override(unsigned long mfn)
1051 {
1052         unsigned long flags;
1053         struct list_head *bucket = &m2p_overrides[mfn_hash(mfn)];
1054         struct page *p, *ret;
1055
1056         ret = NULL;
1057
1058         spin_lock_irqsave(&m2p_override_lock, flags);
1059
1060         list_for_each_entry(p, bucket, lru) {
1061                 if (page_private(p) == mfn) {
1062                         ret = p;
1063                         break;
1064                 }
1065         }
1066
1067         spin_unlock_irqrestore(&m2p_override_lock, flags);
1068
1069         return ret;
1070 }
1071
1072 unsigned long m2p_find_override_pfn(unsigned long mfn, unsigned long pfn)
1073 {
1074         struct page *p = m2p_find_override(mfn);
1075         unsigned long ret = pfn;
1076
1077         if (p)
1078                 ret = page_to_pfn(p);
1079
1080         return ret;
1081 }
1082 EXPORT_SYMBOL_GPL(m2p_find_override_pfn);
1083
1084 #ifdef CONFIG_XEN_DEBUG_FS
1085 #include <linux/debugfs.h>
1086 #include "debugfs.h"
1087 static int p2m_dump_show(struct seq_file *m, void *v)
1088 {
1089         static const char * const level_name[] = { "top", "middle",
1090                                                 "entry", "abnormal", "error"};
1091 #define TYPE_IDENTITY 0
1092 #define TYPE_MISSING 1
1093 #define TYPE_PFN 2
1094 #define TYPE_UNKNOWN 3
1095         static const char * const type_name[] = {
1096                                 [TYPE_IDENTITY] = "identity",
1097                                 [TYPE_MISSING] = "missing",
1098                                 [TYPE_PFN] = "pfn",
1099                                 [TYPE_UNKNOWN] = "abnormal"};
1100         unsigned long pfn, prev_pfn_type = 0, prev_pfn_level = 0;
1101         unsigned int uninitialized_var(prev_level);
1102         unsigned int uninitialized_var(prev_type);
1103
1104         if (!p2m_top)
1105                 return 0;
1106
1107         for (pfn = 0; pfn < MAX_DOMAIN_PAGES; pfn++) {
1108                 unsigned topidx = p2m_top_index(pfn);
1109                 unsigned mididx = p2m_mid_index(pfn);
1110                 unsigned idx = p2m_index(pfn);
1111                 unsigned lvl, type;
1112
1113                 lvl = 4;
1114                 type = TYPE_UNKNOWN;
1115                 if (p2m_top[topidx] == p2m_mid_missing) {
1116                         lvl = 0; type = TYPE_MISSING;
1117                 } else if (p2m_top[topidx] == NULL) {
1118                         lvl = 0; type = TYPE_UNKNOWN;
1119                 } else if (p2m_top[topidx][mididx] == NULL) {
1120                         lvl = 1; type = TYPE_UNKNOWN;
1121                 } else if (p2m_top[topidx][mididx] == p2m_identity) {
1122                         lvl = 1; type = TYPE_IDENTITY;
1123                 } else if (p2m_top[topidx][mididx] == p2m_missing) {
1124                         lvl = 1; type = TYPE_MISSING;
1125                 } else if (p2m_top[topidx][mididx][idx] == 0) {
1126                         lvl = 2; type = TYPE_UNKNOWN;
1127                 } else if (p2m_top[topidx][mididx][idx] == IDENTITY_FRAME(pfn)) {
1128                         lvl = 2; type = TYPE_IDENTITY;
1129                 } else if (p2m_top[topidx][mididx][idx] == INVALID_P2M_ENTRY) {
1130                         lvl = 2; type = TYPE_MISSING;
1131                 } else if (p2m_top[topidx][mididx][idx] == pfn) {
1132                         lvl = 2; type = TYPE_PFN;
1133                 } else if (p2m_top[topidx][mididx][idx] != pfn) {
1134                         lvl = 2; type = TYPE_PFN;
1135                 }
1136                 if (pfn == 0) {
1137                         prev_level = lvl;
1138                         prev_type = type;
1139                 }
1140                 if (pfn == MAX_DOMAIN_PAGES-1) {
1141                         lvl = 3;
1142                         type = TYPE_UNKNOWN;
1143                 }
1144                 if (prev_type != type) {
1145                         seq_printf(m, " [0x%lx->0x%lx] %s\n",
1146                                 prev_pfn_type, pfn, type_name[prev_type]);
1147                         prev_pfn_type = pfn;
1148                         prev_type = type;
1149                 }
1150                 if (prev_level != lvl) {
1151                         seq_printf(m, " [0x%lx->0x%lx] level %s\n",
1152                                 prev_pfn_level, pfn, level_name[prev_level]);
1153                         prev_pfn_level = pfn;
1154                         prev_level = lvl;
1155                 }
1156         }
1157         return 0;
1158 #undef TYPE_IDENTITY
1159 #undef TYPE_MISSING
1160 #undef TYPE_PFN
1161 #undef TYPE_UNKNOWN
1162 }
1163
1164 static int p2m_dump_open(struct inode *inode, struct file *filp)
1165 {
1166         return single_open(filp, p2m_dump_show, NULL);
1167 }
1168
1169 static const struct file_operations p2m_dump_fops = {
1170         .open           = p2m_dump_open,
1171         .read           = seq_read,
1172         .llseek         = seq_lseek,
1173         .release        = single_release,
1174 };
1175
1176 static struct dentry *d_mmu_debug;
1177
1178 static int __init xen_p2m_debugfs(void)
1179 {
1180         struct dentry *d_xen = xen_init_debugfs();
1181
1182         if (d_xen == NULL)
1183                 return -ENOMEM;
1184
1185         d_mmu_debug = debugfs_create_dir("mmu", d_xen);
1186
1187         debugfs_create_file("p2m", 0600, d_mmu_debug, NULL, &p2m_dump_fops);
1188         return 0;
1189 }
1190 fs_initcall(xen_p2m_debugfs);
1191 #endif /* CONFIG_XEN_DEBUG_FS */