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 logical flat p2m table is mapped to a linear kernel memory area.
   7  * For accesses by Xen a three-level tree linked via mfns only is set up to
   8  * allow the address space to be sparse.
   9  *
  10  *               Xen
  11  *                |
  12  *          p2m_top_mfn
  13  *              /   \
  14  * p2m_mid_mfn p2m_mid_mfn
  15  *         /           /
  16  *  p2m p2m p2m ...
  17  *
  18  * The p2m_mid_mfn pages are mapped by p2m_top_mfn_p.
  19  *
  20  * The p2m_top_mfn level is limited to 1 page, so the maximum representable
  21  * pseudo-physical address space is:
  22  *  P2M_TOP_PER_PAGE * P2M_MID_PER_PAGE * P2M_PER_PAGE pages
  23  *
  24  * P2M_PER_PAGE depends on the architecture, as a mfn is always
  25  * unsigned long (8 bytes on 64-bit, 4 bytes on 32), leading to
  26  * 512 and 1024 entries respectively.
  27  *
  28  * In short, these structures contain the Machine Frame Number (MFN) of the PFN.
  29  *
  30  * However not all entries are filled with MFNs. Specifically for all other
  31  * leaf entries, or for the top  root, or middle one, for which there is a void
  32  * entry, we assume it is  "missing". So (for example)
  33  *  pfn_to_mfn(0x90909090)=INVALID_P2M_ENTRY.
  34  * We have a dedicated page p2m_missing with all entries being
  35  * INVALID_P2M_ENTRY. This page may be referenced multiple times in the p2m
  36  * list/tree in case there are multiple areas with P2M_PER_PAGE invalid pfns.
  37  *
  38  * We also have the possibility of setting 1-1 mappings on certain regions, so
  39  * that:
  40  *  pfn_to_mfn(0xc0000)=0xc0000
  41  *
  42  * The benefit of this is, that we can assume for non-RAM regions (think
  43  * PCI BARs, or ACPI spaces), we can create mappings easily because we
  44  * get the PFN value to match the MFN.
  45  *
  46  * For this to work efficiently we have one new page p2m_identity. All entries
  47  * in p2m_identity are set to INVALID_P2M_ENTRY type (Xen toolstack only
  48  * recognizes that and MFNs, no other fancy value).
  49  *
  50  * On lookup we spot that the entry points to p2m_identity and return the
  51  * identity value instead of dereferencing and returning INVALID_P2M_ENTRY.
  52  * If the entry points to an allocated page, we just proceed as before and
  53  * return the PFN. If the PFN has IDENTITY_FRAME_BIT set we unmask that in
  54  * appropriate functions (pfn_to_mfn).
  55  *
  56  * The reason for having the IDENTITY_FRAME_BIT instead of just returning the
  57  * PFN is that we could find ourselves where pfn_to_mfn(pfn)==pfn for a
  58  * non-identity pfn. To protect ourselves against we elect to set (and get) the
  59  * IDENTITY_FRAME_BIT on all identity mapped PFNs.
  60  */
  61
  62 #include <linux/init.h>
  63 #include <linux/module.h>
  64 #include <linux/list.h>
  65 #include <linux/hash.h>
  66 #include <linux/sched.h>
  67 #include <linux/seq_file.h>
  68 #include <linux/bootmem.h>
  69 #include <linux/slab.h>
  70 #include <linux/vmalloc.h>
  71
  72 #include <asm/cache.h>
  73 #include <asm/setup.h>
  74 #include <asm/uaccess.h>
  75
  76 #include <asm/xen/page.h>
  77 #include <asm/xen/hypercall.h>
  78 #include <asm/xen/hypervisor.h>
  79 #include <xen/balloon.h>
  80 #include <xen/grant_table.h>
  81
  82 #include "p2m.h"
  83 #include "multicalls.h"
  84 #include "xen-ops.h"
  85
  86 #define PMDS_PER_MID_PAGE       (P2M_MID_PER_PAGE / PTRS_PER_PTE)
  87
  88 unsigned long *xen_p2m_addr __read_mostly;
  89 EXPORT_SYMBOL_GPL(xen_p2m_addr);
  90 unsigned long xen_p2m_size __read_mostly;
  91 EXPORT_SYMBOL_GPL(xen_p2m_size);
  92 unsigned long xen_max_p2m_pfn __read_mostly;
  93 EXPORT_SYMBOL_GPL(xen_max_p2m_pfn);
  94
  95 #ifdef CONFIG_XEN_BALLOON_MEMORY_HOTPLUG_LIMIT
  96 #define P2M_LIMIT CONFIG_XEN_BALLOON_MEMORY_HOTPLUG_LIMIT
  97 #else
  98 #define P2M_LIMIT 0
  99 #endif
 100
 101 static DEFINE_SPINLOCK(p2m_update_lock);
 102
 103 static unsigned long *p2m_mid_missing_mfn;
 104 static unsigned long *p2m_top_mfn;
 105 static unsigned long **p2m_top_mfn_p;
 106 static unsigned long *p2m_missing;
 107 static unsigned long *p2m_identity;
 108 static pte_t *p2m_missing_pte;
 109 static pte_t *p2m_identity_pte;
 110
 111 static inline unsigned p2m_top_index(unsigned long pfn)
 112 {
 113         BUG_ON(pfn >= MAX_P2M_PFN);
 114         return pfn / (P2M_MID_PER_PAGE * P2M_PER_PAGE);
 115 }
 116
 117 static inline unsigned p2m_mid_index(unsigned long pfn)
 118 {
 119         return (pfn / P2M_PER_PAGE) % P2M_MID_PER_PAGE;
 120 }
 121
 122 static inline unsigned p2m_index(unsigned long pfn)
 123 {
 124         return pfn % P2M_PER_PAGE;
 125 }
 126
 127 static void p2m_top_mfn_init(unsigned long *top)
 128 {
 129         unsigned i;
 130
 131         for (i = 0; i < P2M_TOP_PER_PAGE; i++)
 132                 top[i] = virt_to_mfn(p2m_mid_missing_mfn);
 133 }
 134
 135 static void p2m_top_mfn_p_init(unsigned long **top)
 136 {
 137         unsigned i;
 138
 139         for (i = 0; i < P2M_TOP_PER_PAGE; i++)
 140                 top[i] = p2m_mid_missing_mfn;
 141 }
 142
 143 static void p2m_mid_mfn_init(unsigned long *mid, unsigned long *leaf)
 144 {
 145         unsigned i;
 146
 147         for (i = 0; i < P2M_MID_PER_PAGE; i++)
 148                 mid[i] = virt_to_mfn(leaf);
 149 }
 150
 151 static void p2m_init(unsigned long *p2m)
 152 {
 153         unsigned i;
 154
 155         for (i = 0; i < P2M_PER_PAGE; i++)
 156                 p2m[i] = INVALID_P2M_ENTRY;
 157 }
 158
 159 static void p2m_init_identity(unsigned long *p2m, unsigned long pfn)
 160 {
 161         unsigned i;
 162
 163         for (i = 0; i < P2M_PER_PAGE; i++)
 164                 p2m[i] = IDENTITY_FRAME(pfn + i);
 165 }
 166
 167 static void * __ref alloc_p2m_page(void)
 168 {
 169         if (unlikely(!slab_is_available()))
 170                 return alloc_bootmem_align(PAGE_SIZE, PAGE_SIZE);
 171
 172         return (void *)__get_free_page(GFP_KERNEL | __GFP_REPEAT);
 173 }
 174
 175 static void __ref free_p2m_page(void *p)
 176 {
 177         if (unlikely(!slab_is_available())) {
 178                 free_bootmem((unsigned long)p, PAGE_SIZE);
 179                 return;
 180         }
 181
 182         free_page((unsigned long)p);
 183 }
 184
 185 /*
 186  * Build the parallel p2m_top_mfn and p2m_mid_mfn structures
 187  *
 188  * This is called both at boot time, and after resuming from suspend:
 189  * - At boot time we're called rather early, and must use alloc_bootmem*()
 190  *   to allocate memory.
 191  *
 192  * - After resume we're called from within stop_machine, but the mfn
 193  *   tree should already be completely allocated.
 194  */
 195 void __ref xen_build_mfn_list_list(void)
 196 {
 197         unsigned long pfn, mfn;
 198         pte_t *ptep;
 199         unsigned int level, topidx, mididx;
 200         unsigned long *mid_mfn_p;
 201
 202         if (xen_feature(XENFEAT_auto_translated_physmap))
 203                 return;
 204
 205         /* Pre-initialize p2m_top_mfn to be completely missing */
 206         if (p2m_top_mfn == NULL) {
 207                 p2m_mid_missing_mfn = alloc_p2m_page();
 208                 p2m_mid_mfn_init(p2m_mid_missing_mfn, p2m_missing);
 209
 210                 p2m_top_mfn_p = alloc_p2m_page();
 211                 p2m_top_mfn_p_init(p2m_top_mfn_p);
 212
 213                 p2m_top_mfn = alloc_p2m_page();
 214                 p2m_top_mfn_init(p2m_top_mfn);
 215         } else {
 216                 /* Reinitialise, mfn's all change after migration */
 217                 p2m_mid_mfn_init(p2m_mid_missing_mfn, p2m_missing);
 218         }
 219
 220         for (pfn = 0; pfn < xen_max_p2m_pfn && pfn < MAX_P2M_PFN;
 221              pfn += P2M_PER_PAGE) {
 222                 topidx = p2m_top_index(pfn);
 223                 mididx = p2m_mid_index(pfn);
 224
 225                 mid_mfn_p = p2m_top_mfn_p[topidx];
 226                 ptep = lookup_address((unsigned long)(xen_p2m_addr + pfn),
 227                                       &level);
 228                 BUG_ON(!ptep || level != PG_LEVEL_4K);
 229                 mfn = pte_mfn(*ptep);
 230                 ptep = (pte_t *)((unsigned long)ptep & ~(PAGE_SIZE - 1));
 231
 232                 /* Don't bother allocating any mfn mid levels if
 233                  * they're just missing, just update the stored mfn,
 234                  * since all could have changed over a migrate.
 235                  */
 236                 if (ptep == p2m_missing_pte || ptep == p2m_identity_pte) {
 237                         BUG_ON(mididx);
 238                         BUG_ON(mid_mfn_p != p2m_mid_missing_mfn);
 239                         p2m_top_mfn[topidx] = virt_to_mfn(p2m_mid_missing_mfn);
 240                         pfn += (P2M_MID_PER_PAGE - 1) * P2M_PER_PAGE;
 241                         continue;
 242                 }
 243
 244                 if (mid_mfn_p == p2m_mid_missing_mfn) {
 245                         mid_mfn_p = alloc_p2m_page();
 246                         p2m_mid_mfn_init(mid_mfn_p, p2m_missing);
 247
 248                         p2m_top_mfn_p[topidx] = mid_mfn_p;
 249                 }
 250
 251                 p2m_top_mfn[topidx] = virt_to_mfn(mid_mfn_p);
 252                 mid_mfn_p[mididx] = mfn;
 253         }
 254 }
 255
 256 void xen_setup_mfn_list_list(void)
 257 {
 258         if (xen_feature(XENFEAT_auto_translated_physmap))
 259                 return;
 260
 261         BUG_ON(HYPERVISOR_shared_info == &xen_dummy_shared_info);
 262
 263         HYPERVISOR_shared_info->arch.pfn_to_mfn_frame_list_list =
 264                 virt_to_mfn(p2m_top_mfn);
 265         HYPERVISOR_shared_info->arch.max_pfn = xen_max_p2m_pfn;
 266         HYPERVISOR_shared_info->arch.p2m_generation = 0;
 267         HYPERVISOR_shared_info->arch.p2m_vaddr = (unsigned long)xen_p2m_addr;
 268         HYPERVISOR_shared_info->arch.p2m_cr3 =
 269                 xen_pfn_to_cr3(virt_to_mfn(swapper_pg_dir));
 270 }
 271
 272 /* Set up p2m_top to point to the domain-builder provided p2m pages */
 273 void __init xen_build_dynamic_phys_to_machine(void)
 274 {
 275         unsigned long pfn;
 276
 277          if (xen_feature(XENFEAT_auto_translated_physmap))
 278                 return;
 279
 280         xen_p2m_addr = (unsigned long *)xen_start_info->mfn_list;
 281         xen_p2m_size = ALIGN(xen_start_info->nr_pages, P2M_PER_PAGE);
 282
 283         for (pfn = xen_start_info->nr_pages; pfn < xen_p2m_size; pfn++)
 284                 xen_p2m_addr[pfn] = INVALID_P2M_ENTRY;
 285
 286         xen_max_p2m_pfn = xen_p2m_size;
 287 }
 288
 289 #define P2M_TYPE_IDENTITY       0
 290 #define P2M_TYPE_MISSING        1
 291 #define P2M_TYPE_PFN            2
 292 #define P2M_TYPE_UNKNOWN        3
 293
 294 static int xen_p2m_elem_type(unsigned long pfn)
 295 {
 296         unsigned long mfn;
 297
 298         if (pfn >= xen_p2m_size)
 299                 return P2M_TYPE_IDENTITY;
 300
 301         mfn = xen_p2m_addr[pfn];
 302
 303         if (mfn == INVALID_P2M_ENTRY)
 304                 return P2M_TYPE_MISSING;
 305
 306         if (mfn & IDENTITY_FRAME_BIT)
 307                 return P2M_TYPE_IDENTITY;
 308
 309         return P2M_TYPE_PFN;
 310 }
 311
 312 static void __init xen_rebuild_p2m_list(unsigned long *p2m)
 313 {
 314         unsigned int i, chunk;
 315         unsigned long pfn;
 316         unsigned long *mfns;
 317         pte_t *ptep;
 318         pmd_t *pmdp;
 319         int type;
 320
 321         p2m_missing = alloc_p2m_page();
 322         p2m_init(p2m_missing);
 323         p2m_identity = alloc_p2m_page();
 324         p2m_init(p2m_identity);
 325
 326         p2m_missing_pte = alloc_p2m_page();
 327         paravirt_alloc_pte(&init_mm, __pa(p2m_missing_pte) >> PAGE_SHIFT);
 328         p2m_identity_pte = alloc_p2m_page();
 329         paravirt_alloc_pte(&init_mm, __pa(p2m_identity_pte) >> PAGE_SHIFT);
 330         for (i = 0; i < PTRS_PER_PTE; i++) {
 331                 set_pte(p2m_missing_pte + i,
 332                         pfn_pte(PFN_DOWN(__pa(p2m_missing)), PAGE_KERNEL_RO));
 333                 set_pte(p2m_identity_pte + i,
 334                         pfn_pte(PFN_DOWN(__pa(p2m_identity)), PAGE_KERNEL_RO));
 335         }
 336
 337         for (pfn = 0; pfn < xen_max_p2m_pfn; pfn += chunk) {
 338                 /*
 339                  * Try to map missing/identity PMDs or p2m-pages if possible.
 340                  * We have to respect the structure of the mfn_list_list
 341                  * which will be built just afterwards.
 342                  * Chunk size to test is one p2m page if we are in the middle
 343                  * of a mfn_list_list mid page and the complete mid page area
 344                  * if we are at index 0 of the mid page. Please note that a
 345                  * mid page might cover more than one PMD, e.g. on 32 bit PAE
 346                  * kernels.
 347                  */
 348                 chunk = (pfn & (P2M_PER_PAGE * P2M_MID_PER_PAGE - 1)) ?
 349                         P2M_PER_PAGE : P2M_PER_PAGE * P2M_MID_PER_PAGE;
 350
 351                 type = xen_p2m_elem_type(pfn);
 352                 i = 0;
 353                 if (type != P2M_TYPE_PFN)
 354                         for (i = 1; i < chunk; i++)
 355                                 if (xen_p2m_elem_type(pfn + i) != type)
 356                                         break;
 357                 if (i < chunk)
 358                         /* Reset to minimal chunk size. */
 359                         chunk = P2M_PER_PAGE;
 360
 361                 if (type == P2M_TYPE_PFN || i < chunk) {
 362                         /* Use initial p2m page contents. */
 363 #ifdef CONFIG_X86_64
 364                         mfns = alloc_p2m_page();
 365                         copy_page(mfns, xen_p2m_addr + pfn);
 366 #else
 367                         mfns = xen_p2m_addr + pfn;
 368 #endif
 369                         ptep = populate_extra_pte((unsigned long)(p2m + pfn));
 370                         set_pte(ptep,
 371                                 pfn_pte(PFN_DOWN(__pa(mfns)), PAGE_KERNEL));
 372                         continue;
 373                 }
 374
 375                 if (chunk == P2M_PER_PAGE) {
 376                         /* Map complete missing or identity p2m-page. */
 377                         mfns = (type == P2M_TYPE_MISSING) ?
 378                                 p2m_missing : p2m_identity;
 379                         ptep = populate_extra_pte((unsigned long)(p2m + pfn));
 380                         set_pte(ptep,
 381                                 pfn_pte(PFN_DOWN(__pa(mfns)), PAGE_KERNEL_RO));
 382                         continue;
 383                 }
 384
 385                 /* Complete missing or identity PMD(s) can be mapped. */
 386                 ptep = (type == P2M_TYPE_MISSING) ?
 387                         p2m_missing_pte : p2m_identity_pte;
 388                 for (i = 0; i < PMDS_PER_MID_PAGE; i++) {
 389                         pmdp = populate_extra_pmd(
 390                                 (unsigned long)(p2m + pfn) + i * PMD_SIZE);
 391                         set_pmd(pmdp, __pmd(__pa(ptep) | _KERNPG_TABLE));
 392                 }
 393         }
 394 }
 395
 396 void __init xen_vmalloc_p2m_tree(void)
 397 {
 398         static struct vm_struct vm;
 399         unsigned long p2m_limit;
 400
 401         p2m_limit = (phys_addr_t)P2M_LIMIT * 1024 * 1024 * 1024 / PAGE_SIZE;
 402         vm.flags = VM_ALLOC;
 403         vm.size = ALIGN(sizeof(unsigned long) * max(xen_max_p2m_pfn, p2m_limit),
 404                         PMD_SIZE * PMDS_PER_MID_PAGE);
 405         vm_area_register_early(&vm, PMD_SIZE * PMDS_PER_MID_PAGE);
 406         pr_notice("p2m virtual area at %p, size is %lx\n", vm.addr, vm.size);
 407
 408         xen_max_p2m_pfn = vm.size / sizeof(unsigned long);
 409
 410         xen_rebuild_p2m_list(vm.addr);
 411
 412         xen_p2m_addr = vm.addr;
 413         xen_p2m_size = xen_max_p2m_pfn;
 414
 415         xen_inv_extra_mem();
 416 }
 417
 418 unsigned long get_phys_to_machine(unsigned long pfn)
 419 {
 420         pte_t *ptep;
 421         unsigned int level;
 422
 423         if (unlikely(pfn >= xen_p2m_size)) {
 424                 if (pfn < xen_max_p2m_pfn)
 425                         return xen_chk_extra_mem(pfn);
 426
 427                 return IDENTITY_FRAME(pfn);
 428         }
 429
 430         ptep = lookup_address((unsigned long)(xen_p2m_addr + pfn), &level);
 431         BUG_ON(!ptep || level != PG_LEVEL_4K);
 432
 433         /*
 434          * The INVALID_P2M_ENTRY is filled in both p2m_*identity
 435          * and in p2m_*missing, so returning the INVALID_P2M_ENTRY
 436          * would be wrong.
 437          */
 438         if (pte_pfn(*ptep) == PFN_DOWN(__pa(p2m_identity)))
 439                 return IDENTITY_FRAME(pfn);
 440
 441         return xen_p2m_addr[pfn];
 442 }
 443 EXPORT_SYMBOL_GPL(get_phys_to_machine);
 444
 445 /*
 446  * Allocate new pmd(s). It is checked whether the old pmd is still in place.
 447  * If not, nothing is changed. This is okay as the only reason for allocating
 448  * a new pmd is to replace p2m_missing_pte or p2m_identity_pte by a individual
 449  * pmd. In case of PAE/x86-32 there are multiple pmds to allocate!
 450  */
 451 static pte_t *alloc_p2m_pmd(unsigned long addr, pte_t *pte_pg)
 452 {
 453         pte_t *ptechk;
 454         pte_t *pte_newpg[PMDS_PER_MID_PAGE];
 455         pmd_t *pmdp;
 456         unsigned int level;
 457         unsigned long flags;
 458         unsigned long vaddr;
 459         int i;
 460
 461         /* Do all allocations first to bail out in error case. */
 462         for (i = 0; i < PMDS_PER_MID_PAGE; i++) {
 463                 pte_newpg[i] = alloc_p2m_page();
 464                 if (!pte_newpg[i]) {
 465                         for (i--; i >= 0; i--)
 466                                 free_p2m_page(pte_newpg[i]);
 467
 468                         return NULL;
 469                 }
 470         }
 471
 472         vaddr = addr & ~(PMD_SIZE * PMDS_PER_MID_PAGE - 1);
 473
 474         for (i = 0; i < PMDS_PER_MID_PAGE; i++) {
 475                 copy_page(pte_newpg[i], pte_pg);
 476                 paravirt_alloc_pte(&init_mm, __pa(pte_newpg[i]) >> PAGE_SHIFT);
 477
 478                 pmdp = lookup_pmd_address(vaddr);
 479                 BUG_ON(!pmdp);
 480
 481                 spin_lock_irqsave(&p2m_update_lock, flags);
 482
 483                 ptechk = lookup_address(vaddr, &level);
 484                 if (ptechk == pte_pg) {
 485                         HYPERVISOR_shared_info->arch.p2m_generation++;
 486                         wmb(); /* Tools are synchronizing via p2m_generation. */
 487                         set_pmd(pmdp,
 488                                 __pmd(__pa(pte_newpg[i]) | _KERNPG_TABLE));
 489                         wmb(); /* Tools are synchronizing via p2m_generation. */
 490                         HYPERVISOR_shared_info->arch.p2m_generation++;
 491                         pte_newpg[i] = NULL;
 492                 }
 493
 494                 spin_unlock_irqrestore(&p2m_update_lock, flags);
 495
 496                 if (pte_newpg[i]) {
 497                         paravirt_release_pte(__pa(pte_newpg[i]) >> PAGE_SHIFT);
 498                         free_p2m_page(pte_newpg[i]);
 499                 }
 500
 501                 vaddr += PMD_SIZE;
 502         }
 503
 504         return lookup_address(addr, &level);
 505 }
 506
 507 /*
 508  * Fully allocate the p2m structure for a given pfn.  We need to check
 509  * that both the top and mid levels are allocated, and make sure the
 510  * parallel mfn tree is kept in sync.  We may race with other cpus, so
 511  * the new pages are installed with cmpxchg; if we lose the race then
 512  * simply free the page we allocated and use the one that's there.
 513  */
 514 static bool alloc_p2m(unsigned long pfn)
 515 {
 516         unsigned topidx, mididx;
 517         unsigned long *top_mfn_p, *mid_mfn;
 518         pte_t *ptep, *pte_pg;
 519         unsigned int level;
 520         unsigned long flags;
 521         unsigned long addr = (unsigned long)(xen_p2m_addr + pfn);
 522         unsigned long p2m_pfn;
 523
 524         topidx = p2m_top_index(pfn);
 525         mididx = p2m_mid_index(pfn);
 526
 527         ptep = lookup_address(addr, &level);
 528         BUG_ON(!ptep || level != PG_LEVEL_4K);
 529         pte_pg = (pte_t *)((unsigned long)ptep & ~(PAGE_SIZE - 1));
 530
 531         if (pte_pg == p2m_missing_pte || pte_pg == p2m_identity_pte) {
 532                 /* PMD level is missing, allocate a new one */
 533                 ptep = alloc_p2m_pmd(addr, pte_pg);
 534                 if (!ptep)
 535                         return false;
 536         }
 537
 538         if (p2m_top_mfn) {
 539                 top_mfn_p = &p2m_top_mfn[topidx];
 540                 mid_mfn = ACCESS_ONCE(p2m_top_mfn_p[topidx]);
 541
 542                 BUG_ON(virt_to_mfn(mid_mfn) != *top_mfn_p);
 543
 544                 if (mid_mfn == p2m_mid_missing_mfn) {
 545                         /* Separately check the mid mfn level */
 546                         unsigned long missing_mfn;
 547                         unsigned long mid_mfn_mfn;
 548                         unsigned long old_mfn;
 549
 550                         mid_mfn = alloc_p2m_page();
 551                         if (!mid_mfn)
 552                                 return false;
 553
 554                         p2m_mid_mfn_init(mid_mfn, p2m_missing);
 555
 556                         missing_mfn = virt_to_mfn(p2m_mid_missing_mfn);
 557                         mid_mfn_mfn = virt_to_mfn(mid_mfn);
 558                         old_mfn = cmpxchg(top_mfn_p, missing_mfn, mid_mfn_mfn);
 559                         if (old_mfn != missing_mfn) {
 560                                 free_p2m_page(mid_mfn);
 561                                 mid_mfn = mfn_to_virt(old_mfn);
 562                         } else {
 563                                 p2m_top_mfn_p[topidx] = mid_mfn;
 564                         }
 565                 }
 566         } else {
 567                 mid_mfn = NULL;
 568         }
 569
 570         p2m_pfn = pte_pfn(READ_ONCE(*ptep));
 571         if (p2m_pfn == PFN_DOWN(__pa(p2m_identity)) ||
 572             p2m_pfn == PFN_DOWN(__pa(p2m_missing))) {
 573                 /* p2m leaf page is missing */
 574                 unsigned long *p2m;
 575
 576                 p2m = alloc_p2m_page();
 577                 if (!p2m)
 578                         return false;
 579
 580                 if (p2m_pfn == PFN_DOWN(__pa(p2m_missing)))
 581                         p2m_init(p2m);
 582                 else
 583                         p2m_init_identity(p2m, pfn & ~(P2M_PER_PAGE - 1));
 584
 585                 spin_lock_irqsave(&p2m_update_lock, flags);
 586
 587                 if (pte_pfn(*ptep) == p2m_pfn) {
 588                         HYPERVISOR_shared_info->arch.p2m_generation++;
 589                         wmb(); /* Tools are synchronizing via p2m_generation. */
 590                         set_pte(ptep,
 591                                 pfn_pte(PFN_DOWN(__pa(p2m)), PAGE_KERNEL));
 592                         wmb(); /* Tools are synchronizing via p2m_generation. */
 593                         HYPERVISOR_shared_info->arch.p2m_generation++;
 594                         if (mid_mfn)
 595                                 mid_mfn[mididx] = virt_to_mfn(p2m);
 596                         p2m = NULL;
 597                 }
 598
 599                 spin_unlock_irqrestore(&p2m_update_lock, flags);
 600
 601                 if (p2m)
 602                         free_p2m_page(p2m);
 603         }
 604
 605         return true;
 606 }
 607
 608 unsigned long __init set_phys_range_identity(unsigned long pfn_s,
 609                                       unsigned long pfn_e)
 610 {
 611         unsigned long pfn;
 612
 613         if (unlikely(pfn_s >= xen_p2m_size))
 614                 return 0;
 615
 616         if (unlikely(xen_feature(XENFEAT_auto_translated_physmap)))
 617                 return pfn_e - pfn_s;
 618
 619         if (pfn_s > pfn_e)
 620                 return 0;
 621
 622         if (pfn_e > xen_p2m_size)
 623                 pfn_e = xen_p2m_size;
 624
 625         for (pfn = pfn_s; pfn < pfn_e; pfn++)
 626                 xen_p2m_addr[pfn] = IDENTITY_FRAME(pfn);
 627
 628         return pfn - pfn_s;
 629 }
 630
 631 bool __set_phys_to_machine(unsigned long pfn, unsigned long mfn)
 632 {
 633         pte_t *ptep;
 634         unsigned int level;
 635
 636         /* don't track P2M changes in autotranslate guests */
 637         if (unlikely(xen_feature(XENFEAT_auto_translated_physmap)))
 638                 return true;
 639
 640         if (unlikely(pfn >= xen_p2m_size)) {
 641                 BUG_ON(mfn != INVALID_P2M_ENTRY);
 642                 return true;
 643         }
 644
 645         /*
 646          * The interface requires atomic updates on p2m elements.
 647          * xen_safe_write_ulong() is using __put_user which does an atomic
 648          * store via asm().
 649          */
 650         if (likely(!xen_safe_write_ulong(xen_p2m_addr + pfn, mfn)))
 651                 return true;
 652
 653         ptep = lookup_address((unsigned long)(xen_p2m_addr + pfn), &level);
 654         BUG_ON(!ptep || level != PG_LEVEL_4K);
 655
 656         if (pte_pfn(*ptep) == PFN_DOWN(__pa(p2m_missing)))
 657                 return mfn == INVALID_P2M_ENTRY;
 658
 659         if (pte_pfn(*ptep) == PFN_DOWN(__pa(p2m_identity)))
 660                 return mfn == IDENTITY_FRAME(pfn);
 661
 662         return false;
 663 }
 664
 665 bool set_phys_to_machine(unsigned long pfn, unsigned long mfn)
 666 {
 667         if (unlikely(!__set_phys_to_machine(pfn, mfn))) {
 668                 if (!alloc_p2m(pfn))
 669                         return false;
 670
 671                 return __set_phys_to_machine(pfn, mfn);
 672         }
 673
 674         return true;
 675 }
 676
 677 int set_foreign_p2m_mapping(struct gnttab_map_grant_ref *map_ops,
 678                             struct gnttab_map_grant_ref *kmap_ops,
 679                             struct page **pages, unsigned int count)
 680 {
 681         int i, ret = 0;
 682         pte_t *pte;
 683
 684         if (xen_feature(XENFEAT_auto_translated_physmap))
 685                 return 0;
 686
 687         if (kmap_ops) {
 688                 ret = HYPERVISOR_grant_table_op(GNTTABOP_map_grant_ref,
 689                                                 kmap_ops, count);
 690                 if (ret)
 691                         goto out;
 692         }
 693
 694         for (i = 0; i < count; i++) {
 695                 unsigned long mfn, pfn;
 696
 697                 /* Do not add to override if the map failed. */
 698                 if (map_ops[i].status)
 699                         continue;
 700
 701                 if (map_ops[i].flags & GNTMAP_contains_pte) {
 702                         pte = (pte_t *)(mfn_to_virt(PFN_DOWN(map_ops[i].host_addr)) +
 703                                 (map_ops[i].host_addr & ~PAGE_MASK));
 704                         mfn = pte_mfn(*pte);
 705                 } else {
 706                         mfn = PFN_DOWN(map_ops[i].dev_bus_addr);
 707                 }
 708                 pfn = page_to_pfn(pages[i]);
 709
 710                 WARN(pfn_to_mfn(pfn) != INVALID_P2M_ENTRY, "page must be ballooned");
 711
 712                 if (unlikely(!set_phys_to_machine(pfn, FOREIGN_FRAME(mfn)))) {
 713                         ret = -ENOMEM;
 714                         goto out;
 715                 }
 716         }
 717
 718 out:
 719         return ret;
 720 }
 721 EXPORT_SYMBOL_GPL(set_foreign_p2m_mapping);
 722
 723 int clear_foreign_p2m_mapping(struct gnttab_unmap_grant_ref *unmap_ops,
 724                               struct gnttab_unmap_grant_ref *kunmap_ops,
 725                               struct page **pages, unsigned int count)
 726 {
 727         int i, ret = 0;
 728
 729         if (xen_feature(XENFEAT_auto_translated_physmap))
 730                 return 0;
 731
 732         for (i = 0; i < count; i++) {
 733                 unsigned long mfn = __pfn_to_mfn(page_to_pfn(pages[i]));
 734                 unsigned long pfn = page_to_pfn(pages[i]);
 735
 736                 if (mfn == INVALID_P2M_ENTRY || !(mfn & FOREIGN_FRAME_BIT)) {
 737                         ret = -EINVAL;
 738                         goto out;
 739                 }
 740
 741                 set_phys_to_machine(pfn, INVALID_P2M_ENTRY);
 742         }
 743         if (kunmap_ops)
 744                 ret = HYPERVISOR_grant_table_op(GNTTABOP_unmap_grant_ref,
 745                                                 kunmap_ops, count);
 746 out:
 747         return ret;
 748 }
 749 EXPORT_SYMBOL_GPL(clear_foreign_p2m_mapping);
 750
 751 #ifdef CONFIG_XEN_DEBUG_FS
 752 #include <linux/debugfs.h>
 753 #include "debugfs.h"
 754 static int p2m_dump_show(struct seq_file *m, void *v)
 755 {
 756         static const char * const type_name[] = {
 757                                 [P2M_TYPE_IDENTITY] = "identity",
 758                                 [P2M_TYPE_MISSING] = "missing",
 759                                 [P2M_TYPE_PFN] = "pfn",
 760                                 [P2M_TYPE_UNKNOWN] = "abnormal"};
 761         unsigned long pfn, first_pfn;
 762         int type, prev_type;
 763
 764         prev_type = xen_p2m_elem_type(0);
 765         first_pfn = 0;
 766
 767         for (pfn = 0; pfn < xen_p2m_size; pfn++) {
 768                 type = xen_p2m_elem_type(pfn);
 769                 if (type != prev_type) {
 770                         seq_printf(m, " [0x%lx->0x%lx] %s\n", first_pfn, pfn,
 771                                    type_name[prev_type]);
 772                         prev_type = type;
 773                         first_pfn = pfn;
 774                 }
 775         }
 776         seq_printf(m, " [0x%lx->0x%lx] %s\n", first_pfn, pfn,
 777                    type_name[prev_type]);
 778         return 0;
 779 }
 780
 781 static int p2m_dump_open(struct inode *inode, struct file *filp)
 782 {
 783         return single_open(filp, p2m_dump_show, NULL);
 784 }
 785
 786 static const struct file_operations p2m_dump_fops = {
 787         .open           = p2m_dump_open,
 788         .read           = seq_read,
 789         .llseek         = seq_lseek,
 790         .release        = single_release,
 791 };
 792
 793 static struct dentry *d_mmu_debug;
 794
 795 static int __init xen_p2m_debugfs(void)
 796 {
 797         struct dentry *d_xen = xen_init_debugfs();
 798
 799         if (d_xen == NULL)
 800                 return -ENOMEM;
 801
 802         d_mmu_debug = debugfs_create_dir("mmu", d_xen);
 803
 804         debugfs_create_file("p2m", 0600, d_mmu_debug, NULL, &p2m_dump_fops);
 805         return 0;
 806 }
 807 fs_initcall(xen_p2m_debugfs);
 808 #endif /* CONFIG_XEN_DEBUG_FS */