arch/x86/xen/p2m.c

   1 // SPDX-License-Identifier: GPL-2.0
   2
   3 /*
   4  * Xen leaves the responsibility for maintaining p2m mappings to the
   5  * guests themselves, but it must also access and update the p2m array
   6  * during suspend/resume when all the pages are reallocated.
   7  *
   8  * The logical flat p2m table is mapped to a linear kernel memory area.
   9  * For accesses by Xen a three-level tree linked via mfns only is set up to
  10  * allow the address space to be sparse.
  11  *
  12  *               Xen
  13  *                |
  14  *          p2m_top_mfn
  15  *              /   \
  16  * p2m_mid_mfn p2m_mid_mfn
  17  *         /           /
  18  *  p2m p2m p2m ...
  19  *
  20  * The p2m_mid_mfn pages are mapped by p2m_top_mfn_p.
  21  *
  22  * The p2m_top_mfn level is limited to 1 page, so the maximum representable
  23  * pseudo-physical address space is:
  24  *  P2M_TOP_PER_PAGE * P2M_MID_PER_PAGE * P2M_PER_PAGE pages
  25  *
  26  * P2M_PER_PAGE depends on the architecture, as a mfn is always
  27  * unsigned long (8 bytes on 64-bit, 4 bytes on 32), leading to
  28  * 512 and 1024 entries respectively.
  29  *
  30  * In short, these structures contain the Machine Frame Number (MFN) of the PFN.
  31  *
  32  * However not all entries are filled with MFNs. Specifically for all other
  33  * leaf entries, or for the top  root, or middle one, for which there is a void
  34  * entry, we assume it is  "missing". So (for example)
  35  *  pfn_to_mfn(0x90909090)=INVALID_P2M_ENTRY.
  36  * We have a dedicated page p2m_missing with all entries being
  37  * INVALID_P2M_ENTRY. This page may be referenced multiple times in the p2m
  38  * list/tree in case there are multiple areas with P2M_PER_PAGE invalid pfns.
  39  *
  40  * We also have the possibility of setting 1-1 mappings on certain regions, so
  41  * that:
  42  *  pfn_to_mfn(0xc0000)=0xc0000
  43  *
  44  * The benefit of this is, that we can assume for non-RAM regions (think
  45  * PCI BARs, or ACPI spaces), we can create mappings easily because we
  46  * get the PFN value to match the MFN.
  47  *
  48  * For this to work efficiently we have one new page p2m_identity. All entries
  49  * in p2m_identity are set to INVALID_P2M_ENTRY type (Xen toolstack only
  50  * recognizes that and MFNs, no other fancy value).
  51  *
  52  * On lookup we spot that the entry points to p2m_identity and return the
  53  * identity value instead of dereferencing and returning INVALID_P2M_ENTRY.
  54  * If the entry points to an allocated page, we just proceed as before and
  55  * return the PFN. If the PFN has IDENTITY_FRAME_BIT set we unmask that in
  56  * appropriate functions (pfn_to_mfn).
  57  *
  58  * The reason for having the IDENTITY_FRAME_BIT instead of just returning the
  59  * PFN is that we could find ourselves where pfn_to_mfn(pfn)==pfn for a
  60  * non-identity pfn. To protect ourselves against we elect to set (and get) the
  61  * IDENTITY_FRAME_BIT on all identity mapped PFNs.
  62  */
  63
  64 #include <linux/init.h>
  65 #include <linux/export.h>
  66 #include <linux/list.h>
  67 #include <linux/hash.h>
  68 #include <linux/sched.h>
  69 #include <linux/seq_file.h>
  70 #include <linux/bootmem.h>
  71 #include <linux/memblock.h>
  72 #include <linux/slab.h>
  73 #include <linux/vmalloc.h>
  74
  75 #include <asm/cache.h>
  76 #include <asm/setup.h>
  77 #include <linux/uaccess.h>
  78
  79 #include <asm/xen/page.h>
  80 #include <asm/xen/hypercall.h>
  81 #include <asm/xen/hypervisor.h>
  82 #include <xen/balloon.h>
  83 #include <xen/grant_table.h>
  84
  85 #include "multicalls.h"
  86 #include "xen-ops.h"
  87
  88 #define P2M_MID_PER_PAGE        (PAGE_SIZE / sizeof(unsigned long *))
  89 #define P2M_TOP_PER_PAGE        (PAGE_SIZE / sizeof(unsigned long **))
  90
  91 #define MAX_P2M_PFN     (P2M_TOP_PER_PAGE * P2M_MID_PER_PAGE * P2M_PER_PAGE)
  92
  93 #define PMDS_PER_MID_PAGE       (P2M_MID_PER_PAGE / PTRS_PER_PTE)
  94
  95 unsigned long *xen_p2m_addr __read_mostly;
  96 EXPORT_SYMBOL_GPL(xen_p2m_addr);
  97 unsigned long xen_p2m_size __read_mostly;
  98 EXPORT_SYMBOL_GPL(xen_p2m_size);
  99 unsigned long xen_max_p2m_pfn __read_mostly;
 100 EXPORT_SYMBOL_GPL(xen_max_p2m_pfn);
 101
 102 #ifdef CONFIG_XEN_BALLOON_MEMORY_HOTPLUG_LIMIT
 103 #define P2M_LIMIT CONFIG_XEN_BALLOON_MEMORY_HOTPLUG_LIMIT
 104 #else
 105 #define P2M_LIMIT 0
 106 #endif
 107
 108 static DEFINE_SPINLOCK(p2m_update_lock);
 109
 110 static unsigned long *p2m_mid_missing_mfn;
 111 static unsigned long *p2m_top_mfn;
 112 static unsigned long **p2m_top_mfn_p;
 113 static unsigned long *p2m_missing;
 114 static unsigned long *p2m_identity;
 115 static pte_t *p2m_missing_pte;
 116 static pte_t *p2m_identity_pte;
 117
 118 /*
 119  * Hint at last populated PFN.
 120  *
 121  * Used to set HYPERVISOR_shared_info->arch.max_pfn so the toolstack
 122  * can avoid scanning the whole P2M (which may be sized to account for
 123  * hotplugged memory).
 124  */
 125 static unsigned long xen_p2m_last_pfn;
 126
 127 static inline unsigned p2m_top_index(unsigned long pfn)
 128 {
 129         BUG_ON(pfn >= MAX_P2M_PFN);
 130         return pfn / (P2M_MID_PER_PAGE * P2M_PER_PAGE);
 131 }
 132
 133 static inline unsigned p2m_mid_index(unsigned long pfn)
 134 {
 135         return (pfn / P2M_PER_PAGE) % P2M_MID_PER_PAGE;
 136 }
 137
 138 static inline unsigned p2m_index(unsigned long pfn)
 139 {
 140         return pfn % P2M_PER_PAGE;
 141 }
 142
 143 static void p2m_top_mfn_init(unsigned long *top)
 144 {
 145         unsigned i;
 146
 147         for (i = 0; i < P2M_TOP_PER_PAGE; i++)
 148                 top[i] = virt_to_mfn(p2m_mid_missing_mfn);
 149 }
 150
 151 static void p2m_top_mfn_p_init(unsigned long **top)
 152 {
 153         unsigned i;
 154
 155         for (i = 0; i < P2M_TOP_PER_PAGE; i++)
 156                 top[i] = p2m_mid_missing_mfn;
 157 }
 158
 159 static void p2m_mid_mfn_init(unsigned long *mid, unsigned long *leaf)
 160 {
 161         unsigned i;
 162
 163         for (i = 0; i < P2M_MID_PER_PAGE; i++)
 164                 mid[i] = virt_to_mfn(leaf);
 165 }
 166
 167 static void p2m_init(unsigned long *p2m)
 168 {
 169         unsigned i;
 170
 171         for (i = 0; i < P2M_PER_PAGE; i++)
 172                 p2m[i] = INVALID_P2M_ENTRY;
 173 }
 174
 175 static void p2m_init_identity(unsigned long *p2m, unsigned long pfn)
 176 {
 177         unsigned i;
 178
 179         for (i = 0; i < P2M_PER_PAGE; i++)
 180                 p2m[i] = IDENTITY_FRAME(pfn + i);
 181 }
 182
 183 static void * __ref alloc_p2m_page(void)
 184 {
 185         if (unlikely(!slab_is_available()))
 186                 return memblock_alloc(PAGE_SIZE, PAGE_SIZE);
 187
 188         return (void *)__get_free_page(GFP_KERNEL);
 189 }
 190
 191 static void __ref free_p2m_page(void *p)
 192 {
 193         if (unlikely(!slab_is_available())) {
 194                 memblock_free((unsigned long)p, PAGE_SIZE);
 195                 return;
 196         }
 197
 198         free_page((unsigned long)p);
 199 }
 200
 201 /*
 202  * Build the parallel p2m_top_mfn and p2m_mid_mfn structures
 203  *
 204  * This is called both at boot time, and after resuming from suspend:
 205  * - At boot time we're called rather early, and must use alloc_bootmem*()
 206  *   to allocate memory.
 207  *
 208  * - After resume we're called from within stop_machine, but the mfn
 209  *   tree should already be completely allocated.
 210  */
 211 void __ref xen_build_mfn_list_list(void)
 212 {
 213         unsigned long pfn, mfn;
 214         pte_t *ptep;
 215         unsigned int level, topidx, mididx;
 216         unsigned long *mid_mfn_p;
 217
 218         if (xen_start_info->flags & SIF_VIRT_P2M_4TOOLS)
 219                 return;
 220
 221         /* Pre-initialize p2m_top_mfn to be completely missing */
 222         if (p2m_top_mfn == NULL) {
 223                 p2m_mid_missing_mfn = alloc_p2m_page();
 224                 p2m_mid_mfn_init(p2m_mid_missing_mfn, p2m_missing);
 225
 226                 p2m_top_mfn_p = alloc_p2m_page();
 227                 p2m_top_mfn_p_init(p2m_top_mfn_p);
 228
 229                 p2m_top_mfn = alloc_p2m_page();
 230                 p2m_top_mfn_init(p2m_top_mfn);
 231         } else {
 232                 /* Reinitialise, mfn's all change after migration */
 233                 p2m_mid_mfn_init(p2m_mid_missing_mfn, p2m_missing);
 234         }
 235
 236         for (pfn = 0; pfn < xen_max_p2m_pfn && pfn < MAX_P2M_PFN;
 237              pfn += P2M_PER_PAGE) {
 238                 topidx = p2m_top_index(pfn);
 239                 mididx = p2m_mid_index(pfn);
 240
 241                 mid_mfn_p = p2m_top_mfn_p[topidx];
 242                 ptep = lookup_address((unsigned long)(xen_p2m_addr + pfn),
 243                                       &level);
 244                 BUG_ON(!ptep || level != PG_LEVEL_4K);
 245                 mfn = pte_mfn(*ptep);
 246                 ptep = (pte_t *)((unsigned long)ptep & ~(PAGE_SIZE - 1));
 247
 248                 /* Don't bother allocating any mfn mid levels if
 249                  * they're just missing, just update the stored mfn,
 250                  * since all could have changed over a migrate.
 251                  */
 252                 if (ptep == p2m_missing_pte || ptep == p2m_identity_pte) {
 253                         BUG_ON(mididx);
 254                         BUG_ON(mid_mfn_p != p2m_mid_missing_mfn);
 255                         p2m_top_mfn[topidx] = virt_to_mfn(p2m_mid_missing_mfn);
 256                         pfn += (P2M_MID_PER_PAGE - 1) * P2M_PER_PAGE;
 257                         continue;
 258                 }
 259
 260                 if (mid_mfn_p == p2m_mid_missing_mfn) {
 261                         mid_mfn_p = alloc_p2m_page();
 262                         p2m_mid_mfn_init(mid_mfn_p, p2m_missing);
 263
 264                         p2m_top_mfn_p[topidx] = mid_mfn_p;
 265                 }
 266
 267                 p2m_top_mfn[topidx] = virt_to_mfn(mid_mfn_p);
 268                 mid_mfn_p[mididx] = mfn;
 269         }
 270 }
 271
 272 void xen_setup_mfn_list_list(void)
 273 {
 274         BUG_ON(HYPERVISOR_shared_info == &xen_dummy_shared_info);
 275
 276         if (xen_start_info->flags & SIF_VIRT_P2M_4TOOLS)
 277                 HYPERVISOR_shared_info->arch.pfn_to_mfn_frame_list_list = ~0UL;
 278         else
 279                 HYPERVISOR_shared_info->arch.pfn_to_mfn_frame_list_list =
 280                         virt_to_mfn(p2m_top_mfn);
 281         HYPERVISOR_shared_info->arch.max_pfn = xen_p2m_last_pfn;
 282         HYPERVISOR_shared_info->arch.p2m_generation = 0;
 283         HYPERVISOR_shared_info->arch.p2m_vaddr = (unsigned long)xen_p2m_addr;
 284         HYPERVISOR_shared_info->arch.p2m_cr3 =
 285                 xen_pfn_to_cr3(virt_to_mfn(swapper_pg_dir));
 286 }
 287
 288 /* Set up p2m_top to point to the domain-builder provided p2m pages */
 289 void __init xen_build_dynamic_phys_to_machine(void)
 290 {
 291         unsigned long pfn;
 292
 293         xen_p2m_addr = (unsigned long *)xen_start_info->mfn_list;
 294         xen_p2m_size = ALIGN(xen_start_info->nr_pages, P2M_PER_PAGE);
 295
 296         for (pfn = xen_start_info->nr_pages; pfn < xen_p2m_size; pfn++)
 297                 xen_p2m_addr[pfn] = INVALID_P2M_ENTRY;
 298
 299         xen_max_p2m_pfn = xen_p2m_size;
 300 }
 301
 302 #define P2M_TYPE_IDENTITY       0
 303 #define P2M_TYPE_MISSING        1
 304 #define P2M_TYPE_PFN            2
 305 #define P2M_TYPE_UNKNOWN        3
 306
 307 static int xen_p2m_elem_type(unsigned long pfn)
 308 {
 309         unsigned long mfn;
 310
 311         if (pfn >= xen_p2m_size)
 312                 return P2M_TYPE_IDENTITY;
 313
 314         mfn = xen_p2m_addr[pfn];
 315
 316         if (mfn == INVALID_P2M_ENTRY)
 317                 return P2M_TYPE_MISSING;
 318
 319         if (mfn & IDENTITY_FRAME_BIT)
 320                 return P2M_TYPE_IDENTITY;
 321
 322         return P2M_TYPE_PFN;
 323 }
 324
 325 static void __init xen_rebuild_p2m_list(unsigned long *p2m)
 326 {
 327         unsigned int i, chunk;
 328         unsigned long pfn;
 329         unsigned long *mfns;
 330         pte_t *ptep;
 331         pmd_t *pmdp;
 332         int type;
 333
 334         p2m_missing = alloc_p2m_page();
 335         p2m_init(p2m_missing);
 336         p2m_identity = alloc_p2m_page();
 337         p2m_init(p2m_identity);
 338
 339         p2m_missing_pte = alloc_p2m_page();
 340         paravirt_alloc_pte(&init_mm, __pa(p2m_missing_pte) >> PAGE_SHIFT);
 341         p2m_identity_pte = alloc_p2m_page();
 342         paravirt_alloc_pte(&init_mm, __pa(p2m_identity_pte) >> PAGE_SHIFT);
 343         for (i = 0; i < PTRS_PER_PTE; i++) {
 344                 set_pte(p2m_missing_pte + i,
 345                         pfn_pte(PFN_DOWN(__pa(p2m_missing)), PAGE_KERNEL_RO));
 346                 set_pte(p2m_identity_pte + i,
 347                         pfn_pte(PFN_DOWN(__pa(p2m_identity)), PAGE_KERNEL_RO));
 348         }
 349
 350         for (pfn = 0; pfn < xen_max_p2m_pfn; pfn += chunk) {
 351                 /*
 352                  * Try to map missing/identity PMDs or p2m-pages if possible.
 353                  * We have to respect the structure of the mfn_list_list
 354                  * which will be built just afterwards.
 355                  * Chunk size to test is one p2m page if we are in the middle
 356                  * of a mfn_list_list mid page and the complete mid page area
 357                  * if we are at index 0 of the mid page. Please note that a
 358                  * mid page might cover more than one PMD, e.g. on 32 bit PAE
 359                  * kernels.
 360                  */
 361                 chunk = (pfn & (P2M_PER_PAGE * P2M_MID_PER_PAGE - 1)) ?
 362                         P2M_PER_PAGE : P2M_PER_PAGE * P2M_MID_PER_PAGE;
 363
 364                 type = xen_p2m_elem_type(pfn);
 365                 i = 0;
 366                 if (type != P2M_TYPE_PFN)
 367                         for (i = 1; i < chunk; i++)
 368                                 if (xen_p2m_elem_type(pfn + i) != type)
 369                                         break;
 370                 if (i < chunk)
 371                         /* Reset to minimal chunk size. */
 372                         chunk = P2M_PER_PAGE;
 373
 374                 if (type == P2M_TYPE_PFN || i < chunk) {
 375                         /* Use initial p2m page contents. */
 376 #ifdef CONFIG_X86_64
 377                         mfns = alloc_p2m_page();
 378                         copy_page(mfns, xen_p2m_addr + pfn);
 379 #else
 380                         mfns = xen_p2m_addr + pfn;
 381 #endif
 382                         ptep = populate_extra_pte((unsigned long)(p2m + pfn));
 383                         set_pte(ptep,
 384                                 pfn_pte(PFN_DOWN(__pa(mfns)), PAGE_KERNEL));
 385                         continue;
 386                 }
 387
 388                 if (chunk == P2M_PER_PAGE) {
 389                         /* Map complete missing or identity p2m-page. */
 390                         mfns = (type == P2M_TYPE_MISSING) ?
 391                                 p2m_missing : p2m_identity;
 392                         ptep = populate_extra_pte((unsigned long)(p2m + pfn));
 393                         set_pte(ptep,
 394                                 pfn_pte(PFN_DOWN(__pa(mfns)), PAGE_KERNEL_RO));
 395                         continue;
 396                 }
 397
 398                 /* Complete missing or identity PMD(s) can be mapped. */
 399                 ptep = (type == P2M_TYPE_MISSING) ?
 400                         p2m_missing_pte : p2m_identity_pte;
 401                 for (i = 0; i < PMDS_PER_MID_PAGE; i++) {
 402                         pmdp = populate_extra_pmd(
 403                                 (unsigned long)(p2m + pfn) + i * PMD_SIZE);
 404                         set_pmd(pmdp, __pmd(__pa(ptep) | _KERNPG_TABLE));
 405                 }
 406         }
 407 }
 408
 409 void __init xen_vmalloc_p2m_tree(void)
 410 {
 411         static struct vm_struct vm;
 412         unsigned long p2m_limit;
 413
 414         xen_p2m_last_pfn = xen_max_p2m_pfn;
 415
 416         p2m_limit = (phys_addr_t)P2M_LIMIT * 1024 * 1024 * 1024 / PAGE_SIZE;
 417         vm.flags = VM_ALLOC;
 418         vm.size = ALIGN(sizeof(unsigned long) * max(xen_max_p2m_pfn, p2m_limit),
 419                         PMD_SIZE * PMDS_PER_MID_PAGE);
 420         vm_area_register_early(&vm, PMD_SIZE * PMDS_PER_MID_PAGE);
 421         pr_notice("p2m virtual area at %p, size is %lx\n", vm.addr, vm.size);
 422
 423         xen_max_p2m_pfn = vm.size / sizeof(unsigned long);
 424
 425         xen_rebuild_p2m_list(vm.addr);
 426
 427         xen_p2m_addr = vm.addr;
 428         xen_p2m_size = xen_max_p2m_pfn;
 429
 430         xen_inv_extra_mem();
 431 }
 432
 433 unsigned long get_phys_to_machine(unsigned long pfn)
 434 {
 435         pte_t *ptep;
 436         unsigned int level;
 437
 438         if (unlikely(pfn >= xen_p2m_size)) {
 439                 if (pfn < xen_max_p2m_pfn)
 440                         return xen_chk_extra_mem(pfn);
 441
 442                 return IDENTITY_FRAME(pfn);
 443         }
 444
 445         ptep = lookup_address((unsigned long)(xen_p2m_addr + pfn), &level);
 446         BUG_ON(!ptep || level != PG_LEVEL_4K);
 447
 448         /*
 449          * The INVALID_P2M_ENTRY is filled in both p2m_*identity
 450          * and in p2m_*missing, so returning the INVALID_P2M_ENTRY
 451          * would be wrong.
 452          */
 453         if (pte_pfn(*ptep) == PFN_DOWN(__pa(p2m_identity)))
 454                 return IDENTITY_FRAME(pfn);
 455
 456         return xen_p2m_addr[pfn];
 457 }
 458 EXPORT_SYMBOL_GPL(get_phys_to_machine);
 459
 460 /*
 461  * Allocate new pmd(s). It is checked whether the old pmd is still in place.
 462  * If not, nothing is changed. This is okay as the only reason for allocating
 463  * a new pmd is to replace p2m_missing_pte or p2m_identity_pte by a individual
 464  * pmd. In case of PAE/x86-32 there are multiple pmds to allocate!
 465  */
 466 static pte_t *alloc_p2m_pmd(unsigned long addr, pte_t *pte_pg)
 467 {
 468         pte_t *ptechk;
 469         pte_t *pte_newpg[PMDS_PER_MID_PAGE];
 470         pmd_t *pmdp;
 471         unsigned int level;
 472         unsigned long flags;
 473         unsigned long vaddr;
 474         int i;
 475
 476         /* Do all allocations first to bail out in error case. */
 477         for (i = 0; i < PMDS_PER_MID_PAGE; i++) {
 478                 pte_newpg[i] = alloc_p2m_page();
 479                 if (!pte_newpg[i]) {
 480                         for (i--; i >= 0; i--)
 481                                 free_p2m_page(pte_newpg[i]);
 482
 483                         return NULL;
 484                 }
 485         }
 486
 487         vaddr = addr & ~(PMD_SIZE * PMDS_PER_MID_PAGE - 1);
 488
 489         for (i = 0; i < PMDS_PER_MID_PAGE; i++) {
 490                 copy_page(pte_newpg[i], pte_pg);
 491                 paravirt_alloc_pte(&init_mm, __pa(pte_newpg[i]) >> PAGE_SHIFT);
 492
 493                 pmdp = lookup_pmd_address(vaddr);
 494                 BUG_ON(!pmdp);
 495
 496                 spin_lock_irqsave(&p2m_update_lock, flags);
 497
 498                 ptechk = lookup_address(vaddr, &level);
 499                 if (ptechk == pte_pg) {
 500                         HYPERVISOR_shared_info->arch.p2m_generation++;
 501                         wmb(); /* Tools are synchronizing via p2m_generation. */
 502                         set_pmd(pmdp,
 503                                 __pmd(__pa(pte_newpg[i]) | _KERNPG_TABLE));
 504                         wmb(); /* Tools are synchronizing via p2m_generation. */
 505                         HYPERVISOR_shared_info->arch.p2m_generation++;
 506                         pte_newpg[i] = NULL;
 507                 }
 508
 509                 spin_unlock_irqrestore(&p2m_update_lock, flags);
 510
 511                 if (pte_newpg[i]) {
 512                         paravirt_release_pte(__pa(pte_newpg[i]) >> PAGE_SHIFT);
 513                         free_p2m_page(pte_newpg[i]);
 514                 }
 515
 516                 vaddr += PMD_SIZE;
 517         }
 518
 519         return lookup_address(addr, &level);
 520 }
 521
 522 /*
 523  * Fully allocate the p2m structure for a given pfn.  We need to check
 524  * that both the top and mid levels are allocated, and make sure the
 525  * parallel mfn tree is kept in sync.  We may race with other cpus, so
 526  * the new pages are installed with cmpxchg; if we lose the race then
 527  * simply free the page we allocated and use the one that's there.
 528  */
 529 int xen_alloc_p2m_entry(unsigned long pfn)
 530 {
 531         unsigned topidx;
 532         unsigned long *top_mfn_p, *mid_mfn;
 533         pte_t *ptep, *pte_pg;
 534         unsigned int level;
 535         unsigned long flags;
 536         unsigned long addr = (unsigned long)(xen_p2m_addr + pfn);
 537         unsigned long p2m_pfn;
 538
 539         ptep = lookup_address(addr, &level);
 540         BUG_ON(!ptep || level != PG_LEVEL_4K);
 541         pte_pg = (pte_t *)((unsigned long)ptep & ~(PAGE_SIZE - 1));
 542
 543         if (pte_pg == p2m_missing_pte || pte_pg == p2m_identity_pte) {
 544                 /* PMD level is missing, allocate a new one */
 545                 ptep = alloc_p2m_pmd(addr, pte_pg);
 546                 if (!ptep)
 547                         return -ENOMEM;
 548         }
 549
 550         if (p2m_top_mfn && pfn < MAX_P2M_PFN) {
 551                 topidx = p2m_top_index(pfn);
 552                 top_mfn_p = &p2m_top_mfn[topidx];
 553                 mid_mfn = READ_ONCE(p2m_top_mfn_p[topidx]);
 554
 555                 BUG_ON(virt_to_mfn(mid_mfn) != *top_mfn_p);
 556
 557                 if (mid_mfn == p2m_mid_missing_mfn) {
 558                         /* Separately check the mid mfn level */
 559                         unsigned long missing_mfn;
 560                         unsigned long mid_mfn_mfn;
 561                         unsigned long old_mfn;
 562
 563                         mid_mfn = alloc_p2m_page();
 564                         if (!mid_mfn)
 565                                 return -ENOMEM;
 566
 567                         p2m_mid_mfn_init(mid_mfn, p2m_missing);
 568
 569                         missing_mfn = virt_to_mfn(p2m_mid_missing_mfn);
 570                         mid_mfn_mfn = virt_to_mfn(mid_mfn);
 571                         old_mfn = cmpxchg(top_mfn_p, missing_mfn, mid_mfn_mfn);
 572                         if (old_mfn != missing_mfn) {
 573                                 free_p2m_page(mid_mfn);
 574                                 mid_mfn = mfn_to_virt(old_mfn);
 575                         } else {
 576                                 p2m_top_mfn_p[topidx] = mid_mfn;
 577                         }
 578                 }
 579         } else {
 580                 mid_mfn = NULL;
 581         }
 582
 583         p2m_pfn = pte_pfn(READ_ONCE(*ptep));
 584         if (p2m_pfn == PFN_DOWN(__pa(p2m_identity)) ||
 585             p2m_pfn == PFN_DOWN(__pa(p2m_missing))) {
 586                 /* p2m leaf page is missing */
 587                 unsigned long *p2m;
 588
 589                 p2m = alloc_p2m_page();
 590                 if (!p2m)
 591                         return -ENOMEM;
 592
 593                 if (p2m_pfn == PFN_DOWN(__pa(p2m_missing)))
 594                         p2m_init(p2m);
 595                 else
 596                         p2m_init_identity(p2m, pfn & ~(P2M_PER_PAGE - 1));
 597
 598                 spin_lock_irqsave(&p2m_update_lock, flags);
 599
 600                 if (pte_pfn(*ptep) == p2m_pfn) {
 601                         HYPERVISOR_shared_info->arch.p2m_generation++;
 602                         wmb(); /* Tools are synchronizing via p2m_generation. */
 603                         set_pte(ptep,
 604                                 pfn_pte(PFN_DOWN(__pa(p2m)), PAGE_KERNEL));
 605                         wmb(); /* Tools are synchronizing via p2m_generation. */
 606                         HYPERVISOR_shared_info->arch.p2m_generation++;
 607                         if (mid_mfn)
 608                                 mid_mfn[p2m_mid_index(pfn)] = virt_to_mfn(p2m);
 609                         p2m = NULL;
 610                 }
 611
 612                 spin_unlock_irqrestore(&p2m_update_lock, flags);
 613
 614                 if (p2m)
 615                         free_p2m_page(p2m);
 616         }
 617
 618         /* Expanded the p2m? */
 619         if (pfn > xen_p2m_last_pfn) {
 620                 xen_p2m_last_pfn = pfn;
 621                 HYPERVISOR_shared_info->arch.max_pfn = xen_p2m_last_pfn;
 622         }
 623
 624         return 0;
 625 }
 626 EXPORT_SYMBOL(xen_alloc_p2m_entry);
 627
 628 unsigned long __init set_phys_range_identity(unsigned long pfn_s,
 629                                       unsigned long pfn_e)
 630 {
 631         unsigned long pfn;
 632
 633         if (unlikely(pfn_s >= xen_p2m_size))
 634                 return 0;
 635
 636         if (pfn_s > pfn_e)
 637                 return 0;
 638
 639         if (pfn_e > xen_p2m_size)
 640                 pfn_e = xen_p2m_size;
 641
 642         for (pfn = pfn_s; pfn < pfn_e; pfn++)
 643                 xen_p2m_addr[pfn] = IDENTITY_FRAME(pfn);
 644
 645         return pfn - pfn_s;
 646 }
 647
 648 bool __set_phys_to_machine(unsigned long pfn, unsigned long mfn)
 649 {
 650         pte_t *ptep;
 651         unsigned int level;
 652
 653         if (unlikely(pfn >= xen_p2m_size)) {
 654                 BUG_ON(mfn != INVALID_P2M_ENTRY);
 655                 return true;
 656         }
 657
 658         /*
 659          * The interface requires atomic updates on p2m elements.
 660          * xen_safe_write_ulong() is using __put_user which does an atomic
 661          * store via asm().
 662          */
 663         if (likely(!xen_safe_write_ulong(xen_p2m_addr + pfn, mfn)))
 664                 return true;
 665
 666         ptep = lookup_address((unsigned long)(xen_p2m_addr + pfn), &level);
 667         BUG_ON(!ptep || level != PG_LEVEL_4K);
 668
 669         if (pte_pfn(*ptep) == PFN_DOWN(__pa(p2m_missing)))
 670                 return mfn == INVALID_P2M_ENTRY;
 671
 672         if (pte_pfn(*ptep) == PFN_DOWN(__pa(p2m_identity)))
 673                 return mfn == IDENTITY_FRAME(pfn);
 674
 675         return false;
 676 }
 677
 678 bool set_phys_to_machine(unsigned long pfn, unsigned long mfn)
 679 {
 680         if (unlikely(!__set_phys_to_machine(pfn, mfn))) {
 681                 int ret;
 682
 683                 ret = xen_alloc_p2m_entry(pfn);
 684                 if (ret < 0)
 685                         return false;
 686
 687                 return __set_phys_to_machine(pfn, mfn);
 688         }
 689
 690         return true;
 691 }
 692
 693 int set_foreign_p2m_mapping(struct gnttab_map_grant_ref *map_ops,
 694                             struct gnttab_map_grant_ref *kmap_ops,
 695                             struct page **pages, unsigned int count)
 696 {
 697         int i, ret = 0;
 698         pte_t *pte;
 699
 700         if (xen_feature(XENFEAT_auto_translated_physmap))
 701                 return 0;
 702
 703         if (kmap_ops) {
 704                 ret = HYPERVISOR_grant_table_op(GNTTABOP_map_grant_ref,
 705                                                 kmap_ops, count);
 706                 if (ret)
 707                         goto out;
 708         }
 709
 710         for (i = 0; i < count; i++) {
 711                 unsigned long mfn, pfn;
 712
 713                 /* Do not add to override if the map failed. */
 714                 if (map_ops[i].status)
 715                         continue;
 716
 717                 if (map_ops[i].flags & GNTMAP_contains_pte) {
 718                         pte = (pte_t *)(mfn_to_virt(PFN_DOWN(map_ops[i].host_addr)) +
 719                                 (map_ops[i].host_addr & ~PAGE_MASK));
 720                         mfn = pte_mfn(*pte);
 721                 } else {
 722                         mfn = PFN_DOWN(map_ops[i].dev_bus_addr);
 723                 }
 724                 pfn = page_to_pfn(pages[i]);
 725
 726                 WARN(pfn_to_mfn(pfn) != INVALID_P2M_ENTRY, "page must be ballooned");
 727
 728                 if (unlikely(!set_phys_to_machine(pfn, FOREIGN_FRAME(mfn)))) {
 729                         ret = -ENOMEM;
 730                         goto out;
 731                 }
 732         }
 733
 734 out:
 735         return ret;
 736 }
 737 EXPORT_SYMBOL_GPL(set_foreign_p2m_mapping);
 738
 739 int clear_foreign_p2m_mapping(struct gnttab_unmap_grant_ref *unmap_ops,
 740                               struct gnttab_unmap_grant_ref *kunmap_ops,
 741                               struct page **pages, unsigned int count)
 742 {
 743         int i, ret = 0;
 744
 745         if (xen_feature(XENFEAT_auto_translated_physmap))
 746                 return 0;
 747
 748         for (i = 0; i < count; i++) {
 749                 unsigned long mfn = __pfn_to_mfn(page_to_pfn(pages[i]));
 750                 unsigned long pfn = page_to_pfn(pages[i]);
 751
 752                 if (mfn == INVALID_P2M_ENTRY || !(mfn & FOREIGN_FRAME_BIT)) {
 753                         ret = -EINVAL;
 754                         goto out;
 755                 }
 756
 757                 set_phys_to_machine(pfn, INVALID_P2M_ENTRY);
 758         }
 759         if (kunmap_ops)
 760                 ret = HYPERVISOR_grant_table_op(GNTTABOP_unmap_grant_ref,
 761                                                 kunmap_ops, count);
 762 out:
 763         return ret;
 764 }
 765 EXPORT_SYMBOL_GPL(clear_foreign_p2m_mapping);
 766
 767 #ifdef CONFIG_XEN_DEBUG_FS
 768 #include <linux/debugfs.h>
 769 #include "debugfs.h"
 770 static int p2m_dump_show(struct seq_file *m, void *v)
 771 {
 772         static const char * const type_name[] = {
 773                                 [P2M_TYPE_IDENTITY] = "identity",
 774                                 [P2M_TYPE_MISSING] = "missing",
 775                                 [P2M_TYPE_PFN] = "pfn",
 776                                 [P2M_TYPE_UNKNOWN] = "abnormal"};
 777         unsigned long pfn, first_pfn;
 778         int type, prev_type;
 779
 780         prev_type = xen_p2m_elem_type(0);
 781         first_pfn = 0;
 782
 783         for (pfn = 0; pfn < xen_p2m_size; pfn++) {
 784                 type = xen_p2m_elem_type(pfn);
 785                 if (type != prev_type) {
 786                         seq_printf(m, " [0x%lx->0x%lx] %s\n", first_pfn, pfn,
 787                                    type_name[prev_type]);
 788                         prev_type = type;
 789                         first_pfn = pfn;
 790                 }
 791         }
 792         seq_printf(m, " [0x%lx->0x%lx] %s\n", first_pfn, pfn,
 793                    type_name[prev_type]);
 794         return 0;
 795 }
 796
 797 static int p2m_dump_open(struct inode *inode, struct file *filp)
 798 {
 799         return single_open(filp, p2m_dump_show, NULL);
 800 }
 801
 802 static const struct file_operations p2m_dump_fops = {
 803         .open           = p2m_dump_open,
 804         .read           = seq_read,
 805         .llseek         = seq_lseek,
 806         .release        = single_release,
 807 };
 808
 809 static struct dentry *d_mmu_debug;
 810
 811 static int __init xen_p2m_debugfs(void)
 812 {
 813         struct dentry *d_xen = xen_init_debugfs();
 814
 815         if (d_xen == NULL)
 816                 return -ENOMEM;
 817
 818         d_mmu_debug = debugfs_create_dir("mmu", d_xen);
 819
 820         debugfs_create_file("p2m", 0600, d_mmu_debug, NULL, &p2m_dump_fops);
 821         return 0;
 822 }
 823 fs_initcall(xen_p2m_debugfs);
 824 #endif /* CONFIG_XEN_DEBUG_FS */