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