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
  71 #include <asm/cache.h>
  72 #include <asm/setup.h>
  73 #include <asm/uaccess.h>
  74
  75 #include <asm/xen/page.h>
  76 #include <asm/xen/hypercall.h>
  77 #include <asm/xen/hypervisor.h>
  78 #include <xen/balloon.h>
  79 #include <xen/grant_table.h>
  80
  81 #include "p2m.h"
  82 #include "multicalls.h"
  83 #include "xen-ops.h"
  84
  85 #define PMDS_PER_MID_PAGE       (P2M_MID_PER_PAGE / PTRS_PER_PTE)
  86
  87 static void __init m2p_override_init(void);
  88
  89 unsigned long *xen_p2m_addr __read_mostly;
  90 EXPORT_SYMBOL_GPL(xen_p2m_addr);
  91 unsigned long xen_p2m_size __read_mostly;
  92 EXPORT_SYMBOL_GPL(xen_p2m_size);
  93 unsigned long xen_max_p2m_pfn __read_mostly;
  94 EXPORT_SYMBOL_GPL(xen_max_p2m_pfn);
  95
  96 static DEFINE_SPINLOCK(p2m_update_lock);
  97
  98 static unsigned long *p2m_mid_missing_mfn;
  99 static unsigned long *p2m_top_mfn;
 100 static unsigned long **p2m_top_mfn_p;
 101 static unsigned long *p2m_missing;
 102 static unsigned long *p2m_identity;
 103 static pte_t *p2m_missing_pte;
 104 static pte_t *p2m_identity_pte;
 105
 106 static inline unsigned p2m_top_index(unsigned long pfn)
 107 {
 108         BUG_ON(pfn >= MAX_P2M_PFN);
 109         return pfn / (P2M_MID_PER_PAGE * P2M_PER_PAGE);
 110 }
 111
 112 static inline unsigned p2m_mid_index(unsigned long pfn)
 113 {
 114         return (pfn / P2M_PER_PAGE) % P2M_MID_PER_PAGE;
 115 }
 116
 117 static inline unsigned p2m_index(unsigned long pfn)
 118 {
 119         return pfn % P2M_PER_PAGE;
 120 }
 121
 122 static void p2m_top_mfn_init(unsigned long *top)
 123 {
 124         unsigned i;
 125
 126         for (i = 0; i < P2M_TOP_PER_PAGE; i++)
 127                 top[i] = virt_to_mfn(p2m_mid_missing_mfn);
 128 }
 129
 130 static void p2m_top_mfn_p_init(unsigned long **top)
 131 {
 132         unsigned i;
 133
 134         for (i = 0; i < P2M_TOP_PER_PAGE; i++)
 135                 top[i] = p2m_mid_missing_mfn;
 136 }
 137
 138 static void p2m_mid_mfn_init(unsigned long *mid, unsigned long *leaf)
 139 {
 140         unsigned i;
 141
 142         for (i = 0; i < P2M_MID_PER_PAGE; i++)
 143                 mid[i] = virt_to_mfn(leaf);
 144 }
 145
 146 static void p2m_init(unsigned long *p2m)
 147 {
 148         unsigned i;
 149
 150         for (i = 0; i < P2M_PER_PAGE; i++)
 151                 p2m[i] = INVALID_P2M_ENTRY;
 152 }
 153
 154 static void p2m_init_identity(unsigned long *p2m, unsigned long pfn)
 155 {
 156         unsigned i;
 157
 158         for (i = 0; i < P2M_PER_PAGE; i++)
 159                 p2m[i] = IDENTITY_FRAME(pfn + i);
 160 }
 161
 162 static void * __ref alloc_p2m_page(void)
 163 {
 164         if (unlikely(!slab_is_available()))
 165                 return alloc_bootmem_align(PAGE_SIZE, PAGE_SIZE);
 166
 167         return (void *)__get_free_page(GFP_KERNEL | __GFP_REPEAT);
 168 }
 169
 170 /* Only to be called in case of a race for a page just allocated! */
 171 static void free_p2m_page(void *p)
 172 {
 173         BUG_ON(!slab_is_available());
 174         free_page((unsigned long)p);
 175 }
 176
 177 /*
 178  * Build the parallel p2m_top_mfn and p2m_mid_mfn structures
 179  *
 180  * This is called both at boot time, and after resuming from suspend:
 181  * - At boot time we're called rather early, and must use alloc_bootmem*()
 182  *   to allocate memory.
 183  *
 184  * - After resume we're called from within stop_machine, but the mfn
 185  *   tree should already be completely allocated.
 186  */
 187 void __ref xen_build_mfn_list_list(void)
 188 {
 189         unsigned long pfn, mfn;
 190         pte_t *ptep;
 191         unsigned int level, topidx, mididx;
 192         unsigned long *mid_mfn_p;
 193
 194         if (xen_feature(XENFEAT_auto_translated_physmap))
 195                 return;
 196
 197         /* Pre-initialize p2m_top_mfn to be completely missing */
 198         if (p2m_top_mfn == NULL) {
 199                 p2m_mid_missing_mfn = alloc_p2m_page();
 200                 p2m_mid_mfn_init(p2m_mid_missing_mfn, p2m_missing);
 201
 202                 p2m_top_mfn_p = alloc_p2m_page();
 203                 p2m_top_mfn_p_init(p2m_top_mfn_p);
 204
 205                 p2m_top_mfn = alloc_p2m_page();
 206                 p2m_top_mfn_init(p2m_top_mfn);
 207         } else {
 208                 /* Reinitialise, mfn's all change after migration */
 209                 p2m_mid_mfn_init(p2m_mid_missing_mfn, p2m_missing);
 210         }
 211
 212         for (pfn = 0; pfn < xen_max_p2m_pfn && pfn < MAX_P2M_PFN;
 213              pfn += P2M_PER_PAGE) {
 214                 topidx = p2m_top_index(pfn);
 215                 mididx = p2m_mid_index(pfn);
 216
 217                 mid_mfn_p = p2m_top_mfn_p[topidx];
 218                 ptep = lookup_address((unsigned long)(xen_p2m_addr + pfn),
 219                                       &level);
 220                 BUG_ON(!ptep || level != PG_LEVEL_4K);
 221                 mfn = pte_mfn(*ptep);
 222                 ptep = (pte_t *)((unsigned long)ptep & ~(PAGE_SIZE - 1));
 223
 224                 /* Don't bother allocating any mfn mid levels if
 225                  * they're just missing, just update the stored mfn,
 226                  * since all could have changed over a migrate.
 227                  */
 228                 if (ptep == p2m_missing_pte || ptep == p2m_identity_pte) {
 229                         BUG_ON(mididx);
 230                         BUG_ON(mid_mfn_p != p2m_mid_missing_mfn);
 231                         p2m_top_mfn[topidx] = virt_to_mfn(p2m_mid_missing_mfn);
 232                         pfn += (P2M_MID_PER_PAGE - 1) * P2M_PER_PAGE;
 233                         continue;
 234                 }
 235
 236                 if (mid_mfn_p == p2m_mid_missing_mfn) {
 237                         mid_mfn_p = alloc_p2m_page();
 238                         p2m_mid_mfn_init(mid_mfn_p, p2m_missing);
 239
 240                         p2m_top_mfn_p[topidx] = mid_mfn_p;
 241                 }
 242
 243                 p2m_top_mfn[topidx] = virt_to_mfn(mid_mfn_p);
 244                 mid_mfn_p[mididx] = mfn;
 245         }
 246 }
 247
 248 void xen_setup_mfn_list_list(void)
 249 {
 250         if (xen_feature(XENFEAT_auto_translated_physmap))
 251                 return;
 252
 253         BUG_ON(HYPERVISOR_shared_info == &xen_dummy_shared_info);
 254
 255         HYPERVISOR_shared_info->arch.pfn_to_mfn_frame_list_list =
 256                 virt_to_mfn(p2m_top_mfn);
 257         HYPERVISOR_shared_info->arch.max_pfn = xen_max_p2m_pfn;
 258 }
 259
 260 /* Set up p2m_top to point to the domain-builder provided p2m pages */
 261 void __init xen_build_dynamic_phys_to_machine(void)
 262 {
 263         unsigned long pfn;
 264
 265          if (xen_feature(XENFEAT_auto_translated_physmap))
 266                 return;
 267
 268         xen_p2m_addr = (unsigned long *)xen_start_info->mfn_list;
 269         xen_p2m_size = ALIGN(xen_start_info->nr_pages, P2M_PER_PAGE);
 270
 271         for (pfn = xen_start_info->nr_pages; pfn < xen_p2m_size; pfn++)
 272                 xen_p2m_addr[pfn] = INVALID_P2M_ENTRY;
 273
 274         xen_max_p2m_pfn = xen_p2m_size;
 275 }
 276
 277 #define P2M_TYPE_IDENTITY       0
 278 #define P2M_TYPE_MISSING        1
 279 #define P2M_TYPE_PFN            2
 280 #define P2M_TYPE_UNKNOWN        3
 281
 282 static int xen_p2m_elem_type(unsigned long pfn)
 283 {
 284         unsigned long mfn;
 285
 286         if (pfn >= xen_p2m_size)
 287                 return P2M_TYPE_IDENTITY;
 288
 289         mfn = xen_p2m_addr[pfn];
 290
 291         if (mfn == INVALID_P2M_ENTRY)
 292                 return P2M_TYPE_MISSING;
 293
 294         if (mfn & IDENTITY_FRAME_BIT)
 295                 return P2M_TYPE_IDENTITY;
 296
 297         return P2M_TYPE_PFN;
 298 }
 299
 300 static void __init xen_rebuild_p2m_list(unsigned long *p2m)
 301 {
 302         unsigned int i, chunk;
 303         unsigned long pfn;
 304         unsigned long *mfns;
 305         pte_t *ptep;
 306         pmd_t *pmdp;
 307         int type;
 308
 309         p2m_missing = alloc_p2m_page();
 310         p2m_init(p2m_missing);
 311         p2m_identity = alloc_p2m_page();
 312         p2m_init(p2m_identity);
 313
 314         p2m_missing_pte = alloc_p2m_page();
 315         paravirt_alloc_pte(&init_mm, __pa(p2m_missing_pte) >> PAGE_SHIFT);
 316         p2m_identity_pte = alloc_p2m_page();
 317         paravirt_alloc_pte(&init_mm, __pa(p2m_identity_pte) >> PAGE_SHIFT);
 318         for (i = 0; i < PTRS_PER_PTE; i++) {
 319                 set_pte(p2m_missing_pte + i,
 320                         pfn_pte(PFN_DOWN(__pa(p2m_missing)), PAGE_KERNEL_RO));
 321                 set_pte(p2m_identity_pte + i,
 322                         pfn_pte(PFN_DOWN(__pa(p2m_identity)), PAGE_KERNEL_RO));
 323         }
 324
 325         for (pfn = 0; pfn < xen_max_p2m_pfn; pfn += chunk) {
 326                 /*
 327                  * Try to map missing/identity PMDs or p2m-pages if possible.
 328                  * We have to respect the structure of the mfn_list_list
 329                  * which will be built just afterwards.
 330                  * Chunk size to test is one p2m page if we are in the middle
 331                  * of a mfn_list_list mid page and the complete mid page area
 332                  * if we are at index 0 of the mid page. Please note that a
 333                  * mid page might cover more than one PMD, e.g. on 32 bit PAE
 334                  * kernels.
 335                  */
 336                 chunk = (pfn & (P2M_PER_PAGE * P2M_MID_PER_PAGE - 1)) ?
 337                         P2M_PER_PAGE : P2M_PER_PAGE * P2M_MID_PER_PAGE;
 338
 339                 type = xen_p2m_elem_type(pfn);
 340                 i = 0;
 341                 if (type != P2M_TYPE_PFN)
 342                         for (i = 1; i < chunk; i++)
 343                                 if (xen_p2m_elem_type(pfn + i) != type)
 344                                         break;
 345                 if (i < chunk)
 346                         /* Reset to minimal chunk size. */
 347                         chunk = P2M_PER_PAGE;
 348
 349                 if (type == P2M_TYPE_PFN || i < chunk) {
 350                         /* Use initial p2m page contents. */
 351 #ifdef CONFIG_X86_64
 352                         mfns = alloc_p2m_page();
 353                         copy_page(mfns, xen_p2m_addr + pfn);
 354 #else
 355                         mfns = xen_p2m_addr + pfn;
 356 #endif
 357                         ptep = populate_extra_pte((unsigned long)(p2m + pfn));
 358                         set_pte(ptep,
 359                                 pfn_pte(PFN_DOWN(__pa(mfns)), PAGE_KERNEL));
 360                         continue;
 361                 }
 362
 363                 if (chunk == P2M_PER_PAGE) {
 364                         /* Map complete missing or identity p2m-page. */
 365                         mfns = (type == P2M_TYPE_MISSING) ?
 366                                 p2m_missing : p2m_identity;
 367                         ptep = populate_extra_pte((unsigned long)(p2m + pfn));
 368                         set_pte(ptep,
 369                                 pfn_pte(PFN_DOWN(__pa(mfns)), PAGE_KERNEL_RO));
 370                         continue;
 371                 }
 372
 373                 /* Complete missing or identity PMD(s) can be mapped. */
 374                 ptep = (type == P2M_TYPE_MISSING) ?
 375                         p2m_missing_pte : p2m_identity_pte;
 376                 for (i = 0; i < PMDS_PER_MID_PAGE; i++) {
 377                         pmdp = populate_extra_pmd(
 378                                 (unsigned long)(p2m + pfn + i * PTRS_PER_PTE));
 379                         set_pmd(pmdp, __pmd(__pa(ptep) | _KERNPG_TABLE));
 380                 }
 381         }
 382 }
 383
 384 void __init xen_vmalloc_p2m_tree(void)
 385 {
 386         static struct vm_struct vm;
 387
 388         vm.flags = VM_ALLOC;
 389         vm.size = ALIGN(sizeof(unsigned long) * xen_max_p2m_pfn,
 390                         PMD_SIZE * PMDS_PER_MID_PAGE);
 391         vm_area_register_early(&vm, PMD_SIZE * PMDS_PER_MID_PAGE);
 392         pr_notice("p2m virtual area at %p, size is %lx\n", vm.addr, vm.size);
 393
 394         xen_max_p2m_pfn = vm.size / sizeof(unsigned long);
 395
 396         xen_rebuild_p2m_list(vm.addr);
 397
 398         xen_p2m_addr = vm.addr;
 399         xen_p2m_size = xen_max_p2m_pfn;
 400
 401         xen_inv_extra_mem();
 402
 403         m2p_override_init();
 404 }
 405
 406 unsigned long get_phys_to_machine(unsigned long pfn)
 407 {
 408         pte_t *ptep;
 409         unsigned int level;
 410
 411         if (unlikely(pfn >= xen_p2m_size)) {
 412                 if (pfn < xen_max_p2m_pfn)
 413                         return xen_chk_extra_mem(pfn);
 414
 415                 return IDENTITY_FRAME(pfn);
 416         }
 417
 418         ptep = lookup_address((unsigned long)(xen_p2m_addr + pfn), &level);
 419         BUG_ON(!ptep || level != PG_LEVEL_4K);
 420
 421         /*
 422          * The INVALID_P2M_ENTRY is filled in both p2m_*identity
 423          * and in p2m_*missing, so returning the INVALID_P2M_ENTRY
 424          * would be wrong.
 425          */
 426         if (pte_pfn(*ptep) == PFN_DOWN(__pa(p2m_identity)))
 427                 return IDENTITY_FRAME(pfn);
 428
 429         return xen_p2m_addr[pfn];
 430 }
 431 EXPORT_SYMBOL_GPL(get_phys_to_machine);
 432
 433 /*
 434  * Allocate new pmd(s). It is checked whether the old pmd is still in place.
 435  * If not, nothing is changed. This is okay as the only reason for allocating
 436  * a new pmd is to replace p2m_missing_pte or p2m_identity_pte by a individual
 437  * pmd. In case of PAE/x86-32 there are multiple pmds to allocate!
 438  */
 439 static pte_t *alloc_p2m_pmd(unsigned long addr, pte_t *ptep, pte_t *pte_pg)
 440 {
 441         pte_t *ptechk;
 442         pte_t *pteret = ptep;
 443         pte_t *pte_newpg[PMDS_PER_MID_PAGE];
 444         pmd_t *pmdp;
 445         unsigned int level;
 446         unsigned long flags;
 447         unsigned long vaddr;
 448         int i;
 449
 450         /* Do all allocations first to bail out in error case. */
 451         for (i = 0; i < PMDS_PER_MID_PAGE; i++) {
 452                 pte_newpg[i] = alloc_p2m_page();
 453                 if (!pte_newpg[i]) {
 454                         for (i--; i >= 0; i--)
 455                                 free_p2m_page(pte_newpg[i]);
 456
 457                         return NULL;
 458                 }
 459         }
 460
 461         vaddr = addr & ~(PMD_SIZE * PMDS_PER_MID_PAGE - 1);
 462
 463         for (i = 0; i < PMDS_PER_MID_PAGE; i++) {
 464                 copy_page(pte_newpg[i], pte_pg);
 465                 paravirt_alloc_pte(&init_mm, __pa(pte_newpg[i]) >> PAGE_SHIFT);
 466
 467                 pmdp = lookup_pmd_address(vaddr);
 468                 BUG_ON(!pmdp);
 469
 470                 spin_lock_irqsave(&p2m_update_lock, flags);
 471
 472                 ptechk = lookup_address(vaddr, &level);
 473                 if (ptechk == pte_pg) {
 474                         set_pmd(pmdp,
 475                                 __pmd(__pa(pte_newpg[i]) | _KERNPG_TABLE));
 476                         if (vaddr == (addr & ~(PMD_SIZE - 1)))
 477                                 pteret = pte_offset_kernel(pmdp, addr);
 478                         pte_newpg[i] = NULL;
 479                 }
 480
 481                 spin_unlock_irqrestore(&p2m_update_lock, flags);
 482
 483                 if (pte_newpg[i]) {
 484                         paravirt_release_pte(__pa(pte_newpg[i]) >> PAGE_SHIFT);
 485                         free_p2m_page(pte_newpg[i]);
 486                 }
 487
 488                 vaddr += PMD_SIZE;
 489         }
 490
 491         return pteret;
 492 }
 493
 494 /*
 495  * Fully allocate the p2m structure for a given pfn.  We need to check
 496  * that both the top and mid levels are allocated, and make sure the
 497  * parallel mfn tree is kept in sync.  We may race with other cpus, so
 498  * the new pages are installed with cmpxchg; if we lose the race then
 499  * simply free the page we allocated and use the one that's there.
 500  */
 501 static bool alloc_p2m(unsigned long pfn)
 502 {
 503         unsigned topidx, mididx;
 504         unsigned long *top_mfn_p, *mid_mfn;
 505         pte_t *ptep, *pte_pg;
 506         unsigned int level;
 507         unsigned long flags;
 508         unsigned long addr = (unsigned long)(xen_p2m_addr + pfn);
 509         unsigned long p2m_pfn;
 510
 511         topidx = p2m_top_index(pfn);
 512         mididx = p2m_mid_index(pfn);
 513
 514         ptep = lookup_address(addr, &level);
 515         BUG_ON(!ptep || level != PG_LEVEL_4K);
 516         pte_pg = (pte_t *)((unsigned long)ptep & ~(PAGE_SIZE - 1));
 517
 518         if (pte_pg == p2m_missing_pte || pte_pg == p2m_identity_pte) {
 519                 /* PMD level is missing, allocate a new one */
 520                 ptep = alloc_p2m_pmd(addr, ptep, pte_pg);
 521                 if (!ptep)
 522                         return false;
 523         }
 524
 525         if (p2m_top_mfn) {
 526                 top_mfn_p = &p2m_top_mfn[topidx];
 527                 mid_mfn = ACCESS_ONCE(p2m_top_mfn_p[topidx]);
 528
 529                 BUG_ON(virt_to_mfn(mid_mfn) != *top_mfn_p);
 530
 531                 if (mid_mfn == p2m_mid_missing_mfn) {
 532                         /* Separately check the mid mfn level */
 533                         unsigned long missing_mfn;
 534                         unsigned long mid_mfn_mfn;
 535                         unsigned long old_mfn;
 536
 537                         mid_mfn = alloc_p2m_page();
 538                         if (!mid_mfn)
 539                                 return false;
 540
 541                         p2m_mid_mfn_init(mid_mfn, p2m_missing);
 542
 543                         missing_mfn = virt_to_mfn(p2m_mid_missing_mfn);
 544                         mid_mfn_mfn = virt_to_mfn(mid_mfn);
 545                         old_mfn = cmpxchg(top_mfn_p, missing_mfn, mid_mfn_mfn);
 546                         if (old_mfn != missing_mfn) {
 547                                 free_p2m_page(mid_mfn);
 548                                 mid_mfn = mfn_to_virt(old_mfn);
 549                         } else {
 550                                 p2m_top_mfn_p[topidx] = mid_mfn;
 551                         }
 552                 }
 553         } else {
 554                 mid_mfn = NULL;
 555         }
 556
 557         p2m_pfn = pte_pfn(ACCESS_ONCE(*ptep));
 558         if (p2m_pfn == PFN_DOWN(__pa(p2m_identity)) ||
 559             p2m_pfn == PFN_DOWN(__pa(p2m_missing))) {
 560                 /* p2m leaf page is missing */
 561                 unsigned long *p2m;
 562
 563                 p2m = alloc_p2m_page();
 564                 if (!p2m)
 565                         return false;
 566
 567                 if (p2m_pfn == PFN_DOWN(__pa(p2m_missing)))
 568                         p2m_init(p2m);
 569                 else
 570                         p2m_init_identity(p2m, pfn);
 571
 572                 spin_lock_irqsave(&p2m_update_lock, flags);
 573
 574                 if (pte_pfn(*ptep) == p2m_pfn) {
 575                         set_pte(ptep,
 576                                 pfn_pte(PFN_DOWN(__pa(p2m)), PAGE_KERNEL));
 577                         if (mid_mfn)
 578                                 mid_mfn[mididx] = virt_to_mfn(p2m);
 579                         p2m = NULL;
 580                 }
 581
 582                 spin_unlock_irqrestore(&p2m_update_lock, flags);
 583
 584                 if (p2m)
 585                         free_p2m_page(p2m);
 586         }
 587
 588         return true;
 589 }
 590
 591 unsigned long __init set_phys_range_identity(unsigned long pfn_s,
 592                                       unsigned long pfn_e)
 593 {
 594         unsigned long pfn;
 595
 596         if (unlikely(pfn_s >= xen_p2m_size))
 597                 return 0;
 598
 599         if (unlikely(xen_feature(XENFEAT_auto_translated_physmap)))
 600                 return pfn_e - pfn_s;
 601
 602         if (pfn_s > pfn_e)
 603                 return 0;
 604
 605         if (pfn_e > xen_p2m_size)
 606                 pfn_e = xen_p2m_size;
 607
 608         for (pfn = pfn_s; pfn < pfn_e; pfn++)
 609                 xen_p2m_addr[pfn] = IDENTITY_FRAME(pfn);
 610
 611         return pfn - pfn_s;
 612 }
 613
 614 bool __set_phys_to_machine(unsigned long pfn, unsigned long mfn)
 615 {
 616         pte_t *ptep;
 617         unsigned int level;
 618
 619         /* don't track P2M changes in autotranslate guests */
 620         if (unlikely(xen_feature(XENFEAT_auto_translated_physmap)))
 621                 return true;
 622
 623         if (unlikely(pfn >= xen_p2m_size)) {
 624                 BUG_ON(mfn != INVALID_P2M_ENTRY);
 625                 return true;
 626         }
 627
 628         if (likely(!xen_safe_write_ulong(xen_p2m_addr + pfn, mfn)))
 629                 return true;
 630
 631         ptep = lookup_address((unsigned long)(xen_p2m_addr + pfn), &level);
 632         BUG_ON(!ptep || level != PG_LEVEL_4K);
 633
 634         if (pte_pfn(*ptep) == PFN_DOWN(__pa(p2m_missing)))
 635                 return mfn == INVALID_P2M_ENTRY;
 636
 637         if (pte_pfn(*ptep) == PFN_DOWN(__pa(p2m_identity)))
 638                 return mfn == IDENTITY_FRAME(pfn);
 639
 640         return false;
 641 }
 642
 643 bool set_phys_to_machine(unsigned long pfn, unsigned long mfn)
 644 {
 645         if (unlikely(!__set_phys_to_machine(pfn, mfn))) {
 646                 if (!alloc_p2m(pfn))
 647                         return false;
 648
 649                 return __set_phys_to_machine(pfn, mfn);
 650         }
 651
 652         return true;
 653 }
 654
 655 #define M2P_OVERRIDE_HASH_SHIFT 10
 656 #define M2P_OVERRIDE_HASH       (1 << M2P_OVERRIDE_HASH_SHIFT)
 657
 658 static struct list_head *m2p_overrides;
 659 static DEFINE_SPINLOCK(m2p_override_lock);
 660
 661 static void __init m2p_override_init(void)
 662 {
 663         unsigned i;
 664
 665         m2p_overrides = alloc_bootmem_align(
 666                                 sizeof(*m2p_overrides) * M2P_OVERRIDE_HASH,
 667                                 sizeof(unsigned long));
 668
 669         for (i = 0; i < M2P_OVERRIDE_HASH; i++)
 670                 INIT_LIST_HEAD(&m2p_overrides[i]);
 671 }
 672
 673 static unsigned long mfn_hash(unsigned long mfn)
 674 {
 675         return hash_long(mfn, M2P_OVERRIDE_HASH_SHIFT);
 676 }
 677
 678 /* Add an MFN override for a particular page */
 679 static int m2p_add_override(unsigned long mfn, struct page *page,
 680                             struct gnttab_map_grant_ref *kmap_op)
 681 {
 682         unsigned long flags;
 683         unsigned long pfn;
 684         unsigned long uninitialized_var(address);
 685         unsigned level;
 686         pte_t *ptep = NULL;
 687
 688         pfn = page_to_pfn(page);
 689         if (!PageHighMem(page)) {
 690                 address = (unsigned long)__va(pfn << PAGE_SHIFT);
 691                 ptep = lookup_address(address, &level);
 692                 if (WARN(ptep == NULL || level != PG_LEVEL_4K,
 693                          "m2p_add_override: pfn %lx not mapped", pfn))
 694                         return -EINVAL;
 695         }
 696
 697         if (kmap_op != NULL) {
 698                 if (!PageHighMem(page)) {
 699                         struct multicall_space mcs =
 700                                 xen_mc_entry(sizeof(*kmap_op));
 701
 702                         MULTI_grant_table_op(mcs.mc,
 703                                         GNTTABOP_map_grant_ref, kmap_op, 1);
 704
 705                         xen_mc_issue(PARAVIRT_LAZY_MMU);
 706                 }
 707         }
 708         spin_lock_irqsave(&m2p_override_lock, flags);
 709         list_add(&page->lru,  &m2p_overrides[mfn_hash(mfn)]);
 710         spin_unlock_irqrestore(&m2p_override_lock, flags);
 711
 712         /* p2m(m2p(mfn)) == mfn: the mfn is already present somewhere in
 713          * this domain. Set the FOREIGN_FRAME_BIT in the p2m for the other
 714          * pfn so that the following mfn_to_pfn(mfn) calls will return the
 715          * pfn from the m2p_override (the backend pfn) instead.
 716          * We need to do this because the pages shared by the frontend
 717          * (xen-blkfront) can be already locked (lock_page, called by
 718          * do_read_cache_page); when the userspace backend tries to use them
 719          * with direct_IO, mfn_to_pfn returns the pfn of the frontend, so
 720          * do_blockdev_direct_IO is going to try to lock the same pages
 721          * again resulting in a deadlock.
 722          * As a side effect get_user_pages_fast might not be safe on the
 723          * frontend pages while they are being shared with the backend,
 724          * because mfn_to_pfn (that ends up being called by GUPF) will
 725          * return the backend pfn rather than the frontend pfn. */
 726         pfn = mfn_to_pfn_no_overrides(mfn);
 727         if (__pfn_to_mfn(pfn) == mfn)
 728                 set_phys_to_machine(pfn, FOREIGN_FRAME(mfn));
 729
 730         return 0;
 731 }
 732
 733 int set_foreign_p2m_mapping(struct gnttab_map_grant_ref *map_ops,
 734                             struct gnttab_map_grant_ref *kmap_ops,
 735                             struct page **pages, unsigned int count)
 736 {
 737         int i, ret = 0;
 738         bool lazy = false;
 739         pte_t *pte;
 740
 741         if (xen_feature(XENFEAT_auto_translated_physmap))
 742                 return 0;
 743
 744         if (kmap_ops &&
 745             !in_interrupt() &&
 746             paravirt_get_lazy_mode() == PARAVIRT_LAZY_NONE) {
 747                 arch_enter_lazy_mmu_mode();
 748                 lazy = true;
 749         }
 750
 751         for (i = 0; i < count; i++) {
 752                 unsigned long mfn, pfn;
 753
 754                 /* Do not add to override if the map failed. */
 755                 if (map_ops[i].status)
 756                         continue;
 757
 758                 if (map_ops[i].flags & GNTMAP_contains_pte) {
 759                         pte = (pte_t *)(mfn_to_virt(PFN_DOWN(map_ops[i].host_addr)) +
 760                                 (map_ops[i].host_addr & ~PAGE_MASK));
 761                         mfn = pte_mfn(*pte);
 762                 } else {
 763                         mfn = PFN_DOWN(map_ops[i].dev_bus_addr);
 764                 }
 765                 pfn = page_to_pfn(pages[i]);
 766
 767                 WARN_ON(PagePrivate(pages[i]));
 768                 SetPagePrivate(pages[i]);
 769                 set_page_private(pages[i], mfn);
 770                 pages[i]->index = pfn_to_mfn(pfn);
 771
 772                 if (unlikely(!set_phys_to_machine(pfn, FOREIGN_FRAME(mfn)))) {
 773                         ret = -ENOMEM;
 774                         goto out;
 775                 }
 776
 777                 if (kmap_ops) {
 778                         ret = m2p_add_override(mfn, pages[i], &kmap_ops[i]);
 779                         if (ret)
 780                                 goto out;
 781                 }
 782         }
 783
 784 out:
 785         if (lazy)
 786                 arch_leave_lazy_mmu_mode();
 787
 788         return ret;
 789 }
 790 EXPORT_SYMBOL_GPL(set_foreign_p2m_mapping);
 791
 792 static struct page *m2p_find_override(unsigned long mfn)
 793 {
 794         unsigned long flags;
 795         struct list_head *bucket;
 796         struct page *p, *ret;
 797
 798         if (unlikely(!m2p_overrides))
 799                 return NULL;
 800
 801         ret = NULL;
 802         bucket = &m2p_overrides[mfn_hash(mfn)];
 803
 804         spin_lock_irqsave(&m2p_override_lock, flags);
 805
 806         list_for_each_entry(p, bucket, lru) {
 807                 if (page_private(p) == mfn) {
 808                         ret = p;
 809                         break;
 810                 }
 811         }
 812
 813         spin_unlock_irqrestore(&m2p_override_lock, flags);
 814
 815         return ret;
 816 }
 817
 818 static int m2p_remove_override(struct page *page,
 819                                struct gnttab_map_grant_ref *kmap_op,
 820                                unsigned long mfn)
 821 {
 822         unsigned long flags;
 823         unsigned long pfn;
 824         unsigned long uninitialized_var(address);
 825         unsigned level;
 826         pte_t *ptep = NULL;
 827
 828         pfn = page_to_pfn(page);
 829
 830         if (!PageHighMem(page)) {
 831                 address = (unsigned long)__va(pfn << PAGE_SHIFT);
 832                 ptep = lookup_address(address, &level);
 833
 834                 if (WARN(ptep == NULL || level != PG_LEVEL_4K,
 835                          "m2p_remove_override: pfn %lx not mapped", pfn))
 836                         return -EINVAL;
 837         }
 838
 839         spin_lock_irqsave(&m2p_override_lock, flags);
 840         list_del(&page->lru);
 841         spin_unlock_irqrestore(&m2p_override_lock, flags);
 842
 843         if (kmap_op != NULL) {
 844                 if (!PageHighMem(page)) {
 845                         struct multicall_space mcs;
 846                         struct gnttab_unmap_and_replace *unmap_op;
 847                         struct page *scratch_page = get_balloon_scratch_page();
 848                         unsigned long scratch_page_address = (unsigned long)
 849                                 __va(page_to_pfn(scratch_page) << PAGE_SHIFT);
 850
 851                         /*
 852                          * It might be that we queued all the m2p grant table
 853                          * hypercalls in a multicall, then m2p_remove_override
 854                          * get called before the multicall has actually been
 855                          * issued. In this case handle is going to -1 because
 856                          * it hasn't been modified yet.
 857                          */
 858                         if (kmap_op->handle == -1)
 859                                 xen_mc_flush();
 860                         /*
 861                          * Now if kmap_op->handle is negative it means that the
 862                          * hypercall actually returned an error.
 863                          */
 864                         if (kmap_op->handle == GNTST_general_error) {
 865                                 pr_warn("m2p_remove_override: pfn %lx mfn %lx, failed to modify kernel mappings",
 866                                         pfn, mfn);
 867                                 put_balloon_scratch_page();
 868                                 return -1;
 869                         }
 870
 871                         xen_mc_batch();
 872
 873                         mcs = __xen_mc_entry(
 874                                 sizeof(struct gnttab_unmap_and_replace));
 875                         unmap_op = mcs.args;
 876                         unmap_op->host_addr = kmap_op->host_addr;
 877                         unmap_op->new_addr = scratch_page_address;
 878                         unmap_op->handle = kmap_op->handle;
 879
 880                         MULTI_grant_table_op(mcs.mc,
 881                                 GNTTABOP_unmap_and_replace, unmap_op, 1);
 882
 883                         mcs = __xen_mc_entry(0);
 884                         MULTI_update_va_mapping(mcs.mc, scratch_page_address,
 885                                         pfn_pte(page_to_pfn(scratch_page),
 886                                         PAGE_KERNEL_RO), 0);
 887
 888                         xen_mc_issue(PARAVIRT_LAZY_MMU);
 889
 890                         kmap_op->host_addr = 0;
 891                         put_balloon_scratch_page();
 892                 }
 893         }
 894
 895         /* p2m(m2p(mfn)) == FOREIGN_FRAME(mfn): the mfn is already present
 896          * somewhere in this domain, even before being added to the
 897          * m2p_override (see comment above in m2p_add_override).
 898          * If there are no other entries in the m2p_override corresponding
 899          * to this mfn, then remove the FOREIGN_FRAME_BIT from the p2m for
 900          * the original pfn (the one shared by the frontend): the backend
 901          * cannot do any IO on this page anymore because it has been
 902          * unshared. Removing the FOREIGN_FRAME_BIT from the p2m entry of
 903          * the original pfn causes mfn_to_pfn(mfn) to return the frontend
 904          * pfn again. */
 905         mfn &= ~FOREIGN_FRAME_BIT;
 906         pfn = mfn_to_pfn_no_overrides(mfn);
 907         if (__pfn_to_mfn(pfn) == FOREIGN_FRAME(mfn) &&
 908                         m2p_find_override(mfn) == NULL)
 909                 set_phys_to_machine(pfn, mfn);
 910
 911         return 0;
 912 }
 913
 914 int clear_foreign_p2m_mapping(struct gnttab_unmap_grant_ref *unmap_ops,
 915                               struct gnttab_map_grant_ref *kmap_ops,
 916                               struct page **pages, unsigned int count)
 917 {
 918         int i, ret = 0;
 919         bool lazy = false;
 920
 921         if (xen_feature(XENFEAT_auto_translated_physmap))
 922                 return 0;
 923
 924         if (kmap_ops &&
 925             !in_interrupt() &&
 926             paravirt_get_lazy_mode() == PARAVIRT_LAZY_NONE) {
 927                 arch_enter_lazy_mmu_mode();
 928                 lazy = true;
 929         }
 930
 931         for (i = 0; i < count; i++) {
 932                 unsigned long mfn = __pfn_to_mfn(page_to_pfn(pages[i]));
 933                 unsigned long pfn = page_to_pfn(pages[i]);
 934
 935                 if (mfn == INVALID_P2M_ENTRY || !(mfn & FOREIGN_FRAME_BIT)) {
 936                         ret = -EINVAL;
 937                         goto out;
 938                 }
 939
 940                 set_page_private(pages[i], INVALID_P2M_ENTRY);
 941                 WARN_ON(!PagePrivate(pages[i]));
 942                 ClearPagePrivate(pages[i]);
 943                 set_phys_to_machine(pfn, pages[i]->index);
 944
 945                 if (kmap_ops)
 946                         ret = m2p_remove_override(pages[i], &kmap_ops[i], mfn);
 947                 if (ret)
 948                         goto out;
 949         }
 950
 951 out:
 952         if (lazy)
 953                 arch_leave_lazy_mmu_mode();
 954         return ret;
 955 }
 956 EXPORT_SYMBOL_GPL(clear_foreign_p2m_mapping);
 957
 958 unsigned long m2p_find_override_pfn(unsigned long mfn, unsigned long pfn)
 959 {
 960         struct page *p = m2p_find_override(mfn);
 961         unsigned long ret = pfn;
 962
 963         if (p)
 964                 ret = page_to_pfn(p);
 965
 966         return ret;
 967 }
 968 EXPORT_SYMBOL_GPL(m2p_find_override_pfn);
 969
 970 #ifdef CONFIG_XEN_DEBUG_FS
 971 #include <linux/debugfs.h>
 972 #include "debugfs.h"
 973 static int p2m_dump_show(struct seq_file *m, void *v)
 974 {
 975         static const char * const type_name[] = {
 976                                 [P2M_TYPE_IDENTITY] = "identity",
 977                                 [P2M_TYPE_MISSING] = "missing",
 978                                 [P2M_TYPE_PFN] = "pfn",
 979                                 [P2M_TYPE_UNKNOWN] = "abnormal"};
 980         unsigned long pfn, first_pfn;
 981         int type, prev_type;
 982
 983         prev_type = xen_p2m_elem_type(0);
 984         first_pfn = 0;
 985
 986         for (pfn = 0; pfn < xen_p2m_size; pfn++) {
 987                 type = xen_p2m_elem_type(pfn);
 988                 if (type != prev_type) {
 989                         seq_printf(m, " [0x%lx->0x%lx] %s\n", first_pfn, pfn,
 990                                    type_name[prev_type]);
 991                         prev_type = type;
 992                         first_pfn = pfn;
 993                 }
 994         }
 995         seq_printf(m, " [0x%lx->0x%lx] %s\n", first_pfn, pfn,
 996                    type_name[prev_type]);
 997         return 0;
 998 }
 999
1000 static int p2m_dump_open(struct inode *inode, struct file *filp)
1001 {
1002         return single_open(filp, p2m_dump_show, NULL);
1003 }
1004
1005 static const struct file_operations p2m_dump_fops = {
1006         .open           = p2m_dump_open,
1007         .read           = seq_read,
1008         .llseek         = seq_lseek,
1009         .release        = single_release,
1010 };
1011
1012 static struct dentry *d_mmu_debug;
1013
1014 static int __init xen_p2m_debugfs(void)
1015 {
1016         struct dentry *d_xen = xen_init_debugfs();
1017
1018         if (d_xen == NULL)
1019                 return -ENOMEM;
1020
1021         d_mmu_debug = debugfs_create_dir("mmu", d_xen);
1022
1023         debugfs_create_file("p2m", 0600, d_mmu_debug, NULL, &p2m_dump_fops);
1024         return 0;
1025 }
1026 fs_initcall(xen_p2m_debugfs);
1027 #endif /* CONFIG_XEN_DEBUG_FS */