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