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/export.h>
  64 #include <linux/list.h>
  65 #include <linux/hash.h>
  66 #include <linux/sched.h>
  67 #include <linux/seq_file.h>
  68 #include <linux/bootmem.h>
  69 #include <linux/slab.h>
  70 #include <linux/vmalloc.h>
  71
  72 #include <asm/cache.h>
  73 #include <asm/setup.h>
  74 #include <linux/uaccess.h>
  75
  76 #include <asm/xen/page.h>
  77 #include <asm/xen/hypercall.h>
  78 #include <asm/xen/hypervisor.h>
  79 #include <xen/balloon.h>
  80 #include <xen/grant_table.h>
  81
  82 #include "multicalls.h"
  83 #include "xen-ops.h"
  84
  85 #define P2M_MID_PER_PAGE        (PAGE_SIZE / sizeof(unsigned long *))
  86 #define P2M_TOP_PER_PAGE        (PAGE_SIZE / sizeof(unsigned long **))
  87
  88 #define MAX_P2M_PFN     (P2M_TOP_PER_PAGE * P2M_MID_PER_PAGE * P2M_PER_PAGE)
  89
  90 #define PMDS_PER_MID_PAGE       (P2M_MID_PER_PAGE / PTRS_PER_PTE)
  91
  92 unsigned long *xen_p2m_addr __read_mostly;
  93 EXPORT_SYMBOL_GPL(xen_p2m_addr);
  94 unsigned long xen_p2m_size __read_mostly;
  95 EXPORT_SYMBOL_GPL(xen_p2m_size);
  96 unsigned long xen_max_p2m_pfn __read_mostly;
  97 EXPORT_SYMBOL_GPL(xen_max_p2m_pfn);
  98
  99 #ifdef CONFIG_XEN_BALLOON_MEMORY_HOTPLUG_LIMIT
 100 #define P2M_LIMIT CONFIG_XEN_BALLOON_MEMORY_HOTPLUG_LIMIT
 101 #else
 102 #define P2M_LIMIT 0
 103 #endif
 104
 105 static DEFINE_SPINLOCK(p2m_update_lock);
 106
 107 static unsigned long *p2m_mid_missing_mfn;
 108 static unsigned long *p2m_top_mfn;
 109 static unsigned long **p2m_top_mfn_p;
 110 static unsigned long *p2m_missing;
 111 static unsigned long *p2m_identity;
 112 static pte_t *p2m_missing_pte;
 113 static pte_t *p2m_identity_pte;
 114
 115 /*
 116  * Hint at last populated PFN.
 117  *
 118  * Used to set HYPERVISOR_shared_info->arch.max_pfn so the toolstack
 119  * can avoid scanning the whole P2M (which may be sized to account for
 120  * hotplugged memory).
 121  */
 122 static unsigned long xen_p2m_last_pfn;
 123
 124 static inline unsigned p2m_top_index(unsigned long pfn)
 125 {
 126         BUG_ON(pfn >= MAX_P2M_PFN);
 127         return pfn / (P2M_MID_PER_PAGE * P2M_PER_PAGE);
 128 }
 129
 130 static inline unsigned p2m_mid_index(unsigned long pfn)
 131 {
 132         return (pfn / P2M_PER_PAGE) % P2M_MID_PER_PAGE;
 133 }
 134
 135 static inline unsigned p2m_index(unsigned long pfn)
 136 {
 137         return pfn % P2M_PER_PAGE;
 138 }
 139
 140 static void p2m_top_mfn_init(unsigned long *top)
 141 {
 142         unsigned i;
 143
 144         for (i = 0; i < P2M_TOP_PER_PAGE; i++)
 145                 top[i] = virt_to_mfn(p2m_mid_missing_mfn);
 146 }
 147
 148 static void p2m_top_mfn_p_init(unsigned long **top)
 149 {
 150         unsigned i;
 151
 152         for (i = 0; i < P2M_TOP_PER_PAGE; i++)
 153                 top[i] = p2m_mid_missing_mfn;
 154 }
 155
 156 static void p2m_mid_mfn_init(unsigned long *mid, unsigned long *leaf)
 157 {
 158         unsigned i;
 159
 160         for (i = 0; i < P2M_MID_PER_PAGE; i++)
 161                 mid[i] = virt_to_mfn(leaf);
 162 }
 163
 164 static void p2m_init(unsigned long *p2m)
 165 {
 166         unsigned i;
 167
 168         for (i = 0; i < P2M_PER_PAGE; i++)
 169                 p2m[i] = INVALID_P2M_ENTRY;
 170 }
 171
 172 static void p2m_init_identity(unsigned long *p2m, unsigned long pfn)
 173 {
 174         unsigned i;
 175
 176         for (i = 0; i < P2M_PER_PAGE; i++)
 177                 p2m[i] = IDENTITY_FRAME(pfn + i);
 178 }
 179
 180 static void * __ref alloc_p2m_page(void)
 181 {
 182         if (unlikely(!slab_is_available()))
 183                 return alloc_bootmem_align(PAGE_SIZE, PAGE_SIZE);
 184
 185         return (void *)__get_free_page(GFP_KERNEL);
 186 }
 187
 188 static void __ref free_p2m_page(void *p)
 189 {
 190         if (unlikely(!slab_is_available())) {
 191                 free_bootmem((unsigned long)p, PAGE_SIZE);
 192                 return;
 193         }
 194
 195         free_page((unsigned long)p);
 196 }
 197
 198 /*
 199  * Build the parallel p2m_top_mfn and p2m_mid_mfn structures
 200  *
 201  * This is called both at boot time, and after resuming from suspend:
 202  * - At boot time we're called rather early, and must use alloc_bootmem*()
 203  *   to allocate memory.
 204  *
 205  * - After resume we're called from within stop_machine, but the mfn
 206  *   tree should already be completely allocated.
 207  */
 208 void __ref xen_build_mfn_list_list(void)
 209 {
 210         unsigned long pfn, mfn;
 211         pte_t *ptep;
 212         unsigned int level, topidx, mididx;
 213         unsigned long *mid_mfn_p;
 214
 215         if (xen_start_info->flags & SIF_VIRT_P2M_4TOOLS)
 216                 return;
 217
 218         /* Pre-initialize p2m_top_mfn to be completely missing */
 219         if (p2m_top_mfn == NULL) {
 220                 p2m_mid_missing_mfn = alloc_p2m_page();
 221                 p2m_mid_mfn_init(p2m_mid_missing_mfn, p2m_missing);
 222
 223                 p2m_top_mfn_p = alloc_p2m_page();
 224                 p2m_top_mfn_p_init(p2m_top_mfn_p);
 225
 226                 p2m_top_mfn = alloc_p2m_page();
 227                 p2m_top_mfn_init(p2m_top_mfn);
 228         } else {
 229                 /* Reinitialise, mfn's all change after migration */
 230                 p2m_mid_mfn_init(p2m_mid_missing_mfn, p2m_missing);
 231         }
 232
 233         for (pfn = 0; pfn < xen_max_p2m_pfn && pfn < MAX_P2M_PFN;
 234              pfn += P2M_PER_PAGE) {
 235                 topidx = p2m_top_index(pfn);
 236                 mididx = p2m_mid_index(pfn);
 237
 238                 mid_mfn_p = p2m_top_mfn_p[topidx];
 239                 ptep = lookup_address((unsigned long)(xen_p2m_addr + pfn),
 240                                       &level);
 241                 BUG_ON(!ptep || level != PG_LEVEL_4K);
 242                 mfn = pte_mfn(*ptep);
 243                 ptep = (pte_t *)((unsigned long)ptep & ~(PAGE_SIZE - 1));
 244
 245                 /* Don't bother allocating any mfn mid levels if
 246                  * they're just missing, just update the stored mfn,
 247                  * since all could have changed over a migrate.
 248                  */
 249                 if (ptep == p2m_missing_pte || ptep == p2m_identity_pte) {
 250                         BUG_ON(mididx);
 251                         BUG_ON(mid_mfn_p != p2m_mid_missing_mfn);
 252                         p2m_top_mfn[topidx] = virt_to_mfn(p2m_mid_missing_mfn);
 253                         pfn += (P2M_MID_PER_PAGE - 1) * P2M_PER_PAGE;
 254                         continue;
 255                 }
 256
 257                 if (mid_mfn_p == p2m_mid_missing_mfn) {
 258                         mid_mfn_p = alloc_p2m_page();
 259                         p2m_mid_mfn_init(mid_mfn_p, p2m_missing);
 260
 261                         p2m_top_mfn_p[topidx] = mid_mfn_p;
 262                 }
 263
 264                 p2m_top_mfn[topidx] = virt_to_mfn(mid_mfn_p);
 265                 mid_mfn_p[mididx] = mfn;
 266         }
 267 }
 268
 269 void xen_setup_mfn_list_list(void)
 270 {
 271         BUG_ON(HYPERVISOR_shared_info == &xen_dummy_shared_info);
 272
 273         if (xen_start_info->flags & SIF_VIRT_P2M_4TOOLS)
 274                 HYPERVISOR_shared_info->arch.pfn_to_mfn_frame_list_list = ~0UL;
 275         else
 276                 HYPERVISOR_shared_info->arch.pfn_to_mfn_frame_list_list =
 277                         virt_to_mfn(p2m_top_mfn);
 278         HYPERVISOR_shared_info->arch.max_pfn = xen_p2m_last_pfn;
 279         HYPERVISOR_shared_info->arch.p2m_generation = 0;
 280         HYPERVISOR_shared_info->arch.p2m_vaddr = (unsigned long)xen_p2m_addr;
 281         HYPERVISOR_shared_info->arch.p2m_cr3 =
 282                 xen_pfn_to_cr3(virt_to_mfn(swapper_pg_dir));
 283 }
 284
 285 /* Set up p2m_top to point to the domain-builder provided p2m pages */
 286 void __init xen_build_dynamic_phys_to_machine(void)
 287 {
 288         unsigned long pfn;
 289
 290         xen_p2m_addr = (unsigned long *)xen_start_info->mfn_list;
 291         xen_p2m_size = ALIGN(xen_start_info->nr_pages, P2M_PER_PAGE);
 292
 293         for (pfn = xen_start_info->nr_pages; pfn < xen_p2m_size; pfn++)
 294                 xen_p2m_addr[pfn] = INVALID_P2M_ENTRY;
 295
 296         xen_max_p2m_pfn = xen_p2m_size;
 297 }
 298
 299 #define P2M_TYPE_IDENTITY       0
 300 #define P2M_TYPE_MISSING        1
 301 #define P2M_TYPE_PFN            2
 302 #define P2M_TYPE_UNKNOWN        3
 303
 304 static int xen_p2m_elem_type(unsigned long pfn)
 305 {
 306         unsigned long mfn;
 307
 308         if (pfn >= xen_p2m_size)
 309                 return P2M_TYPE_IDENTITY;
 310
 311         mfn = xen_p2m_addr[pfn];
 312
 313         if (mfn == INVALID_P2M_ENTRY)
 314                 return P2M_TYPE_MISSING;
 315
 316         if (mfn & IDENTITY_FRAME_BIT)
 317                 return P2M_TYPE_IDENTITY;
 318
 319         return P2M_TYPE_PFN;
 320 }
 321
 322 static void __init xen_rebuild_p2m_list(unsigned long *p2m)
 323 {
 324         unsigned int i, chunk;
 325         unsigned long pfn;
 326         unsigned long *mfns;
 327         pte_t *ptep;
 328         pmd_t *pmdp;
 329         int type;
 330
 331         p2m_missing = alloc_p2m_page();
 332         p2m_init(p2m_missing);
 333         p2m_identity = alloc_p2m_page();
 334         p2m_init(p2m_identity);
 335
 336         p2m_missing_pte = alloc_p2m_page();
 337         paravirt_alloc_pte(&init_mm, __pa(p2m_missing_pte) >> PAGE_SHIFT);
 338         p2m_identity_pte = alloc_p2m_page();
 339         paravirt_alloc_pte(&init_mm, __pa(p2m_identity_pte) >> PAGE_SHIFT);
 340         for (i = 0; i < PTRS_PER_PTE; i++) {
 341                 set_pte(p2m_missing_pte + i,
 342                         pfn_pte(PFN_DOWN(__pa(p2m_missing)), PAGE_KERNEL_RO));
 343                 set_pte(p2m_identity_pte + i,
 344                         pfn_pte(PFN_DOWN(__pa(p2m_identity)), PAGE_KERNEL_RO));
 345         }
 346
 347         for (pfn = 0; pfn < xen_max_p2m_pfn; pfn += chunk) {
 348                 /*
 349                  * Try to map missing/identity PMDs or p2m-pages if possible.
 350                  * We have to respect the structure of the mfn_list_list
 351                  * which will be built just afterwards.
 352                  * Chunk size to test is one p2m page if we are in the middle
 353                  * of a mfn_list_list mid page and the complete mid page area
 354                  * if we are at index 0 of the mid page. Please note that a
 355                  * mid page might cover more than one PMD, e.g. on 32 bit PAE
 356                  * kernels.
 357                  */
 358                 chunk = (pfn & (P2M_PER_PAGE * P2M_MID_PER_PAGE - 1)) ?
 359                         P2M_PER_PAGE : P2M_PER_PAGE * P2M_MID_PER_PAGE;
 360
 361                 type = xen_p2m_elem_type(pfn);
 362                 i = 0;
 363                 if (type != P2M_TYPE_PFN)
 364                         for (i = 1; i < chunk; i++)
 365                                 if (xen_p2m_elem_type(pfn + i) != type)
 366                                         break;
 367                 if (i < chunk)
 368                         /* Reset to minimal chunk size. */
 369                         chunk = P2M_PER_PAGE;
 370
 371                 if (type == P2M_TYPE_PFN || i < chunk) {
 372                         /* Use initial p2m page contents. */
 373 #ifdef CONFIG_X86_64
 374                         mfns = alloc_p2m_page();
 375                         copy_page(mfns, xen_p2m_addr + pfn);
 376 #else
 377                         mfns = xen_p2m_addr + pfn;
 378 #endif
 379                         ptep = populate_extra_pte((unsigned long)(p2m + pfn));
 380                         set_pte(ptep,
 381                                 pfn_pte(PFN_DOWN(__pa(mfns)), PAGE_KERNEL));
 382                         continue;
 383                 }
 384
 385                 if (chunk == P2M_PER_PAGE) {
 386                         /* Map complete missing or identity p2m-page. */
 387                         mfns = (type == P2M_TYPE_MISSING) ?
 388                                 p2m_missing : p2m_identity;
 389                         ptep = populate_extra_pte((unsigned long)(p2m + pfn));
 390                         set_pte(ptep,
 391                                 pfn_pte(PFN_DOWN(__pa(mfns)), PAGE_KERNEL_RO));
 392                         continue;
 393                 }
 394
 395                 /* Complete missing or identity PMD(s) can be mapped. */
 396                 ptep = (type == P2M_TYPE_MISSING) ?
 397                         p2m_missing_pte : p2m_identity_pte;
 398                 for (i = 0; i < PMDS_PER_MID_PAGE; i++) {
 399                         pmdp = populate_extra_pmd(
 400                                 (unsigned long)(p2m + pfn) + i * PMD_SIZE);
 401                         set_pmd(pmdp, __pmd(__pa(ptep) | _KERNPG_TABLE));
 402                 }
 403         }
 404 }
 405
 406 void __init xen_vmalloc_p2m_tree(void)
 407 {
 408         static struct vm_struct vm;
 409         unsigned long p2m_limit;
 410
 411         xen_p2m_last_pfn = xen_max_p2m_pfn;
 412
 413         p2m_limit = (phys_addr_t)P2M_LIMIT * 1024 * 1024 * 1024 / PAGE_SIZE;
 414         vm.flags = VM_ALLOC;
 415         vm.size = ALIGN(sizeof(unsigned long) * max(xen_max_p2m_pfn, p2m_limit),
 416                         PMD_SIZE * PMDS_PER_MID_PAGE);
 417         vm_area_register_early(&vm, PMD_SIZE * PMDS_PER_MID_PAGE);
 418         pr_notice("p2m virtual area at %p, size is %lx\n", vm.addr, vm.size);
 419
 420         xen_max_p2m_pfn = vm.size / sizeof(unsigned long);
 421
 422         xen_rebuild_p2m_list(vm.addr);
 423
 424         xen_p2m_addr = vm.addr;
 425         xen_p2m_size = xen_max_p2m_pfn;
 426
 427         xen_inv_extra_mem();
 428 }
 429
 430 unsigned long get_phys_to_machine(unsigned long pfn)
 431 {
 432         pte_t *ptep;
 433         unsigned int level;
 434
 435         if (unlikely(pfn >= xen_p2m_size)) {
 436                 if (pfn < xen_max_p2m_pfn)
 437                         return xen_chk_extra_mem(pfn);
 438
 439                 return IDENTITY_FRAME(pfn);
 440         }
 441
 442         ptep = lookup_address((unsigned long)(xen_p2m_addr + pfn), &level);
 443         BUG_ON(!ptep || level != PG_LEVEL_4K);
 444
 445         /*
 446          * The INVALID_P2M_ENTRY is filled in both p2m_*identity
 447          * and in p2m_*missing, so returning the INVALID_P2M_ENTRY
 448          * would be wrong.
 449          */
 450         if (pte_pfn(*ptep) == PFN_DOWN(__pa(p2m_identity)))
 451                 return IDENTITY_FRAME(pfn);
 452
 453         return xen_p2m_addr[pfn];
 454 }
 455 EXPORT_SYMBOL_GPL(get_phys_to_machine);
 456
 457 /*
 458  * Allocate new pmd(s). It is checked whether the old pmd is still in place.
 459  * If not, nothing is changed. This is okay as the only reason for allocating
 460  * a new pmd is to replace p2m_missing_pte or p2m_identity_pte by a individual
 461  * pmd. In case of PAE/x86-32 there are multiple pmds to allocate!
 462  */
 463 static pte_t *alloc_p2m_pmd(unsigned long addr, pte_t *pte_pg)
 464 {
 465         pte_t *ptechk;
 466         pte_t *pte_newpg[PMDS_PER_MID_PAGE];
 467         pmd_t *pmdp;
 468         unsigned int level;
 469         unsigned long flags;
 470         unsigned long vaddr;
 471         int i;
 472
 473         /* Do all allocations first to bail out in error case. */
 474         for (i = 0; i < PMDS_PER_MID_PAGE; i++) {
 475                 pte_newpg[i] = alloc_p2m_page();
 476                 if (!pte_newpg[i]) {
 477                         for (i--; i >= 0; i--)
 478                                 free_p2m_page(pte_newpg[i]);
 479
 480                         return NULL;
 481                 }
 482         }
 483
 484         vaddr = addr & ~(PMD_SIZE * PMDS_PER_MID_PAGE - 1);
 485
 486         for (i = 0; i < PMDS_PER_MID_PAGE; i++) {
 487                 copy_page(pte_newpg[i], pte_pg);
 488                 paravirt_alloc_pte(&init_mm, __pa(pte_newpg[i]) >> PAGE_SHIFT);
 489
 490                 pmdp = lookup_pmd_address(vaddr);
 491                 BUG_ON(!pmdp);
 492
 493                 spin_lock_irqsave(&p2m_update_lock, flags);
 494
 495                 ptechk = lookup_address(vaddr, &level);
 496                 if (ptechk == pte_pg) {
 497                         HYPERVISOR_shared_info->arch.p2m_generation++;
 498                         wmb(); /* Tools are synchronizing via p2m_generation. */
 499                         set_pmd(pmdp,
 500                                 __pmd(__pa(pte_newpg[i]) | _KERNPG_TABLE));
 501                         wmb(); /* Tools are synchronizing via p2m_generation. */
 502                         HYPERVISOR_shared_info->arch.p2m_generation++;
 503                         pte_newpg[i] = NULL;
 504                 }
 505
 506                 spin_unlock_irqrestore(&p2m_update_lock, flags);
 507
 508                 if (pte_newpg[i]) {
 509                         paravirt_release_pte(__pa(pte_newpg[i]) >> PAGE_SHIFT);
 510                         free_p2m_page(pte_newpg[i]);
 511                 }
 512
 513                 vaddr += PMD_SIZE;
 514         }
 515
 516         return lookup_address(addr, &level);
 517 }
 518
 519 /*
 520  * Fully allocate the p2m structure for a given pfn.  We need to check
 521  * that both the top and mid levels are allocated, and make sure the
 522  * parallel mfn tree is kept in sync.  We may race with other cpus, so
 523  * the new pages are installed with cmpxchg; if we lose the race then
 524  * simply free the page we allocated and use the one that's there.
 525  */
 526 int xen_alloc_p2m_entry(unsigned long pfn)
 527 {
 528         unsigned topidx;
 529         unsigned long *top_mfn_p, *mid_mfn;
 530         pte_t *ptep, *pte_pg;
 531         unsigned int level;
 532         unsigned long flags;
 533         unsigned long addr = (unsigned long)(xen_p2m_addr + pfn);
 534         unsigned long p2m_pfn;
 535
 536         ptep = lookup_address(addr, &level);
 537         BUG_ON(!ptep || level != PG_LEVEL_4K);
 538         pte_pg = (pte_t *)((unsigned long)ptep & ~(PAGE_SIZE - 1));
 539
 540         if (pte_pg == p2m_missing_pte || pte_pg == p2m_identity_pte) {
 541                 /* PMD level is missing, allocate a new one */
 542                 ptep = alloc_p2m_pmd(addr, pte_pg);
 543                 if (!ptep)
 544                         return -ENOMEM;
 545         }
 546
 547         if (p2m_top_mfn && pfn < MAX_P2M_PFN) {
 548                 topidx = p2m_top_index(pfn);
 549                 top_mfn_p = &p2m_top_mfn[topidx];
 550                 mid_mfn = ACCESS_ONCE(p2m_top_mfn_p[topidx]);
 551
 552                 BUG_ON(virt_to_mfn(mid_mfn) != *top_mfn_p);
 553
 554                 if (mid_mfn == p2m_mid_missing_mfn) {
 555                         /* Separately check the mid mfn level */
 556                         unsigned long missing_mfn;
 557                         unsigned long mid_mfn_mfn;
 558                         unsigned long old_mfn;
 559
 560                         mid_mfn = alloc_p2m_page();
 561                         if (!mid_mfn)
 562                                 return -ENOMEM;
 563
 564                         p2m_mid_mfn_init(mid_mfn, p2m_missing);
 565
 566                         missing_mfn = virt_to_mfn(p2m_mid_missing_mfn);
 567                         mid_mfn_mfn = virt_to_mfn(mid_mfn);
 568                         old_mfn = cmpxchg(top_mfn_p, missing_mfn, mid_mfn_mfn);
 569                         if (old_mfn != missing_mfn) {
 570                                 free_p2m_page(mid_mfn);
 571                                 mid_mfn = mfn_to_virt(old_mfn);
 572                         } else {
 573                                 p2m_top_mfn_p[topidx] = mid_mfn;
 574                         }
 575                 }
 576         } else {
 577                 mid_mfn = NULL;
 578         }
 579
 580         p2m_pfn = pte_pfn(READ_ONCE(*ptep));
 581         if (p2m_pfn == PFN_DOWN(__pa(p2m_identity)) ||
 582             p2m_pfn == PFN_DOWN(__pa(p2m_missing))) {
 583                 /* p2m leaf page is missing */
 584                 unsigned long *p2m;
 585
 586                 p2m = alloc_p2m_page();
 587                 if (!p2m)
 588                         return -ENOMEM;
 589
 590                 if (p2m_pfn == PFN_DOWN(__pa(p2m_missing)))
 591                         p2m_init(p2m);
 592                 else
 593                         p2m_init_identity(p2m, pfn & ~(P2M_PER_PAGE - 1));
 594
 595                 spin_lock_irqsave(&p2m_update_lock, flags);
 596
 597                 if (pte_pfn(*ptep) == p2m_pfn) {
 598                         HYPERVISOR_shared_info->arch.p2m_generation++;
 599                         wmb(); /* Tools are synchronizing via p2m_generation. */
 600                         set_pte(ptep,
 601                                 pfn_pte(PFN_DOWN(__pa(p2m)), PAGE_KERNEL));
 602                         wmb(); /* Tools are synchronizing via p2m_generation. */
 603                         HYPERVISOR_shared_info->arch.p2m_generation++;
 604                         if (mid_mfn)
 605                                 mid_mfn[p2m_mid_index(pfn)] = virt_to_mfn(p2m);
 606                         p2m = NULL;
 607                 }
 608
 609                 spin_unlock_irqrestore(&p2m_update_lock, flags);
 610
 611                 if (p2m)
 612                         free_p2m_page(p2m);
 613         }
 614
 615         /* Expanded the p2m? */
 616         if (pfn >= xen_p2m_last_pfn) {
 617                 xen_p2m_last_pfn = ALIGN(pfn + 1, P2M_PER_PAGE);
 618                 HYPERVISOR_shared_info->arch.max_pfn = xen_p2m_last_pfn;
 619         }
 620
 621         return 0;
 622 }
 623 EXPORT_SYMBOL(xen_alloc_p2m_entry);
 624
 625 unsigned long __init set_phys_range_identity(unsigned long pfn_s,
 626                                       unsigned long pfn_e)
 627 {
 628         unsigned long pfn;
 629
 630         if (unlikely(pfn_s >= xen_p2m_size))
 631                 return 0;
 632
 633         if (pfn_s > pfn_e)
 634                 return 0;
 635
 636         if (pfn_e > xen_p2m_size)
 637                 pfn_e = xen_p2m_size;
 638
 639         for (pfn = pfn_s; pfn < pfn_e; pfn++)
 640                 xen_p2m_addr[pfn] = IDENTITY_FRAME(pfn);
 641
 642         return pfn - pfn_s;
 643 }
 644
 645 bool __set_phys_to_machine(unsigned long pfn, unsigned long mfn)
 646 {
 647         pte_t *ptep;
 648         unsigned int level;
 649
 650         if (unlikely(pfn >= xen_p2m_size)) {
 651                 BUG_ON(mfn != INVALID_P2M_ENTRY);
 652                 return true;
 653         }
 654
 655         /*
 656          * The interface requires atomic updates on p2m elements.
 657          * xen_safe_write_ulong() is using __put_user which does an atomic
 658          * store via asm().
 659          */
 660         if (likely(!xen_safe_write_ulong(xen_p2m_addr + pfn, mfn)))
 661                 return true;
 662
 663         ptep = lookup_address((unsigned long)(xen_p2m_addr + pfn), &level);
 664         BUG_ON(!ptep || level != PG_LEVEL_4K);
 665
 666         if (pte_pfn(*ptep) == PFN_DOWN(__pa(p2m_missing)))
 667                 return mfn == INVALID_P2M_ENTRY;
 668
 669         if (pte_pfn(*ptep) == PFN_DOWN(__pa(p2m_identity)))
 670                 return mfn == IDENTITY_FRAME(pfn);
 671
 672         return false;
 673 }
 674
 675 bool set_phys_to_machine(unsigned long pfn, unsigned long mfn)
 676 {
 677         if (unlikely(!__set_phys_to_machine(pfn, mfn))) {
 678                 int ret;
 679
 680                 ret = xen_alloc_p2m_entry(pfn);
 681                 if (ret < 0)
 682                         return false;
 683
 684                 return __set_phys_to_machine(pfn, mfn);
 685         }
 686
 687         return true;
 688 }
 689
 690 int set_foreign_p2m_mapping(struct gnttab_map_grant_ref *map_ops,
 691                             struct gnttab_map_grant_ref *kmap_ops,
 692                             struct page **pages, unsigned int count)
 693 {
 694         int i, ret = 0;
 695         pte_t *pte;
 696
 697         if (xen_feature(XENFEAT_auto_translated_physmap))
 698                 return 0;
 699
 700         if (kmap_ops) {
 701                 ret = HYPERVISOR_grant_table_op(GNTTABOP_map_grant_ref,
 702                                                 kmap_ops, count);
 703                 if (ret)
 704                         goto out;
 705         }
 706
 707         for (i = 0; i < count; i++) {
 708                 unsigned long mfn, pfn;
 709                 struct gnttab_unmap_grant_ref unmap[2];
 710                 int rc;
 711
 712                 /* Do not add to override if the map failed. */
 713                 if (map_ops[i].status != GNTST_okay ||
 714                     (kmap_ops && kmap_ops[i].status != GNTST_okay))
 715                         continue;
 716
 717                 if (map_ops[i].flags & GNTMAP_contains_pte) {
 718                         pte = (pte_t *)(mfn_to_virt(PFN_DOWN(map_ops[i].host_addr)) +
 719                                 (map_ops[i].host_addr & ~PAGE_MASK));
 720                         mfn = pte_mfn(*pte);
 721                 } else {
 722                         mfn = PFN_DOWN(map_ops[i].dev_bus_addr);
 723                 }
 724                 pfn = page_to_pfn(pages[i]);
 725
 726                 WARN(pfn_to_mfn(pfn) != INVALID_P2M_ENTRY, "page must be ballooned");
 727
 728                 if (likely(set_phys_to_machine(pfn, FOREIGN_FRAME(mfn))))
 729                         continue;
 730
 731                 /*
 732                  * Signal an error for this slot. This in turn requires
 733                  * immediate unmapping.
 734                  */
 735                 map_ops[i].status = GNTST_general_error;
 736                 unmap[0].host_addr = map_ops[i].host_addr,
 737                 unmap[0].handle = map_ops[i].handle;
 738                 map_ops[i].handle = ~0;
 739                 if (map_ops[i].flags & GNTMAP_device_map)
 740                         unmap[0].dev_bus_addr = map_ops[i].dev_bus_addr;
 741                 else
 742                         unmap[0].dev_bus_addr = 0;
 743
 744                 if (kmap_ops) {
 745                         kmap_ops[i].status = GNTST_general_error;
 746                         unmap[1].host_addr = kmap_ops[i].host_addr,
 747                         unmap[1].handle = kmap_ops[i].handle;
 748                         kmap_ops[i].handle = ~0;
 749                         if (kmap_ops[i].flags & GNTMAP_device_map)
 750                                 unmap[1].dev_bus_addr = kmap_ops[i].dev_bus_addr;
 751                         else
 752                                 unmap[1].dev_bus_addr = 0;
 753                 }
 754
 755                 /*
 756                  * Pre-populate both status fields, to be recognizable in
 757                  * the log message below.
 758                  */
 759                 unmap[0].status = 1;
 760                 unmap[1].status = 1;
 761
 762                 rc = HYPERVISOR_grant_table_op(GNTTABOP_unmap_grant_ref,
 763                                                unmap, 1 + !!kmap_ops);
 764                 if (rc || unmap[0].status != GNTST_okay ||
 765                     unmap[1].status != GNTST_okay)
 766                         pr_err_once("gnttab unmap failed: rc=%d st0=%d st1=%d\n",
 767                                     rc, unmap[0].status, unmap[1].status);
 768         }
 769
 770 out:
 771         return ret;
 772 }
 773 EXPORT_SYMBOL_GPL(set_foreign_p2m_mapping);
 774
 775 int clear_foreign_p2m_mapping(struct gnttab_unmap_grant_ref *unmap_ops,
 776                               struct gnttab_unmap_grant_ref *kunmap_ops,
 777                               struct page **pages, unsigned int count)
 778 {
 779         int i, ret = 0;
 780
 781         if (xen_feature(XENFEAT_auto_translated_physmap))
 782                 return 0;
 783
 784         for (i = 0; i < count; i++) {
 785                 unsigned long mfn = __pfn_to_mfn(page_to_pfn(pages[i]));
 786                 unsigned long pfn = page_to_pfn(pages[i]);
 787
 788                 if (mfn != INVALID_P2M_ENTRY && (mfn & FOREIGN_FRAME_BIT))
 789                         set_phys_to_machine(pfn, INVALID_P2M_ENTRY);
 790                 else
 791                         ret = -EINVAL;
 792         }
 793         if (kunmap_ops)
 794                 ret = HYPERVISOR_grant_table_op(GNTTABOP_unmap_grant_ref,
 795                                                 kunmap_ops, count) ?: ret;
 796
 797         return ret;
 798 }
 799 EXPORT_SYMBOL_GPL(clear_foreign_p2m_mapping);
 800
 801 #ifdef CONFIG_XEN_DEBUG_FS
 802 #include <linux/debugfs.h>
 803 #include "debugfs.h"
 804 static int p2m_dump_show(struct seq_file *m, void *v)
 805 {
 806         static const char * const type_name[] = {
 807                                 [P2M_TYPE_IDENTITY] = "identity",
 808                                 [P2M_TYPE_MISSING] = "missing",
 809                                 [P2M_TYPE_PFN] = "pfn",
 810                                 [P2M_TYPE_UNKNOWN] = "abnormal"};
 811         unsigned long pfn, first_pfn;
 812         int type, prev_type;
 813
 814         prev_type = xen_p2m_elem_type(0);
 815         first_pfn = 0;
 816
 817         for (pfn = 0; pfn < xen_p2m_size; pfn++) {
 818                 type = xen_p2m_elem_type(pfn);
 819                 if (type != prev_type) {
 820                         seq_printf(m, " [0x%lx->0x%lx] %s\n", first_pfn, pfn,
 821                                    type_name[prev_type]);
 822                         prev_type = type;
 823                         first_pfn = pfn;
 824                 }
 825         }
 826         seq_printf(m, " [0x%lx->0x%lx] %s\n", first_pfn, pfn,
 827                    type_name[prev_type]);
 828         return 0;
 829 }
 830
 831 static int p2m_dump_open(struct inode *inode, struct file *filp)
 832 {
 833         return single_open(filp, p2m_dump_show, NULL);
 834 }
 835
 836 static const struct file_operations p2m_dump_fops = {
 837         .open           = p2m_dump_open,
 838         .read           = seq_read,
 839         .llseek         = seq_lseek,
 840         .release        = single_release,
 841 };
 842
 843 static struct dentry *d_mmu_debug;
 844
 845 static int __init xen_p2m_debugfs(void)
 846 {
 847         struct dentry *d_xen = xen_init_debugfs();
 848
 849         if (d_xen == NULL)
 850                 return -ENOMEM;
 851
 852         d_mmu_debug = debugfs_create_dir("mmu", d_xen);
 853
 854         debugfs_create_file("p2m", 0600, d_mmu_debug, NULL, &p2m_dump_fops);
 855         return 0;
 856 }
 857 fs_initcall(xen_p2m_debugfs);
 858 #endif /* CONFIG_XEN_DEBUG_FS */