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 #include <linux/vmalloc.h>
  71
  72 #include <asm/cache.h>
  73 #include <asm/setup.h>
  74 #include <asm/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 static inline unsigned p2m_top_index(unsigned long pfn)
 116 {
 117         BUG_ON(pfn >= MAX_P2M_PFN);
 118         return pfn / (P2M_MID_PER_PAGE * P2M_PER_PAGE);
 119 }
 120
 121 static inline unsigned p2m_mid_index(unsigned long pfn)
 122 {
 123         return (pfn / P2M_PER_PAGE) % P2M_MID_PER_PAGE;
 124 }
 125
 126 static inline unsigned p2m_index(unsigned long pfn)
 127 {
 128         return pfn % P2M_PER_PAGE;
 129 }
 130
 131 static void p2m_top_mfn_init(unsigned long *top)
 132 {
 133         unsigned i;
 134
 135         for (i = 0; i < P2M_TOP_PER_PAGE; i++)
 136                 top[i] = virt_to_mfn(p2m_mid_missing_mfn);
 137 }
 138
 139 static void p2m_top_mfn_p_init(unsigned long **top)
 140 {
 141         unsigned i;
 142
 143         for (i = 0; i < P2M_TOP_PER_PAGE; i++)
 144                 top[i] = p2m_mid_missing_mfn;
 145 }
 146
 147 static void p2m_mid_mfn_init(unsigned long *mid, unsigned long *leaf)
 148 {
 149         unsigned i;
 150
 151         for (i = 0; i < P2M_MID_PER_PAGE; i++)
 152                 mid[i] = virt_to_mfn(leaf);
 153 }
 154
 155 static void p2m_init(unsigned long *p2m)
 156 {
 157         unsigned i;
 158
 159         for (i = 0; i < P2M_PER_PAGE; i++)
 160                 p2m[i] = INVALID_P2M_ENTRY;
 161 }
 162
 163 static void p2m_init_identity(unsigned long *p2m, unsigned long pfn)
 164 {
 165         unsigned i;
 166
 167         for (i = 0; i < P2M_PER_PAGE; i++)
 168                 p2m[i] = IDENTITY_FRAME(pfn + i);
 169 }
 170
 171 static void * __ref alloc_p2m_page(void)
 172 {
 173         if (unlikely(!slab_is_available()))
 174                 return alloc_bootmem_align(PAGE_SIZE, PAGE_SIZE);
 175
 176         return (void *)__get_free_page(GFP_KERNEL | __GFP_REPEAT);
 177 }
 178
 179 static void __ref free_p2m_page(void *p)
 180 {
 181         if (unlikely(!slab_is_available())) {
 182                 free_bootmem((unsigned long)p, PAGE_SIZE);
 183                 return;
 184         }
 185
 186         free_page((unsigned long)p);
 187 }
 188
 189 /*
 190  * Build the parallel p2m_top_mfn and p2m_mid_mfn structures
 191  *
 192  * This is called both at boot time, and after resuming from suspend:
 193  * - At boot time we're called rather early, and must use alloc_bootmem*()
 194  *   to allocate memory.
 195  *
 196  * - After resume we're called from within stop_machine, but the mfn
 197  *   tree should already be completely allocated.
 198  */
 199 void __ref xen_build_mfn_list_list(void)
 200 {
 201         unsigned long pfn, mfn;
 202         pte_t *ptep;
 203         unsigned int level, topidx, mididx;
 204         unsigned long *mid_mfn_p;
 205
 206         if (xen_feature(XENFEAT_auto_translated_physmap) ||
 207             xen_start_info->flags & SIF_VIRT_P2M_4TOOLS)
 208                 return;
 209
 210         /* Pre-initialize p2m_top_mfn to be completely missing */
 211         if (p2m_top_mfn == NULL) {
 212                 p2m_mid_missing_mfn = alloc_p2m_page();
 213                 p2m_mid_mfn_init(p2m_mid_missing_mfn, p2m_missing);
 214
 215                 p2m_top_mfn_p = alloc_p2m_page();
 216                 p2m_top_mfn_p_init(p2m_top_mfn_p);
 217
 218                 p2m_top_mfn = alloc_p2m_page();
 219                 p2m_top_mfn_init(p2m_top_mfn);
 220         } else {
 221                 /* Reinitialise, mfn's all change after migration */
 222                 p2m_mid_mfn_init(p2m_mid_missing_mfn, p2m_missing);
 223         }
 224
 225         for (pfn = 0; pfn < xen_max_p2m_pfn && pfn < MAX_P2M_PFN;
 226              pfn += P2M_PER_PAGE) {
 227                 topidx = p2m_top_index(pfn);
 228                 mididx = p2m_mid_index(pfn);
 229
 230                 mid_mfn_p = p2m_top_mfn_p[topidx];
 231                 ptep = lookup_address((unsigned long)(xen_p2m_addr + pfn),
 232                                       &level);
 233                 BUG_ON(!ptep || level != PG_LEVEL_4K);
 234                 mfn = pte_mfn(*ptep);
 235                 ptep = (pte_t *)((unsigned long)ptep & ~(PAGE_SIZE - 1));
 236
 237                 /* Don't bother allocating any mfn mid levels if
 238                  * they're just missing, just update the stored mfn,
 239                  * since all could have changed over a migrate.
 240                  */
 241                 if (ptep == p2m_missing_pte || ptep == p2m_identity_pte) {
 242                         BUG_ON(mididx);
 243                         BUG_ON(mid_mfn_p != p2m_mid_missing_mfn);
 244                         p2m_top_mfn[topidx] = virt_to_mfn(p2m_mid_missing_mfn);
 245                         pfn += (P2M_MID_PER_PAGE - 1) * P2M_PER_PAGE;
 246                         continue;
 247                 }
 248
 249                 if (mid_mfn_p == p2m_mid_missing_mfn) {
 250                         mid_mfn_p = alloc_p2m_page();
 251                         p2m_mid_mfn_init(mid_mfn_p, p2m_missing);
 252
 253                         p2m_top_mfn_p[topidx] = mid_mfn_p;
 254                 }
 255
 256                 p2m_top_mfn[topidx] = virt_to_mfn(mid_mfn_p);
 257                 mid_mfn_p[mididx] = mfn;
 258         }
 259 }
 260
 261 void xen_setup_mfn_list_list(void)
 262 {
 263         if (xen_feature(XENFEAT_auto_translated_physmap))
 264                 return;
 265
 266         BUG_ON(HYPERVISOR_shared_info == &xen_dummy_shared_info);
 267
 268         if (xen_start_info->flags & SIF_VIRT_P2M_4TOOLS)
 269                 HYPERVISOR_shared_info->arch.pfn_to_mfn_frame_list_list = ~0UL;
 270         else
 271                 HYPERVISOR_shared_info->arch.pfn_to_mfn_frame_list_list =
 272                         virt_to_mfn(p2m_top_mfn);
 273         HYPERVISOR_shared_info->arch.max_pfn = xen_max_p2m_pfn;
 274         HYPERVISOR_shared_info->arch.p2m_generation = 0;
 275         HYPERVISOR_shared_info->arch.p2m_vaddr = (unsigned long)xen_p2m_addr;
 276         HYPERVISOR_shared_info->arch.p2m_cr3 =
 277                 xen_pfn_to_cr3(virt_to_mfn(swapper_pg_dir));
 278 }
 279
 280 /* Set up p2m_top to point to the domain-builder provided p2m pages */
 281 void __init xen_build_dynamic_phys_to_machine(void)
 282 {
 283         unsigned long pfn;
 284
 285          if (xen_feature(XENFEAT_auto_translated_physmap))
 286                 return;
 287
 288         xen_p2m_addr = (unsigned long *)xen_start_info->mfn_list;
 289         xen_p2m_size = ALIGN(xen_start_info->nr_pages, P2M_PER_PAGE);
 290
 291         for (pfn = xen_start_info->nr_pages; pfn < xen_p2m_size; pfn++)
 292                 xen_p2m_addr[pfn] = INVALID_P2M_ENTRY;
 293
 294         xen_max_p2m_pfn = xen_p2m_size;
 295 }
 296
 297 #define P2M_TYPE_IDENTITY       0
 298 #define P2M_TYPE_MISSING        1
 299 #define P2M_TYPE_PFN            2
 300 #define P2M_TYPE_UNKNOWN        3
 301
 302 static int xen_p2m_elem_type(unsigned long pfn)
 303 {
 304         unsigned long mfn;
 305
 306         if (pfn >= xen_p2m_size)
 307                 return P2M_TYPE_IDENTITY;
 308
 309         mfn = xen_p2m_addr[pfn];
 310
 311         if (mfn == INVALID_P2M_ENTRY)
 312                 return P2M_TYPE_MISSING;
 313
 314         if (mfn & IDENTITY_FRAME_BIT)
 315                 return P2M_TYPE_IDENTITY;
 316
 317         return P2M_TYPE_PFN;
 318 }
 319
 320 static void __init xen_rebuild_p2m_list(unsigned long *p2m)
 321 {
 322         unsigned int i, chunk;
 323         unsigned long pfn;
 324         unsigned long *mfns;
 325         pte_t *ptep;
 326         pmd_t *pmdp;
 327         int type;
 328
 329         p2m_missing = alloc_p2m_page();
 330         p2m_init(p2m_missing);
 331         p2m_identity = alloc_p2m_page();
 332         p2m_init(p2m_identity);
 333
 334         p2m_missing_pte = alloc_p2m_page();
 335         paravirt_alloc_pte(&init_mm, __pa(p2m_missing_pte) >> PAGE_SHIFT);
 336         p2m_identity_pte = alloc_p2m_page();
 337         paravirt_alloc_pte(&init_mm, __pa(p2m_identity_pte) >> PAGE_SHIFT);
 338         for (i = 0; i < PTRS_PER_PTE; i++) {
 339                 set_pte(p2m_missing_pte + i,
 340                         pfn_pte(PFN_DOWN(__pa(p2m_missing)), PAGE_KERNEL_RO));
 341                 set_pte(p2m_identity_pte + i,
 342                         pfn_pte(PFN_DOWN(__pa(p2m_identity)), PAGE_KERNEL_RO));
 343         }
 344
 345         for (pfn = 0; pfn < xen_max_p2m_pfn; pfn += chunk) {
 346                 /*
 347                  * Try to map missing/identity PMDs or p2m-pages if possible.
 348                  * We have to respect the structure of the mfn_list_list
 349                  * which will be built just afterwards.
 350                  * Chunk size to test is one p2m page if we are in the middle
 351                  * of a mfn_list_list mid page and the complete mid page area
 352                  * if we are at index 0 of the mid page. Please note that a
 353                  * mid page might cover more than one PMD, e.g. on 32 bit PAE
 354                  * kernels.
 355                  */
 356                 chunk = (pfn & (P2M_PER_PAGE * P2M_MID_PER_PAGE - 1)) ?
 357                         P2M_PER_PAGE : P2M_PER_PAGE * P2M_MID_PER_PAGE;
 358
 359                 type = xen_p2m_elem_type(pfn);
 360                 i = 0;
 361                 if (type != P2M_TYPE_PFN)
 362                         for (i = 1; i < chunk; i++)
 363                                 if (xen_p2m_elem_type(pfn + i) != type)
 364                                         break;
 365                 if (i < chunk)
 366                         /* Reset to minimal chunk size. */
 367                         chunk = P2M_PER_PAGE;
 368
 369                 if (type == P2M_TYPE_PFN || i < chunk) {
 370                         /* Use initial p2m page contents. */
 371 #ifdef CONFIG_X86_64
 372                         mfns = alloc_p2m_page();
 373                         copy_page(mfns, xen_p2m_addr + pfn);
 374 #else
 375                         mfns = xen_p2m_addr + pfn;
 376 #endif
 377                         ptep = populate_extra_pte((unsigned long)(p2m + pfn));
 378                         set_pte(ptep,
 379                                 pfn_pte(PFN_DOWN(__pa(mfns)), PAGE_KERNEL));
 380                         continue;
 381                 }
 382
 383                 if (chunk == P2M_PER_PAGE) {
 384                         /* Map complete missing or identity p2m-page. */
 385                         mfns = (type == P2M_TYPE_MISSING) ?
 386                                 p2m_missing : p2m_identity;
 387                         ptep = populate_extra_pte((unsigned long)(p2m + pfn));
 388                         set_pte(ptep,
 389                                 pfn_pte(PFN_DOWN(__pa(mfns)), PAGE_KERNEL_RO));
 390                         continue;
 391                 }
 392
 393                 /* Complete missing or identity PMD(s) can be mapped. */
 394                 ptep = (type == P2M_TYPE_MISSING) ?
 395                         p2m_missing_pte : p2m_identity_pte;
 396                 for (i = 0; i < PMDS_PER_MID_PAGE; i++) {
 397                         pmdp = populate_extra_pmd(
 398                                 (unsigned long)(p2m + pfn) + i * PMD_SIZE);
 399                         set_pmd(pmdp, __pmd(__pa(ptep) | _KERNPG_TABLE));
 400                 }
 401         }
 402 }
 403
 404 void __init xen_vmalloc_p2m_tree(void)
 405 {
 406         static struct vm_struct vm;
 407         unsigned long p2m_limit;
 408
 409         p2m_limit = (phys_addr_t)P2M_LIMIT * 1024 * 1024 * 1024 / PAGE_SIZE;
 410         vm.flags = VM_ALLOC;
 411         vm.size = ALIGN(sizeof(unsigned long) * max(xen_max_p2m_pfn, p2m_limit),
 412                         PMD_SIZE * PMDS_PER_MID_PAGE);
 413         vm_area_register_early(&vm, PMD_SIZE * PMDS_PER_MID_PAGE);
 414         pr_notice("p2m virtual area at %p, size is %lx\n", vm.addr, vm.size);
 415
 416         xen_max_p2m_pfn = vm.size / sizeof(unsigned long);
 417
 418         xen_rebuild_p2m_list(vm.addr);
 419
 420         xen_p2m_addr = vm.addr;
 421         xen_p2m_size = xen_max_p2m_pfn;
 422
 423         xen_inv_extra_mem();
 424 }
 425
 426 unsigned long get_phys_to_machine(unsigned long pfn)
 427 {
 428         pte_t *ptep;
 429         unsigned int level;
 430
 431         if (unlikely(pfn >= xen_p2m_size)) {
 432                 if (pfn < xen_max_p2m_pfn)
 433                         return xen_chk_extra_mem(pfn);
 434
 435                 return IDENTITY_FRAME(pfn);
 436         }
 437
 438         ptep = lookup_address((unsigned long)(xen_p2m_addr + pfn), &level);
 439         BUG_ON(!ptep || level != PG_LEVEL_4K);
 440
 441         /*
 442          * The INVALID_P2M_ENTRY is filled in both p2m_*identity
 443          * and in p2m_*missing, so returning the INVALID_P2M_ENTRY
 444          * would be wrong.
 445          */
 446         if (pte_pfn(*ptep) == PFN_DOWN(__pa(p2m_identity)))
 447                 return IDENTITY_FRAME(pfn);
 448
 449         return xen_p2m_addr[pfn];
 450 }
 451 EXPORT_SYMBOL_GPL(get_phys_to_machine);
 452
 453 /*
 454  * Allocate new pmd(s). It is checked whether the old pmd is still in place.
 455  * If not, nothing is changed. This is okay as the only reason for allocating
 456  * a new pmd is to replace p2m_missing_pte or p2m_identity_pte by a individual
 457  * pmd. In case of PAE/x86-32 there are multiple pmds to allocate!
 458  */
 459 static pte_t *alloc_p2m_pmd(unsigned long addr, pte_t *pte_pg)
 460 {
 461         pte_t *ptechk;
 462         pte_t *pte_newpg[PMDS_PER_MID_PAGE];
 463         pmd_t *pmdp;
 464         unsigned int level;
 465         unsigned long flags;
 466         unsigned long vaddr;
 467         int i;
 468
 469         /* Do all allocations first to bail out in error case. */
 470         for (i = 0; i < PMDS_PER_MID_PAGE; i++) {
 471                 pte_newpg[i] = alloc_p2m_page();
 472                 if (!pte_newpg[i]) {
 473                         for (i--; i >= 0; i--)
 474                                 free_p2m_page(pte_newpg[i]);
 475
 476                         return NULL;
 477                 }
 478         }
 479
 480         vaddr = addr & ~(PMD_SIZE * PMDS_PER_MID_PAGE - 1);
 481
 482         for (i = 0; i < PMDS_PER_MID_PAGE; i++) {
 483                 copy_page(pte_newpg[i], pte_pg);
 484                 paravirt_alloc_pte(&init_mm, __pa(pte_newpg[i]) >> PAGE_SHIFT);
 485
 486                 pmdp = lookup_pmd_address(vaddr);
 487                 BUG_ON(!pmdp);
 488
 489                 spin_lock_irqsave(&p2m_update_lock, flags);
 490
 491                 ptechk = lookup_address(vaddr, &level);
 492                 if (ptechk == pte_pg) {
 493                         HYPERVISOR_shared_info->arch.p2m_generation++;
 494                         wmb(); /* Tools are synchronizing via p2m_generation. */
 495                         set_pmd(pmdp,
 496                                 __pmd(__pa(pte_newpg[i]) | _KERNPG_TABLE));
 497                         wmb(); /* Tools are synchronizing via p2m_generation. */
 498                         HYPERVISOR_shared_info->arch.p2m_generation++;
 499                         pte_newpg[i] = NULL;
 500                 }
 501
 502                 spin_unlock_irqrestore(&p2m_update_lock, flags);
 503
 504                 if (pte_newpg[i]) {
 505                         paravirt_release_pte(__pa(pte_newpg[i]) >> PAGE_SHIFT);
 506                         free_p2m_page(pte_newpg[i]);
 507                 }
 508
 509                 vaddr += PMD_SIZE;
 510         }
 511
 512         return lookup_address(addr, &level);
 513 }
 514
 515 /*
 516  * Fully allocate the p2m structure for a given pfn.  We need to check
 517  * that both the top and mid levels are allocated, and make sure the
 518  * parallel mfn tree is kept in sync.  We may race with other cpus, so
 519  * the new pages are installed with cmpxchg; if we lose the race then
 520  * simply free the page we allocated and use the one that's there.
 521  */
 522 static bool alloc_p2m(unsigned long pfn)
 523 {
 524         unsigned topidx;
 525         unsigned long *top_mfn_p, *mid_mfn;
 526         pte_t *ptep, *pte_pg;
 527         unsigned int level;
 528         unsigned long flags;
 529         unsigned long addr = (unsigned long)(xen_p2m_addr + pfn);
 530         unsigned long p2m_pfn;
 531
 532         ptep = lookup_address(addr, &level);
 533         BUG_ON(!ptep || level != PG_LEVEL_4K);
 534         pte_pg = (pte_t *)((unsigned long)ptep & ~(PAGE_SIZE - 1));
 535
 536         if (pte_pg == p2m_missing_pte || pte_pg == p2m_identity_pte) {
 537                 /* PMD level is missing, allocate a new one */
 538                 ptep = alloc_p2m_pmd(addr, pte_pg);
 539                 if (!ptep)
 540                         return false;
 541         }
 542
 543         if (p2m_top_mfn && pfn < MAX_P2M_PFN) {
 544                 topidx = p2m_top_index(pfn);
 545                 top_mfn_p = &p2m_top_mfn[topidx];
 546                 mid_mfn = ACCESS_ONCE(p2m_top_mfn_p[topidx]);
 547
 548                 BUG_ON(virt_to_mfn(mid_mfn) != *top_mfn_p);
 549
 550                 if (mid_mfn == p2m_mid_missing_mfn) {
 551                         /* Separately check the mid mfn level */
 552                         unsigned long missing_mfn;
 553                         unsigned long mid_mfn_mfn;
 554                         unsigned long old_mfn;
 555
 556                         mid_mfn = alloc_p2m_page();
 557                         if (!mid_mfn)
 558                                 return false;
 559
 560                         p2m_mid_mfn_init(mid_mfn, p2m_missing);
 561
 562                         missing_mfn = virt_to_mfn(p2m_mid_missing_mfn);
 563                         mid_mfn_mfn = virt_to_mfn(mid_mfn);
 564                         old_mfn = cmpxchg(top_mfn_p, missing_mfn, mid_mfn_mfn);
 565                         if (old_mfn != missing_mfn) {
 566                                 free_p2m_page(mid_mfn);
 567                                 mid_mfn = mfn_to_virt(old_mfn);
 568                         } else {
 569                                 p2m_top_mfn_p[topidx] = mid_mfn;
 570                         }
 571                 }
 572         } else {
 573                 mid_mfn = NULL;
 574         }
 575
 576         p2m_pfn = pte_pfn(READ_ONCE(*ptep));
 577         if (p2m_pfn == PFN_DOWN(__pa(p2m_identity)) ||
 578             p2m_pfn == PFN_DOWN(__pa(p2m_missing))) {
 579                 /* p2m leaf page is missing */
 580                 unsigned long *p2m;
 581
 582                 p2m = alloc_p2m_page();
 583                 if (!p2m)
 584                         return false;
 585
 586                 if (p2m_pfn == PFN_DOWN(__pa(p2m_missing)))
 587                         p2m_init(p2m);
 588                 else
 589                         p2m_init_identity(p2m, pfn & ~(P2M_PER_PAGE - 1));
 590
 591                 spin_lock_irqsave(&p2m_update_lock, flags);
 592
 593                 if (pte_pfn(*ptep) == p2m_pfn) {
 594                         HYPERVISOR_shared_info->arch.p2m_generation++;
 595                         wmb(); /* Tools are synchronizing via p2m_generation. */
 596                         set_pte(ptep,
 597                                 pfn_pte(PFN_DOWN(__pa(p2m)), PAGE_KERNEL));
 598                         wmb(); /* Tools are synchronizing via p2m_generation. */
 599                         HYPERVISOR_shared_info->arch.p2m_generation++;
 600                         if (mid_mfn)
 601                                 mid_mfn[p2m_mid_index(pfn)] = virt_to_mfn(p2m);
 602                         p2m = NULL;
 603                 }
 604
 605                 spin_unlock_irqrestore(&p2m_update_lock, flags);
 606
 607                 if (p2m)
 608                         free_p2m_page(p2m);
 609         }
 610
 611         return true;
 612 }
 613
 614 unsigned long __init set_phys_range_identity(unsigned long pfn_s,
 615                                       unsigned long pfn_e)
 616 {
 617         unsigned long pfn;
 618
 619         if (unlikely(pfn_s >= xen_p2m_size))
 620                 return 0;
 621
 622         if (unlikely(xen_feature(XENFEAT_auto_translated_physmap)))
 623                 return pfn_e - pfn_s;
 624
 625         if (pfn_s > pfn_e)
 626                 return 0;
 627
 628         if (pfn_e > xen_p2m_size)
 629                 pfn_e = xen_p2m_size;
 630
 631         for (pfn = pfn_s; pfn < pfn_e; pfn++)
 632                 xen_p2m_addr[pfn] = IDENTITY_FRAME(pfn);
 633
 634         return pfn - pfn_s;
 635 }
 636
 637 bool __set_phys_to_machine(unsigned long pfn, unsigned long mfn)
 638 {
 639         pte_t *ptep;
 640         unsigned int level;
 641
 642         /* don't track P2M changes in autotranslate guests */
 643         if (unlikely(xen_feature(XENFEAT_auto_translated_physmap)))
 644                 return true;
 645
 646         if (unlikely(pfn >= xen_p2m_size)) {
 647                 BUG_ON(mfn != INVALID_P2M_ENTRY);
 648                 return true;
 649         }
 650
 651         /*
 652          * The interface requires atomic updates on p2m elements.
 653          * xen_safe_write_ulong() is using __put_user which does an atomic
 654          * store via asm().
 655          */
 656         if (likely(!xen_safe_write_ulong(xen_p2m_addr + pfn, mfn)))
 657                 return true;
 658
 659         ptep = lookup_address((unsigned long)(xen_p2m_addr + pfn), &level);
 660         BUG_ON(!ptep || level != PG_LEVEL_4K);
 661
 662         if (pte_pfn(*ptep) == PFN_DOWN(__pa(p2m_missing)))
 663                 return mfn == INVALID_P2M_ENTRY;
 664
 665         if (pte_pfn(*ptep) == PFN_DOWN(__pa(p2m_identity)))
 666                 return mfn == IDENTITY_FRAME(pfn);
 667
 668         return false;
 669 }
 670
 671 bool set_phys_to_machine(unsigned long pfn, unsigned long mfn)
 672 {
 673         if (unlikely(!__set_phys_to_machine(pfn, mfn))) {
 674                 if (!alloc_p2m(pfn))
 675                         return false;
 676
 677                 return __set_phys_to_machine(pfn, mfn);
 678         }
 679
 680         return true;
 681 }
 682
 683 int set_foreign_p2m_mapping(struct gnttab_map_grant_ref *map_ops,
 684                             struct gnttab_map_grant_ref *kmap_ops,
 685                             struct page **pages, unsigned int count)
 686 {
 687         int i, ret = 0;
 688         pte_t *pte;
 689
 690         if (xen_feature(XENFEAT_auto_translated_physmap))
 691                 return 0;
 692
 693         if (kmap_ops) {
 694                 ret = HYPERVISOR_grant_table_op(GNTTABOP_map_grant_ref,
 695                                                 kmap_ops, count);
 696                 if (ret)
 697                         goto out;
 698         }
 699
 700         for (i = 0; i < count; i++) {
 701                 unsigned long mfn, pfn;
 702
 703                 /* Do not add to override if the map failed. */
 704                 if (map_ops[i].status)
 705                         continue;
 706
 707                 if (map_ops[i].flags & GNTMAP_contains_pte) {
 708                         pte = (pte_t *)(mfn_to_virt(PFN_DOWN(map_ops[i].host_addr)) +
 709                                 (map_ops[i].host_addr & ~PAGE_MASK));
 710                         mfn = pte_mfn(*pte);
 711                 } else {
 712                         mfn = PFN_DOWN(map_ops[i].dev_bus_addr);
 713                 }
 714                 pfn = page_to_pfn(pages[i]);
 715
 716                 WARN(pfn_to_mfn(pfn) != INVALID_P2M_ENTRY, "page must be ballooned");
 717
 718                 if (unlikely(!set_phys_to_machine(pfn, FOREIGN_FRAME(mfn)))) {
 719                         ret = -ENOMEM;
 720                         goto out;
 721                 }
 722         }
 723
 724 out:
 725         return ret;
 726 }
 727 EXPORT_SYMBOL_GPL(set_foreign_p2m_mapping);
 728
 729 int clear_foreign_p2m_mapping(struct gnttab_unmap_grant_ref *unmap_ops,
 730                               struct gnttab_unmap_grant_ref *kunmap_ops,
 731                               struct page **pages, unsigned int count)
 732 {
 733         int i, ret = 0;
 734
 735         if (xen_feature(XENFEAT_auto_translated_physmap))
 736                 return 0;
 737
 738         for (i = 0; i < count; i++) {
 739                 unsigned long mfn = __pfn_to_mfn(page_to_pfn(pages[i]));
 740                 unsigned long pfn = page_to_pfn(pages[i]);
 741
 742                 if (mfn == INVALID_P2M_ENTRY || !(mfn & FOREIGN_FRAME_BIT)) {
 743                         ret = -EINVAL;
 744                         goto out;
 745                 }
 746
 747                 set_phys_to_machine(pfn, INVALID_P2M_ENTRY);
 748         }
 749         if (kunmap_ops)
 750                 ret = HYPERVISOR_grant_table_op(GNTTABOP_unmap_grant_ref,
 751                                                 kunmap_ops, count);
 752 out:
 753         return ret;
 754 }
 755 EXPORT_SYMBOL_GPL(clear_foreign_p2m_mapping);
 756
 757 #ifdef CONFIG_XEN_DEBUG_FS
 758 #include <linux/debugfs.h>
 759 #include "debugfs.h"
 760 static int p2m_dump_show(struct seq_file *m, void *v)
 761 {
 762         static const char * const type_name[] = {
 763                                 [P2M_TYPE_IDENTITY] = "identity",
 764                                 [P2M_TYPE_MISSING] = "missing",
 765                                 [P2M_TYPE_PFN] = "pfn",
 766                                 [P2M_TYPE_UNKNOWN] = "abnormal"};
 767         unsigned long pfn, first_pfn;
 768         int type, prev_type;
 769
 770         prev_type = xen_p2m_elem_type(0);
 771         first_pfn = 0;
 772
 773         for (pfn = 0; pfn < xen_p2m_size; pfn++) {
 774                 type = xen_p2m_elem_type(pfn);
 775                 if (type != prev_type) {
 776                         seq_printf(m, " [0x%lx->0x%lx] %s\n", first_pfn, pfn,
 777                                    type_name[prev_type]);
 778                         prev_type = type;
 779                         first_pfn = pfn;
 780                 }
 781         }
 782         seq_printf(m, " [0x%lx->0x%lx] %s\n", first_pfn, pfn,
 783                    type_name[prev_type]);
 784         return 0;
 785 }
 786
 787 static int p2m_dump_open(struct inode *inode, struct file *filp)
 788 {
 789         return single_open(filp, p2m_dump_show, NULL);
 790 }
 791
 792 static const struct file_operations p2m_dump_fops = {
 793         .open           = p2m_dump_open,
 794         .read           = seq_read,
 795         .llseek         = seq_lseek,
 796         .release        = single_release,
 797 };
 798
 799 static struct dentry *d_mmu_debug;
 800
 801 static int __init xen_p2m_debugfs(void)
 802 {
 803         struct dentry *d_xen = xen_init_debugfs();
 804
 805         if (d_xen == NULL)
 806                 return -ENOMEM;
 807
 808         d_mmu_debug = debugfs_create_dir("mmu", d_xen);
 809
 810         debugfs_create_file("p2m", 0600, d_mmu_debug, NULL, &p2m_dump_fops);
 811         return 0;
 812 }
 813 fs_initcall(xen_p2m_debugfs);
 814 #endif /* CONFIG_XEN_DEBUG_FS */