drivers/xen/privcmd.c

   1 // SPDX-License-Identifier: GPL-2.0-only
   2 /******************************************************************************
   3  * privcmd.c
   4  *
   5  * Interface to privileged domain-0 commands.
   6  *
   7  * Copyright (c) 2002-2004, K A Fraser, B Dragovic
   8  */
   9
  10 #define pr_fmt(fmt) "xen:" KBUILD_MODNAME ": " fmt
  11
  12 #include <linux/eventfd.h>
  13 #include <linux/file.h>
  14 #include <linux/kernel.h>
  15 #include <linux/module.h>
  16 #include <linux/mutex.h>
  17 #include <linux/poll.h>
  18 #include <linux/sched.h>
  19 #include <linux/slab.h>
  20 #include <linux/string.h>
  21 #include <linux/workqueue.h>
  22 #include <linux/errno.h>
  23 #include <linux/mm.h>
  24 #include <linux/mman.h>
  25 #include <linux/uaccess.h>
  26 #include <linux/swap.h>
  27 #include <linux/highmem.h>
  28 #include <linux/pagemap.h>
  29 #include <linux/seq_file.h>
  30 #include <linux/miscdevice.h>
  31 #include <linux/moduleparam.h>
  32 #include <linux/virtio_mmio.h>
  33
  34 #include <asm/xen/hypervisor.h>
  35 #include <asm/xen/hypercall.h>
  36
  37 #include <xen/xen.h>
  38 #include <xen/events.h>
  39 #include <xen/privcmd.h>
  40 #include <xen/interface/xen.h>
  41 #include <xen/interface/memory.h>
  42 #include <xen/interface/hvm/dm_op.h>
  43 #include <xen/interface/hvm/ioreq.h>
  44 #include <xen/features.h>
  45 #include <xen/page.h>
  46 #include <xen/xen-ops.h>
  47 #include <xen/balloon.h>
  48
  49 #include "privcmd.h"
  50
  51 MODULE_LICENSE("GPL");
  52
  53 #define PRIV_VMA_LOCKED ((void *)1)
  54
  55 static unsigned int privcmd_dm_op_max_num = 16;
  56 module_param_named(dm_op_max_nr_bufs, privcmd_dm_op_max_num, uint, 0644);
  57 MODULE_PARM_DESC(dm_op_max_nr_bufs,
  58                  "Maximum number of buffers per dm_op hypercall");
  59
  60 static unsigned int privcmd_dm_op_buf_max_size = 4096;
  61 module_param_named(dm_op_buf_max_size, privcmd_dm_op_buf_max_size, uint,
  62                    0644);
  63 MODULE_PARM_DESC(dm_op_buf_max_size,
  64                  "Maximum size of a dm_op hypercall buffer");
  65
  66 struct privcmd_data {
  67         domid_t domid;
  68 };
  69
  70 static int privcmd_vma_range_is_mapped(
  71                struct vm_area_struct *vma,
  72                unsigned long addr,
  73                unsigned long nr_pages);
  74
  75 static long privcmd_ioctl_hypercall(struct file *file, void __user *udata)
  76 {
  77         struct privcmd_data *data = file->private_data;
  78         struct privcmd_hypercall hypercall;
  79         long ret;
  80
  81         /* Disallow arbitrary hypercalls if restricted */
  82         if (data->domid != DOMID_INVALID)
  83                 return -EPERM;
  84
  85         if (copy_from_user(&hypercall, udata, sizeof(hypercall)))
  86                 return -EFAULT;
  87
  88         xen_preemptible_hcall_begin();
  89         ret = privcmd_call(hypercall.op,
  90                            hypercall.arg[0], hypercall.arg[1],
  91                            hypercall.arg[2], hypercall.arg[3],
  92                            hypercall.arg[4]);
  93         xen_preemptible_hcall_end();
  94
  95         return ret;
  96 }
  97
  98 static void free_page_list(struct list_head *pages)
  99 {
 100         struct page *p, *n;
 101
 102         list_for_each_entry_safe(p, n, pages, lru)
 103                 __free_page(p);
 104
 105         INIT_LIST_HEAD(pages);
 106 }
 107
 108 /*
 109  * Given an array of items in userspace, return a list of pages
 110  * containing the data.  If copying fails, either because of memory
 111  * allocation failure or a problem reading user memory, return an
 112  * error code; its up to the caller to dispose of any partial list.
 113  */
 114 static int gather_array(struct list_head *pagelist,
 115                         unsigned nelem, size_t size,
 116                         const void __user *data)
 117 {
 118         unsigned pageidx;
 119         void *pagedata;
 120         int ret;
 121
 122         if (size > PAGE_SIZE)
 123                 return 0;
 124
 125         pageidx = PAGE_SIZE;
 126         pagedata = NULL;        /* quiet, gcc */
 127         while (nelem--) {
 128                 if (pageidx > PAGE_SIZE-size) {
 129                         struct page *page = alloc_page(GFP_KERNEL);
 130
 131                         ret = -ENOMEM;
 132                         if (page == NULL)
 133                                 goto fail;
 134
 135                         pagedata = page_address(page);
 136
 137                         list_add_tail(&page->lru, pagelist);
 138                         pageidx = 0;
 139                 }
 140
 141                 ret = -EFAULT;
 142                 if (copy_from_user(pagedata + pageidx, data, size))
 143                         goto fail;
 144
 145                 data += size;
 146                 pageidx += size;
 147         }
 148
 149         ret = 0;
 150
 151 fail:
 152         return ret;
 153 }
 154
 155 /*
 156  * Call function "fn" on each element of the array fragmented
 157  * over a list of pages.
 158  */
 159 static int traverse_pages(unsigned nelem, size_t size,
 160                           struct list_head *pos,
 161                           int (*fn)(void *data, void *state),
 162                           void *state)
 163 {
 164         void *pagedata;
 165         unsigned pageidx;
 166         int ret = 0;
 167
 168         BUG_ON(size > PAGE_SIZE);
 169
 170         pageidx = PAGE_SIZE;
 171         pagedata = NULL;        /* hush, gcc */
 172
 173         while (nelem--) {
 174                 if (pageidx > PAGE_SIZE-size) {
 175                         struct page *page;
 176                         pos = pos->next;
 177                         page = list_entry(pos, struct page, lru);
 178                         pagedata = page_address(page);
 179                         pageidx = 0;
 180                 }
 181
 182                 ret = (*fn)(pagedata + pageidx, state);
 183                 if (ret)
 184                         break;
 185                 pageidx += size;
 186         }
 187
 188         return ret;
 189 }
 190
 191 /*
 192  * Similar to traverse_pages, but use each page as a "block" of
 193  * data to be processed as one unit.
 194  */
 195 static int traverse_pages_block(unsigned nelem, size_t size,
 196                                 struct list_head *pos,
 197                                 int (*fn)(void *data, int nr, void *state),
 198                                 void *state)
 199 {
 200         void *pagedata;
 201         int ret = 0;
 202
 203         BUG_ON(size > PAGE_SIZE);
 204
 205         while (nelem) {
 206                 int nr = (PAGE_SIZE/size);
 207                 struct page *page;
 208                 if (nr > nelem)
 209                         nr = nelem;
 210                 pos = pos->next;
 211                 page = list_entry(pos, struct page, lru);
 212                 pagedata = page_address(page);
 213                 ret = (*fn)(pagedata, nr, state);
 214                 if (ret)
 215                         break;
 216                 nelem -= nr;
 217         }
 218
 219         return ret;
 220 }
 221
 222 struct mmap_gfn_state {
 223         unsigned long va;
 224         struct vm_area_struct *vma;
 225         domid_t domain;
 226 };
 227
 228 static int mmap_gfn_range(void *data, void *state)
 229 {
 230         struct privcmd_mmap_entry *msg = data;
 231         struct mmap_gfn_state *st = state;
 232         struct vm_area_struct *vma = st->vma;
 233         int rc;
 234
 235         /* Do not allow range to wrap the address space. */
 236         if ((msg->npages > (LONG_MAX >> PAGE_SHIFT)) ||
 237             ((unsigned long)(msg->npages << PAGE_SHIFT) >= -st->va))
 238                 return -EINVAL;
 239
 240         /* Range chunks must be contiguous in va space. */
 241         if ((msg->va != st->va) ||
 242             ((msg->va+(msg->npages<<PAGE_SHIFT)) > vma->vm_end))
 243                 return -EINVAL;
 244
 245         rc = xen_remap_domain_gfn_range(vma,
 246                                         msg->va & PAGE_MASK,
 247                                         msg->mfn, msg->npages,
 248                                         vma->vm_page_prot,
 249                                         st->domain, NULL);
 250         if (rc < 0)
 251                 return rc;
 252
 253         st->va += msg->npages << PAGE_SHIFT;
 254
 255         return 0;
 256 }
 257
 258 static long privcmd_ioctl_mmap(struct file *file, void __user *udata)
 259 {
 260         struct privcmd_data *data = file->private_data;
 261         struct privcmd_mmap mmapcmd;
 262         struct mm_struct *mm = current->mm;
 263         struct vm_area_struct *vma;
 264         int rc;
 265         LIST_HEAD(pagelist);
 266         struct mmap_gfn_state state;
 267
 268         /* We only support privcmd_ioctl_mmap_batch for non-auto-translated. */
 269         if (xen_feature(XENFEAT_auto_translated_physmap))
 270                 return -ENOSYS;
 271
 272         if (copy_from_user(&mmapcmd, udata, sizeof(mmapcmd)))
 273                 return -EFAULT;
 274
 275         /* If restriction is in place, check the domid matches */
 276         if (data->domid != DOMID_INVALID && data->domid != mmapcmd.dom)
 277                 return -EPERM;
 278
 279         rc = gather_array(&pagelist,
 280                           mmapcmd.num, sizeof(struct privcmd_mmap_entry),
 281                           mmapcmd.entry);
 282
 283         if (rc || list_empty(&pagelist))
 284                 goto out;
 285
 286         mmap_write_lock(mm);
 287
 288         {
 289                 struct page *page = list_first_entry(&pagelist,
 290                                                      struct page, lru);
 291                 struct privcmd_mmap_entry *msg = page_address(page);
 292
 293                 vma = vma_lookup(mm, msg->va);
 294                 rc = -EINVAL;
 295
 296                 if (!vma || (msg->va != vma->vm_start) || vma->vm_private_data)
 297                         goto out_up;
 298                 vma->vm_private_data = PRIV_VMA_LOCKED;
 299         }
 300
 301         state.va = vma->vm_start;
 302         state.vma = vma;
 303         state.domain = mmapcmd.dom;
 304
 305         rc = traverse_pages(mmapcmd.num, sizeof(struct privcmd_mmap_entry),
 306                             &pagelist,
 307                             mmap_gfn_range, &state);
 308
 309
 310 out_up:
 311         mmap_write_unlock(mm);
 312
 313 out:
 314         free_page_list(&pagelist);
 315
 316         return rc;
 317 }
 318
 319 struct mmap_batch_state {
 320         domid_t domain;
 321         unsigned long va;
 322         struct vm_area_struct *vma;
 323         int index;
 324         /* A tristate:
 325          *      0 for no errors
 326          *      1 if at least one error has happened (and no
 327          *          -ENOENT errors have happened)
 328          *      -ENOENT if at least 1 -ENOENT has happened.
 329          */
 330         int global_error;
 331         int version;
 332
 333         /* User-space gfn array to store errors in the second pass for V1. */
 334         xen_pfn_t __user *user_gfn;
 335         /* User-space int array to store errors in the second pass for V2. */
 336         int __user *user_err;
 337 };
 338
 339 /* auto translated dom0 note: if domU being created is PV, then gfn is
 340  * mfn(addr on bus). If it's auto xlated, then gfn is pfn (input to HAP).
 341  */
 342 static int mmap_batch_fn(void *data, int nr, void *state)
 343 {
 344         xen_pfn_t *gfnp = data;
 345         struct mmap_batch_state *st = state;
 346         struct vm_area_struct *vma = st->vma;
 347         struct page **pages = vma->vm_private_data;
 348         struct page **cur_pages = NULL;
 349         int ret;
 350
 351         if (xen_feature(XENFEAT_auto_translated_physmap))
 352                 cur_pages = &pages[st->index];
 353
 354         BUG_ON(nr < 0);
 355         ret = xen_remap_domain_gfn_array(st->vma, st->va & PAGE_MASK, gfnp, nr,
 356                                          (int *)gfnp, st->vma->vm_page_prot,
 357                                          st->domain, cur_pages);
 358
 359         /* Adjust the global_error? */
 360         if (ret != nr) {
 361                 if (ret == -ENOENT)
 362                         st->global_error = -ENOENT;
 363                 else {
 364                         /* Record that at least one error has happened. */
 365                         if (st->global_error == 0)
 366                                 st->global_error = 1;
 367                 }
 368         }
 369         st->va += XEN_PAGE_SIZE * nr;
 370         st->index += nr / XEN_PFN_PER_PAGE;
 371
 372         return 0;
 373 }
 374
 375 static int mmap_return_error(int err, struct mmap_batch_state *st)
 376 {
 377         int ret;
 378
 379         if (st->version == 1) {
 380                 if (err) {
 381                         xen_pfn_t gfn;
 382
 383                         ret = get_user(gfn, st->user_gfn);
 384                         if (ret < 0)
 385                                 return ret;
 386                         /*
 387                          * V1 encodes the error codes in the 32bit top
 388                          * nibble of the gfn (with its known
 389                          * limitations vis-a-vis 64 bit callers).
 390                          */
 391                         gfn |= (err == -ENOENT) ?
 392                                 PRIVCMD_MMAPBATCH_PAGED_ERROR :
 393                                 PRIVCMD_MMAPBATCH_MFN_ERROR;
 394                         return __put_user(gfn, st->user_gfn++);
 395                 } else
 396                         st->user_gfn++;
 397         } else { /* st->version == 2 */
 398                 if (err)
 399                         return __put_user(err, st->user_err++);
 400                 else
 401                         st->user_err++;
 402         }
 403
 404         return 0;
 405 }
 406
 407 static int mmap_return_errors(void *data, int nr, void *state)
 408 {
 409         struct mmap_batch_state *st = state;
 410         int *errs = data;
 411         int i;
 412         int ret;
 413
 414         for (i = 0; i < nr; i++) {
 415                 ret = mmap_return_error(errs[i], st);
 416                 if (ret < 0)
 417                         return ret;
 418         }
 419         return 0;
 420 }
 421
 422 /* Allocate pfns that are then mapped with gfns from foreign domid. Update
 423  * the vma with the page info to use later.
 424  * Returns: 0 if success, otherwise -errno
 425  */
 426 static int alloc_empty_pages(struct vm_area_struct *vma, int numpgs)
 427 {
 428         int rc;
 429         struct page **pages;
 430
 431         pages = kvcalloc(numpgs, sizeof(pages[0]), GFP_KERNEL);
 432         if (pages == NULL)
 433                 return -ENOMEM;
 434
 435         rc = xen_alloc_unpopulated_pages(numpgs, pages);
 436         if (rc != 0) {
 437                 pr_warn("%s Could not alloc %d pfns rc:%d\n", __func__,
 438                         numpgs, rc);
 439                 kvfree(pages);
 440                 return -ENOMEM;
 441         }
 442         BUG_ON(vma->vm_private_data != NULL);
 443         vma->vm_private_data = pages;
 444
 445         return 0;
 446 }
 447
 448 static const struct vm_operations_struct privcmd_vm_ops;
 449
 450 static long privcmd_ioctl_mmap_batch(
 451         struct file *file, void __user *udata, int version)
 452 {
 453         struct privcmd_data *data = file->private_data;
 454         int ret;
 455         struct privcmd_mmapbatch_v2 m;
 456         struct mm_struct *mm = current->mm;
 457         struct vm_area_struct *vma;
 458         unsigned long nr_pages;
 459         LIST_HEAD(pagelist);
 460         struct mmap_batch_state state;
 461
 462         switch (version) {
 463         case 1:
 464                 if (copy_from_user(&m, udata, sizeof(struct privcmd_mmapbatch)))
 465                         return -EFAULT;
 466                 /* Returns per-frame error in m.arr. */
 467                 m.err = NULL;
 468                 if (!access_ok(m.arr, m.num * sizeof(*m.arr)))
 469                         return -EFAULT;
 470                 break;
 471         case 2:
 472                 if (copy_from_user(&m, udata, sizeof(struct privcmd_mmapbatch_v2)))
 473                         return -EFAULT;
 474                 /* Returns per-frame error code in m.err. */
 475                 if (!access_ok(m.err, m.num * (sizeof(*m.err))))
 476                         return -EFAULT;
 477                 break;
 478         default:
 479                 return -EINVAL;
 480         }
 481
 482         /* If restriction is in place, check the domid matches */
 483         if (data->domid != DOMID_INVALID && data->domid != m.dom)
 484                 return -EPERM;
 485
 486         nr_pages = DIV_ROUND_UP(m.num, XEN_PFN_PER_PAGE);
 487         if ((m.num <= 0) || (nr_pages > (LONG_MAX >> PAGE_SHIFT)))
 488                 return -EINVAL;
 489
 490         ret = gather_array(&pagelist, m.num, sizeof(xen_pfn_t), m.arr);
 491
 492         if (ret)
 493                 goto out;
 494         if (list_empty(&pagelist)) {
 495                 ret = -EINVAL;
 496                 goto out;
 497         }
 498
 499         if (version == 2) {
 500                 /* Zero error array now to only copy back actual errors. */
 501                 if (clear_user(m.err, sizeof(int) * m.num)) {
 502                         ret = -EFAULT;
 503                         goto out;
 504                 }
 505         }
 506
 507         mmap_write_lock(mm);
 508
 509         vma = find_vma(mm, m.addr);
 510         if (!vma ||
 511             vma->vm_ops != &privcmd_vm_ops) {
 512                 ret = -EINVAL;
 513                 goto out_unlock;
 514         }
 515
 516         /*
 517          * Caller must either:
 518          *
 519          * Map the whole VMA range, which will also allocate all the
 520          * pages required for the auto_translated_physmap case.
 521          *
 522          * Or
 523          *
 524          * Map unmapped holes left from a previous map attempt (e.g.,
 525          * because those foreign frames were previously paged out).
 526          */
 527         if (vma->vm_private_data == NULL) {
 528                 if (m.addr != vma->vm_start ||
 529                     m.addr + (nr_pages << PAGE_SHIFT) != vma->vm_end) {
 530                         ret = -EINVAL;
 531                         goto out_unlock;
 532                 }
 533                 if (xen_feature(XENFEAT_auto_translated_physmap)) {
 534                         ret = alloc_empty_pages(vma, nr_pages);
 535                         if (ret < 0)
 536                                 goto out_unlock;
 537                 } else
 538                         vma->vm_private_data = PRIV_VMA_LOCKED;
 539         } else {
 540                 if (m.addr < vma->vm_start ||
 541                     m.addr + (nr_pages << PAGE_SHIFT) > vma->vm_end) {
 542                         ret = -EINVAL;
 543                         goto out_unlock;
 544                 }
 545                 if (privcmd_vma_range_is_mapped(vma, m.addr, nr_pages)) {
 546                         ret = -EINVAL;
 547                         goto out_unlock;
 548                 }
 549         }
 550
 551         state.domain        = m.dom;
 552         state.vma           = vma;
 553         state.va            = m.addr;
 554         state.index         = 0;
 555         state.global_error  = 0;
 556         state.version       = version;
 557
 558         BUILD_BUG_ON(((PAGE_SIZE / sizeof(xen_pfn_t)) % XEN_PFN_PER_PAGE) != 0);
 559         /* mmap_batch_fn guarantees ret == 0 */
 560         BUG_ON(traverse_pages_block(m.num, sizeof(xen_pfn_t),
 561                                     &pagelist, mmap_batch_fn, &state));
 562
 563         mmap_write_unlock(mm);
 564
 565         if (state.global_error) {
 566                 /* Write back errors in second pass. */
 567                 state.user_gfn = (xen_pfn_t *)m.arr;
 568                 state.user_err = m.err;
 569                 ret = traverse_pages_block(m.num, sizeof(xen_pfn_t),
 570                                            &pagelist, mmap_return_errors, &state);
 571         } else
 572                 ret = 0;
 573
 574         /* If we have not had any EFAULT-like global errors then set the global
 575          * error to -ENOENT if necessary. */
 576         if ((ret == 0) && (state.global_error == -ENOENT))
 577                 ret = -ENOENT;
 578
 579 out:
 580         free_page_list(&pagelist);
 581         return ret;
 582
 583 out_unlock:
 584         mmap_write_unlock(mm);
 585         goto out;
 586 }
 587
 588 static int lock_pages(
 589         struct privcmd_dm_op_buf kbufs[], unsigned int num,
 590         struct page *pages[], unsigned int nr_pages, unsigned int *pinned)
 591 {
 592         unsigned int i, off = 0;
 593
 594         for (i = 0; i < num; ) {
 595                 unsigned int requested;
 596                 int page_count;
 597
 598                 requested = DIV_ROUND_UP(
 599                         offset_in_page(kbufs[i].uptr) + kbufs[i].size,
 600                         PAGE_SIZE) - off;
 601                 if (requested > nr_pages)
 602                         return -ENOSPC;
 603
 604                 page_count = pin_user_pages_fast(
 605                         (unsigned long)kbufs[i].uptr + off * PAGE_SIZE,
 606                         requested, FOLL_WRITE, pages);
 607                 if (page_count <= 0)
 608                         return page_count ? : -EFAULT;
 609
 610                 *pinned += page_count;
 611                 nr_pages -= page_count;
 612                 pages += page_count;
 613
 614                 off = (requested == page_count) ? 0 : off + page_count;
 615                 i += !off;
 616         }
 617
 618         return 0;
 619 }
 620
 621 static void unlock_pages(struct page *pages[], unsigned int nr_pages)
 622 {
 623         unpin_user_pages_dirty_lock(pages, nr_pages, true);
 624 }
 625
 626 static long privcmd_ioctl_dm_op(struct file *file, void __user *udata)
 627 {
 628         struct privcmd_data *data = file->private_data;
 629         struct privcmd_dm_op kdata;
 630         struct privcmd_dm_op_buf *kbufs;
 631         unsigned int nr_pages = 0;
 632         struct page **pages = NULL;
 633         struct xen_dm_op_buf *xbufs = NULL;
 634         unsigned int i;
 635         long rc;
 636         unsigned int pinned = 0;
 637
 638         if (copy_from_user(&kdata, udata, sizeof(kdata)))
 639                 return -EFAULT;
 640
 641         /* If restriction is in place, check the domid matches */
 642         if (data->domid != DOMID_INVALID && data->domid != kdata.dom)
 643                 return -EPERM;
 644
 645         if (kdata.num == 0)
 646                 return 0;
 647
 648         if (kdata.num > privcmd_dm_op_max_num)
 649                 return -E2BIG;
 650
 651         kbufs = kcalloc(kdata.num, sizeof(*kbufs), GFP_KERNEL);
 652         if (!kbufs)
 653                 return -ENOMEM;
 654
 655         if (copy_from_user(kbufs, kdata.ubufs,
 656                            sizeof(*kbufs) * kdata.num)) {
 657                 rc = -EFAULT;
 658                 goto out;
 659         }
 660
 661         for (i = 0; i < kdata.num; i++) {
 662                 if (kbufs[i].size > privcmd_dm_op_buf_max_size) {
 663                         rc = -E2BIG;
 664                         goto out;
 665                 }
 666
 667                 if (!access_ok(kbufs[i].uptr,
 668                                kbufs[i].size)) {
 669                         rc = -EFAULT;
 670                         goto out;
 671                 }
 672
 673                 nr_pages += DIV_ROUND_UP(
 674                         offset_in_page(kbufs[i].uptr) + kbufs[i].size,
 675                         PAGE_SIZE);
 676         }
 677
 678         pages = kcalloc(nr_pages, sizeof(*pages), GFP_KERNEL);
 679         if (!pages) {
 680                 rc = -ENOMEM;
 681                 goto out;
 682         }
 683
 684         xbufs = kcalloc(kdata.num, sizeof(*xbufs), GFP_KERNEL);
 685         if (!xbufs) {
 686                 rc = -ENOMEM;
 687                 goto out;
 688         }
 689
 690         rc = lock_pages(kbufs, kdata.num, pages, nr_pages, &pinned);
 691         if (rc < 0)
 692                 goto out;
 693
 694         for (i = 0; i < kdata.num; i++) {
 695                 set_xen_guest_handle(xbufs[i].h, kbufs[i].uptr);
 696                 xbufs[i].size = kbufs[i].size;
 697         }
 698
 699         xen_preemptible_hcall_begin();
 700         rc = HYPERVISOR_dm_op(kdata.dom, kdata.num, xbufs);
 701         xen_preemptible_hcall_end();
 702
 703 out:
 704         unlock_pages(pages, pinned);
 705         kfree(xbufs);
 706         kfree(pages);
 707         kfree(kbufs);
 708
 709         return rc;
 710 }
 711
 712 static long privcmd_ioctl_restrict(struct file *file, void __user *udata)
 713 {
 714         struct privcmd_data *data = file->private_data;
 715         domid_t dom;
 716
 717         if (copy_from_user(&dom, udata, sizeof(dom)))
 718                 return -EFAULT;
 719
 720         /* Set restriction to the specified domain, or check it matches */
 721         if (data->domid == DOMID_INVALID)
 722                 data->domid = dom;
 723         else if (data->domid != dom)
 724                 return -EINVAL;
 725
 726         return 0;
 727 }
 728
 729 static long privcmd_ioctl_mmap_resource(struct file *file,
 730                                 struct privcmd_mmap_resource __user *udata)
 731 {
 732         struct privcmd_data *data = file->private_data;
 733         struct mm_struct *mm = current->mm;
 734         struct vm_area_struct *vma;
 735         struct privcmd_mmap_resource kdata;
 736         xen_pfn_t *pfns = NULL;
 737         struct xen_mem_acquire_resource xdata = { };
 738         int rc;
 739
 740         if (copy_from_user(&kdata, udata, sizeof(kdata)))
 741                 return -EFAULT;
 742
 743         /* If restriction is in place, check the domid matches */
 744         if (data->domid != DOMID_INVALID && data->domid != kdata.dom)
 745                 return -EPERM;
 746
 747         /* Both fields must be set or unset */
 748         if (!!kdata.addr != !!kdata.num)
 749                 return -EINVAL;
 750
 751         xdata.domid = kdata.dom;
 752         xdata.type = kdata.type;
 753         xdata.id = kdata.id;
 754
 755         if (!kdata.addr && !kdata.num) {
 756                 /* Query the size of the resource. */
 757                 rc = HYPERVISOR_memory_op(XENMEM_acquire_resource, &xdata);
 758                 if (rc)
 759                         return rc;
 760                 return __put_user(xdata.nr_frames, &udata->num);
 761         }
 762
 763         mmap_write_lock(mm);
 764
 765         vma = find_vma(mm, kdata.addr);
 766         if (!vma || vma->vm_ops != &privcmd_vm_ops) {
 767                 rc = -EINVAL;
 768                 goto out;
 769         }
 770
 771         pfns = kcalloc(kdata.num, sizeof(*pfns), GFP_KERNEL | __GFP_NOWARN);
 772         if (!pfns) {
 773                 rc = -ENOMEM;
 774                 goto out;
 775         }
 776
 777         if (IS_ENABLED(CONFIG_XEN_AUTO_XLATE) &&
 778             xen_feature(XENFEAT_auto_translated_physmap)) {
 779                 unsigned int nr = DIV_ROUND_UP(kdata.num, XEN_PFN_PER_PAGE);
 780                 struct page **pages;
 781                 unsigned int i;
 782
 783                 rc = alloc_empty_pages(vma, nr);
 784                 if (rc < 0)
 785                         goto out;
 786
 787                 pages = vma->vm_private_data;
 788
 789                 for (i = 0; i < kdata.num; i++) {
 790                         xen_pfn_t pfn =
 791                                 page_to_xen_pfn(pages[i / XEN_PFN_PER_PAGE]);
 792
 793                         pfns[i] = pfn + (i % XEN_PFN_PER_PAGE);
 794                 }
 795         } else
 796                 vma->vm_private_data = PRIV_VMA_LOCKED;
 797
 798         xdata.frame = kdata.idx;
 799         xdata.nr_frames = kdata.num;
 800         set_xen_guest_handle(xdata.frame_list, pfns);
 801
 802         xen_preemptible_hcall_begin();
 803         rc = HYPERVISOR_memory_op(XENMEM_acquire_resource, &xdata);
 804         xen_preemptible_hcall_end();
 805
 806         if (rc)
 807                 goto out;
 808
 809         if (IS_ENABLED(CONFIG_XEN_AUTO_XLATE) &&
 810             xen_feature(XENFEAT_auto_translated_physmap)) {
 811                 rc = xen_remap_vma_range(vma, kdata.addr, kdata.num << PAGE_SHIFT);
 812         } else {
 813                 unsigned int domid =
 814                         (xdata.flags & XENMEM_rsrc_acq_caller_owned) ?
 815                         DOMID_SELF : kdata.dom;
 816                 int num, *errs = (int *)pfns;
 817
 818                 BUILD_BUG_ON(sizeof(*errs) > sizeof(*pfns));
 819                 num = xen_remap_domain_mfn_array(vma,
 820                                                  kdata.addr & PAGE_MASK,
 821                                                  pfns, kdata.num, errs,
 822                                                  vma->vm_page_prot,
 823                                                  domid);
 824                 if (num < 0)
 825                         rc = num;
 826                 else if (num != kdata.num) {
 827                         unsigned int i;
 828
 829                         for (i = 0; i < num; i++) {
 830                                 rc = errs[i];
 831                                 if (rc < 0)
 832                                         break;
 833                         }
 834                 } else
 835                         rc = 0;
 836         }
 837
 838 out:
 839         mmap_write_unlock(mm);
 840         kfree(pfns);
 841
 842         return rc;
 843 }
 844
 845 #ifdef CONFIG_XEN_PRIVCMD_EVENTFD
 846 /* Irqfd support */
 847 static struct workqueue_struct *irqfd_cleanup_wq;
 848 static DEFINE_MUTEX(irqfds_lock);
 849 static LIST_HEAD(irqfds_list);
 850
 851 struct privcmd_kernel_irqfd {
 852         struct xen_dm_op_buf xbufs;
 853         domid_t dom;
 854         bool error;
 855         struct eventfd_ctx *eventfd;
 856         struct work_struct shutdown;
 857         wait_queue_entry_t wait;
 858         struct list_head list;
 859         poll_table pt;
 860 };
 861
 862 static void irqfd_deactivate(struct privcmd_kernel_irqfd *kirqfd)
 863 {
 864         lockdep_assert_held(&irqfds_lock);
 865
 866         list_del_init(&kirqfd->list);
 867         queue_work(irqfd_cleanup_wq, &kirqfd->shutdown);
 868 }
 869
 870 static void irqfd_shutdown(struct work_struct *work)
 871 {
 872         struct privcmd_kernel_irqfd *kirqfd =
 873                 container_of(work, struct privcmd_kernel_irqfd, shutdown);
 874         u64 cnt;
 875
 876         eventfd_ctx_remove_wait_queue(kirqfd->eventfd, &kirqfd->wait, &cnt);
 877         eventfd_ctx_put(kirqfd->eventfd);
 878         kfree(kirqfd);
 879 }
 880
 881 static void irqfd_inject(struct privcmd_kernel_irqfd *kirqfd)
 882 {
 883         u64 cnt;
 884         long rc;
 885
 886         eventfd_ctx_do_read(kirqfd->eventfd, &cnt);
 887
 888         xen_preemptible_hcall_begin();
 889         rc = HYPERVISOR_dm_op(kirqfd->dom, 1, &kirqfd->xbufs);
 890         xen_preemptible_hcall_end();
 891
 892         /* Don't repeat the error message for consecutive failures */
 893         if (rc && !kirqfd->error) {
 894                 pr_err("Failed to configure irq for guest domain: %d\n",
 895                        kirqfd->dom);
 896         }
 897
 898         kirqfd->error = rc;
 899 }
 900
 901 static int
 902 irqfd_wakeup(wait_queue_entry_t *wait, unsigned int mode, int sync, void *key)
 903 {
 904         struct privcmd_kernel_irqfd *kirqfd =
 905                 container_of(wait, struct privcmd_kernel_irqfd, wait);
 906         __poll_t flags = key_to_poll(key);
 907
 908         if (flags & EPOLLIN)
 909                 irqfd_inject(kirqfd);
 910
 911         if (flags & EPOLLHUP) {
 912                 mutex_lock(&irqfds_lock);
 913                 irqfd_deactivate(kirqfd);
 914                 mutex_unlock(&irqfds_lock);
 915         }
 916
 917         return 0;
 918 }
 919
 920 static void
 921 irqfd_poll_func(struct file *file, wait_queue_head_t *wqh, poll_table *pt)
 922 {
 923         struct privcmd_kernel_irqfd *kirqfd =
 924                 container_of(pt, struct privcmd_kernel_irqfd, pt);
 925
 926         add_wait_queue_priority(wqh, &kirqfd->wait);
 927 }
 928
 929 static int privcmd_irqfd_assign(struct privcmd_irqfd *irqfd)
 930 {
 931         struct privcmd_kernel_irqfd *kirqfd, *tmp;
 932         __poll_t events;
 933         struct fd f;
 934         void *dm_op;
 935         int ret;
 936
 937         kirqfd = kzalloc(sizeof(*kirqfd) + irqfd->size, GFP_KERNEL);
 938         if (!kirqfd)
 939                 return -ENOMEM;
 940         dm_op = kirqfd + 1;
 941
 942         if (copy_from_user(dm_op, u64_to_user_ptr(irqfd->dm_op), irqfd->size)) {
 943                 ret = -EFAULT;
 944                 goto error_kfree;
 945         }
 946
 947         kirqfd->xbufs.size = irqfd->size;
 948         set_xen_guest_handle(kirqfd->xbufs.h, dm_op);
 949         kirqfd->dom = irqfd->dom;
 950         INIT_WORK(&kirqfd->shutdown, irqfd_shutdown);
 951
 952         f = fdget(irqfd->fd);
 953         if (!f.file) {
 954                 ret = -EBADF;
 955                 goto error_kfree;
 956         }
 957
 958         kirqfd->eventfd = eventfd_ctx_fileget(f.file);
 959         if (IS_ERR(kirqfd->eventfd)) {
 960                 ret = PTR_ERR(kirqfd->eventfd);
 961                 goto error_fd_put;
 962         }
 963
 964         /*
 965          * Install our own custom wake-up handling so we are notified via a
 966          * callback whenever someone signals the underlying eventfd.
 967          */
 968         init_waitqueue_func_entry(&kirqfd->wait, irqfd_wakeup);
 969         init_poll_funcptr(&kirqfd->pt, irqfd_poll_func);
 970
 971         mutex_lock(&irqfds_lock);
 972
 973         list_for_each_entry(tmp, &irqfds_list, list) {
 974                 if (kirqfd->eventfd == tmp->eventfd) {
 975                         ret = -EBUSY;
 976                         mutex_unlock(&irqfds_lock);
 977                         goto error_eventfd;
 978                 }
 979         }
 980
 981         list_add_tail(&kirqfd->list, &irqfds_list);
 982         mutex_unlock(&irqfds_lock);
 983
 984         /*
 985          * Check if there was an event already pending on the eventfd before we
 986          * registered, and trigger it as if we didn't miss it.
 987          */
 988         events = vfs_poll(f.file, &kirqfd->pt);
 989         if (events & EPOLLIN)
 990                 irqfd_inject(kirqfd);
 991
 992         /*
 993          * Do not drop the file until the kirqfd is fully initialized, otherwise
 994          * we might race against the EPOLLHUP.
 995          */
 996         fdput(f);
 997         return 0;
 998
 999 error_eventfd:
1000         eventfd_ctx_put(kirqfd->eventfd);
1001
1002 error_fd_put:
1003         fdput(f);
1004
1005 error_kfree:
1006         kfree(kirqfd);
1007         return ret;
1008 }
1009
1010 static int privcmd_irqfd_deassign(struct privcmd_irqfd *irqfd)
1011 {
1012         struct privcmd_kernel_irqfd *kirqfd;
1013         struct eventfd_ctx *eventfd;
1014
1015         eventfd = eventfd_ctx_fdget(irqfd->fd);
1016         if (IS_ERR(eventfd))
1017                 return PTR_ERR(eventfd);
1018
1019         mutex_lock(&irqfds_lock);
1020
1021         list_for_each_entry(kirqfd, &irqfds_list, list) {
1022                 if (kirqfd->eventfd == eventfd) {
1023                         irqfd_deactivate(kirqfd);
1024                         break;
1025                 }
1026         }
1027
1028         mutex_unlock(&irqfds_lock);
1029
1030         eventfd_ctx_put(eventfd);
1031
1032         /*
1033          * Block until we know all outstanding shutdown jobs have completed so
1034          * that we guarantee there will not be any more interrupts once this
1035          * deassign function returns.
1036          */
1037         flush_workqueue(irqfd_cleanup_wq);
1038
1039         return 0;
1040 }
1041
1042 static long privcmd_ioctl_irqfd(struct file *file, void __user *udata)
1043 {
1044         struct privcmd_data *data = file->private_data;
1045         struct privcmd_irqfd irqfd;
1046
1047         if (copy_from_user(&irqfd, udata, sizeof(irqfd)))
1048                 return -EFAULT;
1049
1050         /* No other flags should be set */
1051         if (irqfd.flags & ~PRIVCMD_IRQFD_FLAG_DEASSIGN)
1052                 return -EINVAL;
1053
1054         /* If restriction is in place, check the domid matches */
1055         if (data->domid != DOMID_INVALID && data->domid != irqfd.dom)
1056                 return -EPERM;
1057
1058         if (irqfd.flags & PRIVCMD_IRQFD_FLAG_DEASSIGN)
1059                 return privcmd_irqfd_deassign(&irqfd);
1060
1061         return privcmd_irqfd_assign(&irqfd);
1062 }
1063
1064 static int privcmd_irqfd_init(void)
1065 {
1066         irqfd_cleanup_wq = alloc_workqueue("privcmd-irqfd-cleanup", 0, 0);
1067         if (!irqfd_cleanup_wq)
1068                 return -ENOMEM;
1069
1070         return 0;
1071 }
1072
1073 static void privcmd_irqfd_exit(void)
1074 {
1075         struct privcmd_kernel_irqfd *kirqfd, *tmp;
1076
1077         mutex_lock(&irqfds_lock);
1078
1079         list_for_each_entry_safe(kirqfd, tmp, &irqfds_list, list)
1080                 irqfd_deactivate(kirqfd);
1081
1082         mutex_unlock(&irqfds_lock);
1083
1084         destroy_workqueue(irqfd_cleanup_wq);
1085 }
1086
1087 /* Ioeventfd Support */
1088 #define QUEUE_NOTIFY_VQ_MASK 0xFFFF
1089
1090 static DEFINE_MUTEX(ioreq_lock);
1091 static LIST_HEAD(ioreq_list);
1092
1093 /* per-eventfd structure */
1094 struct privcmd_kernel_ioeventfd {
1095         struct eventfd_ctx *eventfd;
1096         struct list_head list;
1097         u64 addr;
1098         unsigned int addr_len;
1099         unsigned int vq;
1100 };
1101
1102 /* per-guest CPU / port structure */
1103 struct ioreq_port {
1104         int vcpu;
1105         unsigned int port;
1106         struct privcmd_kernel_ioreq *kioreq;
1107 };
1108
1109 /* per-guest structure */
1110 struct privcmd_kernel_ioreq {
1111         domid_t dom;
1112         unsigned int vcpus;
1113         u64 uioreq;
1114         struct ioreq *ioreq;
1115         spinlock_t lock; /* Protects ioeventfds list */
1116         struct list_head ioeventfds;
1117         struct list_head list;
1118         struct ioreq_port ports[] __counted_by(vcpus);
1119 };
1120
1121 static irqreturn_t ioeventfd_interrupt(int irq, void *dev_id)
1122 {
1123         struct ioreq_port *port = dev_id;
1124         struct privcmd_kernel_ioreq *kioreq = port->kioreq;
1125         struct ioreq *ioreq = &kioreq->ioreq[port->vcpu];
1126         struct privcmd_kernel_ioeventfd *kioeventfd;
1127         unsigned int state = STATE_IOREQ_READY;
1128
1129         if (ioreq->state != STATE_IOREQ_READY ||
1130             ioreq->type != IOREQ_TYPE_COPY || ioreq->dir != IOREQ_WRITE)
1131                 return IRQ_NONE;
1132
1133         /*
1134          * We need a barrier, smp_mb(), here to ensure reads are finished before
1135          * `state` is updated. Since the lock implementation ensures that
1136          * appropriate barrier will be added anyway, we can avoid adding
1137          * explicit barrier here.
1138          *
1139          * Ideally we don't need to update `state` within the locks, but we do
1140          * that here to avoid adding explicit barrier.
1141          */
1142
1143         spin_lock(&kioreq->lock);
1144         ioreq->state = STATE_IOREQ_INPROCESS;
1145
1146         list_for_each_entry(kioeventfd, &kioreq->ioeventfds, list) {
1147                 if (ioreq->addr == kioeventfd->addr + VIRTIO_MMIO_QUEUE_NOTIFY &&
1148                     ioreq->size == kioeventfd->addr_len &&
1149                     (ioreq->data & QUEUE_NOTIFY_VQ_MASK) == kioeventfd->vq) {
1150                         eventfd_signal(kioeventfd->eventfd);
1151                         state = STATE_IORESP_READY;
1152                         break;
1153                 }
1154         }
1155         spin_unlock(&kioreq->lock);
1156
1157         /*
1158          * We need a barrier, smp_mb(), here to ensure writes are finished
1159          * before `state` is updated. Since the lock implementation ensures that
1160          * appropriate barrier will be added anyway, we can avoid adding
1161          * explicit barrier here.
1162          */
1163
1164         ioreq->state = state;
1165
1166         if (state == STATE_IORESP_READY) {
1167                 notify_remote_via_evtchn(port->port);
1168                 return IRQ_HANDLED;
1169         }
1170
1171         return IRQ_NONE;
1172 }
1173
1174 static void ioreq_free(struct privcmd_kernel_ioreq *kioreq)
1175 {
1176         struct ioreq_port *ports = kioreq->ports;
1177         int i;
1178
1179         lockdep_assert_held(&ioreq_lock);
1180
1181         list_del(&kioreq->list);
1182
1183         for (i = kioreq->vcpus - 1; i >= 0; i--)
1184                 unbind_from_irqhandler(irq_from_evtchn(ports[i].port), &ports[i]);
1185
1186         kfree(kioreq);
1187 }
1188
1189 static
1190 struct privcmd_kernel_ioreq *alloc_ioreq(struct privcmd_ioeventfd *ioeventfd)
1191 {
1192         struct privcmd_kernel_ioreq *kioreq;
1193         struct mm_struct *mm = current->mm;
1194         struct vm_area_struct *vma;
1195         struct page **pages;
1196         unsigned int *ports;
1197         int ret, size, i;
1198
1199         lockdep_assert_held(&ioreq_lock);
1200
1201         size = struct_size(kioreq, ports, ioeventfd->vcpus);
1202         kioreq = kzalloc(size, GFP_KERNEL);
1203         if (!kioreq)
1204                 return ERR_PTR(-ENOMEM);
1205
1206         kioreq->dom = ioeventfd->dom;
1207         kioreq->vcpus = ioeventfd->vcpus;
1208         kioreq->uioreq = ioeventfd->ioreq;
1209         spin_lock_init(&kioreq->lock);
1210         INIT_LIST_HEAD(&kioreq->ioeventfds);
1211
1212         /* The memory for ioreq server must have been mapped earlier */
1213         mmap_write_lock(mm);
1214         vma = find_vma(mm, (unsigned long)ioeventfd->ioreq);
1215         if (!vma) {
1216                 pr_err("Failed to find vma for ioreq page!\n");
1217                 mmap_write_unlock(mm);
1218                 ret = -EFAULT;
1219                 goto error_kfree;
1220         }
1221
1222         pages = vma->vm_private_data;
1223         kioreq->ioreq = (struct ioreq *)(page_to_virt(pages[0]));
1224         mmap_write_unlock(mm);
1225
1226         ports = memdup_array_user(u64_to_user_ptr(ioeventfd->ports),
1227                                   kioreq->vcpus, sizeof(*ports));
1228         if (IS_ERR(ports)) {
1229                 ret = PTR_ERR(ports);
1230                 goto error_kfree;
1231         }
1232
1233         for (i = 0; i < kioreq->vcpus; i++) {
1234                 kioreq->ports[i].vcpu = i;
1235                 kioreq->ports[i].port = ports[i];
1236                 kioreq->ports[i].kioreq = kioreq;
1237
1238                 ret = bind_evtchn_to_irqhandler_lateeoi(ports[i],
1239                                 ioeventfd_interrupt, IRQF_SHARED, "ioeventfd",
1240                                 &kioreq->ports[i]);
1241                 if (ret < 0)
1242                         goto error_unbind;
1243         }
1244
1245         kfree(ports);
1246
1247         list_add_tail(&kioreq->list, &ioreq_list);
1248
1249         return kioreq;
1250
1251 error_unbind:
1252         while (--i >= 0)
1253                 unbind_from_irqhandler(irq_from_evtchn(ports[i]), &kioreq->ports[i]);
1254
1255         kfree(ports);
1256 error_kfree:
1257         kfree(kioreq);
1258         return ERR_PTR(ret);
1259 }
1260
1261 static struct privcmd_kernel_ioreq *
1262 get_ioreq(struct privcmd_ioeventfd *ioeventfd, struct eventfd_ctx *eventfd)
1263 {
1264         struct privcmd_kernel_ioreq *kioreq;
1265         unsigned long flags;
1266
1267         list_for_each_entry(kioreq, &ioreq_list, list) {
1268                 struct privcmd_kernel_ioeventfd *kioeventfd;
1269
1270                 /*
1271                  * kioreq fields can be accessed here without a lock as they are
1272                  * never updated after being added to the ioreq_list.
1273                  */
1274                 if (kioreq->uioreq != ioeventfd->ioreq) {
1275                         continue;
1276                 } else if (kioreq->dom != ioeventfd->dom ||
1277                            kioreq->vcpus != ioeventfd->vcpus) {
1278                         pr_err("Invalid ioeventfd configuration mismatch, dom (%u vs %u), vcpus (%u vs %u)\n",
1279                                kioreq->dom, ioeventfd->dom, kioreq->vcpus,
1280                                ioeventfd->vcpus);
1281                         return ERR_PTR(-EINVAL);
1282                 }
1283
1284                 /* Look for a duplicate eventfd for the same guest */
1285                 spin_lock_irqsave(&kioreq->lock, flags);
1286                 list_for_each_entry(kioeventfd, &kioreq->ioeventfds, list) {
1287                         if (eventfd == kioeventfd->eventfd) {
1288                                 spin_unlock_irqrestore(&kioreq->lock, flags);
1289                                 return ERR_PTR(-EBUSY);
1290                         }
1291                 }
1292                 spin_unlock_irqrestore(&kioreq->lock, flags);
1293
1294                 return kioreq;
1295         }
1296
1297         /* Matching kioreq isn't found, allocate a new one */
1298         return alloc_ioreq(ioeventfd);
1299 }
1300
1301 static void ioeventfd_free(struct privcmd_kernel_ioeventfd *kioeventfd)
1302 {
1303         list_del(&kioeventfd->list);
1304         eventfd_ctx_put(kioeventfd->eventfd);
1305         kfree(kioeventfd);
1306 }
1307
1308 static int privcmd_ioeventfd_assign(struct privcmd_ioeventfd *ioeventfd)
1309 {
1310         struct privcmd_kernel_ioeventfd *kioeventfd;
1311         struct privcmd_kernel_ioreq *kioreq;
1312         unsigned long flags;
1313         struct fd f;
1314         int ret;
1315
1316         /* Check for range overflow */
1317         if (ioeventfd->addr + ioeventfd->addr_len < ioeventfd->addr)
1318                 return -EINVAL;
1319
1320         /* Vhost requires us to support length 1, 2, 4, and 8 */
1321         if (!(ioeventfd->addr_len == 1 || ioeventfd->addr_len == 2 ||
1322               ioeventfd->addr_len == 4 || ioeventfd->addr_len == 8))
1323                 return -EINVAL;
1324
1325         /* 4096 vcpus limit enough ? */
1326         if (!ioeventfd->vcpus || ioeventfd->vcpus > 4096)
1327                 return -EINVAL;
1328
1329         kioeventfd = kzalloc(sizeof(*kioeventfd), GFP_KERNEL);
1330         if (!kioeventfd)
1331                 return -ENOMEM;
1332
1333         f = fdget(ioeventfd->event_fd);
1334         if (!f.file) {
1335                 ret = -EBADF;
1336                 goto error_kfree;
1337         }
1338
1339         kioeventfd->eventfd = eventfd_ctx_fileget(f.file);
1340         fdput(f);
1341
1342         if (IS_ERR(kioeventfd->eventfd)) {
1343                 ret = PTR_ERR(kioeventfd->eventfd);
1344                 goto error_kfree;
1345         }
1346
1347         kioeventfd->addr = ioeventfd->addr;
1348         kioeventfd->addr_len = ioeventfd->addr_len;
1349         kioeventfd->vq = ioeventfd->vq;
1350
1351         mutex_lock(&ioreq_lock);
1352         kioreq = get_ioreq(ioeventfd, kioeventfd->eventfd);
1353         if (IS_ERR(kioreq)) {
1354                 mutex_unlock(&ioreq_lock);
1355                 ret = PTR_ERR(kioreq);
1356                 goto error_eventfd;
1357         }
1358
1359         spin_lock_irqsave(&kioreq->lock, flags);
1360         list_add_tail(&kioeventfd->list, &kioreq->ioeventfds);
1361         spin_unlock_irqrestore(&kioreq->lock, flags);
1362
1363         mutex_unlock(&ioreq_lock);
1364
1365         return 0;
1366
1367 error_eventfd:
1368         eventfd_ctx_put(kioeventfd->eventfd);
1369
1370 error_kfree:
1371         kfree(kioeventfd);
1372         return ret;
1373 }
1374
1375 static int privcmd_ioeventfd_deassign(struct privcmd_ioeventfd *ioeventfd)
1376 {
1377         struct privcmd_kernel_ioreq *kioreq, *tkioreq;
1378         struct eventfd_ctx *eventfd;
1379         unsigned long flags;
1380         int ret = 0;
1381
1382         eventfd = eventfd_ctx_fdget(ioeventfd->event_fd);
1383         if (IS_ERR(eventfd))
1384                 return PTR_ERR(eventfd);
1385
1386         mutex_lock(&ioreq_lock);
1387         list_for_each_entry_safe(kioreq, tkioreq, &ioreq_list, list) {
1388                 struct privcmd_kernel_ioeventfd *kioeventfd, *tmp;
1389                 /*
1390                  * kioreq fields can be accessed here without a lock as they are
1391                  * never updated after being added to the ioreq_list.
1392                  */
1393                 if (kioreq->dom != ioeventfd->dom ||
1394                     kioreq->uioreq != ioeventfd->ioreq ||
1395                     kioreq->vcpus != ioeventfd->vcpus)
1396                         continue;
1397
1398                 spin_lock_irqsave(&kioreq->lock, flags);
1399                 list_for_each_entry_safe(kioeventfd, tmp, &kioreq->ioeventfds, list) {
1400                         if (eventfd == kioeventfd->eventfd) {
1401                                 ioeventfd_free(kioeventfd);
1402                                 spin_unlock_irqrestore(&kioreq->lock, flags);
1403
1404                                 if (list_empty(&kioreq->ioeventfds))
1405                                         ioreq_free(kioreq);
1406                                 goto unlock;
1407                         }
1408                 }
1409                 spin_unlock_irqrestore(&kioreq->lock, flags);
1410                 break;
1411         }
1412
1413         pr_err("Ioeventfd isn't already assigned, dom: %u, addr: %llu\n",
1414                ioeventfd->dom, ioeventfd->addr);
1415         ret = -ENODEV;
1416
1417 unlock:
1418         mutex_unlock(&ioreq_lock);
1419         eventfd_ctx_put(eventfd);
1420
1421         return ret;
1422 }
1423
1424 static long privcmd_ioctl_ioeventfd(struct file *file, void __user *udata)
1425 {
1426         struct privcmd_data *data = file->private_data;
1427         struct privcmd_ioeventfd ioeventfd;
1428
1429         if (copy_from_user(&ioeventfd, udata, sizeof(ioeventfd)))
1430                 return -EFAULT;
1431
1432         /* No other flags should be set */
1433         if (ioeventfd.flags & ~PRIVCMD_IOEVENTFD_FLAG_DEASSIGN)
1434                 return -EINVAL;
1435
1436         /* If restriction is in place, check the domid matches */
1437         if (data->domid != DOMID_INVALID && data->domid != ioeventfd.dom)
1438                 return -EPERM;
1439
1440         if (ioeventfd.flags & PRIVCMD_IOEVENTFD_FLAG_DEASSIGN)
1441                 return privcmd_ioeventfd_deassign(&ioeventfd);
1442
1443         return privcmd_ioeventfd_assign(&ioeventfd);
1444 }
1445
1446 static void privcmd_ioeventfd_exit(void)
1447 {
1448         struct privcmd_kernel_ioreq *kioreq, *tmp;
1449         unsigned long flags;
1450
1451         mutex_lock(&ioreq_lock);
1452         list_for_each_entry_safe(kioreq, tmp, &ioreq_list, list) {
1453                 struct privcmd_kernel_ioeventfd *kioeventfd, *tmp;
1454
1455                 spin_lock_irqsave(&kioreq->lock, flags);
1456                 list_for_each_entry_safe(kioeventfd, tmp, &kioreq->ioeventfds, list)
1457                         ioeventfd_free(kioeventfd);
1458                 spin_unlock_irqrestore(&kioreq->lock, flags);
1459
1460                 ioreq_free(kioreq);
1461         }
1462         mutex_unlock(&ioreq_lock);
1463 }
1464 #else
1465 static inline long privcmd_ioctl_irqfd(struct file *file, void __user *udata)
1466 {
1467         return -EOPNOTSUPP;
1468 }
1469
1470 static inline int privcmd_irqfd_init(void)
1471 {
1472         return 0;
1473 }
1474
1475 static inline void privcmd_irqfd_exit(void)
1476 {
1477 }
1478
1479 static inline long privcmd_ioctl_ioeventfd(struct file *file, void __user *udata)
1480 {
1481         return -EOPNOTSUPP;
1482 }
1483
1484 static inline void privcmd_ioeventfd_exit(void)
1485 {
1486 }
1487 #endif /* CONFIG_XEN_PRIVCMD_EVENTFD */
1488
1489 static long privcmd_ioctl(struct file *file,
1490                           unsigned int cmd, unsigned long data)
1491 {
1492         int ret = -ENOTTY;
1493         void __user *udata = (void __user *) data;
1494
1495         switch (cmd) {
1496         case IOCTL_PRIVCMD_HYPERCALL:
1497                 ret = privcmd_ioctl_hypercall(file, udata);
1498                 break;
1499
1500         case IOCTL_PRIVCMD_MMAP:
1501                 ret = privcmd_ioctl_mmap(file, udata);
1502                 break;
1503
1504         case IOCTL_PRIVCMD_MMAPBATCH:
1505                 ret = privcmd_ioctl_mmap_batch(file, udata, 1);
1506                 break;
1507
1508         case IOCTL_PRIVCMD_MMAPBATCH_V2:
1509                 ret = privcmd_ioctl_mmap_batch(file, udata, 2);
1510                 break;
1511
1512         case IOCTL_PRIVCMD_DM_OP:
1513                 ret = privcmd_ioctl_dm_op(file, udata);
1514                 break;
1515
1516         case IOCTL_PRIVCMD_RESTRICT:
1517                 ret = privcmd_ioctl_restrict(file, udata);
1518                 break;
1519
1520         case IOCTL_PRIVCMD_MMAP_RESOURCE:
1521                 ret = privcmd_ioctl_mmap_resource(file, udata);
1522                 break;
1523
1524         case IOCTL_PRIVCMD_IRQFD:
1525                 ret = privcmd_ioctl_irqfd(file, udata);
1526                 break;
1527
1528         case IOCTL_PRIVCMD_IOEVENTFD:
1529                 ret = privcmd_ioctl_ioeventfd(file, udata);
1530                 break;
1531
1532         default:
1533                 break;
1534         }
1535
1536         return ret;
1537 }
1538
1539 static int privcmd_open(struct inode *ino, struct file *file)
1540 {
1541         struct privcmd_data *data = kzalloc(sizeof(*data), GFP_KERNEL);
1542
1543         if (!data)
1544                 return -ENOMEM;
1545
1546         /* DOMID_INVALID implies no restriction */
1547         data->domid = DOMID_INVALID;
1548
1549         file->private_data = data;
1550         return 0;
1551 }
1552
1553 static int privcmd_release(struct inode *ino, struct file *file)
1554 {
1555         struct privcmd_data *data = file->private_data;
1556
1557         kfree(data);
1558         return 0;
1559 }
1560
1561 static void privcmd_close(struct vm_area_struct *vma)
1562 {
1563         struct page **pages = vma->vm_private_data;
1564         int numpgs = vma_pages(vma);
1565         int numgfns = (vma->vm_end - vma->vm_start) >> XEN_PAGE_SHIFT;
1566         int rc;
1567
1568         if (!xen_feature(XENFEAT_auto_translated_physmap) || !numpgs || !pages)
1569                 return;
1570
1571         rc = xen_unmap_domain_gfn_range(vma, numgfns, pages);
1572         if (rc == 0)
1573                 xen_free_unpopulated_pages(numpgs, pages);
1574         else
1575                 pr_crit("unable to unmap MFN range: leaking %d pages. rc=%d\n",
1576                         numpgs, rc);
1577         kvfree(pages);
1578 }
1579
1580 static vm_fault_t privcmd_fault(struct vm_fault *vmf)
1581 {
1582         printk(KERN_DEBUG "privcmd_fault: vma=%p %lx-%lx, pgoff=%lx, uv=%p\n",
1583                vmf->vma, vmf->vma->vm_start, vmf->vma->vm_end,
1584                vmf->pgoff, (void *)vmf->address);
1585
1586         return VM_FAULT_SIGBUS;
1587 }
1588
1589 static const struct vm_operations_struct privcmd_vm_ops = {
1590         .close = privcmd_close,
1591         .fault = privcmd_fault
1592 };
1593
1594 static int privcmd_mmap(struct file *file, struct vm_area_struct *vma)
1595 {
1596         /* DONTCOPY is essential for Xen because copy_page_range doesn't know
1597          * how to recreate these mappings */
1598         vm_flags_set(vma, VM_IO | VM_PFNMAP | VM_DONTCOPY |
1599                          VM_DONTEXPAND | VM_DONTDUMP);
1600         vma->vm_ops = &privcmd_vm_ops;
1601         vma->vm_private_data = NULL;
1602
1603         return 0;
1604 }
1605
1606 /*
1607  * For MMAPBATCH*. This allows asserting the singleshot mapping
1608  * on a per pfn/pte basis. Mapping calls that fail with ENOENT
1609  * can be then retried until success.
1610  */
1611 static int is_mapped_fn(pte_t *pte, unsigned long addr, void *data)
1612 {
1613         return pte_none(ptep_get(pte)) ? 0 : -EBUSY;
1614 }
1615
1616 static int privcmd_vma_range_is_mapped(
1617                    struct vm_area_struct *vma,
1618                    unsigned long addr,
1619                    unsigned long nr_pages)
1620 {
1621         return apply_to_page_range(vma->vm_mm, addr, nr_pages << PAGE_SHIFT,
1622                                    is_mapped_fn, NULL) != 0;
1623 }
1624
1625 const struct file_operations xen_privcmd_fops = {
1626         .owner = THIS_MODULE,
1627         .unlocked_ioctl = privcmd_ioctl,
1628         .open = privcmd_open,
1629         .release = privcmd_release,
1630         .mmap = privcmd_mmap,
1631 };
1632 EXPORT_SYMBOL_GPL(xen_privcmd_fops);
1633
1634 static struct miscdevice privcmd_dev = {
1635         .minor = MISC_DYNAMIC_MINOR,
1636         .name = "xen/privcmd",
1637         .fops = &xen_privcmd_fops,
1638 };
1639
1640 static int __init privcmd_init(void)
1641 {
1642         int err;
1643
1644         if (!xen_domain())
1645                 return -ENODEV;
1646
1647         err = misc_register(&privcmd_dev);
1648         if (err != 0) {
1649                 pr_err("Could not register Xen privcmd device\n");
1650                 return err;
1651         }
1652
1653         err = misc_register(&xen_privcmdbuf_dev);
1654         if (err != 0) {
1655                 pr_err("Could not register Xen hypercall-buf device\n");
1656                 goto err_privcmdbuf;
1657         }
1658
1659         err = privcmd_irqfd_init();
1660         if (err != 0) {
1661                 pr_err("irqfd init failed\n");
1662                 goto err_irqfd;
1663         }
1664
1665         return 0;
1666
1667 err_irqfd:
1668         misc_deregister(&xen_privcmdbuf_dev);
1669 err_privcmdbuf:
1670         misc_deregister(&privcmd_dev);
1671         return err;
1672 }
1673
1674 static void __exit privcmd_exit(void)
1675 {
1676         privcmd_ioeventfd_exit();
1677         privcmd_irqfd_exit();
1678         misc_deregister(&privcmd_dev);
1679         misc_deregister(&xen_privcmdbuf_dev);
1680 }
1681
1682 module_init(privcmd_init);
1683 module_exit(privcmd_exit);