fuse_sideload.cpp

/*
 * Copyright (C) 2014 The Android Open Source Project
 * Copyright (C) 2019 The LineageOS Project
 *
 * Licensed under the Apache License, Version 2.0 (the "License");
 * you may not use this file except in compliance with the License.
 * You may obtain a copy of the License at
 *
 *      http://www.apache.org/licenses/LICENSE-2.0
 *
 * Unless required by applicable law or agreed to in writing, software
 * distributed under the License is distributed on an "AS IS" BASIS,
 * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 * See the License for the specific language governing permissions and
 * limitations under the License.
 */

// This module creates a special filesystem containing two files.
//
// "/sideload/package.zip" appears to be a normal file, but reading
// from it causes data to be fetched from the adb host.  We can use
// this to sideload packages over an adb connection without having to
// store the entire package in RAM on the device.
//
// Because we may not trust the adb host, this filesystem maintains
// the following invariant: each read of a given position returns the
// same data as the first read at that position.  That is, once a
// section of the file is read, future reads of that section return
// the same data.  (Otherwise, a malicious adb host process could
// return one set of bits when the package is read for signature
// verification, and then different bits for when the package is
// accessed by the installer.)  If the adb host returns something
// different than it did on the first read, the reader of the file
// will see their read fail with EINVAL.
//
// The other file, "/sideload/exit", is used to control the subprocess
// that creates this filesystem.  Calling stat() on the exit file
// causes the filesystem to be unmounted and the adb process on the
// device shut down.
//
// Note that only the minimal set of file operations needed for these
// two files is implemented.  In particular, you can't opendir() or
// readdir() on the "/sideload" directory; ls on it won't work.

#include "fuse_sideload.h"

#include <errno.h>
#include <fcntl.h>
#include <limits.h>  // PATH_MAX
#include <linux/fuse.h>
#include <stdint.h>
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
#include <sys/mount.h>
#include <sys/param.h>  // MIN
#include <sys/stat.h>
#include <sys/uio.h>
#include <unistd.h>

#include <array>
#include <string>
#include <vector>

#include <android-base/stringprintf.h>
#include <android-base/unique_fd.h>
#include <openssl/sha.h>

static constexpr uint64_t PACKAGE_FILE_ID = FUSE_ROOT_ID + 1;

static constexpr int NO_STATUS = 1;

using SHA256Digest = std::array<uint8_t, SHA256_DIGEST_LENGTH>;

#define INSTALL_REQUIRED_MEMORY (100 * 1024 * 1024)

struct fuse_data {
  android::base::unique_fd ffd;  // file descriptor for the fuse socket

  provider_vtab vtab;

  uint64_t file_size;  // bytes

  uint32_t block_size;   // block size that the adb host is using to send the file to us
  uint32_t file_blocks;  // file size in block_size blocks

  uid_t uid;
  gid_t gid;

  uint32_t curr_block;  // cache the block most recently used
  uint8_t* block_data;

  uint8_t* extra_block;  // another block of storage for reads that span two blocks

  std::vector<SHA256Digest>
      hashes;  // SHA-256 hash of each block (all zeros if block hasn't been read yet)

  // Block cache
  uint32_t block_cache_max_size;  // Max allowed block cache size
  uint32_t block_cache_size;      // Current block cache size
  uint8_t** block_cache;          // Block cache data
};

static uint64_t free_memory() {
  uint64_t mem = 0;
  FILE* fp = fopen("/proc/meminfo", "r");
  if (fp) {
    char buf[256];
    char* linebuf = buf;
    size_t buflen = sizeof(buf);
    while (getline(&linebuf, &buflen, fp) > 0) {
      char* key = buf;
      char* val = strchr(buf, ':');
      *val = '\0';
      ++val;
      if (strcmp(key, "MemFree") == 0) {
        mem += strtoul(val, nullptr, 0) * 1024;
      }
      if (strcmp(key, "Buffers") == 0) {
        mem += strtoul(val, nullptr, 0) * 1024;
      }
      if (strcmp(key, "Cached") == 0) {
        mem += strtoul(val, nullptr, 0) * 1024;
      }
    }
    fclose(fp);
  }
  return mem;
}

static int block_cache_fetch(struct fuse_data* fd, uint32_t block) {
  if (fd->block_cache == nullptr) {
    return -1;
  }
  if (fd->block_cache[block] == nullptr) {
    return -1;
  }
  memcpy(fd->block_data, fd->block_cache[block], fd->block_size);
  return 0;
}

static void block_cache_enter(struct fuse_data* fd, uint32_t block) {
  if (!fd->block_cache) return;
  if (fd->block_cache_size == fd->block_cache_max_size) {
    // Evict a block from the cache.  Since the file is typically read
    // sequentially, start looking from the block behind the current
    // block and proceed backward.
    int n;
    for (n = fd->curr_block - 1; n != (int)fd->curr_block; --n) {
      if (n < 0) {
        n = fd->file_blocks - 1;
      }
      if (fd->block_cache[n]) {
        free(fd->block_cache[n]);
        fd->block_cache[n] = nullptr;
        fd->block_cache_size--;
        break;
      }
    }
  }

  fd->block_cache[block] = (uint8_t*)malloc(fd->block_size);
  memcpy(fd->block_cache[block], fd->block_data, fd->block_size);

  fd->block_cache_size++;
}

static void fuse_reply(const fuse_data* fd, uint64_t unique, const void* data, size_t len) {
  fuse_out_header hdr;
  hdr.len = len + sizeof(hdr);
  hdr.error = 0;
  hdr.unique = unique;

  struct iovec vec[2];
  vec[0].iov_base = &hdr;
  vec[0].iov_len = sizeof(hdr);
  vec[1].iov_base = const_cast<void*>(data);
  vec[1].iov_len = len;

  int res = writev(fd->ffd, vec, 2);
  if (res == -1) {
    printf("*** REPLY FAILED *** %s\n", strerror(errno));
  }
}

static int handle_init(void* data, fuse_data* fd, const fuse_in_header* hdr) {
  const fuse_init_in* req = static_cast<const fuse_init_in*>(data);

  // Kernel 2.6.16 is the first stable kernel with struct fuse_init_out defined (fuse version 7.6).
  // The structure is the same from 7.6 through 7.22. Beginning with 7.23, the structure increased
  // in size and added new parameters.
  if (req->major != FUSE_KERNEL_VERSION || req->minor < 6) {
    printf("Fuse kernel version mismatch: Kernel version %d.%d, Expected at least %d.6", req->major,
           req->minor, FUSE_KERNEL_VERSION);
    return -1;
  }

  fuse_init_out out;
  out.minor = MIN(req->minor, FUSE_KERNEL_MINOR_VERSION);
  size_t fuse_struct_size = sizeof(out);
#if defined(FUSE_COMPAT_22_INIT_OUT_SIZE)
  /* FUSE_KERNEL_VERSION >= 23. */

  // If the kernel only works on minor revs older than or equal to 22, then use the older structure
  // size since this code only uses the 7.22 version of the structure.
  if (req->minor <= 22) {
    fuse_struct_size = FUSE_COMPAT_22_INIT_OUT_SIZE;
  }
#endif

  out.major = FUSE_KERNEL_VERSION;
  out.max_readahead = req->max_readahead;
  out.flags = 0;
  out.max_background = 32;
  out.congestion_threshold = 32;
  out.max_write = 4096;
  fuse_reply(fd, hdr->unique, &out, fuse_struct_size);

  return NO_STATUS;
}

static void fill_attr(fuse_attr* attr, const fuse_data* fd, uint64_t nodeid, uint64_t size,
                      uint32_t mode) {
  *attr = {};
  attr->nlink = 1;
  attr->uid = fd->uid;
  attr->gid = fd->gid;
  attr->blksize = 4096;

  attr->ino = nodeid;
  attr->size = size;
  attr->blocks = (size == 0) ? 0 : (((size - 1) / attr->blksize) + 1);
  attr->mode = mode;
}

static int handle_getattr(void* /* data */, const fuse_data* fd, const fuse_in_header* hdr) {
  fuse_attr_out out = {};
  out.attr_valid = 10;

  if (hdr->nodeid == FUSE_ROOT_ID) {
    fill_attr(&(out.attr), fd, hdr->nodeid, 4096, S_IFDIR | 0555);
  } else if (hdr->nodeid == PACKAGE_FILE_ID) {
    fill_attr(&(out.attr), fd, PACKAGE_FILE_ID, fd->file_size, S_IFREG | 0444);
  } else {
    return -ENOENT;
  }

  fuse_reply(fd, hdr->unique, &out, sizeof(out));
  return NO_STATUS;
}

static int handle_lookup(void* data, const fuse_data* fd, const fuse_in_header* hdr) {
  if (data == nullptr) return -ENOENT;

  fuse_entry_out out = {};
  out.entry_valid = 10;
  out.attr_valid = 10;

  std::string filename(static_cast<const char*>(data));
  if (filename == FUSE_SIDELOAD_HOST_FILENAME) {
    out.nodeid = PACKAGE_FILE_ID;
    out.generation = PACKAGE_FILE_ID;
    fill_attr(&(out.attr), fd, PACKAGE_FILE_ID, fd->file_size, S_IFREG | 0444);
  } else {
    return -ENOENT;
  }

  fuse_reply(fd, hdr->unique, &out, sizeof(out));
  return NO_STATUS;
}

static int handle_open(void* /* data */, const fuse_data* fd, const fuse_in_header* hdr) {
  if (hdr->nodeid != PACKAGE_FILE_ID) return -ENOENT;

  fuse_open_out out = {};
  out.fh = 10;  // an arbitrary number; we always use the same handle
  fuse_reply(fd, hdr->unique, &out, sizeof(out));
  return NO_STATUS;
}

static int handle_flush(void* /* data */, fuse_data* /* fd */, const fuse_in_header* /* hdr */) {
  return 0;
}

static int handle_release(void* /* data */, fuse_data* /* fd */, const fuse_in_header* /* hdr */) {
  return 0;
}

// Fetch a block from the host into fd->curr_block and fd->block_data.
// Returns 0 on successful fetch, negative otherwise.
static int fetch_block(fuse_data* fd, uint32_t block) {
  if (block == fd->curr_block) {
    return 0;
  }

  if (block >= fd->file_blocks) {
    memset(fd->block_data, 0, fd->block_size);
    fd->curr_block = block;
    return 0;
  }

  if (block_cache_fetch(fd, block) == 0) {
    fd->curr_block = block;
    return 0;
  }

  size_t fetch_size = fd->block_size;
  if (block * fd->block_size + fetch_size > fd->file_size) {
    // If we're reading the last (partial) block of the file, expect a shorter response from the
    // host, and pad the rest of the block with zeroes.
    fetch_size = fd->file_size - (block * fd->block_size);
    memset(fd->block_data + fetch_size, 0, fd->block_size - fetch_size);
  }

  int result = fd->vtab.read_block(block, fd->block_data, fetch_size);
  if (result < 0) return result;

  fd->curr_block = block;

  // Verify the hash of the block we just got from the host.
  //
  // - If the hash of the just-received data matches the stored hash for the block, accept it.
  // - If the stored hash is all zeroes, store the new hash and accept the block (this is the first
  //   time we've read this block).
  // - Otherwise, return -EINVAL for the read.

  SHA256Digest hash;
  SHA256(fd->block_data, fd->block_size, hash.data());

  const SHA256Digest& blockhash = fd->hashes[block];
  if (hash == blockhash) {
    return 0;
  }

  for (uint8_t i : blockhash) {
    if (i != 0) {
      fd->curr_block = -1;
      return -EIO;
    }
  }

  fd->hashes[block] = hash;
  block_cache_enter(fd, block);
  return 0;
}

static int handle_read(void* data, fuse_data* fd, const fuse_in_header* hdr) {
  if (hdr->nodeid != PACKAGE_FILE_ID) return -ENOENT;

  const fuse_read_in* req = static_cast<const fuse_read_in*>(data);
  uint64_t offset = req->offset;
  uint32_t size = req->size;

  // The docs on the fuse kernel interface are vague about what to do when a read request extends
  // past the end of the file. We can return a short read -- the return structure does include a
  // length field -- but in testing that caused the program using the file to segfault. (I
  // speculate that this is due to the reading program accessing it via mmap; maybe mmap dislikes
  // when you return something short of a whole page?) To fix this we zero-pad reads that extend
  // past the end of the file so we're always returning exactly as many bytes as were requested.
  // (Users of the mapped file have to know its real length anyway.)

  fuse_out_header outhdr;
  outhdr.len = sizeof(outhdr) + size;
  outhdr.error = 0;
  outhdr.unique = hdr->unique;

  struct iovec vec[3];
  vec[0].iov_base = &outhdr;
  vec[0].iov_len = sizeof(outhdr);

  uint32_t block = offset / fd->block_size;
  int result = fetch_block(fd, block);
  if (result != 0) return result;

  // Two cases:
  //
  //   - the read request is entirely within this block. In this case we can reply immediately.
  //
  //   - the read request goes over into the next block. Note that since we mount the filesystem
  //     with max_read=block_size, a read can never span more than two blocks. In this case we copy
  //     the block to extra_block and issue a fetch for the following block.

  uint32_t block_offset = offset - (block * fd->block_size);

  int vec_used;
  if (size + block_offset <= fd->block_size) {
    // First case: the read fits entirely in the first block.

    vec[1].iov_base = fd->block_data + block_offset;
    vec[1].iov_len = size;
    vec_used = 2;
  } else {
    // Second case: the read spills over into the next block.

    memcpy(fd->extra_block, fd->block_data + block_offset, fd->block_size - block_offset);
    vec[1].iov_base = fd->extra_block;
    vec[1].iov_len = fd->block_size - block_offset;

    result = fetch_block(fd, block + 1);
    if (result != 0) return result;
    vec[2].iov_base = fd->block_data;
    vec[2].iov_len = size - vec[1].iov_len;
    vec_used = 3;
  }

  if (writev(fd->ffd, vec, vec_used) == -1) {
    printf("*** READ REPLY FAILED: %s ***\n", strerror(errno));
  }
  return NO_STATUS;
}

static volatile int terminated = 0;
static void sig_term(int) {
  terminated = 1;
}

int run_fuse_sideload(const provider_vtab& vtab, uint64_t file_size, uint32_t block_size,
                      const char* mount_point) {
  // If something's already mounted on our mountpoint, try to remove it. (Mostly in case of a
  // previous abnormal exit.)
  umount2(mount_point, MNT_FORCE);

  // fs/fuse/inode.c in kernel code uses the greater of 4096 and the passed-in max_read.
  if (block_size < 4096) {
    fprintf(stderr, "block size (%u) is too small\n", block_size);
    return -1;
  }
  if (block_size > (1 << 22)) {  // 4 MiB
    fprintf(stderr, "block size (%u) is too large\n", block_size);
    return -1;
  }

  fuse_data fd = {};
  fd.vtab = vtab;
  fd.file_size = file_size;
  fd.block_size = block_size;
  fd.file_blocks = (file_size == 0) ? 0 : (((file_size - 1) / block_size) + 1);

  uint64_t mem = free_memory();
  uint64_t avail = mem - (INSTALL_REQUIRED_MEMORY + fd.file_blocks * sizeof(uint8_t*));

  int result;
  if (fd.file_blocks > (1 << 18)) {
    fprintf(stderr, "file has too many blocks (%u)\n", fd.file_blocks);
    result = -1;
    goto done;
  }

  // All hashes will be zero-initialized.
  fd.hashes.resize(fd.file_blocks);
  fd.uid = getuid();
  fd.gid = getgid();

  fd.curr_block = -1;
  fd.block_data = static_cast<uint8_t*>(malloc(block_size));
  if (fd.block_data == nullptr) {
    fprintf(stderr, "failed to allocate %d bites for block_data\n", block_size);
    result = -1;
    goto done;
  }
  fd.extra_block = static_cast<uint8_t*>(malloc(block_size));
  if (fd.extra_block == nullptr) {
    fprintf(stderr, "failed to allocate %d bites for extra_block\n", block_size);
    result = -1;
    goto done;
  }

  fd.block_cache_max_size = 0;
  fd.block_cache_size = 0;
  fd.block_cache = nullptr;
  if (mem > avail) {
    uint32_t max_size = avail / fd.block_size;
    if (max_size > fd.file_blocks) {
      max_size = fd.file_blocks;
    }
    // The cache must be at least 1% of the file size or two blocks,
    // whichever is larger.
    if (max_size >= fd.file_blocks / 100 && max_size >= 2) {
      fd.block_cache_max_size = max_size;
      fd.block_cache = (uint8_t**)calloc(fd.file_blocks, sizeof(uint8_t*));
    }
  }

  signal(SIGTERM, sig_term);

  fd.ffd.reset(open("/dev/fuse", O_RDWR));
  if (!fd.ffd) {
    perror("open /dev/fuse");
    result = -1;
    goto done;
  }

  {
    std::string opts = android::base::StringPrintf(
        "fd=%d,user_id=%d,group_id=%d,max_read=%u,allow_other,rootmode=040000", fd.ffd.get(),
        fd.uid, fd.gid, block_size);

    result = mount("/dev/fuse", mount_point, "fuse", MS_NOSUID | MS_NODEV | MS_RDONLY | MS_NOEXEC,
                   opts.c_str());
    if (result == -1) {
      perror("mount");
      goto done;
    }
  }

  uint8_t request_buffer[sizeof(fuse_in_header) + PATH_MAX * 8];
  while (!terminated) {
    fd_set fds;
    struct timeval tv;
    FD_ZERO(&fds);
    FD_SET(fd.ffd, &fds);
    tv.tv_sec = 1;
    tv.tv_usec = 0;
    int rc = select(fd.ffd + 1, &fds, nullptr, nullptr, &tv);
    if (rc <= 0) {
      continue;
    }
    ssize_t len = TEMP_FAILURE_RETRY(read(fd.ffd, request_buffer, sizeof(request_buffer)));
    if (len == -1) {
      perror("read request");
      if (errno == ENODEV) {
        result = -1;
        break;
      }
      continue;
    }

    if (static_cast<size_t>(len) < sizeof(fuse_in_header)) {
      fprintf(stderr, "request too short: len=%zd\n", len);
      continue;
    }

    fuse_in_header* hdr = reinterpret_cast<fuse_in_header*>(request_buffer);
    void* data = request_buffer + sizeof(fuse_in_header);

    result = -ENOSYS;

    switch (hdr->opcode) {
      case FUSE_INIT:
        result = handle_init(data, &fd, hdr);
        break;

      case FUSE_LOOKUP:
        result = handle_lookup(data, &fd, hdr);
        break;

      case FUSE_GETATTR:
        result = handle_getattr(data, &fd, hdr);
        break;

      case FUSE_OPEN:
        result = handle_open(data, &fd, hdr);
        break;

      case FUSE_READ:
        result = handle_read(data, &fd, hdr);
        break;

      case FUSE_FLUSH:
        result = handle_flush(data, &fd, hdr);
        break;

      case FUSE_RELEASE:
        result = handle_release(data, &fd, hdr);
        break;

      default:
        fprintf(stderr, "unknown fuse request opcode %d\n", hdr->opcode);
        break;
    }

    if (result != NO_STATUS) {
      fuse_out_header outhdr;
      outhdr.len = sizeof(outhdr);
      outhdr.error = result;
      outhdr.unique = hdr->unique;
      TEMP_FAILURE_RETRY(write(fd.ffd, &outhdr, sizeof(outhdr)));
    }
  }

done:
  fd.vtab.close();

  if (umount2(mount_point, MNT_DETACH) == -1) {
    fprintf(stderr, "fuse_sideload umount failed: %s\n", strerror(errno));
  }

  if (fd.block_cache) {
    uint32_t n;
    for (n = 0; n < fd.file_blocks; ++n) {
      free(fd.block_cache[n]);
    }
    free(fd.block_cache);
  }

  free(fd.block_data);
  free(fd.extra_block);

  return result;
}