• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 // Copyright (C) 2019 The Android Open Source Project
2 //
3 // Licensed under the Apache License, Version 2.0 (the "License");
4 // you may not use this file except in compliance with the License.
5 // You may obtain a copy of the License at
6 //
7 //      http://www.apache.org/licenses/LICENSE-2.0
8 //
9 // Unless required by applicable law or agreed to in writing, software
10 // distributed under the License is distributed on an "AS IS" BASIS,
11 // WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
12 // See the License for the specific language governing permissions and
13 // limitations under the License.
14 
15 #define ATRACE_TAG ATRACE_TAG_APP
16 #define LOG_TAG "FuseDaemon"
17 #define LIBFUSE_LOG_TAG "libfuse"
18 
19 #include "FuseDaemon.h"
20 #include "android-base/strings.h"
21 
22 #include <android-base/logging.h>
23 #include <android-base/properties.h>
24 #include <android/log.h>
25 #include <android/trace.h>
26 #include <ctype.h>
27 #include <dirent.h>
28 #include <errno.h>
29 #include <fcntl.h>
30 #include <fuse_i.h>
31 #include <fuse_log.h>
32 #include <fuse_lowlevel.h>
33 #include <inttypes.h>
34 #include <limits.h>
35 #include <linux/fuse.h>
36 #include <stdbool.h>
37 #include <stdio.h>
38 #include <stdlib.h>
39 #include <string.h>
40 #include <sys/inotify.h>
41 #include <sys/mman.h>
42 #include <sys/mount.h>
43 #include <sys/param.h>
44 #include <sys/resource.h>
45 #include <sys/stat.h>
46 #include <sys/statfs.h>
47 #include <sys/statvfs.h>
48 #include <sys/time.h>
49 #include <sys/types.h>
50 #include <sys/uio.h>
51 #include <unistd.h>
52 
53 #include <iostream>
54 #include <list>
55 #include <map>
56 #include <mutex>
57 #include <queue>
58 #include <regex>
59 #include <thread>
60 #include <unordered_map>
61 #include <unordered_set>
62 #include <vector>
63 
64 #include "MediaProviderWrapper.h"
65 #include "libfuse_jni/FuseUtils.h"
66 #include "libfuse_jni/ReaddirHelper.h"
67 #include "libfuse_jni/RedactionInfo.h"
68 #include "node-inl.h"
69 
70 using mediaprovider::fuse::DirectoryEntry;
71 using mediaprovider::fuse::dirhandle;
72 using mediaprovider::fuse::handle;
73 using mediaprovider::fuse::node;
74 using mediaprovider::fuse::RedactionInfo;
75 using std::list;
76 using std::string;
77 using std::vector;
78 
79 // logging macros to avoid duplication.
80 #define TRACE_NODE(__node, __req)                                                  \
81     LOG(VERBOSE) << __FUNCTION__ << " : " << #__node << " = [" << get_name(__node) \
82                  << "] (uid=" << (__req)->ctx.uid << ") "
83 
84 #define ATRACE_NAME(name) ScopedTrace ___tracer(name)
85 #define ATRACE_CALL() ATRACE_NAME(__FUNCTION__)
86 
87 class ScopedTrace {
88   public:
ScopedTrace(const char * name)89     explicit inline ScopedTrace(const char *name) {
90       ATrace_beginSection(name);
91     }
92 
~ScopedTrace()93     inline ~ScopedTrace() {
94       ATrace_endSection();
95     }
96 };
97 
98 const bool IS_OS_DEBUGABLE = android::base::GetIntProperty("ro.debuggable", 0);
99 
100 #define FUSE_UNKNOWN_INO 0xffffffff
101 
102 // Stolen from: android_filesystem_config.h
103 #define AID_APP_START 10000
104 
105 constexpr size_t MAX_READ_SIZE = 128 * 1024;
106 // Stolen from: UserHandle#getUserId
107 constexpr int PER_USER_RANGE = 100000;
108 
109 // Stolen from: UserManagerService
110 constexpr int MAX_USER_ID = UINT32_MAX / PER_USER_RANGE;
111 
112 const int MY_UID = getuid();
113 const int MY_USER_ID = MY_UID / PER_USER_RANGE;
114 const std::string MY_USER_ID_STRING(std::to_string(MY_UID / PER_USER_RANGE));
115 
116 // Regex copied from FileUtils.java in MediaProvider, but without media directory.
117 const std::regex PATTERN_OWNED_PATH(
118         "^/storage/[^/]+/(?:[0-9]+/)?Android/(?:data|obb)/([^/]+)(/?.*)?",
119         std::regex_constants::icase);
120 
121 static constexpr char TRANSFORM_SYNTHETIC_DIR[] = "synthetic";
122 static constexpr char TRANSFORM_TRANSCODE_DIR[] = "transcode";
123 
124 /*
125  * In order to avoid double caching with fuse, call fadvise on the file handles
126  * in the underlying file system. However, if this is done on every read/write,
127  * the fadvises cause a very significant slowdown in tests (specifically fio
128  * seq_write). So call fadvise on the file handles with the most reads/writes
129  * only after a threshold is passed.
130  */
131 class FAdviser {
132   public:
FAdviser()133     FAdviser() : thread_(MessageLoop, this), total_size_(0) {}
134 
~FAdviser()135     ~FAdviser() {
136         SendMessage(Message::quit);
137         thread_.join();
138     }
139 
Record(int fd,size_t size)140     void Record(int fd, size_t size) { SendMessage(Message::record, fd, size); }
141 
Close(int fd)142     void Close(int fd) { SendMessage(Message::close, fd); }
143 
144   private:
145     struct Message {
146         enum Type { record, close, quit };
147         Type type;
148         int fd;
149         size_t size;
150     };
151 
RecordImpl(int fd,size_t size)152     void RecordImpl(int fd, size_t size) {
153         total_size_ += size;
154 
155         // Find or create record in files_
156         // Remove record from sizes_ if it exists, adjusting size appropriately
157         auto file = files_.find(fd);
158         if (file != files_.end()) {
159             auto old_size = file->second;
160             size += old_size->first;
161             sizes_.erase(old_size);
162         } else {
163             file = files_.insert(Files::value_type(fd, sizes_.end())).first;
164         }
165 
166         // Now (re) insert record in sizes_
167         auto new_size = sizes_.insert(Sizes::value_type(size, fd));
168         file->second = new_size;
169 
170         if (total_size_ < threshold_) return;
171 
172         LOG(INFO) << "Threshold exceeded - fadvising " << total_size_;
173         while (!sizes_.empty() && total_size_ > target_) {
174             auto size = --sizes_.end();
175             total_size_ -= size->first;
176             posix_fadvise(size->second, 0, 0, POSIX_FADV_DONTNEED);
177             files_.erase(size->second);
178             sizes_.erase(size);
179         }
180         LOG(INFO) << "Threshold now " << total_size_;
181     }
182 
CloseImpl(int fd)183     void CloseImpl(int fd) {
184         auto file = files_.find(fd);
185         if (file == files_.end()) return;
186 
187         total_size_ -= file->second->first;
188         sizes_.erase(file->second);
189         files_.erase(file);
190     }
191 
MessageLoopImpl()192     void MessageLoopImpl() {
193         while (1) {
194             Message message;
195 
196             {
197                 std::unique_lock<std::mutex> lock(mutex_);
198                 cv_.wait(lock, [this] { return !queue_.empty(); });
199                 message = queue_.front();
200                 queue_.pop();
201             }
202 
203             switch (message.type) {
204                 case Message::record:
205                     RecordImpl(message.fd, message.size);
206                     break;
207 
208                 case Message::close:
209                     CloseImpl(message.fd);
210                     break;
211 
212                 case Message::quit:
213                     return;
214             }
215         }
216     }
217 
MessageLoop(FAdviser * ptr)218     static int MessageLoop(FAdviser* ptr) {
219         ptr->MessageLoopImpl();
220         return 0;
221     }
222 
SendMessage(Message::Type type,int fd=-1,size_t size=0)223     void SendMessage(Message::Type type, int fd = -1, size_t size = 0) {
224         {
225             std::unique_lock<std::mutex> lock(mutex_);
226             Message message = {type, fd, size};
227             queue_.push(message);
228         }
229         cv_.notify_one();
230     }
231 
232     std::mutex mutex_;
233     std::condition_variable cv_;
234     std::queue<Message> queue_;
235     std::thread thread_;
236 
237     typedef std::multimap<size_t, int> Sizes;
238     typedef std::map<int, Sizes::iterator> Files;
239 
240     Files files_;
241     Sizes sizes_;
242     size_t total_size_;
243 
244     const size_t threshold_ = 64 * 1024 * 1024;
245     const size_t target_ = 32 * 1024 * 1024;
246 };
247 
248 /* Single FUSE mount */
249 struct fuse {
fusefuse250     explicit fuse(const std::string& _path, ino_t _ino)
251         : path(_path),
252           tracker(mediaprovider::fuse::NodeTracker(&lock)),
253           root(node::CreateRoot(_path, &lock, _ino, &tracker)),
254           mp(0),
255           zero_addr(0),
256           disable_dentry_cache(false),
257           passthrough(false) {}
258 
IsRootfuse259     inline bool IsRoot(const node* node) const { return node == root; }
260 
GetEffectiveRootPathfuse261     inline string GetEffectiveRootPath() {
262         if (android::base::StartsWith(path, "/storage/emulated")) {
263             return path + "/" + MY_USER_ID_STRING;
264         }
265         return path;
266     }
267 
GetTransformsDirfuse268     inline string GetTransformsDir() { return GetEffectiveRootPath() + "/.transforms"; }
269 
270     // Note that these two (FromInode / ToInode) conversion wrappers are required
271     // because fuse_lowlevel_ops documents that the root inode is always one
272     // (see FUSE_ROOT_ID in fuse_lowlevel.h). There are no particular requirements
273     // on any of the other inodes in the FS.
FromInodefuse274     inline node* FromInode(__u64 inode) {
275         if (inode == FUSE_ROOT_ID) {
276             return root;
277         }
278 
279         return node::FromInode(inode, &tracker);
280     }
281 
ToInodefuse282     inline __u64 ToInode(node* node) const {
283         if (IsRoot(node)) {
284             return FUSE_ROOT_ID;
285         }
286 
287         return node::ToInode(node);
288     }
289 
290     std::recursive_mutex lock;
291     const string path;
292     // The Inode tracker associated with this FUSE instance.
293     mediaprovider::fuse::NodeTracker tracker;
294     node* const root;
295     struct fuse_session* se;
296 
297     /*
298      * Used to make JNI calls to MediaProvider.
299      * Responsibility of freeing this object falls on corresponding
300      * FuseDaemon object.
301      */
302     mediaprovider::fuse::MediaProviderWrapper* mp;
303 
304     /*
305      * Points to a range of zeroized bytes, used by pf_read to represent redacted ranges.
306      * The memory is read only and should never be modified.
307      */
308     /* const */ char* zero_addr;
309 
310     FAdviser fadviser;
311 
312     std::atomic_bool* active;
313     std::atomic_bool disable_dentry_cache;
314     std::atomic_bool passthrough;
315     // FUSE device id.
316     std::atomic_uint dev;
317 };
318 
319 enum class FuseOp { lookup, readdir, mknod, mkdir, create };
320 
get_name(node * n)321 static inline string get_name(node* n) {
322     if (n) {
323         std::string name = IS_OS_DEBUGABLE ? "real_path: " + n->BuildPath() + " " : "";
324         name += "node_path: " + n->BuildSafePath();
325         return name;
326     }
327     return "?";
328 }
329 
ptr_to_id(void * ptr)330 static inline __u64 ptr_to_id(void* ptr) {
331     return (__u64)(uintptr_t) ptr;
332 }
333 
334 /*
335  * Set an F_RDLCK or F_WRLCKK on fd with fcntl(2).
336  *
337  * This is called before the MediaProvider returns fd from the lower file
338  * system to an app over the ContentResolver interface. This allows us
339  * check with is_file_locked if any reference to that fd is still open.
340  */
set_file_lock(int fd,bool for_read,const std::string & path)341 static int set_file_lock(int fd, bool for_read, const std::string& path) {
342     std::string lock_str = (for_read ? "read" : "write");
343 
344     struct flock fl{};
345     fl.l_type = for_read ? F_RDLCK : F_WRLCK;
346     fl.l_whence = SEEK_SET;
347 
348     int res = fcntl(fd, F_OFD_SETLK, &fl);
349     if (res) {
350         PLOG(WARNING) << "Failed to set lock: " << lock_str;
351         return res;
352     }
353     return res;
354 }
355 
356 /*
357  * Check if an F_RDLCK or F_WRLCK is set on fd with fcntl(2).
358  *
359  * This is used to determine if the MediaProvider has given an fd to the lower fs to an app over
360  * the ContentResolver interface. Before that happens, we always call set_file_lock on the file
361  * allowing us to know if any reference to that fd is still open here.
362  *
363  * Returns true if fd may have a lock, false otherwise
364  */
is_file_locked(int fd,const std::string & path)365 static bool is_file_locked(int fd, const std::string& path) {
366     struct flock fl{};
367     fl.l_type = F_WRLCK;
368     fl.l_whence = SEEK_SET;
369 
370     int res = fcntl(fd, F_OFD_GETLK, &fl);
371     if (res) {
372         PLOG(WARNING) << "Failed to check lock";
373         // Assume worst
374         return true;
375     }
376     bool locked = fl.l_type != F_UNLCK;
377     return locked;
378 }
379 
get_fuse(fuse_req_t req)380 static struct fuse* get_fuse(fuse_req_t req) {
381     return reinterpret_cast<struct fuse*>(fuse_req_userdata(req));
382 }
383 
is_package_owned_path(const string & path,const string & fuse_path)384 static bool is_package_owned_path(const string& path, const string& fuse_path) {
385     if (path.rfind(fuse_path, 0) != 0) {
386         return false;
387     }
388     return std::regex_match(path, PATTERN_OWNED_PATH);
389 }
390 
391 // See fuse_lowlevel.h fuse_lowlevel_notify_inval_entry for how to call this safetly without
392 // deadlocking the kernel
fuse_inval(fuse_session * se,fuse_ino_t parent_ino,fuse_ino_t child_ino,const string & child_name,const string & path)393 static void fuse_inval(fuse_session* se, fuse_ino_t parent_ino, fuse_ino_t child_ino,
394                        const string& child_name, const string& path) {
395     if (mediaprovider::fuse::containsMount(path, MY_USER_ID_STRING)) {
396         LOG(WARNING) << "Ignoring attempt to invalidate dentry for FUSE mounts";
397         return;
398     }
399 
400     if (fuse_lowlevel_notify_inval_entry(se, parent_ino, child_name.c_str(), child_name.size())) {
401         // Invalidating the dentry can fail if there's no dcache entry, however, there may still
402         // be cached attributes, so attempt to invalidate those by invalidating the inode
403         fuse_lowlevel_notify_inval_inode(se, child_ino, 0, 0);
404     }
405 }
406 
get_entry_timeout(const string & path,node * node,struct fuse * fuse)407 static double get_entry_timeout(const string& path, node* node, struct fuse* fuse) {
408     string media_path = fuse->GetEffectiveRootPath() + "/Android/media";
409     if (fuse->disable_dentry_cache || node->ShouldInvalidate() ||
410         is_package_owned_path(path, fuse->path) || android::base::StartsWith(path, media_path)) {
411         // We set dentry timeout to 0 for the following reasons:
412         // 1. The dentry cache was completely disabled
413         // 2.1 Case-insensitive lookups need to invalidate other case-insensitive dentry matches
414         // 2.2 Nodes supporting transforms need to be invalidated, so that subsequent lookups by a
415         // uid requiring a transform is guaranteed to come to the FUSE daemon.
416         // 3. With app data isolation enabled, app A should not guess existence of app B from the
417         // Android/{data,obb}/<package> paths, hence we prevent the kernel from caching that
418         // information.
419         // 4. Installd might delete Android/media/<package> dirs when app data is cleared.
420         // This can leave a stale entry in the kernel dcache, and break subsequent creation of the
421         // dir via FUSE.
422         return 0;
423     }
424     return std::numeric_limits<double>::max();
425 }
426 
get_path(node * node)427 static std::string get_path(node* node) {
428     const string& io_path = node->GetIoPath();
429     return io_path.empty() ? node->BuildPath() : io_path;
430 }
431 
432 // Returns true if the path resides under .transforms/synthetic.
433 // NOTE: currently only file paths corresponding to redacted URIs reside under this folder. The path
434 // itself never exists and just a link for transformation.
is_synthetic_path(const string & path,struct fuse * fuse)435 static inline bool is_synthetic_path(const string& path, struct fuse* fuse) {
436     return android::base::StartsWithIgnoreCase(
437             path, fuse->GetTransformsDir() + "/" + TRANSFORM_SYNTHETIC_DIR);
438 }
439 
is_transcode_supported_path(const string & path,struct fuse * fuse)440 static inline bool is_transcode_supported_path(const string& path, struct fuse* fuse) {
441     // Keep in sync with MediaProvider#supportsTranscode
442     return android::base::EndsWithIgnoreCase(path, ".mp4") &&
443            android::base::StartsWithIgnoreCase(path,
444                                                fuse->GetEffectiveRootPath() + "/dcim/camera/");
445 }
446 
is_transforms_dir_path(const string & path,struct fuse * fuse)447 static inline bool is_transforms_dir_path(const string& path, struct fuse* fuse) {
448     return android::base::StartsWithIgnoreCase(path, fuse->GetTransformsDir());
449 }
450 
validate_node_path(const std::string & path,const std::string & name,fuse_req_t req,int * error_code,struct fuse_entry_param * e,const FuseOp op)451 static std::unique_ptr<mediaprovider::fuse::FileLookupResult> validate_node_path(
452         const std::string& path, const std::string& name, fuse_req_t req, int* error_code,
453         struct fuse_entry_param* e, const FuseOp op) {
454     struct fuse* fuse = get_fuse(req);
455     const struct fuse_ctx* ctx = fuse_req_ctx(req);
456     memset(e, 0, sizeof(*e));
457 
458     const bool synthetic_path = is_synthetic_path(path, fuse);
459     if (lstat(path.c_str(), &e->attr) < 0 && !(op == FuseOp::lookup && synthetic_path)) {
460         *error_code = errno;
461         return nullptr;
462     }
463 
464     if (is_transforms_dir_path(path, fuse)) {
465         if (op == FuseOp::lookup) {
466             // Lookups are only allowed under .transforms/synthetic dir
467             if (!(android::base::EqualsIgnoreCase(path, fuse->GetTransformsDir()) ||
468                   android::base::StartsWithIgnoreCase(
469                           path, fuse->GetTransformsDir() + "/" + TRANSFORM_SYNTHETIC_DIR))) {
470                 *error_code = ENONET;
471                 return nullptr;
472             }
473         } else {
474             // user-code is only allowed to make lookups under .transforms dir, and that too only
475             // under .transforms/synthetic dir
476             *error_code = ENOENT;
477             return nullptr;
478         }
479     }
480 
481     if (S_ISDIR(e->attr.st_mode)) {
482         // now that we have reached this point, ops on directories are safe and require no
483         // transformation.
484         return std::make_unique<mediaprovider::fuse::FileLookupResult>(0, 0, 0, true, false, "");
485     }
486 
487     if (!synthetic_path && !is_transcode_supported_path(path, fuse)) {
488         // Transforms are only supported for synthetic or transcode-supported paths
489         return std::make_unique<mediaprovider::fuse::FileLookupResult>(0, 0, 0, true, false, "");
490     }
491 
492     // Handle potential file transforms
493     std::unique_ptr<mediaprovider::fuse::FileLookupResult> file_lookup_result =
494             fuse->mp->FileLookup(path, req->ctx.uid, req->ctx.pid);
495 
496     if (!file_lookup_result) {
497         // Fail lookup if we can't fetch FileLookupResult for path
498         LOG(WARNING) << "Failed to fetch FileLookupResult for " << path;
499         *error_code = ENOENT;
500         return nullptr;
501     }
502 
503     const string& io_path = file_lookup_result->io_path;
504     // Update size with io_path size if io_path is not same as path
505     if (!io_path.empty() && (io_path != path) && (lstat(io_path.c_str(), &e->attr) < 0)) {
506         *error_code = errno;
507         return nullptr;
508     }
509 
510     return file_lookup_result;
511 }
512 
make_node_entry(fuse_req_t req,node * parent,const string & name,const string & path,struct fuse_entry_param * e,int * error_code,const FuseOp op)513 static node* make_node_entry(fuse_req_t req, node* parent, const string& name, const string& path,
514                              struct fuse_entry_param* e, int* error_code, const FuseOp op) {
515     struct fuse* fuse = get_fuse(req);
516     const struct fuse_ctx* ctx = fuse_req_ctx(req);
517     node* node;
518 
519     memset(e, 0, sizeof(*e));
520 
521     std::unique_ptr<mediaprovider::fuse::FileLookupResult> file_lookup_result =
522             validate_node_path(path, name, req, error_code, e, op);
523     if (!file_lookup_result) {
524         // Fail lookup if we can't validate |path, |errno| would have already been set
525         return nullptr;
526     }
527 
528     const bool should_invalidate = file_lookup_result->transforms_supported;
529     const bool transforms_complete = file_lookup_result->transforms_complete;
530     const int transforms = file_lookup_result->transforms;
531     const int transforms_reason = file_lookup_result->transforms_reason;
532     const string& io_path = file_lookup_result->io_path;
533 
534     node = parent->LookupChildByName(name, true /* acquire */, transforms);
535     if (!node) {
536         ino_t ino = e->attr.st_ino;
537         node = ::node::Create(parent, name, io_path, should_invalidate, transforms_complete,
538                               transforms, transforms_reason, &fuse->lock, ino, &fuse->tracker);
539     } else if (!mediaprovider::fuse::containsMount(path, std::to_string(getuid() / PER_USER_RANGE))) {
540         // Only invalidate a path if it does not contain mount.
541         // Invalidate both names to ensure there's no dentry left in the kernel after the following
542         // operations:
543         // 1) touch foo, touch FOO, unlink *foo*
544         // 2) touch foo, touch FOO, unlink *FOO*
545         // Invalidating lookup_name fixes (1) and invalidating node_name fixes (2)
546         // SetShouldInvalidate invalidates lookup_name by using 0 timeout below and we explicitly
547         // invalidate node_name if different case
548         // Note that we invalidate async otherwise we will deadlock the kernel
549         if (name != node->GetName()) {
550             // Record that we have made a case insensitive lookup, this allows us invalidate nodes
551             // correctly on subsequent lookups for the case of |node|
552             node->SetShouldInvalidate();
553 
554             // Make copies of the node name and path so we're not attempting to acquire
555             // any node locks from the invalidation thread. Depending on timing, we may end
556             // up invalidating the wrong inode but that shouldn't result in correctness issues.
557             const fuse_ino_t parent_ino = fuse->ToInode(parent);
558             const fuse_ino_t child_ino = fuse->ToInode(node);
559             const std::string& node_name = node->GetName();
560             std::thread t([=]() { fuse_inval(fuse->se, parent_ino, child_ino, node_name, path); });
561             t.detach();
562         }
563 
564         // This updated value allows us correctly decide if to keep_cache and use direct_io during
565         // FUSE_OPEN. Between the last lookup and this lookup, we might have deleted a cached
566         // transcoded file on the lower fs. A subsequent transcode at FUSE_READ should ensure we
567         // don't reuse any stale transcode page cache content.
568         node->SetTransformsComplete(transforms_complete);
569     }
570     TRACE_NODE(node, req);
571 
572     // This FS is not being exported via NFS so just a fixed generation number
573     // for now. If we do need this, we need to increment the generation ID each
574     // time the fuse daemon restarts because that's what it takes for us to
575     // reuse inode numbers.
576     e->generation = 0;
577     e->ino = fuse->ToInode(node);
578     e->entry_timeout = get_entry_timeout(path, node, fuse);
579     e->attr_timeout = std::numeric_limits<double>::max();
580     return node;
581 }
582 
is_requesting_write(int flags)583 static inline bool is_requesting_write(int flags) {
584     return flags & (O_WRONLY | O_RDWR);
585 }
586 
587 namespace mediaprovider {
588 namespace fuse {
589 
590 /**
591  * Function implementations
592  *
593  * These implement the various functions in fuse_lowlevel_ops
594  *
595  */
596 
pf_init(void * userdata,struct fuse_conn_info * conn)597 static void pf_init(void* userdata, struct fuse_conn_info* conn) {
598     struct fuse* fuse = reinterpret_cast<struct fuse*>(userdata);
599 
600     // We don't want a getattr request with every read request
601     conn->want &= ~FUSE_CAP_AUTO_INVAL_DATA & ~FUSE_CAP_READDIRPLUS_AUTO;
602     unsigned mask = (FUSE_CAP_SPLICE_WRITE | FUSE_CAP_SPLICE_MOVE | FUSE_CAP_SPLICE_READ |
603                      FUSE_CAP_ASYNC_READ | FUSE_CAP_ATOMIC_O_TRUNC | FUSE_CAP_WRITEBACK_CACHE |
604                      FUSE_CAP_EXPORT_SUPPORT | FUSE_CAP_FLOCK_LOCKS);
605 
606     bool disable_splice_write = false;
607     if (fuse->passthrough) {
608         if (conn->capable & FUSE_CAP_PASSTHROUGH) {
609             mask |= FUSE_CAP_PASSTHROUGH;
610 
611             // SPLICE_WRITE seems to cause linux kernel cache corruption with passthrough enabled.
612             // It is still under investigation but while running
613             // ScopedStorageDeviceTest#testAccessMediaLocationInvalidation, we notice test flakes
614             // of about 1/20 for the following reason:
615             // 1. App without ACCESS_MEDIA_LOCATION permission reads redacted bytes via FUSE cache
616             // 2. App with ACCESS_MEDIA_LOCATION permission reads non-redacted bytes via passthrough
617             // cache
618             // (2) fails because bytes from (1) sneak into the passthrough cache??
619             // To workaround, we disable splice for write when passthrough is enabled.
620             // This shouldn't have any performance regression if comparing passthrough devices to
621             // no-passthrough devices for the following reasons:
622             // 1. No-op for no-passthrough devices
623             // 2. Passthrough devices
624             //   a. Files not requiring redaction use passthrough which bypasses FUSE_READ entirely
625             //   b. Files requiring redaction are still faster than no-passthrough devices that use
626             //      direct_io
627             disable_splice_write = true;
628         } else {
629             LOG(WARNING) << "Passthrough feature not supported by the kernel";
630             fuse->passthrough = false;
631         }
632     }
633 
634     conn->want |= conn->capable & mask;
635     if (disable_splice_write) {
636         conn->want &= ~FUSE_CAP_SPLICE_WRITE;
637     }
638 
639     conn->max_read = MAX_READ_SIZE;
640 
641     fuse->active->store(true, std::memory_order_release);
642 }
643 
pf_destroy(void * userdata)644 static void pf_destroy(void* userdata) {
645     struct fuse* fuse = reinterpret_cast<struct fuse*>(userdata);
646     LOG(INFO) << "DESTROY " << fuse->path;
647 
648     node::DeleteTree(fuse->root);
649 }
650 
651 // Return true if the path is accessible for that uid.
is_app_accessible_path(MediaProviderWrapper * mp,const string & path,uid_t uid)652 static bool is_app_accessible_path(MediaProviderWrapper* mp, const string& path, uid_t uid) {
653     if (uid < AID_APP_START || uid == MY_UID) {
654         return true;
655     }
656 
657     if (path == "/storage/emulated") {
658         // Apps should never refer to /storage/emulated - they should be using the user-spcific
659         // subdirs, eg /storage/emulated/0
660         return false;
661     }
662 
663     std::smatch match;
664     if (std::regex_match(path, match, PATTERN_OWNED_PATH)) {
665         const std::string& pkg = match[1];
666         // .nomedia is not a valid package. .nomedia always exists in /Android/data directory,
667         // and it's not an external file/directory of any package
668         if (pkg == ".nomedia") {
669             return true;
670         }
671         if (android::base::StartsWith(path, "/storage/emulated")) {
672             // Emulated storage bind-mounts app-private data directories, and so these
673             // should not be accessible through FUSE anyway.
674             LOG(WARNING) << "Rejected access to app-private dir on FUSE: " << path
675                          << " from uid: " << uid;
676             return false;
677         }
678         if (!mp->isUidAllowedAccessToDataOrObbPath(uid, path)) {
679             PLOG(WARNING) << "Invalid other package file access from " << uid << "(: " << path;
680             return false;
681         }
682     }
683     return true;
684 }
685 
686 static std::regex storage_emulated_regex("^\\/storage\\/emulated\\/([0-9]+)");
do_lookup(fuse_req_t req,fuse_ino_t parent,const char * name,struct fuse_entry_param * e,int * error_code,const FuseOp op)687 static node* do_lookup(fuse_req_t req, fuse_ino_t parent, const char* name,
688                        struct fuse_entry_param* e, int* error_code, const FuseOp op) {
689     struct fuse* fuse = get_fuse(req);
690     node* parent_node = fuse->FromInode(parent);
691     if (!parent_node) {
692         *error_code = ENOENT;
693         return nullptr;
694     }
695     string parent_path = parent_node->BuildPath();
696     // We should always allow lookups on the root, because failing them could cause
697     // bind mounts to be invalidated.
698     if (!fuse->IsRoot(parent_node) && !is_app_accessible_path(fuse->mp, parent_path, req->ctx.uid)) {
699         *error_code = ENOENT;
700         return nullptr;
701     }
702 
703     TRACE_NODE(parent_node, req);
704 
705     const string child_path = parent_path + "/" + name;
706     std::smatch match;
707     std::regex_search(child_path, match, storage_emulated_regex);
708 
709     // Ensure the FuseDaemon user id matches the user id or cross-user lookups are allowed in
710     // requested path
711     if (match.size() == 2 && MY_USER_ID_STRING != match[1].str()) {
712         // If user id mismatch, check cross-user lookups
713         long userId = strtol(match[1].str().c_str(), nullptr, 10);
714         if (userId < 0 || userId > MAX_USER_ID ||
715             !fuse->mp->ShouldAllowLookup(req->ctx.uid, userId)) {
716             *error_code = EACCES;
717             return nullptr;
718         }
719     }
720 
721     return make_node_entry(req, parent_node, name, child_path, e, error_code, op);
722 }
723 
pf_lookup(fuse_req_t req,fuse_ino_t parent,const char * name)724 static void pf_lookup(fuse_req_t req, fuse_ino_t parent, const char* name) {
725     ATRACE_CALL();
726     struct fuse_entry_param e;
727 
728     int error_code = 0;
729     if (do_lookup(req, parent, name, &e, &error_code, FuseOp::lookup)) {
730         fuse_reply_entry(req, &e);
731     } else {
732         CHECK(error_code != 0);
733         fuse_reply_err(req, error_code);
734     }
735 }
736 
do_forget(fuse_req_t req,struct fuse * fuse,fuse_ino_t ino,uint64_t nlookup)737 static void do_forget(fuse_req_t req, struct fuse* fuse, fuse_ino_t ino, uint64_t nlookup) {
738     node* node = fuse->FromInode(ino);
739     TRACE_NODE(node, req);
740     if (node) {
741         // This is a narrowing conversion from an unsigned 64bit to a 32bit value. For
742         // some reason we only keep 32 bit refcounts but the kernel issues
743         // forget requests with a 64 bit counter.
744         node->Release(static_cast<uint32_t>(nlookup));
745     }
746 }
747 
pf_forget(fuse_req_t req,fuse_ino_t ino,uint64_t nlookup)748 static void pf_forget(fuse_req_t req, fuse_ino_t ino, uint64_t nlookup) {
749     // Always allow to forget so no need to check is_app_accessible_path()
750     ATRACE_CALL();
751     node* node;
752     struct fuse* fuse = get_fuse(req);
753 
754     do_forget(req, fuse, ino, nlookup);
755     fuse_reply_none(req);
756 }
757 
pf_forget_multi(fuse_req_t req,size_t count,struct fuse_forget_data * forgets)758 static void pf_forget_multi(fuse_req_t req,
759                             size_t count,
760                             struct fuse_forget_data* forgets) {
761     ATRACE_CALL();
762     struct fuse* fuse = get_fuse(req);
763 
764     for (int i = 0; i < count; i++) {
765         do_forget(req, fuse, forgets[i].ino, forgets[i].nlookup);
766     }
767     fuse_reply_none(req);
768 }
769 
pf_fallocate(fuse_req_t req,fuse_ino_t ino,int mode,off_t offset,off_t length,fuse_file_info * fi)770 static void pf_fallocate(fuse_req_t req, fuse_ino_t ino, int mode, off_t offset, off_t length,
771                          fuse_file_info* fi) {
772     ATRACE_CALL();
773     struct fuse* fuse = get_fuse(req);
774 
775     handle* h = reinterpret_cast<handle*>(fi->fh);
776     auto err = fallocate(h->fd, mode, offset, length);
777     fuse_reply_err(req, err ? errno : 0);
778 }
779 
pf_getattr(fuse_req_t req,fuse_ino_t ino,struct fuse_file_info * fi)780 static void pf_getattr(fuse_req_t req,
781                        fuse_ino_t ino,
782                        struct fuse_file_info* fi) {
783     ATRACE_CALL();
784     struct fuse* fuse = get_fuse(req);
785     node* node = fuse->FromInode(ino);
786     if (!node) {
787         fuse_reply_err(req, ENOENT);
788         return;
789     }
790     const string& path = get_path(node);
791     if (!is_app_accessible_path(fuse->mp, path, req->ctx.uid)) {
792         fuse_reply_err(req, ENOENT);
793         return;
794     }
795     TRACE_NODE(node, req);
796 
797     struct stat s;
798     memset(&s, 0, sizeof(s));
799     if (lstat(path.c_str(), &s) < 0) {
800         fuse_reply_err(req, errno);
801     } else {
802         fuse_reply_attr(req, &s, std::numeric_limits<double>::max());
803     }
804 }
805 
pf_setattr(fuse_req_t req,fuse_ino_t ino,struct stat * attr,int to_set,struct fuse_file_info * fi)806 static void pf_setattr(fuse_req_t req,
807                        fuse_ino_t ino,
808                        struct stat* attr,
809                        int to_set,
810                        struct fuse_file_info* fi) {
811     ATRACE_CALL();
812     struct fuse* fuse = get_fuse(req);
813     node* node = fuse->FromInode(ino);
814     if (!node) {
815         fuse_reply_err(req, ENOENT);
816         return;
817     }
818     const string& path = get_path(node);
819     if (!is_app_accessible_path(fuse->mp, path, req->ctx.uid)) {
820         fuse_reply_err(req, ENOENT);
821         return;
822     }
823 
824     int fd = -1;
825     if (fi) {
826         // If we have a file_info, setattr was called with an fd so use the fd instead of path
827         handle* h = reinterpret_cast<handle*>(fi->fh);
828         fd = h->fd;
829     } else {
830         const struct fuse_ctx* ctx = fuse_req_ctx(req);
831         std::unique_ptr<FileOpenResult> result = fuse->mp->OnFileOpen(
832                 path, path, ctx->uid, ctx->pid, node->GetTransformsReason(), true /* for_write */,
833                 false /* redact */, false /* log_transforms_metrics */);
834 
835         if (!result) {
836             fuse_reply_err(req, EFAULT);
837             return;
838         }
839 
840         if (result->status) {
841             fuse_reply_err(req, EACCES);
842             return;
843         }
844     }
845     struct timespec times[2];
846     TRACE_NODE(node, req);
847 
848     /* XXX: incomplete implementation on purpose.
849      * chmod/chown should NEVER be implemented.*/
850 
851     if ((to_set & FUSE_SET_ATTR_SIZE)) {
852         int res = 0;
853         if (fd == -1) {
854             res = truncate64(path.c_str(), attr->st_size);
855         } else {
856             res = ftruncate64(fd, attr->st_size);
857         }
858 
859         if (res < 0) {
860             fuse_reply_err(req, errno);
861             return;
862         }
863     }
864 
865     /* Handle changing atime and mtime.  If FATTR_ATIME_and FATTR_ATIME_NOW
866      * are both set, then set it to the current time.  Else, set it to the
867      * time specified in the request.  Same goes for mtime.  Use utimensat(2)
868      * as it allows ATIME and MTIME to be changed independently, and has
869      * nanosecond resolution which fuse also has.
870      */
871     if (to_set & (FATTR_ATIME | FATTR_MTIME)) {
872         times[0].tv_nsec = UTIME_OMIT;
873         times[1].tv_nsec = UTIME_OMIT;
874         if (to_set & FATTR_ATIME) {
875             if (to_set & FATTR_ATIME_NOW) {
876                 times[0].tv_nsec = UTIME_NOW;
877             } else {
878                 times[0] = attr->st_atim;
879             }
880         }
881 
882         if (to_set & FATTR_MTIME) {
883             if (to_set & FATTR_MTIME_NOW) {
884                 times[1].tv_nsec = UTIME_NOW;
885             } else {
886                 times[1] = attr->st_mtim;
887             }
888         }
889 
890         TRACE_NODE(node, req);
891         int res = 0;
892         if (fd == -1) {
893             res = utimensat(-1, path.c_str(), times, 0);
894         } else {
895             res = futimens(fd, times);
896         }
897 
898         if (res < 0) {
899             fuse_reply_err(req, errno);
900             return;
901         }
902     }
903 
904     lstat(path.c_str(), attr);
905     fuse_reply_attr(req, attr, std::numeric_limits<double>::max());
906 }
907 
pf_canonical_path(fuse_req_t req,fuse_ino_t ino)908 static void pf_canonical_path(fuse_req_t req, fuse_ino_t ino)
909 {
910     struct fuse* fuse = get_fuse(req);
911     node* node = fuse->FromInode(ino);
912     const string& path = node ? get_path(node) : "";
913 
914     if (node && is_app_accessible_path(fuse->mp, path, req->ctx.uid)) {
915         // TODO(b/147482155): Check that uid has access to |path| and its contents
916         fuse_reply_canonical_path(req, path.c_str());
917         return;
918     }
919     fuse_reply_err(req, ENOENT);
920 }
921 
pf_mknod(fuse_req_t req,fuse_ino_t parent,const char * name,mode_t mode,dev_t rdev)922 static void pf_mknod(fuse_req_t req,
923                      fuse_ino_t parent,
924                      const char* name,
925                      mode_t mode,
926                      dev_t rdev) {
927     ATRACE_CALL();
928     struct fuse* fuse = get_fuse(req);
929     node* parent_node = fuse->FromInode(parent);
930     if (!parent_node) {
931         fuse_reply_err(req, ENOENT);
932         return;
933     }
934     string parent_path = parent_node->BuildPath();
935     if (!is_app_accessible_path(fuse->mp, parent_path, req->ctx.uid)) {
936         fuse_reply_err(req, ENOENT);
937         return;
938     }
939 
940     TRACE_NODE(parent_node, req);
941 
942     const string child_path = parent_path + "/" + name;
943 
944     mode = (mode & (~0777)) | 0664;
945     if (mknod(child_path.c_str(), mode, rdev) < 0) {
946         fuse_reply_err(req, errno);
947         return;
948     }
949 
950     int error_code = 0;
951     struct fuse_entry_param e;
952     if (make_node_entry(req, parent_node, name, child_path, &e, &error_code, FuseOp::mknod)) {
953         fuse_reply_entry(req, &e);
954     } else {
955         CHECK(error_code != 0);
956         fuse_reply_err(req, error_code);
957     }
958 }
959 
pf_mkdir(fuse_req_t req,fuse_ino_t parent,const char * name,mode_t mode)960 static void pf_mkdir(fuse_req_t req,
961                      fuse_ino_t parent,
962                      const char* name,
963                      mode_t mode) {
964     ATRACE_CALL();
965     struct fuse* fuse = get_fuse(req);
966     node* parent_node = fuse->FromInode(parent);
967     if (!parent_node) {
968         fuse_reply_err(req, ENOENT);
969         return;
970     }
971     const struct fuse_ctx* ctx = fuse_req_ctx(req);
972     const string parent_path = parent_node->BuildPath();
973     if (!is_app_accessible_path(fuse->mp, parent_path, ctx->uid)) {
974         fuse_reply_err(req, ENOENT);
975         return;
976     }
977 
978     TRACE_NODE(parent_node, req);
979 
980     const string child_path = parent_path + "/" + name;
981 
982     int status = fuse->mp->IsCreatingDirAllowed(child_path, ctx->uid);
983     if (status) {
984         fuse_reply_err(req, status);
985         return;
986     }
987 
988     mode = (mode & (~0777)) | 0775;
989     if (mkdir(child_path.c_str(), mode) < 0) {
990         fuse_reply_err(req, errno);
991         return;
992     }
993 
994     int error_code = 0;
995     struct fuse_entry_param e;
996     if (make_node_entry(req, parent_node, name, child_path, &e, &error_code, FuseOp::mkdir)) {
997         fuse_reply_entry(req, &e);
998     } else {
999         CHECK(error_code != 0);
1000         fuse_reply_err(req, error_code);
1001     }
1002 }
1003 
pf_unlink(fuse_req_t req,fuse_ino_t parent,const char * name)1004 static void pf_unlink(fuse_req_t req, fuse_ino_t parent, const char* name) {
1005     ATRACE_CALL();
1006     struct fuse* fuse = get_fuse(req);
1007     node* parent_node = fuse->FromInode(parent);
1008     if (!parent_node) {
1009         fuse_reply_err(req, ENOENT);
1010         return;
1011     }
1012     const struct fuse_ctx* ctx = fuse_req_ctx(req);
1013     const string parent_path = parent_node->BuildPath();
1014     if (!is_app_accessible_path(fuse->mp, parent_path, ctx->uid)) {
1015         fuse_reply_err(req, ENOENT);
1016         return;
1017     }
1018 
1019     TRACE_NODE(parent_node, req);
1020 
1021     const string child_path = parent_path + "/" + name;
1022 
1023     int status = fuse->mp->DeleteFile(child_path, ctx->uid);
1024     if (status) {
1025         fuse_reply_err(req, status);
1026         return;
1027     }
1028 
1029     // TODO(b/169306422): Log each deleted node
1030     parent_node->SetDeletedForChild(name);
1031     fuse_reply_err(req, 0);
1032 }
1033 
pf_rmdir(fuse_req_t req,fuse_ino_t parent,const char * name)1034 static void pf_rmdir(fuse_req_t req, fuse_ino_t parent, const char* name) {
1035     ATRACE_CALL();
1036     struct fuse* fuse = get_fuse(req);
1037     node* parent_node = fuse->FromInode(parent);
1038     if (!parent_node) {
1039         fuse_reply_err(req, ENOENT);
1040         return;
1041     }
1042     const string parent_path = parent_node->BuildPath();
1043     if (!is_app_accessible_path(fuse->mp, parent_path, req->ctx.uid)) {
1044         fuse_reply_err(req, ENOENT);
1045         return;
1046     }
1047 
1048     if (is_transforms_dir_path(parent_path, fuse)) {
1049         // .transforms is a special daemon controlled dir so apps shouldn't be able to see it via
1050         // readdir, and any dir operations attempted on it should fail
1051         fuse_reply_err(req, ENOENT);
1052         return;
1053     }
1054 
1055     TRACE_NODE(parent_node, req);
1056 
1057     const string child_path = parent_path + "/" + name;
1058 
1059     int status = fuse->mp->IsDeletingDirAllowed(child_path, req->ctx.uid);
1060     if (status) {
1061         fuse_reply_err(req, status);
1062         return;
1063     }
1064 
1065     if (rmdir(child_path.c_str()) < 0) {
1066         fuse_reply_err(req, errno);
1067         return;
1068     }
1069 
1070     node* child_node = parent_node->LookupChildByName(name, false /* acquire */);
1071     TRACE_NODE(child_node, req);
1072     if (child_node) {
1073         child_node->SetDeleted();
1074     }
1075 
1076     fuse_reply_err(req, 0);
1077 }
1078 /*
1079 static void pf_symlink(fuse_req_t req, const char* link, fuse_ino_t parent,
1080                          const char* name)
1081 {
1082     cout << "TODO:" << __func__;
1083 }
1084 */
do_rename(fuse_req_t req,fuse_ino_t parent,const char * name,fuse_ino_t new_parent,const char * new_name,unsigned int flags)1085 static int do_rename(fuse_req_t req, fuse_ino_t parent, const char* name, fuse_ino_t new_parent,
1086                      const char* new_name, unsigned int flags) {
1087     ATRACE_CALL();
1088     struct fuse* fuse = get_fuse(req);
1089 
1090     if (flags != 0) {
1091         return EINVAL;
1092     }
1093 
1094     node* old_parent_node = fuse->FromInode(parent);
1095     if (!old_parent_node) return ENOENT;
1096     const struct fuse_ctx* ctx = fuse_req_ctx(req);
1097     const string old_parent_path = old_parent_node->BuildPath();
1098     if (!is_app_accessible_path(fuse->mp, old_parent_path, ctx->uid)) {
1099         return ENOENT;
1100     }
1101 
1102     if (is_transforms_dir_path(old_parent_path, fuse)) {
1103         // .transforms is a special daemon controlled dir so apps shouldn't be able to see it via
1104         // readdir, and any dir operations attempted on it should fail
1105         return ENOENT;
1106     }
1107 
1108     node* new_parent_node = fuse->FromInode(new_parent);
1109     if (!new_parent_node) return ENOENT;
1110     const string new_parent_path = new_parent_node->BuildPath();
1111     if (!is_app_accessible_path(fuse->mp, new_parent_path, ctx->uid)) {
1112         return ENOENT;
1113     }
1114 
1115     if (!old_parent_node || !new_parent_node) {
1116         return ENOENT;
1117     } else if (parent == new_parent && name == new_name) {
1118         // No rename required.
1119         return 0;
1120     }
1121 
1122     TRACE_NODE(old_parent_node, req);
1123     TRACE_NODE(new_parent_node, req);
1124 
1125     const string old_child_path = old_parent_path + "/" + name;
1126     const string new_child_path = new_parent_path + "/" + new_name;
1127 
1128     if (android::base::EqualsIgnoreCase(fuse->GetEffectiveRootPath() + "/android", old_child_path)) {
1129         // Prevent renaming Android/ dir since it contains bind-mounts on the primary volume
1130         return EACCES;
1131     }
1132 
1133     // TODO(b/147408834): Check ENOTEMPTY & EEXIST error conditions before JNI call.
1134     const int res = fuse->mp->Rename(old_child_path, new_child_path, req->ctx.uid);
1135     // TODO(b/145663158): Lookups can go out of sync if file/directory is actually moved but
1136     // EFAULT/EIO is reported due to JNI exception.
1137     if (res == 0) {
1138         // TODO(b/169306422): Log each renamed node
1139         old_parent_node->RenameChild(name, new_name, new_parent_node);
1140     }
1141     return res;
1142 }
1143 
pf_rename(fuse_req_t req,fuse_ino_t parent,const char * name,fuse_ino_t new_parent,const char * new_name,unsigned int flags)1144 static void pf_rename(fuse_req_t req, fuse_ino_t parent, const char* name, fuse_ino_t new_parent,
1145                       const char* new_name, unsigned int flags) {
1146     int res = do_rename(req, parent, name, new_parent, new_name, flags);
1147     fuse_reply_err(req, res);
1148 }
1149 
1150 /*
1151 static void pf_link(fuse_req_t req, fuse_ino_t ino, fuse_ino_t new_parent,
1152                       const char* new_name)
1153 {
1154     cout << "TODO:" << __func__;
1155 }
1156 */
1157 
create_handle_for_node(struct fuse * fuse,const string & path,int fd,uid_t uid,uid_t transforms_uid,node * node,const RedactionInfo * ri,int * keep_cache)1158 static handle* create_handle_for_node(struct fuse* fuse, const string& path, int fd, uid_t uid,
1159                                       uid_t transforms_uid, node* node, const RedactionInfo* ri,
1160                                       int* keep_cache) {
1161     std::lock_guard<std::recursive_mutex> guard(fuse->lock);
1162 
1163     bool redaction_needed = ri->isRedactionNeeded();
1164     handle* handle = nullptr;
1165     int transforms = node->GetTransforms();
1166     bool transforms_complete = node->IsTransformsComplete();
1167     if (transforms_uid > 0) {
1168         CHECK(transforms);
1169     }
1170 
1171     if (fuse->passthrough) {
1172         *keep_cache = transforms_complete;
1173         // We only enabled passthrough iff these 2 conditions hold
1174         // 1. Redaction is not needed
1175         // 2. Node transforms are completed, e.g transcoding.
1176         // (2) is important because we transcode lazily (on the first read) and with passthrough,
1177         // we will never get a read into the FUSE daemon, so passthrough would have returned
1178         // arbitrary bytes the first time around. However, if we ensure that transforms are
1179         // completed, then it's safe to use passthrough. Additionally, transcoded nodes never
1180         // require redaction so (2) implies (1)
1181         handle = new struct handle(fd, ri, true /* cached */,
1182                                    !redaction_needed && transforms_complete /* passthrough */, uid,
1183                                    transforms_uid);
1184     } else {
1185         // Without fuse->passthrough, we don't want to use the FUSE VFS cache in two cases:
1186         // 1. When redaction is needed because app A with EXIF access might access
1187         // a region that should have been redacted for app B without EXIF access, but app B on
1188         // a subsequent read, will be able to see the EXIF data because the read request for
1189         // that region will be served from cache and not get to the FUSE daemon
1190         // 2. When the file has a read or write lock on it. This means that the MediaProvider
1191         // has given an fd to the lower file system to an app. There are two cases where using
1192         // the cache in this case can be a problem:
1193         // a. Writing to a FUSE fd with caching enabled will use the write-back cache and a
1194         // subsequent read from the lower fs fd will not see the write.
1195         // b. Reading from a FUSE fd with caching enabled may not see the latest writes using
1196         // the lower fs fd because those writes did not go through the FUSE layer and reads from
1197         // FUSE after that write may be served from cache
1198         bool has_redacted = node->HasRedactedCache();
1199         bool is_redaction_change =
1200                 (redaction_needed && !has_redacted) || (!redaction_needed && has_redacted);
1201         bool is_cached_file_open = node->HasCachedHandle();
1202         bool direct_io = (is_cached_file_open && is_redaction_change) || is_file_locked(fd, path);
1203 
1204         if (!is_cached_file_open && is_redaction_change) {
1205             node->SetRedactedCache(redaction_needed);
1206             // Purges stale page cache before open
1207             *keep_cache = 0;
1208         } else {
1209             *keep_cache = transforms_complete;
1210         }
1211         handle = new struct handle(fd, ri, !direct_io /* cached */, false /* passthrough */, uid,
1212                                    transforms_uid);
1213     }
1214 
1215     node->AddHandle(handle);
1216     return handle;
1217 }
1218 
do_passthrough_enable(fuse_req_t req,struct fuse_file_info * fi,unsigned int fd)1219 bool do_passthrough_enable(fuse_req_t req, struct fuse_file_info* fi, unsigned int fd) {
1220     int passthrough_fh = fuse_passthrough_enable(req, fd);
1221 
1222     if (passthrough_fh <= 0) {
1223         return false;
1224     }
1225 
1226     fi->passthrough_fh = passthrough_fh;
1227     return true;
1228 }
1229 
pf_open(fuse_req_t req,fuse_ino_t ino,struct fuse_file_info * fi)1230 static void pf_open(fuse_req_t req, fuse_ino_t ino, struct fuse_file_info* fi) {
1231     ATRACE_CALL();
1232     struct fuse* fuse = get_fuse(req);
1233     node* node = fuse->FromInode(ino);
1234     if (!node) {
1235         fuse_reply_err(req, ENOENT);
1236         return;
1237     }
1238     const struct fuse_ctx* ctx = fuse_req_ctx(req);
1239     const string& io_path = get_path(node);
1240     const string& build_path = node->BuildPath();
1241     if (!is_app_accessible_path(fuse->mp, io_path, ctx->uid)) {
1242         fuse_reply_err(req, ENOENT);
1243         return;
1244     }
1245 
1246     bool for_write = is_requesting_write(fi->flags);
1247 
1248     if (for_write && node->GetTransforms()) {
1249         TRACE_NODE(node, req) << "write with transforms";
1250     } else {
1251         TRACE_NODE(node, req) << (for_write ? "write" : "read");
1252     }
1253 
1254     if (fi->flags & O_DIRECT) {
1255         fi->flags &= ~O_DIRECT;
1256         fi->direct_io = true;
1257     }
1258 
1259     // Force permission check with the build path because the MediaProvider database might not be
1260     // aware of the io_path
1261     // We don't redact if the caller was granted write permission for this file
1262     std::unique_ptr<FileOpenResult> result = fuse->mp->OnFileOpen(
1263             build_path, io_path, ctx->uid, ctx->pid, node->GetTransformsReason(), for_write,
1264             !for_write /* redact */, true /* log_transforms_metrics */);
1265     if (!result) {
1266         fuse_reply_err(req, EFAULT);
1267         return;
1268     }
1269 
1270     if (result->status) {
1271         fuse_reply_err(req, result->status);
1272         return;
1273     }
1274 
1275     // With the writeback cache enabled, FUSE may generate READ requests even for files that
1276     // were opened O_WRONLY; so make sure we open it O_RDWR instead.
1277     int open_flags = fi->flags;
1278     if (open_flags & O_WRONLY) {
1279         open_flags &= ~O_WRONLY;
1280         open_flags |= O_RDWR;
1281     }
1282 
1283     if (open_flags & O_APPEND) {
1284         open_flags &= ~O_APPEND;
1285     }
1286 
1287     const int fd = open(io_path.c_str(), open_flags);
1288     if (fd < 0) {
1289         fuse_reply_err(req, errno);
1290         return;
1291     }
1292 
1293     int keep_cache = 1;
1294     handle* h = create_handle_for_node(fuse, io_path, fd, result->uid, result->transforms_uid, node,
1295                                        result->redaction_info.release(), &keep_cache);
1296     fi->fh = ptr_to_id(h);
1297     fi->keep_cache = keep_cache;
1298     fi->direct_io = !h->cached;
1299 
1300     // TODO(b/173190192) ensuring that h->cached must be enabled in order to
1301     // user FUSE passthrough is a conservative rule and might be dropped as
1302     // soon as demonstrated its correctness.
1303     if (h->passthrough) {
1304         if (!do_passthrough_enable(req, fi, fd)) {
1305             // TODO: Should we crash here so we can find errors easily?
1306             PLOG(ERROR) << "Passthrough OPEN failed for " << io_path;
1307             fuse_reply_err(req, EFAULT);
1308             return;
1309         }
1310     }
1311 
1312     fuse_reply_open(req, fi);
1313 }
1314 
do_read(fuse_req_t req,size_t size,off_t off,struct fuse_file_info * fi)1315 static void do_read(fuse_req_t req, size_t size, off_t off, struct fuse_file_info* fi) {
1316     handle* h = reinterpret_cast<handle*>(fi->fh);
1317     struct fuse_bufvec buf = FUSE_BUFVEC_INIT(size);
1318 
1319     buf.buf[0].fd = h->fd;
1320     buf.buf[0].pos = off;
1321     buf.buf[0].flags =
1322             (enum fuse_buf_flags) (FUSE_BUF_IS_FD | FUSE_BUF_FD_SEEK);
1323 
1324     fuse_reply_data(req, &buf, (enum fuse_buf_copy_flags) 0);
1325 }
1326 
1327 /**
1328  * Sets the parameters for a fuse_buf that reads from memory, including flags.
1329  * Makes buf->mem point to an already mapped region of zeroized memory.
1330  * This memory is read only.
1331  */
create_mem_fuse_buf(size_t size,fuse_buf * buf,struct fuse * fuse)1332 static void create_mem_fuse_buf(size_t size, fuse_buf* buf, struct fuse* fuse) {
1333     buf->size = size;
1334     buf->mem = fuse->zero_addr;
1335     buf->flags = static_cast<fuse_buf_flags>(0 /*read from fuse_buf.mem*/);
1336     buf->pos = -1;
1337     buf->fd = -1;
1338 }
1339 
1340 /**
1341  * Sets the parameters for a fuse_buf that reads from file, including flags.
1342  */
create_file_fuse_buf(size_t size,off_t pos,int fd,fuse_buf * buf)1343 static void create_file_fuse_buf(size_t size, off_t pos, int fd, fuse_buf* buf) {
1344     buf->size = size;
1345     buf->fd = fd;
1346     buf->pos = pos;
1347     buf->flags = static_cast<fuse_buf_flags>(FUSE_BUF_IS_FD | FUSE_BUF_FD_SEEK);
1348     buf->mem = nullptr;
1349 }
1350 
do_read_with_redaction(fuse_req_t req,size_t size,off_t off,fuse_file_info * fi)1351 static void do_read_with_redaction(fuse_req_t req, size_t size, off_t off, fuse_file_info* fi) {
1352     handle* h = reinterpret_cast<handle*>(fi->fh);
1353 
1354     std::vector<ReadRange> ranges;
1355     h->ri->getReadRanges(off, size, &ranges);
1356 
1357     // As an optimization, return early if there are no ranges to redact.
1358     if (ranges.size() == 0) {
1359         do_read(req, size, off, fi);
1360         return;
1361     }
1362 
1363     const size_t num_bufs = ranges.size();
1364     auto bufvec_ptr = std::unique_ptr<fuse_bufvec, decltype(free)*>{
1365             reinterpret_cast<fuse_bufvec*>(
1366                     malloc(sizeof(fuse_bufvec) + (num_bufs - 1) * sizeof(fuse_buf))),
1367             free};
1368     fuse_bufvec& bufvec = *bufvec_ptr;
1369 
1370     // initialize bufvec
1371     bufvec.count = num_bufs;
1372     bufvec.idx = 0;
1373     bufvec.off = 0;
1374 
1375     for (int i = 0; i < num_bufs; ++i) {
1376         const ReadRange& range = ranges[i];
1377         if (range.is_redaction) {
1378             create_mem_fuse_buf(range.size, &(bufvec.buf[i]), get_fuse(req));
1379         } else {
1380             create_file_fuse_buf(range.size, range.start, h->fd, &(bufvec.buf[i]));
1381         }
1382     }
1383 
1384     fuse_reply_data(req, &bufvec, static_cast<fuse_buf_copy_flags>(0));
1385 }
1386 
pf_read(fuse_req_t req,fuse_ino_t ino,size_t size,off_t off,struct fuse_file_info * fi)1387 static void pf_read(fuse_req_t req, fuse_ino_t ino, size_t size, off_t off,
1388                     struct fuse_file_info* fi) {
1389     ATRACE_CALL();
1390     handle* h = reinterpret_cast<handle*>(fi->fh);
1391     struct fuse* fuse = get_fuse(req);
1392 
1393     node* node = fuse->FromInode(ino);
1394 
1395     if (!node->IsTransformsComplete()) {
1396         if (!fuse->mp->Transform(node->BuildPath(), node->GetIoPath(), node->GetTransforms(),
1397                                  node->GetTransformsReason(), req->ctx.uid, h->uid,
1398                                  h->transforms_uid)) {
1399             fuse_reply_err(req, EFAULT);
1400             return;
1401         }
1402         node->SetTransformsComplete(true);
1403     }
1404 
1405     fuse->fadviser.Record(h->fd, size);
1406 
1407     if (h->ri->isRedactionNeeded()) {
1408         do_read_with_redaction(req, size, off, fi);
1409     } else {
1410         do_read(req, size, off, fi);
1411     }
1412 }
1413 
1414 /*
1415 static void pf_write(fuse_req_t req, fuse_ino_t ino, const char* buf,
1416                        size_t size, off_t off, struct fuse_file_info* fi)
1417 {
1418     cout << "TODO:" << __func__;
1419 }
1420 */
1421 
pf_write_buf(fuse_req_t req,fuse_ino_t ino,struct fuse_bufvec * bufv,off_t off,struct fuse_file_info * fi)1422 static void pf_write_buf(fuse_req_t req,
1423                          fuse_ino_t ino,
1424                          struct fuse_bufvec* bufv,
1425                          off_t off,
1426                          struct fuse_file_info* fi) {
1427     ATRACE_CALL();
1428     handle* h = reinterpret_cast<handle*>(fi->fh);
1429     struct fuse_bufvec buf = FUSE_BUFVEC_INIT(fuse_buf_size(bufv));
1430     ssize_t size;
1431     struct fuse* fuse = get_fuse(req);
1432 
1433     buf.buf[0].fd = h->fd;
1434     buf.buf[0].pos = off;
1435     buf.buf[0].flags =
1436             (enum fuse_buf_flags) (FUSE_BUF_IS_FD | FUSE_BUF_FD_SEEK);
1437     size = fuse_buf_copy(&buf, bufv, (enum fuse_buf_copy_flags) 0);
1438 
1439     if (size < 0)
1440         fuse_reply_err(req, -size);
1441     else {
1442         // Execute Record *before* fuse_reply_write to avoid the following ordering:
1443         // fuse_reply_write -> pf_release (destroy handle) -> Record (use handle after free)
1444         fuse->fadviser.Record(h->fd, size);
1445         fuse_reply_write(req, size);
1446     }
1447 }
1448 // Haven't tested this one. Not sure what calls it.
1449 #if 0
1450 static void pf_copy_file_range(fuse_req_t req, fuse_ino_t ino_in,
1451                                  off_t off_in, struct fuse_file_info* fi_in,
1452                                  fuse_ino_t ino_out, off_t off_out,
1453                                  struct fuse_file_info* fi_out, size_t len,
1454                                  int flags)
1455 {
1456     handle* h_in = reinterpret_cast<handle *>(fi_in->fh);
1457     handle* h_out = reinterpret_cast<handle *>(fi_out->fh);
1458     struct fuse_bufvec buf_in = FUSE_BUFVEC_INIT(len);
1459     struct fuse_bufvec buf_out = FUSE_BUFVEC_INIT(len);
1460     ssize_t size;
1461 
1462     buf_in.buf[0].fd = h_in->fd;
1463     buf_in.buf[0].pos = off_in;
1464     buf_in.buf[0].flags = (enum fuse_buf_flags)(FUSE_BUF_IS_FD|FUSE_BUF_FD_SEEK);
1465 
1466     buf_out.buf[0].fd = h_out->fd;
1467     buf_out.buf[0].pos = off_out;
1468     buf_out.buf[0].flags = (enum fuse_buf_flags)(FUSE_BUF_IS_FD|FUSE_BUF_FD_SEEK);
1469     size = fuse_buf_copy(&buf_out, &buf_in, (enum fuse_buf_copy_flags) 0);
1470 
1471     if (size < 0) {
1472         fuse_reply_err(req, -size);
1473     }
1474 
1475     fuse_reply_write(req, size);
1476 }
1477 #endif
1478 
pf_release(fuse_req_t req,fuse_ino_t ino,struct fuse_file_info * fi)1479 static void pf_release(fuse_req_t req,
1480                        fuse_ino_t ino,
1481                        struct fuse_file_info* fi) {
1482     ATRACE_CALL();
1483     struct fuse* fuse = get_fuse(req);
1484 
1485     node* node = fuse->FromInode(ino);
1486     handle* h = reinterpret_cast<handle*>(fi->fh);
1487     TRACE_NODE(node, req);
1488 
1489     fuse->fadviser.Close(h->fd);
1490     if (node) {
1491         node->DestroyHandle(h);
1492     }
1493 
1494     fuse_reply_err(req, 0);
1495 }
1496 
do_sync_common(int fd,bool datasync)1497 static int do_sync_common(int fd, bool datasync) {
1498     int res = datasync ? fdatasync(fd) : fsync(fd);
1499 
1500     if (res == -1) return errno;
1501     return 0;
1502 }
1503 
pf_fsync(fuse_req_t req,fuse_ino_t ino,int datasync,struct fuse_file_info * fi)1504 static void pf_fsync(fuse_req_t req,
1505                      fuse_ino_t ino,
1506                      int datasync,
1507                      struct fuse_file_info* fi) {
1508     ATRACE_CALL();
1509     handle* h = reinterpret_cast<handle*>(fi->fh);
1510     int err = do_sync_common(h->fd, datasync);
1511 
1512     fuse_reply_err(req, err);
1513 }
1514 
pf_fsyncdir(fuse_req_t req,fuse_ino_t ino,int datasync,struct fuse_file_info * fi)1515 static void pf_fsyncdir(fuse_req_t req,
1516                         fuse_ino_t ino,
1517                         int datasync,
1518                         struct fuse_file_info* fi) {
1519     dirhandle* h = reinterpret_cast<dirhandle*>(fi->fh);
1520     int err = do_sync_common(dirfd(h->d), datasync);
1521 
1522     fuse_reply_err(req, err);
1523 }
1524 
pf_opendir(fuse_req_t req,fuse_ino_t ino,struct fuse_file_info * fi)1525 static void pf_opendir(fuse_req_t req,
1526                        fuse_ino_t ino,
1527                        struct fuse_file_info* fi) {
1528     ATRACE_CALL();
1529     struct fuse* fuse = get_fuse(req);
1530     node* node = fuse->FromInode(ino);
1531     if (!node) {
1532         fuse_reply_err(req, ENOENT);
1533         return;
1534     }
1535     const struct fuse_ctx* ctx = fuse_req_ctx(req);
1536     const string path = node->BuildPath();
1537     if (!is_app_accessible_path(fuse->mp, path, ctx->uid)) {
1538         fuse_reply_err(req, ENOENT);
1539         return;
1540     }
1541 
1542     TRACE_NODE(node, req);
1543 
1544     int status = fuse->mp->IsOpendirAllowed(path, ctx->uid, /* forWrite */ false);
1545     if (status) {
1546         fuse_reply_err(req, status);
1547         return;
1548     }
1549 
1550     DIR* dir = opendir(path.c_str());
1551     if (!dir) {
1552         fuse_reply_err(req, errno);
1553         return;
1554     }
1555 
1556     dirhandle* h = new dirhandle(dir);
1557     node->AddDirHandle(h);
1558 
1559     fi->fh = ptr_to_id(h);
1560     fuse_reply_open(req, fi);
1561 }
1562 
1563 #define READDIR_BUF 8192LU
1564 
do_readdir_common(fuse_req_t req,fuse_ino_t ino,size_t size,off_t off,struct fuse_file_info * fi,bool plus)1565 static void do_readdir_common(fuse_req_t req,
1566                               fuse_ino_t ino,
1567                               size_t size,
1568                               off_t off,
1569                               struct fuse_file_info* fi,
1570                               bool plus) {
1571     struct fuse* fuse = get_fuse(req);
1572     dirhandle* h = reinterpret_cast<dirhandle*>(fi->fh);
1573     size_t len = std::min<size_t>(size, READDIR_BUF);
1574     char buf[READDIR_BUF];
1575     size_t used = 0;
1576     std::shared_ptr<DirectoryEntry> de;
1577 
1578     struct fuse_entry_param e;
1579     size_t entry_size = 0;
1580 
1581     node* node = fuse->FromInode(ino);
1582     if (!node) {
1583         fuse_reply_err(req, ENOENT);
1584         return;
1585     }
1586     const string path = node->BuildPath();
1587     if (!is_app_accessible_path(fuse->mp, path, req->ctx.uid)) {
1588         fuse_reply_err(req, ENOENT);
1589         return;
1590     }
1591 
1592     TRACE_NODE(node, req);
1593     // Get all directory entries from MediaProvider on first readdir() call of
1594     // directory handle. h->next_off = 0 indicates that current readdir() call
1595     // is first readdir() call for the directory handle, Avoid multiple JNI calls
1596     // for single directory handle.
1597     if (h->next_off == 0) {
1598         h->de = fuse->mp->GetDirectoryEntries(req->ctx.uid, path, h->d);
1599     }
1600     // If the last entry in the previous readdir() call was rejected due to
1601     // buffer capacity constraints, update directory offset to start from
1602     // previously rejected entry. Directory offset can also change if there was
1603     // a seekdir() on the given directory handle.
1604     if (off != h->next_off) {
1605         h->next_off = off;
1606     }
1607     const int num_directory_entries = h->de.size();
1608     // Check for errors. Any error/exception occurred while obtaining directory
1609     // entries will be indicated by marking first directory entry name as empty
1610     // string. In the erroneous case corresponding d_type will hold error number.
1611     if (num_directory_entries && h->de[0]->d_name.empty()) {
1612         fuse_reply_err(req, h->de[0]->d_type);
1613         return;
1614     }
1615 
1616     while (h->next_off < num_directory_entries) {
1617         de = h->de[h->next_off];
1618         entry_size = 0;
1619         h->next_off++;
1620         if (plus) {
1621             int error_code = 0;
1622             if (do_lookup(req, ino, de->d_name.c_str(), &e, &error_code, FuseOp::readdir)) {
1623                 entry_size = fuse_add_direntry_plus(req, buf + used, len - used, de->d_name.c_str(),
1624                                                     &e, h->next_off);
1625             } else {
1626                 // Ignore lookup errors on
1627                 // 1. non-existing files returned from MediaProvider database.
1628                 // 2. path that doesn't match FuseDaemon UID and calling uid.
1629                 if (error_code == ENOENT || error_code == EPERM || error_code == EACCES
1630                     || error_code == EIO) continue;
1631                 fuse_reply_err(req, error_code);
1632                 return;
1633             }
1634         } else {
1635             // This should never happen because we have readdir_plus enabled without adaptive
1636             // readdir_plus, FUSE_CAP_READDIRPLUS_AUTO
1637             LOG(WARNING) << "Handling plain readdir for " << de->d_name << ". Invalid d_ino";
1638             e.attr.st_ino = FUSE_UNKNOWN_INO;
1639             e.attr.st_mode = de->d_type << 12;
1640             entry_size = fuse_add_direntry(req, buf + used, len - used, de->d_name.c_str(), &e.attr,
1641                                            h->next_off);
1642         }
1643         // If buffer in fuse_add_direntry[_plus] is not large enough then
1644         // the entry is not added to buffer but the size of the entry is still
1645         // returned. Check available buffer size + returned entry size is less
1646         // than actual buffer size to confirm entry is added to buffer.
1647         if (used + entry_size > len) {
1648             // When an entry is rejected, lookup called by readdir_plus will not be tracked by
1649             // kernel. Call forget on the rejected node to decrement the reference count.
1650             if (plus) {
1651                 do_forget(req, fuse, e.ino, 1);
1652             }
1653             break;
1654         }
1655         used += entry_size;
1656     }
1657     fuse_reply_buf(req, buf, used);
1658 }
1659 
pf_readdir(fuse_req_t req,fuse_ino_t ino,size_t size,off_t off,struct fuse_file_info * fi)1660 static void pf_readdir(fuse_req_t req, fuse_ino_t ino, size_t size, off_t off,
1661                        struct fuse_file_info* fi) {
1662     ATRACE_CALL();
1663     do_readdir_common(req, ino, size, off, fi, false);
1664 }
1665 
pf_readdirplus(fuse_req_t req,fuse_ino_t ino,size_t size,off_t off,struct fuse_file_info * fi)1666 static void pf_readdirplus(fuse_req_t req,
1667                            fuse_ino_t ino,
1668                            size_t size,
1669                            off_t off,
1670                            struct fuse_file_info* fi) {
1671     ATRACE_CALL();
1672     do_readdir_common(req, ino, size, off, fi, true);
1673 }
1674 
pf_releasedir(fuse_req_t req,fuse_ino_t ino,struct fuse_file_info * fi)1675 static void pf_releasedir(fuse_req_t req,
1676                           fuse_ino_t ino,
1677                           struct fuse_file_info* fi) {
1678     ATRACE_CALL();
1679     struct fuse* fuse = get_fuse(req);
1680 
1681     node* node = fuse->FromInode(ino);
1682 
1683     dirhandle* h = reinterpret_cast<dirhandle*>(fi->fh);
1684     TRACE_NODE(node, req);
1685     if (node) {
1686         node->DestroyDirHandle(h);
1687     }
1688 
1689     fuse_reply_err(req, 0);
1690 }
1691 
pf_statfs(fuse_req_t req,fuse_ino_t ino)1692 static void pf_statfs(fuse_req_t req, fuse_ino_t ino) {
1693     ATRACE_CALL();
1694     struct statvfs st;
1695     struct fuse* fuse = get_fuse(req);
1696 
1697     if (statvfs(fuse->root->GetName().c_str(), &st))
1698         fuse_reply_err(req, errno);
1699     else
1700         fuse_reply_statfs(req, &st);
1701 }
1702 /*
1703 static void pf_setxattr(fuse_req_t req, fuse_ino_t ino, const char* name,
1704                           const char* value, size_t size, int flags)
1705 {
1706     cout << "TODO:" << __func__;
1707 }
1708 
1709 static void pf_getxattr(fuse_req_t req, fuse_ino_t ino, const char* name,
1710                           size_t size)
1711 {
1712     cout << "TODO:" << __func__;
1713 }
1714 
1715 static void pf_listxattr(fuse_req_t req, fuse_ino_t ino, size_t size)
1716 {
1717     cout << "TODO:" << __func__;
1718 }
1719 
1720 static void pf_removexattr(fuse_req_t req, fuse_ino_t ino, const char* name)
1721 {
1722     cout << "TODO:" << __func__;
1723 }*/
1724 
pf_access(fuse_req_t req,fuse_ino_t ino,int mask)1725 static void pf_access(fuse_req_t req, fuse_ino_t ino, int mask) {
1726     ATRACE_CALL();
1727     struct fuse* fuse = get_fuse(req);
1728 
1729     node* node = fuse->FromInode(ino);
1730     if (!node) {
1731         fuse_reply_err(req, ENOENT);
1732         return;
1733     }
1734     const string path = node->BuildPath();
1735     if (path != "/storage/emulated" && !is_app_accessible_path(fuse->mp, path, req->ctx.uid)) {
1736         fuse_reply_err(req, ENOENT);
1737         return;
1738     }
1739     TRACE_NODE(node, req);
1740 
1741     // exists() checks are always allowed.
1742     if (mask == F_OK) {
1743         int res = access(path.c_str(), F_OK);
1744         fuse_reply_err(req, res ? errno : 0);
1745         return;
1746     }
1747     struct stat stat;
1748     if (lstat(path.c_str(), &stat)) {
1749         // File doesn't exist
1750         fuse_reply_err(req, ENOENT);
1751         return;
1752     }
1753 
1754     // For read and write permission checks we go to MediaProvider.
1755     int status = 0;
1756     bool for_write = mask & W_OK;
1757     bool is_directory = S_ISDIR(stat.st_mode);
1758     if (is_directory) {
1759         if (path == "/storage/emulated" && mask == X_OK) {
1760             // Special case for this path: apps should be allowed to enter it,
1761             // but not list directory contents (which would be user numbers).
1762             int res = access(path.c_str(), X_OK);
1763             fuse_reply_err(req, res ? errno : 0);
1764             return;
1765         }
1766         status = fuse->mp->IsOpendirAllowed(path, req->ctx.uid, for_write);
1767     } else {
1768         if (mask & X_OK) {
1769             // Fuse is mounted with MS_NOEXEC.
1770             fuse_reply_err(req, EACCES);
1771             return;
1772         }
1773 
1774         std::unique_ptr<FileOpenResult> result = fuse->mp->OnFileOpen(
1775                 path, path, req->ctx.uid, req->ctx.pid, node->GetTransformsReason(), for_write,
1776                 false /* redact */, false /* log_transforms_metrics */);
1777         if (!result) {
1778             status = EFAULT;
1779         } else if (result->status) {
1780             status = EACCES;
1781         }
1782     }
1783 
1784     fuse_reply_err(req, status);
1785 }
1786 
pf_create(fuse_req_t req,fuse_ino_t parent,const char * name,mode_t mode,struct fuse_file_info * fi)1787 static void pf_create(fuse_req_t req,
1788                       fuse_ino_t parent,
1789                       const char* name,
1790                       mode_t mode,
1791                       struct fuse_file_info* fi) {
1792     ATRACE_CALL();
1793     struct fuse* fuse = get_fuse(req);
1794     node* parent_node = fuse->FromInode(parent);
1795     if (!parent_node) {
1796         fuse_reply_err(req, ENOENT);
1797         return;
1798     }
1799     const string parent_path = parent_node->BuildPath();
1800     if (!is_app_accessible_path(fuse->mp, parent_path, req->ctx.uid)) {
1801         fuse_reply_err(req, ENOENT);
1802         return;
1803     }
1804 
1805     TRACE_NODE(parent_node, req);
1806 
1807     const string child_path = parent_path + "/" + name;
1808 
1809     int mp_return_code = fuse->mp->InsertFile(child_path.c_str(), req->ctx.uid);
1810     if (mp_return_code) {
1811         fuse_reply_err(req, mp_return_code);
1812         return;
1813     }
1814 
1815     // With the writeback cache enabled, FUSE may generate READ requests even for files that
1816     // were opened O_WRONLY; so make sure we open it O_RDWR instead.
1817     int open_flags = fi->flags;
1818     if (open_flags & O_WRONLY) {
1819         open_flags &= ~O_WRONLY;
1820         open_flags |= O_RDWR;
1821     }
1822 
1823     if (open_flags & O_APPEND) {
1824         open_flags &= ~O_APPEND;
1825     }
1826 
1827     mode = (mode & (~0777)) | 0664;
1828     int fd = open(child_path.c_str(), open_flags, mode);
1829     if (fd < 0) {
1830         int error_code = errno;
1831         // We've already inserted the file into the MP database before the
1832         // failed open(), so that needs to be rolled back here.
1833         fuse->mp->DeleteFile(child_path.c_str(), req->ctx.uid);
1834         fuse_reply_err(req, error_code);
1835         return;
1836     }
1837 
1838     int error_code = 0;
1839     struct fuse_entry_param e;
1840     node* node =
1841             make_node_entry(req, parent_node, name, child_path, &e, &error_code, FuseOp::create);
1842     TRACE_NODE(node, req);
1843     if (!node) {
1844         CHECK(error_code != 0);
1845         fuse_reply_err(req, error_code);
1846         return;
1847     }
1848 
1849     // Let MediaProvider know we've created a new file
1850     fuse->mp->OnFileCreated(child_path);
1851 
1852     // TODO(b/147274248): Assume there will be no EXIF to redact.
1853     // This prevents crashing during reads but can be a security hole if a malicious app opens an fd
1854     // to the file before all the EXIF content is written. We could special case reads before the
1855     // first close after a file has just been created.
1856     int keep_cache = 1;
1857     handle* h = create_handle_for_node(fuse, child_path, fd, req->ctx.uid, 0 /* transforms_uid */,
1858                                        node, new RedactionInfo(), &keep_cache);
1859     fi->fh = ptr_to_id(h);
1860     fi->keep_cache = keep_cache;
1861     fi->direct_io = !h->cached;
1862 
1863     // TODO(b/173190192) ensuring that h->cached must be enabled in order to
1864     // user FUSE passthrough is a conservative rule and might be dropped as
1865     // soon as demonstrated its correctness.
1866     if (h->passthrough) {
1867         if (!do_passthrough_enable(req, fi, fd)) {
1868             PLOG(ERROR) << "Passthrough CREATE failed for " << child_path;
1869             fuse_reply_err(req, EFAULT);
1870             return;
1871         }
1872     }
1873 
1874     fuse_reply_create(req, &e, fi);
1875 }
1876 /*
1877 static void pf_getlk(fuse_req_t req, fuse_ino_t ino,
1878                        struct fuse_file_info* fi, struct flock* lock)
1879 {
1880     cout << "TODO:" << __func__;
1881 }
1882 
1883 static void pf_setlk(fuse_req_t req, fuse_ino_t ino,
1884                        struct fuse_file_info* fi,
1885                        struct flock* lock, int sleep)
1886 {
1887     cout << "TODO:" << __func__;
1888 }
1889 
1890 static void pf_bmap(fuse_req_t req, fuse_ino_t ino, size_t blocksize,
1891                       uint64_t idx)
1892 {
1893     cout << "TODO:" << __func__;
1894 }
1895 
1896 static void pf_ioctl(fuse_req_t req, fuse_ino_t ino, unsigned int cmd,
1897                        void* arg, struct fuse_file_info* fi, unsigned flags,
1898                        const void* in_buf, size_t in_bufsz, size_t out_bufsz)
1899 {
1900     cout << "TODO:" << __func__;
1901 }
1902 
1903 static void pf_poll(fuse_req_t req, fuse_ino_t ino, struct fuse_file_info* fi,
1904                       struct fuse_pollhandle* ph)
1905 {
1906     cout << "TODO:" << __func__;
1907 }
1908 
1909 static void pf_retrieve_reply(fuse_req_t req, void* cookie, fuse_ino_t ino,
1910                                 off_t offset, struct fuse_bufvec* bufv)
1911 {
1912     cout << "TODO:" << __func__;
1913 }
1914 
1915 static void pf_flock(fuse_req_t req, fuse_ino_t ino,
1916                        struct fuse_file_info* fi, int op)
1917 {
1918     cout << "TODO:" << __func__;
1919 }
1920 
1921 static void pf_fallocate(fuse_req_t req, fuse_ino_t ino, int mode,
1922                        off_t offset, off_t length, struct fuse_file_info* fi)
1923 {
1924     cout << "TODO:" << __func__;
1925 }
1926 */
1927 
1928 static struct fuse_lowlevel_ops ops{
1929     .init = pf_init, .destroy = pf_destroy, .lookup = pf_lookup, .forget = pf_forget,
1930     .getattr = pf_getattr, .setattr = pf_setattr, .canonical_path = pf_canonical_path,
1931     .mknod = pf_mknod, .mkdir = pf_mkdir, .unlink = pf_unlink, .rmdir = pf_rmdir,
1932     /*.symlink = pf_symlink,*/
1933     .rename = pf_rename,
1934     /*.link = pf_link,*/
1935     .open = pf_open, .read = pf_read,
1936     /*.write = pf_write,*/
1937     /*.flush = pf_flush,*/
1938     .release = pf_release, .fsync = pf_fsync, .opendir = pf_opendir, .readdir = pf_readdir,
1939     .releasedir = pf_releasedir, .fsyncdir = pf_fsyncdir, .statfs = pf_statfs,
1940     /*.setxattr = pf_setxattr,
1941     .getxattr = pf_getxattr,
1942     .listxattr = pf_listxattr,
1943     .removexattr = pf_removexattr,*/
1944     .access = pf_access, .create = pf_create,
1945     /*.getlk = pf_getlk,
1946     .setlk = pf_setlk,
1947     .bmap = pf_bmap,
1948     .ioctl = pf_ioctl,
1949     .poll = pf_poll,*/
1950     .write_buf = pf_write_buf,
1951     /*.retrieve_reply = pf_retrieve_reply,*/
1952     .forget_multi = pf_forget_multi,
1953     /*.flock = pf_flock,*/
1954     .fallocate = pf_fallocate,
1955     .readdirplus = pf_readdirplus,
1956     /*.copy_file_range = pf_copy_file_range,*/
1957 };
1958 
1959 static struct fuse_loop_config config = {
1960         .clone_fd = 1,
1961         .max_idle_threads = 10,
1962 };
1963 
1964 static std::unordered_map<enum fuse_log_level, enum android_LogPriority> fuse_to_android_loglevel({
1965     {FUSE_LOG_EMERG, ANDROID_LOG_FATAL},
1966     {FUSE_LOG_ALERT, ANDROID_LOG_ERROR},
1967     {FUSE_LOG_CRIT, ANDROID_LOG_ERROR},
1968     {FUSE_LOG_ERR, ANDROID_LOG_ERROR},
1969     {FUSE_LOG_WARNING, ANDROID_LOG_WARN},
1970     {FUSE_LOG_NOTICE, ANDROID_LOG_INFO},
1971     {FUSE_LOG_INFO, ANDROID_LOG_DEBUG},
1972     {FUSE_LOG_DEBUG, ANDROID_LOG_VERBOSE},
1973     });
1974 
fuse_logger(enum fuse_log_level level,const char * fmt,va_list ap)1975 static void fuse_logger(enum fuse_log_level level, const char* fmt, va_list ap) {
1976     __android_log_vprint(fuse_to_android_loglevel.at(level), LIBFUSE_LOG_TAG, fmt, ap);
1977 }
1978 
ShouldOpenWithFuse(int fd,bool for_read,const std::string & path)1979 bool FuseDaemon::ShouldOpenWithFuse(int fd, bool for_read, const std::string& path) {
1980     if (fuse->passthrough) {
1981         // Always open with FUSE if passthrough is enabled. This avoids the delicate file lock
1982         // acquisition below to ensure VFS cache consistency and doesn't impact filesystem
1983         // performance since read(2)/write(2) happen in the kernel
1984         return true;
1985     }
1986 
1987     bool use_fuse = false;
1988 
1989     if (active.load(std::memory_order_acquire)) {
1990         std::lock_guard<std::recursive_mutex> guard(fuse->lock);
1991         const node* node = node::LookupAbsolutePath(fuse->root, path);
1992         if (node && node->HasCachedHandle()) {
1993             use_fuse = true;
1994         } else {
1995             // If we are unable to set a lock, we should use fuse since we can't track
1996             // when all fd references (including dups) are closed. This can happen when
1997             // we try to set a write lock twice on the same file
1998             use_fuse = set_file_lock(fd, for_read, path);
1999         }
2000     } else {
2001         LOG(WARNING) << "FUSE daemon is inactive. Cannot open file with FUSE";
2002     }
2003 
2004     return use_fuse;
2005 }
2006 
InvalidateFuseDentryCache(const std::string & path)2007 void FuseDaemon::InvalidateFuseDentryCache(const std::string& path) {
2008     LOG(VERBOSE) << "Invalidating FUSE dentry cache";
2009     if (active.load(std::memory_order_acquire)) {
2010         string name;
2011         fuse_ino_t parent;
2012         fuse_ino_t child;
2013         {
2014             std::lock_guard<std::recursive_mutex> guard(fuse->lock);
2015             const node* node = node::LookupAbsolutePath(fuse->root, path);
2016             if (node) {
2017                 name = node->GetName();
2018                 child = fuse->ToInode(const_cast<class node*>(node));
2019                 parent = fuse->ToInode(node->GetParent());
2020             }
2021         }
2022 
2023         if (!name.empty()) {
2024             fuse_inval(fuse->se, parent, child, name, path);
2025         }
2026     } else {
2027         LOG(WARNING) << "FUSE daemon is inactive. Cannot invalidate dentry";
2028     }
2029 }
2030 
FuseDaemon(JNIEnv * env,jobject mediaProvider)2031 FuseDaemon::FuseDaemon(JNIEnv* env, jobject mediaProvider) : mp(env, mediaProvider),
2032                                                              active(false), fuse(nullptr) {}
2033 
IsStarted() const2034 bool FuseDaemon::IsStarted() const {
2035     return active.load(std::memory_order_acquire);
2036 }
2037 
Start(android::base::unique_fd fd,const std::string & path)2038 void FuseDaemon::Start(android::base::unique_fd fd, const std::string& path) {
2039     android::base::SetDefaultTag(LOG_TAG);
2040 
2041     struct fuse_args args;
2042     struct fuse_cmdline_opts opts;
2043 
2044     struct stat stat;
2045 
2046     if (lstat(path.c_str(), &stat)) {
2047         PLOG(ERROR) << "ERROR: failed to stat source " << path;
2048         return;
2049     }
2050 
2051     if (!S_ISDIR(stat.st_mode)) {
2052         PLOG(ERROR) << "ERROR: source is not a directory";
2053         return;
2054     }
2055 
2056     args = FUSE_ARGS_INIT(0, nullptr);
2057     if (fuse_opt_add_arg(&args, path.c_str()) || fuse_opt_add_arg(&args, "-odebug") ||
2058         fuse_opt_add_arg(&args, ("-omax_read=" + std::to_string(MAX_READ_SIZE)).c_str())) {
2059         LOG(ERROR) << "ERROR: failed to set options";
2060         return;
2061     }
2062 
2063     struct fuse fuse_default(path, stat.st_ino);
2064     fuse_default.mp = &mp;
2065     // fuse_default is stack allocated, but it's safe to save it as an instance variable because
2066     // this method blocks and FuseDaemon#active tells if we are currently blocking
2067     fuse = &fuse_default;
2068 
2069     // Used by pf_read: redacted ranges are represented by zeroized ranges of bytes,
2070     // so we mmap the maximum length of redacted ranges in the beginning and save memory allocations
2071     // on each read.
2072     fuse_default.zero_addr = static_cast<char*>(mmap(
2073             NULL, MAX_READ_SIZE, PROT_READ, MAP_ANONYMOUS | MAP_PRIVATE, /*fd*/ -1, /*off*/ 0));
2074     if (fuse_default.zero_addr == MAP_FAILED) {
2075         LOG(FATAL) << "mmap failed - could not start fuse! errno = " << errno;
2076     }
2077 
2078     // Custom logging for libfuse
2079     if (android::base::GetBoolProperty("persist.sys.fuse.log", false)) {
2080         fuse_set_log_func(fuse_logger);
2081     }
2082 
2083     if (MY_USER_ID != 0 && mp.IsAppCloneUser(MY_USER_ID)) {
2084         // Disable dentry caching for the app clone user
2085         fuse->disable_dentry_cache = true;
2086     }
2087 
2088     fuse->passthrough = android::base::GetBoolProperty("persist.sys.fuse.passthrough.enable", false);
2089     if (fuse->passthrough) {
2090         LOG(INFO) << "Using FUSE passthrough";
2091     }
2092 
2093     struct fuse_session
2094             * se = fuse_session_new(&args, &ops, sizeof(ops), &fuse_default);
2095     if (!se) {
2096         PLOG(ERROR) << "Failed to create session ";
2097         return;
2098     }
2099     fuse_default.se = se;
2100     fuse_default.active = &active;
2101     se->fd = fd.release();  // libfuse owns the FD now
2102     se->mountpoint = strdup(path.c_str());
2103 
2104     // Single thread. Useful for debugging
2105     // fuse_session_loop(se);
2106     // Multi-threaded
2107     LOG(INFO) << "Starting fuse...";
2108     fuse_session_loop_mt(se, &config);
2109     fuse->active->store(false, std::memory_order_release);
2110     LOG(INFO) << "Ending fuse...";
2111 
2112     if (munmap(fuse_default.zero_addr, MAX_READ_SIZE)) {
2113         PLOG(ERROR) << "munmap failed!";
2114     }
2115 
2116     fuse_opt_free_args(&args);
2117     fuse_session_destroy(se);
2118     LOG(INFO) << "Ended fuse";
2119     return;
2120 }
2121 
GetOriginalMediaFormatFilePath(int fd) const2122 const string FuseDaemon::GetOriginalMediaFormatFilePath(int fd) const {
2123     struct stat s;
2124     memset(&s, 0, sizeof(s));
2125     if (fstat(fd, &s) < 0) {
2126         PLOG(DEBUG) << "GetOriginalMediaFormatFilePath fstat failed.";
2127         return string();
2128     }
2129 
2130     ino_t ino = s.st_ino;
2131     dev_t dev = s.st_dev;
2132 
2133     dev_t fuse_dev = fuse->dev.load(std::memory_order_acquire);
2134     if (dev != fuse_dev) {
2135         PLOG(DEBUG) << "GetOriginalMediaFormatFilePath FUSE device id does not match.";
2136         return string();
2137     }
2138 
2139     const node* node = node::LookupInode(fuse->root, ino);
2140     if (!node) {
2141         PLOG(DEBUG) << "GetOriginalMediaFormatFilePath no node found with given ino";
2142         return string();
2143     }
2144 
2145     return node->BuildPath();
2146 }
2147 
InitializeDeviceId(const std::string & path)2148 void FuseDaemon::InitializeDeviceId(const std::string& path) {
2149     struct stat stat;
2150 
2151     if (lstat(path.c_str(), &stat)) {
2152         PLOG(ERROR) << "InitializeDeviceId failed to stat given path " << path;
2153         return;
2154     }
2155 
2156     fuse->dev.store(stat.st_dev, std::memory_order_release);
2157 }
2158 } //namespace fuse
2159 }  // namespace mediaprovider
2160