Allows for easy override of fuse-bpf for testing without a rebuild
Test: Set this property with ro.fuse.bpf.enabled both true and false
Make sure ro.fuse.bpf.is_running is expected result
Bug: 219958836
Change-Id: I589511ea5cda76db1d55bdc2124fb546907d8acd
With the way the FUSE mount point are currently setup for emulated
volumes, there can be multiple paths that serve the same files on the
lower filesystem; eg
* /mnt/user/0/emulated/0/Android
* /mnt/user/10/emulated/0/Android
both refer to the same file on the lower filesystem:
* /data/media/0/Android
this is normally not a problem, because cross-user file access is not
allowed, and so the FUSE daemon won't serve files for other users.
With clone profiles this is no longer true however, as their volumes
are accessible by each other.
So, it can happen that an app running in clone profile 10 accesses
"/mnt/user/10/emulated/0/Android", which would be served by the FUSE
daemon for the user 10 filesystem.
At the same time, an app running in the owner profile 0 accesses
"mnt/user/0/emulated/0/Android", which would be served by the FUSE
daemon for the user 0 filesystem.
This can cause page cache inconsistencies, because multiple FUSE daemons
can be running on top of the same entries in the lower filesystem.
To prevent this, use bind mounts to make sure that cross-profile
accesses actually end up in the FUSE daemon to which the volume
belongs: "/mnt/user/10/emulated/0" is bind-mounted to
"/mnt/user/0/emulated/0", and vice-versa.
Bug: 228271997
Test: manual
Change-Id: Iefcbc813670628b329a1a5d408b6126b84991e09
persist.sys.fuse.bpf.enable and ro.fuse.bpf.enabled are both used to
decide if FUSE-BPF must be enabled or not.
- ro.fuse.bpf.enabled is a read-only property that is set in the device
makefile and would allow dogfooding devices to turn the feature
on/off.
- persist.sys.fuse.bpf.enable is a system property that overrides
ro.fuse.bpf.enabled and can only be set manually during the
development to simplify the testing of FUSE-BPF, mostly to compare if
those tests that are failing with FUSE-BPF were failing also without
the feature.
Bug: 202785178
Test: adb logcat | grep "FuseDaemon" | grep BPF
Ignore-AOSP-First: FUSE-BPF is not available in AOSP
Signed-off-by: Alessio Balsini <balsini@google.com>
Change-Id: I23f9d27172907f6c72c73bea22e4a7e0ac643888
FUSE BPF aims at achieving comparable performance to bind-mounts, with
the flexibility of FUSE.
Disable data and obb bind-mounts in favor of the FUSE filesystem if the
system implements the feature.
Bug: 202785178
Test: mount | grep obb
Signed-off-by: Alessio Balsini <balsini@google.com>
Change-Id: Ia8b289b84542125831a857b559bb6f93afbee494
IVold.MOUNT_FLAG_VISIBLE is split into MOUNT_FLAG_VISIBLE_FOR_READ and
MOUNT_FLAG_VISIBLE_FOR_WRITE.
Accordingly, VolumeBase::MountFlags::kVisible is split into
kVisibleForRead and kVisibleForWrite.
Bug: 206019156
Test: m
Change-Id: Ia55673400d9f713f221650e1335a46ba11f6f027
Merged-In: Ia55673400d9f713f221650e1335a46ba11f6f027
Otherwise, when system removes user's volume, it will hang
as there are mounts (obb and data mounts) still remain mounted in system.
Bug: 187122943
Test: atest UserLifecycleTests#managedProfileUnlock_stopped, it's not blocked anymore
Change-Id: Ic37985f98e6cbfe4fa38b981d3332c4dfc40c5b8
Originally it kills all the apps with obb and data mounted.
Due to recent changes, all apps will have obb and data dirs mounted
in default root namespace. Hence all apps will be killed by
by KillProcessesWithMounts().
To fix this, we also check if the dir is mounted as tmpfs,
as the default namespace one is bind mounted to lowerfs,
which app data isolation is mounted as tmpfs, so we only
kill the process that have obb dir mounted as tmpfs.
Bug: 148049767
Test: Able to boot without warnings
Change-Id: I5f862ad6f64f5df739b68ea7c9815352bae3be5c
Merged-In: I45d9a63ed47cbc27aebb63357a43f51ad62275db
So shell / root will always access to them directly not via fuse.
And zygote will be unmount these directories to prevent them being
abused for leaking app visibility.
Also, /mnt/androidwritable is not very useful now as it's the same as
/mnt/installer, but we should make shell / root to access /mnt/androidwritable
later and /mnt/installer should only access obb but not data dir.
Bug: 182997439
Test: Able to boot without errors
Test: df on /sdcard/Android/data shows it's no on fuse.
Change-Id: I2ad10b1e80c135f637d37ddf502ee010f89f4946
Since Android R, the FUSE prop is always on and FUSE-off is no longer
supported
Test: m
Bug: 160159282
Merged-In: Ic4414b850511fe3b4fc6df3f8b736d21335db820
Change-Id: I5a7643f9ca2f37cd7f264331df76b42df31988d5
By default FUSE filesystems have a max_ratio of 1%, meaning only 1% of
dirty pages on the system can belong to a FUSE filesystem before we
start writing back pages (and throttling, if writeback can't keep up).
This limit is useful for untrusted filesystems, but in our case, we
trust the FUSE filesystem. Since FUSE writes result in writes to the
lower filesystem, FUSE should take at most 50%. Let's start with
changing max_ratio to 40%, to avoid needless throttling.
Bug: 159254170
Bug: 159770752
Test: inspect /sys/class/bdi manually after boot
Change-Id: I467e3770fc4afba0a08fa480c0b86aa054c8b875
For fuse read ahead can be configured by writing a value to the
/sys/class/bdi/{MAJOR}:{MINOR}/read_ahead_kb file.
There are several different ways of getting {MAJOR}:{MINOR} values of
the filesystem:
* Look at st_dev of stat("/mnt/user/0/emulated").
* Parse /proc/self/mountinfo.
Stat'ing approach is used since it's easier to implement.
Bug: 157982297
Test: atest vold_tests
Test: adb shell cat /proc/self/mountinfo to get MAJOR:MINOR
Test: adb shell cat /sys/class/bdi/{MAJOR}:{MINOR}/read_ahead_kb
Test: created public volume, checked it's read_ahead_kb is also 256
Change-Id: Id0c149c4af1ceabf3afc33b4100563a512b38316
This allows devices that have sdcardfs enabled in the kernel to not use
it. When external_storage.sdcardfs.enabled=0, sdcardfs will not be
mounted. This is treated as default true to not affect upgrading
devices. It does not use the old ro.sys.sdcardfs as that has been
repurposed over time and no longer can be relied on to turn off
sdcardfs. This is included within emulated_storage.mk
Bug: 155222498
Test: mount|grep "type sdcardfs" should find nothing after boot complete
if external_storage.sdcardfs.enabled=0
Change-Id: I23d75fb1225aeabbcb1a035ad62fd042b6b3c7b5
To improvement performance, and also making them able to list
the dirs.
This should also be fine under b/151055432, as the whole obb
directory is mounted, renameTo() from installer to apps should be
a move not copy.
Bug: 153422990
Bug: 153540919
Test: atest AdoptableHostTest
Change-Id: Ia18fd4393db14a0f11d6e5b947dd716515bdeeef
Because we want all other paths (in particular Android/media) to go
through FUSE.
Also use scope_guard to make unwinding some failures easier.
Bug: 151272568
Test: atest AdoptableHostTest
Change-Id: Ib487b9071b5b212c7bb12ce54f80c96d98acaef5
In EmulatedVolume#doMount, if some operations fail, we call
EmulatedVolume#doUnmount.
During this unmount we try to unmount Android/ causing a FUSE_LOOKUP
on the FUSE mount. If the FUSE mount is not up, this can hang.
Now we introduce a new state to prevent unmounting Android/ if it
wasn't mounted.
Test: atest AdoptableHostTest
Bug: 151685786
Change-Id: I6246d3910c352034d2a4fb09ad9c1e7fd91cba5e
Also, use the pids provided by system server to remount all existing
processes, so we don't need to do the heavy and unreliable scanning in
/proc anymore.
Bug: 149548518
Test: atest AdoptableHostTest
Change-Id: Ifb5b79a3bc5438f36e0d61ec8aec96bdbc60ca13
We should mount Android/data also, not only Android/obb.
Test: After flag is enabled, AdoptableHostTest still pass.
Bug: 148049767
Bug: 150584566
Change-Id: I26dc3756aa5843b85565495e9c2698130113f49a
Merged-In: I26dc3756aa5843b85565495e9c2698130113f49a
(cherry picked from commit d88e090098)
- Remove bind mounting Android/ code as we want to bind mount obb dir
for each process instead.
- Set property "vold.vold.fuse_running_users" as an array of user id
for which fuse is ready to use.
- After fuse is ready for a user, fork a background process in vold
to bind mount all direct boot apps for that user so its direct boot
apps obb dir will be mounted to lower fs for imporoved performance.
Bug: 148049767
Bug: 137890172
Test: After flag is enabled, AdoptableHostTest still pass.
Change-Id: I90079fbeed1c91f9780ca71e37b0012884680b7c
Normally sdcardfs takes care of setting up these directories on-demand,
for example when an app requests its private data directory to be
created. On devices without sdcardfs however, we ourselves need to make
sure to setup the UID/GID of these directories correctly.
Introduce a new PrepareAndroidDirs() function which sets the dirs up
correctly. On devices without sdcardfs, that means:
Path UID GID mode
/Android media_rw media_rw 771
/Android/data media_rw ext_data_rw 771
/Android/obb media_rw ext_obb_rw 771
Bug: 146419093
Test: wipe Android/, reboot, with and without sdcardfs, verify
contents
Change-Id: I3a879089422c7fc449b6a3e6f1c4b386b86687a4
When we try mounting an EmulatedVolume, we may mount sdcardfs but fail
in any of the FUSE mounts, in this case we should unmount whatever
mounts we made during the mount.
Test: Intentionally causing a partial failure, verified that sdcardfs
gets unmounted
Bug: 147610762
Change-Id: I29ed044ed8ab8aa3dd83bc97a49eb3140ce4fe27
Installers will be allowed to write OBB for any application; this is not
easy to achieve on sdcardfs, where the GID of Android/obb is the same as
the GID of Android/data (app-private data), meaning giving installers
write access to Android/obb would also give them write access to
Android/data.
Instead, we create a /mnt/installer view, which is exactly the same as
/mnt/user, with the sole exception that the write sdcardfs view of
Android/obb is mounted on top. This is what will allow installers to
write there, while still being restricted with respect to app-private
data in Android/data.
Bug: 134706060
Test: atest AdoptableHostTest
Change-Id: If2b93870a877efef182bdc06466552a7527499ad
In preparation of sdcardfs going away on devices launching with R,
conditionally use it.
Bug: 146419093
Test: cuttlefish with sdcardfs, cuttlefish without sdcardfs but with
FUSE
Change-Id: I2c1d4b428dcb43c3fd274dde84d5088984161993
Now that StorageManager asks vold to create package directories again,
apps only need write access in their own app-private directory. Both
app-private and OBB dirs will be created by privileged daemons as
needed.
This means we can use the "default" sdcardfs view for the Android/
bind-mount again. This has the added benefit that it fixes shell, which
wasn't in the "everybody" group, and therefore suddenly couldn't enter
/sdcard/Android anymore.
Bug: 146189163
Test: atest AdoptableHostTest
Change-Id: I37ca3b19ea4d11ed866efa808f51be945a4dc080
With the newly added flag for Settings developer options, which is now
used to change the state of FUSE, PROP_FUSE now acts as the snapshot
feature flag for the current boot.
Bug: 145391093
Test: atest AdoptableHostTest
Change-Id: I22363b088e88c764294cbd61c6d94160c907fae3
The path for emulated volumes is set to /storage/emulated,
not /storage/emulated/<userid> . When unmounting, we only
want to kill process with a reference to /<userid>; this prevents
killing processed needlessly.
Bug: 137890172
Test: atest AdoptableHostTest
Change-Id: I70e36e87077e87db9b3c7e07dc0e481ba06c2c14
For apps seeing the FUSE filesystem, we want to bind-mount the Android/
directory to the lower filesystem. The main reason for this is game
performance - Android/ contains both OBBs and app-private external data,
and both are heavily accessed during game startup. This is a pretty
straightforward bind-mount on top of /mnt/user.
Bug: 137890172
Test: Running the following:
df /storge/emulated/0 ==> /dev/fuse (FUSE)
df /storage/emulated/0/Android ==> /data/media (sdcardfs)
Test: atest AdoptableHostTest
Change-Id: Ic17a5751b5a94846ee565ff935644a078044ab06
The pass-through mount is used by MediaProvider to access external
storage. Previously, it was the raw filesystem (eg ext4/f2fs); the
problem with that is that the permissions on that filesystem don't allow
MediaProvider to access all the files it needs to - in particular
directories under Android/
To solve this problem, we can have the pass-through mount sit on top of
sdcardfs instead of the raw filesystem. This means we need to mount
sdcardfs even in case we're using FUSE, but we already needed to do this
anyway for other performance reasons.
Bug: 135341433
Test: atest AdoptableHostTest
Change-Id: I893d5e5076c5096d2d55212f643c9a857242e964
Long before scoped storage, FUSE was used to create several views of the
filesystem; this was later replaced by sdcardfs, yet the code still
references FUSE in variables and log messages. Fix those up to avoid
confusion with the "new FUSE".
Bug: 135341433
Test: atest AdoptableHostTest
Change-Id: I2966c5f95d6f38df36a20cdfdd0b95b3051f3f1e
LOG() can hang, because another vold thread may have been holding
a libc lock at the time we forked. Simply moving the log line to
before the fork().
Bug: 135341433
Test: atest AdoptableHostTest
Change-Id: I6063eb33f35c835a226bc1e56de97ace426747ff
Mounting a FUSE path needs two steps:
1. Mounting the filesystem
2. Starting the FUSE session in the FUSE daemon
The second part requires retriving an fd from (1) and the mount paths
and passing it to the FUSE daemon.
Previously, we'd return from the Vold mount call and mark the volume
as mounted while we scramble to do (2). This means there's a time
period where the Volume is marked as MOUNTED but not actually ready
and any IO access on the paths will hang forever. This could also be
misleading when interpreting bug reports.
Now, we block the Vold mount call until the FUSE session is started
Test: atest AdoptableHostTest
Bug: 144275217
Change-Id: I45238a31df71286f67ef1c65c711d0085d72e97f
Ifabaa12368e5a591fbcdce4ee71c83ff35fdac6b introduced individual
emulated volumes for each Android user. The change however didn't
create the volumes for the secondary users on user start in vold
without the persist.sys.fuse flag
Now we always create the volumes but only mount sdcardfs volumes
for user 0 because the sdcardfs mount paths do not change with for
different users unlike the FUSE mount paths.
Bug: 144473552
Test: atest AdoptableHostTest
Test: Start a guest user in Settings and launch chrome browser in that
user, verify that chrome does not crash
Change-Id: I89f3591d0197d86267f0e3934f496273e2f9fd7e
Up until now, the FUSE mount logic has made two assumptions:
1. The primary external volume is an emulated volume on /data/media
2. Only the primary user is running, as user zero
These assumptions are fixed by the following changes
creating an EmulatedVolume per Android user and changing the
VolumeBase id format to append the user to the id, so
s/emulated/emulated-0/. This allows us mount separate volumes per user
Some additional refactorings to re-use/clean up code.
Test: adb shell sm set-virtual-disk and partition disk operations work
even after setting up a work profile
Bug: 135341433
Change-Id: Ifabaa12368e5a591fbcdce4ee71c83ff35fdac6b
This will allow us to receive values from server flags and store them in
persist.sys.fuse without risking flag consistency during a boot.
Test: manual - flip persist.sys.fuse both ways and make sure FuseDaemon
works as expected.
Bug: 140803239
Change-Id: I839a1973c98b4eda982226d20be48d1c08e7464a
When mounting a FUSE device on /mnt/user/<userid>/<vol>,
bind mount the correspoinding lower filesystem path to
/mnt/pass_through/<userid>/<vol>. At Zygote fork time, an
app with the right privilege will have the pass_through path
bind mounted into /storage instead of the /mnt/user path.
This provides such an app direct access to the lower filesystem
without going through FUSE.
Bug: 140064376
Test: mount(8) shows /mnt/pass_through/0/emulated is a bind
mount of the lower fs
Change-Id: I32c3cad64138910fcec9fb8f66b206706b5fd139
Also allow the state just before doMount() as a valid state for setting
fuse fd.
Test: manual
BUG:140173712
Change-Id: I012f8a83fef00e68f33010954fbc2ebc53cf8f1d
Since system_server cannot mount devices by itself,
add a binder interface to vold that system_server
can call to initiate this mount when required.
BUG: 135341433
Test: manual
Test: atest --test-mapping packages/providers/MediaProvider
Test: ExternalStorageHostTest DownloadProviderTests
Change-Id: If4fd02a1f1a8d921a3f96783d8c73e085c5b7ca1
This will be used for system internals to access
secondary volumes without having to bypass sdcardfs.
Bug: 121277410
Test: manual
Exempt-From-Owner-Approval: Got approval on internal master
Change-Id: I9b3e33f6c6e426e83469b8030b02d59c02183946
- Also update vold to create sandboxes for secondary storage devices.
- Since bind mounts are created in the process specific namespaces, we
don't need /mnt/storage anymore which we were using it to prevent
some bind mounts from propagating onto /mnt/runtime/write.
- Create bind mounts for {media,obb} dirs similar to data dir in
per process namespace.
- Also fix a bug where we are not passing correct packages to vold when
a new user starts.
Bug: 111890351
Test: manual
Change-Id: I7849efc4fbf3c654606fa30de7ab2de0236d766f
This will allow adding lots of verbose logs which can be enabled
only during local testing/debugging. Update the existing verbose
level logs to debug level since we want those to be logged by
default.
Test: manual
Change-Id: Ib05e2b6efa71308458d49affb6ed81d3975b28ab
Apply clang-format to fix the remaining files not fixed by
change I23cde3f0bbcac13bef555d13514e922c79d5ad48
Test: Format-only changes; treehugger suffices.
Change-Id: I1bfd5c8d68d298596875d5edae26cdfe27c03489
