Currently the fsync() inside write_all() may be called multiple times
when performing a command. Move that to the outer loop and call it
only after completing the command.
Also remove the O_SYNC flag when writing a stash.
Change-Id: I71e51d76051a2f7f504eef1aa585d2cb7a000d80
This reverts commit b65f0272c8.
It slows down the update too much on some devices (e.g. increased
from 8 mins to 40 mins to take a full OTA update).
Bug: 22129621
Change-Id: I016e3b47313e3113f01bb4f8eb3c14856bdc35e5
(cherry picked from commit 7125f9594d)
A RangeSet has half-closed half-open bounds. For example, "3,5" contains
blocks 3 and 4. So "3,5" and "5,7" are actually not overlapped.
Bug: 22098085
Change-Id: I75e54a6506f2a20255d782ee710e889fad2eaf29
Due to observed BLKDISCARD flakiness, overwrite blocks that we want
to discard with zeros first to avoid later issues with dm-verity if
BLKDISCARD is not successful.
Bug: 20614277
Bug: 20881595
Change-Id: I4f6f2db39db990879ff10468c9db41606497bd6f
Due to observed BLKDISCARD flakiness, overwrite blocks that we want
to discard with zeros first to avoid later issues with dm-verity if
BLKDISCARD is not successful.
Bug: 20614277
Bug: 20881595
Change-Id: I0280fe115b020dcab35f49041fb55b7f8e793da3
Although stdout and stderr are both redirected to log file with no
buffering, we are seeing some outputs are mixed in random order.
This is because ui_print commands from the updater are passed to the
recovery binary via a pipe, which may interleave with other outputs
that go to stderr directly.
In recovery, adding ui::PrintOnScreenOnly() function to handle
ui_print command, which skips printing to stdout. Meanwhile, updater
prints the contents to stderr in addition to piping them to recovery.
Change-Id: Idda93ea940d2e23a0276bb8ead4aa70a3cb97700
In the block updater, if BLKDISCARD fails, the error is silently
ignored and some of the blocks may not be erased. This means the
target partition will have inconsistent contents.
If the ioctl fails, return an error and abort the update.
Bug: 20614277
Change-Id: I33867ba9337c514de8ffae59f28584b285324067
These are already getting libc++, so it isn't necessary. If any of the
other static libraries (such as adb) use new or delete from libc++,
there will be symbol collisions.
Change-Id: I55e43ec60006d3c2403122fa1174bde06f18e09f
Only trusted input is passed to parse_range, but check for invalid
input to catch possible problems in transfer lists.
Bug: 21033983
Bug: 21034030
Bug: 21034172
Bug: 21034406
Change-Id: Ia17537a2d23d5f701522fbc42ed38924e1ee3366
When automatically stashing overlapping blocks, should the stash
file already exist due to an explicit stash command, it's not safe
to remove the stash file after the command has completed.
Note that it is safe to assume that the stash file will remain in
place during the execution of the next command, so we don't have
take other measures to preserve overlapping blocks.
The stash file itself will be removed by a free command when it's
no longer needed.
Bug: 20297065
Change-Id: I8ff1a798b94086adff183c5aac03260eb947ae2c
I've added explanatory comments to mzExtractRecursive because
that function will live on as a utility even after we move the
zip format related logic to libziparchive.
bug: 19472796
(cherry-picked from commit c9ccdfd7a42de08c47ab771b94dc5b9d1f957b95)
Change-Id: I8b7fb6fa3eafb2e7ac080ef7a7eceb691b252d8a
Add support for transfer list version 3, which allows us to
verify the status of each command and resume an interrupted
block based OTA update. Notes on the changes:
- Move the previous BlockImageUpdateFn to a shorter and
reusable PerformBlockImageUpdate, which can be used also
in BlockImageVerifyFn for verification.
- Split individual transfer list commands into separate
functions with unified parameters for clarity, and use
a hash table to locate them during execution.
- Move common block reading and writing to ReadBlocks and
WriteBlocks to reduce code duplication, and rename the
readblock and writeblock to less confusing read_all and
write_all.
The coding style of the new functions follows the existing
style in the updater/edify code.
Needs matching changes from
Ia5c56379f570047f10f0aa7373a1025439495c98
Bug: 18262110
Change-Id: I1e752464134aeb2d396946348e6041acabe13942
When building for 32p, we need to be explicit that we wish to build
the 32bit version of the binaries that will be placed in the recovery
image. The recovery image doesn't actually care... but if we are not
explicit in this, the makefiles will ask for the 64bit binaries but the
Android.mk for the binaries will supply the 32bit images (causing the
build to fail).
Change-Id: Iea2d5f412740c082795da4358765751138a4b167
This allows tune2fs to be executed from within OTA scripts,
allowing for file system modifications without formatting the
partition
Bug: 18430740
Change-Id: I0c2e05b5ef4a81ecea043e9b7b99b545d18fe5e6
At the end of the OTA script, we walk through /system, updating
all the permissions on the filesystem, including the UID, GID,
standard UNIX permissions, capabilities, and SELinux labels.
In the case of a symbolic link, however, we want to skip most of
those operations. The UID, GID, UNIX permissions, and capabilities
don't meaningfully apply to symbolic links.
However, that's not true with SELinux labels. The SELinux label on
a symbolic link is important. We need to make sure the label on the
symbolic link is always updated, even if none of the other attributes
are updated.
This change unconditionally updates the SELinux label on the symbolic
link itself. lsetfilecon() is used, so that the link itself is updated,
not what it's pointing to.
In addition, drop the ENOTSUP special case. SELinux has been a
requirement since Android 4.4. Running without filesystem extended
attributes is no longer supported, and we shouldn't even try to handle
non-SELinux updates anymore. (Note: this could be problematic if
these scripts are ever used to produce OTA images for 4.2 devices)
Bug: 18079773
Change-Id: I87f99a1c88fe02bb2914f1884cac23ce1b385f91
In version 2 of block image diffs, we support a new command to load
data from the image and store it in the "stash table" and then
subsequently use entries in the stash table to fill in missing bits of
source data we're not allowed to read when doing move/bsdiff/imgdiff
commands.
This leads to smaller update packages because we can break cycles in
the ordering of how pieces are updated by storing data away and using
it later, rather than not using the data as input to the patch system
at all. This comes at the cost of the RAM or scratch disk needed to
store the data.
The implementation is backwards compatible; it can still handle the
existing version 1 of the transfer file format.
Change-Id: I4559bfd76d5403859637aeac832f3a5e9e13b63a
