Haiku monthly activity report - 08/2019
Hi there, it's time for the monthly report again! This report covers hrev53338-hrev53461. It's been a busy month!
Andrew Lindesay continue his work on HaikuDepot, tweaking the BarberPole look, adding a display of "usage conditions" (EULA, license, etc) from packages,
Ryan Leavengood also worked in this area, making sure if you open an existing hpkg file with HaikuDepot, it will offer you to uninstall the package if it's currently installed.
humdinger improved the colors and icons used in DriveSetup to indicate read-only, BFS, and encrypted volumes, after attempting to document the existing ones and finding they didn't make that much sense.
The package kit is still relatively new code, and its error handling is showing its limitations as we investigate problems and often get bugreports with not-so-helpful error messages. Kallisti5 fixed some of these, as part of an ongoing effort, with some progress made every time a new unhelpful error message is encountered.
Ryan Leavengood also optimized the way we handle "old" states (allowing to rollback the system to an older state from the boot menu), so that a normal boot avoids running this code and is now much faster.
AGMS is working on a pre-uninstall script feature, allowing packages that needs to run some code on uninstall to do so. This requires an addition to the package format, so this code is merged in gradually to avoid breaking installs. Make sure your system is up to date to at least hrev53373 to avoid problems in the future.
Pascal Abresch got the first part of his work to handle "media" keys (play, pause, and other additional keys) recognised by Haiku. The PS/2 driver has been adjusted, but adding all these new keys to the keymap means we now have more than 128 possible keys, which the BeOS keymap format does not allow. So we will need a new one, and this will break compatibility with old apps using the keymap directly (as the API allows).
Simon South added a way to notify applications when the keymap is changed, and used this in his work implementing "Meta" key support in Terminal (allowing access to Bash line editing features as well as making emacs and other apps usable to their full extent)
Fixes and cleanups
PulkoMandy continued his work on reviewing issues reported by PVS studio, identifying minor issues in FreeBSD network drivers, a bug in the KDL command to dump semaphores (it is possible to dump them by name again, now), a memory leak in the Bluetooth preferences and confusing variable names in the boot menu system.
CodeForEvolution made coding style cleanups in CodyCam and reworked the resource file a bit. CodyCam is however still in need for a working UVC webcam driver.
MT improved support for keymaps and keyboard layouts localization, and fixed various warnings in many places in Haiku codebase.
Another round of bugfixes from waddlesplash: his work on optiminzing the kernel uncovered some bugs and introduced the occasional regression (given our lack of automated tests, this is unsurprising). These problems have now been solved and Haiku is running nice and stable (and faster than before) again.
X512 fixed an off-by-one error in the BMenuBar class that led to a 1 pixel glitch on the right side, that no one had noticed so far.
PulkoMandy fixed some error handling problems in BRoster leading to an infinite loop, and reworked the way we handle launching applications from symlinks, allowing the application to know when this happens and access the symlink file to get information about it. Some applications rely on the name they are launched at to behave differently, and other may also use FS attributes from the symlink to do specific things.
oco fixed a crash that could happen when unplugging an USB midi device
Kallisti5 reworked the handling of VPN and dial-up interfaces in the Network Prefs. While these are still not supported fully, at least the UI is in place for when we get them going.
Waddlesplash improved the BSecureSocket implementation to allow better hostname validation on SSL sockets, and changed the API to reject untrusted certificates by default (previously, apps were required to override a method if they wanted that, the expectation being that they would in some way ask the user what to do).
Some of the work for the SPARC port by PulkoMandy was merged, fixing most of our OpenFirmware code to handle the 64bit version of it used on SPARC machines (this code was initially written for PowerPC, which uses 32bit there).
On the RISC-V side, TLS was disabled in the buildtools by kallisti5 as we are currently lacking the required primitives to support it.
Jaroslaw Pelczar has also started upstreaming his work on support for ARM64 CPUs. This led to a lot of cleanups in various part of the code for warnings that had been previously ignored, and the initial work to add the new CPU in buildtools, haikuporter, and the initial support files required to build a bootstrap image. Jaroslaw is a new contributor, but not unknown to Haiku as he had already worked on a proof of concept ARM64 port as well as an earlier one targeting the AVR32 architecture. This time with things being upstreamed, we hope that the port will be more maintainable and reach an usable state.
Waddlesplash has reviewed some parts of the code for the most obvious security problems. This led to some improvements to the areas flag, making sure it is not possible for a process to easily clone another one's areas without specific flags being set. In particular, since the rpmalloc allocator is using areas for the heap, it was important to provide some level of protection for these. Note that apps compiled under BeOS will not benefit from this, as it is not possible for them to use the new flag, and we automatically apply compatibility fixes for them.
He also made the ioctl system call check if the argument is a pointer, and if that's the case, wether it points into userland memory and is not a (malicious or accidental) attempt to get an ioctl to read or write into kernel memory. The check is imperfect because ioctl allows to pass both pointers and plain integers, so drivers should also do some additional checks in their ioctl handlers. But we now have at least a minimal level of safety for the most common cases.
Waddlesplash made some more cleanups and bugfixes to the XHCI stack, fixing some stalls in USB3 devices.
Axeld added various options to the netstat tool to more easily filter data to be analyzed.
Simon South fixed the "minimal" build profile to include OpenSSL and libedit (the latter needed to run the Debugger). These are now considered required for any build of Haiku.
CodeForEvolution made the test_app_server system build on 64bit Haiku. This tool allows to run a secondary app_server in a window, allowing to easily experiment with changes to the app_server itself, as well as window decorators, without having to reboot the system to try changes.
Waddlesplash fixed a typo in the cleanup code that prevented node from crashed apps from being removed from the media system. No more "ghost" sliders leftover in media preferences from long gone apps!
Ryan Leavengood made some other changes, in particular to the handling of the "port pool" used to optimize communication between apps and the media server, and to the cleanup of BBuffer instances (fixing a 10 year old problem).
Extrowerk improved the support for tracked music "module" files (MOD, S3M, IT, ...) in our ffmpeg plugin and in MediaPlayer and file types.
app_server and Interface Kit
PulkoMandy worked on adding some error handling in app_server, which is currently not really worried about going out of memory and will most likely crash when it happens. Since it is a rather critical system component, we should try to do better. These problems are currently hit pretty easily by running the WebKit test suite, which hints at problems in our WebKit port, but still, nothing should be able to take down app_server with it that easily.
Mikael Konradsson patched DiskUsage to look nicer when using a "dark" (light text on darker background) color scheme, and made some fixes to various controls in the interface kit to improve their looks in this case.
Storage and Filesystems
Some patches from GSoC projects for btrfs write support (from BRJ) and the much older BFS partition resizer (from ahenriksson) were merged. There are still a lot of these waiting on Gerrit for someone to adopt them and finish them up, so your help on going through these is very welcome!
Likewise, the MMC driver stack originally developped by Krishnan Iyer is still in need for a bit of cleanup and completing the implementation. This month, the support in the Device Manager was merged, so the driver can be loaded as expected.
Korli added support for some EXT4 features, in particular support for 64bit filesystems and metadata checksums, which tend to be enabled by default in current Linux distributions.
Following this, PulkoMandy hit a problem with some of his partitions being recognized as both ext4 and BFS. It turns out each of these put their superblock in different places on the disk, so the two superblocks can co-exist if a partition has been switched from one to the other format. The EXT2 driver was fixed to not crash in this case, and give priority to the BFS one. And the mkbfs tool now makes sure to erase any leftover ext2 superblock from the disk when creating a new BFS volume. This should avoid a lot of "no bootable volume" found problems for disks where there previously was a Linux installation.
Waddlesplash made some fixes to the still new NVMe driver.
Waddlesplash also got the ramfs up and running again. This is a filesystem that stores data directly into RAM. Unlike a ramdisk, it does not use a RAM-based block device and a traditional filesystem above it, which means it only uses RAM for allocated files and automatically frees it up when the files are removed. This means it is even faster and more flexible than a RAMdisk. However, the implementation is still not complete, in particular there seem to be issues with timestamps as well as queries.
He also updated our NTFS tools to the latest version.
mmlr worked on the virtio SCSI driver, which can be used in virtual machines to avoid emulating an actual storage device, and instead use a more appropriate high-level interface. In particular, the driver is now able to handle timeout on requests if they turn out to be too slow to complete.
Finally, mounting BFS partitions will now be read-write by default, and will not anymore ask the user if they only want to mount read-only. The filesystem is now stable enough that this should be fine to do.
Ryan Leavengood investigated one of the long-standing problems in Tracker. In this case, it was the rather slow display of query results, especially when there were a lot of them. The issue turned out to be a regression dating back to the early days of OpenTracker, before the Haiku project even existed, and was eventually fixed by optimizing a case where a lot of redrawing on screen was done when not needed. Displaying query results is now even faster than before.
PulkoMandy reworked the file information dialog to fix some glitches (which were visible only in some font sizes and/or locales) and add a view of file attributes (similar to listattr), so it is not needed anymore to resort to the command line to explore these.
Waddlesplash added some errata patching in the 64bit version of Haiku for AMD Ryzen CPUs. These have various problems that can be fixed by writing magic values to internal CPU registers. This should make the system run a lot better on such machines.
- Haiku field trip report - January 2020
- Google Code-in 2019 finished
- No, I'm not Haiku's lead developer
- Haiku monthly activity report - 12/2019
- Haiku almost-monthly activity report - October and November 2019
- Haiku monthly activity report - September 2019
- Node.js now available in Haiku
- Haiku monthly activity report - 08/2019
- GSOC 2019 Final Report
- Haiku Activity Report: Performance Edition