The opendir and closedir/free_dircookie hooks were called with
mismatched vnode. It seems only googlefs is actually affected by this,
since all other fs without a get_vnode_name just don't are about the
passed vnode arg to closedir and free_dircookie.
Now I should really get some sleep!
See http://www.freelists.org/post/haiku-development/Unifying-visuals-of-BTabView-usage
Besides the tab bar going the full width of the window, also some
layouting tweaks to several apps and prefs.
Still missing: the first (and last) tabs in the tab bar should be inset by
B_USE_WINDOW_SPACING so the controls in the tab view line up nicely.
I think I remember stippi wanting to look into it... :)
* This removes the CenteredViewContainer class, as it is no longer
being used.
* However, it also removes its functionality, ie. the Sudoku view now
fills the complete window (even without any borders), again.
The lock was only acquired when paths to watch were added or removed,
protecting the data structures against concurrent modification due
to addition/removal of entries by the API user.
Locking is also required for node monitor messages since these can
trigger the data structures to be modified (due to recursive watching
and new directories becoming available or due to resyncing of modified
ancestor chains).
Previously it was possible to corrupt the data structures when node
monitor messages were received while still starting to watch a directory
structure. This was especially likely in the case of watching devfs
directories, as accessing these can trigger device scanning which in
turn could possibly add new device entries. Either the path monitor
looper or the API user would then trip over the corrupted data
structures.
Probably fixes#11280. Although I was only able to reproduce crashes
on the API side, corruption of the hash tables and corresponding endless
loops are quite plausible.
Possibly also fixes#12412 if the input_server was in the process of
starting to watch entries. It's hard to tell due to the lack of a back
trace but would fit the crashes I was able to reproduce with a synthetic
test case.
Reduce duplication of code by
* Removing from elf_common.h definitions available in os/kernel/elf.h
* Deleting elf32.h and elf64.h
* Renaming elf_common.h to elf_private.h
* Updating source to build using public and private ELF header files
together
Signed-off-by: Jessica Hamilton <jessica.l.hamilton@gmail.com>
* Services that end are now automatically restarted.
* However, this is very basic at the moment, there is no throttling,
and no support for shutting down the system.
* Ie. a listing of all targets/jobs, as well as specific (basic) info
on each.
* Also added a bit of optional debug output.
* Moved translating the path to launch time -- we should take the job's
environment into account here at some point.
Also print the locker sem (for manual name lookup) and the involved
threads. It was also missing the line terminator which messed up the
following output.
Also fix a typo in a comment.
Since a BLocker can be unlocked from other threads than the one holding
the lock, it can also be further unlocked even when already unlocked.
This caused the recursive count to become negative. The first lock then
needs to reinitialize the count to 1 for the lock balance to work again.
Just incrementing the negative recursive count lead to it never
counting back down from one to zero in the unlock case, which made the
BLocker impossible to unlock.
This makes the Haiku BLocker behave exactly like the BeOS one, including
the negative recursive count and reinitialization, as evidenced by its
debugging features showing the internal counts.
Alternatively to reinitializing the recursive count it could be
prevented from going below zero in the first place, but I don't see why
we should deviate from BeOS there while allowing its awkward unlock
behaviour.
This makes some more exotic use cases work like the BGLView <-> SDL
combination that previously would always just hang. While these abuses
should be reviewed/corrected, just hanging the BLocker doesn't seem
useful.
This adds to the "unittests" target a dependency on the
"AppTestRunApp3a" (etc.) files meant to be copied as part of the
BApplication test suite so they are generated when "jam unittests" is
run.
* TestsRules: Add "UnitTestDependency" rule.
* testapps/Jamfile: Make unit tests depend on copied files.
Fixes#12441.
Signed-off-by: Adrien Destugues <pulkomandy@pulkomandy.tk>
* There is a small API change, and only the x86_gcc2 package was
updated. This means the x86 and x86_64 builds are broken until someone
updates the packages.
* Fix several small bugs, the most important one being that cookies are
written to disk again.
The use of an unreliable test for relocatability effectively broke
runtime_loader's support for non-position-independent executables, as it
would insist on randomly positioning these files' segments in memory
anyway causing the program to quickly crash.
With this change runtime_loader uses the object type specified in the
file's header to determine whether its segments can be safely relocated,
restoring support for non-PI executables.
Fixes#12427.
Signed-off-by: Adrien Destugues <pulkomandy@pulkomandy.tk>
* Add -fgnu89-inline flag for libroot/posix/glibc
* Change __GNUC__ == 4 to __GNUC__ >= 4
Signed-off-by: Adrien Destugues <pulkomandy@pulkomandy.tk>
Fixes#11990, most of the changes had already been done.
Several errors in the html output caused WebPositive to not render
the file completely.
Signed-off-by: Adrien Destugues <pulkomandy@pulkomandy.tk>
Fixes#12387.
Splitting the almost 70 MiB userguide package into one package per
language at ~4.5 MiB saves time and space when installing/updating.
At the online tool that manages the user guide pages, we have to
make changes to fall back to the online page if you try to switch
to a language you haven't installed locally. See #9322.
* Save the package state before attempting to install/uninstall the package.
* In the case of failure set the package state back to what it was.
* Use the synchronous version of BAlert so that the package state won't
change until the user clicks the Close button on the dialog.
Fixes#10838
This allows the Cmd+A key shortcut to select all text on
the subject control. You could already Select All from the
View menu but this is more convinient.
Fixes#12361
* Since we're disk bound, and not CPU bound, it doesn't make much sense
to restrict the number of threads on the number of CPUs.
* It's still not completely independent of the number of CPUs now,
though: we'll have 3 * CPU count worker threads.
* This was the harmless part: a job was been requeued that already was
being launched.
* I was already aware of this one, and only accidentally stumbled over
the non-harmless case in the JobQueue code when I tried to fix that
little issue... (ie. never ignore warnings, even if you think you
know what's going on).
* A dependent job was requeued even if it wasn't part of the queue
before. The code relied on dependent jobs being already enqueued;
but that cannot be guaranteed.
* If a job failed, its dependent jobs are now also set to failed, so
that they won't be requeued at a later point.
* This caused some of the "Launching xxx failed: Operation not allowed"
messages in the boot process. Those actually weren't harmless, and
could mess up the natural job order.
* Triggered whenever a volume is mounted (surprise!).
* There is no way to specify which volume you are interested in for now
(if someone knows a good use case, I'd be willing to add that,
though :-)).
* Sticky events are events that keep their signal raised, ie. even if
a job is initialized afterwards, it will still be triggered.
* Consolidated naming for external events.
* Events are now registered once they are actually being used. This
allows them to allocate the resources they need to do their thing.
* With events being registered, this proved to be confusing, and it also
helps to differentiate between event objects within the daemon, and
events coming from external sources.
* No impact to non-ValleyView chipsets
* Bump some register locations for VLV
* Only have HDMI port to test with on my ValleyView GPU
and our driver seems to be missing all HDMI and
sideband functionality.
* As ValleyView chipsets seem to be UEFI only, we don't
have VESA fallback, so this shouldn't cause regressions.
(unless we get UEFI framebuffer support)
This issue only manifested itself when the navigation toolbar was shown.
The scrollbar appeared to have no border while the rest did. This issue
manifested when the scrollbar insets were adjusted in hrev49654. The
scroll bar insets were really hiding the bug underlying bug though.
I'll try to explain what was happening and how I fixed it. The PoseView
container, called BorderedView, was showing its top border when the
navigation bar was hidden, and hiding its top border when the navigation
bar was shown. This (almost) worked because the menu bar provided a
bottom border while the navigation toolbar didn't. However hiding
BorderedView's top border also hid the scroll bar border.
My solution was to draw a bottom border on the navigation toolbar and
then remove the top border from BorderedView unconditionally. So either
the menu bar or the navigation toolbar provides a bottom border and the
BorderedView has no top border.
Fixes#12392
Highlighting was added in hrev45983, stopped working in hrev49614, specifically
d891923650.
Add comment to explain what is happening and hopefully deter this from being
removed again in the future.
Fixes#12359
* Properly flag missing devices
* Do away with shifts and define ssts and sctl masks
* Tested working on 6 different systems with a
combination of drive configurations.
* Empty media on AHCI devices still cause port change
storms. (the issue that was attempted fixed in
5584c22fdd)
* Seems to work fine, although it should probably also be triggered when
there are still jobs in the queue -- someone more knowledgeable might
want to chime in here, please :-)
* If this turns out to be problematic, we can just drop the "on_demand"
job config again.
* Was broken in two ways: if only the shortcut "on_demand" was used,
the event didn't get created at all due to a bug in
Events::AddOnDemand().
* Furthermore, _LaunchJob() always triggered a demand, but it should
only do this when not called from a target.
* devfs: set st_rdev to the inode number of the node being queried. This
may not be the best thing to do, as it does not match what is set in
st_dev for other files, so it can't be used to find which device
stores a particular file. I'm not sure if st_rdev is actually used that
way anywhere, however.
* vfs: do not clobber st_rdev with -1 for "special" (device) files.
Refactor the code a little so setting the common attributes is done in
a single place.
Fixes#12390.
src/tools/makebootable defines _USER_MODE, which we can check
for in gpt/Header.cpp to remove the dependency on libuuid, as
makebootable only needs read-only access to partition maps
anyway.
Previously, deleting a partition was lazy and only changed
the partition type to an invalid UUID, and would still
show up as a partition in many partitioning tools.
POSIX requires SIGTTOU to be sent to a process in a background process
group that tries to change the foreground process group ID associated
with its controlling terminal, unless the process is ignoring SIGTTOU or
the calling thread is blocking it. Previously the code checked the
former condition but not the latter, making it possible for calls to
tcsetpgrp() to get stuck in a loop and never return.
Should fix#3417.
Signed-off-by: Augustin Cavalier <waddlesplash@gmail.com>
- Correct a TRACE_LOCALS statement in ValueWriter.cpp and update it to
match the current definition of ValuePieceLocation.
Fixes#12411.
Signed-off-by: Rene Gollent <rene@gollent.com>
- Add jamfile for a build variant of libuuid.
- Adjust the build version of makebootable to link to that instead, as it would
otherwise link in the target platform's version of libuuid_kernel, which
wasn't necessarily compatible with the build platform's objects (i.e. when
building a 32-bit Haiku image on a 64-bit host OS), and would consequently
fail.
When writing an individual entry to disk, the offset of the entry
was omitted, which resulted in entries not block-aligned to
overwrite previously valid entries. This in turn resulted in the
stored entries CRC no longer matching what was on disk, causing
the partitioning system to fail to identify it as a valid GPT
when read from disk later (e.g. after a reboot).
E.g. if the first entry is the ESP (which it typically is), and
then the second being an entry for a BFS partition, updating the
BFS partition entry would overwrite the entry for the ESP, thus
corrupting the entries table.
* Move to more standardized functions matching AHCI spec
* Don't perform unnecessary double port resets
* Begin implementing a software reset to try first per spec.
Software reset needs more work, falls through to port reset
for the moment which is stable.
* Don't duplicate ATA defines, use what we already provide.
* Tested working on VirtualBox 1-16 AHCI ports, Intel C200,
and AMD FCH.
* For the moment i still remain with the classic lateness calculus.
My code wasn't perfect, but this commit fix the remaining
problems from my perspective.
* The first reason is that if we have a patologic latency
such as adding for experimental reasons a snooze() before a SendBuffer or
in the BufferReceived callback, we still can't do anything about it.
If we use enqueue_time and don't send a LateProducer notice, this latency
will never be detected by the API client. We can't do anything about it,
and it's even better that systems with such problems are recognized as
soon as possible IMO.
* The second reason is that the lateness calculus described in the BeBook
is done this way because the media_kit want us to adjust our timing in both
early and late situations.
* Realtime expect that things are always delivered under a certain time
limit, if the software at the bottom doesn't meet with this requirement,
it's just not realtime and things can't work in realtime.
* enqueue_time has nothing to do with the performance_time. But we can
still add this to the media_timed_event struct so that applications can
make use of it.
* Lateness was probably not used a lot in BeOS programs as it looks like
a relatively new feature but i have the concern to complete our API
implementation to be close to what i see was reasonably the designers aim.
In preparation for implementing Undo/Redo support, we need _Insert() to
take a TextDocument instead of a BString, CharacterStyle and ParagraphStyle.
When a chunk of the TextDocument has been removed, we need to be able to
Insert() that as part of the Undo operation. Not well tested, but typing
still works.
Added TODO about which data the TextChangedEvent should have in order to
know how to respond in the TextDocumentLayout. For now, update all paragraphs
in the layout, add new ones as necessary, and remove the ones we have in
excess by using Invalidate().
Changed from white on red background, to yellow on black
(neutral) background. Red on black made it look like a
warning, whereas yellow feels more informative.
* The previous code was broken, and did not take the actual window
size into account; it's now using BWindow::MoveOnScreen().
* Also, make the width of the time zone list relative to the font size.
* The monitor size now scales with the font size.
* It now uses BWindow::MoveOnScreen() to find a good position on screen.
* The monitor info is now properly hidden when there is no info (this
removes a blank line at the top of the left box).
* As the specs say, this causes undefined behavior.
* Should help with the recently introduced boot issues, but cannot
be considered a full fix -- as mmlr pointed out, one has to detect
unsolicited COMINITs.
* This should fix ticket #4157, although I probably have missed
something.
* In any case, it no longer messes with the ref counts of the
file descriptor, and the race condition in put_fd() should be
gone.
* It's still rather messy all in all.
* Each io_context now has a "inherit_fds" member that decides whether
or not this context allows to inherit FDs to its children.
* This replaces the former O_CLOEXEC mechanism.
Make sure that the caret starts blinking half a second after it last moved.
The previous solution using Pulse() had a number of problems:
* The caret could be hidden during moving it or during typing. It would then
be shown again very soon after, i.e. when typing the next character or
when moving it to the next offset. But it looks bad anyway.
* When the caret stopped moving, it started blinking a random amount of
time afterwards, getting back into the rhythm of Pulse() messages.
However, starting to blink a constant time after the caret last moved,
looks much more satisfying.
* It now uses a font that's 3/4 the size of the plain font; ie. there
shouldn't be any change with the default font size.
* Also cleaned up some weird layout code on the way.
* You can specify which borders will be drawn using the
BControlLook::B_TOP_BORDER, ... constants.
* Adapted Mail to no longer need the SetInsets() hack.
* Use a scale factor depending on the font size.
* Be more generous when it comes to the max width.
* Use StringWidth() for the default size of the size and status columns.
* Moved entirely into MainWindow.
* Moved duplicated code into separate methods.
* Resize the main window on larger screens by default, as we can make
use of the extra space.
* Use BWindow::MoveOnScreen() instead of make_sure_frame_is_on_screen()
as the former has more info. And is even smarter now as it can
optionally resize windows to fit on screen.
* Center window on screen by default (ie. when there are no settings).
... and invalidate based on TextChangedEvents. I am not yet sure whether
TextChangeEvent needs separate counts for removed and changed paragraphs.
It is most likely not yet correct and may either update too many paragraph
layouts or miss updating some at the end.
TextDocument:
* Moved implementation of Remove() and Insert() into private methods.
* Reimplement all public Insert() methods and Remove() on top of Replace().
* In Replace(), send a TextChangedEvent. Added TODO for sending a
TextChangingEvent, although at this point, I am not sure if it will be
needed at all.
The int32 was cast to a ssize_t which has a different size on 64 bit,
therefore clobbering the stack.
Also remove the use of basic type references in arguments, which
probably was the reason for doing the above in the first place.
The cleanup commit df48d3f9a8 broke
constructing a BPicture from an archive due to an incomplete rename.
The passed in BMessage was used as the data buffer instead of the
extracted data field.
Fixes the application side crash of #12340. Seeing how long this was
broken without anyone noticing, the feature doesn't seem to be very
popular.
This introduces a more sane API (currently private) that allows for
safer and possibly more efficient implementations:
* It uses a struct of named and typed function pointers instead of just
a void pointer array. This adds type safety to the callbacks so the
compiler can figure out if things match up before subtle bugs get
introduced.
* It provides bounds for all strings/buffers passed to the callbacks.
* It uses const references instead of implicitly copying arguments.
* It folds stroke_x/fill_x pairs into draw_x functions with a fill
argument to reduce the amount of functions needed.
* It uses unsigned values where negative values make no sense.
The old API has been implemented on top of the new one using adapter
functions. It makes copies of all data passed to the callbacks which
effectively keeps the picture data from being modified. This matches
with the R5 behaviour.
This also reimplements the buffer parsing to be safe against corrupted
data by validating that the types actually fit in the provided sizes
and buffers (using a templated reader).
Since this class is used from the app_server with user provided data,
making it more safe is important even though it comes with a slight
overhead (replicating R5 behaviour, i.e. crashing the app_server when
corrupted data is fed, doesn't seem very appropriate here).
While not thread-safe, it should still be possible to use it in
non-threaded programs, or with locking on the application side.
The "thread-safe" implementation we got from NetBSD called abort(),
which is not a good solution. Restore the non-thread-safe
implementation, which should still work.
Fixes issues in openssh and mtr, and possibly other ported apps.
