The scheduler uses the load tracking logic to compute the load of
threads to be enqueued into the run queue. The time delta between the
last enqueue and the next enqueue may grow very large for threads
that mostly wait on conditions. In such cases the int "n" period count
variable would become too small and wrap around, leading to an
assertion failure.
For this to happen, the thread in question would have to have slept for
at least ~25 days and then wake up. Threads often affected would be ones
waiting for some other process to end, for example shell threads waiting
for a long running process to exit.
Fixes#13558.
Following recent changes to use libroot_build on Haiku also, it is now
actually impossible to build Haiku components on non-Haiku platforms
(BeOS R5, Dan0, BONE, Zeta), so we can remove any logic related to this.
This is only the first part; still to be removed are:
* SetSubDirSupportedPlatformsBeOSCompatible
* HOST_PLATFORM_BEOS_COMPATIBLE
* TARGET_PLATFORM_BEOS_COMPATIBLE
As it turns out, using the xattr emulation layer plus "libgnu"
causes some strange mixups at package build time, and so packages
built with it were winding up with no attributes at all.
So I've just bitten the bullet and written a full passthrough layer
to the system attributes. Verified using a full build of haiku.hpkg
this time ... after a lot of painful debugging of symlink mixups.
Hopefully I am finally rid of this plague...
To quote jscipione (from 95e8362c52),
"Let me tell you a story about a bug" -- though this tale spans a much
lesser time than that one did.
In 5e19679ea3, I enabled libroot_build for
Haiku, instead of using the system libroot as we had before. There were
a number of bugs introduced along with this that I hadn't fixed (and there
may be more after this), but most of the obvious ones (crashes on x86_64...)
were fixed shortly enough.
Attribute usage, though, was a different story. Unlike most of the POSIX
calls in libroot, which were aliasing system functions no matter what the
platform, the attribute calls were not, as they are specific to Haiku.
Initially I had completely forgot about them, and it wasn't until a few days
later when I noticed that I had an "attributes" directory in my generated
that I realized that the "generic" attribute layer was being used on Haiku.
I attempted a fix for this in 5e19679ea3,
thinking that would clear the problem up, but I didn't actually run a test
beyond seeing that my BuildConfig had been updated properly. In fact,
BuildSetup was hard-wired to not even pass that definition through on
Haiku, and so that commit had in effect caused nothing.
My initial "fix" of just changing BuildSetup then caused a build failure,
as while libroot_build itself compiled, it ran into errors whenever attributes
were used, because in letting the real libroot's attribute calls shine
through, I had bypassed libroot_build's FD emulation/shim layer.
Then I tried and failed at three separate attempts to solve this with code:
- a version of the "fs_attr_...h" interface for Haiku. This proved possible
in theory, but in practice I would need to reimplement a lot of attribute
handling code in it, because all I had access to from there was syscalls.
- a version of "fs_attr_untyped" that bypassed its reimplementations of
the "fs*attr" functions for the libroot ones, only using the FD shim layer.
This proved possibly not even theoretically possible because it would have
caused preprocessor hell in some of the build headers, and also assumptions
about how attributes are read were totally different.
- a completely new "fs_attr_haiku" that was a completely new interface to
the fs*attr functions. This proved practically impossible because of the
need to include structures from the system libroot to call out to readdir,
etc. that attempts to solve would also have caused preprocessor hell.
Then I realized that the Linux xattr emulation library, which I'd used
as a reference when attempting the first solution, was shipped by default
as a system library in all builds of Haiku ... and so I could just tell
fs_attr_untyped to use the Linux xattr handler, and then link against libgnu.
So that is how I arrived at this strange and decidedly unorthodox solution
to a problem of my own creation.
Not in the POSIX specification, but defined (not behind any guards)
in (at least) FreeBSD, NetBSD, glibc, and macOS.
Found by miqlas and myself while working on porting GNU inetutils.
I've updated the docs to match the current BMailComponent, documented
new functions, and cleaned up the MailComponent.h file to at least
somewhat match our coding style.
First in a series (there are 3 more old API docs on the Mail Kit in that
"Public API" folder.)
It was needed on macOS for a time when BUrl used regexes for parsing.
Now it does not, and so we can remove libshared's RegExp from build
libshared, and thus also libgnuregex.
The feature gives user ability to choose the position of notifications
out of Follow Deskbar, Lower Right, Lower Left, Upper Right and Upper
Left. Fixes#9749 - Notification_Server: add the ability to choose the
position of notifications (easy).
Signed-off-by: Adrien Destugues <pulkomandy@pulkomandy.tk>
I didn't notice this in the previous commit because apparently GCC2
just links against libroot's versions of them. On GCC5, however,
the version from libroot_build was used even for calls from libroot itself,
which led to infinite loops and then stack overflows.
So instead we must have the "syscall" functions in libroot_build shadow
the real ones by being named differently, which I did by changing their
prefix from "_kern" to "_kernbuild" via preprocessor macros.
Since the build syscalls.h is now substantially different than the non-
build one (and has not been synchronized in nearly a decade anyway),
I've just stripped out all the syscall defns except for the ones actually used
in the build.
Thanks to kallisti5 for helping me debug and test.
Previously we just used the system libroot, which of course meant
that when libroot's ABI changed, the build broke. Now we use the full
libroot_build that we do on non-Haiku platforms. The logic for "BeOS-compatible
but not Haiku" does not really apply anymore, so it has been gutted where
appropriate (and libhaikucompat has been decoupled from the build.)
The only caveat here is the change to Errors.h -- we really should be using
the system's one where I included the one from the tree, but for whatever
reason, GCC2 refused to handle the #include_next properly.
Fixes the build breakage of Haiku-on-Haiku by my prior commits (sorry).
Technically a "hack" (but then again most of the config/build stuff is);
as we need to use the system's config/types.h in order to get stdint
definitions and the like.
Previously there was a config_build directory which allowed the existence
of two types.h -- the system one, and the headers/build one, but seeing
as we only need this to provide Haiku-specific core types on other platforms,
using the system's one should be fine.
Our core type definitions have not changed in some time (and it's unclear
when they would change aside from potential new platforms), breakage of the
Haiku-on-Haiku build due to this should not happen often (if ever.)
Since it's just a C compiler "technically" the ABI does not matter,
but since it also can target other ABIs from one toolchain (e.g. x64),
just treat it as GCC4 ABI unilaterally.
Fixes#13847.
Now that we do not target BeOS and also do not include the main headers
directory when building "build" binaries, we can drop the separate
config_build directory and thus also the separate BeBuild.h, and just
..-include the regular one.
The build BeBuild.h defined empty _IMPEXP_ROOT and _IMPEXP_BE preprocessor
macros that the regular one does not; so I also re-synchronized
headers which used these as needed.
This file was apparenly based off the BeOS one (as is evidenced
by the "Be Incorporated" copyright ... which is problematic.)
Now it's directly based off of the non-build one.
The base VMCache class changed to the generic_ types with their
introduction in in *2011* (435c43f591),
but these classes were never properly adapted. These functions should not
be called here (they panic() -- but the base class only returns B_ERROR,
so that is a difference at least.)
Found by Clang's -Woverloaded-virtual.
That requires more padding (1 byte vs 4 or 8 depending on integer size),
so just use regular loops and chained ==s.
Caught by Clang. No functional change intended.
* Expects its config files in /boot/home/test_launch.
* Uses standard I/O, and is always in user mode.
* Also added test_launch_roster command that is able to talk to the test
server like it does to the real thing.
Add methods to get and set "Always on top", "Auto raise", and "auto hide"
which are all booleans which control aspects of the Deskbar window to
BDeskbar.
Set the bool to the default value initially. Check if sending the
message succeeds, if so check the reply which also fills out the bool.
Don't check to see if reply succeeded because the bool will only be
overwritten if it did.
Follow the BDeskbar convention Is...() for getter, Set...() for setter
e.g IsAlwaysOnTop() is the getter, SetAlwaysOnTop() is the setter.
Define new message constants to call the newly created methods.
Follow BDeskbar convention: 'gtla' is used for getter, 'stla' for setter.
g/s for getter/setter, tla is an all-lowercase code unique to each
getter/setter pair.
Copy/paste these message constants into BarApp.h unchanged. Replace four
letter codes with imported message constants in BarApp.cpp and
BarWindow.cpp. Much nicer than using bare codes.
The new BDeskbar methods are all handled by TBarApp. The getters send
back a reply message containing the bool while the setters fall through
to existing setter cases.
Fixes#13733.
Unlike written in the ticket, the string length is 136 (not 135),
and the code checks for rejects anything greater or equal to the max
value. So set the constant to 137.
At present, does not work (it fails to properly set up interrupts,
resulting in thousands of unhandled ones which all but grinds the system
to a halt) but this at least is some progress.
It is not a good idea to have a thread get an address from the request
cache, while another thread is deleting said address as the cache has
grown too large. Add a lock around the cache access to make it safe.
