Before 2019, the entire ConditionVariable system was "giant"-locked:
that is, there was a single global lock that all ConditionVariable
and ConditionVariableEntry operations had to pass through. This of
course was not very performant on multicore systems and when
ConditionVariables see significant use, so I reworked it then to have
more granular locking.
Those patches took a number of attempts to get right, as having two
objects in separate threads that can each access the other not turn
into a deadlock or use-after-free is not easy to say the least,
and the ultimate solution I came up with erased most of the performance
gains I initially saw on the first (partially broken) patchsets.
So I have wanted to revisit this and see if there was a better way
even since then. Recently there have been a few reports of
ConditionVariable-related panics (apparently double unlocks),
notably #16894, and so that was reason enough to actually revisit
this code and see if a better solution could be found.
Well, I think I have come up with one: after this commit, Entries
no longer have their own lock, and instead accesses to Entry members
are almost always atomic; and there is now a case where we spin inside
Variable::_NotifyLocked as well as one in Entry::_RemoveFromVariable.
This leads to somewhat simpler code (no more lock/unlock dance in Notify),
though it is significantly more difficult to understand the nuances of it,
so I have left a sizable number of comments explaining the intricacies
of the new logic.
Note: I initially tried 1000 for "tries", but on a few instances I did see
the panic hit, strangely. I don't think the code that is waited on can
be reasonably reduced any further, so I have just increased the limit to
10000 (which is still well below what spinlocks use.) Hopefully this suffices.
Quick benchmark, x86, compiling HaikuDepot and the mime_db in VMware, 2 cores:
before:
real 0m23.627s
user 0m25.152s
sys 0m7.319s
after:
real 0m23.962s
user 0m25.229s
sys 0m7.330s
Though I occasionally I saw sys times as low as 7.171s, so this seems
to be at least not a regression if not a definitive improvement.
Change-Id: Id042947976885cd5c1433cc4290bdf41b01ed10e
Reviewed-on: https://review.haiku-os.org/c/haiku/+/4727
Tested-by: Commit checker robot <no-reply+buildbot@haiku-os.org>
Reviewed-by: Alex von Gluck IV <kallisti5@unixzen.com>
Check RX buffer status when GetChar is called in no-wait mode.
This fixes an 'infinite keypress' issue that used to occur
after 16 keypresses.
Change-Id: I47762de387b07c4fed46cc192cd3b81fdabfb270
Reviewed-on: https://review.haiku-os.org/c/haiku/+/4732
Reviewed-by: Adrien Destugues <pulkomandy@gmail.com>
GCC still assumes that the dirent has no data past the end for some
scenarios here and still mis-optimizes things. Therefore, drop the
usages of unions altogether, and instead use a casted character array.
Additionally, use B_FILE_NAME_LENGTH for the array, not B_PATH_NAME_LENGTH,
and make sure to add 1 for the NULL terminator.
The lock entry is the first thing in the struct, so this is a no-op
change, but it is safer to do in case of changes, of course.
Spinlocks have been structures for quite a long time, so this was
probably just missed in the conversion.
* Move IsDownloadComplete call into DownloadFileRequest; this way
we will always revalidate checksums even if the file is fully
downloaded instead of skipping that.
* Treat ERANGE as a "bad data" error in PackageManager (it usually means
we have a size mismatch between the local and the server's file)
* If we fail checksum validation or get ERANGE, and we reused a download,
immediately try again without reusing the download. This fixes most
problems that previously required you to delete old transaction
directories.
Change-Id: Ia3079655691b871e0b206e366b59fca0f832c63d
Reviewed-on: https://review.haiku-os.org/c/haiku/+/4730
Reviewed-by: waddlesplash <waddlesplash@gmail.com>
This way, we will get a more coherent crash instead of
an unceremonious one. Follow-up to #17389.
Change-Id: Iffbf421ce85d638628243d5785ba61ff6b9a8043
Reviewed-on: https://review.haiku-os.org/c/haiku/+/4729
Reviewed-by: waddlesplash <waddlesplash@gmail.com>
we don't sample if the last sample is too recent and use the cached result.
Change-Id: I17ed29bda7fe7276f1a4148b3e1985c9d32ae032
Reviewed-on: https://review.haiku-os.org/c/haiku/+/4101
Reviewed-by: Jérôme Duval <jerome.duval@gmail.com>
Reviewed-by: Adrien Destugues <pulkomandy@gmail.com>
This file should ideally contain only those things needed
across all system headers, even POSIX ones, and all other
declarations (B_* ones especially) should go in SupportDefs.h.
However, as nothing but riscv64 uses this right now, I've just
moved it to there.
GCC 11 treats [1] as a fixed-length array and not a flexible-length
array, and so some things that used direct strcmp("..", ent->d_name),
for instance, would be optimized out as being always unequal,
which was the cause of #17389. Using a real FLA informs GCC that
there is going to be more than one byte of data, and thus this
fixes that bug.
BeOS used [1] and not [0], possibly because it had to deal with
compilers (MetroWerks? Early GCC2?) that did not support FLAs.
GCC 2.95 does, using [0], and GCC 4 does, using [], so we can go
with that here.
(I did try using [0] for both, which seems to be OK with GCC 11,
but GCC 8 throws errors when d_name is dereferenced directly
as being-out-of-bounds. So, we have to use the #if here and give
newer GCC the [] syntax and not [0] to avoid that problem.)
The real question probably is whether or not we should backport
some variant of these changes to R1/beta3, as software at HaikuPorts
very well may run in to the same issue. (The alternative workaround
is to compile with -O1 and not -O2 for any affected software.) But
maybe this is an argument for keeping with the beta4 schedule of
this coming January...
At present, it does, but that is an oddity we have preserved from BeOS
that the next commit is going to remove. (This commit thus wastes 1 byte
without the following one.)
Most changes are pretty straightforward: only a +1 is needed,
and a few removed from sizing calculations. Some filesystems like UDF
originally passed back the length with the \0 included, so they have
been adjusted further. UFS2 had some other sizing problems which are also
corrected in this commit.
Our dirent structure is "slim": it has a flexible-length array at the
end which must be allocated to whatever size the consumer wants. However,
we use [1] there and not [0] or [], which meant GCC thought it was not
a flexible-length array, and so it optimized various string accesses
that it assumed must be always false. Among these was BDirectory's
check for "." and "..", and so that resulted in infinite loops.
When changing our dirent structure to a proper FLA instead of [1],
GCC then throws errors on LongDirEntry as it has data "after" the
FLA; which is what we want, but there is no way to tell GCC that.
So now we use a union instead, which is the proper way to statically
allocate a FLA.
This is part of #17389, but the real fix requires changing our dirent
structure, which is coming in a separate commit.
