than that of the current thread has been woken up. I didn't see the
reason why the thread should otherwise relinquish the rest of its
quantum. I noticed for instance that client and app server window
threads were ping-ponging more than seemed necessary. In most cases
when the client sent a port message it would be unscheduled although it
had run only for a few microseconds and had still stuff to do.
I measured a relatively Terminal-heavy "find /boot" (second run), which
does now take 5-10% less time.
git-svn-id: file:///srv/svn/repos/haiku/haiku/trunk@27314 a95241bf-73f2-0310-859d-f6bbb57e9c96
unscheduled was incorrect.
* Introduced _kern_analyze_scheduling() syscall. It requires scheduler
kernel tracing to be enabled. It uses the tracing entries for a given
period of time to do a similar analysis the "scheduler" debugger
command does (i.e. number of runs, run time, latencies, preemption
times) for each thread. Additionally the analysis includes for each
thread how long the thread waited on each locking primitive in total.
* Added kernel tracing for the creation of semaphores and initialization
of condition variables, mutexes, and rw locks. The enabling macro is
SCHEDULING_ANALYSIS_TRACING. The only purpose is to provide
_kern_analyze_scheduling() with more info on the locking primitives
(the name in particular).
git-svn-id: file:///srv/svn/repos/haiku/haiku/trunk@27304 a95241bf-73f2-0310-859d-f6bbb57e9c96
tracing buffer entries even when not in the kernel debugger.
git-svn-id: file:///srv/svn/repos/haiku/haiku/trunk@27302 a95241bf-73f2-0310-859d-f6bbb57e9c96
the ipc hash table lock along with the semaphore set hash table lock were
hold, thinking (wrongly) that the semaphore set lock itself was not needed.
What could happen was that another process on semop could have gained the lock
of the set itself, and then release the semaphore set hash table lock.
This would make it think that the set was still valid, while it could have
actually been deleted right after it release the semaphore set hash table lock.
Same would have happened for any other processes waiting on the semaphore set
mutex queue. By calling the lock on the mutex when deleting the set, it
*should be* safe to assume that there is no one else waiting on its queue,
since the list of waiters is handled in a FIFO way.
As far as I can see from the mutex_destroy code, it looks safe to hold the lock
when calling this function. Please confirm.
git-svn-id: file:///srv/svn/repos/haiku/haiku/trunk@27268 a95241bf-73f2-0310-859d-f6bbb57e9c96
for message queue
* Removed unnecessary header Vector.h
* Removed HasSemaphoreSet method: since there will be an IPC table for each
subsystem, if a key exist, it already has a semaphore set associated
with it
git-svn-id: file:///srv/svn/repos/haiku/haiku/trunk@27256 a95241bf-73f2-0310-859d-f6bbb57e9c96
* Implemented _kern_msgget()
Work in progress, some stuff may be removed.
git-svn-id: file:///srv/svn/repos/haiku/haiku/trunk@27254 a95241bf-73f2-0310-859d-f6bbb57e9c96
variable. pthread_cond_broadcast() incorrectly returned an error
when no one was waiting.
git-svn-id: file:///srv/svn/repos/haiku/haiku/trunk@27253 a95241bf-73f2-0310-859d-f6bbb57e9c96
inherit it.
* IOScheduler::ScheduleRequest() uses the request's thread and team now
instead of the current one. Otherwise for requests processed
iteratively this would always be the I/O scheduler's notifier thread.
* Also get the thread's I/O priority now. It's still ignored later,
though.
git-svn-id: file:///srv/svn/repos/haiku/haiku/trunk@27248 a95241bf-73f2-0310-859d-f6bbb57e9c96
vector at the end of the file will be aligned to the given value.
* BFS uses an alignment of 512 bytes (should be block size of the
underlying device or BFS block size, whatever is less), which should
be fine, since file data are only stored in BFS blocks. This totally
avoids any partial operations at the I/O scheduler level, thus saving
disk operations. Not that I could measure any performance difference.
Theoretically it should help a lot though, particularly when dealing
with lots of small files, since we avoid using bounce buffers, which
are (a) limited in number and (b) require copying of the data.
git-svn-id: file:///srv/svn/repos/haiku/haiku/trunk@27246 a95241bf-73f2-0310-859d-f6bbb57e9c96
* swap_file_add() open()s the swap file now with O_NOCACHE and
swap_file_delete() closes it. This squashes a TODO (the file cache
wasn't kept disabled for the swap file before).
* swap_file_add() only adds swap files that can actually be used (i.e.
non 0-sized ones).
git-svn-id: file:///srv/svn/repos/haiku/haiku/trunk@27232 a95241bf-73f2-0310-859d-f6bbb57e9c96
overcommitting caches. If the page in question was just not mapped or
swapped out, we would increase the committment unnecessarily
(potentially even beyond the size of the cache).
git-svn-id: file:///srv/svn/repos/haiku/haiku/trunk@27229 a95241bf-73f2-0310-859d-f6bbb57e9c96
* Implement color palette generation for the boot splash images in the
generate_boot_screen build tool. Only 4-bit screen support is missing now.
* Adopted images.h with the new results from generate_boot_screen.
This should fix black boot screens for graphics cards that don't support
true color modes for the native resolution. I've tried to find the ticket,
#2177 almost looks like the one, but it looks more like the mode is out
of range if I understand the ticket right.
Thanks a lot, David, and sorry it took so very long to apply your patch!
git-svn-id: file:///srv/svn/repos/haiku/haiku/trunk@27223 a95241bf-73f2-0310-859d-f6bbb57e9c96
operation length.
* IORequest: Added owner (IORequestOwner). Also added a SetUnfinished()
method, which is invoked by the I/O scheduler after all operations of
the request have been finished, but the request isn't done yet.
* Added debugger commands "io_request_owner" and "io_scheduler" printing
information for a IORequestOwner and IOScheduler object respectively.
* Implemented an actual I/O scheduling algorithm. It's a simple round
robin strategy (a queue per thread) with a unidirectional elevator
serializing the operations. ATM priorities are ignored, the bandwidth
isn't adjusted to the device, and there are TODOs all over the place.
git-svn-id: file:///srv/svn/repos/haiku/haiku/trunk@27216 a95241bf-73f2-0310-859d-f6bbb57e9c96
* The VMAnonymousCache destructor was unreserving too much physical
memory.
* Addressed TODO in _SwapBlockBuild(): When the swap block couldn't be
allocated we wait and loop until it can.
git-svn-id: file:///srv/svn/repos/haiku/haiku/trunk@27202 a95241bf-73f2-0310-859d-f6bbb57e9c96
left hand side of assignments. IOW it is possible to modify memory, now.
git-svn-id: file:///srv/svn/repos/haiku/haiku/trunk@27193 a95241bf-73f2-0310-859d-f6bbb57e9c96
referencing iframe registers. Their prefix is "$". E.g. "$eax" refers to
the eax register of the current iframe. The features cooperates with the
"in_context" command, i.e. "in_context 92 $eip = 0" will set the eip
register of thread 92 to 0 (thus sealing its fate ;-)).
git-svn-id: file:///srv/svn/repos/haiku/haiku/trunk@27192 a95241bf-73f2-0310-859d-f6bbb57e9c96
debugged thread, if set. This makes "dis" without specified address work
when used with "in_context".
git-svn-id: file:///srv/svn/repos/haiku/haiku/trunk@27190 a95241bf-73f2-0310-859d-f6bbb57e9c96
executing them. Doesn't really make any difference ATM.
* Handle B_BUSY returned by DMAResource::TranslateNext() correctly.
git-svn-id: file:///srv/svn/repos/haiku/haiku/trunk@27186 a95241bf-73f2-0310-859d-f6bbb57e9c96
dynamically assign one when needed. Under the assumption that in most
cases a bounce buffer isn't needed, we can thus prepare a lot more
operations.
git-svn-id: file:///srv/svn/repos/haiku/haiku/trunk@27185 a95241bf-73f2-0310-859d-f6bbb57e9c96
TRACK_PAGE_USAGE_STATS): The page daemon tracks the distribution of page
usage counts. Can be printed using the "page_usage" command.
git-svn-id: file:///srv/svn/repos/haiku/haiku/trunk@27182 a95241bf-73f2-0310-859d-f6bbb57e9c96
RAM Haiku uses (for debugging purposes).
* "page_stats" prints some more infos now.
* page_writer():
- Moved the low_resource_state() invocation out of the inner loop.
Reduces lock contention on the sLowResourceLock recursive lock.
- Additional debug output: Every 1024 written pages the page writer
prints a message.
* steal_pages(): Addressed the TODO: When there should be pages to
steal, but we can't get them ATM, we now timeout on the free pages
condition variable instead of snoozing unconditionally, so that we
wake up earlier when someone frees pages in the meantime.
git-svn-id: file:///srv/svn/repos/haiku/haiku/trunk@27181 a95241bf-73f2-0310-859d-f6bbb57e9c96
due to alignment padding before.
* Reorganized merging of caches a bit. Renamed MergeStore() to Merge()
and moved some more functionality into it. The method also moves the
pages from source to consumer, now. This is necessary, since
VMAnonymousCache needs to consider both physical pages and swap pages
at the same time. Before we first moved the physical pages and the
swap pages later, which was broken for two reasons: (1) A swap page in
the consumer cache shadows a physical page of the source cache, which
we ignored. (2) A source cache's physical page that also had a swap
page would lose the latter in the process when moved to the consumer
cache, i.e. if the page was not marked modified, it could be stolen
and its data would be lost.
These changes improve the situation when building Haiku with 256 MB RAM
in that jam doesn't crash anymore, but in my test the system became
totally unusable after about an hour or 7000 targets (GUI froze). For
some reason it didn't manage to free pages anymore although swapping
heavily.
git-svn-id: file:///srv/svn/repos/haiku/haiku/trunk@27179 a95241bf-73f2-0310-859d-f6bbb57e9c96
more information of the unscheduled thread, namely its new state and, if
waiting, also the object it is going to wait on. When scheduler tracing
level 2 is enabled, additionally the address where the thread was
preempted is recorded. The latter could be interesting for analyzing
race condition bugs.
git-svn-id: file:///srv/svn/repos/haiku/haiku/trunk@27168 a95241bf-73f2-0310-859d-f6bbb57e9c96
* It can now also lookup userland symbols.
* By respecting the currently debugged thread it smoothly cooperates
with the "in_context" command. IOW it can lookup symbols in any team.
git-svn-id: file:///srv/svn/repos/haiku/haiku/trunk@27167 a95241bf-73f2-0310-859d-f6bbb57e9c96
PC of the innermost iframe.
* The "in_context" command does now set the currently debugged thread
respectively.
git-svn-id: file:///srv/svn/repos/haiku/haiku/trunk@27166 a95241bf-73f2-0310-859d-f6bbb57e9c96
debugger and added respective getter/setter methods
debug_{get,set}_debugged_thread(). By default the currently debugged
thread is the thread that dropped into the kernel debugger, but commands
like "in_context" can change it.
git-svn-id: file:///srv/svn/repos/haiku/haiku/trunk@27164 a95241bf-73f2-0310-859d-f6bbb57e9c96
state_to_text().
* state_to_text() can now deal with a NULL thread argument.
git-svn-id: file:///srv/svn/repos/haiku/haiku/trunk@27163 a95241bf-73f2-0310-859d-f6bbb57e9c96
line and executes the command line in the context of the specified
thread. E.g. "in_context <thread> db <addr>" allows to read memory from
a team that is not active ATM.
git-svn-id: file:///srv/svn/repos/haiku/haiku/trunk@27156 a95241bf-73f2-0310-859d-f6bbb57e9c96
first command line argument of a given command line string.
git-svn-id: file:///srv/svn/repos/haiku/haiku/trunk@27155 a95241bf-73f2-0310-859d-f6bbb57e9c96
short form (like "threads").
* "thread" allows to specify more than one thread now.
git-svn-id: file:///srv/svn/repos/haiku/haiku/trunk@27152 a95241bf-73f2-0310-859d-f6bbb57e9c96
that possibility.
* It now has a "remove" argument instead, that decides whether or not the
image has to be removed from the hash still.
* Moved locking to put_module_image(), ie. both, {load|unload}_module_image()
need to be called with the sModulesLock held now.
* module_init_post_boot_device() needs to remove the image from the hash itself,
or else it messes up its hash iterator.
* check_module_image() could call unload_module_image() under certain
conditions - but this was completely wrong, and could have caused crashes.
* search_module()/check_module_image() now keep a reference to the module
image on success, thus a caller doesn't need to call get_module_image()
again afterwards - this also fixes another round of potential module
unloading to happen.
* That reference is leaked for built-in modules in get_module(), but it doesn't
matter for them, anyway.
git-svn-id: file:///srv/svn/repos/haiku/haiku/trunk@27140 a95241bf-73f2-0310-859d-f6bbb57e9c96
as they aren't unloaded when there is no boot device yet.
git-svn-id: file:///srv/svn/repos/haiku/haiku/trunk@27136 a95241bf-73f2-0310-859d-f6bbb57e9c96
dirent without the name first, and will then check if the d_reclen member
is valid before copying the name.
* This shows the problem with the FAT file system that was revealed by r26859.
git-svn-id: file:///srv/svn/repos/haiku/haiku/trunk@27128 a95241bf-73f2-0310-859d-f6bbb57e9c96
synchronously; we use the resource resizer instead. This avoids another
potential deadlock.
git-svn-id: file:///srv/svn/repos/haiku/haiku/trunk@27123 a95241bf-73f2-0310-859d-f6bbb57e9c96
* Added a second kernel daemon service, resource resizer, which is
mainly supposed to be used for resizing allocations asynchrounously.
git-svn-id: file:///srv/svn/repos/haiku/haiku/trunk@27122 a95241bf-73f2-0310-859d-f6bbb57e9c96
without object depot it should be complete. For the other stuff
internal_alloc() would need to pass the flag on to block_alloc(), but
that isn't possible yet.
* Adjusted the low memory handler to respect the minimum object reserve
of the object caches.
The swap_test_heap test does seem to pass reliably with 128 MB RAM and
128 MB allocation, now. It's quite slow, though. Particularly while the
allocation is filled, the system is completely unusable.
git-svn-id: file:///srv/svn/repos/haiku/haiku/trunk@27118 a95241bf-73f2-0310-859d-f6bbb57e9c96
create_area_etc().
* When the new flag CREATE_AREA_DONT_WAIT is specified, the functions
don't wait for memory or pages to become available. They fail
immediately instead.
git-svn-id: file:///srv/svn/repos/haiku/haiku/trunk@27117 a95241bf-73f2-0310-859d-f6bbb57e9c96
eventually the VIP heap would be exhausted.
* WriteAsync() didn't call the provided callback when an error occurred
before invoking vfs_asynchronous_write_pages(). The page writer would
get stuck in those cases.
* The object cache used for the swap blocks does now use the
asynchronous resizing feature to avoid deadlocks.
git-svn-id: file:///srv/svn/repos/haiku/haiku/trunk@27101 a95241bf-73f2-0310-859d-f6bbb57e9c96
of free objects an object cache should try to have ready. If the number
of free objects drops below the threshold, a new urgent priority thread
is asked to asynchronously resize the object cache (pretty similar to
the heap grower thread). Such a mechanism is necessary for code paths
that are supposed to free pages, but may need memory themselves (like
the swap support).
git-svn-id: file:///srv/svn/repos/haiku/haiku/trunk@27100 a95241bf-73f2-0310-859d-f6bbb57e9c96
calls the method unchecked, and the buffer can actually be NULL, if
Init() failed.
* panic() when running out of VIP memory, at least when KDEBUG is
set.
* Use heap_set_get_caller() for the VIP heap, so the heap leak checking
produces useful caller addresses.
git-svn-id: file:///srv/svn/repos/haiku/haiku/trunk@27099 a95241bf-73f2-0310-859d-f6bbb57e9c96
is enabled to set a per-heap get_caller() function.
* Added "-h <heap>" option to the "allocations_per_caller" command. If
given only the allocation for the specified heap are considered.
git-svn-id: file:///srv/svn/repos/haiku/haiku/trunk@27097 a95241bf-73f2-0310-859d-f6bbb57e9c96
we must use {read|write}_pages() instead of {read|write}().
* This should fix accessing other file systems than BFS (ie. those that doesn't
have an io() function) like ext3 and FAT.
git-svn-id: file:///srv/svn/repos/haiku/haiku/trunk@27088 a95241bf-73f2-0310-859d-f6bbb57e9c96
global cache list before locking it. This allows to get rid of the
block_cache::deleting field, as well as simplifies some code.
* This also fixes a possible deadlock I recently introduced (on destruction,
the locking order was wrong).
* Now uses an anonymous condition variable instead.
* Moved the block_cache initialization code into a dedicated method that will
now also fail in case the low resource handler couldn't be registered (as
pointed out by Salvatore).
git-svn-id: file:///srv/svn/repos/haiku/haiku/trunk@27074 a95241bf-73f2-0310-859d-f6bbb57e9c96
* Fixed lock leak in Write().
* Fixed bug in _Commit(): swap_space_reserve() was fed with the wrong
value (could even be negative).
* swap_free_page_swap_space(): Removed incrementing of sAvailSwapSpace.
The function is only supposed to deallocate swap space, not to
unreserve it.
git-svn-id: file:///srv/svn/repos/haiku/haiku/trunk@27071 a95241bf-73f2-0310-859d-f6bbb57e9c96
operation is available ATM, it will return NULL.
* _Finisher() does now re-schedule a request, if all of its operations
finished successfully, but there are still remaining bytes.
* _Scheduler() does now operate in two passes. First it creates as many
operations for a given request as possible, then it executes the
operations. This fixes bug #2644. The problem was that by creating and
executing the operations in a single loop, an operation could be
finished before the next one was added. The request would thus be
considered finished and the request owner be notified. This would
usually lead to the destruction of the request while it was still in
use.
* _Scheduler(): In case we don't have a DMA resource also advance the
request.
git-svn-id: file:///srv/svn/repos/haiku/haiku/trunk@27070 a95241bf-73f2-0310-859d-f6bbb57e9c96
* Moved memory unlocking from IORequest::OperationFinished() to
IORequest::NotifyFinished(). This way we can reschedule a request,
e.g. if we didn't have enough unused IOOperations at hand the first
time.
* Added some more debug output and asserts.
git-svn-id: file:///srv/svn/repos/haiku/haiku/trunk@27069 a95241bf-73f2-0310-859d-f6bbb57e9c96
default).
* Enlarged the serial buffer that is handed over to the kernel to 8192 bytes.
* Minor cleanup.
git-svn-id: file:///srv/svn/repos/haiku/haiku/trunk@27067 a95241bf-73f2-0310-859d-f6bbb57e9c96
* Some renaming: A location in a swap file where a page can be stored is
now called "slot" instead of "page" or "swap page".
* swap_slot_alloc(): Update the hint more correctly after allocating
slots at the hint.
* swap_space_reserve(): When less than the requested space could be
reserved, it always returned 0 and leaked the remaining pages.
* swap_file_delete(): sSwapFileListLock wasn't unlocked in error cases.
Use MutexLocker now.
* swap_free_page_swap_space(): sAvailSwapSpace wasn't updated.
git-svn-id: file:///srv/svn/repos/haiku/haiku/trunk@27057 a95241bf-73f2-0310-859d-f6bbb57e9c96
be passed on to the IORequest. Most relevantly physical pages can now
be written directly by passing B_PHYSICAL_IO_REQUEST.
* Added VMCache::WriteAsync() which is supposed to write pages
asynchronously. The base class version version falls back to the
synchronous Write(). Only VMVnodeCache implements WriteAsync() ATM,
VMAnonymousCache (swap support) still has to be adjusted accordingly.
* write_page() doesn't need to map the page anymore as it can write the
physical page directly.
* Modified the page writer to write pages asynchronously. This shouldn't
have any noticeable effect yet. It will though as soon as the I/O
scheduler reorders I/O operations.
git-svn-id: file:///srv/svn/repos/haiku/haiku/trunk@27056 a95241bf-73f2-0310-859d-f6bbb57e9c96
writes the given page iovecs asynchronously. The new class
AsyncIOCallback is used to inform the caller when the request has been
finished.
git-svn-id: file:///srv/svn/repos/haiku/haiku/trunk@27055 a95241bf-73f2-0310-859d-f6bbb57e9c96
deleted automatically when it's finished.
* Added IORequest::Create() for creating a IORequest on the heap
(respectively the VIP heap).
git-svn-id: file:///srv/svn/repos/haiku/haiku/trunk@27054 a95241bf-73f2-0310-859d-f6bbb57e9c96
many bytes to be read/written, which, among other things, could trigger
an assert in the IORequest code:
* In case of a partial (i.e. non-block-aligned) begin, transferLeft was
not adjusted correctly.
* The main loop was lacking a transferLeft check.
* Main loop: When finally using a bounce buffer, the unrestricted vec
length was used as base length for the bounce buffer.
git-svn-id: file:///srv/svn/repos/haiku/haiku/trunk@27042 a95241bf-73f2-0310-859d-f6bbb57e9c96
stuff at all, anymore.
* Removed now unused get_device_for_path().
* unpublish_driver() now only deletes the device if unpublishing actually
worked out okay.
* Minor cleanup.
git-svn-id: file:///srv/svn/repos/haiku/haiku/trunk@27038 a95241bf-73f2-0310-859d-f6bbb57e9c96
* Added inode ID member to BaseDevice to make this possible.
* Removed unused and unmaintained legacy_driver::devices_published field.
* Implemented legacy driver's unpublish_driver().
* Reenabled legacy driver reloading on changes.
* Renamed devfs_driver_{added|removed}() to driver_{added|removed}(), and
made them private.
* Simplified deletion of device_node lists (no need to use an iterator here),
added device unpublishing.
* Minor cleanup.
git-svn-id: file:///srv/svn/repos/haiku/haiku/trunk@27033 a95241bf-73f2-0310-859d-f6bbb57e9c96
maximal size.
* Turned off debug output accidently enabled with the last commit.
git-svn-id: file:///srv/svn/repos/haiku/haiku/trunk@27006 a95241bf-73f2-0310-859d-f6bbb57e9c96
