The test application lets run a thread at the highest priority that calls
yield all the time - the system stays responsible when it runs, so it seems
to work fine :)
Changed the malloc implementation to use _kern_thread_yield() instead of
snoozing.
We should think about making this call public, too.
git-svn-id: file:///srv/svn/repos/haiku/haiku/trunk@16166 a95241bf-73f2-0310-859d-f6bbb57e9c96
enqueueing a thread to the run queue.
This mechanism is now used for the thread priority boost on semaphore
release. Also, those threads are no longer made real time threads, they
now get a temporary priority of B_FIRST_REAL_TIME_PRIORITY - 1.
git-svn-id: file:///srv/svn/repos/haiku/haiku/trunk@16165 a95241bf-73f2-0310-859d-f6bbb57e9c96
device tree for PCI controllers and make them known to the bus manager,
if we know how to talk with them. ATM we support only the UniNorth chip,
which can be found in G4 Macs (code ported from FreeBSD).
As far as I can judge it, all attached devices are identified correctly
on all three host bridges of my Mac mini.
git-svn-id: file:///srv/svn/repos/haiku/haiku/trunk@16102 a95241bf-73f2-0310-859d-f6bbb57e9c96
* Moved the Open Firmware function platform_get_next_device() from
the boot loader into the kernel (renamed to of_get_next_device()).
git-svn-id: file:///srv/svn/repos/haiku/haiku/trunk@16101 a95241bf-73f2-0310-859d-f6bbb57e9c96
* BEntry::Remove() now uses _kern_remove_dir() for directories.
* Added fd parameter to _kern_remove_dir().
* Fixed LibBeAdapter's _kern_unlink() to only work on files, and
added _kern_remove_dir() for directories.
git-svn-id: file:///srv/svn/repos/haiku/haiku/trunk@16078 a95241bf-73f2-0310-859d-f6bbb57e9c96
can now safely unmount volumes that are still in use by some applications.
Minor fixes to the FD disconnection implementation:
* put_fd() checked the condition for being able to disconnect a descriptor
incorrectly (causing the FD to never be disconnected).
* remove_fd() would hand out disconnected descriptors (but should have
returned NULL for them).
git-svn-id: file:///srv/svn/repos/haiku/haiku/trunk@15952 a95241bf-73f2-0310-859d-f6bbb57e9c96
code. The stack pointer was not adjusted, hence we were
overwriting the previous register value. But it looks like I
missed to check in the arch_cpu.h with the iframe structure
including the floating point registers anyway.
* Backported the ELF PPC relocation code from the boot loader to
the kernel.
* Fixed the PPC version of arch_thread_switch_kstack_and_call().
Apparently the signature had changed, but the assembly
implementation was not adjusted accordingly.
* sc prints more registers now (LR, CR, CTR, XER,...).
* Fixed several occurences of not-working fault handlers.
Apparently the compiler realized, that the "error" label was
never jumped to (by the code it knew), and optimized the
respective code away. Now we use a trick to make it think the
error label might actually be jumped to. I wonder whether the
x86 version has the same problem when being compiled with GCC4.
* Adopted the x86 page fault handling interrupt code.
git-svn-id: file:///srv/svn/repos/haiku/haiku/trunk@15933 a95241bf-73f2-0310-859d-f6bbb57e9c96
We don't do anything with it yet, though, so the BIOS will probably ignore us since
we are supposed to poll for events.
git-svn-id: file:///srv/svn/repos/haiku/haiku/trunk@15900 a95241bf-73f2-0310-859d-f6bbb57e9c96
* Cloned iframe stack management from x86.
* Reimplemented arch_thread_{get,set}_current_thread(). The
thread structure is stored in SPRG2. It is set to NULL in
arch_cpu_preboot_init(), now. A non-null current thread
causes all kinds of undesired behavior in early boot code.
* We establish the address space mappings we know from the
Open Firmware as areas. At least those in kernel address
space. The ones in userland address space are tougher.
Fortunately on my Mac mini there aren't any save the
boot_loader stack, which is not needed any longer anyway.
* Added stack trace support to the kernel debugger. Mostly
cloned and adjusted the x86 code. Some bits are still
missing, like stack traces for other threads.
git-svn-id: file:///srv/svn/repos/haiku/haiku/trunk@15890 a95241bf-73f2-0310-859d-f6bbb57e9c96
void after turning off BAT for the segment containing itself.
The monster macro for the exception vector code was not really
elegant besides being too long for the 32 byte performance
monitor exception slot. Furthermore wasting three of the SPRG*
registers as cheap scratch memory wasn't that nice either.
We now have a three-step approach: The exception vectors
themselves contain only five instructions which branch to common
code at the beginning of the same physical page. That one sets
up BAT for itself, turns address translation back on and jumps
into the kernel. There we turn off BAT again, dump an iframe,
and enter the actual exception handler (/dispatcher). Upon return
the registers are restored from the iframe and we get back to the
place where the exception occurred.
git-svn-id: file:///srv/svn/repos/haiku/haiku/trunk@15881 a95241bf-73f2-0310-859d-f6bbb57e9c96
The time base conversion factor is the 32 bit value
2^32 * 1000000 / time base frequency,
so the system time can be computed by
system time = time base * conversion factor / 2^32.
The expression in system_time() looks more complicated now, but is
actually much faster (factor 2.5 on my Mac mini). I'm positively
surprised, how good the assembly looks, that GCC 4 generates. There's
not that much potential for optimization by hand-coding the function.
git-svn-id: file:///srv/svn/repos/haiku/haiku/trunk@15863 a95241bf-73f2-0310-859d-f6bbb57e9c96
parts, too. Fixed a potential overflow.
* The generic physical page mapper reserves the virtual address range
for the IO space now, so that noone can interfere until an area has
been created. The location of the IO space is no longer fixed; it
didn't look to me like it was necessary for x86, and we definitely
need to be flexible for PPC.
git-svn-id: file:///srv/svn/repos/haiku/haiku/trunk@15855 a95241bf-73f2-0310-859d-f6bbb57e9c96
now run under BeOS as well (as long as they don't use any functions that are not
available under R5).
The solution is a bit messy, but we have to live with it :-)
The runtime loader now patches the __gRuntimeLoader symbol in libroot.so to point
to its exported structure instead of passing it to the init functions as an
argument.
(Hax0red by axeld and bonefish on stippi's assimilated machine -- resistence is futile)
git-svn-id: file:///srv/svn/repos/haiku/haiku/trunk@15848 a95241bf-73f2-0310-859d-f6bbb57e9c96
device in the Open Firmware implementation of boot loader and
pass its path to the kernel, where it's opened and used for
getting/setting the real time. The expensive atomic_*64() on PPC
32-bit make things a bit more complicated. Moreover, missing
64 bit multiplication and division instructions won't really
allow system_time() to be anywhere near as fast as on x86. :-/
git-svn-id: file:///srv/svn/repos/haiku/haiku/trunk@15837 a95241bf-73f2-0310-859d-f6bbb57e9c96
substructure now (that's the only member actually). The system time
offset is therefore accessed via architecture specific accessor
functions.
Note, that this commit breaks the PPC build. Since I want to rename at
least one file I've already changed, I can't avoid that.
git-svn-id: file:///srv/svn/repos/haiku/haiku/trunk@15835 a95241bf-73f2-0310-859d-f6bbb57e9c96
(year, month,...) representation out of the x86 specific code and put
respective support functions into real_time_clock.c. We'll need those
for the PPC specific part too.
git-svn-id: file:///srv/svn/repos/haiku/haiku/trunk@15827 a95241bf-73f2-0310-859d-f6bbb57e9c96
Basically the architecture specific code is now responsible to
init and make use of the platform specific code, now. The reason
being that we have only one kernel per platform and thus cannot
decide at compile time, which platform to use (if any).
The PPC implementation features an abstract base class PPCPlatform
(implemented for all supported platforms) through which platform
support is provided.
git-svn-id: file:///srv/svn/repos/haiku/haiku/trunk@15824 a95241bf-73f2-0310-859d-f6bbb57e9c96
bus, and time base frequency) in the PPC boot loader, and propagate
them to the kernel via kernel_args.
* Now we use the correct time base frequency for timer calculations.
* Implemented PPC specific system info stuff. Added a few PPC CPU
types to <OS.h>.
git-svn-id: file:///srv/svn/repos/haiku/haiku/trunk@15817 a95241bf-73f2-0310-859d-f6bbb57e9c96
remapping stuff into separate functions and made them available to
others.
* Remap the exception handler space in arch_int_init_post_vm() into the
kernel address space (same issue as with the page table).
git-svn-id: file:///srv/svn/repos/haiku/haiku/trunk@15783 a95241bf-73f2-0310-859d-f6bbb57e9c96
According to the spec we need to set it before taking over the MMU,
but we can't call it before arch_mmu_init(), since we need the OF
to allocate the page table. So we do it after we have allocated
the new page table.
* Added PPC specific kernel_args: The virtual address ranges we want
to keep in the kernel. We fill that in with the translations we
find when initializing the MMU stuff. We remove the memory the
boot loader occupies from those. Besides the stack for the boot
loader only the OF stuff remains.
* arch_mmu_allocate() now starts to search at KERNEL_BASE for a free
virtual address when no particular address is requested. This saves
us further trouble in the kernel, since those allocations would
need to be remapped otherwise.
git-svn-id: file:///srv/svn/repos/haiku/haiku/trunk@15780 a95241bf-73f2-0310-859d-f6bbb57e9c96
the fact that I couldn't find ptesync in an otherwise more complete
documentation I downloaded yesterday made me suspicious.
* arch_cpu_global_TLB_invalidate() uses tlbia now. The instruction is
optional, but so is tlbie (how I understood it is that both exist,
when the architecture implementation has a TLB). And the former loop
looked just scary.
* Implemented arch_cpu_user_TLB_invalidate(). It does just the same as
arch_cpu_global_TLB_invalidate().
* Some changes with respect to synchronization required on page table
and segment register updates.
* Some more minor renaming. Pulled a new function
remove_page_table_entry() out of unmap_tmap().
* In arch_vm_translation_map_init_post_area() we do now remap the page
table into the kernel address space, if it was without before. The
page table might actually be a good application for BAT, though.
git-svn-id: file:///srv/svn/repos/haiku/haiku/trunk@15773 a95241bf-73f2-0310-859d-f6bbb57e9c96
{HAIKU,HOST,TARGET}_KERNEL_PIC_{CC,LINK}FLAGS which define the
compiler/linker flags specifying the kind of position independence
the kernel shall have. For x86 we had and still have -fno-pic, but the
PPC kernel has -fPIE (position independent executable) now, as we
need to relocate it.
* The boot loader relocates the kernel now. Mostly copied the relocation
code from the kernel ELF loader. Almost completely rewrote the PPC
specific relocation code, though. It's more correct and more complete now
(some things are still missing though).
* Added boot platform awareness to the kernel. Moved the generic
Open Firmware code (openfirmware.c/h) from the boot loader to the kernel.
* The kernel PPC serial debug output is sent to the console for the time
being.
* The PPC boot loader counts the CPUs now and allocates the kernel stacks
(made OF device iteration a bit more flexible on the way -- the search
can be restricted to subtree). Furthermore we really enter the kernel...
(Yay! :-) ... and crash in the first dprintf() (in the atomic_set()
called by acquire_spinlock()). kprintf() works, though.
git-svn-id: file:///srv/svn/repos/haiku/haiku/trunk@15756 a95241bf-73f2-0310-859d-f6bbb57e9c96
specifying whether only the exact supplied address is acceptable. If
false, the address is considered a hint only. It will be picked, if
available, otherwise a greater address is tried to be acquired, and
as last resort any address. This feature is only implemented for PPC.
It is needed since the preferred kernel text base address 0x80000000
might not be available (and actually isn't on my Mac mini).
* Fixed a bug in the PPC memory management code:
is_{virtual,physical}_allocated() were checking whether the given
range was completely contained by an existing range instead of
checking for intersection. As a consequence we could (and did) allocate
a range intersecting with already allocated ranges. The kernel segment
thus overwrote OF memory for instance.
* The ELF loader makes sure that it got both text and data segment of
the image to be loaded.
The PPC boot loader successfully loads kernel and modules now. Next
comes the hard part, I'm afraid.
git-svn-id: file:///srv/svn/repos/haiku/haiku/trunk@15708 a95241bf-73f2-0310-859d-f6bbb57e9c96
IP, and UDP, as well as a home brewn UDP based protocol, "remote disk",
which provides random access to a single remote file/device. The Open
Firmware flavored boot loader automatically initializes the net stack,
searches for a remote disk, and tries to boot from it, if the boot
device is a network device (e.g. when loading the boot loader via
TFTP).
This is quite nice for developing with a two-machine setup, since one
doesn't even need to install Haiku on the test machine anymore, but can
serve it directly from the development machine. When the networking
support in the kernel is working, this method could even be used to
fully boot, not just for loading kernel and initial modules.
git-svn-id: file:///srv/svn/repos/haiku/haiku/trunk@15689 a95241bf-73f2-0310-859d-f6bbb57e9c96
vm_cache_ref starts with a reference count of 1. When acquiring a vm_cache,
you no longer need to worry if that should go through the vm_store, or not;
as it now always does.
* map_backing_store() no longer needs to play with the vm_cache_ref
references.
* that simplified some code.
* vfs_get_vnode_cache() now grabs a reference to the cache, if successful.
* better balanced vnode ownership on vnode_store creation (vnode_store
released the vnode before if its creation failed).
git-svn-id: file:///srv/svn/repos/haiku/haiku/trunk@15641 a95241bf-73f2-0310-859d-f6bbb57e9c96
this saves 4 bytes per page. To compensate the loss of bytes, the offset is now
stored in page size units, that's enough to address 2^44 or 16 TB (which is now
the maximal supported file size!).
* Renamed vm_page::ppn to physical_page_number.
git-svn-id: file:///srv/svn/repos/haiku/haiku/trunk@15637 a95241bf-73f2-0310-859d-f6bbb57e9c96
arch/*/thread_struct.h to arch_thread_types.h, so that it can directly
be included without having to specify the architecure.
git-svn-id: file:///srv/svn/repos/haiku/haiku/trunk@15616 a95241bf-73f2-0310-859d-f6bbb57e9c96
them apart (this even saves a pointer from vm_virtual_map to its address space)
* aspace -> address_space
* vm_create_address_space() did not check if creating the semaphore succeeded
* Removed team::kaspace - was not really needed (introduced a new vm_kernel_address_space()
function that doesn't grab a reference to the address space)
* Removed vm_address_space::name - it was just a copy of the team name, anyway,
and there is always only one address space per team
* Removed aspace_id - the address space is now using the team_id
* Some cleanup.
git-svn-id: file:///srv/svn/repos/haiku/haiku/trunk@15609 a95241bf-73f2-0310-859d-f6bbb57e9c96
written the page, we now do it before, so that it cannot lose any changed data
anymore; it doesn't matter if the page is written to while writing it back, the
worst thing that can happen is that we write the same page twice. Also, we don't
rely on the PAGE_MODIFIED bit anymore, we now check all mappings of that page
to find all modified pages, no matter how far the (currently disabled) page
daemon had come.
Also, destroying an area will now result in writing back changed pages - this
is only really important for memory mapped files, though, and should probably
be avoided for other vm_store types.
Minor cleanup.
git-svn-id: file:///srv/svn/repos/haiku/haiku/trunk@15597 a95241bf-73f2-0310-859d-f6bbb57e9c96
