| Age | Commit message (Collapse) | Author |
|---|
|
uncached. To be used in the drm code.
ok mpi@
|
|
comments
ok millert@
|
|
Same diff as guenther@ committed for alpha.
ok guenther@
|
|
|
|
pv lists with a mutex. This should make pmap_enter(9), pmap_remove(9) and
pmap_page_protect(9) safe to use without holding the kernel lock.
ok visa@, mpi@, deraadt@
|
|
* pool_allocator_single: single page allocator, always interrupt safe
* pool_allocator_multi: multi-page allocator, interrupt safe
* pool_allocator_multi_ni: multi-page allocator, not interrupt-safe
ok deraadt@, dlg@
|
|
|
|
segement size constraint.
Make xhci(4) work on my G5.
ok dlg@
|
|
symbols in ASM
ok deraadt@ mpi@
|
|
ok deraadt@
|
|
support on powerpc as it uses a non-executable GOT and PLT.
"start slamming stuff in" deraadt@
|
|
there's no real functional advantage to this, except that it will
make it easier to add deadlock detection to the code.
this is modelled on the c mutex implementation thats on alpha,
mips64, and hppa.
ok mpi@ kettenis@
|
|
faultbuf. But 1/ sr was only restored for machine check exceptions, and 2/ the
way it was saved was unsafe if interrupts were enabled, and could cause %r2
to be lost.
Discussing this with deraadt@ at the end of c2k15, this was probably needed
for the old VI boards which were the target of the original powerpc port,
came with a worse-than-Genesi openfirmware. Since then, machine check
exceptions have been unheard of; or, if they happen, they do not need the
status register to be restored.
ok mpi@ deraadt@
|
|
machines work again with the unlocked reaper.
ok mpi@, deraadt@
no objection from miod@
|
|
|
|
"shared reference pointers".
srp allows concurrent access to a data structure by multiple cpus
while avoiding interlocking cpu opcodes. it manages its own reference
counts and the garbage collection of those data structure to avoid
use after frees.
internally srp is a twisted version of hazard pointers, which are
a relative of RCU.
jmatthew wrote the bulk of a hazard pointer implementation and
changed bpf to use it to allow mpsafe access to bpfilters. however,
at s2k15 we were trying to apply it to other data structures but
the memory overhead of every hazard pointer would have blown out
significantly in several uses cases. a bulk of our time at s2k15
was spent reworking hazard pointers into srp.
this diff adds the srp api and adds the necessary metadata to struct
cpuinfo on our MP architectures. srp on uniprocessor platforms has
alternate code that is optimised because it knows there'll be no
concurrent access to data by multiple cpus.
srp is made available to the system via param.h, so it should be
available everywhere in the kernel.
the docs likely need improvement cos im too close to the implementation.
ok mpi@
|
|
ok mpi@
|
|
|
|
ok mpi@
|
|
|
|
|
|
reused by a CPU while another CPU is manipulating it.
This races occurs because the virtual spill handlers are run without
taking the KERNEL_LOCK for obvious reasons. So use a per-pmap mutex
that CPUs must hold when modifying a pted in order to guarantee the
atomicity of operations *and* the coherence between pmap VPs tree and
what's in the HASH.
Thanks to dlg@ for assisting me debugging this. This change ends your
PowerPC pmap SMP show of the week. GENERIC.MP on macppc should now be
stable enough to build ports without corrupting its own memory.
ok kettenis@, deraadt@, dlg@
|
|
Needed for upcoming locking.
|
|
Since this lock is recursive we can now guarantee the atomicity of
pte_inser{32,64}() when a pted has to be removed first. This fixes
one of the races.
Using a __mp_lock here also allowed dlg@ to provide me useful traces
to fix the next race. Thanks for your help!
ok kettenis@, deraadt@, dlg@
|
|
it. This will reduce the number of places to audit for locking.
Note that for profiling purposes pte_spill_v() is now marked a __noprof
since per-CPU profiling buffers are not guaranteed to be 1:1 mapped and
cannot be accessed from the real mode fault handler.
ok kettenis@, deraadt@, dlg@
|
|
Document every operation, make sure to call "sync" when appropriate so
that other CPUs see the bit changes and finally grab a lock where it was
missing to grantee atomicity.
ok kettenis@, deraadt@, dlg@
|
|
This should not introduce any behavior change but makes the code easier
to read and later easier to protect. This also brings this pmap closer
to what others do.
Thanks to kettenis@ for spotting a bad typo!
ok kettenis@, deraadt@, dlg@
|
|
If you wonder why pte_insert{32,64}() is not using pmap_hash_remove() if
it finds a conflicting PTE in the HASH, it's because in the current state
trying to grab the same lock a second time would lead to a deadlock.
This is much easier to reproduce on G5 (or G4 with BAT disabled).
ok kettenis@, deraadt@, dlg@
|
|
|
|
|
|
if a PTE is present in the HASH.
Note that atomicity is currently not guaranteed between this check and
the following operations.
ok kettenis@, deraadt@, dlg@
|
|
that does not call pmap_vp_lookup().
Carreful readers would have notice the removal of the bits on the virtual
address with a page mask, this change allows me to find the 13 years old
bug fixed in r1.145.
ok kettenis@, deraadt@, dlg@
|
|
This simplify pmap_remove() & friends by re-using an already fetched PTE
descriptor.
There's currently a race on MP system where one CPU can reuse a pted
while another one is still trying to insert it in the HASH. This commit
starts reducing the number of pmap_vp_lookup() calls to help fix this
race.
ok kettenis@, deraadt@, dlg@
|
|
Even if this change is not strickly needed, because the memory will be
returned to the pool it helped me track the use-after-free.
|
|
By instructing spl(9) calls on MP machines I figured out that their high
cost was hiding a race condition involving PTE reuse in our pmap. Thanks
to deraadt@ for finding a way to trigger such panic by adding a couple of
splvm().
This should make the races easier to trigger but will be addressed
shortly.
This commit starts your PowerPC pmap SMP show of the week.
ok kettenis@, deraadt@, dlg@
|
|
sub, inc, and dec) fixed.
the asm read the value from memory into a register, added to it,
and then tried to write it back. after succeeding it doesnt have
to add again before returning.
this splits sub, inc, and dec off from add. sub can use the subf
opcode, and inc and dec can use the addic opcode. explicitely
identify where the modified memory is so we can avoid using "memory"
as a clobber.
ok mpi@
|
|
delete coredump_trad(), uvm_coredump(), cpu_coredump(), struct md_coredump,
and various #includes that are superfluous.
This leaves compat_linux processes without a coredump callback. If that
ability is desired, someone should update it to use coredump_elf32() and
verify the results...
ok kettenis@
|
|
had a proper stdint.h. No ports fallout. OK guenther@ miod@
|
|
if callers of save_vec() weren't expected to pass a non NULL pointer
as an argument.
ok kettenis@
|
|
This makes the secondary cpu of my PowerMac as fast as the primary one,
and divide the build time by 3 with a GENERIC.MP kernel on MP G5s
Found thanks to MP kernel profiling.
ok dlg@, miod@
|
|
device_unref() as found by deraadt@.
|
|
of UVM on PowerPC architectures by breaking pmap_is_referenced() and
friends.
ok kettenis@
|
|
include extra sync operations.
ok kettenis@
|
|
ok mpi@ deraadt@
|
|
processors that support it.
Due to the way trap code is patched it is currently not possible to
enabled/disable BAT at runtime.
ok miod@, kettenis@
|
|
ok miod@, kettenis@ as part of a larger diff
|
|
Use this on vax to correctly pick the end of the stack area now that the
stackgap adjustment code will no longer guarantee it is a fixed location.
|
|
with and ok miod@
|
|
each arch used to have to provide an rw_cas operation, but now we
have the rwlock code build its own version. on smp machines it uses
atomic_cas_ulong. on uniproc machines it avoids interlocked
instructions by using straight loads and stores. this is safe because
rwlocks are only used from process context and processes are currently
not preemptible in our kernel. so alpha/ppc/etc might get a benefit.
ok miod@ kettenis@ deraadt@
|
|
|
