| Age | Commit message (Collapse) | Author |
|---|
|
A number of memory allocation optimizations have been implemented. Most
optimizations reduce contention caused by synchronization between
threads during allocation and deallocation of memory. Most notably:
* Synchronization of memory management in scheduler specific allocator
instances has been rewritten to use lock-free synchronization.
* Synchronization of memory management in scheduler specific
pre-allocators has been rewritten to use lock-free synchronization.
* The 'mseg_alloc' memory segment allocator now use scheduler specific
instances instead of one instance. Apart from reducing contention
this also ensures that memory allocators always create memory
segments on the local NUMA node on a NUMA system.
|
|
Conflicts:
erts/aclocal.m4
erts/emulator/beam/erl_db.c
erts/emulator/sys/win32/sys.c
erts/include/internal/ethread_header_config.h.in
|
|
* pg/fix-hibernate-scheduling-with-hipe:
Fix bug related to hibernate and HiPE (clear F_HIBERNATE_SCHED flag)
OTP-9452
|
|
All uses of the old deprecated atomic API in the runtime system
have been replaced with the use of the new atomic API. In a lot of
places this change imply a relaxation of memory barriers used.
|
|
F_HIBERNATE_SCHED flag that was introduced in b7ecdcd1ae9e11b8f75e must
be cleared in hipe_mode_switch as well. Otherwise, processes running
HiPE code that hibernate, wake up and then trap into a BIF will not be
rescheduled.
|
|
|
|
|
|
|
|
* sverker/erts_printf-halfword:
erts_printf %be to print integers of size Eterm
Fix use of type BeamInstr in hipe_debug.c
Conflicts:
erts/emulator/hipe/hipe_debug.c
|
|
* ks/hipe-ppc64:
Enable HiPE by default when compiling for PPC64
Translate RTL to PPC code on PPC64 too
Changes in ppc files for PPC64
Additions for the PPC64 backend
Changes for the PPC64 backend
Added loader for ppc64
New files for the 64-bit backends
Cleanup tags
OTP-9198
|
|
Existing %bp to print pointer size integers does not work in halfword
emulator to print Eterm size integers.
|
|
|
|
|
|
* pg/hipe-remove-constants-pool:
Remove hipe constants pool
OTP-9128
|
|
|
|
|
|
|
|
|
|
|
|
Hipe constants used to be allocated within a single, fixed-size pool for
interaction with the garbage collector. However, the garbage collector
no longer depends on constants being allocated within a single pool, and
the fixed size of the pool both meant unnecessary allocations on most
deployments and crashes on deployments requiring more constants.
The code was simplified to directly invoke erts_alloc.
Debugging and undocumented function hipe_bifs:show_literals/0 was
removed (it returned true and output text to the console), and
debugging and undocumented function hipe_bifs:constants_size/0 was
rewritten with a global to count the size of allocated constants.
|
|
This commit fixes four related bugs:
- calling hibernate/3 using a dynamic call would fail with badarg
as hibernate/3 as a BIF was not implemented. hibernate/3 is generally
provided as a Beam instruction, and code is translated to use this
instruction when loaded.
- calling hibernate/3 from HiPE would fail with badarg because this
would call the aforementioned BIF which was not implemented.
- calling hibernate/3 with some HiPE-native garbage in the process heap
would randomly crash at the next garbage collect. This bug only
happened in a complex, yet reproduceable scenarios, where native code
calls beam code that calls hibernate/3, and the process has some
garbage when being hibernated and the process generates garbage when
awaken.
- when entering HiPE, the process current_function can be set and be
inaccurate.
The fix is three folded:
- hibernate_3 BIF now actually works instead of throwing a badarg. While
hibernate_3 BIF was (usually) not called from BEAM, it is called from
HiPE. hibernate behaviour is very close to the scheduler and this is why
it is implemented as an instruction in BEAM. The fix consists in doing
the actual hibernation (through the now exported erts_hibernate
function) and setting the process flag to TRAP as well as the process
status to P_WAITING. On BIF epilogue in both BEAM and HiPE, this status
is tested on TRAP and if set, the scheduler is invoked. The i_hibernate
instruction and translation code is now redundant and could be deleted.
- hibernation now also empties the HiPE native stack, with a new
function hipe_empty_nstack provided by Mikael Pettersson.
- when entering HiPE through hipe_mode_switch, p->current is cleared,
as suggested by Mikael Pettersson. p->current normally hold a pointer to
the {M,F,A} of the current function if it exists. When hibernating, it
is set to {erlang,hibernate,3}, and all stdlib hibernate tests
(gen_server_SUITE:hibernate/1, proc_lib_suite:hibernate/1, etc.)
actually rely on this information. Clearing p->current fixes the tests
and avoids the surprise one might have when querying the process info
of a process that hibernated and woke up in a native function.
Non-regression tests are provided, a test for the dynamic call as well
as a Makefile-handled duplication of the hibernate_SUITE into
hibernate_native_SUITE for the HiPE case.
|
|
In the HiPE part of the runtime system's Process struct
there is a state field which is 32 bits wide even on 64-bit
machines.
There is a single instruction in the HiPE AMD64 runtime
where this field is incorrectly written with a 64-bit store.
Luckily the extraneous 32 bits are written as zeros to 4
bytes of tail-padding at the end of the struct, so nothing
should have broken because of this.
The same bug exists in the HiPE PowerPC64 runtime (in
development), but on the big-endian PPC64 the effect is
to write the actual value to the tail-padding and zero
to the struct field, which potentially breaks TRAPs from
BIFs (depending on BIF arities and how many parameter
registers the runtime has been configured to use).
Thanks to Paul Guyot for noticing the oversized write on AMD64.
|
|
* pg/fix-hipe-crash-in-gc_after_bif:
Fix call to erts_gc_after_bif_call in hipe glue
|
|
* mk/net-dragonfly-bsd-patches:
Remove unused variables
Use proper install method
Add support for DragonFly BSD
Add support for NetBSD
|
|
R12B-0 changed the signature of erts_gc_after_bif_call and it now
takes 4 parameters instead of 2 in R11B-5. Yet, the glue code was not
updated accordingly. As a result, the function erts_gc_after_bif_call
was called with garbage and would randomly cause a crash later in the
garbage collector code.
The fix consists in passing NULL and 0 for the third and fourth
parameters, since there is no term to add to rootset, recovering the
behaviour of R11B-5
(see otp_src_R11B-5/erts/emulator/beam/erl_gc.c, line 314).
(Includes assembly language fixes and code style improvements
suggested by Mikael Pettersson.)
|
|
The HiPE runtime system has a hipe_bifs:get_hrvtime/0 BIF which
mimics the non-standard gethrvtime() C API. It's possible to
configure the implementation to use the "perfctr" Linux kernel
extension for performance-monitoring counters, in which case
get_hrvtime has very high precision and low overhead. Otherwise
it uses the same code as runtime(statistics).
This patch changes the get_hrvtime implementation to do a runtime
check to see if perfctr is available, and to use the fallback code
rather than returning a dummy value if perfctr is unavailable,
which is common.
The current dummy value return is a bug. It messes up the API
and either breaks callers (they get badarg when trying to compute
on the value) or forces them to implement checks and fallbacks
themselves. Timing code in HiPE's test suites and benchmarks
is known to be affected.
|
|
These are the current NetBSD pkgsrc patches.
|
|
It's been reported that HiPE-enabled Erlang VMs running on BSD
systems sometimes generate messages like
Yikes! erts_alloc() returned misaligned address 0x8016a512c
These originate from hipe_bif0.c:hipe_bifs_alloc_data_2().
A native code module has an associated data area of some
size and alignment. In the case where the size is zero,
the alignment is irrelevant, but the allocation BIF checks
it anyway. The warning then triggers on systems where
malloc(0) returns blocks with less alignment than we
(erroneously) expected.
The fix is to simply skip the allocation in this case and
return NULL. The loader won't actually use the address in
this case so that's safe. This is also an optimization since
it avoids allocating memory that cannot be used, and it avoids
fragmenting the system heap with useless tiny blocks.
A second problem is that the warning message failed to
identify its origin. Fixed by prefixing the message by
the BIF's name rather than the silly Yikes! string.
Tested and confirmed to solve the original reporter's problem.
|
|
Merging the three off-heap lists (binaries, funs and externals) into
one list. This reduces memory consumption by two words (pointers) per
ETS object.
|
|
New NIF features:
Send messages from a NIF, or from thread created by NIF, to any local
process (enif_send)
Store terms between NIF calls (enif_alloc_env, enif_make_copy)
Create binary terms with user defined memory management
(enif_make_resource_binary)
|
|
|
|
* gc/hipe_darwin_amd64:
Fix hipe memory allocation problems on darwin/amd64
Porting x86 darwin fixes to amd64 darwin hipe asm/m4 code
Automatically enable hipe for darwin/amd64 builds
Allow configure to enable_hipe for darwin/amd64 builds
OTP-8416 HiPE now works in the 64-bit emulator on Mac OS X. (Thanks to
Geoff Cant.)
|
|
Compared to GNU as, Mac OS X's assembler uses different directives for
text and global sections, and omits type and size directives
entirely. We also need to mangle symbol names in order to allow
linking with C object files.
|
|
* mp/hipe-smp-fixes:
work around hipe_mfa_info_table lock omission
fix hipe loader SMP non-atomicity error
OTP-8397 The loading of native code was not properly atomic in the SMP
emulator, which could cause crashes. Also a per-MFA information
table for the native code has now been protected with a lock
since it turns that it could be accessed concurrently in the SMP
emulator. (Thanks to Mikael Pettersson.)
|
|
HiPE maintains per-MFA information such as native code entry
point in a table. This table was thought to be read-only at
runtime, except when the loader populates it, so it employed
no locking. That turned out to be incorrect: if there is an
apply of a previously unseen MFA, a native code stub for that
MFA is created and recorded in the table, causing it to grow.
Work around this for now by slapping a mutex around accesses
to that table.
Signed-off-by: Mikael Pettersson <[email protected]>
|
|
|
