Age | Commit message (Collapse) | Author |
|
This is a partial revert of
bf7c722bd3b30e04a56c2d369fd9f5501a8316d0.
It does not make sense to treat 0 as a wild card in cmp function
for the hash table as creation is part of the hash calculation.
|
|
Previously, all ETS tables used centralized counter variables to keep
track of the number of items stored and the amount of memory
consumed. These counters can cause scalability problems (especially on
big NUMA systems). This commit adds an implementation of a
decentralized counter and modifies the implementation of ETS so that
ETS tables of type ordered_set with write_concurrency enabled use the
decentralized counter. [Experiments][1] indicate that this change
substantially improves the scalability of ETS ordered_set tables with
write_concurrency enabled in scenarios with frequent `ets:insert/2`
and `ets:delete/2` calls.
The new counter is implemented in the module erts_flxctr
(`erts_flxctr.h` and `erts_flxctr.c`). The module has the suffix
flxctr as it contains the implementation of a flexible counter (i.e.,
counter instances can be configured to be either centralized or
decentralized). Counters that are configured to be centralized are
implemented with a single counter variable which is modified with
atomic operations. Decentralized counters are spread over several
cache lines (how many can be configured with the parameter
`+dcg`). The scheduler threads are mapped to cache lines so that there
is no single point of contention when decentralized counters are
updated. The thread progress functionality of the Erlang VM is
utilized to implement support for linearizable snapshots of
decentralized counters. The snapshot functionality is used by the
`ets:info/1` and `ets:info/2` functions.
[1]: http://winsh.me/ets_catree_benchmark/flxctr_res.html
|
|
|
|
erts: Optimize arithmetic ops using overflow intrinsics
|
|
* john/merge-re-replace-fix:
erts: Skip ERTS_NOINLINE on non-GCC-compatible compilers
cleanup: beam_emu.c: use ERTS_NOINLINE instead of NOINLINE
stdlib: fix re:replace on LTO builds
|
|
This reverts commit 321dc6ee0241f802c940def174c0a77262e11f21.
|
|
This reverts commit 0b74eec2a1fd5716ba6e21d8c6715433d8d34510.
|
|
This reverts commit bd8f6106d44a58c261920eef72842bb3bc5a4968.
PLUS a little change in epmd_srv.c:750 ("4" -> "replylen")
that was part of e2cf4a8a4b03b9f430ba228276c3b2629159e832
by mistake.
|
|
The wrong allocator type was given to erts_free.
|
|
Complementory fix to 922fd355831575965
|
|
The edep->data values need to be 64-bit aligned as it
contains 64-bit values.
|
|
|
|
|
|
Cannot do unaligned word writes on sparc!
|
|
|
|
Add erts_debug:interpreter_size/0 for retrieving the approximate
size of the BEAM interpreter (process_main()).
|
|
|
|
__declspec(noinline) works fine on MSVC but requires us to place
the macro before a function rather than after, which in turn causes
early versions of GCC to puke since they only accept __attribute__
at the end of a function declaration.
Since this is a new macro that previously only saw use in beam_emu,
I figured it's easiest to leave it disabled on MSVC.
|
|
Fail when we cannot encode term in binary instead of producing a
faulty result.
|
|
Measurements show that i_minus instructions (subtraction) are
frequent enough to warrant creating specialized i_minus instructions.
Thanks to José Valim for doing instruction counting on
Elixir code.
|
|
PR 2194, stdlib: fix re:replace on LTO builds
* trofi/master:
cleanup: beam_emu.c: use ERTS_NOINLINE instead of NOINLINE
stdlib: fix re:replace on LTO builds
|
|
* john/kernel/cuddle-file-suite:
file_name_SUITE: Darwin 18 broke support for 'icky' names
file_SUITE: Ignore EILSEQ on Darwin in +fnu/latin1 test
|
|
Signed-off-by: Sergei Trofimovich <[email protected]>
|
|
Fabio Coatti reported elixir build failure in https://bugs.gentoo.org/681778.
The minimal reproducer looks like that (from otp git tree):
$ ./configure CFLAGS='-O2 -flto' LDFLAGS='-O2 -flto=8'
$ make
$ ERL_TOP=$PWD \
PATH=$ERL_TOP/bin:$PATH \
\
bin/erl \
\
-noshell -eval 're:replace("a","b","c",[{return,list}]).' \
-s erlang halt
{"init terminating in do_boot",{badarg,[{re,replace,["a","b","c",[{return,list}]],
[{file,"re.erl"},{line,362}]},
{erl_eval,do_apply,6,[{file,"erl_eval.erl"},{line,680}]},
{init,start_it,1,[]},
{init,start_em,1,[]},
{init,do_boot,3,[]}]}}
init terminating in do_boot ({badarg,[{re,replace,[[_],[_],[_],[_]],[{_},{_}]},
{erl_eval,do_apply,6,[{_},{_}]},{init,start_it,1,[]},{init,start_em,1,[]},{init,do_boot,3,[]}]})
Crash dump is being written to: erl_crash.dump...done
The failure happens in libpcre2 where stack overflow is mis-identified
at function entry of
erts_pcre_compile2()
compile_regex()
if (PUBL(stack_guard) != NULL && PUBL(stack_guard)())
{
*errorcodeptr= ERR85;
return FALSE;
}
The stack "overflow" detection happens in
thr_wrapper()
ethr_set_stacklimit__()
because the stack usage code relies on the fact that ethr_set_stacklimit__()
and similar functions don't get inlined into callers for stack growth
measurement.
Before the change inlining avoidance was achieved by putting functions
into standalone translation units. LTO makes this technique inefficient.
The change marks functions explicitly as __attribute__((__noinline__)) on gcc.
Reported-by: Fabio Coatti
Bug: https://bugs.gentoo.org/681778
Signed-off-by: Sergei Trofimovich <[email protected]>
|
|
* sverker/process_info-reductions-fix/OTP-15709:
erts: Fix bug in process_info(reductions)
erts: Use ptr_val for pointer in gc msg copy
|
|
|
|
* sverker/process_info-reductions-fix/OTP-15709:
erts: Fix bug in process_info(reductions)
erts: Use ptr_val for pointer in gc msg copy
|
|
This commit fixes a bug that could cause a crash or memory usage to
grow until the machine ran out of memory when adding a key-value pair
to a map. This could happen when inserting a new key-value pair with a
key K1 containing a binary B1 into a map M having a key K2 with a
binary B2 if the following conditions were met:
* size(B1) >= 4294967296
* size(B2) >= 4294967296
* size(M) >= 32
* (size(B1) rem 4294967296) == (size(B2) rem 4294967296)
* the first (size(B1) rem 4294967296) bytes are the same both in B1
and B2
* substituting B1 in K1 with B2 would result in a term with the same
value as K2
The root cause of the bug is that the map implementation only hashed
the first (X modulo 4294967296) bytes of binaries so that different
binaries could get hashed to the same hash value independently of the
hash seed.
|
|
There's no way to create the file in question, so we may as well
skip the test.
|
|
returning incorrect result as
* current process might not be RUNNING in which case REDS_IN
is actually used as def_arg_reg[5]
* FCALLS might not have been swapped out
* the SAVED_CALLS case was wrong and returned number of reds left
|
|
Cherry-pick: 3d7d66b84b69f3eb42e8bcd8dc510b6676e31cac
Without this fix asserts would trigger in debug build
but nothing else would break.
|
|
* john/erts/remove-destructive-bs_get_binary2/ERL-901:
erts: Remove unsafe bs_get_binary2 optimization from loader
|
|
A load-time optimization assumed that match contexts had no further
uses when a bs_get_binary2 overwrote the match context's register,
and figured it would be safe to reuse the match context's memory
for the resulting binary.
This is no longer safe as of OTP 22, as a match context may be
reused after being passed to another function.
|
|
* john/erts/seq_trace-sigq-clearing/OTP-15704:
erts: Handle all signal types when clearing seq_trace token
|
|
We walk through all signal queues in the system when clearing
seq_trace, so we must be prepared to encounter all kinds of
signals.
|
|
Fix the unsafe load-time optimization introduced in 07bdbb3a1edc.
https://bugs.erlang.org/browse/ERL-899
|
|
* john/erts/explain-empty-stack-check-in-next_catch:
erts: Add an explanation for the empty stack check in next_catch
|
|
This reverts commit 0118046f9bda8dcf3594184fa439b79267db8017.
|
|
|
|
* sverker/bug-fixing:
erts: Fix erts_debug:set_internal_state(reds_left)
erts: Fix binary_SUITE:cmp_old_impl
|
|
to avoid failed ERTS_DBG_CHK_REDS by clearing virtual_reds.
|
|
'lukas/erts/fragment-dist-messages/OTP-13397/OTP-15610/OTP-15611/OTP-15612/OTP-15613/OTP-15703'
* lukas/erts/fragment-dist-messages/OTP-13397/OTP-15610/OTP-15611/OTP-15612/OTP-15613/OTP-15703: (26 commits)
erts: Yield when exiting/free process is suspended by de
doc: Fix Design Princ, statem incorrect anchor
erts: Include external msg in need of GC
erts: Include dist header in return from dist_ctrl_get_data
erts: Fix so that exit/down terms stay alive
erts: Pending signals can be for free processes
erts: Fix faulty assert in reference_table_term
etp: Don't crash etp-stack* when c_p->i is null
erts: Adjust dist obuf size correctly after hdr fin
erts: Fix non-payload dist exit signals
erts: Always do trylock on proc locks when crash dumping
Fix tests to work better in debug emulator
erts: erts_factory_proc_init should not set hole marker
erts: Skip heavy process tab tests in debug emu
ts: Use same dynlinking logic for all bsd
Revert "erts: Always run fds check after each testcase"
erts: Always run fds check after each testcase
erts: Always stop any testnodes before testcase exits
erts: Add crash dumping of EXITING and FREE processes
wx: Remove ERL_FLAGS from Makefile erlc command
...
|
|
|
|
|
|
When many external messages suddenly appear in the mailbox
the young gen size is adjusted accordingly but it was
immediately shrunk as the data was not counted towards the
shrink size. This commit includes the ext dist size in the
shrink calculation which means that the decode of the external
messages will not trigger as many GCs which means much better
performance.
|
|
|
|
When the GC is run or the link/monitor node is deleted,
the terms associated with it will also be deallocated
so we have to make sure that we don't do any GCs and that
all link/monitor nodes are copied onto the heap before
the node is deallocated.
|
|
Since we no schedule free processes multiple times, pending
signals can be from free as well as current processes.
|
|
|
|
|