IF YOU WOULD LIKE TO GET AN ACCOUNT, please write an
email to Administrator. User accounts are meant only to access repo
and report issues and/or generate pull requests.
This is a purpose-specific Git hosting for
BaseALT
projects. Thank you for your understanding!
Только зарегистрированные пользователи имеют доступ к сервису!
Для получения аккаунта, обратитесь к администратору.
which have a ulimit -n of 100. Drop the number of searches to 20.
That should limit the maximum total fd count to about 76, most of
which is all the listening sockets on the various interfaces
talloc(NULL, xxx) to allocate the registry context. That had two
consequences
1) it was a massive memory leak, as all winreg operations leaked their
entire context (including an open ldb database) every time
2) event_context_find() never found the exsting event context, so we
used a new event context each time, which called epoll_create()
each time, which caused a fd to be allocated
descriptors in smbd on all build farm machines. I suspect we have a fd
leak somewhere, but its hard to track down.
I know part of the problem is the number of (fake) network interfaces
we define in the build farm tests, with each of them listening on a
whole bunch of different protocol ports. That chews up around 48 file
descriptors just to startup. I don't think thats the real problem
though, and I suspect something else is leaking file descriptors on
some hosts.
filesystems and the tests are taking more than 1.5 hours. We will
still be testing the sync code paths, as the ldb tests still run
synchronous, but doing all of our other tests synchronously just costs
too much time.
completes in a minute on my laptop (the benchmark I usually use).
Simo, I removed th ldb tests, as ldb is pretty well tested by lots of
the other tests, and its a pretty slow test. We could instead add a
"ldb-testquick.sh" which does only minimal testing. The full tests
will of course be run with 'make test' and in the build farm.
when I changed the test output to be easier to read, I also broke it
so that it always gives a zero exit status. Fix now, and unfortunately
I expect to see some more red in the farm as a result :(
Jerry, there is a big difference on the wire between these two:
[out] uint32 x;
and
[out] uint32 *x;
if you change from
[out] uint32 x;
then you need to change to:
[out,ref] uint32 *x;
otherwise it changes the format on the wire, which means we are no
longer compatible with MS servers.
but be aware that even if you change to a ref ptr, you also need to
change all the client code to set all the return variables in the out
part of the structure. That's why I don't like the MIDL restriction of
forcing the use of ref pointers for output variables - it makes life
much harder when writing client code, and makes the code much more
error prone (just look at all the extra code needed to make this work
again).
I know we could auto-allocate these variables in the generated client
side NDR code, but if we did that then we would have no way of doing a
_real_ ref out pointer, which we really wanted to set to some already
allocated variable.
So please hold off on changing our idl to use the MIDL convention for
output variables until Jelmer and I have had a good "chat" about this :-)
- change the test for net/if.h to do a full compile, not just an
existance test. net/if.h is completely broken on hpux, and can never
compile (it uses stuff before it defines it), so by using a
AC_TRY_COMPILE() test we avoid using net/if.h on hpux, which should
fix the build
Also cope with the fact that we define the FSTYPE as NTFS by default.
We never use this anywhere else, so we may just change it, but just
detect the fact and return DISK in share_classic for now.
- undefine _XOPEN_SOURCE_EXTENDED for the AIX interface detection test
#define _XOPEN_SOURCE_EXTENDED 1 brings in sa_len to sockaddr on Tru64
which means the AIX code compiles...
metze
try to find the problem on Tru64...where configure says the AIX method finds 1 interface
but later can't compile netif.c. (revision 18486 was the last that detects ifconf with 2 interfaces)
metze
libreplace. This should fix the standalone build of tdb on HPUX, where
we need to blacklist mmap.
Unfortunately this requires that we have a copy of config.guess and
config.sub in each of our project subdirectories. I tried to find a
way to use something like AC_CONFIG_AUX_DIR($libreplacedir) and just
put config.{guess,sub} in the lib/replace/ directory, but I couldn't
figure out how to do that in a way that kept autoconf happy for each
of our separate builds. Any autoconf guru out there see a way to do
this?
caused the RPC-SECRETS test to crash smbd in an inlined version of
this memcmp() call. This patch should have absolutely no effect at
all, but in fact it prevents the crash.
Disassembling at the point of the crash, it shows that gcc is inlining
the memcmp(). I don't know enough MIPS assembler to actually spot the
bug. In case anyone reading this does know MIPS assembler, here is the
gcc generated code that crashes:
0x105e0218 <gssapi_krb5_verify_header+168>: lw $t1,52($sp)
0x105e021c <gssapi_krb5_verify_header+172>: lw $t1,0($t1)
0x105e0220 <gssapi_krb5_verify_header+176>: lhu $t1,0($t1)
0x105e0224 <gssapi_krb5_verify_header+180>: lw $t2,68($sp)
0x105e0228 <gssapi_krb5_verify_header+184>: lhu $t2,0($t2)
0x105e022c <gssapi_krb5_verify_header+188>: subu $t1,$t1,$t2
it gets a segv at 0x105e0220.
lha, what do you think of this? The change should be innocuous on all
other platforms, apart from making the code harder to read :(
