Age | Commit message (Collapse) | Author | Files | Lines |
|
rate from 20k/sec to 55k/sec. The irpc rate goes from 16k/sec to
34k/sec.
I should have run strace -T on this a long time ago.
(This used to be commit b9281668bb0c971af14df37ec3e979b9d5ef276e)
|
|
compared to about 20k messages/sec for the raw messaging layer. I
think that is quite acceptable given the extra functionality.
(This used to be commit a05d38d1d91f1f54d3e3794a596b468992594852)
|
|
management system I proposed on samba-technical a couple of days
ago. Essentially it is a very lightweight way for any code in Samba to
make IDL based rpc calls to anywhere else in the code, without the
client or server having to go to the trouble of setting up a full rpc
service.
It can be used with any of our existing IDL, but I expect it will
mostly be used for a new set of Samba specific management calls.
The LOCAL-IRPC torture test demonstrates how it can be used by calling
the echo_AddOne() call over this transport.
(This used to be commit 3d589a09954eb8b318f567e1150b0c27412fb942)
|
|
(This used to be commit ca3c4e961713ffd8952fca328e27196ba2ccc0c9)
|
|
- avoid a memcpy
(This used to be commit 753839c358f45c02927b137092828f3a79158083)
|
|
send on the unix domain datagram socket. This gains us about 50% in
speed, and also means that we don't run the risk of running out of
file descriptors due to heavy messaging traffic. We now use a single
file descriptor no matter how many messages are pending to any number
of servers.
(This used to be commit 2369170fc1b4ff9a48284779fa6d3a6254177ec1)
|
|
(This used to be commit d465ff67bd10340ecd46480039f483daa82c7ebe)
|
|
sockets. This gains us about 40% in messaging speed.
(This used to be commit f244a64ed537447e44229172427b5b6a5c64800c)
|
|
(This used to be commit b902ea546d2d1327b23f40ddaeeaa8e7e3662454)
|
|
less likely that anyone will use pstring for new code
- got rid of winbind_client.h from includes.h. This one triggered a
huge change, as winbind_client.h was including system/filesys.h and
defining the old uint32 and uint16 types, as well as its own
pstring and fstring.
(This used to be commit 9db6c79e902ec538108d6b7d3324039aabe1704f)
|
|
(This used to be commit 7f54c8a339f36aa43c9340be70ab7f0067593ef2)
|
|
gettimeofday() call
and just use timeval_current() when its actually needed
(This used to be commit 236403cc4dc2924ed6a898acae0bb44cc1688dcc)
|
|
make it possible to add optimisations to the events code such as
keeping the next timed event in a sorted list, and using epoll for
file descriptor events.
I also removed the loop events code, as it wasn't being used anywhere,
and changed timed events to always be one-shot (as adding a new timed
event in the event handler is so easy to do if needed)
(This used to be commit d7b4b6de51342a65bf46fce772d313f92f8d73d3)
|
|
servers in smbd. The old code still contained a fairly bit of legacy
from the time when smbd was only handling SMB connection. The new code
gets rid of all of the smb_server specific code in smbd/, and creates
a much simpler infrastructures for new server code.
Major changes include:
- simplified the process model code a lot.
- got rid of the top level server and service structures
completely. The top level context is now the event_context. This
got rid of service.h and server.h completely (they were the most
confusing parts of the old code)
- added service_stream.[ch] for the helper functions that are
specific to stream type services (services that handle streams, and
use a logically separate process per connection)
- got rid of the builtin idle_handler code in the service logic, as
none of the servers were using it, and it can easily be handled by
a server in future by adding its own timed_event to the event
context.
- fixed some major memory leaks in the rpc server code.
- added registration of servers, rather than hard coding our list of
possible servers. This allows for servers as modules in the future.
- temporarily disabled the winbind code until I add the helper
functions for that type of server
- added error checking on service startup. If a configured server
fails to startup then smbd doesn't startup.
- cleaned up the command line handling in smbd, removing unused options
(This used to be commit cf6a46c3cbde7b1eb1b86bd3882b953a2de3a42e)
|
|
large commit. I thought this was worthwhile to get done for
consistency.
(This used to be commit ec32b22ed5ec224f6324f5e069d15e92e38e15c0)
|
|
control of the event, so instead build that into the function. If you
pass NULL as mem_ctx then it leaves it as a child of the events
structure.
(This used to be commit 7f981b9ed96f39027cbfd500f41e0c2be64cbb50)
|
|
complexity was that events didn't automatically cleanup
themselves. This was because the events code was written before we had
talloc destructors, so you needed to call event_remove_XX() to clean
the event out of the event lists from every piece of code that used
events. I have now added automatic event destructors, which in turn
allowed me to simplify a lot of the calling code.
The 2nd source of complexity was caused by the ref_count, which was
needed to cope with event handlers destroying events while handling
them, which meant the linked lists became invalid, so the ref_count ws
used to mark events for later destruction.
The new system is much simpler. I now have a ev->destruction_count,
which is incremented in all event destructors. The event dispatch code
checks for changes to this and handles it.
(This used to be commit a3c7417cfeab429ffb22d5546b205818f531a7b4)
|
|
to that the destructor don't double free the registered events
when the event context is first in the talloc_free() hierarchie
fixes a bug with process_model_thread and the talloc_steal(conn, ev)
metze
(This used to be commit 05c3d1c4a6a6350cc8b5fb2ee8201ae6feed3d3d)
|
|
- change smbcli_read/write to take void * for the buffers to match read(2)/write(2)
all this fixes a lot of gcc-4 warnings
metze
(This used to be commit b94f92bc6637f748d6f7049f4f9a30b0b8d18a7a)
|
|
deferred reply is short-circuited immediately when the file is
closed by another user, allowing it to be opened by the waiting user.
- added a sane set of timeval manipulation routines
- converted all the events code and code that uses it to use struct
timeval instead of time_t, which allows for microsecond resolution
instead of 1 second resolution. This was needed for doing the pvfs
deferred open code, and is why the patch is so big.
(This used to be commit 0d51511d408d91eb5f68a35e980e0875299b1831)
|
|
(This used to be commit c6f486574470a311e0d336c026103f131451e21e)
|
|
ioctl.h)
(This used to be commit b97e395c814762024336c1cf4d7c25be8da5813a)
|
|
(This used to be commit 264ce9181089922547e8f6f67116f2d7277a5105)
|
|
var/locks/smbd.tmp/
and deletes that dir on startup.
(This used to be commit 7e942e7f1bd2c293a0e6648df43a96f8b8a2a295)
|
|
you set this option (either on the command line using --option or in
smb.conf) then every socket recv or send will return short by random
amounts. This allows you to test that the non-blocking socket logic in
your code works correctly.
I also removed the flags argument to socket_accept(), and instead made
the new socket inherit the flags of the old socket, which makes more
sense to me.
(This used to be commit 406d356e698da01c84e8aa5b7894752b4403f63c)
|
|
The main change is to make socket_recv() take a pre-allocated buffer,
rather than allocating one itself. This allows non-blocking users of
this API to avoid a memcpy(). As a result our messaging code is now
about 10% faster, and the ncacn_ip_tcp and ncalrpc code is also
faster.
The second change was to remove the unused mem_ctx argument from
socket_send(). Having it there implied that memory could be allocated,
which meant the caller had to worry about freeing that memory (if for
example it is sending in a tight loop using the same memory
context). Removing that unused argument keeps life simpler for users.
(This used to be commit a16e4756cd68ca8aab4ffc59d4d9db0b6e44dbd1)
|
|
(This used to be commit cc93813e4a09c538ad485dc2b3cb4c9be34f3d18)
|
|
destructor
(This used to be commit ab222b236a091d31b1f5f2cba150a11585ab5836)
|
|
messaging_deregister()
- added a pvfs_lock_close_pending() hook to remove pending locks on file close
- fixed the private ptr argument to messaging_deregister() in pvfs_wait
- fixed a bug in continuing lock requests after a lock that is blocking a pending lock is removed
- removed bogus brl_unlock() call in lock continue
- corrected error code for LOCKING_ANDX_CHANGE_LOCKTYPE
- expanded the lock cancel test suite to test lock cancel by unlock and by close
- added a testsuite for LOCKING_ANDX_CHANGE_LOCKTYPE
(This used to be commit 5ef80f034d4aa4dd6810532c63ad041bfc019cb8)
|
|
This adds a pvfs_wait_message() routine which uses the new messaging
system, event timers and talloc destructors to give a nice generic
async event handling system with a easy to use interface. The
extensions to pvfs_lock.c are based on calls to pvfs_wait_message()
routines.
We now pass all of our smbtorture locking tests, although while
writing this code I have thought of some additonal tests that should
be added, particularly for lock cancel operations. I'll work on that
soon.
This commit also extends the smbtorture lock tests to test the rather
weird 0xEEFFFFFF locking semantics that I have discovered in
win2003. Win2003 treats the 0xEEFFFFFF boundary as special, and will
give different error codes on either side of it. Locks on both sides
are allowed, the only difference is which error code is given when a
lock is denied. Anyone like to hazard a guess as to why? It has
me stumped.
(This used to be commit 4395c0557ab175d6a8dd99df03c266325949ffa5)
|
|
full. This means callers can just "send and forget" rather than
having to check for a temporary failure. The mechanism takes nice
advantage of the timed events handling is our events code. A message
will only fail now if we completely run out of some resource (such
as memory).
- changed the test code not to do retries itself, but only to warn on real failures
(This used to be commit 8cddc610a25e64c1ad39dd6a2fc2e7f467e04fc9)
|
|
makes things a bit more efficient
(This used to be commit 8380225d326e4bfb3f15fddc72c097870713132a)
|
|
(This used to be commit 13739b68d8357d5d330f12b851d0311feb81e545)
|
|
- added the new messaging system, based on unix domain sockets. It
gets over 10k messages/second on my laptop without any socket
cacheing, which is better than I expected.
- added a LOCAL-MESSAGING torture test
(This used to be commit 3af06478da7ab34a272226d8d9ac87e0a4940cfb)
|