glusterfs.git/rpc/rpc-lib/src/rpcsvc.c, branch v8.2

posix: Implement a janitor thread to close fd

2020-08-21T10:38:07+00:00

Problem: In the commit fb20713b380e1df8d7f9e9df96563be2f9144fd6 we use
         syntask to close fd but we have found the patch is reducing the
         performance

Solution: Use janitor thread to close fd's and save the pfd ctx into
          ctx janitor list and also save the posix_xlator into pfd object to
          avoid the race condition during cleanup in brick_mux environment

Change-Id: Ifb3d18a854b267333a3a9e39845bfefb83fbc092
Fixes: #1396
Signed-off-by: Mohit Agrawal 
(cherry picked from commit 41b9616435cbdf671805856e487e373060c9455b)

Posix: Use simple approach to close fd

2020-03-20T04:08:42+00:00

Problem: posix_release(dir) functions add the fd's into a ctx->janitor_fds
         and janitor thread closes the fd's.In brick_mux environment it is
         difficult to handle race condition in janitor threads because brick
         spawns a single janitor thread for all bricks.

Solution: Use synctask to execute posix_release(dir) functions instead of 
          using background a thread to close fds.

Credits: Pranith Karampuri 
Change-Id: Iffb031f0695a7da83d5a2f6bac8863dad225317e
Fixes: bz#1811631
Signed-off-by: Mohit Agrawal

feature/changelog: Avoid thread creation if xlator is not enabled

2020-02-09T12:42:56+00:00

Problem:
Changelog creates threads even if the changelog is not enabled

Background:
Changelog xlator broadly does two things
  1. Journalling - Cosumers are geo-rep and glusterfind
  2. Event Notification for registered events like (open, release etc) -
     Consumers are bitrot, geo-rep

The existing option "changelog.changelog" controls journalling and
there is no option to control event notification and is enabled by
default. So when bitrot/geo-rep is not enabled on the volume, threads
and resources(rpc and rbuf) related to event notifications consumes
resources and cpu cycle which is unnecessary.

Solution:
The solution is to have two different options as below.
 1. changelog-notification : Event notifications
 2. changelog : Journalling

This patch introduces the option "changelog-notification" which is
not exposed to user. When either bitrot or changelog (journalling)
is enabled, it internally enbales 'changelog-notification'. But
once the 'changelog-notification' is enabled, it will not be disabled
for the life time of the brick process even after bitrot and changelog
is disabled. As of now, rpc resource cleanup has lot of races and is
difficult to cleanup cleanly. If allowed, it leads to memory leaks
and crashes on enable/disable of bitrot or changelog (journal) in a
loop. Hence to be safer, the event notification is not disabled within
lifetime of process once enabled.

Change-Id: Ifd00286e0966049e8eb9f21567fe407cf11bb02a
Updates: #475
Signed-off-by: Mohit Agrawal

pcsvc: fix subnet_mask_v4 check

2019-11-27T07:16:24+00:00

The check we had for subnet mask validation wasn't checking in
proper sequence. Corrected the order of calling `inet_pton()` as
the fix.

Fixes: #765
Change-Id: I5d31468eb917aa94cbb85f573b37c60023e9daf3
Signed-off-by: Amar Tumballi

rpc: align structs

2019-10-17T06:34:27+00:00

squash tens of warnings on padding of structs in afr structures.
The warnings were found by manually added '-Wpadded' to the GCC
command line.

Also made relevant structs and definitions static, where it
was applicable.

Change-Id: Ib71a7e9c6179378f072d796d11172d086c343e53
updates: bz#1193929
Signed-off-by: Yaniv Kaul

multiple files: another attempt to remove includes

2019-06-14T16:50:32+00:00

There are many include statements that are not needed.
A previous more ambitious attempt failed because of *BSD plafrom
(see https://review.gluster.org/#/c/glusterfs/+/21929/ )

Now trying a more conservative reduction.
It does not solve all circular deps that we have, but it
does reduce some of them. There is just too much to handle
reasonably (dht-common.h includes dht-lock.h which includes
dht-common.h ...), but it does reduce the overall number of lines
of include we need to look at in the future to understand and fix
the mess later one.

Change-Id: I550cd001bdefb8be0fe67632f783c0ef6bee3f9f
updates: bz#1193929
Signed-off-by: Yaniv Kaul

Fix some "Null pointer dereference" coverity issues

2019-05-26T13:59:13+00:00

This patch fixes the following CID's:

  * 1124829
  * 1274075
  * 1274083
  * 1274128
  * 1274135
  * 1274141
  * 1274143
  * 1274197
  * 1274205
  * 1274210
  * 1274211
  * 1288801
  * 1398629

Change-Id: Ia7c86cfab3245b20777ffa296e1a59748040f558
Updates: bz#789278
Signed-off-by: Xavi Hernandez

rpclib: slow floating point math and libm

2019-04-03T04:40:15+00:00

In release-6 rpc/rpc-lib (libgfrpc) added the function
get_rightmost_set_bit() which calls log2(3), a call that takes
a floating point parameter.

It's used thusly:
    right_most_unset_bit = get_rightmost_set_bit(...);

(So is it really the right-most unset bit, or the right-most set bit?)

It's unclear to me whether this is in the data path or not. If it is,
it's rather scary to think about integer-to-float conversions and slow
calls to libm functions in the data path.

gcc and clang have __builtin_ctz() which returns the same result as
get_rightmost_set_bit(), and does it substantially faster. Approx
20M iterations of get_rightmost_set_bit() took ~33sec of wall clock
time on my devel machine, while 20M iterations of __builtin_ctz()
took < 9sec; get_rightmost_set_bit() is 3x slower than __builtin_ctz().

And as a side benefit, we can again eliminate the need to link libgfrpc
with libm.

Change-Id: If9e7e80874577c52223f8125b385fc930de20699
updates: bz#1193929
Signed-off-by: Kaleb S. KEITHLEY

rpc/transport: Missing a ref on dict while creating transport object

2019-03-20T13:24:44+00:00

while creating rpc_tranpsort object, we store a dictionary without
taking a ref on dict but it does an unref during the cleaning of the
transport object.

So the rpc layer expect the caller to take a ref on the dictionary
before passing dict to rpc layer. This leads to a lot of confusion
across the code base and leads to ref leaks.

Semantically, this is not correct. It is the rpc layer responsibility
to take a ref when storing it, and free during the cleanup.

I'm listing down the total issues or leaks across the code base because
of this confusion. These issues are currently present in the upstream
master.

1) changelog_rpc_client_init

2) quota_enforcer_init

3) rpcsvc_create_listeners : when there are two transport, like tcp,rdma.

4) quotad_aggregator_init

5) glusterd: init

6) nfs3_init_state

7) server: init

8) client:init

This patch does the cleanup according to the semantics.

Change-Id: I46373af9630373eb375ee6de0e6f2bbe2a677425
updates: bz#1659708
Signed-off-by: Mohammed Rafi KC

core: implement a global thread pool

2019-02-18T02:58:24+00:00

This patch implements a thread pool that is wait-free for adding jobs to
the queue and uses a very small locked region to get jobs. This makes it
possible to decrease contention drastically. It's based on wfcqueue
structure provided by urcu library.

It automatically enables more threads when load demands it, and stops
them when not needed. There's a maximum number of threads that can be
used. This value can be configured.

Depending on the workload, the maximum number of threads plays an
important role. So it needs to be configured for optimal performance.
Currently the thread pool doesn't self adjust the maximum for the
workload, so this configuration needs to be changed manually.

For this reason, the global thread pool has been made optional, so that
volumes can still use the thread pool provided by io-threads.

To enable it for bricks, the following option needs to be set:

   config.global-threading = on

This option has no effect if bricks are already running. A restart is
required to activate it. It's recommended to also enable the following
option when running bricks with the global thread pool:

   performance.iot-pass-through = on

To enable it for a FUSE mount point, the option '--global-threading'
must be added to the mount command. To change it, an umount and remount
is needed. It's recommended to disable the following option when using
global threading on a mount point:

   performance.client-io-threads = off

To enable it for services managed by glusterd, glusterd needs to be
started with option '--global-threading'. In this case all daemons, like
self-heal, will be using the global thread pool.

Currently it can only be enabled for bricks, FUSE mounts and glusterd
services.

The maximum number of threads for clients and bricks can be configured
using the following options:

   config.client-threads
   config.brick-threads

These options can be applied online and its effect is immediate most of
the times. If one of them is set to 0, the maximum number of threads
will be calcutated as #cores * 2.

Some distributions use a very old userspace-rcu library (version 0.7)
for this reason, some header files from version 0.10 have been copied
into contrib/userspace-rcu and are used if the detected version is 0.7
or older.

An additional change has been made to io-threads to prevent that threads
are started when iot-pass-through is set.

Change-Id: I09d19e246b9e6d53c6247b29dfca6af6ee00a24b
updates: #532
Signed-off-by: Xavi Hernandez