vfs: Remove aio_linux
authorVolker Lendecke <vl@samba.org>
Fri, 10 Nov 2017 11:09:36 +0000 (12:09 +0100)
committerJeremy Allison <jra@samba.org>
Sat, 11 Nov 2017 03:49:27 +0000 (04:49 +0100)
Triggered by https://bugzilla.samba.org/show_bug.cgi?id=13128 I think
this module should go. Once Linux aio will do what Samba needs, this
might be worth another look.

What we should instead do soon is support Linux preadv2 and the
RWF_NOWAIT flag to avoid the thread context switch whenever possible.

Signed-off-by: Volker Lendecke <vl@samba.org>
Reviewed-by: Jeremy Allison <jra@samba.org>
WHATSNEW.txt
docs-xml/manpages/vfs_aio_linux.8.xml [deleted file]
docs-xml/wscript_build
source3/modules/vfs_aio_linux.c [deleted file]
source3/modules/wscript_build
source3/wscript

index cc03f2016c8995a583e0a4267c461d984ed8d547..475ebcbdc840d38135f8d8f355f676e7a9a14559 100644 (file)
@@ -65,6 +65,15 @@ commands have been removed from rpcclient.
 "net rpc vampire keytab" from Active Directory domains continues to be
 supported.
 
+vfs_aio_linux module removed
+============================
+
+The current Linux kernel aio does not match what Samba would
+do. Shipping code that uses it leads people to false
+assumptions. Samba implements async I/O based on threads by default,
+there is no special module required to see benefits of read and write
+request being sent do the disk in parallel.
+
 KNOWN ISSUES
 ============
 
diff --git a/docs-xml/manpages/vfs_aio_linux.8.xml b/docs-xml/manpages/vfs_aio_linux.8.xml
deleted file mode 100644 (file)
index 3a00950..0000000
+++ /dev/null
@@ -1,115 +0,0 @@
-<?xml version="1.0" encoding="iso-8859-1"?>
-<!DOCTYPE refentry PUBLIC "-//Samba-Team//DTD DocBook V4.2-Based Variant V1.0//EN" "http://www.samba.org/samba/DTD/samba-doc">
-<refentry id="vfs_aio_linux.8">
-
-<refmeta>
-       <refentrytitle>vfs_aio_linux</refentrytitle>
-       <manvolnum>8</manvolnum>
-       <refmiscinfo class="source">Samba</refmiscinfo>
-       <refmiscinfo class="manual">System Administration tools</refmiscinfo>
-       <refmiscinfo class="version">4.7</refmiscinfo>
-</refmeta>
-
-
-<refnamediv>
-       <refname>vfs_aio_linux</refname>
-       <refpurpose>implement async I/O in Samba vfs using Linux kernel aio calls</refpurpose>
-</refnamediv>
-
-<refsynopsisdiv>
-       <cmdsynopsis>
-               <command>vfs objects = aio_linux</command>
-       </cmdsynopsis>
-</refsynopsisdiv>
-
-<refsect1>
-       <title>DESCRIPTION</title>
-
-       <para>This VFS module is part of the
-       <citerefentry><refentrytitle>samba</refentrytitle>
-       <manvolnum>7</manvolnum></citerefentry> suite.</para>
-
-       <para>The <command>aio_linux</command> VFS module enables asynchronous
-       I/O for Samba on Linux kernels that have the kernel AIO calls available
-       without using the Posix AIO interface. Posix AIO can suffer from severe
-       limitations.  For example, on some Linux versions the
-       real-time signals that it uses are broken under heavy load.
-       Other systems only allow AIO when special kernel modules are
-       loaded or only allow a certain system-wide amount of async
-       requests being scheduled. Systems based on glibc (most Linux
-       systems) only allow a single outstanding request per file
-       descriptor which essentially makes Posix AIO useless on systems
-       using the glibc implementation.</para>
-
-       <para>To work around all these limitations, the aio_linux module
-       was written. It uses the Linux kernel AIO interface instead of the
-       internal Posix AIO interface to allow read and write calls
-       to be processed asynchronously. A queue size of 128 events
-       is used by default. To change this limit set the "aio num events"
-       parameter below.
-       </para>
-
-       <para>
-       Note that the smb.conf parameters <command>aio read size</command>
-       and <command>aio write size</command> must also be set appropriately
-       for this module to be active.
-       </para>
-
-       <para>This module MUST be listed last in any module stack as
-       the Samba VFS pread/pwrite interface is not thread-safe. This
-       module makes direct pread and pwrite system calls and does
-       NOT call the Samba VFS pread and pwrite interfaces.</para>
-
-</refsect1>
-
-
-<refsect1>
-       <title>EXAMPLES</title>
-
-       <para>Straight forward use:</para>
-
-<programlisting>
-        <smbconfsection name="[cooldata]"/>
-       <smbconfoption name="path">/data/ice</smbconfoption>
-       <smbconfoption name="aio read size">1024</smbconfoption>
-       <smbconfoption name="aio write size">1024</smbconfoption>
-       <smbconfoption name="vfs objects">aio_linux</smbconfoption>
-</programlisting>
-
-</refsect1>
-
-<refsect1>
-       <title>OPTIONS</title>
-
-       <variablelist>
-
-               <varlistentry>
-               <term>aio_linux:aio num events = INTEGER</term>
-               <listitem>
-               <para>Set the maximum size of the event queue
-               that is used to limit outstanding IO requests.
-               </para>
-               <para>By default this is set to 128.</para>
-               </listitem>
-               </varlistentry>
-
-       </variablelist>
-</refsect1>
-<refsect1>
-       <title>VERSION</title>
-
-       <para>This man page is correct for version 4.0 of the Samba suite.
-       </para>
-</refsect1>
-
-<refsect1>
-       <title>AUTHOR</title>
-
-       <para>The original Samba software and related utilities
-       were created by Andrew Tridgell. Samba is now developed
-       by the Samba Team as an Open Source project similar
-       to the way the Linux kernel is developed.</para>
-
-</refsect1>
-
-</refentry>
index afba0b9e8416fc6209e7271b3b7eb460adb87536..0bd08d11cdb9e8e81671bef7e266d0cc2de8ef2f 100644 (file)
@@ -52,7 +52,6 @@ manpages='''
          manpages/vfs_acl_tdb.8
          manpages/vfs_acl_xattr.8
          manpages/vfs_aio_fork.8
-         manpages/vfs_aio_linux.8
          manpages/vfs_aio_pthread.8
          manpages/vfs_audit.8
         manpages/vfs_btrfs.8
diff --git a/source3/modules/vfs_aio_linux.c b/source3/modules/vfs_aio_linux.c
deleted file mode 100644 (file)
index e89c290..0000000
+++ /dev/null
@@ -1,346 +0,0 @@
-/*
- * Simulate Posix AIO using Linux kernel AIO.
- *
- * Copyright (C) Jeremy Allison 2012
- * Copyright (C) Volker Lendecke 2012
- *
- * This program is free software; you can redistribute it and/or modify
- * it under the terms of the GNU General Public License as published by
- * the Free Software Foundation; either version 3 of the License, or
- * (at your option) any later version.
- *
- * This program is distributed in the hope that it will be useful,
- * but WITHOUT ANY WARRANTY; without even the implied warranty of
- * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
- * GNU General Public License for more details.
- *
- * You should have received a copy of the GNU General Public License
- * along with this program; if not, write to the Free Software
- * Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
- */
-
-#include "includes.h"
-#include "system/filesys.h"
-#include "smbd/smbd.h"
-#include "smbd/globals.h"
-#include "lib/util/tevent_unix.h"
-#include "lib/util/sys_rw.h"
-#include <sys/eventfd.h>
-#include <libaio.h>
-#include "smbprofile.h"
-
-static int event_fd = -1;
-static io_context_t io_ctx;
-static struct tevent_fd *aio_read_event;
-static bool used;
-static unsigned num_busy;
-
-static void aio_linux_done(struct tevent_context *event_ctx,
-                          struct tevent_fd *event,
-                          uint16_t flags, void *private_data);
-
-/************************************************************************
- Housekeeping. Cleanup if no activity for 30 seconds.
-***********************************************************************/
-
-static void aio_linux_housekeeping(struct tevent_context *event_ctx,
-                                        struct tevent_timer *te,
-                                        struct timeval now,
-                                        void *private_data)
-{
-       /* Remove this timed event handler. */
-       TALLOC_FREE(te);
-
-       if ((num_busy != 0) || used) {
-               used = false;
-
-               /* Still busy. Look again in 30 seconds. */
-               (void)tevent_add_timer(event_ctx,
-                                       NULL,
-                                       timeval_current_ofs(30, 0),
-                                       aio_linux_housekeeping,
-                                       NULL);
-               return;
-       }
-
-       /* No activity for 30 seconds. Close out kernel resources. */
-       io_queue_release(io_ctx);
-       memset(&io_ctx, '\0', sizeof(io_ctx));
-
-       if (event_fd != -1) {
-               close(event_fd);
-               event_fd = -1;
-       }
-
-       TALLOC_FREE(aio_read_event);
-}
-
-/************************************************************************
- Ensure event fd and aio context are initialized.
-***********************************************************************/
-
-static bool init_aio_linux(struct vfs_handle_struct *handle)
-{
-       struct tevent_timer *te = NULL;
-
-       if (event_fd != -1) {
-               /* Already initialized. */
-               return true;
-       }
-
-       /* Schedule a shutdown event for 30 seconds from now. */
-       te = tevent_add_timer(handle->conn->sconn->ev_ctx,
-                               NULL,
-                               timeval_current_ofs(30, 0),
-                               aio_linux_housekeeping,
-                               NULL);
-
-       if (te == NULL) {
-               goto fail;
-       }
-
-       event_fd = eventfd(0, EFD_NONBLOCK | EFD_CLOEXEC);
-       if (event_fd == -1) {
-               goto fail;
-       }
-
-       aio_read_event = tevent_add_fd(server_event_context(),
-                               NULL,
-                               event_fd,
-                               TEVENT_FD_READ,
-                               aio_linux_done,
-                               NULL);
-       if (aio_read_event == NULL) {
-               goto fail;
-       }
-
-       if (io_queue_init(lp_aio_max_threads(), &io_ctx)) {
-               goto fail;
-       }
-
-       DEBUG(10,("init_aio_linux: initialized with up to %d events\n",
-                 (int)lp_aio_max_threads()));
-
-       return true;
-
-  fail:
-
-       DEBUG(10,("init_aio_linux: initialization failed\n"));
-
-       TALLOC_FREE(te);
-       TALLOC_FREE(aio_read_event);
-       if (event_fd != -1) {
-               close(event_fd);
-               event_fd = -1;
-       }
-       memset(&io_ctx, '\0', sizeof(io_ctx));
-       return false;
-}
-
-struct aio_linux_state {
-       struct iocb event_iocb;
-       ssize_t ret;
-       struct vfs_aio_state vfs_aio_state;
-       struct timespec start;
-};
-
-static struct tevent_req *aio_linux_pread_send(
-       struct vfs_handle_struct *handle, TALLOC_CTX *mem_ctx,
-       struct tevent_context *ev, struct files_struct *fsp,
-       void *data, size_t n, off_t offset)
-{
-       struct tevent_req *req;
-       struct aio_linux_state *state;
-       struct iocb *piocb;
-       int ret;
-
-       req = tevent_req_create(mem_ctx, &state, struct aio_linux_state);
-       if (req == NULL) {
-               return NULL;
-       }
-       if (!init_aio_linux(handle)) {
-               tevent_req_error(req, EIO);
-               return tevent_req_post(req, ev);
-       }
-
-       io_prep_pread(&state->event_iocb, fsp->fh->fd, data, n, offset);
-       io_set_eventfd(&state->event_iocb, event_fd);
-       state->event_iocb.data = req;
-
-       piocb = &state->event_iocb;
-
-       PROFILE_TIMESTAMP(&state->start);
-       ret = io_submit(io_ctx, 1, &piocb);
-       if (ret < 0) {
-               tevent_req_error(req, -ret);
-               return tevent_req_post(req, ev);
-       }
-       num_busy += 1;
-       used = true;
-       return req;
-}
-
-static struct tevent_req *aio_linux_pwrite_send(
-       struct vfs_handle_struct *handle, TALLOC_CTX *mem_ctx,
-       struct tevent_context *ev, struct files_struct *fsp,
-       const void *data, size_t n, off_t offset)
-{
-       struct tevent_req *req;
-       struct aio_linux_state *state;
-       struct iocb *piocb;
-       int ret;
-
-       req = tevent_req_create(mem_ctx, &state, struct aio_linux_state);
-       if (req == NULL) {
-               return NULL;
-       }
-       if (!init_aio_linux(handle)) {
-               tevent_req_error(req, EIO);
-               return tevent_req_post(req, ev);
-       }
-
-       io_prep_pwrite(&state->event_iocb, fsp->fh->fd, discard_const(data),
-                      n, offset);
-       io_set_eventfd(&state->event_iocb, event_fd);
-       state->event_iocb.data = req;
-
-       piocb = &state->event_iocb;
-
-       PROFILE_TIMESTAMP(&state->start);
-       ret = io_submit(io_ctx, 1, &piocb);
-       if (ret < 0) {
-               tevent_req_error(req, -ret);
-               return tevent_req_post(req, ev);
-       }
-       num_busy += 1;
-       used = true;
-       return req;
-}
-
-static struct tevent_req *aio_linux_fsync_send(
-       struct vfs_handle_struct *handle, TALLOC_CTX *mem_ctx,
-       struct tevent_context *ev, struct files_struct *fsp)
-{
-       struct tevent_req *req;
-       struct aio_linux_state *state;
-       struct iocb *piocb;
-       int ret;
-
-       req = tevent_req_create(mem_ctx, &state, struct aio_linux_state);
-       if (req == NULL) {
-               return NULL;
-       }
-       if (!init_aio_linux(handle)) {
-               tevent_req_error(req, EIO);
-               return tevent_req_post(req, ev);
-       }
-
-       io_prep_fsync(&state->event_iocb, fsp->fh->fd);
-       io_set_eventfd(&state->event_iocb, event_fd);
-       state->event_iocb.data = req;
-
-       piocb = &state->event_iocb;
-
-       PROFILE_TIMESTAMP(&state->start);
-       ret = io_submit(io_ctx, 1, &piocb);
-       if (ret < 0) {
-               tevent_req_error(req, -ret);
-               return tevent_req_post(req, ev);
-       }
-       num_busy += 1;
-       used = true;
-       return req;
-}
-
-static void aio_linux_done(struct tevent_context *event_ctx,
-                          struct tevent_fd *event,
-                          uint16_t flags, void *private_data)
-{
-       uint64_t num_events = 0;
-       struct timespec end;
-
-       DEBUG(10, ("aio_linux_done called with flags=%d\n",
-                  (int)flags));
-
-       PROFILE_TIMESTAMP(&end);
-
-       /* Read the number of events available. */
-       if (sys_read(event_fd, &num_events, sizeof(num_events)) !=
-                       sizeof(num_events)) {
-               smb_panic("aio_linux_handle_completion: invalid read");
-       }
-
-       while (num_events > 0) {
-               struct timespec ts = { 0, };
-               struct io_event finished;
-               struct tevent_req *req;
-               struct aio_linux_state *state;
-               int ret;
-
-               ret = io_getevents(io_ctx, 1, 1, &finished, &ts);
-               if (ret < 0) {
-                       DEBUG(1, ("aio_linux_done: io_getevents returned %s\n",
-                                 strerror(-ret)));
-                       return;
-               }
-               if (ret == 0) {
-                       DEBUG(10, ("aio_linux_done: io_getvents returned "
-                                  "0\n"));
-                       continue;
-               }
-
-               num_busy -= 1;
-
-               req = talloc_get_type_abort(finished.data,
-                                           struct tevent_req);
-               state = tevent_req_data(req, struct aio_linux_state);
-
-               if (finished.res < 0) {
-                       state->ret = -1;
-                       state->vfs_aio_state.error = -finished.res;
-               } else {
-                       state->ret = finished.res;
-               }
-               state->vfs_aio_state.duration = nsec_time_diff(&end, &state->start);
-               tevent_req_done(req);
-               num_events -= 1;
-       }
-}
-
-static ssize_t aio_linux_recv(struct tevent_req *req,
-                             struct vfs_aio_state *vfs_aio_state)
-{
-       struct aio_linux_state *state = tevent_req_data(
-               req, struct aio_linux_state);
-
-       if (tevent_req_is_unix_error(req, &vfs_aio_state->error)) {
-               return -1;
-       }
-       *vfs_aio_state = state->vfs_aio_state;
-       return state->ret;
-}
-
-static int aio_linux_int_recv(struct tevent_req *req,
-                             struct vfs_aio_state *vfs_aio_state)
-{
-       /*
-        * Use implicit conversion ssize_t->int
-        */
-       return aio_linux_recv(req, vfs_aio_state);
-}
-
-static struct vfs_fn_pointers vfs_aio_linux_fns = {
-       .pread_send_fn = aio_linux_pread_send,
-       .pread_recv_fn = aio_linux_recv,
-       .pwrite_send_fn = aio_linux_pwrite_send,
-       .pwrite_recv_fn = aio_linux_recv,
-       .fsync_send_fn = aio_linux_fsync_send,
-       .fsync_recv_fn = aio_linux_int_recv,
-};
-
-static_decl_vfs;
-NTSTATUS vfs_aio_linux_init(TALLOC_CTX *ctx)
-{
-       return smb_register_vfs(SMB_VFS_INTERFACE_VERSION,
-                               "aio_linux", &vfs_aio_linux_fns);
-}
index ab9be384189ad5e50eddd15e144939dfc12a666b..4e4fdb98a1d964171b31dc89a6fec090f858d485 100644 (file)
@@ -357,14 +357,6 @@ bld.SAMBA3_MODULE('vfs_aio_pthread',
                  internal_module=bld.SAMBA3_IS_STATIC_MODULE('vfs_aio_pthread'),
                  enabled=bld.SAMBA3_IS_ENABLED_MODULE('vfs_aio_pthread'))
 
-bld.SAMBA3_MODULE('vfs_aio_linux',
-                 subsystem='vfs',
-                 source='vfs_aio_linux.c',
-                 deps='samba-util aio',
-                 init_function='',
-                 internal_module=bld.SAMBA3_IS_STATIC_MODULE('vfs_aio_linux'),
-                 enabled=bld.SAMBA3_IS_ENABLED_MODULE('vfs_aio_linux'))
-
 bld.SAMBA3_MODULE('vfs_preopen',
                  subsystem='vfs',
                  source='vfs_preopen.c',
index e311f5f6efe7dc76bea4f5b92d2cbd3fb797be1a..d45a7d9ec606a2f638f09f3b448253a3ca72c1a0 100644 (file)
@@ -551,29 +551,6 @@ return acl_get_perm_np(permset_d, perm);
                 msg='Checking for openat',
                 headers='fcntl.h')
 
-    if host_os.rfind('linux') > -1:
-        conf.CHECK_FUNCS_IN('io_submit', 'aio')
-        conf.CHECK_CODE('''
-struct io_event ioev;
-struct iocb *ioc;
-io_context_t ctx;
-struct timespec ts;
-int fd;
-char *buf;
-fd = eventfd(0, EFD_NONBLOCK | EFD_CLOEXEC);
-io_queue_init(128,&ctx);
-io_prep_pwrite(ioc, 1, buf, 1, 0);
-io_prep_pread(ioc, 1, buf, 1, 0);
-io_set_eventfd(ioc, fd);
-io_set_callback(ioc, (io_callback_t)(0));
-io_submit(ctx, 1, &ioc);
-io_getevents(ctx, 1, 1, &ioev, &ts);
-''',
-            'HAVE_LINUX_KERNEL_AIO',
-            msg='Checking for linux kernel asynchronous io support',
-            headers='unistd.h stdlib.h sys/types.h fcntl.h sys/eventfd.h libaio.h',
-            lib='aio')
-
     conf.CHECK_CODE('''
 struct msghdr msg;
 union {
@@ -1717,9 +1694,6 @@ main() {
     if Options.options.with_pthreadpool:
         default_shared_modules.extend(TO_LIST('vfs_aio_pthread'))
 
-    if conf.CONFIG_SET('HAVE_LINUX_KERNEL_AIO'):
-        default_shared_modules.extend(TO_LIST('vfs_aio_linux'))
-
     if conf.CONFIG_SET('HAVE_LDAP'):
         default_static_modules.extend(TO_LIST('pdb_ldapsam idmap_ldap'))