On 24/04/2021 21:56 Joan Moreau wrote:

chroot= does not resolve the issue I have "chroot = login" in my conf

Thanks! The chroot was needed to get the core dump. Can you try if this does fix the crash? Aki

From 1df4e02cbff710ce8938480b07a5690e37f661f6 Mon Sep 17 00:00:00 2001 From: Timo Sirainen Date: Fri, 23 Apr 2021 16:43:36 +0300 Subject: [PATCH] login-common: Fix handling destroyed_clients linked list

The client needs to be removed from destroyed_clients linked list before it's added to client_fd_proxies linked list. Broken by 1c622cdbe08df2f642e28923c39894516143ae2a --- src/login-common/client-common.c | 11 +++++++---- 1 file changed, 7 insertions(+), 4 deletions(-) diff --git a/src/login-common/client-common.c b/src/login-common/client-common.c index bdb6e9c798..1d264d9f75 100644 --- a/src/login-common/client-common.c +++ b/src/login-common/client-common.c @@ -289,8 +289,9 @@ void client_disconnect(struct client *client, const char *reason, /* Login was successful. We may now be proxying the connection, so don't disconnect the client until client_unref(). */ if (client->iostream_fd_proxy != NULL) { + i_assert(!client->fd_proxying); client->fd_proxying = TRUE; - i_assert(client->prev == NULL && client->next == NULL); + DLLIST_REMOVE(&destroyed_clients, client); DLLIST_PREPEND(&client_fd_proxies, client); client_fd_proxies_count++; } @@ -307,8 +308,9 @@ void client_destroy(struct client *client, const char *reason) if (last_client == client) last_client = client->prev; - /* remove from clients linked list before it's added to - client_fd_proxies. */ + /* move to destroyed_clients linked list before it's potentially + added to client_fd_proxies. */ + i_assert(!client->fd_proxying); DLLIST_REMOVE(&clients, client); DLLIST_PREPEND(&destroyed_clients, client); @@ -409,13 +411,14 @@ bool client_unref(struct client **_client) DLLIST_REMOVE(&client_fd_proxies, client); i_assert(client_fd_proxies_count > 0); client_fd_proxies_count--; + } else { + DLLIST_REMOVE(&destroyed_clients, client); } i_stream_unref(&client->input); o_stream_unref(&client->output); i_close_fd(&client->fd); event_unref(&client->event); - DLLIST_REMOVE(&destroyed_clients, client); i_free(client->proxy_user); i_free(client->proxy_master_user); i_free(client->virtual_user); -- GitLab

This is because you are not compiling with libsystemd-dev installed. I guess we need to make some service template that use type simple when you don't use libsystemd. Aki

On 25/04/2021 22:53 Joan Moreau wrote:

Yes, it seems fixed with this patch :)

Another bug with git, is the "type=" in systemd is switched from "simple" to "notify". The later does not work and reverting to "simple" does work

On 2021-04-25 17:53, Aki Tuomi wrote:

...

...

On 24/04/2021 21:56 Joan Moreau wrote:

chroot= does not resolve the issue I have "chroot = login" in my conf

Thanks!

The chroot was needed to get the core dump.

Can you try if this does fix the crash?

Aki

From 1df4e02cbff710ce8938480b07a5690e37f661f6 Mon Sep 17 00:00:00 2001 From: Timo Sirainen Date: Fri, 23 Apr 2021 16:43:36 +0300 Subject: [PATCH] login-common: Fix handling destroyed_clients linked list

The client needs to be removed from destroyed_clients linked list before it's added to client_fd_proxies linked list.

Broken by 1c622cdbe08df2f642e28923c39894516143ae2a ---  src/login-common/client-common.c | 11 +++++++----  1 file changed, 7 insertions(+), 4 deletions(-)

diff --git a/src/login-common/client-common.c b/src/login-common/client-common.c index bdb6e9c798..1d264d9f75 100644 --- a/src/login-common/client-common.c +++ b/src/login-common/client-common.c @@ -289,8 +289,9 @@ void client_disconnect(struct client *client, const char *reason, /* Login was successful. We may now be proxying the connection, so don't disconnect the client until client_unref(). */ if (client->iostream_fd_proxy != NULL) { + i_assert(!client->fd_proxying); client->fd_proxying = TRUE; - i_assert(client->prev == NULL && client->next == NULL); + DLLIST_REMOVE(&destroyed_clients, client); DLLIST_PREPEND(&client_fd_proxies, client); client_fd_proxies_count++; } @@ -307,8 +308,9 @@ void client_destroy(struct client *client, const char *reason)   if (last_client == client) last_client = client->prev; - /* remove from clients linked list before it's added to - client_fd_proxies. */ + /* move to destroyed_clients linked list before it's potentially + added to client_fd_proxies. */ + i_assert(!client->fd_proxying); DLLIST_REMOVE(&clients, client); DLLIST_PREPEND(&destroyed_clients, client);   @@ -409,13 +411,14 @@ bool client_unref(struct client **_client) DLLIST_REMOVE(&client_fd_proxies, client); i_assert(client_fd_proxies_count > 0); client_fd_proxies_count--; + } else { + DLLIST_REMOVE(&destroyed_clients, client); } i_stream_unref(&client->input); o_stream_unref(&client->output); i_close_fd(&client->fd); event_unref(&client->event);   - DLLIST_REMOVE(&destroyed_clients, client); i_free(client->proxy_user); i_free(client->proxy_master_user); i_free(client->virtual_user);

The current autoconf code is bit buggy, but if you do indeed have libsystemd-dev installed it should do the right thing and will work with systemd even if you have Type=notify. This has been actually tested, so if it's not working, then something else is wrong. Did you remember to run ./autogen.sh after pulling from git to make sure you get new configure script? Aki

On 26/04/2021 10:11 Joan Moreau wrote:

Yes systemd is installed (and the "dev" files as well)

On 2021-04-26 06:23, Aki Tuomi wrote:

...

This is because you are not compiling with libsystemd-dev installed. I guess we need to make some service template that use type simple when you don't use libsystemd.

Aki

...

On 25/04/2021 22:53 Joan Moreau wrote:

Yes, it seems fixed with this patch :)

Another bug with git, is the "type=" in systemd is switched from "simple" to "notify". The later does not work and reverting to "simple" does work

On 2021-04-25 17:53, Aki Tuomi wrote:

...

...

On 24/04/2021 21:56 Joan Moreau wrote:

chroot= does not resolve the issue I have "chroot = login" in my conf

Thanks!

The chroot was needed to get the core dump.

Can you try if this does fix the crash?

Aki

From 1df4e02cbff710ce8938480b07a5690e37f661f6 Mon Sep 17 00:00:00 2001 From: Timo Sirainen Date: Fri, 23 Apr 2021 16:43:36 +0300 Subject: [PATCH] login-common: Fix handling destroyed_clients linked list

The client needs to be removed from destroyed_clients linked list before it's added to client_fd_proxies linked list.

Broken by 1c622cdbe08df2f642e28923c39894516143ae2a ---  src/login-common/client-common.c | 11 +++++++----  1 file changed, 7 insertions(+), 4 deletions(-)

diff --git a/src/login-common/client-common.c b/src/login-common/client-common.c index bdb6e9c798..1d264d9f75 100644 --- a/src/login-common/client-common.c +++ b/src/login-common/client-common.c @@ -289,8 +289,9 @@ void client_disconnect(struct client *client, const char *reason,  /* Login was successful. We may now be proxying the connection,  so don't disconnect the client until client_unref(). */  if (client->iostream_fd_proxy != NULL) { + i_assert(!client->fd_proxying);  client->fd_proxying = TRUE; - i_assert(client->prev == NULL && client->next == NULL); + DLLIST_REMOVE(&destroyed_clients, client);  DLLIST_PREPEND(&client_fd_proxies, client);  client_fd_proxies_count++;  } @@ -307,8 +308,9 @@ void client_destroy(struct client *client, const char *reason)    if (last_client == client)  last_client = client->prev; - /* remove from clients linked list before it's added to - client_fd_proxies. */ + /* move to destroyed_clients linked list before it's potentially + added to client_fd_proxies. */ + i_assert(!client->fd_proxying);  DLLIST_REMOVE(&clients, client);  DLLIST_PREPEND(&destroyed_clients, client);   @@ -409,13 +411,14 @@ bool client_unref(struct client **_client)  DLLIST_REMOVE(&client_fd_proxies, client);  i_assert(client_fd_proxies_count > 0);  client_fd_proxies_count--; + } else { + DLLIST_REMOVE(&destroyed_clients, client);  }  i_stream_unref(&client->input);  o_stream_unref(&client->output);  i_close_fd(&client->fd);  event_unref(&client->event);   - DLLIST_REMOVE(&destroyed_clients, client);  i_free(client->proxy_user);  i_free(client->proxy_master_user);  i_free(client->virtual_user);

I don't know then. It works for me and I just tried it again. The only reason it would fail would be that HAVE_LIBSYSTEMD is not defined, so it would not be using libsystemd for notify support. $ sudo systemctl status dovecot ● dovecot.service - Dovecot IMAP/POP3 email server Loaded: loaded (/lib/systemd/system/dovecot.service; disabled; vendor preset: enabled) Active: active (running) since Mon 2021-04-26 10:30:02 EEST; 2s ago Docs: man:dovecot(1) https://doc.dovecot.org/ Main PID: 30213 (dovecot) Status: "v2.4.devel (98a1cca054) running" Tasks: 4 (limit: 4701) Memory: 3.3M CGroup: /system.slice/dovecot.service ├─30213 /home/cmouse/dovecot/sbin/dovecot -F ├─30214 dovecot/anvil ├─30215 dovecot/log └─30216 dovecot/config You can tell from the "Status" line that it's using Type=notify. Aki

On 26/04/2021 10:29 Joan Moreau wrote:

Yes, I do run autogen.sh after every "git pull"

On 2021-04-26 08:21, Aki Tuomi wrote:

...

The current autoconf code is bit buggy, but if you do indeed have libsystemd-dev installed it should do the right thing and will work with systemd even if you have Type=notify.

This has been actually tested, so if it's not working, then something else is wrong.

Did you remember to run ./autogen.sh after pulling from git to make sure you get new configure script?

Aki

...

On 26/04/2021 10:11 Joan Moreau wrote:

Yes systemd is installed (and the "dev" files as well)

On 2021-04-26 06:23, Aki Tuomi wrote:

...

This is because you are not compiling with libsystemd-dev installed. I guess we need to make some service template that use type simple when you don't use libsystemd.

Aki

...

On 25/04/2021 22:53 Joan Moreau wrote:

Yes, it seems fixed with this patch :)

Another bug with git, is the "type=" in systemd is switched from "simple" to "notify". The later does not work and reverting to "simple" does work

On 2021-04-25 17:53, Aki Tuomi wrote:

...

> On 24/04/2021 21:56 Joan Moreau wrote: > > > chroot= does not resolve the issue > I have "chroot = login" in my conf >

Thanks!

The chroot was needed to get the core dump.

Can you try if this does fix the crash?

Aki

From 1df4e02cbff710ce8938480b07a5690e37f661f6 Mon Sep 17 00:00:00 2001 From: Timo Sirainen Date: Fri, 23 Apr 2021 16:43:36 +0300 Subject: [PATCH] login-common: Fix handling destroyed_clients linked list

The client needs to be removed from destroyed_clients linked list before it's added to client_fd_proxies linked list.

Broken by 1c622cdbe08df2f642e28923c39894516143ae2a ---  src/login-common/client-common.c | 11 +++++++----  1 file changed, 7 insertions(+), 4 deletions(-)

diff --git a/src/login-common/client-common.c b/src/login-common/client-common.c index bdb6e9c798..1d264d9f75 100644 --- a/src/login-common/client-common.c +++ b/src/login-common/client-common.c @@ -289,8 +289,9 @@ void client_disconnect(struct client *client, const char *reason,  /* Login was successful. We may now be proxying the connection,  so don't disconnect the client until client_unref(). */  if (client->iostream_fd_proxy != NULL) { + i_assert(!client->fd_proxying);  client->fd_proxying = TRUE; - i_assert(client->prev == NULL && client->next == NULL); + DLLIST_REMOVE(&destroyed_clients, client);  DLLIST_PREPEND(&client_fd_proxies, client);  client_fd_proxies_count++;  } @@ -307,8 +308,9 @@ void client_destroy(struct client *client, const char *reason)    if (last_client == client)  last_client = client->prev; - /* remove from clients linked list before it's added to - client_fd_proxies. */ + /* move to destroyed_clients linked list before it's potentially + added to client_fd_proxies. */ + i_assert(!client->fd_proxying);  DLLIST_REMOVE(&clients, client);  DLLIST_PREPEND(&destroyed_clients, client);   @@ -409,13 +411,14 @@ bool client_unref(struct client **_client)  DLLIST_REMOVE(&client_fd_proxies, client);  i_assert(client_fd_proxies_count > 0);  client_fd_proxies_count--; + } else { + DLLIST_REMOVE(&destroyed_clients, client);  }  i_stream_unref(&client->input);  o_stream_unref(&client->output);  i_close_fd(&client->fd);  event_unref(&client->event);   - DLLIST_REMOVE(&destroyed_clients, client);  i_free(client->proxy_user);  i_free(client->proxy_master_user);  i_free(client->virtual_user);

Looking at config.log, there is #define HAVE_LIBSYSTEMD 1 But "Type=notify" does not appear My systemd is version 248 On 2021-04-26 12:05, Joan Moreau wrote:

I have

# sudo systemctl status dovecot ● dovecot.service - Dovecot IMAP/POP3 email server Loaded: loaded (/usr/lib/systemd/system/dovecot.service; enabled; vendor preset: disabled) Active: active (running) since Sun 2021-04-25 20:13:25 UTC; 14h ago Docs: man:dovecot(1) https://doc.dovecot.org/ Main PID: 2559364 (dovecot) Tasks: 28 (limit: 76912) Memory: 1.0G CPU: 7min 18.342s CGroup: /system.slice/dovecot.service ├─2559364 /usr/sbin/dovecot -F ├─2559366 dovecot/imap-login ├─2559367 dovecot/anvil [11 connections] ├─2559368 dovecot/log

On 2021-04-26 08:32, Aki Tuomi wrote: I don't know then. It works for me and I just tried it again. The only reason it would fail would be that HAVE_LIBSYSTEMD is not defined, so it would not be using libsystemd for notify support.

$ sudo systemctl status dovecot ● dovecot.service - Dovecot IMAP/POP3 email server Loaded: loaded (/lib/systemd/system/dovecot.service; disabled; vendor preset: enabled) Active: active (running) since Mon 2021-04-26 10:30:02 EEST; 2s ago Docs: man:dovecot(1) https://doc.dovecot.org/ Main PID: 30213 (dovecot) Status: "v2.4.devel (98a1cca054) running" Tasks: 4 (limit: 4701) Memory: 3.3M CGroup: /system.slice/dovecot.service ├─30213 /home/cmouse/dovecot/sbin/dovecot -F ├─30214 dovecot/anvil ├─30215 dovecot/log └─30216 dovecot/config

You can tell from the "Status" line that it's using Type=notify.

Aki

On 26/04/2021 10:29 Joan Moreau wrote:

Yes, I do run autogen.sh after every "git pull"

On 2021-04-26 08:21, Aki Tuomi wrote: The current autoconf code is bit buggy, but if you do indeed have libsystemd-dev installed it should do the right thing and will work with systemd even if you have Type=notify.

This has been actually tested, so if it's not working, then something else is wrong.

Did you remember to run ./autogen.sh after pulling from git to make sure you get new configure script?

Aki

On 26/04/2021 10:11 Joan Moreau wrote:

Yes systemd is installed (and the "dev" files as well)

On 2021-04-26 06:23, Aki Tuomi wrote: This is because you are not compiling with libsystemd-dev installed. I guess we need to make some service template that use type simple when you don't use libsystemd.

Aki

On 25/04/2021 22:53 Joan Moreau wrote:

Yes, it seems fixed with this patch :)

Another bug with git, is the "type=" in systemd is switched from "simple" to "notify". The later does not work and reverting to "simple" does work

On 2021-04-25 17:53, Aki Tuomi wrote: On 24/04/2021 21:56 Joan Moreau wrote:

chroot= does not resolve the issue I have "chroot = login" in my conf

Thanks!

The chroot was needed to get the core dump.

Can you try if this does fix the crash?

Aki

From 1df4e02cbff710ce8938480b07a5690e37f661f6 Mon Sep 17 00:00:00 2001 From: Timo Sirainen Date: Fri, 23 Apr 2021 16:43:36 +0300 Subject: [PATCH] login-common: Fix handling destroyed_clients linked list

The client needs to be removed from destroyed_clients linked list before it's added to client_fd_proxies linked list.

Broken by 1c622cdbe08df2f642e28923c39894516143ae2a --- src/login-common/client-common.c | 11 +++++++---- 1 file changed, 7 insertions(+), 4 deletions(-)

diff --git a/src/login-common/client-common.c b/src/login-common/client-common.c index bdb6e9c798..1d264d9f75 100644 --- a/src/login-common/client-common.c +++ b/src/login-common/client-common.c @@ -289,8 +289,9 @@ void client_disconnect(struct client *client, const char *reason, /* Login was successful. We may now be proxying the connection, so don't disconnect the client until client_unref(). */ if (client->iostream_fd_proxy != NULL) { + i_assert(!client->fd_proxying); client->fd_proxying = TRUE; - i_assert(client->prev == NULL && client->next == NULL); + DLLIST_REMOVE(&destroyed_clients, client); DLLIST_PREPEND(&client_fd_proxies, client); client_fd_proxies_count++; } @@ -307,8 +308,9 @@ void client_destroy(struct client *client, const char *reason)

if (last_client == client) last_client = client->prev; - /* remove from clients linked list before it's added to - client_fd_proxies. */ + /* move to destroyed_clients linked list before it's potentially + added to client_fd_proxies. */ + i_assert(!client->fd_proxying); DLLIST_REMOVE(&clients, client); DLLIST_PREPEND(&destroyed_clients, client);

@@ -409,13 +411,14 @@ bool client_unref(struct client **_client) DLLIST_REMOVE(&client_fd_proxies, client); i_assert(client_fd_proxies_count > 0); client_fd_proxies_count--; + } else { + DLLIST_REMOVE(&destroyed_clients, client); } i_stream_unref(&client->input); o_stream_unref(&client->output); i_close_fd(&client->fd); event_unref(&client->event);

- DLLIST_REMOVE(&destroyed_clients, client); i_free(client->proxy_user); i_free(client->proxy_master_user); i_free(client->virtual_user);

Indeed, latest git works much better :) On 2021-04-27 05:58, Aki Tuomi wrote:

Can you try with latest git? We did some improvements on the systemd configure parts.

Aki

On 26/04/2021 23:32 Joan Moreau wrote:

Looking at config.log, there is #define HAVE_LIBSYSTEMD 1 But "Type=notify" does not appear My systemd is version 248

On 2021-04-26 12:05, Joan Moreau wrote: I have # sudo systemctl status dovecot ● dovecot.service - Dovecot IMAP/POP3 email server Loaded: loaded (/usr/lib/systemd/system/dovecot.service; enabled; vendor preset: disabled) Active: active (running) since Sun 2021-04-25 20:13:25 UTC; 14h ago Docs: man:dovecot(1) https://doc.dovecot.org/ Main PID: 2559364 (dovecot) Tasks: 28 (limit: 76912) Memory: 1.0G CPU: 7min 18.342s CGroup: /system.slice/dovecot.service ├─2559364 /usr/sbin/dovecot -F ├─2559366 dovecot/imap-login ├─2559367 dovecot/anvil [11 connections] ├─2559368 dovecot/log

On 2021-04-26 08:32, Aki Tuomi wrote: I don't know then. It works for me and I just tried it again. The only reason it would fail would be that HAVE_LIBSYSTEMD is not defined, so it would not be using libsystemd for notify support.

$ sudo systemctl status dovecot ● dovecot.service - Dovecot IMAP/POP3 email server Loaded: loaded (/lib/systemd/system/dovecot.service; disabled; vendor preset: enabled) Active: active (running) since Mon 2021-04-26 10:30:02 EEST; 2s ago Docs: man:dovecot(1) https://doc.dovecot.org/ Main PID: 30213 (dovecot) Status: "v2.4.devel (98a1cca054) running" Tasks: 4 (limit: 4701) Memory: 3.3M CGroup: /system.slice/dovecot.service ├─30213 /home/cmouse/dovecot/sbin/dovecot -F ├─30214 dovecot/anvil ├─30215 dovecot/log └─30216 dovecot/config

You can tell from the "Status" line that it's using Type=notify.

Aki

On 26/04/2021 10:29 Joan Moreau wrote:

Yes, I do run autogen.sh after every "git pull"

On 2021-04-26 08:21, Aki Tuomi wrote: The current autoconf code is bit buggy, but if you do indeed have libsystemd-dev installed it should do the right thing and will work with systemd even if you have Type=notify.

This has been actually tested, so if it's not working, then something else is wrong.

Did you remember to run ./autogen.sh after pulling from git to make sure you get new configure script?

Aki

On 26/04/2021 10:11 Joan Moreau wrote:

Yes systemd is installed (and the "dev" files as well)

On 2021-04-26 06:23, Aki Tuomi wrote: This is because you are not compiling with libsystemd-dev installed. I guess we need to make some service template that use type simple when you don't use libsystemd.

Aki

On 25/04/2021 22:53 Joan Moreau wrote:

Yes, it seems fixed with this patch :)

Another bug with git, is the "type=" in systemd is switched from "simple" to "notify". The later does not work and reverting to "simple" does work

On 2021-04-25 17:53, Aki Tuomi wrote: On 24/04/2021 21:56 Joan Moreau wrote:

chroot= does not resolve the issue I have "chroot = login" in my conf

Thanks!

The chroot was needed to get the core dump.

Can you try if this does fix the crash?

Aki

From 1df4e02cbff710ce8938480b07a5690e37f661f6 Mon Sep 17 00:00:00 2001 From: Timo Sirainen Date: Fri, 23 Apr 2021 16:43:36 +0300 Subject: [PATCH] login-common: Fix handling destroyed_clients linked list

The client needs to be removed from destroyed_clients linked list before it's added to client_fd_proxies linked list.

Broken by 1c622cdbe08df2f642e28923c39894516143ae2a --- src/login-common/client-common.c | 11 +++++++---- 1 file changed, 7 insertions(+), 4 deletions(-)

diff --git a/src/login-common/client-common.c b/src/login-common/client-common.c index bdb6e9c798..1d264d9f75 100644 --- a/src/login-common/client-common.c +++ b/src/login-common/client-common.c @@ -289,8 +289,9 @@ void client_disconnect(struct client *client, const char *reason, /* Login was successful. We may now be proxying the connection, so don't disconnect the client until client_unref(). */ if (client->iostream_fd_proxy != NULL) { + i_assert(!client->fd_proxying); client->fd_proxying = TRUE; - i_assert(client->prev == NULL && client->next == NULL); + DLLIST_REMOVE(&destroyed_clients, client); DLLIST_PREPEND(&client_fd_proxies, client); client_fd_proxies_count++; } @@ -307,8 +308,9 @@ void client_destroy(struct client *client, const char *reason)

if (last_client == client) last_client = client->prev; - /* remove from clients linked list before it's added to - client_fd_proxies. */ + /* move to destroyed_clients linked list before it's potentially + added to client_fd_proxies. */ + i_assert(!client->fd_proxying); DLLIST_REMOVE(&clients, client); DLLIST_PREPEND(&destroyed_clients, client);

@@ -409,13 +411,14 @@ bool client_unref(struct client **_client) DLLIST_REMOVE(&client_fd_proxies, client); i_assert(client_fd_proxies_count > 0); client_fd_proxies_count--; + } else { + DLLIST_REMOVE(&destroyed_clients, client); } i_stream_unref(&client->input); o_stream_unref(&client->output); i_close_fd(&client->fd); event_unref(&client->event);

- DLLIST_REMOVE(&destroyed_clients, client); i_free(client->proxy_user); i_free(client->proxy_master_user); i_free(client->virtual_user);

Can you provide any details on this instability? Aki On April 27, 2021 7:58:01 PM UTC, Joan Moreau wrote:

Ok, a third regression is that it becomes highly unstable with the patch you sent

I had to get back to 2.3.14

On 2021-04-27 17:07, Joan Moreau wrote:

...

Indeed, latest git works much better :)

On 2021-04-27 05:58, Aki Tuomi wrote: Can you try with latest git? We did some improvements on the systemd configure parts.

Aki

On 26/04/2021 23:32 Joan Moreau wrote:

Looking at config.log, there is #define HAVE_LIBSYSTEMD 1 But "Type=notify" does not appear My systemd is version 248

On 2021-04-26 12:05, Joan Moreau wrote: I have # sudo systemctl status dovecot ● dovecot.service - Dovecot IMAP/POP3 email server Loaded: loaded (/usr/lib/systemd/system/dovecot.service; enabled; vendor preset: disabled) Active: active (running) since Sun 2021-04-25 20:13:25 UTC; 14h ago Docs: man:dovecot(1) https://doc.dovecot.org/ Main PID: 2559364 (dovecot) Tasks: 28 (limit: 76912) Memory: 1.0G CPU: 7min 18.342s CGroup: /system.slice/dovecot.service ├─2559364 /usr/sbin/dovecot -F ├─2559366 dovecot/imap-login ├─2559367 dovecot/anvil [11 connections] ├─2559368 dovecot/log

On 2021-04-26 08:32, Aki Tuomi wrote: I don't know then. It works for

...

me and I just tried it again. The only reason it would fail would be that HAVE_LIBSYSTEMD is not defined, so it would not be using libsystemd for notify support.

$ sudo systemctl status dovecot ● dovecot.service - Dovecot IMAP/POP3 email server Loaded: loaded (/lib/systemd/system/dovecot.service; disabled; vendor

...

preset: enabled) Active: active (running) since Mon 2021-04-26 10:30:02 EEST; 2s ago Docs: man:dovecot(1) https://doc.dovecot.org/ Main PID: 30213 (dovecot) Status: "v2.4.devel (98a1cca054) running" Tasks: 4 (limit: 4701) Memory: 3.3M CGroup: /system.slice/dovecot.service ├─30213 /home/cmouse/dovecot/sbin/dovecot -F ├─30214 dovecot/anvil ├─30215 dovecot/log └─30216 dovecot/config

You can tell from the "Status" line that it's using Type=notify.

Aki

On 26/04/2021 10:29 Joan Moreau wrote:

Yes, I do run autogen.sh after every "git pull"

On 2021-04-26 08:21, Aki Tuomi wrote: The current autoconf code is bit buggy, but if you do indeed have libsystemd-dev installed it should do the right thing and will work with systemd even if you have Type=notify.

This has been actually tested, so if it's not working, then something

...

else is wrong.

Did you remember to run ./autogen.sh after pulling from git to make sure you get new configure script?

Aki

On 26/04/2021 10:11 Joan Moreau wrote:

Yes systemd is installed (and the "dev" files as well)

On 2021-04-26 06:23, Aki Tuomi wrote: This is because you are not compiling with libsystemd-dev installed. I guess we need to make some

...

service template that use type simple when you don't use libsystemd.

Aki

On 25/04/2021 22:53 Joan Moreau wrote:

Yes, it seems fixed with this patch :)

Another bug with git, is the "type=" in systemd is switched from "simple" to "notify". The later does not work and reverting to "simple" does work

On 2021-04-25 17:53, Aki Tuomi wrote: On 24/04/2021 21:56 Joan Moreau

...

wrote:

chroot= does not resolve the issue I have "chroot = login" in my conf

Thanks!

The chroot was needed to get the core dump.

Can you try if this does fix the crash?

Aki

From 1df4e02cbff710ce8938480b07a5690e37f661f6 Mon Sep 17 00:00:00 2001 From: Timo Sirainen Date: Fri, 23 Apr 2021 16:43:36 +0300 Subject: [PATCH] login-common: Fix handling destroyed_clients linked list

The client needs to be removed from destroyed_clients linked list before it's added to client_fd_proxies linked list.

Broken by 1c622cdbe08df2f642e28923c39894516143ae2a --- src/login-common/client-common.c | 11 +++++++---- 1 file changed, 7 insertions(+), 4 deletions(-)

diff --git a/src/login-common/client-common.c b/src/login-common/client-common.c index bdb6e9c798..1d264d9f75 100644 --- a/src/login-common/client-common.c +++ b/src/login-common/client-common.c @@ -289,8 +289,9 @@ void client_disconnect(struct client *client, const char *reason, /* Login was successful. We may now be proxying the connection, so don't disconnect the client until client_unref(). */ if (client->iostream_fd_proxy != NULL) { + i_assert(!client->fd_proxying); client->fd_proxying = TRUE; - i_assert(client->prev == NULL && client->next == NULL); + DLLIST_REMOVE(&destroyed_clients, client); DLLIST_PREPEND(&client_fd_proxies, client); client_fd_proxies_count++; } @@ -307,8 +308,9 @@ void client_destroy(struct client *client, const char *reason)

if (last_client == client) last_client = client->prev; - /* remove from clients linked list before it's added to - client_fd_proxies. */ + /* move to destroyed_clients linked list before it's potentially + added to client_fd_proxies. */ + i_assert(!client->fd_proxying); DLLIST_REMOVE(&clients, client); DLLIST_PREPEND(&destroyed_clients, client);

@@ -409,13 +411,14 @@ bool client_unref(struct client **_client) DLLIST_REMOVE(&client_fd_proxies, client); i_assert(client_fd_proxies_count > 0); client_fd_proxies_count--; + } else { + DLLIST_REMOVE(&destroyed_clients, client); } i_stream_unref(&client->input); o_stream_unref(&client->output); i_close_fd(&client->fd); event_unref(&client->event);

- DLLIST_REMOVE(&destroyed_clients, client); i_free(client->proxy_user); i_free(client->proxy_master_user); i_free(client->virtual_user);

Did you see if the problem was that the imap-login process was using 100% CPU, or was the issue something else? I can't find a bug with the patch itself. But attached is another patch that adds some more asserts to make sure the linked lists are being used correctly, so if there is some bug it should now assert-crash instead of doing something else weird like go to infinite loop. But maybe the high CPU usage was something unrelated to this patch?

On 28. Apr 2021, at 21.57, Joan Moreau wrote:

Not much details

Git version (including the patch you sent) raised CPU load very very high.

Can't play too much on my production server.

Let me know if I can help

On 2021-04-28 06:12, Aki Tuomi wrote:

...

Can you provide any details on this instability?

Aki

On April 27, 2021 7:58:01 PM UTC, Joan Moreau wrote: Ok, a third regression is that it becomes highly unstable with the patch you sent

I had to get back to 2.3.14

On 2021-04-27 17:07, Joan Moreau wrote:

Indeed, latest git works much better :)

On 2021-04-27 05:58, Aki Tuomi wrote:

Can you try with latest git? We did some improvements on the systemd configure parts.

Aki

On 26/04/2021 23:32 Joan Moreau mailto:jom@grosjo.net> wrote:

Looking at config.log, there is #define HAVE_LIBSYSTEMD 1 But "Type=notify" does not appear My systemd is version 248

On 2021-04-26 12:05, Joan Moreau wrote: I have # sudo systemctl status dovecot ● dovecot.service - Dovecot IMAP/POP3 email server Loaded: loaded (/usr/lib/systemd/system/dovecot.service; enabled; vendor preset: disabled) Active: active (running) since Sun 2021-04-25 20:13:25 UTC; 14h ago Docs: man:dovecot(1) https://doc.dovecot.org/ https://doc.dovecot.org/ Main PID: 2559364 (dovecot) Tasks: 28 (limit: 76912) Memory: 1.0G CPU: 7min 18.342s CGroup: /system.slice/dovecot.service ├─2559364 /usr/sbin/dovecot -F ├─2559366 dovecot/imap-login ├─2559367 dovecot/anvil [11 connections] ├─2559368 dovecot/log

On 2021-04-26 08:32, Aki Tuomi wrote: I don't know then. It works for me and I just tried it again. The only reason it would fail would be that HAVE_LIBSYSTEMD is not defined, so it would not be using libsystemd for notify support.

$ sudo systemctl status dovecot ● dovecot.service - Dovecot IMAP/POP3 email server Loaded: loaded (/lib/systemd/system/dovecot.service; disabled; vendor preset: enabled) Active: active (running) since Mon 2021-04-26 10:30:02 EEST; 2s ago Docs: man:dovecot(1) https://doc.dovecot.org/ https://doc.dovecot.org/ Main PID: 30213 (dovecot) Status: "v2.4.devel (98a1cca054) running" Tasks: 4 (limit: 4701) Memory: 3.3M CGroup: /system.slice/dovecot.service ├─30213 /home/cmouse/dovecot/sbin/dovecot -F ├─30214 dovecot/anvil ├─30215 dovecot/log └─30216 dovecot/config

You can tell from the "Status" line that it's using Type=notify.

Aki

On 26/04/2021 10:29 Joan Moreau mailto:jom@grosjo.net> wrote:

Yes, I do run autogen.sh after every "git pull"

On 2021-04-26 08:21, Aki Tuomi wrote: The current autoconf code is bit buggy, but if you do indeed have libsystemd-dev installed it should do the right thing and will work with systemd even if you have Type=notify.

This has been actually tested, so if it's not working, then something else is wrong.

Did you remember to run ./autogen.sh after pulling from git to make sure you get new configure script?

Aki

On 26/04/2021 10:11 Joan Moreau mailto:jom@grosjo.net> wrote:

Yes systemd is installed (and the "dev" files as well)

On 2021-04-26 06:23, Aki Tuomi wrote: This is because you are not compiling with libsystemd-dev installed. I guess we need to make some service template that use type simple when you don't use libsystemd.

Aki

On 25/04/2021 22:53 Joan Moreau mailto:jom@grosjo.net> wrote:

Yes, it seems fixed with this patch :)

Another bug with git, is the "type=" in systemd is switched from "simple" to "notify". The later does not work and reverting to "simple" does work

On 2021-04-25 17:53, Aki Tuomi wrote: On 24/04/2021 21:56 Joan Moreau mailto:jom@grosjo.net> wrote:

chroot= does not resolve the issue I have "chroot = login" in my conf

Thanks!

The chroot was needed to get the core dump.

Can you try if this does fix the crash?

Aki

From 1df4e02cbff710ce8938480b07a5690e37f661f6 Mon Sep 17 00:00:00 2001 From: Timo Sirainen mailto:timo.sirainen@open-xchange.com> Date: Fri, 23 Apr 2021 16:43:36 +0300 Subject: [PATCH] login-common: Fix handling destroyed_clients linked list

The client needs to be removed from destroyed_clients linked list before it's added to client_fd_proxies linked list.

Broken by 1c622cdbe08df2f642e28923c39894516143ae2a --- src/login-common/client-common.c | 11 +++++++---- 1 file changed, 7 insertions(+), 4 deletions(-)

diff --git a/src/login-common/client-common.c b/src/login-common/client-common.c index bdb6e9c798..1d264d9f75 100644 --- a/src/login-common/client-common.c +++ b/src/login-common/client-common.c @@ -289,8 +289,9 @@ void client_disconnect(struct client *client, const char *reason, /* Login was successful. We may now be proxying the connection, so don't disconnect the client until client_unref(). */ if (client->iostream_fd_proxy != NULL) { + i_assert(!client->fd_proxying); client->fd_proxying = TRUE; - i_assert(client->prev == NULL && client->next == NULL); + DLLIST_REMOVE(&destroyed_clients, client); DLLIST_PREPEND(&client_fd_proxies, client); client_fd_proxies_count++; } @@ -307,8 +308,9 @@ void client_destroy(struct client *client, const char *reason)

if (last_client == client) last_client = client->prev; - /* remove from clients linked list before it's added to - client_fd_proxies. */ + /* move to destroyed_clients linked list before it's potentially + added to client_fd_proxies. */ + i_assert(!client->fd_proxying); DLLIST_REMOVE(&clients, client); DLLIST_PREPEND(&destroyed_clients, client);

@@ -409,13 +411,14 @@ bool client_unref(struct client **_client) DLLIST_REMOVE(&client_fd_proxies, client); i_assert(client_fd_proxies_count > 0); client_fd_proxies_count--; + } else { + DLLIST_REMOVE(&destroyed_clients, client); } i_stream_unref(&client->input); o_stream_unref(&client->output); i_close_fd(&client->fd); event_unref(&client->event);

- DLLIST_REMOVE(&destroyed_clients, client); i_free(client->proxy_user); i_free(client->proxy_master_user); i_free(client->virtual_user);