Hi,

On Wed, Feb 7, 2024 at 1:33 PM Jordan Rife jrife@google.com wrote:

On Wed, Feb 7, 2024 at 2:39 AM Salvatore Bonaccorso carnil@debian.org wrote:

...

Hi Valentin, hi all

[This is about a regression reported in Debian for 6.1.67]

On Tue, Feb 06, 2024 at 01:00:11PM +0100, Valentin Kleibel wrote:

...

Package: linux-image-amd64 Version: 6.1.76+1 Source: linux Source-Version: 6.1.76+1 Severity: important Control: notfound -1 6.6.15-2

Dear Maintainers,

We discovered a bug affecting dlm that prevents any tcp communications by dlm when booted with debian kernel 6.1.76-1.

Dlm startup works (corosync-cpgtool shows the dlm:controld group with all expected nodes) but as soon as we try to add a lockspace dmesg shows:

dlm: Using TCP for communications
dlm: cannot start dlm midcomms -97

It seems that commit "dlm: use kernel_connect() and kernel_bind()" (e9cdebbe) was merged to 6.1.

Checking the code it seems that the changed function dlm_tcp_listen_bind() fails with exit code 97 (EAFNOSUPPORT) It is called from

dlm/lockspace.c: threads_start() -> dlm_midcomms_start() dlm/midcomms.c: dlm_midcomms_start() -> dlm_lowcomms_start() dlm/lowcomms.c: dlm_lowcomms_start() -> dlm_listen_for_all() -> dlm_proto_ops->listen_bind() = dlm_tcp_listen_bind()

The error code is returned all the way to threads_start() where the error message is emmitted.

Booting with the unsigned kernel from testing (6.6.15-2), which also contains this commit, works without issues.

I'm not sure what additional changes are required to get this working or if rolling back this change is an option.

We'd be happy to test patches that might fix this issue.

Thanks for your report. So we have a 6.1.76 specific regression for the backport of e9cdebbe23f1 ("dlm: use kernel_connect() and kernel_bind()") .

Let's loop in the upstream regression list for tracking and people involved for the subsystem to see if the issue can be identified. As it is working for 6.6.15 which includes the commit backport as well it might be very well that a prerequisite is missing.

# annotate regression with 6.1.y specific commit #regzbot ^introduced e11dea8f503341507018b60906c4a9e7332f3663 #regzbot link: https://bugs.debian.org/1063338

Any ideas?

Regards, Salvatore

Just a quick look comparing dlm_tcp_listen_bind between the latest 6.1 and 6.6 stable branches, it looks like there is a mismatch here with the dlm_local_addr[0] parameter.

6.1

static int dlm_tcp_listen_bind(struct socket *sock) { int addr_len;

/* Bind to our port */ make_sockaddr(dlm_local_addr[0], dlm_config.ci_tcp_port, &addr_len); return kernel_bind(sock, (struct sockaddr *)&dlm_local_addr[0], addr_len); }

6.6

static int dlm_tcp_listen_bind(struct socket *sock) { int addr_len;

/* Bind to our port */ make_sockaddr(&dlm_local_addr[0], dlm_config.ci_tcp_port, &addr_len); return kernel_bind(sock, (struct sockaddr *)&dlm_local_addr[0], addr_len); }

6.6 contains commit c51c9cd8 (fs: dlm: don't put dlm_local_addrs on heap) which changed

static struct sockaddr_storage *dlm_local_addr[DLM_MAX_ADDR_COUNT];

to

static struct sockaddr_storage dlm_local_addr[DLM_MAX_ADDR_COUNT];

It looks like kernel_bind() in 6.1 needs to be modified to match.

makes sense. I tried to cherry-pick e9cdebbe23f1 ("dlm: use kernel_connect() and kernel_bind()") on v6.1.67 as I don't see it there. It failed and does not apply cleanly.

Are we talking here about a debian kernel specific backport? If so, maybe somebody missed to modify those parts you mentioned.

- Alex

Hi Jordan, hi all

Just a quick look comparing dlm_tcp_listen_bind between the latest 6.1 and 6.6 stable branches, it looks like there is a mismatch here with the dlm_local_addr[0] parameter.

6.1

static int dlm_tcp_listen_bind(struct socket *sock) { int addr_len;

/* Bind to our port */ make_sockaddr(dlm_local_addr[0], dlm_config.ci_tcp_port, &addr_len); return kernel_bind(sock, (struct sockaddr *)&dlm_local_addr[0], addr_len); }

6.6

static int dlm_tcp_listen_bind(struct socket *sock) { int addr_len;

/* Bind to our port */ make_sockaddr(&dlm_local_addr[0], dlm_config.ci_tcp_port, &addr_len); return kernel_bind(sock, (struct sockaddr *)&dlm_local_addr[0], addr_len); }

6.6 contains commit c51c9cd8 (fs: dlm: don't put dlm_local_addrs on heap) which changed

static struct sockaddr_storage *dlm_local_addr[DLM_MAX_ADDR_COUNT];

to

static struct sockaddr_storage dlm_local_addr[DLM_MAX_ADDR_COUNT];

It looks like kernel_bind() in 6.1 needs to be modified to match.

We tried to apply commit c51c9cd8 (fs: dlm: don't put dlm_local_addrs on heap) to the debian kernel 6.1.76 and came up with the attached patch. Besides the different offsets there is a slight change dlm_tcp_bind() where in 6.1.76 kernel_bind() is used instead of sock->ops->bind() in the original commit.

This patch solves the issue we experienced.

Thanks for your help, Valentin

Hi Valentin,

Would you be able to confirm that the attached patch fixes your issue as well?

-Jordan

On Thu, Feb 8, 2024 at 9:42 AM Jordan Rife jrife@google.com wrote:

On Thu, Feb 8, 2024 at 3:37 AM Valentin Kleibel valentin@vrvis.at wrote:

...

Hi Jordan, hi all

...

Just a quick look comparing dlm_tcp_listen_bind between the latest 6.1 and 6.6 stable branches, it looks like there is a mismatch here with the dlm_local_addr[0] parameter.

6.1

static int dlm_tcp_listen_bind(struct socket *sock) { int addr_len;

/* Bind to our port */ make_sockaddr(dlm_local_addr[0], dlm_config.ci_tcp_port, &addr_len); return kernel_bind(sock, (struct sockaddr *)&dlm_local_addr[0], addr_len); }

6.6

static int dlm_tcp_listen_bind(struct socket *sock) { int addr_len;

/* Bind to our port */ make_sockaddr(&dlm_local_addr[0], dlm_config.ci_tcp_port, &addr_len); return kernel_bind(sock, (struct sockaddr *)&dlm_local_addr[0], addr_len); }

6.6 contains commit c51c9cd8 (fs: dlm: don't put dlm_local_addrs on heap) which changed

static struct sockaddr_storage *dlm_local_addr[DLM_MAX_ADDR_COUNT];

to

static struct sockaddr_storage dlm_local_addr[DLM_MAX_ADDR_COUNT];

It looks like kernel_bind() in 6.1 needs to be modified to match.

We tried to apply commit c51c9cd8 (fs: dlm: don't put dlm_local_addrs on heap) to the debian kernel 6.1.76 and came up with the attached patch. Besides the different offsets there is a slight change dlm_tcp_bind() where in 6.1.76 kernel_bind() is used instead of sock->ops->bind() in the original commit.

This patch solves the issue we experienced.

Thanks for your help, Valentin

Good to hear that works for you! We should fix this in the 6.1 stable kernel as well.

IMO it may be less risky and simpler to fix the backport of my patch e9cdebbe23f1 ("dlm: use kernel_connect() and kernel_bind()") and just switch (struct sockaddr *)&dlm_local_addr[0] to (struct sockaddr *)dlm_local_addr[0] in the call to kernel_bind() rather than backporting c51c9cd8 (fs: dlm: don't put dlm_local_addrs on heap) to 6.1.

I will have some time soon to fix the 6.1 backport, but it may make sense just to revert in the meantime.

-Jordan

I sent this patch out to stable@vger.kernel.org. Everyone should be CCd. Thanks for your help in confirming the fix works.

-Jordan