From shanlei at asiainfo.com  Sat May  7 06:57:56 2022
From: shanlei at asiainfo.com (lei shan)
Date: Sat, 7 May 2022 14:57:56 +0800
Subject: [PATCH]: unlink legacy unix socket when unit starts
In-Reply-To: <CAFXm0sty1+kWz_UqaNeEmOd_9QvUzNhjihM77pKVySJ+7-Q9sQ@mail.gmail.com>
References: <b7b60585adb2be56b3f331ca5f5037cf@asiainfo.com>
 <fd3e06f2-8e70-eaad-5943-c67b21694768@over.ru> <95ff104c5c17783dee0898b4665be957@asiainfo.com>
 <CAE-G2q1p85UbbBMofzNz7Grzp+1xJqL8_PPB81t1rMJ56gp6tw@mail.gmail.com>
 <9a4bd8763f964196469c7ba64a8a44d3@asiainfo.com> <CAE-G2q2sOfJ7b24swH1nZOrcird6jUwc3uRhBybvJS2eX_rmHw@mail.gmail.com>
 <20220426214510.cuwucvv7hgx42m6n@xenon.mamontov.net> <CAFXm0svsy1DAKf-UNM_GVFbj5Ky3=O2ZmMnH_GnyBvAZBbs+wA@mail.gmail.com>
 <e5e150ba-9926-39e1-740b-fd041d7d0374@over.ru> <CAFXm0sv4xU9Aa6HWrjLW_TfiuA_J0357Zpo9Wrs2w0mg9df6Vg@mail.gmail.com>
 <CAFXm0sty1+kWz_UqaNeEmOd_9QvUzNhjihM77pKVySJ+7-Q9sQ@mail.gmail.com>
Message-ID: <CAFXm0svbq=jP9-BtAzK6gsdS0kuZKBF6ZAWdsbSkByeuDKbGzQ@mail.gmail.com>

Hello, after discussion, I change the patch below,  store the socket when
the main process creates it, then delete it in runtime_exit.

# HG changeset patch
# User stdanley <shanlei at asiainfo.com>
# Date 1651904970 -28800
#      Sat May 07 14:29:30 2022 +0800
# Node ID c6aa8691ca3c7f92b5645547c8a182f34503a758
# Parent  df887748fffb281302ef02aef1d13037355d749c
clean unit socket file before leave

diff -r df887748fffb -r c6aa8691ca3c src/nxt_main_process.c
--- a/src/nxt_main_process.c    Wed Dec 15 12:18:59 2021 +0100
+++ b/src/nxt_main_process.c    Sat May 07 14:29:30 2022 +0800
@@ -1070,6 +1070,7 @@

         type = NXT_PORT_MSG_RPC_READY_LAST | NXT_PORT_MSG_CLOSE_FD;

+        nxt_runtime_listen_socket_add(task->thread->runtime,sa);
     } else {
         size = ls.end - ls.start;

diff -r df887748fffb -r c6aa8691ca3c src/nxt_runtime.c
--- a/src/nxt_runtime.c Wed Dec 15 12:18:59 2021 +0100
+++ b/src/nxt_runtime.c Sat May 07 14:29:30 2022 +0800
# HG changeset patch
# User stdanley <shanlei at asiainfo.com>
# Date 1651904970 -28800
#      Sat May 07 14:29:30 2022 +0800
# Node ID c6aa8691ca3c7f92b5645547c8a182f34503a758
# Parent  df887748fffb281302ef02aef1d13037355d749c
clean unit socket file before leave

diff -r df887748fffb -r c6aa8691ca3c src/nxt_main_process.c
--- a/src/nxt_main_process.c    Wed Dec 15 12:18:59 2021 +0100
+++ b/src/nxt_main_process.c    Sat May 07 14:29:30 2022 +0800
@@ -1070,6 +1070,7 @@

         type = NXT_PORT_MSG_RPC_READY_LAST | NXT_PORT_MSG_CLOSE_FD;

+        nxt_runtime_listen_socket_add(task->thread->runtime,sa);
     } else {
         size = ls.end - ls.start;

diff -r df887748fffb -r c6aa8691ca3c src/nxt_runtime.c
--- a/src/nxt_runtime.c Wed Dec 15 12:18:59 2021 +0100
+++ b/src/nxt_runtime.c Sat May 07 14:29:30 2022 +0800
+
+            curr = rt->listen_sockets->elts;
+            ncurr = rt->listen_sockets->nelts;
+            for ( c=0;c<ncurr; c++) {
+                sa= curr[c].sockaddr;
+                if (sa->u.sockaddr.sa_family == AF_UNIX) {
+                    name = (nxt_file_name_t *) sa->u.sockaddr_un.sun_path;
+                    (void) nxt_file_delete(name);
+                }
+            }
         }
 #endif
     }

On Fri, Apr 29, 2022 at 4:37 PM lei shan <shanlei at asiainfo.com> wrote:

>
> ---------- Forwarded message ---------
> From: lei shan <shanlei at asiainfo.com>
> Date: Fri, 29 Apr 2022, 16:36
> Subject: Re: [PATCH]: unlink legacy unix socket when unit starts
> To: Александр Поволоцкий <tarkhil at over.ru>
>
>
> I am working on it, just as control socket did.
> The idea is when the main process create the pipe, it stores the socket in
> runtime' listening_conntions, then remove it in runtime_exit of the main
> process.
>
> The cleaning can't be done by router is :
> The Unix socket owned by root user
>
>
>
>
> On Fri, 29 Apr 2022, 14:09 Александр Поволоцкий, <tarkhil at over.ru> wrote:
>
>> No patch will remove pipe on kill -9. We need to implement the logic used
>> for control socket.
>> On 29.04.2022 8:34, lei shan wrote:
>>
>> Hello guys,
>>
>> The process may crash, or other may misconfig and cause the file
>> confliction.
>> But at least, unit should remove the  pipe file on exit,as it already
>> does to the control pipe.
>>
>> So a patch should remove the Unix pipe on stop , instead of on startup.
>>
>> Can we agree on this?
>>
>> Thanks & Best Regards
>>
>>
>> On Wed, 27 Apr 2022, 05:45 Oleg A. Mamontov, <oleg at mamontov.net> wrote:
>>
>>> On Wed, Apr 27, 2022 at 05:11:26AM +1000, Scott van Kalken wrote:
>>> >This will work well for cases where unit shuts down gracefully. If unit
>>> does
>>> >not shut down gracefully, then i think you will have the same startup
>>> problem.
>>> >
>>> >It’s a good point about the TCP socket.
>>>
>>> It seems the key point is the way to check on startup whether the unix
>>> socket is stale or currently in use and then to decide remove it and
>>> re-create or just fail the start, isn't it?
>>>
>>> >On Tue, 26 Apr 2022 at 11:50 am, <shanlei at asiainfo.com> wrote:
>>> >
>>> >    I have reconsided the situation:
>>> >    compare to TCP sockets, if a TCP bind fails, we know other programs
>>> are
>>> >    using the socket ,the ops team then fix the mis-configured port
>>> number,
>>> >    change unit or other programs.
>>> >    the unix socket should behave similarly.
>>> >    the diffrence is when a tcp server is done, the socket is recycled
>>> by
>>> >    system, while the unix socket file is left in disk, which block
>>> "unit"'s
>>> >    next boot.
>>> >    may the idea patch is to unlink the unix socket file on shutdown,
>>> >    instead of startup?
>>> >
>>> >
>>> >    On 2022-04-25 14:38，Scott van Kalken wrote：
>>> >    > What if it were a command line option to delete on start? This
>>> way you
>>> >    > preserve the existing behaviour while introducing the desired
>>> >    > behaviour. It’s up to the user / admin to decide if they want to
>>> >    > “force delete” as an example.
>>> >    >
>>> >    > On Mon, 25 Apr 2022 at 4:33 pm, <shanlei at asiainfo.com> wrote:
>>> >    >
>>> >    >> I know the concern, that if we misconfigure the socket file,
>>> other
>>> >    >> file
>>> >    >> may be deleted accidentally.
>>> >    >> But if the program wont delete the file, bind will fail , that
>>> means
>>> >    >> we
>>> >    >> have to delete the file manually before every start. that not a
>>> good
>>> >    >>
>>> >    >> option to maintain a system.
>>> >    >>
>>> >    >> 在 2022-04-24 18:05，Александр Поволоцкий
>>> >    >> wrote：
>>> >    >>> Unlink socket that can be in use? I think it's not a good idea.
>>> >    >>>
>>> >    >>> On 24.04.2022 12:25, shanlei at asiainfo.com wrote:
>>> >    >>>>
>>> >    >>>> hello,guys,
>>> >    >>>>
>>> >    >>>> We found if we add a unix socket listener, the file was left in
>>> >    >> disk
>>> >    >>>> when unit got shutdown.
>>> >    >>>> If the unit starts again and load configuration from its state
>>> >    >> folder,
>>> >    >>>> it would fail with error msg "file exists",
>>> >    >>>> so We think the proper process is to "delete the legacy unix
>>> >    >> socket
>>> >    >>>> before start", and we present following patch:
>>> >    >>>>
>>> >    >>>>
>>> >    >>>> # HG changeset patch
>>> >    >>>> # User stdanley <shanlei at asiainfo.com>
>>> >    >>>> # Date 1650792235 -28800
>>> >    >>>> #      Sun Apr 24 17:23:55 2022 +0800
>>> >    >>>> # Node ID 29738840c393f12c7d7aa8f3959ed1655b39cd4d
>>> >    >>>> # Parent  8e06a879600e83af11345d023c53c2ca445cf82c
>>> >    >>>> unlink legacy unix socket file before binding.
>>> >    >>>>
>>> >    >>>> diff -r 8e06a879600e -r 29738840c393 src/nxt_main_process.c
>>> >    >>>> --- a/src/nxt_main_process.c    Tue Apr 12 04:16:00 2022 +0100
>>> >    >>>> +++ b/src/nxt_main_process.c    Sun Apr 24 17:23:55 2022 +0800
>>> >    >>>> @@ -1142,6 +1142,15 @@
>>> >    >>>> }
>>> >    >>>>
>>> >    >>>> #endif
>>> >    >>>> +#if (NXT_HAVE_UNIX_DOMAIN)
>>> >    >>>> +
>>> >    >>>> +    if (sa->u.sockaddr.sa_family == AF_UNIX
>>> >    >>>> +        && sa->type == SOCK_STREAM
>>> >    >>>> +        && sa->u.sockaddr_un.sun_path[0] != '\0') {
>>> >    >>>> +
>>> >    >>>> +        (void) unlink(sa->u.sockaddr_un.sun_path);
>>> >    >>>> +}
>>> >    >>>> +#endif
>>> >    >>>>
>>> >    >>>> if (bind(s, &sa->u.sockaddr, sa->socklen) != 0) {
>>> >    >>>> err = nxt_errno;
>>> >    >>>> _______________________________________________
>>> >    >>>> unit mailing list -- unit at nginx.org
>>> >    >>>> To unsubscribe send an email to unit-leave at nginx.org
>>> >    >>> _______________________________________________
>>> >    >>> unit mailing list -- unit at nginx.org
>>> >    >>> To unsubscribe send an email to unit-leave at nginx.org
>>> >    >> _______________________________________________
>>> >    >> unit mailing list -- unit at nginx.org
>>> >    >> To unsubscribe send an email to unit-leave at nginx.org
>>> >    > _______________________________________________
>>> >    > unit mailing list -- unit at nginx.org
>>> >    > To unsubscribe send an email to unit-leave at nginx.org
>>> >
>>>
>>> >_______________________________________________
>>> >unit mailing list -- unit at nginx.org
>>> >To unsubscribe send an email to unit-leave at nginx.org
>>>
>>>
>>> --
>>> Cheers,
>>> Oleg A. Mamontov
>>>
>>> mailto: oleg at mamontov.net
>>>
>>> skype:  lonerr11
>>> cell:   +7 (903) 798-1352
>>> _______________________________________________
>>> unit mailing list -- unit at nginx.org
>>> To unsubscribe send an email to unit-leave at nginx.org
>>>
>>
>> _______________________________________________
>> unit mailing list -- unit at nginx.org
>> To unsubscribe send an email to unit-leave at nginx.org
>>
>>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mailman.nginx.org/pipermail/unit/attachments/20220507/87f19336/attachment.htm>