Horton
Menu

Thread

  1. [PATCH] check kernel version for io_method

    pierre.forstmann@gmail.com — 2026-01-13T22:05:29Z

    Hello,


Please find a patch proposal for bug BUG #19369: Not documented that 
io_uring on kernel versions between 5.1 and below 5.6 does not work.

This is not  a document patch but code patch using a GUC hook to :

- check liburing setup

- check kernel version

to allow/disallow io_method=io_uring


I have made successfull tests on:

- Alma Linux 9.7/ kernel 5.14 without liburing setup

- Debian 12 / kernel 6.1 with liburing-dev


NB:

This is my first patch and I apologize if it is not fully compliant with 
patching rules.


Thanks.

  2. Re: [PATCH] check kernel version for io_method

    Andreas Karlsson <andreas@proxel.se> — 2026-01-14T01:07:43Z

    On 1/13/26 11:05 PM, Pierre wrote:
> Please find a patch proposal for bug BUG #19369: Not documented that 
> io_uring on kernel versions between 5.1 and below 5.6 does not work.
Wouldn't it make more sense to use io_uring_get_probe_ring() and 
io_uring_opcode_supported() to actually check for the required opcode? 
Then if someone uses a Kernel with the required features backported 
everything will still work.

Andreas

  3. Re: [PATCH] check kernel version for io_method

    Steven Niu <niushiji@gmail.com> — 2026-01-14T02:31:24Z

    From: Andreas Karlsson <andreas@proxel.se>
Sent: Wednesday, January 14, 2026 09:07
To: Pierre <pierre.forstmann@gmail.com>; pgsql-hackers@lists.postgresql.org <pgsql-hackers@lists.postgresql.org>
Subject: Re: [PATCH] check kernel version for io_method


On 1/13/26 11:05 PM, Pierre wrote:
> Please find a patch proposal for bug BUG #19369: Not documented that 
> io_uring on kernel versions between 5.1 and below 5.6 does not work.
Wouldn't it make more sense to use io_uring_get_probe_ring() and 
io_uring_opcode_supported() to actually check for the required opcode? 
Then if someone uses a Kernel with the required features backported 
everything will still work.

Andreas

+1 
It would be more reliable to check the availability of io_uring feature than to check OS version. 

BTW, the error message in the following code block is wrong as it locates in branch that IO_URING is not enabled. 

#else
+	if (*newval == IOMETHOD_WORKER || *newval == IOMETHOD_SYNC)
+	{
+		/*
+		 * OK
+		 */
+	} 
+	else 
+	{
+		GUC_check_errdetail("io_uring requires Linux kernel 5.6 or later.");
+		return false;
+	}



Thanks,
Steven

  4. Re: [PATCH] check kernel version for io_method

    Andreas Karlsson <andreas@proxel.se> — 2026-01-14T03:09:35Z

    On 1/14/26 3:31 AM, Steven Niu wrote:
> BTW, the error message in the following code block is wrong as it locates in branch that IO_URING is not enabled.
> 
> #else
> +	if (*newval == IOMETHOD_WORKER || *newval == IOMETHOD_SYNC)
> +	{
> +		/*
> +		 * OK
> +		 */
> +	}
> +	else
> +	{
> +		GUC_check_errdetail("io_uring requires Linux kernel 5.6 or later.");
> +		return false;
> +	}

That code is unreachable so it should just be deleted. Look at the enum.

/* Enum for io_method GUC. */
typedef enum IoMethod
{
	IOMETHOD_SYNC = 0,
	IOMETHOD_WORKER,
#ifdef IOMETHOD_IO_URING_ENABLED
	IOMETHOD_IO_URING,
#endif
}			IoMethod;

Andreas

  5. Re: [PATCH] check kernel version for io_method

    Steven Niu <niushiji@gmail.com> — 2026-01-14T07:22:35Z

    From: Andreas Karlsson <andreas@proxel.se>
Sent: Wednesday, January 14, 2026 11:09
To: Steven Niu <niushiji@gmail.com>; Pierre <pierre.forstmann@gmail.com>; pgsql-hackers@lists.postgresql.org <pgsql-hackers@lists.postgresql.org>
Subject: Re: [PATCH] check kernel version for io_method


On 1/14/26 3:31 AM, Steven Niu wrote:
> BTW, the error message in the following code block is wrong as it locates in branch that IO_URING is not enabled.
> 
> #else
> +     if (*newval == IOMETHOD_WORKER || *newval == IOMETHOD_SYNC)
> +     {
> +             /*
> +              * OK
> +              */
> +     }
> +     else
> +     {
> +             GUC_check_errdetail("io_uring requires Linux kernel 5.6 or later.");
> +             return false;
> +     }

That code is unreachable so it should just be deleted. Look at the enum.

/* Enum for io_method GUC. */
typedef enum IoMethod
{
        IOMETHOD_SYNC = 0,
        IOMETHOD_WORKER,
#ifdef IOMETHOD_IO_URING_ENABLED
        IOMETHOD_IO_URING,
#endif
}                       IoMethod;

Andreas



________________________________________

Hi, Andreas, 

I agree that you are correct for now. But in future, the IoMethod may have new member. Then we have to process it. 

How about change to like this:
 + #else
 +     if (*newval == IOMETHOD_IO_URING)
 +     {
 +             GUC_check_errdetail("io_uring not enabled.");
 +             return false;
 +     }

Thanks,
Steven

  6. Re: [PATCH] check kernel version for io_method

    Andreas Karlsson <andreas@proxel.se> — 2026-01-14T07:27:07Z

    On 1/14/26 8:22 AM, Steven Niu wrote:
> I agree that you are correct for now. But in future, the IoMethod may have new member. Then we have to process it.
> 
> How about change to like this:
>   + #else
>   +     if (*newval == IOMETHOD_IO_URING)
>   +     {
>   +             GUC_check_errdetail("io_uring not enabled.");
>   +             return false;
>   +     }

No, that would be incorrect. It would not even compile. 
IOMETHOD_IO_URING is not even defined if we do not compile with liburing 
enabled. You can try it out yourself by compiling PostgreSQL without 
io_uring support and then try to enable it. You will get an error and 
that code will never be reached.

Andreas

  7. Re: [PATCH] check kernel version for io_method

    Steven Niu <niushiji@gmail.com> — 2026-01-14T07:31:09Z

    From: Andreas Karlsson <andreas@proxel.se>
Sent: Wednesday, January 14, 2026 15:27
To: Steven Niu <niushiji@gmail.com>; Pierre <pierre.forstmann@gmail.com>; pgsql-hackers@lists.postgresql.org <pgsql-hackers@lists.postgresql.org>
Subject: Re: [PATCH] check kernel version for io_method


On 1/14/26 8:22 AM, Steven Niu wrote:
> I agree that you are correct for now. But in future, the IoMethod may have new member. Then we have to process it.
> 
> How about change to like this:
>   + #else
>   +     if (*newval == IOMETHOD_IO_URING)
>   +     {
>   +             GUC_check_errdetail("io_uring not enabled.");
>   +             return false;
>   +     }

No, that would be incorrect. It would not even compile. 
IOMETHOD_IO_URING is not even defined if we do not compile with liburing 
enabled. You can try it out yourself by compiling PostgreSQL without 
io_uring support and then try to enable it. You will get an error and 
that code will never be reached.

Andreas




________________________________________  


Yes, you are definitely correct. 
I'd like to withdwar my proposal.

  8. Re: [PATCH] check kernel version for io_method

    pierre.forstmann@gmail.com — 2026-01-14T10:18:00Z

    Le 14/01/2026 à 08:31, Steven Niu a écrit :
> From: Andreas Karlsson <andreas@proxel.se>
> Sent: Wednesday, January 14, 2026 15:27
> To: Steven Niu <niushiji@gmail.com>; Pierre <pierre.forstmann@gmail.com>; pgsql-hackers@lists.postgresql.org <pgsql-hackers@lists.postgresql.org>
> Subject: Re: [PATCH] check kernel version for io_method
>
>
> On 1/14/26 8:22 AM, Steven Niu wrote:
>> I agree that you are correct for now. But in future, the IoMethod may have new member. Then we have to process it.
>>
>> How about change to like this:
>>     + #else
>>     +     if (*newval == IOMETHOD_IO_URING)
>>     +     {
>>     +             GUC_check_errdetail("io_uring not enabled.");
>>     +             return false;
>>     +     }
> No, that would be incorrect. It would not even compile.
> IOMETHOD_IO_URING is not even defined if we do not compile with liburing
> enabled. You can try it out yourself by compiling PostgreSQL without
> io_uring support and then try to enable it. You will get an error and
> that code will never be reached.
>
> Andreas
>
>
>
>
> ________________________________________
>
>
> Yes, you are definitely correct.
> I'd like to withdwar my proposal.

Hello,

I have modified message in case io_uring is not enabled when building 
PG. But currently I cannot test it, GUC parameter is rejected in earlier 
steps:

pg_ctl start
waiting for server to start....2026-01-14 09:35:37.046 GMT [16336] LOG:  
invalid value for parameter "io_method": "io_uring"
2026-01-14 09:35:37.046 GMT [16336] HINT:  Available values: sync, worker.
2026-01-14 10:35:37.047 CET [16336] FATAL:  configuration file 
"/home/pierre/pgdata/postgresql.conf" contains errors
  stopped waiting
pg_ctl: could not start server

I don't know how to test io_uring features in code because if io_uring 
is not enabled at compilation time, I don't think it is possible to 
compile code calling  io_uring_get_probe_ring() or 
io_uring_opcode_supported() ?


Thanks.

  9. Re: [PATCH] check kernel version for io_method

    Jakub Wartak <jakub.wartak@enterprisedb.com> — 2026-01-14T11:14:35Z

    On Wed, Jan 14, 2026 at 3:31 AM Steven Niu <niushiji@gmail.com> wrote:
>
> From: Andreas Karlsson <andreas@proxel.se>
> Sent: Wednesday, January 14, 2026 09:07
> To: Pierre <pierre.forstmann@gmail.com>; pgsql-hackers@lists.postgresql.org <pgsql-hackers@lists.postgresql.org>
> Subject: Re: [PATCH] check kernel version for io_method
>
>
> On 1/13/26 11:05 PM, Pierre wrote:
> > Please find a patch proposal for bug BUG #19369: Not documented that
> > io_uring on kernel versions between 5.1 and below 5.6 does not work.
> Wouldn't it make more sense to use io_uring_get_probe_ring() and
> io_uring_opcode_supported() to actually check for the required opcode?
> Then if someone uses a Kernel with the required features backported
> everything will still work.
>
> Andreas
>
> +1
> It would be more reliable to check the availability of io_uring feature than to check OS version.
>

Hi,

I haven't looked at this patch, however the above statement is not
completely true. There is a parallel problem [1] related to kernel
version, where if you do not run proper kernel version (>= 6.5) or
proper liburing version, then fork() (-> all connections established)
are going to be slow slugging under more than basic load due to lack
of "combined memory mapping creation" (so technically speaking
recommending someone to go to 5.6.x but < 6.5 IMHO is also not good
advice). See first message in that [1] for a performance report about
this. IMHVO if we are checking for kernel versions we could also warn
about performance regression (something like merge those two patches
if one wants to have a good io_uring experience).

-J.

[1] - https://www.postgresql.org/message-id/CAKZiRmzxj6Lt1w2ffDoUmN533TgyDeYVULEH1PQFLRyBJSFP6w%40mail.gmail.com

  10. Re: [PATCH] check kernel version for io_method

    pierre.forstmann@gmail.com — 2026-01-14T15:29:43Z

    OK I was wrong: I can use io_uring_get_probe() if postgres has been 
built with liburing flag.

So I have replaced uname() call by io_uring_get_probe() call and I have 
modified the error messages: Linux kernel version is neither directly 
used  nor mentioned.


Le 14/01/2026 à 11:18, Pierre a écrit :
>
> Le 14/01/2026 à 08:31, Steven Niu a écrit :
>> From: Andreas Karlsson <andreas@proxel.se>
>> Sent: Wednesday, January 14, 2026 15:27
>> To: Steven Niu <niushiji@gmail.com>; Pierre 
>> <pierre.forstmann@gmail.com>; pgsql-hackers@lists.postgresql.org 
>> <pgsql-hackers@lists.postgresql.org>
>> Subject: Re: [PATCH] check kernel version for io_method
>>
>>
>> On 1/14/26 8:22 AM, Steven Niu wrote:
>>> I agree that you are correct for now. But in future, the IoMethod 
>>> may have new member. Then we have to process it.
>>>
>>> How about change to like this:
>>>     + #else
>>>     +     if (*newval == IOMETHOD_IO_URING)
>>>     +     {
>>>     +             GUC_check_errdetail("io_uring not enabled.");
>>>     +             return false;
>>>     +     }
>> No, that would be incorrect. It would not even compile.
>> IOMETHOD_IO_URING is not even defined if we do not compile with liburing
>> enabled. You can try it out yourself by compiling PostgreSQL without
>> io_uring support and then try to enable it. You will get an error and
>> that code will never be reached.
>>
>> Andreas
>>
>>
>>
>>
>> ________________________________________
>>
>>
>> Yes, you are definitely correct.
>> I'd like to withdwar my proposal.
>
> Hello,
>
> I have modified message in case io_uring is not enabled when building 
> PG. But currently I cannot test it, GUC parameter is rejected in 
> earlier steps:
>
> pg_ctl start
> waiting for server to start....2026-01-14 09:35:37.046 GMT [16336] 
> LOG:  invalid value for parameter "io_method": "io_uring"
> 2026-01-14 09:35:37.046 GMT [16336] HINT:  Available values: sync, 
> worker.
> 2026-01-14 10:35:37.047 CET [16336] FATAL:  configuration file 
> "/home/pierre/pgdata/postgresql.conf" contains errors
>  stopped waiting
> pg_ctl: could not start server
>
> I don't know how to test io_uring features in code because if io_uring 
> is not enabled at compilation time, I don't think it is possible to 
> compile code calling  io_uring_get_probe_ring() or 
> io_uring_opcode_supported() ?
>
>
> Thanks.
>

  11. Re: [PATCH] check kernel version for io_method

    Andreas Karlsson <andreas@proxel.se> — 2026-01-14T17:26:27Z

    On 1/14/26 12:14 PM, Jakub Wartak wrote:
> I haven't looked at this patch, however the above statement is not
> completely true. There is a parallel problem [1] related to kernel
> version, where if you do not run proper kernel version (>= 6.5) or
> proper liburing version, then fork() (-> all connections established)
> are going to be slow slugging under more than basic load due to lack
> of "combined memory mapping creation" (so technically speaking
> recommending someone to go to 5.6.x but < 6.5 IMHO is also not good
> advice). See first message in that [1] for a performance report about
> this. IMHVO if we are checking for kernel versions we could also warn
> about performance regression (something like merge those two patches
> if one wants to have a good io_uring experience).

We can probe for that too, which we already do. If you call 
pgaio_uring_ring_shmem_size() it will return 0 on Linux <6.5. Which I 
think eve further supports probing for the features we need rather than 
looking at the kernel version.

Andreas