Re: Swap partition priority

I can do an experiment.

I currently swap to a GEOM mirror. I can create an unmirrored fstab and take one disk offline and run the machine off a single disk and see if the swap writes require less memory for block IO.

-Joshua

> On Aug 18, 2026, at 10:44 AM, Mark Millard <[email protected]> wrote:
> 
> On 8/18/26 08:21, Joshua Armstrong wrote:
>> Short answer, yes - interleaving is hardcoded and the swapon(2) syscall
>> only accepts the device node name (or file name). Here’s the doc header
>> comment for the allocation code in sys/vm/swap_pager.c:
>> 
>>     915 /*
>> 
>>     916  * SWP_PAGER_GETSWAPSPACE() -   allocate raw swap space
>> 
>>     917  *
>> 
>>     918  *      Allocate swap for up to the requested number of pages.  The
>> 
>>     919  *      starting swap block number (a page index) is returned or
>> 
>>     920  *      SWAPBLK_NONE if the allocation failed.
>> 
>>     921  *
>> 
>>     922  *      Also has the side effect of advising that somebody made
>> a mistake
>> 
>>     923  *      when they configured swap and didn't configure enough.
>> 
>>     924  *
>> 
>>     925  *      This routine may not sleep.
>> 
>>     926  *
>> 
>>     927  *      We allocate in round-robin fashion from the configured
>> devices.
>> 
>>     928  */
>> 
>> 
>> However, the suggestion to use a concatenated GEOM should work great
>> because that’s a single device according to the swap and the GEOM
>> allocator is (should be) entirely transparent to the swap strategy:
>> 
>> 
>> Relevant snippet from swapgeom_strategy:
>> 
>> 
>>    3212         bp->b_caller1 = bio;
>> 
>>    3213         bio->bio_caller1 = sp;
>> 
>>    3214         bio->bio_caller2 = bp;
>> 
>>    3215         bio->bio_cmd = bp->b_iocmd;
>> 
>>    3216         bio->bio_offset = (bp->b_blkno - sp->sw_first) * PAGE_SIZE;
>> 
>>    3217         bio->bio_length = bp->b_bcount;
>> 
>>    3218         bio->bio_done = swapgeom_done;
>> 
>>    3219         bio->bio_flags |= BIO_SWAP;
>> 
>>    3220         if (!buf_mapped(bp)) {
>> 
>>    3221                 bio->bio_ma = bp->b_pages;
>> 
>>    3222                 bio->bio_data = unmapped_buf;
>> 
>>    3223                 bio->bio_ma_offset = (vm_offset_t)bp->b_offset &
>> PAGE_MASK;
>> 
>>    3224                 bio->bio_ma_n = bp->b_npages;
>> 
>>    3225                 bio->bio_flags |= BIO_UNMAPPED;
>> 
>>    3226         } else {
>> 
>>    3227                 bio->bio_data = bp->b_data;
>> 
>>    3228                 bio->bio_ma = NULL;
>> 
>>    3229         }
>> 
>>    3230         g_io_request(bio, cp);
>> 
>>    3231         return;
>> 
>> 
>> 
>>> On Aug 18, 2026, at 9:00 AM, Joshua Armstrong <[email protected]>
>>> wrote:
>>> 
>>> I’ll throw my hat into this ring.
>>> 
>>> As soon as I’m finished with my meeting, I’ll check the swap
>>> management / virtual memory code to see if interleaving is hardcoded
>>> or not. I believe the current (and long-standing, according to my
>>> ancient copy of the Unix System Administration Handbook) behavior is
>>> to interleave swap writes between configured devices for performance.
>>> 
>>> I agree, however, that having a priority system would make more sense
>>> in today’s world. I sometimes end up creating a temporary zvol and
>>> using that as emergency extra swap if I’m building world, but I would
>>> assume this has a performance penalty (though perhaps not massive)
>>> versus swap partitions on bare disks.
>>> 
>>> -Joshua
>>> 
>>>> On Aug 18, 2026, at 3:48 AM, Miroslav Lachman <[email protected]> wrote:
>>>> 
>>>> On 18/08/2026 09:49, Stéphane Rochoy wrote:
>>>>> Andrea Venturoli <[email protected]> writes:
>>>>>> Is it possible to set swap partition priority?
>>>>>> According to swapon(8), the answer is no, but I wonder if the man page
>>>>>> might be outdated, if patches are available, etc...
>>>>>> 
>>>>>> The reason is I have SSDs and HDDs in the same system: I normally use
>>>>>> swap on SSDs (for speed obviously), but sometimes the amount is not
>>>>>> enough.
>>>>>> So I'm ponderind adding swap on HDDS too, but this would slow down
>>>>>> everyrting unless it's really used as last resort.
>>>>> Hello,
>>>>> Long ago I came accross a setup were the swap device was a GEOM mirror
>>>>> (see gmirror(8)). AFAIK it was working except for vmcores: the kernel
>>>>> was always using the first device which required to help savecore(8) a
>>>>> bit.
>>>>> So maybe there's some GEOM class matching your needs.
>>>> There is geom_concat - see man gconcat. It may be possible to use it
>>>> for swap (never tried it).
>>>> 
>>>> Best regards
>>>> Miroslav Lachman
>>>> 
>>> 
>> 
> 
> Konstantin Belousov once wrote on the freebsd-arm list:
> 
> QUOTE
> . . .
> swapfile write requires the write request to come through the filesystem
> write path, which might require the filesystem to allocate more memory
> and read some data. E.g. it is known that any ZFS write request
> allocates memory, and that write request on large UFS file might require
> allocating and reading an indirect block buffer to find the block number
> of the written block, if the indirect block was not yet read.
> 
> As result, swapfile swapping is more prone to the trivial and
> unavoidable deadlocks where the pagedaemon thread, which produces free
> memory, needs more free memory to make a progress.  Swap write on the
> raw partition over simple partitioning scheme directly over HBA are
> usually safe, while e.g. zfs over geli over umass is the worst construction.
> END QOUTE
> 
> I do not know if the suggested use of GEOM is an example of that type of
> issue getting involved.
> 
> -- 
> ===
> Mark Millard
> marklmi at yahoo.com