[CFR][ZFS] merge Illumos rev. 13379 (writing to imbalanced vdevs)

[Martin Matuska]

Dňa 29.06.2011 19:42, Pawel Jakub Dawidek  wrote / napísal(a):

> On Mon, Jun 20, 2011 at 11:16:50AM +0200, Martin Matuska wrote:
>> zfs tries to allocate blocks evenly across all devices. This means when
>> devices are imbalanced zfs will lots of CPU searching for space on devices
>> which tend to be pretty full.
>>
>> https://www.illumos.org/issues/1051
>> https://www.illumos.org/projects/illumos-gate/repository/revisions/13379
>>
>> If there are no objections, I will commit this to -HEAD
> Some comments inline.
>
>> @@ -5137,6 +5138,7 @@
>>  	 */
>>  	kernel_init(FREAD | FWRITE);
>>  	VERIFY(spa_open(zs->zs_pool, &spa, FTAG) == 0);
>> +	spa->spa_debug = B_TRUE;
>>  	zs->zs_spa = spa;
>>  
>>  	spa->spa_dedup_ditto = 2 * ZIO_DEDUPDITTO_MIN;
> Do we want spa debug to be enabled by default? If we don't then please
> enable it only when DEBUG is defined.
>
We are here in ztest.c, spa_debug (new boolean) is disabled by default.

>>  /*
>> + * This value defines the number of allowed allocation failures per vdev.
>> + * If a device reaches this threshold in a given txg then we consider skipping
>> + * allocations on that device.
>> + */
>> +int zfs_mg_alloc_failures;
> In FreeBSD we probably want sysctl and loader tunable for this.
> I think it should be fine to make sysctl read/write, but I haven't
> looked very closely.
zio.c:
This is called in zio_init(), called by spa_init():
        /*
         * The zio write taskqs have 1 thread per cpu, allow 1/2 of the
taskqs
         * to fail 3 times per txg or 8 failures, whichever is greater.
         */
        zfs_mg_alloc_failures = MAX((3 * max_ncpus / 2), 8);

Looks like it is set on runtime so we have to change the code to make it
properly tunable and set a default value.
Maye comment out the code above and directly initialize with:

int zfs_mg_alloc_failures = MAX((3 * max_ncpus / 2), 8);


The setup might look like this:
TUNABLE_INT("vfs.zfs.mg_alloc_failures", &zfs_mg_alloc_failures);
SYSCTL_INT(_vfs_zfs, OID_AUTO, mg_alloc_failures, CTLFLAG_RW,
&zfs_mg_alloc_failures, 0,
    "Number of allowed allocation failures per vdev");

Another question is if it should be under vfs.zfs. or vfs.zfs.vdev.

-- 
Martin Matuska
FreeBSD committer
http://blog.vx.sk