svn commit: r304142 - head/usr.sbin/bsdinstall/partedit

[Warner Losh]

On Fri, Aug 19, 2016 at 12:51 AM, Dag-Erling Smørgrav <des at des.no> wrote:
> Warner Losh <imp at bsdimp.com> writes:
>> Allan Jude <allanjude at freebsd.org> writes:
>> > Which makes more sense:
>> >
>> > A) If stripesize == 0, use some sane value like 4096
>>
>> I don't like this.
>>
>> > B) Some other combination that uses the reported stripe size, unless it
>> > is 0, in which case it uses 4096 (or some other value controlled by a
>> > different new sysctl)
>>
>> Don't like this so much.
>>
>> > C) create kern.geom.min_stripe_size with a default of 512, but users can
>> > set 4096 if they use only 4k devices. (doesn't really solve the problem
>> > for the installer)
>>
>> Default it to 4k, and allow users to set it to 512. If the drive
>> reports < this value
>> report this value instead.
>
> I don't like either option.  Option D (which I don't like either, but
> which should at least work in most cases) is a sysctl that specifies a
> minimum factor, and set the reported stripe size to the least common
> multiple of that number and the actual stripe or sector size.  This is
> what my bsdinstall patch does.  However, I think that pushing this down
> to a layer where it will affect all applications is a terrible idea,
> because we have no way of knowing what will break[*], and it can
> seriously mislead users and hinder troubleshooting - especially if it is
> enabled by default rather than only when necessary.

I took a look into the implications of doing a 4k stripesize 'automatically'
this morning. I found a few places in g_part where it would actively
hurt when coupled with gpart's insistence on aligning things. So I
now think it's a bad idea. This will make it harder for FreeBSD to
generate arbitrary disk layouts. And I'm not too sure about what
things like gstripe would report as a result and if this would actually
interfere if you had a large, but not power of two stripe size.

> I don't think it's a good idea to enforce stripe alignment everywhere,
> either.  It works for partitions because they are very large relative to
> the stripe size, and at worst we will waste a few millionths of the disk
> on inter-partition gaps, which should only occur between the partition
> table and the boot partition, and possibly, if the stripe size is very
> large, between the boot partition and the swap partition.  But forcing
> filesystems to respect the stripe size will lead to no end of trouble,
> because RAID volumes can have stripe sizes of 16 kB or more.  I think it
> is important to align partitions during installation because of the huge
> performance impact of misaligned partitions on AF disks, but despite
> what Nathan claims, I never advocated applying the same logic
> everywhere.

Yea, having poked at it for just a little while, I agree. The installer is the
right place to make sure we don't cross-thread the 4k sectors. Stripe size
means too many other things to have it be useful in that context.

Warner