svn commit: r304187 - in head: . share/man/man4 sys/conf sys/dev/mcd sys/modules sys/modules/mcd

[Warner Losh]

On Fri, Aug 19, 2016 at 5:27 AM, Konstantin Belousov
<kostikbel at gmail.com> wrote:
> On Fri, Aug 19, 2016 at 09:12:53PM +1000, Bruce Evans wrote:
>> On Fri, 19 Aug 2016, Konstantin Belousov wrote:
>>
>> > On Thu, Aug 18, 2016 at 09:28:57PM -0600, Warner Losh wrote:
>> >> On Thu, Aug 18, 2016 at 12:50 AM, Julian Elischer <julian at freebsd.org> wrote:
>> >>> On 16/08/2016 4:54 AM, John Baldwin wrote:
>> >>>>
>> >>>> On Monday, August 15, 2016 08:38:02 PM John Baldwin wrote:
>> >>>>> ,,,
>> >>>>> Log:
>> >>>>>    Remove the mcd(4) driver for Mitsumi CD-ROM players.
>> >>>>>       This is a driver for a pre-ATAPI ISA CD-ROM adapter.  As noted in
>> >>>>>    the manpage, this driver is only useful as a backend to cdcontrol to
>> >>>>>    play audio CDs since it doesn't use DMA, so its data performance is
>> >>>>>    "abysmal" (and that was true in the mid 90's).
>> >>>>
>> >>>> No one stepped up to test patches for it either when I last posted patches
>> >>>> to
>> >>>> convert it from timeout(9) to callout(9).  I have a few more drivers that
>> >>>> are
>> >>>> ...
>> >>>
>> >>> I would imagine any machine still holding one of these probably has not
>> >>> enough memory to run FreeBSD.
>> >>>
>> >>> would we still run in 2MB?
>> >>
>> >> With insane levels of tuning, we can run in 32MB userland that can do
>> >> things. Even 64MB is tight w/o some tuning. 16MB is almost certainly
>> >> right out except for very specialized situations. 2MB? We can't even
>> >> load the loader in that :(. Oh, and all these memory configs are only
>> >> possible if you tweak the loader's block cache...
>> >
>> > 32MB is quite usable.  Without any tuning, you get slightly less than 10MB
>> > for userspace, which is enough to for many things, and plenty if swap is
>> > added.
>>
>> No, 32MB needs lots of needs tuning.  -current seems to need about 16MB
>> more than just a few months ago when I last discussed this with you.
>> My i386 system doesn't have many drivers or a bloated userland (*),
>> but it took the following tuning plus my PAE tuning fixes to boot in
>> 32MB:
>> - disable [l]em1
>> - reduce tx and buffers to 256 for em0.  They default to 4096 for em.  That
>>    is something like 8K * 1500 bytes = 12MB for em0 alone.  lem[1] wants
>>    another 12MB.  I think this is not all statically allocated, but the
>>    drivers hang onto that much.
>> This now longer works.  -current without my PAE tuning fixes hangs mounting
>> root or in usb initialization with this tuning and 40MB.  -current with my
>> PAE tuning fixes hangs similarly with 32MB; with 40MB it boots to the
>> start of multiuser but then hangs (it starts 1 getty, then 2 sendmails
>> and kills them with "out of swap space") and 1 rpcbind (also killed).  I
>> don't use swap, but it was needed 20 years ago on a system that actually
>> had 32MB of memory.
>>
>> (*) /bin/sh doing nothing much in -current i386: size 6532K res 1924K
>>      /bin/sh                    in my ~5.2  i386: size  864K res  592K
>>
>> The kernel size is 5.5MB text 370K data 2.2MB bss (lots of bloat in bss
>> for debugging and vt).  In single user mode, with 40MB to start, there
>> is 22116K wired and 2516K free.  A few programs can be run in 2516K
>> without swap if they have res 592K and not 1924K.
>>
>> > Note that you cannot boot on such configurations since loader was broken,
>> > but if you do manage to jump to kernel, things were fine several months
>> > ago.  I tested my relatively recent OOM changes on 32MB qemu config.
>>
>> I have no problems booting such configuratations since I don't use the
>> current loader and only use old loader to change the environment to
>> set up special configurations like this.  I use my version of biosboot
>> for boot2.  This requires fixing 2 layers of complicated breakage in
>> init386().  vm86 is now used before the TSS and PIC resources that it
>> uses are initialized, but only in paths that are not normally used
>> because they are for memory sizing that is normally done by loader
>> :-(.
>>
>> I normally use my version of biosboot for boot2.  I improved its caching
>> just a couple of years ago.  It was using 9K buffer optimized for 1440K
>> floppies.  Now it uses a 32K buffer.  Booting a 5.5MB kernel takes a
>> fraction of a second.  Of course I don't use modules, so not many seeks
>> are needed.
>
> Of course I use modules and do not use GENERIC.  I just tried: with today
> HEAD, and old loader on 32MB VM, I get 11MB free in single-user mode.
> Active+inactive is ~4MB, and 1M is eaten by buffers, which is about right
> for init+/bin/sh+top idle system.

The situation on x86 must be a lot better than arm. My old Atmel boards with
32MB have < 1MB free when booted to the login prompt (more at single user)
and need special tuning to reduce the 5MB of network buffers allocated to
be anything approaching useful.

Warner