Date: Sun, 25 Sep 2016 11:19:55 -0600 From: Warner Losh <imp@bsdimp.com> To: Russell Haley <russ.haley@gmail.com> Cc: Mark Millard <markmi@dsl-only.net>, freebsd-arm <freebsd-arm@freebsd.org>, FreeBSD-STABLE Mailing List <freebsd-stable@freebsd.org> Subject: Re: 11.0-RELEASE tier level for arm64/aaarch64 and the officially built arm/armv6 variants? Message-ID: <CANCZdfo9BMThmEvnwmm3rg1OMU=SsETKSBUiX-wwjUDdnT2M5Q@mail.gmail.com> In-Reply-To: <CABx9NuTFobc1e4UxsLSPKLTtMgt37QDgR99xHhkdCkc7qPHCxA@mail.gmail.com> References: <4076CFFA-7BE2-4E1B-A7E8-08FD8FC27D21@dsl-only.net> <332FA120-31E5-4D31-B63E-A0DFDD7DEFC7@dsl-only.net> <CANCZdfqM17qzZg2SqwJRTWO67KCnAC%2BHYKatcb8CBHF3TM7kFg@mail.gmail.com> <4F04B0C0-DAB1-4EEA-A3A0-4FAA89AD93E2@dsl-only.net> <CABx9NuRv69k0u84%2B3bpDGPM1U7DsZJUtBC42jktk3ttinY9p%2BA@mail.gmail.com> <CANCZdfpBzh15fJdQ559MQv1J=5TrbSfgdziKMVUW-kDF%2BEEwww@mail.gmail.com> <CABx9NuTFobc1e4UxsLSPKLTtMgt37QDgR99xHhkdCkc7qPHCxA@mail.gmail.com>
next in thread | previous in thread | raw e-mail | index | archive | help
On Sun, Sep 25, 2016 at 1:13 AM, Russell Haley <russ.haley@gmail.com> wrote= : > On Sat, Sep 24, 2016 at 11:10 PM, Warner Losh <imp@bsdimp.com> wrote: >> On Sat, Sep 24, 2016 at 10:21 PM, Russell Haley <russ.haley@gmail.com> w= rote: >>> On Sat, Sep 24, 2016 at 5:12 PM, Mark Millard <markmi@dsl-only.net> wro= te: >>>> On 2016-Sep-24, at 2:11 PM, Warner Losh <imp@bsdimp.com> wrote: >>>> >>>>> On Fri, Sep 23, 2016 at 8:29 PM, Mark Millard <markmi@dsl-only.net> w= rote: >>>>>> [A resend since I forget to list free-arm in the To: the first time.= ] >>>>>> >>>>>> From https://www.freebsd.org/platforms/arm.html : >>>>>> >>>>>>> 32-bit ARM is officially a Tier 2 architecture, as the FreeBSD proj= ect does not provide official releases or pre-built packages for this platf= orm due to it primarily targeting the embedded arena. However, FreeBSD/ARM = is being actively developed and maintained, is well supported, and provides= an excellent framework for building ARM-based systems. FreeBSD/arm support= s ARMv4 and ARMv5 processors. FreeBSD/armv6 supports ARMv6 and ARMv7 proces= sors, including SMP on the latter. >>>>>> >>>>>> "does not provide official releases or pre-built packages"? >>>>>> >>>>>>> # uname -apKU >>>>>>> FreeBSD rpi2 11.0-PRERELEASE FreeBSD 11.0-PRERELEASE #5 r304943M: S= un Aug 28 03:17:54 PDT 2016 markmi@FreeBSDx64:/usr/obj/clang/arm.armv6/= usr/src/sys/RPI2-NODBG arm armv6 1100502 1100502 >>>>>> >>>>>>> # pkg search '.*' | wc >>>>>>> 21349 155540 1596736 >>>>>> >>>>>> Will 11.0-RELEASE change the tier level for any of the specific arm-= armv6 variants that have FreeBSD-11.0-*-arm-armv6-*.img* files built, such = as for RPI2? >>>>>> >>>>>> Even if all the officially built arm-armv6 variants stay tier 2, the= wording on the web page likely needs to be changed because so much is buil= t and available that the above quote claims is not available. >>>>> >>>>> armv6 is basically Tier 1 right now, though not as Tier 1 as i386 or >>>>> amd64 due to the fragmented nature of the arm world. On the platforms >>>>> we run on and create releases for, however, it's my opinion that it i= s >>>>> Tier 1: it has been running in production a while, things people >>>>> expect from a FreeBSD system are present, you can get decent support >>>>> if you ask questions, there's no known major gotchas in deploying thi= s >>>>> hardware. The only remaining annoying issue is the 'u-boot' problem >>>>> where we have to have a different u-boot image for every board and no >>>>> standardized way to convert a 'generic' image into one that's specifi= c >>>>> for specific boards. >>> >>> I'll point out again that barebox is an excellent alternative to u-boot= (IMHO): >> >> Doesn't matter, still has the same issues that u-boot has. > u-boot has a different sources for almost each board we support (due > to the usual FOSS issues). That is NOT the case in barebox. There is > one source and it's kept up to date by the team, not the vendors. Actually, that's no longer the case. All boards we support have all their bits in u-boot's main repo. But you can't say no one will fork barebox to get their support out faster, which is why we have the current situation in ports... u-boot forked because it was so popular, and the forks rejoined. >> But does it support the u-boot ABI? > I'd have to look into this. >> >>> - Supports most, if not all of the boards that FreeBSD supports, plus >>> many that it doesn't >>> - Single source tree for all boards. Specify build time parameters to >>> build one or all the images >>> - Well supported community with central maintainer-ship >>> - Simple, familiar shell interface (*awesome*) >>> - Excellent documentation (u-boots is good too though) >>> - Has support for (U)EFI >>> - Supports quemu aarch64 (not *quite*sure what the means though) >> >> Right, u-boot has all these things, except maybe the shell interface >> (not sure what you mean by that). > Instead of stringing together variables and commands, it uses a > scripting language like a simplified sh. Want to change how something > boots? Update a script. Save it to disk (it has it's own persistence > mechanisms) and export it. Sounds cool. >>> To be fair, I'm not saying the problem is the fault of denx, but >>> barebox has a lot going for it. The maintainer was very keen to see it >>> ported top FreeBSD and was willing to support the effort. I ran into >>> some build time linux api requirements, but he didn't think that would >>> be much to overcome (and it wasn't I just kept adding the patches he >>> sent me and the build moved forward. As always, I ran out of time for >>> the really fun stuff). While I am a hopeless dreamer and I'm sure I've >>> over simplified the problem, I thought it would be neat to see FreeBSD >>> upstream support for zfs and ufs to the barebox boot loader and do >>> away with ubldr. We would then have a modern, easy to use, boot loader >>> that supports the standard startup toolchain. >> >> We can't easily do away with ubldr if we want to support tunables, kerne= l >> modules loaded at boot time and a few other nifty features like nextboot= . > > Are these things not in standard loader? Should they be? ubldr is the standard loader, just spelled in a funky way and living on a place u-boot can see :). We have different loaders for different environments. even on x86 we have /boot/loader for BIOS boots and /boot/loader.efi for EFI boots. >>> Either way, if installers move to a pkgng based method (so cool) then >>> installing u-boot and arm binaries from pkg-static will be the same as >>> x86 (ha ha I said that with a straight face!). >> >> Yea, not so much. You have to build the bootable image not on the >> target system, like you do on x86. > > Doesn't the current ports and packages cross build everything that's mark= ed? For builds from source, yes. But that's different than installing from FreeBSD's installer. >> We'd have to have something that >> installs uboot onto a generic image (perhaps with hooks for kernels >> since those aren't generic on armv6) and then put that into a bootable >> SD card. > > The x86 installer (that I argue is a platform legacy) has to customize > the bootloader for each installation. If we HAVE to use an installer > on arm, what wrong prompting the user for some input (i.e. what som > are you using) while including all the u-boots in the ports tree and > all the supported kernels, then just installing the correct ones for > the board (with input from the user)? Ummm, you can't boot anything to do that work like you can on x86. >>>>>For x86 this is all done with the installer since >>>>> that boot environment is more standardized. Does this last issue keep >>>>> arm from being Tier 1? That's a judgement call, but I think the >>>>> project should promote w/o this last issue. >>> >>> How does a platform get promoted? Is that something the Core team decid= es? >> >> Yes. >> >>> I see two facts about current Arm support that show platform maturity: >>> a) u-boot is in the ports tree and we have Lego-easy build scripts in >>> crochet that could be called an installation method. Building for arm >>> is not difficult anymore. >> >> Except corchet isn't in the tree, and the solution is horrible. It's >> a script for each SoC, and those scripts are now scattered about. >> Plus that's a creation from source model, not a creation from >> RE produced bits model, which is needed for Tier 1. > Both correctable sins, especially as it's BSD licensed. My point was > that even building a SOM specific image is relatively painless with > the right scripts. Hell, I've even got a custom build script that > could be modified for generic use. Yea, a developer can do it with little pain. However, compared to the pain to boot Linux on these boards, the pain is quite high and you need to know way too much to make it work. BTW, my 'horrible' comment was given the state of things today: we have scattered knowledge of how to make images, custom one-off that work on some boards well, and fail on others. We have crazy, quirky things that aren't well integrated into the tree. Crochet itself solves an interesting problem, but does so a bit in isolation. >>> b)Arm requires images, not installers. Correct me if I'm wrong but, >>> installers are a tool primarily invented for x86 PC type computers. >>> FreeBSD publishes standardized ISOs for all supported Arm platforms >>> that work by simply "xzcat | dd" onto the sd card (or wherever you >>> need it). I'm not sure how standardized or manual that build process >>> is, but I would think that if the Arm platform support is able to keep >>> up with the standard FreeBSD release cycle (i.e. not break every other >>> release) then there would be no reason to NOT call it tier 1? >> >> Creating the images is currently a pita. That's it. > I see. So not so mature. Yes. It's an area I work on when I have time to make better... Warner >>> What I don't know about is "official" documentation for building for >>> arm and supporting cross building to Arm. Will someone need to write >>> an "Arm Handbook" to be promoted? >> >> No. While useful, most of that already exists. > > Thanks for the response Warner, I always appreciate the chance to > learn more about FreeBSD. > Russ > >> Warner >> >>>> Interesting and good to know. Thanks. >>>> >>>> I might have guessed that going along with the u-boot issue would be t= he fanout in: >>>> >>>> 11.0/sys/arm/ . . . >>>> >>>> allwinner/a10/ >>>> allwinner/a20/ >>>> allwinner/a31/ >>>> allwinner/a83t/ >>>> allwinner/h3/ >>>> . . . >>>> broadcom/bcm2835/ >>>> . . . >>>> >>>> (Full list not shown.) >>>> >>>> I was thinking that this might make the tier level specific to the sta= tus of each such directory's content so that it was the combination of that= and the sysutils/u-boot-*/ status that made the difference for assigning t= he level. I'd guess that lack of a usable directory in either place would = not be tier 2 even. Similarly until the required sys/arm/*/* and sysutils/u= -boot-*/ directory-tree content have reached a sufficiently complete status= . >>>> >>>> I'd expect that there will always be a lag for what exists in the worl= d vs. what has these materials worked out in FreeBSD. >>>> >>>> >>>>>> Also from https://www.freebsd.org/platforms/arm.html : >>>>>> >>>>>>> Initial support for 64-bit ARM is complete. 64-bit ARM platforms fo= llow a set of standard conventions, and a single FreeBSD build will work on= hardware from multiple vendors. As a result, FreeBSD will provide official= releases for FreeBSD/arm64 and packages will be available. FreeBSD/arm64 i= s on the path to becoming a Tier 1 architecture. >>>>>> >>>>>> Will 11.0-RELEASE make arm64/aarch64 Tier 1? >>>>>> >>>>>> [I will note that, while there are no official builds for the Pine64= family (A64 based) that are under the Allwinner arm activity, the SOC's in= volved are Cortex-A53 64-bit arm based. They likely do not fit in the "stan= dard conventions" or arm64/aarch64 would be where they would have been supp= orted. Some rewording might be appropriate for the above quote as well.] >>>>> >>>>> No. aarch64 isn't Tier 1 yet. There's many small bits that are >>>>> missing. It is quite solidly Tier 2, but we don't have a linker, we >>>>> don't have widespread hardware availability, we don't have production >>>>> experience with the platform. Most things work, but there's still som= e >>>>> gotchas. There's still the 'u-boot' problem with many arm64 systems >>>>> because for systems that use u-boot to bootstrap UEFI, you need a >>>>> different image for each board (some closely related board families >>>>> can get by with one to be pedantic). All these issues are still >>>>> significant barriers to production use. It's not been officially >>>>> promoted yet and I don't think the time is quite right yet. >>>> >>>> Intersting as well. I'd guess that conceptually this probably would ap= ply to both: >>>> >>>> sys/arm/allwinner/a64/ and sysutils/u-boot/pine64/ >>>> (presuming, contrary to fact, that 11.0 had sys/arm/allwinner/a64/ ) >>>> >>>> and. . . >>>> >>>> sys/arm64/cavium/ >>>> sys/arm64/cloudabi64/ >>>> >>>> So just sys/arm/ vs. sys/arm64/ for an aarch64 would not really make a= difference yet for tier level. >>>> >>>>> Warner >>>> >>>> Thanks again for the notes. >>>> >>>> =3D=3D=3D >>>> Mark Millard >>>> markmi at dsl-only.net >>>> >>>> _______________________________________________ >>>> freebsd-arm@freebsd.org mailing list >>>> https://lists.freebsd.org/mailman/listinfo/freebsd-arm >>>> To unsubscribe, send any mail to "freebsd-arm-unsubscribe@freebsd.org"
Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?CANCZdfo9BMThmEvnwmm3rg1OMU=SsETKSBUiX-wwjUDdnT2M5Q>