Toomas Soome
2017-May-16 16:20 UTC
EFI loader doesn't handle md_preload (md_image) correct?
> On 16. mai 2017, at 19:13, Harry Schmalzbauer <freebsd at omnilan.de> wrote: > > Bez?glich Toomas Soome's Nachricht vom 16.05.2017 18:00 (localtime): >> >>> On 16. mai 2017, at 18:45, Harry Schmalzbauer <freebsd at omnilan.de >>> <mailto:freebsd at omnilan.de>> wrote: >>> >>> Bez?glich Harry Schmalzbauer's Nachricht vom 16.05.2017 17:28 (localtime): >>>> Bez?glich Toomas Soome's Nachricht vom 16.05.2017 16:57 (localtime): >>>>>> On 16. mai 2017, at 17:55, Harry Schmalzbauer <freebsd at omnilan.de >>>>>> <mailto:freebsd at omnilan.de>> wrote: >>>>>> >>>>>> Hello, >>>>>> >>>>>> unfortunately I had some trouble with my preferred MFS-root setups. >>>>>> It seems EFI loader doesn't handle type md_image correctly. >>>>>> >>>>>> If I load any md_image with loader invoked by gptboot or gptzfsboot, >>>>>> 'lsmod' >>>>>> shows "elf kernel", "elf obj module(s)" and "md_image". >>>>>> >>>>>> Using the same loader.conf, but EFI loader, the md_image-file is >>>>>> prompted and sems to be loaded, but not registered. There's no >>>>>> md_image >>>>>> with 'lsmod', hence it's not astonsihing that kernel doesn't attach md0 >>>>>> so booting fails since there's no rootfs. >>>>>> >>>>>> Any help highly appreciated, hope Toomas doesn't mind beeing >>>>>> initially CC'd. >>>>>> >>>>>> Thanks, >>>>>> >>>>>> -harry >>>>> >>>>> The first question is, how large is the md_image and what other >>>>> modules are loaded? >>>> Thanks for your quick response. >>>> >>>> The images are 50-500MB uncompressed (provided by gzip compressed file). >>>> Small ammount of elf modules, 5, each ~50kB. >>> >>> On the real HW, there's vmm and some more: >>> Id Refs Address Size Name >>> 1 46 0xffffffff80200000 16M kernel >>> 2 1 0xffffffff8121d000 86K unionfs.ko >>> 3 1 0xffffffff81233000 3.1M zfs.ko >>> 4 2 0xffffffff81545000 51K opensolaris.ko >>> 5 7 0xffffffff81552000 279K usb.ko >>> 6 1 0xffffffff81598000 67K ukbd.ko >>> 7 1 0xffffffff815a9000 51K umass.ko >>> 8 1 0xffffffff815b6000 46K aesni.ko >>> 9 1 0xffffffff815c3000 54K uhci.ko >>> 10 1 0xffffffff815d1000 65K ehci.ko >>> 11 1 0xffffffff815e2000 15K cc_htcp.ko >>> 12 1 0xffffffff815e6000 3.4M vmm.ko >>> 13 1 0xffffffffa3a21000 12K ums.ko >>> 14 1 0xffffffffa3a24000 9.1K uhid.ko >>> >>> Providing md_image uncompressed doesn't change anything. >>> >>> Will deploy a /usr separated rootfs, which is only ~100MB uncompressed >>> and see if that changes anything. >>> That's all I can provide, code is far beyond my knowledge... >>> >>> -harry >> >> >> The issue is, that current UEFI implementation is using 64MB staging >> memory for loading the kernel and modules and files. When the boot is >> called, the relocation code will put the bits from staging area into the >> final places. The BIOS version does not need such staging area, and that >> will explain the difference. >> >> I actually have different implementation to address the same problem, >> but thats for illumos case, and will need some work to make it usable >> for freebsd; the idea is actually simple - allocate staging area per >> loaded file and relocate the bits into the place by component, not as >> continuous large chunk (this would also allow to avoid the mines like >> planted by hyperv;), but right now there is no very quick real solution >> other than just build efi loader with larger staging size. > > Ic, thanks for the explanation. > While not aware about the purpose of the staging area nor the > consequences of enlarging it, do you think it's feasable increasing it > to 768Mib? > > At least now I have an idea baout the issue and an explanation why > reducing md_imgae to 100MB hasn't helped ? still more than 64... > > Any quick hint where to define the staging area size highly appreciated, > fi there are no hard objections against a 768MB size. > > -harryThe problem is that before UEFI Boot Services are not switched off, the memory is managed (and owned) by the firmware, and you can not write into the random places. So you need to allocate the staging area, read data there, switch BS off, move data into final location and start the kernel. Alternatively, you move only the kernel into place, and let the kernel to relocate and manage the modules and MD image. rgds, toomas
Harry Schmalzbauer
2017-May-16 16:26 UTC
EFI loader doesn't handle md_preload (md_image) correct?
Bez?glich Toomas Soome's Nachricht vom 16.05.2017 18:20 (localtime):> >> On 16. mai 2017, at 19:13, Harry Schmalzbauer <freebsd at omnilan.de> wrote: >> >> Bez?glich Toomas Soome's Nachricht vom 16.05.2017 18:00 (localtime): >>> >>>> On 16. mai 2017, at 18:45, Harry Schmalzbauer <freebsd at omnilan.de >>>> <mailto:freebsd at omnilan.de>> wrote: >>>> >>>> Bez?glich Harry Schmalzbauer's Nachricht vom 16.05.2017 17:28 (localtime): >>>>> Bez?glich Toomas Soome's Nachricht vom 16.05.2017 16:57 (localtime): >>>>>>> On 16. mai 2017, at 17:55, Harry Schmalzbauer <freebsd at omnilan.de >>>>>>> <mailto:freebsd at omnilan.de>> wrote: >>>>>>> >>>>>>> Hello, >>>>>>> >>>>>>> unfortunately I had some trouble with my preferred MFS-root setups. >>>>>>> It seems EFI loader doesn't handle type md_image correctly. >>>>>>> >>>>>>> If I load any md_image with loader invoked by gptboot or gptzfsboot, >>>>>>> 'lsmod' >>>>>>> shows "elf kernel", "elf obj module(s)" and "md_image". >>>>>>> >>>>>>> Using the same loader.conf, but EFI loader, the md_image-file is >>>>>>> prompted and sems to be loaded, but not registered. There's no >>>>>>> md_image >>>>>>> with 'lsmod', hence it's not astonsihing that kernel doesn't attach md0 >>>>>>> so booting fails since there's no rootfs. >>>>>>> >>>>>>> Any help highly appreciated, hope Toomas doesn't mind beeing >>>>>>> initially CC'd. >>>>>>> >>>>>>> Thanks, >>>>>>> >>>>>>> -harry >>>>>> >>>>>> The first question is, how large is the md_image and what other >>>>>> modules are loaded? >>>>> Thanks for your quick response. >>>>> >>>>> The images are 50-500MB uncompressed (provided by gzip compressed file). >>>>> Small ammount of elf modules, 5, each ~50kB. >>>> >>>> On the real HW, there's vmm and some more: >>>> Id Refs Address Size Name >>>> 1 46 0xffffffff80200000 16M kernel >>>> 2 1 0xffffffff8121d000 86K unionfs.ko >>>> 3 1 0xffffffff81233000 3.1M zfs.ko >>>> 4 2 0xffffffff81545000 51K opensolaris.ko >>>> 5 7 0xffffffff81552000 279K usb.ko >>>> 6 1 0xffffffff81598000 67K ukbd.ko >>>> 7 1 0xffffffff815a9000 51K umass.ko >>>> 8 1 0xffffffff815b6000 46K aesni.ko >>>> 9 1 0xffffffff815c3000 54K uhci.ko >>>> 10 1 0xffffffff815d1000 65K ehci.ko >>>> 11 1 0xffffffff815e2000 15K cc_htcp.ko >>>> 12 1 0xffffffff815e6000 3.4M vmm.ko >>>> 13 1 0xffffffffa3a21000 12K ums.ko >>>> 14 1 0xffffffffa3a24000 9.1K uhid.ko >>>> >>>> Providing md_image uncompressed doesn't change anything. >>>> >>>> Will deploy a /usr separated rootfs, which is only ~100MB uncompressed >>>> and see if that changes anything. >>>> That's all I can provide, code is far beyond my knowledge... >>>> >>>> -harry >>> >>> >>> The issue is, that current UEFI implementation is using 64MB staging >>> memory for loading the kernel and modules and files. When the boot is >>> called, the relocation code will put the bits from staging area into the >>> final places. The BIOS version does not need such staging area, and that >>> will explain the difference. >>> >>> I actually have different implementation to address the same problem, >>> but thats for illumos case, and will need some work to make it usable >>> for freebsd; the idea is actually simple - allocate staging area per >>> loaded file and relocate the bits into the place by component, not as >>> continuous large chunk (this would also allow to avoid the mines like >>> planted by hyperv;), but right now there is no very quick real solution >>> other than just build efi loader with larger staging size. >> >> Ic, thanks for the explanation. >> While not aware about the purpose of the staging area nor the >> consequences of enlarging it, do you think it's feasable increasing it >> to 768Mib? >> >> At least now I have an idea baout the issue and an explanation why >> reducing md_imgae to 100MB hasn't helped ? still more than 64... >> >> Any quick hint where to define the staging area size highly appreciated, >> fi there are no hard objections against a 768MB size. >> >> -harry > > The problem is that before UEFI Boot Services are not switched off, the memory is managed (and owned) by the firmware,Hmm, I've been expecting something like that (owend by firmware) ;-) So I'll stay with CSM for now, and will happily be an early adopter if you need someone to try anything (-stable mergable). Thanks, -harry