Re: [UEFI] Boot issues on some UEFI implementations

From: Toomas Soome <>
Date: Thu, 26 Jul 2018 19:23:43 +0300
> On 26 Jul 2018, at 16:58, O. Hartmann <> wrote:
> On Wed, 25 Jul 2018 07:30:32 -0700 (PDT)
> "Rodney W. Grimes" < <>> wrote:
>>>> On 25 Jul 2018, at 12:10, O. Hartmann <> wrote:
>>>> On Wed, 25 Jul 2018 11:46:07 +0300
>>>> Toomas Soome <> wrote:
>>>>>> On 25 Jul 2018, at 10:59, O. Hartmann <> wrote:
>>>>>> On Tue, 24 Jul 2018 08:53:36 +0300
>>>>>> Toomas Soome <> wrote:
>>>>>> Hello  Toomas Soome,
>>>>>> I CC Allan Jude since I discovered something  weird today regarding the
>>>>>> UEFI boot capabilities of USB flash devices and SSDs. See below.
>>>>>>>> On 24 Jul 2018, at 08:16, O. Hartmann <> wrote:
>>>>>>>> On Mon, 23 Jul 2018 10:56:04 +0300
>>>>>>>> Toomas Soome <> wrote:
>>>>>>>>>> On 23 Jul 2018, at 10:27, O. Hartmann <>
>>>>>>>>>> wrote:
>>>>>>>>>> On Fri, 13 Jul 2018 18:44:23 +0300
>>>>>>>>>> Toomas Soome < <>> wrote:
>>>>>>>>>>>> On 13 Jul 2018, at 17:44, O. Hartmann <
>>>>>>>>>>>> <>> wrote:
>>>>>>>>>>>> -----BEGIN PGP SIGNED MESSAGE-----
>>>>>>>>>>>> Hash: SHA512
>>>>>>>>>>>> Am Fri, 13 Jul 2018 14:26:51 +0300
>>>>>>>>>>>> Toomas Soome < <>
>>>>>>>>>>>> < <>>> schrieb:       
>>>>>>>>>>>>>> On 13 Jul 2018, at 14:00, O. Hartmann <>
>>>>>>>>>>>>>> wrote:
>>>>>>>>>>>>>> The problem is some kind of weird. I face UEFI boot problems on
>>>>>>>>>>>>>> GPT drives where the first partition begins at block 40 of the
>>>>>>>>>>>>>> hdd/ssd.
>>>>>>>>>>>>>> I have two host in private use based on an
>>>>>>>>>>>>>> outdated ASRock Z77-Pro4-M and Z77-Pro4 mainboard (IvyBridge,
>>>>>>>>>>>>>> Socket LGA1155). Both boards are equipted with the lates
>>>>>>>>>>>>>> official available AMI firmware revision, dating to 2013. This
>>>>>>>>>>>>>> is for the Z77-Pro4-M revision 2.0 (2013/7/23) and for the Z77
>>>>>>>>>>>>>> Pro4 revision 1.8 (2013/7/17). For both boards a BETA revision
>>>>>>>>>>>>>> for the Spectre/Meltdown mitigation is available, but I didn't
>>>>>>>>>>>>>> test that. But please read.
>>>>>>>>>>>>>> The third box I realised this problem is a brand new Fujitsu
>>>>>>>>>>>>>> Esprimo Q956, also AMI firmware, at V5.0.0.11 R 1.26.0 for
>>>>>>>>>>>>>> 3413-A1x, date 05/25/2018 (or 20180525).
>>>>>>>>>>>>>> Installing on any kind of HDD or SSD manually or via bsdinstall
>>>>>>>>>>>>>> the OS using UEFI-only boot method on a GPT partitioned device
>>>>>>>>>>>>>> fails. The ASRock boards jump immediately into the firmware,
>>>>>>>>>>>>>> the Fujitsu offers some kind of CPU/Memory/HDD test facility.
>>>>>>>>>>>>>> If on both type of vendor/boards CSM is disabled and UEFI boot
>>>>>>>>>>>>>> only is implied, the MBR partitioned FreeBSD installation USB
>>>>>>>>>>>>>> flash device does boot in UEFI! I guess I can assume this when
>>>>>>>>>>>>>> the well known clumsy 80x25 char console suddenly gets bright
>>>>>>>>>>>>>> and shiny with a much higher resoltion as long the GPU supports
>>>>>>>>>>>>>> EFI GOP. Looking with gpart at the USB flash drives reveals
>>>>>>>>>>>>>> that the EFI partition starts at block 1 of the device and the
>>>>>>>>>>>>>> device has a MBR layout. I haven't found a way to force the GPT
>>>>>>>>>>>>>> scheme, when initialised via gpart, to let the partitions start
>>>>>>>>>>>>>> at block 1. This might be a naiv thinking, so please be patient
>>>>>>>>>>>>>> with me.
>>>>>>>>>>>>>> I do not know whether this is a well-known issue. On the ASRock
>>>>>>>>>>>>>> boards, I tried years ago some LinuxRed Hat and Suse with UEFI
>>>>>>>>>>>>>> and that worked - FreeBSD not. I gave up on that that time.
>>>>>>>>>>>>>> Now, having the very same issues with a new Fujitsu system,
>>>>>>>>>>>>>> leaves me with the impression that FreeBSD's UEFI
>>>>>>>>>>>>>> implementation might have problems I'm not aware of.
>>>>>>>>>>>>>> Can someone shed some light onto this? 
>>>>>>>>>>>>> The first thing to check is if the secure boot is disabled. We
>>>>>>>>>>>>> do not support secure boot at all at this time.          
>>>>>>>>>>>> Secure boot is in every scenario disabled!
>>>>>>>>>>>>> If you have efi or bios version running - you can check from
>>>>>>>>>>>>> either console variable value (it can have efi or vidconsole or
>>>>>>>>>>>>> comconsole) or better yet, see if efi-version is set (show
>>>>>>>>>>>>> efi-version) - if efi-version is not set, it is BIOS loader
>>>>>>>>>>>>> running. Another indirect way is to see lsdev -v, with device
>>>>>>>>>>>>> paths present, it is uefi:)          
>>>>>>>>>>>> What are you talking about?
>>>>>>>>>>>> What is the point of entry - running system, loader?
>>>>>>>>>>>> sysct machdep.bootmethod: BIOS
>>>>>>>>>>>> This makes me quite sure that the system has booted via BIOS - as
>>>>>>>>>>>> I'm sure since I've checked that many times. UEFI doesn't work on
>>>>>>>>>>>> those systems with FreeBSD. I'm not sure antmore, but I tried
>>>>>>>>>>>> also Windows 7 on those mainboards booting via UEFI - and I might
>>>>>>>>>>>> recall that they failed also. I also recall that there were
>>>>>>>>>>>> issues with earlier UEFI versions regarding booting only Windows
>>>>>>>>>>>> 8/8.1 - and nothing else, but the fact that Linux worked confuses
>>>>>>>>>>>> me a bit.
>>>>>>>>>>>> If this ASRock crap (never ever again this brand!) doesn't work at
>>>>>>>>>>>> all - who cares, I intend to purchase new server grade hardware.
>>>>>>>>>>>> But the more puzzling issue is with the Fujitsu, which I consider
>>>>>>>>>>>> serious and from the behaviour the Fujitsu failure looks exactly
>>>>>>>>>>>> like the ASRock - Windows 7 works, RedHat 7.5 works (I assume I
>>>>>>>>>>>> can trust the Firmware settings when I disable CSM support, that
>>>>>>>>>>>> the Firmware will only EFI/UEFI capable loader? Or is there a
>>>>>>>>>>>> ghosty override somwhere to be expected?). Also on ASRock
>>>>>>>>>>>> disabling CSM should ensure not booting a dual-bootstrap-capable
>>>>>>>>>>>> system. This said, on the recent Fujitsu, it seems to boil down
>>>>>>>>>>>> to a FreeBSD UEFI-firmware interaction problem, while the ASRock
>>>>>>>>>>>> is still under suspicion to be broken by design.         
>>>>>>>>>>>>> GPT partitions can never start from disk absolute sector 1; this
>>>>>>>>>>>>> is because at sector 0 there is MBR (for compatibility), sector
>>>>>>>>>>>>> 1 is GPT table and then sectors 2-33 have GPT partition table
>>>>>>>>>>>>> entries, so the first possible data sector is 34 (absolute 34).
>>>>>>>>>>>>> Thats assuming 512B sectors. For details see UEFI 2.7 Chapter
>>>>>>>>>>>>> 5.3.1 page 131.          
>>>>>>>>>>>> Thanks for the explanation. That implies the installer did right,
>>>>>>>>>>>> gpart did also right and therefore there must be an issue with the
>>>>>>>>>>>> stuff located within the EFI partition?         
>>>>>>>>>>> Ok, so, it is not about UEFI bootcode but BIOS, and if we reach
>>>>>>>>>>> BIOS loader at all or not - that is, if the BIOS bootstrap is
>>>>>>>>>>> actually caring to read the MBR code and start it, since once the
>>>>>>>>>>> MBR code is started, it is all about our code.        
>>>>>>>>>> I'm getting confused a bit here. Do you mean by "BIOS" the CSM? or
>>>>>>>>>> do you mean that specific portion of the UEFI firmware, which looks
>>>>>>>>>> for the proper UEFI partition?
>>>>>>>>> BIOS as either native or CSM. Note that from boot code point of view
>>>>>>>>> the CSM boot *is* BIOS boot, we have no access to UEFI features.
>>>>>>>>>> The boxes in question, most notably the more recent Fujitsu Esprimo
>>>>>>>>>> Q956, refuse booting UEFI, even if properly setup (in terms of what
>>>>>>>>>> FreeBSD provides on recent CURRENT) is applied and CSM is switched
>>>>>>>>>> off in the firmware. Again: GPT partition scheme.
>>>>>>>>>> The system boots properly if a second partition of type
>>>>>>>>>> "freebsd-boot" is applied and bootcode is properly applied via
>>>>>>>>>> "gpart bootcode -b /boot/pmbr -p /boot/gptboot -i 2 ada0" (ada0 is
>>>>>>>>>> the device).         
>>>>>>>>>>> btw, you can try to validate the installed boot blocks by using
>>>>>>>>>>> recent enough loader (usb or iso) and then you can use from OK
>>>>>>>>>>> prompt:        
>>>>>>>>>> lsdev provides me with the follwoing informations (CSM enabled):
>>>>>>>>>> OK lsdev
>>>>>>>>>> disk devices:
>>>>>>>>>> 	disk0:		BIOS DRIVE C ...
>>>>>>>>>> 		disk0p1:	EFI
>>>>>>>>>> 		disk0p2:	FreeBSD BOOT
>>>>>>>>>> 		disk0p3:	FreeBSD SWAP
>>>>>>>>>> 		disk0p4:	FreeBSD ZFS
>>>>>>>>>> zfs devices:
>>>>>>>>>> 	zfs:zroot
>>>>>>>>>> OK chain disk0
>>>>>>>>>> open failed     (so for disk0p{1-4}.
>>>>>>>>>> OK chain zroot
>>>>>>>>>> failed to read disk (just for completeness)        
>>>>>>>>> chain command does use only device name (such as disk0: or
>>>>>>>>> disk0p2: ), but not zfs pool as device.  I just found I haven?t
>>>>>>>>> ported the code to read the file.      
>>>>>>>> ??
>>>>>>>>> the point for chain command test is to see if the normal read and
>>>>>>>>> execute would work, so in your case please try:
>>>>>>>>> chain disk0:      
>>>>>>>> As stated above, I did so, and the result is also mentioned above, I
>>>>>>>> always get "open failed".
>>>>>>>> This is the same for 
>>>>>>>> chain disk0
>>>>>>>> chain disk0p1
>>>>>>>> chain disk0p2
>>>>>>>> chain disk0p3
>>>>>>>> chain disk0p4
>>>>>>>> as already said. CSM is enabled in this case.      
>>>>>>> sigh? chain command does take device as argument, device must always
>>>>>>> end with colon?. in this case, the devil is in details:) as I wrote
>>>>>>> above, the command should be:
>>>>>>> chain disk0:
>>>>>>> The disk0p1: etc will only work when partition boot code was installed
>>>>>>> (which you most likely do not have - the only possible candidate could
>>>>>>> be FreeBSD ZFS partition).    
>>>>>> The command "chain disk0:" works as expected (CSM enabled, GPT partition
>>>>>> scheme, but with PMBR bootblock installed and freebsd-boot partition
>>>>>> conatining gptzfsboot installed.
>>>>>>>>> to read pmbr (512B) and execute it. The expected outcome would be
>>>>>>>>> that pmbr boot code would browse the GPT, read stage1 from disk0p2:
>>>>>>>>> and execute it; stage1 would detect FreeBSD ZFS from disk0p4: and
>>>>>>>>> load and execute /boot/loader. If that will happen, it means the
>>>>>>>>> boot code in our stages is just fine, but the bios (CSM) does not
>>>>>>>>> load pmbr?.  if thats true, it would mean that you either need to
>>>>>>>>> use UEFI boot or need to have some hack to fool the BIOS or just not
>>>>>>>>> use GPT on that machine with CSM.      
>>>>>>>> To make it clear here: The only way to boot this box is using CSM (as
>>>>>>>> it is the same with the ASRock boards mentioned earlier). But my
>>>>>>>> intention is to disable CSM and use a GPT/UEFI environment only! And
>>>>>>>> GPT/UEFI doesn't work with FreeBSD, neither with 12-CURRENT, nor
>>>>>>>> 11.2-RELENG.
>>>>>>>> It would be nice if this could be fixed. I'm more interested in the
>>>>>>>> fix on the recent Fujitsu device than the outdated ASRock crap, but
>>>>>>>> if the fix for the Fujitsu Firmware could fix older issues as a
>>>>>>>> byproduct, I'd appreciate that.
>>>>>>>> Kind regards,
>>>>>>> ok, somehow I have lost that part of the discussion. Well, you wrote
>>>>>>> that the UEFI boot fails when the first partition starts from sector
>>>>>>> 40 - does it mean you have boot when the partition will start from
>>>>>>> some other sector? I think, there is something else going on.    
>>>>>> Well, I simply try to describe what I "see" to make things
>>>>>> disambiguous. I'm not familiar with the deeper insights of disk layouts
>>>>>> on a binary level. So, you explained to me the reason, why ESP (EGI
>>>>>> partition) starts at block 40. I compared that to the FreeBSD USB flash
>>>>>> image FreeBSD provides, but this is another story since the image uses
>>>>>> MBR scheme as I figured out.
>>>>>>> What you can do is to see if that firmware will offer you EFI shell
>>>>>>> option, from there you can try to start the bootx64.efi manually and
>>>>>>> see what error you will get. However, the number 1 cause for failing
>>>>>>> to start the bootloader in UEFI is secure boot - we do not support it
>>>>>>> and secure boot must be switched off. 
>>>>>>> However, they seem to claim "The Secure Boot option is available in the
>>>>>>> UEFI/BIOS of most if not all ASRock boards. It is disabled by
>>>>>>> default.? 
>>>>>>> Still suggest to double check if thats really the case. Also, if the
>>>>>>> bootx64.efi start will fail and no messages are appearing on screen,
>>>>>>> then either there is something in firmware logs or you could get them
>>>>>>> from trying to start bootx64.efi from UEFI shell.    
>>>>>> Since I'm with this problem since 2014 and try from time to time, be
>>>>>> ausred that I tried every possible permutationof all reasonable
>>>>>> options, even those nonsense, to get rid of that problem.
>>>>>> I never had any problems with any other UEFI capable server/workstation
>>>>>> firmware so far booting FreeBSD off in UEFI-native (GPT partition
>>>>>> scheme, CSM disabled) so far - until now, when I ran into this Fujitsu
>>>>>> ESPRIMO Q956 with the most recent firmware (as of lat week, week 29 of
>>>>>> 2018) having the very same problems. 
>>>>>> I figured out something strange on the Fujitsu - and that is the same
>>>>>> with the ASRock boards.
>>>>>> We/I prepare some USB flash drives to boot a NanoBSD for a very small
>>>>>> appliance, but nevertheless, the USB flash device is booted on Fujitsu
>>>>>> servers with UEFI-only configurations. I assume at this point that
>>>>>> disabling on the most recent Fujitsu firmwares on reasonable "new"
>>>>>> hardware (not older than three years) will disable any(!) legacy BIOS
>>>>>> capabilities. The same is assumed for the Fujitus ESPRIMO Q956. I can
>>>>>> not speak for the ASRock A77 Pro4/m boards mentioned above/earlier,
>>>>>> they are from 2012/2013 and "quite old".
>>>>>> The NanoBSD image of ours doesn't have a "freebsd-boot" partition. The
>>>>>> partition scheme of the flash device is GPT. The layout looks like this:
>>>>>> gpart show -l da4    
>>>>>> =>      40  15425456  da4  GPT  (7.4G)    
>>>>>>      40      2000    1  efiboot0  (1.0M)
>>>>>>    2040   1453584    3  disk1a  (710M)
>>>>>> 1455624      4096    5  disk3  (2.0M)
>>>>>> 1459720  13965776       - free -  (6.7G)
>>>>>> I created the flash with md, gpart and dd straightforward, efiboot0 is
>>>>>> the ESP partition and its format/content is created via dd
>>>>>> if=/boot/boot1.efifat of=/dev/da4p1 - I presume this is very simple.
>>>>>> This USB flash device boots(!) successfully (UEFI!) on both the ASRock
>>>>>> boards and the Esprimo Q956!
>>>>>> But any SSD prepared the same way doesn't. Why? 
>>>>>> On the ASRock, I recall having fiddled around with HDD also for a while
>>>>>> conatining Windows 7/SP1 and FreeBSD. Windows7 booted, FreeBSD - I can't
>>>>>> remember. 
>>>>>> In the lack of proper hardware I'm unable to check whether USB-attached
>>>>>> HDD or SSD will boot or HDD will boot (just in case the local SATA has
>>>>>> problems booting UEFI and USB not).
>>>>>> Kind regards,
>>>>>> Oliver 
>>>>> Am. well. I think the suggestion to test out FAT32 is still good one to
>>>>> test. This is because it is known that some vendors do not support
>>>>> booting FAT12/FAT16 from HDD (the likely reason is that UEFI
>>>>> specification does not tell which FAT must be supported, and only hint
>>>>> about FAT12/FAT16 in context of removable devices).  
>>>> I prepared yesterday a GTP/ZFS-only 11.2-RELENG on the ESPRIMO Q956. It
>>>> took me a time to circumvent the installer and I had to install the
>>>> system manually. In that strain, I also "tried" to establish the ESP with
>>>> FAT32, as Allen Jude suggested earlier. I didn't find any ad hoc help how
>>>> to find out the format (FAT12/16/32) of the ESP, so I assume when using
>>>> 12-CURRENT's or 11.2-RELENG's installer with AUTO-ZFS and GPT (UEFI)
>>>> (only!) the resulting ESP is FAT12 or FAT16 (300mb by default). I also
>>>> assume, that when dd'ing the /boo/boot1.efifat image to a partition, the
>>>> format is FAT, but not FAT32. Therefore, I refomatted the manually
>>>> created ESP (using "gpart add -t efi ...") using "newfs_msdos -F 32 -b
>>>> xxx ...". I had to fiddle around a bit with option -b to fit in a proper
>>>> format to meet a 512mb ESP - I'm not sure whether this is the proper
>>>> option to deal with. When using the default and only -F32, the size of
>>>> the partition has to be 4G at least I assume. Having done that, I copied
>>>> the the content of boot1.efifat (mdconfig -t vnode ..., I guess we know
>>>> the drill ...) to the newly formatted ESP to /boot/efi/ ...
>>>> Having so far no knowledge of how to asure that the created ESP is FAT32,
>>>> I assume it is FAT32.
>>>> The result is negative on the ESPRIMO Q956. When disabling the CSM, the
>>>> box is not willing to boot from SSD with the ESP prepared as decribed.
>>>> So, a chance that this might still be due to a misconfiguration lies now
>>>> within the -b option of newfs_msdos - if the -b option is assumed the
>>>> proper option?
>>>> At the moment, the ESP of the Esprimo is subject to changes, if you wish,
>>>> but not in size, since it is limited to 512mb.
>>>> Thanks and kind regards,
>>>> Oliver  
>>> Yea, i was hoping fstyp command would report the FAT type, but it does not
>>> (request for feature?:)  
>> FYI, the file(1) command is very good at disecting a disk image to tell
>> you what the MBR looks like, and at looking at partitions if pointed at
>> them with the -s option.  It should be able to detect FAT12/16/32.
>> root_at_x230a:/home/ISO/x # file -s /dev/md2s1
>> /dev/md2s1: DOS/MBR boot sector, code offset 0x3c+2, OEM-ID "BSD4.4  ", root
>> entries 512, sectors 1600 (volumes <=32 MB) , sectors/FAT 5, sectors/track
>> 63, heads 1, serial number 0xbd4111ee, label: "EFISYS     ", FAT (12 bit),
>> followed by FAT
>>> However, the more annoying idea would be to install some OS which will boot
>>> with UEFI on this machine, then copy boot1.efi from freebsd to it (the
>>> default program the UEFI will load is ESP:EFI/boot/bootx64.efi  in case of
>>> UEFI64 and ESP:EFI/boot/bootia32.efi for EFI32. However, we do not support
>>> EFI32.
>>> note that boot1.efi alone will not do much but printing on screen how it
>>> will search for freebsd, but for the purpose of the test it would suffice -
>>> that would give us confirmed working ESP file system (since the other os
>>> would be able to boot) and then we can confirm if boot1.efi itself is OK.  
> Some new results.
> I installed RedHat 7.5 and inestigated the ESP.
> - The ESP starts at block 2048, while FreeBSD's ESP starts at block 40.
> - size is both 200mb if installed automatically. I forgit to investigate the
>  FAT format, but this might be unnecessary as shown further in this post.
> - RedHat's ESP contains ~ 10 MB of data in two folders, /efi/boot, /efi/redhat.
> copying FreeBSD's BOOTX64.efi over RedHat's doesn't change anything, also
> renaming /efi/boot/fbx64.efi of RedHat's installation. But renaming /efi/redhat
> renders RedHat to fail the boot process on the Fujitsu with the signs of the
> built-in testprogram as reported.
> I took the liberty and installed 11.2-RELENG again, ZFS only, UEFI boot only
> (CSM in firmware disabled, but there is still a gptzfsboot-prepared partition
> for later use, just for the record). Booting UEFI-only fails as reported. On
> this installation I copied the RedHat ESP completely into FreeBSD's ESP,
> renamed /efi/boot/BOOTX64.efi to /efi/boot/BOOTX64.efi.rh and copied FreeBSD's
> BOOTX64.efi to /efi/boot. 
> The Esprimo Q956 tries then to boot(!) RedHat's kernel. It seems, that
> the /efi/redhat folder of the ESP is important, if renamed, the whole process
> dies as I reported earlier.
> Still unanswered is the question: why is a NanoBSD prepared UEFI-only USB flash
> booting with CSM disabled (so asumingly UEFI only then) on both ASRock and
> Fujitsu (as described in more detail initially and earlier), while SSDs fail?
> Is there a difference? Since FreeBSD boots in UEFI mode from USB flash prepared
> as described (straight forward, 800k ESP starting at block 40, the boot1.efifat
> image dd'ed onto the partition, UFS partition following, no freebsd-boot
> partition or MBR/PMBR bootcode applied ever!), I think BOOTX64.EFI is
> technically all right. There must be then an issue with the SATA/SSD/HDD boot
> pathway.
> Hope I could provide some more details, sorry if it sounds confusing or way too
> long, but I try to descibe the situation as thorough as possible.

OK, this is already good hint. The thing with ESP is that there is “default” file system tree: EFI/BOOT/BOOT<sysname>.EFI, this is noted as default for *removable* media, fortunately it is usable for hard disks as well, or at least in most cases.

Now, for non-removable media, the UEFI does provide boot manager interface to define boot entries, and the fact that renaming efi/redhad directory did break the redhat boot, is very loud hint. And this means, this system is probably ignoring efi/boot tree on non-removable media and is expecting the boot manager entry to be created instead.

UEFI boot manager can be configured /usually/ manually via firmware menu, or by application, such as efibootmgr. The normal approach is to create efi/<vendorname> directory and to copy the application there, then create the boot manager configuration.

See UEFI specification v2.7, chapter 3 Boot Manager, page 79.

What is different in FreeBSD case is that the whole interface to program the UEFI Boot Manager is currently being developed, so either it has to be done manually or from some other OS (see <> for example, first hit from google:D).

Received on Thu Jul 26 2018 - 14:23:57 UTC

This archive was generated by hypermail 2.4.0 : Wed May 19 2021 - 11:41:17 UTC