Re: Trying to kexec on Allwinner A80

From: Russell King - ARM Linux admin
Date: Tue Apr 07 2020 - 08:22:51 EST

Next message: Trond Myklebust: "[GIT PULL] Please pull NFS client changes for 5.7"
Previous message: Mike Marshall: "Re: linux-next: Signed-off-by missing for commit in the orangefs tree"
In reply to: Corentin Labbe: "Re: Trying to kexec on Allwinner A80"
Next in thread: Corentin Labbe: "Re: Trying to kexec on Allwinner A80"
Messages sorted by: [ date ] [ thread ] [ subject ] [ author ]

On Tue, Apr 07, 2020 at 01:34:54PM +0200, Corentin Labbe wrote:
> On Tue, Apr 07, 2020 at 11:19:12AM +0100, Russell King - ARM Linux admin wrote:
> > On Tue, Apr 07, 2020 at 11:02:03AM +0100, Russell King - ARM Linux admin wrote:
> > > On Mon, Apr 06, 2020 at 10:27:20AM +0200, Corentin Labbe wrote:
> > > > Hello
> > > >
> > > > I am trying to add the last missing Allwinner Soc in kernelci: the A80.
> > > > But this SoC does not have any way to be used in kernelci, no USB nor network in uboot.
> > > > So I have tried to fake uboot since the kernel has network support and run the new kernel via kexec.
> > > >
> > > > But kexec 2.0.18 fail to work:
> > > > kexec --force /tmp/kernel --initrd /tmp/ramdisk --dtb /tmp/dtb --command-line='console=ttyS0,115200n8 root=/dev/ram0 earlycon=uart,mmio32,0x7000000 ip=dhcp'
> > >
> > > As I mentioned in my other reply, this apparent "kexec" command line
> > > does not seem to reflect the arguments you actually used to invoke
> > > the kexec output below.
> > >
> > > > syscall kexec_file_load not available.
> > >
> > > This message is produced by do_kexec_file_load(). This function is only
> > > invoked if the do_kexec_file_syscall flag in main() is set. This flag
> > > is only set if one of:
> > >
> > > --kexec-file-syscall
> > > --kexec-syscall
> > > --kexec-syscall-auto
> > > -s
> > > -c
> > > -a
> > >
> > > are provided on the kexec command line. Your command line above does
> > > not contain any of those arguments, so either the command line is not
> > > what you used, or you are using a patched kexec, or your compiler is
> > > grossly miscompiling kexec.
> > >
> > > > Try gzip decompression.
> > > > kernel: 0xb6535008 kernel_size: 0x853200
> > > > MEMORY RANGES
> > > > 0000000020000000-000000009fffffff (0)
> > >
> > > Then there's the debug output, which is only produced if the
> > > kexec_debug global is set, which in turn is only set if --debug or -d
> > > is supplied on the kexec command line - which again, your kexec
> > > command line does not contain this.
> > >
> > > > zImage header: 0x016f2818 0x00000000 0x00853200
> > > > zImage size 0x853200, file size 0x853200
> > > > zImage requires 0x00864200 bytes
> > > > offset 0x0000bae4 tag 0x5a534c4b size 8
> > > > Decompressed kernel sizes:
> > > > text+data 0x0158b3a0 bss 0x000632f0 total 0x015ee690
> > > > Resulting kernel space: 0x01def5a0
> > > > Kernel: address=0x20008000 size=0x01def5a0
> > > > Initrd: address=0x21df8000 size=0x0090b6fa
> > > > DT : address=0x22704000 size=0x00005c09
> > > > kexec_load: entry = 0x20008000 flags = 0x280000
> > > > nr_segments = 3
> > > > segment[0].buf = 0xb6535008
> > > > segment[0].bufsz = 0x853204
> > > > segment[0].mem = 0x20008000
> > > > segment[0].memsz = 0x854000
> > > > segment[1].buf = 0xb5c29008
> > > > segment[1].bufsz = 0x90b6fa
> > > > segment[1].mem = 0x21df8000
> > > > segment[1].memsz = 0x90c000
> > > > segment[2].buf = 0x4db50
> > > > segment[2].bufsz = 0x5c09
> > > > segm[ 71.039126] kexec_core: Starting new kernel
> > > > ent[2].mem = 0[ 71.044712] Disabling non-boot CPUs ...
> > > > x22704000
> > > > segment[2].memsz = 0x6000
> > > > [ 71.489070] Bye!
> > > >
> > > >
> > > > I have tried also kexec-2.0.20
> > > > Try gzip decompression.
> > > > zImage header: 0x00000000 0x000019b4 0x00001000
> > > > zImage requires 0x008641c0 bytes
> > > > Could not find a free area of memory of 0x86c1c0 bytes...
> > > > Cannot load /tmp/kernel
> > >
> > > kexec 2.0.20 doesn't appear to have changed anything to do with how
> > > allocations are done. The above output looks even more strange and
> > > confusing. "zImage header" is produced by debugging prints, which
> > > imply that kexec_debug was set.
> > >
> > > However, the "MEMORY RANGES" output is missing - this has not gone
> > > away with kexec 2.0.20, it's still there, and works for me (I've
> > > just built and tested kexec 2.0.20).
> > >
> > > Also, the values on the "zImage header" line are completely messed
> > > up; the first should be the zImage magic value and it is not - that
> > > suggests that the file being loaded is not a zImage file, or is
> > > corrupted.
> >
> > Under a VM (the kernel doesn't have kexec support - but that's not a
> > problem, because the initial loading stages are the relevant parts
> > which all happen in userspace):
> >
> > # build/sbin/kexec --version
> > kexec-tools 2.0.20
> > # build/sbin/kexec --debug --load virt-vmlinuz-5.4.0+
> > Try gzip decompression.
> > kernel: 0xb6a6c008 kernel_size: 0x407358
> > MEMORY RANGES
> > 0000000040000000-000000007fffffff (0)
> > zImage header: 0x016f2818 0x00000000 0x00407358
> > zImage size 0x407358, file size 0x407358
> > zImage requires 0x00418358 bytes
> > offset 0x00007178 tag 0x5a534c4b size 12
> > Decompressed kernel sizes:
> > text+data 0x00c2ed24 bss 0x000319ec total 0x00c60710
> > Resulting kernel space: 0x0104707c
> > Kernel: address=0x40008000 size=0x0104707c
> > DT : address=0x41051000 size=0x00100000
> > kexec_load: entry = 0x40008000 flags = 0x280000
> > nr_segments = 2
> > segment[0].buf = 0xb6a6c008
> > segment[0].bufsz = 0x40735c
> > segment[0].mem = 0x40008000
> > segment[0].memsz = 0x408000
> > segment[1].buf = 0xb696b008
> > segment[1].bufsz = 0x100000
> > segment[1].mem = 0x41051000
> > segment[1].memsz = 0x100000
> > kexec_load failed: Function not implemented
> > entry = 0x40008000 flags = 0x280000
> > nr_segments = 2
> > segment[0].buf = 0xb6aa0008
> > segment[0].bufsz = 0x40735c
> > segment[0].mem = 0x40008000
> > segment[0].memsz = 0x408000
> > segment[1].buf = 0xb699f008
> > segment[1].bufsz = 0x100000
> > segment[1].mem = 0x41051000
> > segment[1].memsz = 0x100000
> > #
> >
> > On a SolidRun Hummingboard2 (iMX6 based, which has kexec support, same
> > kexec binary, first without an initrd, and then with):
> >
> > # build/sbin/kexec --version
> > kexec-tools 2.0.20
> > # build/sbin/kexec --debug --load multi-vmlinuz-5.6.0+
> > Try gzip decompression.
> > kernel: 0xb6763008 kernel_size: 0x7273a8
> > MEMORY RANGES
> > 0000000010000000-000000004fffffff (0)
> > zImage header: 0x016f2818 0x00000000 0x007273a8
> > zImage size 0x7273a8, file size 0x7273a8
> > zImage requires 0x007383a8 bytes
> > offset 0x00004da8 tag 0x5a534c4b size 12
> > Decompressed kernel sizes:
> > text+data 0x00dbedb8 bss 0x007489f8 total 0x015077b0
> > Resulting kernel space: 0x015077b0
> > Kernel: address=0x10008000 size=0x015077b0
> > DT : address=0x11511000 size=0x0000b000
> > kexec_load: entry = 0x10008000 flags = 0x280000
> > nr_segments = 2
> > segment[0].buf = 0xb6763008
> > segment[0].bufsz = 0x7273ac
> > segment[0].mem = 0x10008000
> > segment[0].memsz = 0x728000
> > segment[1].buf = 0x1207cb0
> > segment[1].bufsz = 0xb000
> > segment[1].mem = 0x11511000
> > segment[1].memsz = 0xb000
> > # build/sbin/kexec --unload
> > # build/sbin/kexec --debug --load multi-vmlinuz-5.6.0+ --initrd /boot/initrd.img-5.4.0+
> > Try gzip decompression.
> > kernel: 0xb65d8008 kernel_size: 0x7273a8
> > MEMORY RANGES
> > 0000000010000000-000000004fffffff (0)
> > zImage header: 0x016f2818 0x00000000 0x007273a8
> > zImage size 0x7273a8, file size 0x7273a8
> > zImage requires 0x007383a8 bytes
> > offset 0x00004da8 tag 0x5a534c4b size 12
> > Decompressed kernel sizes:
> > text+data 0x00dbedb8 bss 0x007489f8 total 0x015077b0
> > Resulting kernel space: 0x015077b0
> > Kernel: address=0x10008000 size=0x015077b0
> > Initrd: address=0x11510000 size=0x0053f46a
> > DT : address=0x11a50000 size=0x0000b044
> > kexec_load: entry = 0x10008000 flags = 0x280000
> > nr_segments = 3
> > segment[0].buf = 0xb65d8008
> > segment[0].bufsz = 0x7273ac
> > segment[0].mem = 0x10008000
> > segment[0].memsz = 0x728000
> > segment[1].buf = 0xb6098008
> > segment[1].bufsz = 0x53f46a
> > segment[1].mem = 0x11510000
> > segment[1].memsz = 0x540000
> > segment[2].buf = 0x993cf0
> > segment[2].bufsz = 0xb044
> > segment[2].mem = 0x11a50000
> > segment[2].memsz = 0xc000
> >
> > On clearfog (Armada 388):
> >
> > # build/sbin/kexec --debug --load multi-vmlinuz-5.6.0+
> > Try gzip decompression.
> > kernel: 0xb6745008 kernel_size: 0x7273a8
> > MEMORY RANGES
> > 0000000000000000-000000003fffffff (0)
> > zImage header: 0x016f2818 0x00000000 0x007273a8
> > zImage size 0x7273a8, file size 0x7273a8
> > zImage requires 0x007383a8 bytes
> > offset 0x00004da8 tag 0x5a534c4b size 12
> > Decompressed kernel sizes:
> > text+data 0x00dbedb8 bss 0x007489f8 total 0x015077b0
> > Resulting kernel space: 0x015077b0
> > Kernel: address=0x00008000 size=0x015077b0
> > DT : address=0x01511000 size=0x00007be3
> > kexec_load: entry = 0x8000 flags = 0x280000
> > nr_segments = 2
> > segment[0].buf = 0xb6745008
> > segment[0].bufsz = 0x7273ac
> > segment[0].mem = 0x8000
> > segment[0].memsz = 0x728000
> > segment[1].buf = 0x1be7f68
> > segment[1].bufsz = 0x7be3
> > segment[1].mem = 0x1511000
> > segment[1].memsz = 0x8000
> >
> > All appears to work fine.
> >
>
> I have found a part of my problem, kexec-tool seems to always use the OPT_KEXEC_FILE_SYSCALL.
> Even if I set --kexec-syscall.

You mention below that you're using buildroot. Does buildroot have
patches on top of the reference kexec-tools codebase as found at
git://git.kernel.org/pub/scm/utils/kernel/kexec/kexec-tools.git ?

Looking at the 2.0.20 code, there's no way for that to be happening;
the only way that _could_ happen is if getopt_long() always returns
's' when we get to the last argument, which is highly unlikely.

> On my early tries I got "syscall kexec_file_load not available."
> When I did a full rebuild of my buildroot to go back to kexec-tool 2.0.18, this syscall become availlable.

32-bit kernels have no support for kexec_file_load(), although the
syscall is reserved.

Irrespective of that, presence of the syscall number allocation is
not sufficient to tell whether the syscall is implemented by the
kernel; that is controlled by the kernel's CONFIG_KEXEC_FILE symbol.
This symbol is never offered for 32-bit ARM kernels.

> The get_memory_ranges seems to be called only from my_load() which is executed when !do_kexec_file_syscall.
>
> Since kexec-tool always set do_kexec_file_syscall, it is never called.
> I have added a print for each syscal option, and it seems that OPT_KEXEC_SYSCALL_AUTO is called after the handling of OPT_KEXEC_SYSCALL.
>
> So I have hack to always set do_kexec_file_syscall=0

It sounds like buildroot is patching kexec-tools in ways that break
32-bit ARM, and it sounds like this needs to be reported as a bug to
buildroot.

That said, I think there's a weakness in the ARM support in
kexec-tools, which this patch should address for 2.0.20. Things have
changed in the merged patches since 2.0.20, so the return needs to be
EFALLBACK not ENOSYS, so please pay attention to exactly which version
of kexec-tools you have (whether it's the 2.0.20 tagged version or has
further patches from kexec-tools on top.)

diff --git a/kexec/arch/arm/kexec-zImage-arm.c b/kexec/arch/arm/kexec-zImage-arm.c
index 9a67b8a4db98..ca4e35382ea1 100644
--- a/kexec/arch/arm/kexec-zImage-arm.c
+++ b/kexec/arch/arm/kexec-zImage-arm.c
@@ -421,6 +421,12 @@ int zImage_arm_load(int argc, char **argv, const char *buf, off_t len,
};
static const char short_options[] = KEXEC_ARCH_OPT_STR "a:r:";

+ /* We do not support file mode */
+ if (info->file_mode) {
+ fprintf(stderr, "syscall kexec_file_load not implemented.\n");
+ return -ENOSYS;
+ }
+
/*
* Parse the command line arguments
*/

> So now my test go further, but the final kernel crash.
> DEBUG: bootz: run kexec with /tmp/kernel --debug --kexec-syscall --force --initrd /tmp/ramdisk --dtb /tmp/dtb --command-line='console=ttyS0,115200n8 root=/dev/ram0 earlycon=uart,mmio32,0x7000000 ip=dhcp'
> Set DEBUG!
> main:1417 OPT_KEXEC_SYSCALL
> main:1422 OPT_KEXEC_SYSCALL_AUTO
> arch_process_options:119
> main:1500
> main:1517 res=0 do_load=1
> main:1519 res=0 do_kexec_file_syscall=0
> my_load:713
> Try gzip decompression.
> kernel: 0xb693b008 kernel_size: 0x443ac0
> get_memory_ranges:36
> MEMORY RANGES
> 0000000020000000-000000009fffffff (0)
> DEBUG: my_load:737 memory_ranges=1
> zImage_arm_load:423
> zImage header: 0x016f2818 0x00000000 0x00443ac0
> zImage size 0x443ac0, file size 0x443ac0
> zImage requires 0x00454ac0 bytes
> offset 0x00006678 tag 0x5a534c4b size 8
> Decompressed kernel sizes:
> text+data 0x00b78080 bss 0x0003d3c0 total 0x00bb5440
> Resulting kernel space: 0x00fccb40
> DEBUG: locate_hole:237 memory_ranges=1
> Check 0/1 536870912 0 type=-1610612737
> Check 0/1 536870912 0
> Kernel: address=0x20008000 size=0x00fccb40
> DEBUG: locate_hole:237 memory_ranges=1
> Check 0/1 536870912 0 type=-1610612737
> Check 0/1 536870912 0
> Initrd: address=0x20fd5000 size=0x0167213d
> DEBUG: locate_hole:237 memory_ranges=1
> Check 0/1 536870912 0 type=-1610612737
> Check 0/2 536870912 0
> Check 1/2 577011712 0
> DT : address=0x22648000 size=0x00006043
> kexec_load: entry = 0x20008000 flags = 0x280000
> nr_segments = 3
> segment[0].buf = 0xb693b008
> segment[0].bufsz = 0x443ac4
> segment[0].mem = 0x20008000
> segment[0].memsz = 0x444000
> segment[1].buf = 0xb52c8008
> segment[1].bufsz = 0x167213d
> segment[1].mem = 0x20fd5000
> segment[1].memsz = 0x1673000
> segment[2].buf = 0x4ef88
> segment[2].bufsz = 0x6043
> segment[2].mem = 0x2[ 31.265096] sun7i-dwmac 830000.ethernet eth0: Link is Down
> 2648000
> segment[2].memsz = 0x7000
> main:1568 res=0
> main:1582 res=0 do_exec=1
> [ 31.277297] kexec_core: Starting new kernel
> [ 31.282700] Disabling non-boot CPUs ...
> [ 31.692085] Bye!
> C:0x200080C0-0x2044BAC0->0x20B80B00-0x20FC4500

This line shows that we entered the new kernel, which was located
(correctly) at 0x200080C0-0x2044BAC0, and it has decided to relocate
itself to 0x20B80B00-0x20FC4500.

The lack of further output suggests that the decompressor itself
wasn't able to run after being relocated higher in memory.

It may be worth booting the same kernel image and see what the C:
line comes out as there.

--
RMK's Patch system: https://www.armlinux.org.uk/developer/patches/
FTTC broadband for 0.8mile line in suburbia: sync at 10.2Mbps down 587kbps up

Next message: Trond Myklebust: "[GIT PULL] Please pull NFS client changes for 5.7"
Previous message: Mike Marshall: "Re: linux-next: Signed-off-by missing for commit in the orangefs tree"
In reply to: Corentin Labbe: "Re: Trying to kexec on Allwinner A80"
Next in thread: Corentin Labbe: "Re: Trying to kexec on Allwinner A80"
Messages sorted by: [ date ] [ thread ] [ subject ] [ author ]