Problems with bhyve's kgdb support and loadable modules

[John Baldwin]

On Saturday, May 03, 2014 11:01:00 pm Peter Grehan wrote:
> Hi Kurt,
> 
> > Is there any reasonable tutorial for using kgdb with the
> > bvmdebug kernel option?
> 
>   Not really (https://wiki.freebsd.org/BHyVe/gdb)
> 
>   In any event, 9.2 doesn't have bvmdebug, though it would be a simple 
> backport. It's also not strictly required - you can use the serial port 
> same as on h/w; see below.
> 
> > A couple of folks I know have run into issues trying to
> > debug a FreeBSD stable/9 kernel from their bhyve
> > hosting machine (running stable/10).
> >
> > In particular, the loadable modules that are in use in
> > the stable/9 kernel are being "troublesome" to get to
> > the point where source-level debugging actually works.
> >
> > Even a pointer to a couple of "worked" examples might be
> > useful.
> >
> > I've read this:
> > http://people.freebsd.org/~jhb/papers/bsdcan/2008/article/node4.html
> > but not all the techniques in there appear to work properly.
> 
>   I tried to repro this with some success from a host running CURRENT.
> 
>   Firstly, I installed a 9.2 VM, with source. I edited GENERIC and added 
> options DDB and GDB, and reinstalled the kernel.
> 
>   The disk was then copied, and mdconfig'd/mounted on the host to 
> provide access to the just-buit 9.2 kernel syms and sources.
> 
>   com2 was set up as a debug port by dropping to the bhyveload prompt and
> 
>     hint.uart.1.flags="0x80"
> 
>   (this could also have been done in the guest's /boot/loader.conf)
> 
>   com2 was then set up in the bhyve command line to point to an nmdm device
> 
>    -l com2,/dev/nmdm0A
> 
>   The guest probed uart1 as a debug port:
> 
> uart1: <16550 or compatible> port 0x2f8-0x2ff irq 3 flags 0x80 on acpi0
> ioapic0: routing intpin 3 (ISA IRQ 3) to lapic 0 vector 64
> uart1: fast interrupt
> uart1: debug port (9600,n,8,1)
> 
>   After booting, I loaded the the tap device in the guest to provide a 
> kld for kgdb to examine:
> 
> root at fbsd9-2:~ # kldload if_tap
> root at fbsd9-2:~ # kldstat
> Id Refs Address            Size     Name
>   1    3 0xffffffff80200000 15f92d8  kernel
>   2    1 0xffffffff81a12000 59e9     if_tap.ko
> 
>   Now time to try kgdb:
> 
> root at fbsd9-2:~ # sysctl debug.kdb.enter=1
> debug.kdb.enter: 0KDB: enter: sysctl debug.kdb.enter
> [ thread pid 577 tid 100054 ]
> Stopped at      kdb_enter+0x3b: movq    $0,0xaf0362(%rip)
> db> gdb
> (ctrl-c will return control to ddb)
> Switching to gdb back-end
> Switching to gdb back-end
> 
>   In another window, cd'd to the mounted copy of the 9.2 disk:
> 
> kgdb -r /dev/nmdm0B kernel.debug
> ...
> This GDB was configured as "amd64-marcel-freebsd"...Switching to remote 
> protocol
> kdb_enter (why=0xffffffff80fbf671 "sysctl", msg=0x80 <Address 0x80 out 
> of bounds>) at ../../../kern/subr_kdb.c:441
> 441			kdb_why = KDB_WHY_UNSET;
> 
> (kgdb)
> 
>   After some experimentation, I found the way to get the correct symbols 
> for the kld was to manually specify it:
> 
> (kgdb) add-kld /mnt/boot/kernel/if_tap.ko
> add symbol table from file "/mnt/boot/kernel/if_tap.ko.symbols" at
> 	.text_addr = 0xffffffff81a12000
> 	.rodata.str1.8_addr = 0xffffffff81a13b50
> 	.rodata.str1.1_addr = 0xffffffff81a13ddb
> 	set_sysinit_set_addr = 0xffffffff81a13f68
> 	set_modmetadata_set_addr = 0xffffffff81a13f80
> 	set_sysctl_set_addr = 0xffffffff81a13f90
> 	set_sysuninit_set_addr = 0xffffffff81a13fc0
> 	.data_addr = 0xffffffff81a13fe0
> 	.bss_addr = 0xffffffff81a14620
> (y or n) y
> 
>   However, 'info sharedlibrary' didn't seem to reflect this:
> 
> (kgdb) info sharedlibrary
>  From                To                  Syms Read   Shared Object Library
> 0xffffffff81a12000  0xffffffff81a13c04  No 
> /boot/kernel/if_tap.ko.symbols
> 
>   Might have been a bug there, since I was able to successfully set 
> breakpoints in if_tap routines and have them trigger.
> 
>   I did have some trouble getting the source path set up correctly, but 
> never fully investigated that: seems like gdb is quite rich in that area 
> and it should be possible to get sorted.

I think you want to set solib-absolute-prefix to '/mnt'.  That will cause
kgdb to prepend '/mnt' to the paths it finds for each kld.  Mounting /mnt
directly on the host is a bit risky though.  What I've done with bhyve is
to build a test kernel on the host and install it to a directory on the
host, then use the '-h' flag to bhyveload to export that directory to the 
loader and explicitly load the test kernel from the 'host0:' filesystem.

I've even been able to build a cross kgdb-i386 to debug a cross-built
kernel for an i386 guest and have this work ok.

-- 
John Baldwin