[mdb-discuss] spurious mdb stops in sigacthandler when debugging b117 Xorg ?
I'm trying to debug an Xorg mouse problem in b116 or newer, on 32-bit x86. The problem is that under unknown conditions the mouse pointer jumps to the upper left screen corner. ( http://www.opensolaris.org/jive/thread.jspa?threadID=105715tstart=0 ) Unfortunately, when I recompile Xorg from source and run that, the problem disappears. So that adding printfs to Xorg, or trying to compile Xorg binaries with certain changesets removed can't be used to narrow down the problem. For that reason I'm now trying to debug the problem with mdb using the original b116 / b117 Xorg 32-bit binary. The idea was to set a conditional mdb write access breakpoint on the mouse driver's y coordinate, that triggers only in case y was changed to a value of 0: # mdb -p `pgrep -x Xorg` Loading modules: [ ld.so.1 libproc.so.1 libnvpair.so.1 libuutil.so.1 libavl.so.1 ] GetPointerEvents+113/i GetPointerEvents+0x113: movl %ecx,-0x20(%ebp) :b :c mdb: stop at GetPointerEvents+0x113 mdb: target stopped at: GetPointerEvents+0x113: movl %ecx,-0x20(%ebp) eax=X 86f5998 At this point, %eax contains a pointer to the mouse driver's state structure, and offset 0xf8 in that structure is the current x coordinate, and offset 0xfc is the y coordinate. The following should set a conditional write access breakpoint, that stops only when y has been changed to 0: 0x86f5998+0xfc::wp -w -L 4 -c ,*(0x86f5998+0xfc)!=0:c :c mdb: stop at GetPointerEvents+0x113 mdb: target stopped at: GetPointerEvents+0x113: movl %ecx,-0x20(%ebp) $b ID S TA HT LM Description Action - - -- -- -- -- [ 1 ] + T 0 0 stop on SIGINT - [ 2 ] + T 0 0 stop on SIGQUIT - [ 3 ] + T 0 0 stop on SIGILL - [ 4 ] + T 0 0 stop on SIGTRAP - [ 5 ] + T 0 0 stop on SIGABRT - [ 6 ] + T 0 0 stop on SIGEMT - [ 7 ] + T 0 0 stop on SIGFPE - [ 8 ] + T 0 0 stop on SIGBUS - [ 9 ] + T 0 0 stop on SIGSEGV - [ 10] + T 0 0 stop on SIGSYS - [ 11] + T 0 0 stop on SIGXCPU - [ 12] + T 0 0 stop on SIGXFSZ - 13 + 2 0 stop at GetPointerEvents+0x113 - [ 14] + 2 0 stop on write of [0x86f5a94, 0x86f5a98) ,*(0x86f5998+0xfc ... ::delete 13 :c ... Xorg is running for some time ... mdb: target stopped at: libc_hwcap1.so.1`sigacthandler+1: movl %esp,%ebp 0x86f5998+0xf8/XX 0x86f5a90: 274 36d Now this is something I don't understand and looks like a mdb bug: After some time using the mouse in Xorg we do stop in mdb, but the y value didn't change to 0 ? Seems we didn't stop because of the write access breakpoint. But why did we stop? $C 08047758 libc_hwcap1.so.1`sigacthandler+1(86f5768) 08047788 xf86SigioReadInput+0x2f(15, 86f5768, 8047858, 80dfe15) 08047858 xf86SIGIO+0x1a9(16, 0, 8047920) 0804786c libc_hwcap1.so.1`__sighndlr+0xf(16, 0, 8047920, 80dfdf8) 080478dc libc_hwcap1.so.1`call_user_handler+0x2af(16) 0804790c libc_hwcap1.so.1`sigacthandler+0xdf(16, 0, 8047920) 08047b38 SecurityLookupIDByType+0xe(878cc00, 1600286, 3, 20) 08047b58 dixLookupGC+0x22(8047b7c, 1600286, 878cc00, 20) 08047b98 ProcChangeGC+0x37(878cc00, 38) 08047c28 Dispatch+0x44f(840a418, 840a41c, 8409930, 840a8b8, 840a8ec, 820694c) 08047d18 main+0x605(9, 8047d50, 8047d78, 8047d0c) 08047d44 _start+0x7d(9, 8047e20, 8047e32, 8047e35, 8047e3f, 8047e43) ::status debugging PID 950 (32-bit) file: /usr/X11/bin/i386/Xorg threading model: native threads status: stopped after a single-step $b ID S TA HT LM Description Action - - -- -- -- -- [ 1 ] + T 0 0 stop on SIGINT - [ 2 ] + T 0 0 stop on SIGQUIT - [ 3 ] + T 0 0 stop on SIGILL - [ 4 ] + T 0 0 stop on SIGTRAP - [ 5 ] + T 0 0 stop on SIGABRT - [ 6 ] + T 0 0 stop on SIGEMT - [ 7 ] + T 0 0 stop on SIGFPE - [ 8 ] + T 0 0 stop on SIGBUS - [ 9 ] + T 0 0 stop on SIGSEGV -
[mdb-discuss] kmdb amd64: parameters not shown in stack backtraces ?
I'm not exactly sure when it happened (somewhere between snv_98 and snv_100), but starting a couple of weeks ago kmdb appears to have lost the ability to display parameter values in stack backtraces on the amd64 platform. E.g. on an old build 82 or wiith an opensolaris snv_97 livecd, mdb -k / kmdb was able to show a stack backtrace like this: ::pgrep init | ::walk thread | ::findstack -v stack pointer for thread ff01c9373b20: ff0007b6fc00 [ ff0007b6fc00 _resume_from_idle+0xf1() ] ff0007b6fc40 swtch+0x17f() ff0007b6fcd0 cv_timedwait_sig+0x194(ff01c939663a, ff01c9396600, 719e2) ff0007b6fd60 cv_waituntil_sig+0xbb(ff01c939663a, ff01c9396600, ff0007b6fe80, 2) ff0007b6fe40 poll_common+0x3dd(806b500, 1, ff0007b6fe80, 0) ff0007b6fec0 pollsys+0xec(806b500, 1, 80475f8, 0) ff0007b6ff10 sys_syscall32+0x101() But with snv_100 or newer, all I get is this: ::pgrep init | ::walk thread | ::findstack -v stack pointer for thread ff02d38dbb20: ff000f917c50 [ ff000f917c50 _resume_from_idle+0xf1() ] ff000f917c80 swtch+0x160() ff000f917cf0 cv_timedwait_sig+0x1c0() ff000f917d60 cv_waituntil_sig+0xb0() ff000f917e40 poll_common+0x465() ff000f917ec0 pollsys+0xe8() ff000f917f10 sys_syscall32+0x101() Is it possible that the switch to the studio 12 compiler did break this kmdb feature? -- This message posted from opensolaris.org
[mdb-discuss] kmdb amd64: parameters not shown in stack backtraces ?
James C. McPherson wrote: J?rgen Keil wrote: ... starting a couple of weeks ago kmdb appears to have lost the ability to display parameter values in stack backtraces on the amd64 platform. Yes, and it was noticed starting with daily.0925. http://bugs.opensolaris.org/bugdatabase/view_bug.do?bug_id=6753543 6753543 is_argsaved() needs to be augmented for SS12 Oh, somehow I missed 6753543 when I tried to find an existing bug report for this problem... This is one of the fixes that's gone into the new compilers, Ok, I expected something like that, that the compiler is using a different code sequence to save the parameters... I'm not sure whether the patch (-08, iirc) is available on sunsolve yet. I don't know whether the tarball via opensolaris.org has been updated either, sorry. Hmm, as far as I can tell, a studio 12 compiler tarball is not yet available for download from the opensolaris web site. I'm using the studio 12 compiler that was included with one the SX:CE 9x release (build 91?). I just checked the web pages bwlow and they tell me that patch 124868-06 is needed for the (x86) C 5.9 compiler; the latest version of that patch from sunsolve is -07 (release two days ago) and according to the description text doesn't seem to fix this problem; -08 is not yet available http://www.opensolaris.org/os/downloads/on/ http://www.opensolaris.org/os/community/tools/sun_studio_tools/sun_studio_12_tools/ -- Sun Studio 12 Compilers and Tools for OpenSolaris Developers! Starting with Build 100, you must use the Sun Studio 12 compiler plus a specific set of patches to build ON. * The Sun Studio 12 compiler can be obtained from: http://developers.sun.com/sunstudio/downloads/index.jsp You must choose the package installer to download, and if you want to do a custom install, you must be sure to include PerfLib. * You must also download and apply the following critical patches: SPARC: 124867-07: Sun Studio 12: Patch for C 5.9 compiler 124863-07: Sun Studio 12: Patch for Sun C++ Compiler 124861-08: Sun Studio 12: Compiler Common patch for Sun C C++ F77 F95 x86: 124868-06: Sun Studio 12x86: Patch for C 5.9 compiler 124864-07: Sun Studio 12x86: Patch for Sun C++ Compiler 126498-09: Sun Studio 12_x86: Sun Compiler Common patch for x86 backend These can be obtained from: http://sunsolve.sun.com/show.do?target=patchpage A tarball containing an install image of Sun Studio 12 with appropriate patches is being created and will be posted soon. -- This message posted from opensolaris.org
[mdb-discuss] mdb wraps lines at 80 characters when input is redirected
echo ::findleaks -v -d | mdb -p $pid | c++filt However, it does not work: it seems that when mdb input is redirected, it wraps lines at 80 characters and C++ names are not demangled correctly. Does setting the output width with the $w command work? echo '300$w ; ::findleaks -v -d' | mdb ... -- This message posted from opensolaris.org
[mdb-discuss] build 38: mdb disassembler broken / segfaults
Is the following a known bug? On a 32-bit x86 platform, running on-20060404 bits, mdb segfaults, like this: $ mdb /bin/date main:b :r mdb: stop at main mdb: target stopped at: main: pushl %ebp puts:b :c mdb: stop at libc.so.1`puts mdb: target stopped at: libc.so.1`puts: pushl %ebp puts?i libc.so.1`puts: Segmentation Fault - core dumped $ pstack core core 'core' of 1033:mdb /bin/date ce5a44a2 memcpy (80456b4, ce74fac0, 0, 804558c, 1) + 22 ce66115a get_byte (80d3ac8) + 32 ce661545 dtrace_get_SIB (80d3adc, 8045618, 8045650, 804564c) + 2a ce6615c3 dtrace_get_modrm (80d3adc, 8045618, 8045650, 804564c) + 3a ce663e5e dtrace_disx86 (80d3adc, 2) + 204e ce6612cc dis_disassemble (80d3ac8, ce74fabf, 0, 8045780, 400) + 3d 08066853 libdisasm_ins2str (80d2fa8, 80dfeb0, fffd, 8045780, 400, ce74fabf) + 50 0806659e mdb_dis_ins2str (80d2fa8, 80dfeb0, fffd, 8045780, 400, ce74fabf) + 24 08069f0d fmt_instr (80dfeb0, fffd, ce74fabf, 0, 1) + 4e 0806a250 mdb_fmt_print (80dfeb0, fffd, ce74fabf, 0, 1, 69) + 55 080609e1 print_arglist (fffd, ce74fabf, 0, 1, 1, 80d69b8) + 1f8 08060aee print_common (fffd, 1, 1, 80d69b8, 805e45c, 10) + 9d 08060b28 cmd_print_object (ce74fabf, 1, 1, 80d69b8) + 16 0805e47a dcmd_invoke (80cc2e0, ce74fabf, 1, 1, 80d69b8, 0) + 53 0805e63e mdb_call_idcmd (80cc2e0, ce74fabf, 0, 1, 0, 1) + 111 0805e07b mdb_call (ce74fabf, 0, 1, 0, 1, 8046334) + 2d3 0808e423 yyparse (8046bbc, 8099450, 8046c7c, 8046300, 80b1b20, 0) + a4b 0805db9b mdb_run (8047280, 0, 0, 7273752f, 8047214, 80470ac) + 266 08075449 main (2, 80470f0, 80470fc) + f29 0805ae0e (2, 804727c, 8047280, 0, 804728a, 804729c) This message posted from opensolaris.org
[mdb-discuss] mdb :r command corrupts arguments for 64 bit x86 debug target
Can anyone reproduce this:
I'm running SX:CE b129 amd64, bfu'ed to current ON bits;
the same problem exists on OpenSolaris dev build b132.
I'm trying to debug a /usr/sbin/amd64/update_drv problem,
using mdb. Problem is that the debug target somehow
receives a :r quoted string argument slightly modified.
Here's an example that reproduces the issue with
a 64-bit echo test program:
% cat x.c
#include stdio.h
int
main(int argc, char **argv)
{
int i;
for (i = 1; argv[i] != NULL; i++)
printf(%s , argv[i]);
printf(\n);
return 0;
}
% cc -m64 -o x x.c
% ./x -d -i 'pci1814,601' rtls
-d -i pci1814,601 rtls
Ok, test program works as expected.
Now the same under mdb control:
% mdb ./x
:r -d -i 'pci1814,601' rtls
-d -i pci1144,601 rtls
mdb: target has terminated
Note how the pci1814,601 string
argument was modified when running
under mdb.
The problem does not happen with a
32-bit debug target.
--
This message posted from opensolaris.org
