[I was wrong: it is code elimination, not motion. volatile is
not a fix.]
On 2019-Jan-1, at 19:37, Mark Millard <marklmi at yahoo.com> wrote:
> [It looks to me like the assembler code has some code moved out of the
> loop that it should not be (by intent). It appears that calling
> tcmpset_32 does not prevent code motion to before the call and
> the variable involved was not declared volatile.]
>
> On 2019-Jan-1, at 18:43, Mark Millard <marklmi at yahoo.com> wrote:
>
>> The below showed up for poudiere-devel bulk getting stuck using one FreeBSD
>> cpu while building graphics/poppler-qt5 . This is not a kevent hang-up,
>> unlike
>> the last hang-up that I analyzed. I do not yet know how repeatable this is
>> but the original hang-up and the one experiment the below is from.
>>
>> From top:
>>
>> PID USERNAME THR PRI NICE SIZE RES SWAP STATE C TIME CPU
>> COMMAND
>> 12789 root 4 52 0 166M 33M 0 uwait 6 36:06 97.22%
>> /usr/local/bin/qemu-arm-static /usr/local/bin/cmake -E cmake_autogen
>> /wrkdirs/usr/ports/graphics/poppler-qt5/work/poppler-0
>>
>> Note: The vast margority of the 36:06 has been stuck in the uwait loop
>> involved.
>>
>> From ps -auxd:
>>
>> root 94075 0.0 0.0 12932 3552 1 S+ 10:42 0:01.21 |
>> `-- sh -e /usr/local/share/poudriere/bulk.sh -jFBSDFSSDjailArmV7 -w
>> graphics/poppler-qt5
>> root 1944 0.0 0.0 12932 3540 1 I+ 10:42 0:00.00 |
>> |-- sh -e /usr/local/share/poudriere/bulk.sh -jFBSDFSSDjailArmV7 -w
>> graphics/poppler-qt5
>> root 1957 0.0 0.0 12932 3556 1 I 10:42 0:00.04 |
>> |-- sh: poudriere[FBSDFSSDjailArmV7-default][01]: build_pkg
>> (poppler-qt5-0.72.0) (sh)
>> root 12328 0.0 0.0 12932 3548 1 I 10:49 0:00.00 |
>> | `-- sh: poudriere[FBSDFSSDjailArmV7-default][01]: build_pkg
>> (poppler-qt5-0.72.0) (sh)
>> root 12329 0.0 0.0 10328 1756 1 IJ 10:49 0:00.01 |
>> | `-- /usr/bin/make -C /usr/ports/graphics/poppler-qt5 stage
>> root 12350 0.0 0.0 9860 1248 1 IJ 10:49 0:00.00 |
>> | `-- /usr/bin/make -f Makefile
>> DESTDIR=/wrkdirs/usr/ports/graphics/poppler-qt5/work/stage install
>> root 12353 0.0 0.0 10236 1664 1 IJ 10:49 0:00.05 |
>> | `-- /nxb-bin/usr/bin/make -f CMakeFiles/Makefile2 qt5/all
>> root 12787 0.0 0.0 9856 1236 1 IJ 10:50 0:00.00 |
>> | `-- /nxb-bin/usr/bin/make -f
>> qt5/tests/CMakeFiles/check_qt5_attachments_autogen.dir/build.make qt5/test
>> root 12789 100.0 0.0 169868 33528 1 IJ 10:50 36:35.26 |
>> | `-- /usr/local/bin/qemu-arm-static /usr/local/bin/cmake
>> -E cmake_autogen /wrkdirs/usr/ports/graphics/
>> root 94423 0.0 0.0 12932 3484 1 S+ 10:42 0:12.91 |
>> `-- sh -e /usr/local/share/poudriere/bulk.sh -jFBSDFSSDjailArmV7 -w
>> graphics/poppler-qt5
>>
>>
>> (gdb) attach 12789
>> Attaching to process 12789
>> Reading symbols from
>> /usr/local/poudriere/data/.m/FBSDFSSDjailArmV7-default/01/usr/local/bin/qemu-arm-static...done.
>> [New LWP 101168 of process 12789]
>> [New LWP 101178 of process 12789]
>> [New LWP 101499 of process 12789]
>> [Switching to LWP 100304 of process 12789]
>> _umtx_op () at _umtx_op.S:3
>> 3 RSYSCALL(_umtx_op)
>> (gdb) info threads
>> Id Target Id Frame
>> * 1 LWP 100304 of process 12789 _umtx_op () at _umtx_op.S:3
>> 2 LWP 101168 of process 12789 _umtx_op_err () at
>> /usr/src/lib/libthr/arch/amd64/amd64/_umtx_op_err.S:37
>> 3 LWP 101178 of process 12789 _umtx_op () at _umtx_op.S:3
>> 4 LWP 101499 of process 12789 0x0000000060051c26 in atomic_cmpset_int
>> (dst=<optimized out>, expect=<optimized out>, src=536870912) at
>> /usr/include/machine/atomic.h:220
>> (gdb) thread 4
>> [Switching to thread 4 (LWP 101499 of process 12789)]
>> #0 0x0000000060051c26 in atomic_cmpset_int (dst=<optimized out>,
>> expect=<optimized out>, src=536870912) at /usr/include/machine/atomic.h:220
>> 220 ATOMIC_CMPSET(int);
>>
>> (gdb) bt
>> #0 0x0000000060051c26 in atomic_cmpset_int (dst=<optimized out>,
>> expect=<optimized out>, src=536870912) at /usr/include/machine/atomic.h:220
>> #1 tcmpset_32 (addr=<optimized out>, a=<optimized out>, b=536870912) at
>> /wrkdirs/usr/ports/emulators/qemu-user-static/work/qemu-bsd-user-4ef7d07/bsd-user/freebsd/os-thread.c:178
>> #2 freebsd_rw_unlock (target_addr=4108246528) at
>> /wrkdirs/usr/ports/emulators/qemu-user-static/work/qemu-bsd-user-4ef7d07/bsd-user/freebsd/os-thread.c:1264
>> #3 0x000000006004ab33 in do_freebsd__umtx_op (obj=<optimized out>,
>> op=536870912, val=<optimized out>, uaddr=<optimized out>,
>> target_time=<optimized out>)
>> at
>> /wrkdirs/usr/ports/emulators/qemu-user-static/work/qemu-bsd-user-4ef7d07/bsd-user/freebsd/os-thread.h:474
>> #4 0x0000000060041b83 in do_freebsd_syscall (cpu_env=0x86159b118, num=454,
>> arg1=<optimized out>, arg2=<optimized out>, arg3=<optimized out>, arg4=0,
>> arg5=0, arg6=-184411592, arg7=-199471616,
>> arg8=-1622188640) at
>> /wrkdirs/usr/ports/emulators/qemu-user-static/work/qemu-bsd-user-4ef7d07/bsd-user/syscall.c:1364
>> #5 0x00000000600392f0 in target_cpu_loop (env=0x86159b118) at
>> /wrkdirs/usr/ports/emulators/qemu-user-static/work/qemu-bsd-user-4ef7d07/bsd-user/arm/target_arch_cpu.h:207
>> #6 0x0000000060038c99 in cpu_loop (env=0xf4dede80) at
>> /wrkdirs/usr/ports/emulators/qemu-user-static/work/qemu-bsd-user-4ef7d07/bsd-user/main.c:121
>> #7 0x0000000060050c1a in new_freebsd_thread_start (arg=<optimized out>) at
>> /wrkdirs/usr/ports/emulators/qemu-user-static/work/qemu-bsd-user-4ef7d07/bsd-user/freebsd/os-thread.c:152
>> #8 0x00000000601ad5f6 in thread_start (curthread=0x861571300) at
>> /usr/src/lib/libthr/thread/thr_create.c:291
>> #9 0x0000000000000000 in ?? ()
>> Backtrace stopped: Cannot access memory at address 0x7fffdfbb3000
>>
>> 181 #define ATOMIC_CMPSET(TYPE) \
>> 182 static __inline int \
>> 183 atomic_cmpset_##TYPE(volatile u_##TYPE *dst, u_##TYPE expect, u_##TYPE
>> src) \
>> 184 { \
>> (gdb)
>> 185 u_char res; \
>> 186 \
>> 187 __asm __volatile( \
>> 188 " " MPLOCKED " " \
>> 189 " cmpxchg %3,%1 ; " \
>> 190 " sete %0 ; " \
>> 191 "# atomic_cmpset_" #TYPE " " \
>> 192 : "=q" (res), /* 0 */ \
>> 193 "+m" (*dst), /* 1 */ \
>> 194 "+a" (expect) /* 2 */ \
>> (gdb)
>> 195 : "r" (src) /* 3 */ \
>> 196 : "memory", "cc"); \
>> 197 return (res); \
>> 198 } \
>> . . .
>> 218 ATOMIC_CMPSET(char);
>> 219 ATOMIC_CMPSET(short);
>> 220 ATOMIC_CMPSET(int);
>> 221 ATOMIC_CMPSET(long);
>>
>> (gdb) up
>> #1 tcmpset_32 (addr=<optimized out>, a=<optimized out>, b=536870912) at
>> /wrkdirs/usr/ports/emulators/qemu-user-static/work/qemu-bsd-user-4ef7d07/bsd-user/freebsd/os-thread.c:178
>> 178 return atomic_cmpset_acq_32(addr, current, new);
>> (gdb) list
>> 173 static int tcmpset_32(uint32_t *addr, uint32_t a, uint32_t b)
>> 174 {
>> 175 uint32_t current = tswap32(a);
>> 176 uint32_t new = tswap32(b);
>> 177
>> 178 return atomic_cmpset_acq_32(addr, current, new);
>> 179 }
>>
>> (gdb) up
>> #2 freebsd_rw_unlock (target_addr=4108246528) at
>> /wrkdirs/usr/ports/emulators/qemu-user-static/work/qemu-bsd-user-4ef7d07/bsd-user/freebsd/os-thread.c:1264
>> 1264 if (!tcmpset_32(&target_urwlock->rw_state, state, (state -
>> 1))) {
>> (gdb) list
>> 1259 }
>> 1260 }
>> 1261 } else if (TARGET_URWLOCK_READER_COUNT(state) != 0) {
>> 1262 /* decrement reader count */
>> 1263 for (;;) {
>> 1264 if (!tcmpset_32(&target_urwlock->rw_state, state, (state -
>> 1))) {
>> 1265 if (TARGET_URWLOCK_READER_COUNT(state) == 0) {
>> 1266 unlock_user_struct(target_urwlock,
>> 1267 target_addr, 1);
>> 1268 return -TARGET_EPERM;
>> (gdb)
>> 1269 }
>> 1270 } else {
>> 1271 break;
>> 1272 }
>> 1273 }
>> 1274 } else {
>> 1275 unlock_user_struct(target_urwlock, target_addr, 1);
>> 1276 return -TARGET_EPERM;
>> 1277 }
>> 1278
>>
>> The above loop is where the code is stuck looping, using most of the
>> capacity of
>> one cpu (as FreeBSD classifies cpus). The value of state is stuck at:
>>
>> (gdb) print state
>> $2 = 536870913
>>
>> I'll note that the loop is inside qemu-arm-atatic, not in the cmake that it
>> is running.
>>
>>
>>
>> To show a little more context before the assembler for the routine that ends
>> up
>> containing that loop:
>>
>> (gdb) up
>> #3 0x000000006004ab33 in do_freebsd__umtx_op (obj=<optimized out>,
>> op=536870912, val=<optimized out>, uaddr=<optimized out>,
>> target_time=<optimized out>)
>> at
>> /wrkdirs/usr/ports/emulators/qemu-user-static/work/qemu-bsd-user-4ef7d07/bsd-user/freebsd/os-thread.h:474
>> 474 ret = freebsd_rw_unlock(obj);
>> (gdb) list
>> 469 ret = freebsd_rw_wrlock(obj, val, 0, NULL);
>> 470 }
>> 471 break;
>> 472
>> 473 case TARGET_UMTX_OP_RW_UNLOCK:
>> 474 ret = freebsd_rw_unlock(obj);
>> 475 break;
>> 476
>> 477 #if defined(__FreeBSD_version) && __FreeBSD_version > 900000
>> 478 #ifdef UMTX_OP_MUTEX_WAKE2
>>
>> (gdb) up
>> #4 0x0000000060041b83 in do_freebsd_syscall (cpu_env=0x86159b118, num=454,
>> arg1=<optimized out>, arg2=<optimized out>, arg3=<optimized out>, arg4=0,
>> arg5=0, arg6=-184411592, arg7=-199471616,
>> arg8=-1622188640) at
>> /wrkdirs/usr/ports/emulators/qemu-user-static/work/qemu-bsd-user-4ef7d07/bsd-user/syscall.c:1364
>> 1364 ret = do_freebsd__umtx_op(arg1, arg2, arg3, arg4, arg5);
>> (gdb) list
>> 1359 ret = do_freebsd__umtx_unlock(arg1);
>> 1360 break;
>> 1361 #endif
>> 1362
>> 1363 case TARGET_FREEBSD_NR__umtx_op: /* undocumented */
>> 1364 ret = do_freebsd__umtx_op(arg1, arg2, arg3, arg4, arg5);
>> 1365 break;
>> 1366
>> 1367 /*
>> 1368 * ioctl(2)
>>
>> As for the assembler level view for the routine containing the
>> loop (via inlining and such):
>>
>> Dump of assembler code for function freebsd_rw_unlock:
>> 0x0000000060051bd0 <+0>: push %rbp
>> 0x0000000060051bd1 <+1>: mov %rsp,%rbp
>> 0x0000000060051bd4 <+4>: push %rbx
>> 0x0000000060051bd5 <+5>: push %rax
>> 0x0000000060051bd6 <+6>: mov %edi,%ebx
>> 0x0000000060051bd8 <+8>: mov $0x20,%esi
>> 0x0000000060051bdd <+13>: mov $0x3,%edx
>> 0x0000000060051be2 <+18>: callq 0x60036b70 <page_check_range>
>> 0x0000000060051be7 <+23>: mov %eax,%ecx
>> 0x0000000060051be9 <+25>: mov $0xfffffff2,%eax
>> 0x0000000060051bee <+30>: test %ecx,%ecx
>> 0x0000000060051bf0 <+32>: jne 0x60051c50 <freebsd_rw_unlock+128>
>> 0x0000000060051bf2 <+34>: mov %ebx,%edi
>> 0x0000000060051bf4 <+36>: add 0x26a6355(%rip),%rdi # 0x626f7f50
>> <guest_base>
>> 0x0000000060051bfb <+43>: je 0x60051c50 <freebsd_rw_unlock+128>
>> 0x0000000060051bfd <+45>: mov (%rdi),%ecx
>> 0x0000000060051bff <+47>: mov 0x4(%rdi),%edx
>> 0x0000000060051c02 <+50>: test %ecx,%ecx
>> 0x0000000060051c04 <+52>: js 0x60051c30 <freebsd_rw_unlock+96>
>> 0x0000000060051c06 <+54>: test $0x1fffffff,%ecx
>> 0x0000000060051c0c <+60>: je 0x60051c4b <freebsd_rw_unlock+123>
>> 0x0000000060051c0e <+62>: lea -0x1(%rcx),%esi
>> 0x0000000060051c11 <+65>: nopw %cs:0x0(%rax,%rax,1)
>> 0x0000000060051c1b <+75>: nopl 0x0(%rax,%rax,1)
>> NOTE: loop is below:
>> 0x0000000060051c20 <+80>: mov %ecx,%eax
>> => 0x0000000060051c22 <+82>: lock cmpxchg %esi,(%rdi)
>> 0x0000000060051c26 <+86>: sete %bl
>> 0x0000000060051c29 <+89>: test %bl,%bl
>> 0x0000000060051c2b <+91>: je 0x60051c20 <freebsd_rw_unlock+80>
>> Note: loop is above
>> 0x0000000060051c2d <+93>: jmp 0x60051c57 <freebsd_rw_unlock+135>
>> 0x0000000060051c2f <+95>: nop
>> 0x0000000060051c30 <+96>: mov %ecx,%esi
>> 0x0000000060051c32 <+98>: and $0x7fffffff,%esi
>> 0x0000000060051c38 <+104>: mov %ecx,%eax
>> 0x0000000060051c3a <+106>: lock cmpxchg %esi,(%rdi)
>> 0x0000000060051c3e <+110>: sete %bl
>> 0x0000000060051c41 <+113>: test %bl,%bl
>> 0x0000000060051c43 <+115>: jne 0x60051c57 <freebsd_rw_unlock+135>
>> 0x0000000060051c45 <+117>: mov (%rdi),%ecx
>> 0x0000000060051c47 <+119>: test %ecx,%ecx
>> 0x0000000060051c49 <+121>: js 0x60051c30 <freebsd_rw_unlock+96>
>> 0x0000000060051c4b <+123>: mov $0xffffffff,%eax
>> 0x0000000060051c50 <+128>: add $0x8,%rsp
>> 0x0000000060051c54 <+132>: pop %rbx
>> 0x0000000060051c55 <+133>: pop %rbp
>> 0x0000000060051c56 <+134>: retq
>> 0x0000000060051c57 <+135>: test $0x2,%dl
>> 0x0000000060051c5a <+138>: jne 0x60051c70 <freebsd_rw_unlock+160>
>> 0x0000000060051c5c <+140>: test $0x40000000,%ecx
>> 0x0000000060051c62 <+146>: jne 0x60051c82 <freebsd_rw_unlock+178>
>> 0x0000000060051c64 <+148>: xor %eax,%eax
>> 0x0000000060051c66 <+150>: test $0x20000000,%ecx
>> 0x0000000060051c6c <+156>: je 0x60051c50 <freebsd_rw_unlock+128>
>> 0x0000000060051c6e <+158>: jmp 0x60051c89 <freebsd_rw_unlock+185>
>> 0x0000000060051c70 <+160>: test $0x20000000,%ecx
>> 0x0000000060051c76 <+166>: jne 0x60051c89 <freebsd_rw_unlock+185>
>> 0x0000000060051c78 <+168>: xor %eax,%eax
>> 0x0000000060051c7a <+170>: test $0x40000000,%ecx
>> 0x0000000060051c80 <+176>: je 0x60051c50 <freebsd_rw_unlock+128>
>> 0x0000000060051c82 <+178>: mov $0x1,%edx
>> 0x0000000060051c87 <+183>: jmp 0x60051c8e <freebsd_rw_unlock+190>
>> 0x0000000060051c89 <+185>: mov $0x7fffffff,%edx
>> 0x0000000060051c8e <+190>: mov $0x3,%esi
>> 0x0000000060051c93 <+195>: xor %ecx,%ecx
>> 0x0000000060051c95 <+197>: xor %r8d,%r8d
>> 0x0000000060051c98 <+200>: callq 0x601b5970 <_umtx_op>
>> 0x0000000060051c9d <+205>: mov %eax,%edi
>> 0x0000000060051c9f <+207>: add $0x8,%rsp
>> 0x0000000060051ca3 <+211>: pop %rbx
>> 0x0000000060051ca4 <+212>: pop %rbp
>> --Type <RET> for more, q to quit, c to continue without paging--
>> 0x0000000060051ca5 <+213>: jmpq 0x60040a40 <get_errno>
>> End of assembler dump.
>>
>> For just before the "lock cmpxchg %esi,(%rdi)" in the loop:
>>
>> rax 0x20000001 536870913
>> rbx 0xf4dede00 4108246528
>> rcx 0x20000001 536870913
>> rdx 0x2 2
>> rsi 0x20000000 536870912
>> rdi 0xf4dede80 4108246656
>>
>> (gdb) x 0xf4dede80
>> 0xf4dede80: 0x40000001
>>
>> 0x0000000060051c20 <+80>: mov %ecx,%eax
>> => 0x0000000060051c22 <+82>: lock cmpxchg %esi,(%rdi)
>> 0x0000000060051c26 <+86>: sete %bl
>> 0x0000000060051c29 <+89>: test %bl,%bl
>> 0x0000000060051c2b <+91>: je 0x60051c20 <freebsd_rw_unlock+80>
>>
>>
>> At this point I do not have interpretation of the details, not even
>> a comparison to the source code.
>
>
> . . . (misleading notes deleted) . . .
I was wrong about the code motion: in . . .
} else if (TARGET_URWLOCK_READER_COUNT(state) != 0) {
/* decrement reader count */
for (;;) {
if (!tcmpset_32(&target_urwlock->rw_state, state, (state - 1))) {
if (TARGET_URWLOCK_READER_COUNT(state) == 0) {
unlock_user_struct(target_urwlock,
target_addr, 1);
return -TARGET_EPERM;
}
} else {
break;
}
}
the outer test was leading to the inner-if being eliminated. While
"volatile state" stopped that from happening it did not fix the hang-up
in the loop. (There are likely other consequences to the use of memory
involved instead of just a register as well.)
===
Mark Millard
marklmi at yahoo.com
( dsl-only.net went
away in early 2018-Mar)
_______________________________________________
[email protected] mailing list
https://lists.freebsd.org/mailman/listinfo/freebsd-emulation
To unsubscribe, send any mail to "[email protected]"
