Update: If I call sbrk() or brk() in my program after my program is restarted the problem goes away. With this extra call, once the (now anonymous) memory region is used up my program will begin to use a new heap region instead of crashing with a segmentation fault.
On Fri, Oct 2, 2020 at 11:25 AM Christine Andress < [email protected]> wrote: > Hello, > > I am encountering an issue using DMTCP to checkpoint/restart my program. > Everything seems good when first restarted, until my program has consumed > the memory on what was the pre-checkpointed heap, and the next call to > new() or malloc() which needs to expand the heap causes the program to > crash. > > For example, immediately prior to checkpointing the /proc/<pid>/maps shows > the heap at: > 03606000-037f8000 rwxp 00000000 00:00 0 > [heap] > > On restore (on dmtcp_checkpoint() == DMTCP_AFTER_RESTART), the same > mapping entry appears, but without the “[heap]” name: > 03606000-037f8000 rwxp 00000000 00:00 0 > > > And once a “new string(“Hello”))” is called where the expected memory > returned spans the address 0x37f8000, my program crashes with a > segmentation fault. > > > #0 sysmalloc (nb=nb@entry=48, av=av@entry=0x7fd76d2b0c40 <main_arena>) > at malloc.c:2723 > #1 0x00007fd76d1782c9 in _int_malloc (av=av@entry=0x7fd76d2b0c40 > <main_arena>, bytes=bytes@entry=32) at malloc.c:4133 > #2 0x00007fd76d17956a in __GI___libc_malloc (bytes=32) at malloc.c:3057 > #3 0x00007fd7719b8208 in malloc (size=32) at alloc/mallocwrappers.cpp:41 > #4 0x00007fd76d4e7fd8 in operator new(unsigned long) () from > /usr/lib/x86_64-linux-gnu/libstdc++.so.6 > > #5 0x00000000014fe6b8 in doMemoryStuff (msg=0x7fd76a5b00b0 "Hello") at > src/mucking/mystuff.cpp > > Looking at DMTCP’s code path when restarting my checkpoint, I find that > src/mtcp/mtcp_restart.c’s restore_brk(..) function exits without trying to > call mtcp_sys_brk() because saved_brk < current_brk. The later attempt to > call mtcp_sys_brk() in restorememoryareas() does not fail, in that it does > not return error code -1, but if I insert a print statement of the > subsequent result of mtcp_sys_brk(NULL), it shows that the value of the > break has not changed to the saved_brk value. Then in > read_one_memory_area() when it is trying to read in the “[heap]” area the > warning > WARNING: break (0x555555565000) not equal to end of heap (0x37f8000) > is emitted. > > Checkpointing with either dmtcp_launch options --disable-alloc-plugin or > --disable-all-plugins does not seem to change the behaviour. > > > Do you have any suggestions on how I could solve or circumvent this issue? > > dmtcp version: 3.0.0 master branch at > 114f5d59961b6b3e178629b961ce58be17282403 > uname info: Linux machine-1 4.19.0-9-cloud-amd64 #1 SMP Debian > 4.19.118-2+deb10u1 (2020-06-07) x86_64 GNU/Linux > > > > Thanks, > > Christine > >
_______________________________________________ Dmtcp-forum mailing list [email protected] https://lists.sourceforge.net/lists/listinfo/dmtcp-forum
