core - process core file
The operating system writes out a core file for a process when the process is terminated due to receiving certain signals. A core file is a disk copy of the contents of the process address space at the time the process received the signal, along with additional information about the state of the process. This information can be consumed by a debugger. Core files can also be generated by applying the gcore(1) utility to a running process.
Typically, core files are produced following abnormal termination of a process resulting from a bug in the corresponding application. Whatever the cause, the core file itself provides invaluable information to the programmer or support engineer to aid in diagnosing the problem. The core file can be inspected using a debugger such as dbx(1) or mdb(1) or by applying one of the proc(1) tools.
The operating system attempts to create up to two core files for each abnormally terminating process, using a global core file name pattern and a per-process core file name pattern. These patterns are expanded to determine the pathname of the resulting core files, and can be configured by coreadm(8). By default, the global core file pattern is disabled and not used, and the per-process core file pattern is set to core. Therefore, by default, the operating system attempts to create a core file named core in the process's current working directory.
A process terminates and produces a core file whenever it receives one of the signals whose default disposition is to cause a core dump. The list of signals that result in generating a core file is shown in signal.h(3HEAD). Therefore, a process might not produce a core file if it has blocked or modified the behavior of the corresponding signal. Additionally, no core dump can be created under the following conditions:
If normal file and directory access permissions prevent the creation or modification of the per-process core file pathname by the current process user and group ID. This test does not apply to the global core file pathname because, regardless of the UID of the process dumping core, the attempt to write the global core file is made as the superuser.
Core files owned by the user nobody will not be produced. For example, core files generated for the superuser on an NFS directory are owned by nobody and are, therefore, not written.
If the core file pattern expands to a pathname that contains intermediate directory components that do not exist. For example, if the global pattern is set to /var/core/%n/core.%p, and no directory /var/core/`uname -n` has been created, no global core files are produced.
If the destination directory is part of a filesystem that is mounted read-only.
If the core file name already exists in the destination directory and is not a regular file (that is, is a symlink, block or character special-file, and so forth).
If the kernel cannot open the destination file O_EXCL, which can occur if same file is being created by another process simultaneously.
If the process's effective user ID is different from its real user ID or if its effective group ID is different from its real group ID. Similarly, set-user-ID and set-group-ID programs do not produce core files as this could potentially compromise system security. These processes can be explicitly granted permission to produce core files using coreadm(8), at the risk of exposing secure information.
The core file contains all the process information pertinent to debugging: contents of hardware registers, process status, and process data. The format of a core file is object file specific.
For ELF executable programs (see a.out(5)), the core file generated is also an ELF file, containing ELF program and file headers. The e_type field in the file header has type ET_CORE. The program header contains an entry for every segment that was part of the process address space, including shared library segments. The contents of the mappings specified by coreadm(8) are also part of the core image. Each program header has its p_memsz field set to the size of the mapping. The program headers that represent mappings whose data is included in the core file have their p_filesz field set the same as p_memsz, otherwise p_filesz is zero.
A mapping's data can be excluded due to the core file content settings (see coreadm(8)), or due to some failure. If the data is excluded because of a failure, the program header entry will have the PF_SUNW_FAILURE flag set in its p_flags field.
The program headers of an ELF core file include an entry for a NOTE segment, containing several note entries as described below. The note entry header and core file note type (n_type) definitions are contained in <sys/elf.h>.
Prior to Oracle Solaris 11.4, a core file contained two NOTE sections, the extra one containing structures defined in the obsolete <sys/old_procfs.h> header file for the old ioctl()-based /proc interface. Programs should recognize and skip this old NOTE segment. It can be recognized by the presence of entries with entry name "CORE" and with these note types:
The one true NOTE segment contains the following entries. Each has entry name “CORE” and presents the contents of a system structure:
n_type: NT_PSINFO. This structure contains information of interest to the ps(1) command, such as process status, CPU usage, nice value, controlling terminal, user-ID, process-ID, the name of the executable, and so forth. The psinfo_t structure is defined in <sys/procfs.h>.
n_type: NT_PSTATUS. This structure contains things of interest to a debugger from the operating system, such as pending signals, state, process-ID, and so forth. The pstatus_t structure is defined in <sys/procfs.h>.
n_type: NT_PLATFORM. This entry contains a string describing the specific model of the hardware platform on which this core file was created. This information is the same as provided by sysinfo(2) when invoked with the command SI_PLATFORM.
n_type: NT_AUXV. This entry contains the array of auxv_t structures that was passed by the operating system as startup information to the dynamic linker. Auxiliary vector information is defined in <sys/auxv.h>.
n_type: NT_UTSNAME. This structure contains the system information that would have been returned to the process if it had performed a uname(2) system call prior to dumping core. The utsname structure is defined in <sys/utsname.h>.
n_type: NT_PRCRED. This structure contains the process credentials, including the real, saved, and effective user and group IDs. The prcred_t structure is defined in <aasys/procfs.h>. Following the structure is an optional array of supplementary group IDs. The total number of supplementary group IDs is given by the pr_ngroups member of the prcred_t structure, and the structure includes space for one supplementary group. If pr_ngroups is greater than 1, there is pr_ngroups - 1 gid_t items following the structure; otherwise, there is no additional data.
n_type: NT_ZONENAME. This entry contains a string which describes the name of the zone in which the process was running. See zones(7). The information is the same as provided by getzonenamebyid(3C) when invoked with the numerical ID returned by getzoneid(3C).
n_type: NT_LDT. This entry is present only on an 32-bit x86 machine and only if the process has set up a Local Descriptor Table (LDT). It contains an array of structures of type struct ssd, each of which was typically used to set up the %gs segment register to be used to fetch the address of the current thread information structure in a multithreaded process. The ssd structure is defined in <sys/sysi86.h>.
n_type: NT_CONTENT. This optional entry indicates which parts of the process image are specified to be included in the core file. See coreadm(8).
n_type: NT_SIGACTION. This entry contains an array of type struct sigaction which contains either the signal handler information or a valid signal for each process. The structure is defined in <sys/signal.h>.
Following these entries, for each active and zombie LWP in the process, the NOTE segment contains an entry with an lwpsinfo_t structure plus, for a non-zombie LWP, an entry with an lwpstatus_t structure, plus other optionally-present entries describing the LWP, as follows. A zombie LWP is a non-detached LWP that has terminated but has not yet been reaped by another LWP in the same process.
n_type: NT_LWPSINFO. This structure contains information of interest to the ps(1) command, such as LWP status, CPU usage, nice value, LWP-ID, and so forth. The lwpsinfo_t structure is defined in <sys/procfs.h>. This is the only entry present for a zombie LWP.
n_type: NT_LWPSTATUS. This structure contains things of interest to a debugger from the operating system, such as the general registers, the floating point registers, state, reason for stopping, LWP-ID, and so forth. The lwpstatus_t structure is defined in <sys/procfs.h>>.
n_type: NT_GWINDOWS. This entry is present only on a SPARC machine and only if the system was unable to flush all of the register windows to the stack. It contains all of the unspilled register windows. The gwindows_t structure is defined in <sys/regset.h>.
n_type: NT_PRXREG. This entry is present only if the machine has extra register state associated with it. It contains the extra register state. The prxregset_t structure is defined in <sys/procfs_isa.h>.
n_type: NT_ASRS. This entry is present only on a SPARC V9 machine and only if the process is a 64-bit process. It contains the ancillary state registers for the LWP. The asrset_t structure is defined in <sys/regset.h>.
n_type: NT_FDINFO. This entry contains information about an open file descriptor in the process. Each open file descriptor in the process has a dedicated NT_FDINFO entry in the corefile. The structure is defined in <sys/procfs.h>
The section header array of an ELF core file will contain entries for unwind sections, the dynamic section, and sections associated with the .dynsym dynamic symbol table, if any were found amongst the process's load objects. Depending on the coreadm(8) settings, the section header may also, include entries for CTF, and sections associated with the .symtab symbol table. In each case, the sh_addr field is set to the base address of the first mapping of the parent load object, enabling one to be matched with the other.
The size of the core file created by a process can be controlled by the user (see getrlimit(2)).
elfdump(1), gcore(1), mdb(1), proc(1), ps(1), getrlimit(2), setrlimit(2), setuid(2), sysinfo(2), uname(2), getzoneid(3C), getzonenamebyid(3C), elf(3ELF), signal.h(3HEAD), a.out(5), proc(5), zones(7), coreadm(8)