Malcom, Gcc supports thread-local variables. See https://gcc.gnu.org/onlinedocs/gcc-3.3.1/gcc/Thread-Local.html
I am not sure about native compilers on Solaris, HPUX, or AIX. In any case, I found out that the Windows native code in Hadoop seems to handle error messages properly. Here is what I found: $ find ~/work/hadoop/hadoop-trunk/ -name '*.c'|xargs grep FormatMessage|awk -F: '{print $1}'|sort -u /home/asokan/work/hadoop/hadoop-trunk/hadoop-common-project/hadoop-common/src/main/native/src/org/apache/hadoop/io/nativeio/NativeIO.c /home/asokan/work/hadoop/hadoop-trunk/hadoop-common-project/hadoop-common/src/main/native/src/org/apache/hadoop/security/JniBasedUnixGroupsMappingWin.c /home/asokan/work/hadoop/hadoop-trunk/hadoop-common-project/hadoop-common/src/main/winutils/libwinutils.c $ find ~/work/hadoop/hadoop-trunk/ -name '*.c'|xargs grep terror|awk -F: '{print $1}'|sort -u /home/asokan/work/hadoop/hadoop-trunk/hadoop-common-project/hadoop-common/src/main/native/src/exception.c /home/asokan/work/hadoop/hadoop-trunk/hadoop-common-project/hadoop-common/src/main/native/src/org/apache/hadoop/io/nativeio/SharedFileDescriptorFactory.c /home/asokan/work/hadoop/hadoop-trunk/hadoop-common-project/hadoop-common/src/main/native/src/org/apache/hadoop/net/unix/DomainSocket.c /home/asokan/work/hadoop/hadoop-trunk/hadoop-common-project/hadoop-common/src/main/native/src/org/apache/hadoop/net/unix/DomainSocketWatcher.c /home/asokan/work/hadoop/hadoop-trunk/hadoop-common-project/hadoop-common/src/main/native/src/org/apache/hadoop/security/JniBasedUnixGroupsMapping.c This means you need not worry about the Windows version of terror(). You need to change five files that contain UNIX specific native code. I have a question on your suggested implementation: How do you initialize the static errlist array in a thread-safe manner? ________________________________ Here is another thread-safe implementation that I could come up with: #include <string.h> #include <stdlib.h> #include <errno.h> #include <stdio.h> #define MESSAGE_BUFFER_SIZE 256 char * getSystemErrorMessage(char * buf, int buf_len, int code) { #if defined(_HPUX_SOURCE) char * msg; errno = 0; msg = strerror(code); if (errno == 0) { strncpy(buf, msg, buf_len-1); buf[buf_len-1] = '\0'; } else { snprintf(buf, buf_len, "%s %d", "Can't get system error message for code", code); } #else if (strerror_r(code, buf, buf_len) != 0) { snprintf(buf, buf_len, "%s %d", "Can't get system error message for code", code); } #endif return buf; } #define TERROR(code) \ getSystemErrorMessage(messageBuffer, sizeof(messageBuffer), code) int main(int argc, char ** argv) { if (argc > 1) { char messageBuffer[MESSAGE_BUFFER_SIZE]; int code = atoi(argv[1]); fprintf(stderr, "System error for code %s: %s\n", argv[1], TERROR(code)); } return 0; } This changes terror to a macro TERROR and requires all functions that call TERROR macro to declare the local variable messageBuffer. Since there are only five files to modify, I think it is not a big effort. What do you think? -- Asokan On 12/13/2014 04:29 AM, malcolm wrote: Colin, I am not sure what you mean by a thread-local buffer (in native code). In Java this is pretty standard, but I couldn't find any implementation for C code. Here is the terror function: const char* terror(int errnum) { if ((errnum < 0) || (errnum >= sys_nerr)) { return "unknown error."; } return sys_errlist[errnum]; } The interface is identical to strerror, but the implementation is actually re-entrant since it returns a pointer to a static string. If I understand your suggestion, the new function would look like this: const char* terror(int errnum) { static char result[65]; strerror_r(errnum, result, 64); return result; } No need for snprintf, strerror_r has the 'n' bounding built-in. Of course, this is still non-re-entrant, so unless the caller copies the returned buffer, before the function is called again, there is a problem. After considerable thought, I have come up with this version of terror, tested OK on Windows, Linux and Solaris: #if defined(_WIN32) #define strerror_r(errno,buf,len) strerror_s(buf,len,errno) #endif #define MAX_ERRORS 256 #define MAX_ERROR_LEN 80 char *terror(int errnum) { static char errlist[MAX_ERRORS][MAX_ERROR_LEN+1]; // cache of error messages if ( errnum >= 0 && errnum < MAX_ERRORS ) { if ( errlist[errnum][0] == 0 ) strerror_r( errnum, errlist[errnum], MAX_ERROR_LEN); return errlist[errnum]; } else { return "Unknown error"; } } This version is portable and re-entrant. On windows, the largest errnum is 43, on Ubuntu 14.04 we have 133, and on Solaris 11.1 we get 151. If this is OK with you, I will open a jira for this. Thanks, Malcolm On 12/12/2014 11:10 PM, Colin McCabe wrote: Just use snprintf to copy the error message from strerror_r into a thread-local buffer of 64 bytes or so. Then preserve the existing terror() interface. Can you open a jira for this? best, Colin On Thu, Dec 11, 2014 at 8:35 PM, malcolm<malcolm.kaval...@oracle.com><mailto:malcolm.kaval...@oracle.com> wrote: So, turns out that if I had naively changed all calls to terror or references to sys_errlist, to using strerror_r, then I would have broken code for Windows and HPUX (and possibly other OSes). If we are to assume that current code runs fine on all platforms (maybe even AIX an MacOS, for example), then any change/additions made to the code and not ifdeffed appropriately can break on other OSes. On the other hand, too many ifdefs can pollute the code source and render it less readable (though possibly less important). In the general case what are code contributors responsibilities to adding code regarding OSes besides Linux ? What OSes does jenkins test on ? I guess maintainers of code on non-tested platforms are responsible for their own testing ? How do we avoid the ping-pong effect, i.e. I make a generic change to code which breaks on Windows, then the Windows maintainer reverts changes to break on Solaris for example ? Or does this not happen in actuality ? On 12/11/2014 11:25 PM, Asokan, M wrote: Hi Malcom, The Windows versions of strerror() and strerror_s() functions are probably meant for ANSI C library functions that set errno. For core Windows API calls (like UNIX system calls), one gets the error number by calling GetLastError() function. In the code snippet I sent earlier, the "code" argument is the value returned by GetLastError(). Neither strerror() nor strerror_s() will give the correct error message for this error code. You could probably look at libwinutils.c in Hadoop source. It uses FormatMessageW (which returns messages in Unicode.) My requirement was to return messages in current system locale. -- Asokan ________________________________________ From: malcolm [malcolm.kaval...@oracle.com<mailto:malcolm.kaval...@oracle.com>] Sent: Thursday, December 11, 2014 4:04 PM To:common-dev@hadoop.apache.org<mailto:To:common-dev@hadoop.apache.org> Subject: Re: Solaris Port Hi Asok, I googled and found that windows has strerror, and strerror_s (which is the strerror_r equivalent). Is there a reason why you didn't use this call ? On 12/11/2014 06:27 PM, Asokan, M wrote: Hi Malcom, Recently, I had to work on a function to get system error message on various systems. Here is the piece of code I came up with. Hope it helps. static void get_system_error_message(char * buf, int buf_len, int code) { #if defined(_WIN32) LPVOID lpMsgBuf; DWORD status = FormatMessage(FORMAT_MESSAGE_ALLOCATE_BUFFER | FORMAT_MESSAGE_FROM_SYSTEM | FORMAT_MESSAGE_IGNORE_INSERTS, NULL, code, MAKELANGID(LANG_NEUTRAL, SUBLANG_DEFAULT), /* Default language */ (LPTSTR) &lpMsgBuf, 0, NULL); if (status > 0) { strncpy(buf, (char *)lpMsgBuf, buf_len-1); buf[buf_len-1] = '\0'; /* Free the buffer returned by system */ LocalFree(lpMsgBuf); } else { _snprintf(buf, buf_len-1 , "%s %d", "Can't get system error message for code", code); buf[buf_len-1] = '\0'; } #else #if defined(_HPUX_SOURCE) { char * msg; errno = 0; msg = strerror(code); if (errno == 0) { strncpy(buf, msg, buf_len-1); buf[buf_len-1] = '\0'; } else { snprintf(buf, buf_len, "%s %d", "Can't get system error message for code", code); } } #else if (strerror_r(code, buf, buf_len) != 0) { snprintf(buf, buf_len, "%s %d", "Can't get system error message for code", code); } #endif #endif } Note that HPUX does not have strerror_r() since strerror() itself is thread-safe. Also Windows does not have snprintf(). The equivalent function _snprintf() has a subtle difference in its interface. -- Asokan ________________________________________ From: malcolm [malcolm.kaval...@oracle.com<mailto:malcolm.kaval...@oracle.com>] Sent: Thursday, December 11, 2014 11:02 AM To:common-dev@hadoop.apache.org<mailto:To:common-dev@hadoop.apache.org> Subject: Re: Solaris Port Fine with me, I volunteer to do this, if accepted. On 12/11/2014 05:48 PM, Allen Wittenauer wrote: sys_errlist was removed for a reason. Creating a fake sys_errlist on Solaris will mean the libhadoop.so will need to be tied a specific build (kernel/include pairing) and therefore limits upward mobility/compatibility. That doesn’t seem like a very good idea. IMO, switching to strerror_r is much preferred, since other than the brain-dead GNU libc version, is highly portable and should work regardless of the kernel or OS in place. On Dec 11, 2014, at 5:20 AM, malcolm<malcolm.kaval...@oracle.com><mailto:malcolm.kaval...@oracle.com> wrote: FYI, there are a couple more files that reference sys_errlist directly (not just terror within exception.c) , but also hdfs_http_client.c and NativeiO.c On 12/11/2014 07:38 AM, malcolm wrote: Hi Colin, Exactly, as you noticed, the problem is the thread-local buffer needed to return from terror. Currently, terror just returns a static string from an array, this is fast, simple and error-proof. In order to use strerror_r inside terror, would require allocating a buffer inside terror and depend on the caller to free the buffer after using it, or to pass a buffer to terrror (which is basically the same as strerror_r, rendering terror redundant). Both cases require modification outside terror itself, as far as I can tell, no simple fix. Unless you have an alternative which I haven't thought of ? As far as I can tell, we have two choices: 1. Remove terror and replace calls with strerror_r, passing a buffer from the callee. Advantage: a more modern portable interface. Disadvantage: All calls to terror need to be modified, though all seem to be in a few files as far as I can tell. 2. Adding a sys_errlist array (ifdeffed for Solaris) Advantage: no change to any calls to terror Disadvantage: 2 additional files added to source tree (.c and .h) and some minor ifdefs only used for Solaris. I think it is more a question of style than anything else, so I leave you to make the call. Thanks for your patience, Malcolm On 12/10/2014 09:54 PM, Colin McCabe wrote: On Wed, Dec 10, 2014 at 2:31 AM, malcolm <malcolm.kaval...@oracle.com><mailto:malcolm.kaval...@oracle.com> wrote: Hi Colin, Thanks for the hints around JIRAs. You are correct errno still exists, however sys_errlist does not. Hadoop uses a function terror (defined in exception.c) which indexes sys_errlist by errno to return the error message from the array. This function is called 26 times in various places (in 2.2) Originally, I thought to replace all calls to terror with strerror, but there can be issues with multi-threading (it returns a buffer which can be overwritten), so it seemed simpler just to recreate the sys_errlist message array. There is also a multi-threaded version strerror_r where you pass the buffer as a parameter, but this would necessitate changing every call to terror with mutiple lines of code. Why don't you just use strerror_r inside terror()? I wrote that code originally. The reason I didn't want to use strerror_r there is because GNU libc provides a non-POSIX definition of strerror_r, and forcing it to use the POSIX one is a pain. But you can do it. You also will require a thread-local buffer to hold the return from strerror_r, since it is not guaranteed to be static (although in practice it is 99% of the time-- another annoyance with the API). ________________________________ ATTENTION: ----- The information contained in this message (including any files transmitted with this message) may contain proprietary, trade secret or other confidential and/or legally privileged information. Any pricing information contained in this message or in any files transmitted with this message is always confidential and cannot be shared with any third parties without prior written approval from Syncsort. This message is intended to be read only by the individual or entity to whom it is addressed or by their designee. If the reader of this message is not the intended recipient, you are on notice that any use, disclosure, copying or distribution of this message, in any form, is strictly prohibited. If you have received this message in error, please immediately notify the sender and/or Syncsort and destroy all copies of this message in your possession, custody or control.