notice error exit from pager

If the pager fails to run, git produces no output, e.g.:

 $ GIT_PAGER=not-a-command git log

The error reporting fails for two reasons:

(1) start_command: There is a mechanism that detects errors during execvp introduced in 2b541bf8 (start_command: detect execvp failures early). The child writes one byte to a pipe only if execvp fails. The parent waits for either EOF, when the successful execvp automatically closes the pipe (see FD_CLOEXEC in fcntl(1)), or it reads a single byte, in which case it knows that the execvp failed. This mechanism is incompatible with the workaround introduced in 35ce8622 (pager: Work around window resizing bug in 'less'), which waits for input from the parent before the exec. Since both the parent and the child are waiting for input from each other, that would result in a deadlock. In order to avoid that, the mechanism is disabled by closing the child_notifier file descriptor. That is my best guess, at least, from reading the code. Hannes, can you confirm if this analysis is correct?

(2) finish_command: The parent correctly detects the 127 exit status from the child, but the error output goes nowhere, since by that time it is already being redirected to the child.

No simple solution for (1) comes to mind. I am not sure if this bug is still relevant. I could not reproduce it here with less version
436. If this bug was fixed a long time ago, maybe we can remove the workaround by now?

Number (2) can be solved by not sending error output to the pager. But I am not completely satisfied with that solution. Not redirecting error output to the pager can result in the pager overwriting error output with standard output. Right now, I can't think of any common situations where this would be the case. But, for example, a git log on a broken repo with some missing commits will abort at some point with "error: could not read <sha1> [...]". With this patch, if there is at least one page of standard output before that error, the error message is not visible at all and the output simply stops as if git log had reached the root commit.

The easiest solution is to produce the error output in the child. Below is a patch for that. The previous code seems to go through some pain to let the parent handle the error output. But I don't see the point. Is there more to this than meets my eye?

Clemens

---
 run-command.c |   11 ++++++-----
 1 files changed, 6 insertions(+), 5 deletions(-)

index 70e8a24..944a882 100644
@@ -127,9 +127,6 @@ static int wait_or_whine(pid_t pid, const char *argv0, int silent_exec_failure)
 		if (code == 127) {
 			code = -1;
 			failed_errno = ENOENT;
-			if (!silent_exec_failure)
-				error("cannot run %s: %s", argv0,
-					strerror(ENOENT));
 		}
 	} else {
 		error("waitpid is confused (%s)", argv0);

@@ -287,10 +284,14 @@ fail_pipe:
 		 * Do not check for cmd->silent_exec_failure; the parent
 		 * process will check it when it sees this exit code.
 		 */
-		if (errno == ENOENT)
+		if (errno == ENOENT) {
+			if (!cmd->silent_exec_failure)
+				error("cannot run %s: %s", cmd->argv[0],
+					strerror(ENOENT));
 			exit(127);
-		else
+		} else {
 			die_errno("cannot exec '%s'", cmd->argv[0]);
+		}
 	}
 	if (cmd->pid < 0)
 		error("cannot fork() for %s: %s", cmd->argv[0],

1.7.3.1.105.g84915

Hmm. I can confirm that it doesn't happen for me either any more, although I don't know if that is an actual less bugfix, or it was timing-dependent.

That said, the workaround has other nice properties too, like delaying the screen cleaning that "less" often does. So I don't know if we really want to get rid of it.

I think your patch looks reasonable.

Linus

Am 26.07.2011 23:04, schrieb Clemens Buchacher:

You analysis is correct. I think the bug in less was fixed shortly after this workaround was introduced. I would like to remove this workaround for a different reason, but it seems there are people who are quite fond of it ;)

The new process's error output may be redirected elsewhere, but if the exec fails, output should still go to the parent's stderr. This has already been done for the die_routine. Do the same for error_routine.

---

Ah, good point. I skipped over that part thinking it was not relevant. This should take care of it.

Clemens

 git-compat-util.h |    2 ++
 run-command.c     |   15 +++++++--------
 usage.c           |   18 ++++++++++++++++++
 3 files changed, 27 insertions(+), 8 deletions(-)

index a75530d..07cbfe9 100644
@@ -240,6 +240,7 @@ extern char *gitbasename(char *);
 
 /* General helper functions */
 extern void vreportf(const char *prefix, const char *err, va_list params);
+extern void vwritef(int fd, const char *prefix, const char *err, va_list params);
 extern NORETURN void usage(const char *err);
 extern NORETURN void usagef(const char *err, ...) __attribute__((format (printf, 1, 2)));
 extern NORETURN void die(const char *err, ...) __attribute__((format (printf, 1, 2)));

@@ -248,6 +249,7 @@ extern int error(const char *err, ...) __attribute__((format (printf, 1, 2)));
 extern void warning(const char *err, ...) __attribute__((format (printf, 1, 2)));
 
 extern void set_die_routine(NORETURN_PTR void (*routine)(const char *err, va_list params));
+extern void set_error_routine(void (*routine)(const char *err, va_list params));
 
 extern int prefixcmp(const char *str, const char *prefix);
 extern int suffixcmp(const char *str, const char *suffix);

index c752174..a2796c4 100644
@@ -77,16 +77,14 @@ static void notify_parent(void)
 
 static NORETURN void die_child(const char *err, va_list params)
 {
-	char msg[4096];
-	int len = vsnprintf(msg, sizeof(msg), err, params);
-	if (len > sizeof(msg))
-		len = sizeof(msg);
-
-	write_in_full(child_err, "fatal: ", 7);
-	write_in_full(child_err, msg, len);
-	write_in_full(child_err, "\n", 1);
+	vwritef(child_err, "fatal: ", err, params);
 	exit(128);
 }
+
+static void error_child(const char *err, va_list params)
+{
+	vwritef(child_err, "error: ", err, params);
+}
 #endif
 
 static inline void set_cloexec(int fd)

@@ -214,6 +212,7 @@ fail_pipe:
 			set_cloexec(child_err);
 		}
 		set_die_routine(die_child);
+		set_error_routine(error_child);
 
 		close(notify_pipe[0]);
 		set_cloexec(notify_pipe[1]);

index b5e67e3..a2a6678 100644
@@ -4,6 +4,7 @@
  * Copyright (C) Linus Torvalds, 2005
  */
 #include "git-compat-util.h"
+#include "cache.h"
 
 void vreportf(const char *prefix, const char *err, va_list params)
 {

@@ -12,6 +13,18 @@ void vreportf(const char *prefix, const char *err, va_list params)
 	fprintf(stderr, "%s%s\n", prefix, msg);
 }
 
+void vwritef(int fd, const char *prefix, const char *err, va_list params)
+{
+	char msg[4096];
+	int len = vsnprintf(msg, sizeof(msg), err, params);
+	if (len > sizeof(msg))
+		len = sizeof(msg);
+
+	write_in_full(fd, prefix, strlen(prefix));
+	write_in_full(fd, msg, len);
+	write_in_full(fd, "\n", 1);
+}
+
 static NORETURN void usage_builtin(const char *err, va_list params)
 {
 	vreportf("usage: ", err, params);

@@ -46,6 +59,11 @@ void set_die_routine(NORETURN_PTR void (*routine)(const char *err, va_list param
 	die_routine = routine;
 }
 
+void set_error_routine(void (*routine)(const char *err, va_list params))
+{
+	error_routine = routine;
+}
+
 void NORETURN usagef(const char *err, ...)
 {
 	va_list params;

1.7.3.1.105.g84915

If the pager fails to run, git produces no output, e.g.:

 $ GIT_PAGER=not-a-command git log

The error reporting fails for two reasons:

(1) start_command: There is a mechanism that detects errors during execvp introduced in 2b541bf8 (start_command: detect execvp failures early). The child writes one byte to a pipe only if execvp fails. The parent waits for either EOF, when the successful execvp automatically closes the pipe (see FD_CLOEXEC in fcntl(1)), or it reads a single byte, in which case it knows that the execvp failed. This mechanism is incompatible with the workaround introduced in 35ce8622 (pager: Work around window resizing bug in 'less'), which waits for input from the parent before the exec. Since both the parent and the child are waiting for input from each other, that would result in a deadlock. In order to avoid that, the mechanism is disabled by closing the child_notifier file descriptor.

(2) finish_command: The parent correctly detects the 127 exit status from the child, but the error output goes nowhere, since by that time it is already being redirected to the child.

No simple solution for (1) comes to mind.

Number (2) can be solved by not sending error output to the pager. Not redirecting error output to the pager can result in the pager overwriting error output with standard output, however.

Since there is no reliable way to handle error reporting in the parent, produce the output in the child instead.

---

Nothing new compared to the RFC, just a slightly trimmed commit message.

 run-command.c |   15 ++++++---------
 1 files changed, 6 insertions(+), 9 deletions(-)

index 5c91f37..a2796c4 100644
@@ -125,9 +125,6 @@ static int wait_or_whine(pid_t pid, const char *argv0, int silent_exec_failure)
 		if (code == 127) {
 			code = -1;
 			failed_errno = ENOENT;
-			if (!silent_exec_failure)
-				error("cannot run %s: %s", argv0,
-					strerror(ENOENT));
 		}
 	} else {
 		error("waitpid is confused (%s)", argv0);

@@ -282,14 +279,14 @@ fail_pipe:
 		} else {
 			execvp(cmd->argv[0], (char *const*) cmd->argv);
 		}
-		/*
-		 * Do not check for cmd->silent_exec_failure; the parent
-		 * process will check it when it sees this exit code.
-		 */
-		if (errno == ENOENT)
+		if (errno == ENOENT) {
+			if (!cmd->silent_exec_failure)
+				error("cannot run %s: %s", cmd->argv[0],
+					strerror(ENOENT));
 			exit(127);
-		else
+		} else {
 			die_errno("cannot exec '%s'", cmd->argv[0]);
+		}
 	}
 	if (cmd->pid < 0)
 		error("cannot fork() for %s: %s", cmd->argv[0],

1.7.3.1.105.g84915

Hmm, how does this interact with your earlier "error_routine: use parent's stderr if exec fails" patch?

The use case addressed in this patch (i.e., the pager) does not strictly dependent on the previous patch. But Johannes noted that in general, error output of the child could be redirected, hiding the execve failure again. The error_routine patch avoids that by writing to the parent's stderr stream.

Clemens