[REGRESSION] Interrupted clone/fetch leaves .lock files around

[REGRESSION] Interrupted clone/fetch leaves .lock files around

[Jonas Fonseca]

Hi,

It used to be possible to continue an interrupted clone when using
cg-clone, by cding into the partial repo and run cg-fetch. However, it
seems that the recent changes in the ref locking code ends up leaving
.lock files around when interrupted.

$ cg clone http://elinks.cz/elinks.git
defaulting to local storage area
Fetching head...
Fetching objects...
...
Getting pack c0e265dab40fa34912c3ee6e02ba29686ab84a7b
 which contains 16f85ec5043966c69e2142198c52e12494dcfc76
progress: 22 objects, 939754 bytes, now fetching c0e265dab40f... (657678 bytes)
cg-clone: interrupted
$ cd elinks
$ cg fetch
Recovering from a previously interrupted initial clone...
Fetching head...
Fetching objects...
error: Couldn't open lock file .git/refs/heads/origin.lock: File exists
error: Can't lock ref heads/origin
progress: 0 objects, 0 bytes
cg-fetch: objects fetch failed

Below is my feeble attempt at a (tested) fix.

diff --git a/fetch.c b/fetch.c
index e040ef9..861dc60 100644
--- a/fetch.c
+++ b/fetch.c
@@ -1,3 +1,5 @@
+#include <signal.h>
+
 #include "fetch.h"
 
 #include "cache.h"
@@ -214,9 +216,19 @@ static int mark_complete(const char *pat
 	return 0;
 }
 
+static struct ref_lock *lock = NULL;
+
+static void remove_lockfile_on_signal(int signo)
+{
+	if (lock)
+		unlock_ref(lock);
+	lock = NULL;
+	signal(SIGINT, SIG_DFL);
+	raise(signo);
+}
+
 int pull(char *target)
 {
-	struct ref_lock *lock = NULL;
 	unsigned char sha1[20];
 	char *msg;
 	int ret;
@@ -229,6 +241,7 @@ int pull(char *target)
 			error("Can't lock ref %s", write_ref);
 			return -1;
 		}
+		signal(SIGINT, remove_lockfile_on_signal);
 	}
 
 	if (!get_recover)
@@ -236,22 +249,11 @@ int pull(char *target)
 
 	if (interpret_target(target, sha1)) {
 		error("Could not interpret %s as something to pull", target);
-		if (lock)
-			unlock_ref(lock);
-		return -1;
-	}
-	if (process(lookup_unknown_object(sha1))) {
-		if (lock)
-			unlock_ref(lock);
-		return -1;
-	}
-	if (loop()) {
-		if (lock)
-			unlock_ref(lock);
-		return -1;
-	}
 
-	if (write_ref) {
+	} else if (process(lookup_unknown_object(sha1)) || loop()) {
+		; /* unlock */
+
+	} else if (write_ref) {
 		if (write_ref_log_details) {
 			msg = xmalloc(strlen(write_ref_log_details) + 12);
 			sprintf(msg, "fetch from %s", write_ref_log_details);
@@ -261,6 +263,10 @@ int pull(char *target)
 		if (msg)
 			free(msg);
 		return ret;
+	} else {
+		return 0;
 	}
-	return 0;
+
+	remove_lockfile_on_signal(0);
+	return -1;
 }

-- 
Jonas Fonseca

Re: [REGRESSION] Interrupted clone/fetch leaves .lock files around

[Jonas Fonseca]

Jonas Fonseca <fonseca@diku.dk> wrote Tue, Jun 06, 2006:
> Below is my feeble attempt at a (tested) fix.

Ok, so maybe I didn't test it so well, other than continuously
interrupting the fetch. ;)
 
> diff --git a/fetch.c b/fetch.c
> index e040ef9..861dc60 100644
> --- a/fetch.c
> +++ b/fetch.c
> @@ -214,9 +216,19 @@ static int mark_complete(const char *pat
>  	return 0;
>  }
>  
> +static struct ref_lock *lock = NULL;
> +
> +static void remove_lockfile_on_signal(int signo)
> +{
> +	if (lock)
> +		unlock_ref(lock);
> +	lock = NULL;
> +	signal(SIGINT, SIG_DFL);
> +	raise(signo);
> +}
> +
>  int pull(char *target)
>  {
> -	struct ref_lock *lock = NULL;
>  	unsigned char sha1[20];
>  	char *msg;
>  	int ret;
...
> @@ -261,6 +263,10 @@ int pull(char *target)
>  		if (msg)
>  			free(msg);
>  		return ret;
> +	} else {
> +		return 0;
>  	}
> -	return 0;
> +
> +	remove_lockfile_on_signal(0);

This will end up calling raise().

> +	return -1;
>  }

-- 
Jonas Fonseca

Re: [REGRESSION] Interrupted clone/fetch leaves .lock files around

[Junio C Hamano]

Jonas Fonseca <fonseca@diku.dk> writes:

> Below is my feeble attempt at a (tested) fix.
>
> diff --git a/fetch.c b/fetch.c
> index e040ef9..861dc60 100644
> --- a/fetch.c
> +++ b/fetch.c
> @@ -1,3 +1,5 @@
> +#include <signal.h>
> +
>  #include "fetch.h"

I suspect you could do something similar to what we already do
for index updates using atexit().  Let me take a look.

Re: [REGRESSION] Interrupted clone/fetch leaves .lock files around

[Junio C Hamano]

Junio C Hamano <junkio@cox.net> writes:

> Jonas Fonseca <fonseca@diku.dk> writes:
>
>> Below is my feeble attempt at a (tested) fix.
>>
>> diff --git a/fetch.c b/fetch.c
>> index e040ef9..861dc60 100644
>> --- a/fetch.c
>> +++ b/fetch.c
>> @@ -1,3 +1,5 @@
>> +#include <signal.h>
>> +
>>  #include "fetch.h"
>
> I suspect you could do something similar to what we already do
> for index updates using atexit().  Let me take a look.

Indeed it turns out that the signal work Pasky did in index.c is
exactly suitable for this.  I've pushed out three patches in
"next" -- a few more eyeballs are appreciated on this one.