Re: If you are one of the cool kids who cranks kern.bufcachepercent up..
@@ -350,7 +359,7 @@ * referencing buffers. */ toggle ^= 1; - if (numvnodes 2 * maxvnodes) + if (numvnodes / 2 maxvnodes) toggle = 0; + if ((numvnodes 1) maxvnodes) slightly better? I find your lack of faith in the compiler disturbing.
Re: If you are one of the cool kids who cranks kern.bufcachepercent up..
On Sat, Jan 09, 2010 at 04:05:13AM -0700, Bob Beck wrote: [dd] -Bob Index: kern/vfs_subr.c === RCS file: /cvs/src/sys/kern/vfs_subr.c,v retrieving revision 1.184 diff -u -r1.184 vfs_subr.c @@ -350,7 +359,7 @@ * referencing buffers. */ toggle ^= 1; - if (numvnodes 2 * maxvnodes) + if (numvnodes / 2 maxvnodes) toggle = 0; + if ((numvnodes 1) maxvnodes) slightly better? --patrick
Re: If you are one of the cool kids who cranks kern.bufcachepercent up..
On Saturday, January 9, 2010, Joerg Sonnenberger jo...@britannica.bec.de wrote: On Sat, Jan 09, 2010 at 08:15:20PM +, Grumpy wrote: @@ -350,7 +359,7 @@ * referencing buffers. */ toggle ^= 1; - if (numvnodes 2 * maxvnodes) + if (numvnodes / 2 maxvnodes) toggle = 0; + if ((numvnodes 1) maxvnodes) slightly better? I find your lack of faith in the compiler disturbing. I find your lack of type checking before trolling disturbing. stares at the code Ummm, what? AFAICT, the only difference between the two variants that I can see is that the shift version might behave differently if numvnodes is negative, which can't happen. Can you be less funny and more clear about what type checking issue you see? Philip Guenther
Re: If you are one of the cool kids who cranks kern.bufcachepercent up..
On Sat, Jan 09, 2010 at 02:41:02PM -0800, Philip Guenther wrote: I find your lack of faith in the compiler disturbing. I find your lack of type checking before trolling disturbing. stares at the code Ummm, what? AFAICT, the only difference between the two variants that I can see is that the shift version might behave differently if numvnodes is negative, which can't happen. Can you be less funny and more clear about what type checking issue you see? You have analysed the situation correctly. The problem is that the compiler does not know that the signed numvnodes is never negative, so it creates different code. E.g. on AMD64 it is 6 instructions for the signed division by 2 compared to one instruction for the explicit shift or the unsigned division. Joerg
Re: If you are one of the cool kids who cranks kern.bufcachepercent up..
On Sat, Jan 9, 2010 at 6:22 PM, Joerg Sonnenberger jo...@britannica.bec.de wrote: You have analysed the situation correctly. The problem is that the compiler does not know that the signed numvnodes is never negative, so it creates different code. E.g. on AMD64 it is 6 instructions for the signed division by 2 compared to one instruction for the explicit shift or the unsigned division. That's a good argument for making numvnodes unsigned, but not for obfuscating the code.
Re: [patch] user/6104: newsyslog count off by one
Hi David,
--- newsyslog.c.bak Thu Oct 22 21:18:12 2009
+++ newsyslog.c Thu Oct 22 22:00:09 2009
@@ -756,6 +756,9 @@
char file1[MAXPATHLEN], file2[MAXPATHLEN], *suffix;
int numdays = ent-numlogs;
+ /* We start counting at 0 so reduce for correct maximum */
+ numdays--;
+
/* Remove oldest log (may not exist) */
(void)snprintf(file1, sizeof(file1), %s.%d, oldlog, numdays);
(void)snprintf(file2, sizeof(file2), %s.%d%s, oldlog, numdays,
That does part of the job, but looks slightly ugly.
When applied to a running system, it lets the excessive archive file
sit around forever, or until you manually delete it.
I like the following better:
Remove all _consecutive_ archive files above the requested limit.
That's also useful after lowering the count in newsyslog.conf.
Note that the loop may be terminated earlier when rotating for real
than in -n mode; specifically, when unlink() fails.
But that's fine because -n is supposed to tell you what would be tried.
Of course, what you try to do may still fail.
Any OKs?
Index: newsyslog.c
===
RCS file: /cvs/src/usr.bin/newsyslog/newsyslog.c,v
retrieving revision 1.86
diff -u -p -r1.86 newsyslog.c
--- newsyslog.c 27 Oct 2009 23:59:40 - 1.86
+++ newsyslog.c 10 Jan 2010 00:36:43 -
@@ -750,22 +750,25 @@ void
rotate(struct conf_entry *ent, const char *oldlog)
{
char file1[MAXPATHLEN], file2[MAXPATHLEN], *suffix;
- int numdays = ent-numlogs;
+ int numdays = ent-numlogs - 1;
+ int done = 0;
- /* Remove oldest log (may not exist) */
- (void)snprintf(file1, sizeof(file1), %s.%d, oldlog, numdays);
- (void)snprintf(file2, sizeof(file2), %s.%d%s, oldlog, numdays,
- COMPRESS_POSTFIX);
-
- if (noaction) {
- printf(\trm -f %s %s\n, file1, file2);
- } else {
- (void)unlink(file1);
- (void)unlink(file2);
- }
+ /* Remove old logs */
+ do {
+ (void)snprintf(file1, sizeof(file1), %s.%d, oldlog, numdays);
+ (void)snprintf(file2, sizeof(file2), %s.%d%s, oldlog,
+ numdays, COMPRESS_POSTFIX);
+ if (noaction) {
+ printf(\trm -f %s %s\n, file1, file2);
+ done = access(file1, 0) access(file2, 0);
+ } else {
+ done = unlink(file1) unlink(file2);
+ }
+ numdays++;
+ } while (done == 0);
/* Move down log files */
- while (numdays--) {
+ for (numdays = ent-numlogs - 1; numdays = 0; numdays--) {
/*
* If both the compressed archive and the non-compressed archive
* exist, we decide which to rotate based on the CE_COMPACT flag
Re: If you are one of the cool kids who cranks kern.bufcachepercent up..
On Sat, Jan 09, 2010 at 06:35:45PM -0500, Ted Unangst wrote: On Sat, Jan 9, 2010 at 6:22 PM, Joerg Sonnenberger jo...@britannica.bec.de wrote: You have analysed the situation correctly. The problem is that the compiler does not know that the signed numvnodes is never negative, so it creates different code. E.g. on AMD64 it is 6 instructions for the signed division by 2 compared to one instruction for the explicit shift or the unsigned division. That's a good argument for making numvnodes unsigned, but not for obfuscating the code. With respect to 'obfuscating', I would agree with you, except using shift to divide (or multiply) by two, or multiples of two, is very common. --patrick
Re: If you are one of the cool kids who cranks kern.bufcachepercent up..
Good god people. just test the fscking diff and stop the chestbeating and cockpulling already.
