X-Original-To: pgsql-performance-postgresql.org@localhost.postgresql.org Received: from localhost (neptune.hub.org [200.46.204.2]) by svr1.postgresql.org (Postfix) with ESMTP id 3806AD1B491 for ; Thu, 4 Dec 2003 21:20:52 +0000 (GMT) Received: from svr1.postgresql.org ([200.46.204.71]) by localhost (neptune.hub.org [200.46.204.2]) (amavisd-new, port 10024) with ESMTP id 26242-05 for ; Thu, 4 Dec 2003 17:20:23 -0400 (AST) Received: from yertle.kcilink.com (yertle.kcilink.com [216.194.193.105]) by svr1.postgresql.org (Postfix) with ESMTP id D9481D1B47A for ; Thu, 4 Dec 2003 17:20:20 -0400 (AST) Received: by yertle.kcilink.com (Postfix, from userid 100) id 5AF5A2178A; Thu, 4 Dec 2003 16:20:22 -0500 (EST) From: Vivek Khera MIME-Version: 1.0 Content-Type: text/plain; charset=us-ascii Content-Transfer-Encoding: 7bit Message-ID: <16335.42262.196435.193875@yertle.int.kciLink.com> Date: Thu, 4 Dec 2003 16:20:22 -0500 To: "Matthew T. O'Connor" Cc: Subject: Re: autovacuum daemon stops doing work after about an hour In-Reply-To: <2348.66.106.27.14.1070571171.squirrel@matth.zeut.net> References: <16332.63508.196365.801418@yertle.int.kciLink.com> <1070516326.1455.11.camel@zedora.zeut.net> <2348.66.106.27.14.1070571171.squirrel@matth.zeut.net> X-Mailer: VM 7.17 under 21.4 (patch 14) "Reasonable Discussion" XEmacs Lucid X-Virus-Scanned: by amavisd-new at postgresql.org X-Archive-Number: 200312/99 X-Sequence-Number: 4959 >>>>> "MTO" == Matthew T O'Connor writes: MTO> Could this be the recently reported bug where time goes backwards on MTO> FreeBSD? Can anyone who knows more about this problem chime in, I know it MTO> was recently discussed on Hackers. Time does not go backwards -- the now and then variables are properly incrementing in time as you see from the debugging output. The error appears to be with the computation of the "diff". It is either a C programming error, or a compiler error. I'm not a C "cop" so I can't tell you which it is. Witness this program, below, compiled as "cc -g -o t t.c" and the output here: % ./t seconds = 3509 seconds1 = 3509000000 useconds = -452486 stepped diff = 3508547514 seconds2 = -785967296 seconds3 = 3509000000 diff = -786419782 long long diff = 3508547514 % apperantly, if you compute (now.tv_sec - then.tv_sec) * 1000000 all at once, it overflows since the RHS is all computed using longs rather than long longs. Fix is to cast at least one of the values to long long on the RHS, as in the computation of seconds3 below. compare that to the computation of seconds2 and you'll see that this is the cause. I'd be curious to see the output of this program on other platforms and other compilers. I'm using gcc 2.95.4 as shipped with FreeBSD 4.8+. That all being said, you should never sleep less than the base time, and never for more than a max amount, perhaps 1 hour? --cut here-- #include #include int main() { struct timeval now, then; long long diff = 0; long long seconds, seconds1, seconds2, seconds3, useconds; now.tv_sec = 1070565077L; now.tv_usec = 216477L; then.tv_sec = 1070561568L; then.tv_usec = 668963L; seconds = now.tv_sec - then.tv_sec; printf("seconds = %lld\n",seconds); seconds1 = seconds * 1000000; printf("seconds1 = %lld\n",seconds1); useconds = now.tv_usec - then.tv_usec; printf("useconds = %lld\n",useconds); diff = seconds1 + useconds; printf("stepped diff = %lld\n",diff); /* this appears to be the culprit... it should be same as seconds1 */ seconds2 = (now.tv_sec - then.tv_sec) * 1000000; printf("seconds2 = %lld\n",seconds2); /* seems we need to cast long's to long long's for this computation */ seconds3 = ((long long)now.tv_sec - (long long)then.tv_sec) * 1000000; printf("seconds3 = %lld\n",seconds3); diff = (now.tv_sec - then.tv_sec) * 1000000 + (now.tv_usec - then.tv_usec); printf ("diff = %lld\n",diff); diff = ((long long)now.tv_sec - (long long)then.tv_sec) * 1000000 + (now.tv_usec - then.tv_usec); printf ("long long diff = %lld\n",diff); exit(0); } --cut here--