From mboxrd@z Thu Jan 1 00:00:00 1970 Return-Path: Received: from matoro.tk (unknown [IPv6:2600:1700:4b10:9d80::2]) by sourceware.org (Postfix) with ESMTP id 2478B3858D28 for ; Mon, 9 Oct 2023 02:01:49 +0000 (GMT) DMARC-Filter: OpenDMARC Filter v1.4.2 sourceware.org 2478B3858D28 Authentication-Results: sourceware.org; dmarc=pass (p=reject dis=none) header.from=matoro.tk Authentication-Results: sourceware.org; spf=pass smtp.mailfrom=matoro.tk DKIM-Signature: a=rsa-sha256; bh=Y8hcaU27U3XcJGywaTs+p7EUE8S6VNpp0Vf72OoHeYo=; c=relaxed/relaxed; d=matoro.tk; h=Subject:Subject:Sender:To:To:Cc:Cc:From:From:Date:Date:MIME-Version:MIME-Version:Content-Type:Content-Type:Content-Transfer-Encoding:Content-Transfer-Encoding:Reply-To:In-Reply-To:In-Reply-To:Message-Id:Message-Id:References:References:Autocrypt:Openpgp; i=@matoro.tk; s=20230917; t=1696816887; v=1; x=1697248887; b=egbs78kFigl+dvBR4wgVWDNW5Mebe51kjiJKAChsNTtLAn+WwVLQ0Jra0lhEoWgLcj0lTIJV ZPZyLhBfJCq+vfADP+r9upO8Nya8Ohk767dbNw2VtABjBt2gek7249MH4vbDW2JtaVT+V2pil3v UZ0kYJRLsWB2Bsleg/5lUNolYNsw3KB37QB3FgRIl2QRci1xz9zpU7DZn2mkBnOI9+BjyNfSXdt y8JGk5m7ty+XriwThLQHGt9g5OMOxrXdLd7uW01nwuRaaVeVb9eSJUSzb43b+6ZyO3MUGvC/LdX iKThtxZwvqC1qI7IVD6479Vt0yeY4B6BTQv48/s5sg1pksfa9uluzvOLnS/yYOnjdbs+bsDP3SP +MM4DNH71t+Nx8D6eYB9ltY16urkc7Tc1u/kZFuJ5zAF0qNX//Us2lSL3w3/5dxpWPLywfFHjh8 bOGSJm8DmpMAQD5k+yuifFKDmiiq9bW5LIs7jmkcbjocNT30ChB5gUuw2rzHCeO3J5Jofk1E2Fx DaliXMwYwt9jZ7T58yqabIQ4XR5qKSIRSznBlAu9gWe09/96wpCq7g/7D6QpRlKEalDPiyrx8WO sAtGgzav+H/Ab4G52PRq3Cngrskl1W7MRbvf7N1Lwh4GU5DvqQuKx3/oXwn+nQMvxPWWJoZTe7k I2Taf8LBWMc= Received: by matoro.tk (envelope-sender ) with ESMTPS id 58cb133c; Sun, 08 Oct 2023 22:01:27 -0400 MIME-Version: 1.0 Date: Sun, 08 Oct 2023 22:01:27 -0400 From: matoro To: aurelien@aurel32.net, libc-alpha@sourceware.org Cc: mattst88@gmail.com, carlos@redhat.com, Sam James Subject: Re: [PATCH] alpha: correct handling of negative *rlimit() args besides -1 In-Reply-To: <10925e7b87d9edef1229db7beb0761b0@matoro.tk> References: <20221008024522.523048-1-mattst88@gmail.com> <10925e7b87d9edef1229db7beb0761b0@matoro.tk> Message-ID: <941585b729f19e14a3143b7be44fbf35@matoro.tk> X-Sender: matoro_mailinglist_glibc@matoro.tk Content-Type: text/plain; charset=US-ASCII; format=flowed Content-Transfer-Encoding: 7bit X-Spam-Status: No, score=-13.1 required=5.0 tests=BAYES_00,DKIM_SIGNED,DKIM_VALID,DKIM_VALID_AU,DKIM_VALID_EF,GIT_PATCH_0,KAM_NUMSUBJECT,SPF_HELO_PASS,SPF_PASS,TXREP autolearn=ham autolearn_force=no version=3.4.6 X-Spam-Checker-Version: SpamAssassin 3.4.6 (2021-04-09) on server2.sourceware.org List-Id: On 2023-02-14 14:38, matoro wrote: > Hi Aurelien, I came up with the idea for this originally. Matt noticed > that it had stalled and asked me to check back in. > >> On 2022-10-07 22:45, Matt Turner via Libc-alpha wrote: >> > The generic version of RLIM_INFINITY in Linux is equal to (rlim_t)-1, >> > which is equal to ULLONG_MAX. On alpha however it is instead defined as >> > 0x7ffffffffffffffful. This was special-cased in 0d0bc78 [BZ #22648] but >> > it specifically used an equality check. >> >> I am not sure this commit is giving the full picture, commits around >> should also be checked to understand it. > > Can you elaborate here? This was my understanding based on what I > read, but you are the original author, so your perspective will surely > be more complete than mine. > >> > There is a cpython test case test_prlimit_refcount which calls >> > setrlimit() with { -2, -2 } as arguments rather than the usual -1, it >> > therefore fails the equality test and is treated as a large arbitrary >> > positive value past the maximum of RLIM_INFINITY and fails with EPERM. >> > This patch changes the behavior of the *rlimit() calls to treat all >> > integers between 0x7ffffffffffffffful and (rlim_t)-1 as (rlim_t)-1, >> > i.e., RLIM_INFINITY. >> >> Basically on alpha, the glibc API is now identical to the prlimit64 >> API, >> which means there is a dead zone with invalid values from >> 0x8000000000000000ul to 0xfffffffffffffffeul. The kernel returns EPERM >> for values in this range. >> >> You suggestion is to consider values is this zone as infinity. I have >> mixed feeling about that. From the setrlimit() side it looks like the >> correct thing to do. But this breaks the assumption that calling >> getrlimit() after a successful setrlimit() call will return the same >> value. > > Is this behavior specified one way or the other? Alternatively, is the > Python unit test making an assumption that is not guaranteed (that > calling setrlimit() with a negative value behaves the same way as > calling it specifically with RLIM_INFINITY)? If this is Python's > mistake, that can be corrected there. The test in question: > https://github.com/python/cpython/blob/main/Lib/test/test_resource.py#L163-L175 > >> > diff --git a/sysdeps/unix/sysv/linux/alpha/getrlimit64.c b/sysdeps/unix/sysv/linux/alpha/getrlimit64.c >> > index c195f5b55c..40f3e6bdff 100644 >> > --- a/sysdeps/unix/sysv/linux/alpha/getrlimit64.c >> > +++ b/sysdeps/unix/sysv/linux/alpha/getrlimit64.c >> > @@ -38,11 +38,11 @@ __old_getrlimit64 (enum __rlimit_resource resource, >> > if (__getrlimit64 (resource, &krlimits) < 0) >> > return -1; >> > >> > - if (krlimits.rlim_cur == RLIM64_INFINITY) >> > + if (krlimits.rlim_cur >= OLD_RLIM64_INFINITY) >> > rlimits->rlim_cur = OLD_RLIM64_INFINITY; >> > else >> > rlimits->rlim_cur = krlimits.rlim_cur; >> > - if (krlimits.rlim_max == RLIM64_INFINITY) >> > + if (krlimits.rlim_max >= OLD_RLIM64_INFINITY) >> > rlimits->rlim_max = OLD_RLIM64_INFINITY; >> > else >> > rlimits->rlim_max = krlimits.rlim_max; >> >> That said, I do not understand the change there. It is done on the >> *compat* symbol which still uses the old glibc API definition. There >> we >> want to keep doing the exact reverse operations as in the >> rlim_to_rlim64() kernel function. >> >> > diff --git a/sysdeps/unix/sysv/linux/alpha/setrlimit64.c b/sysdeps/unix/sysv/linux/alpha/setrlimit64.c >> > index 421616ed20..4e88540a48 100644 >> > --- a/sysdeps/unix/sysv/linux/alpha/setrlimit64.c >> > +++ b/sysdeps/unix/sysv/linux/alpha/setrlimit64.c >> > @@ -35,11 +35,11 @@ __old_setrlimit64 (enum __rlimit_resource resource, >> > { >> > struct rlimit64 krlimits; >> > >> > - if (rlimits->rlim_cur == OLD_RLIM64_INFINITY) >> > + if (rlimits->rlim_cur >= OLD_RLIM64_INFINITY) >> > krlimits.rlim_cur = RLIM64_INFINITY; >> > else >> > krlimits.rlim_cur = rlimits->rlim_cur; >> > - if (rlimits->rlim_max == OLD_RLIM64_INFINITY) >> > + if (rlimits->rlim_max >= OLD_RLIM64_INFINITY) >> > krlimits.rlim_max = RLIM64_INFINITY; >> > else >> > krlimits.rlim_max = rlimits->rlim_max; >> >> Ditto here we want to do the reverse operations as the >> rlim64_to_rlim() >> kernel function. > > I don't quite understand where else the change would go. We don't want > to be touching the generic implementations do we? Or are you saying > this should actually be going in the kernel and not glibc? Hi Aurelien, I know this is a bit of a one-off issue, would you mind elaborating on the comments you left above? Unfortunately it's not clear enough to me where I should go from here.