From mboxrd@z Thu Jan 1 00:00:00 1970 From: Petr Vorel Date: Wed, 29 Nov 2017 22:19:27 +0100 Subject: [Buildroot] Cannot rebuild autobuild In-Reply-To: References: <20171115125857.dcy6j7qvjt32ww43@dell5510> <61134b08-bb0a-4665-4b68-3ea2e4063563@mind.be> <20171116001919.67e98b4b@gmx.net> <58b8ef53-75b6-fc8b-8d37-e21570ae4e3b@mind.be> <20171116072948.smctbcvxwkxgqoah@x230> <20171118223206.069b771a@gmx.net> <20171121201605.tzmasudql66iruxa@dell5510> <20171125150144.GC2798@scaer> Message-ID: <20171129211926.GA16373@x230> List-Id: MIME-Version: 1.0 Content-Type: text/plain; charset="us-ascii" Content-Transfer-Encoding: 7bit To: buildroot@busybox.net Hi, > > So, this is a change in behaviour, and should be reported upstream as a > > bug. Will you do so? > As I explained in the extended commit message when I applied the patch fixing > this (b9d2d4cb4eb: "Fix makefile include order by using sort/wildcard.") make > actually has had this behaviour "forever" (since make 3.82). It calls glob with > GLOB_NOSORT. Note that in most of the cases where we didn't already do $(sort > $(wildcard ...)), the order actually doesn't (or shouldn't) matter. For example, > package/*/*.mk hopefully does not depend on the order. In other words, it's very > likely that nobody notices if things get included in the "wrong" order. Even though I understand that the change in 3.82 you pointed out *should* change the behavior, it's different: 3.82 is working. > As far as I could see, only the external-toolchain-package really could cause > any problems. > So my guess is that the bisected commit really has nothing to do with it, it > just happens to trigger the random ordering issue. I don't understand changes in 4fd5672 ("Use Jenkins hash.") [2], but it does it permanently (I tested it several times). That's why I thought it would be good to notify upstream. Now I'm not sure (I haven't reported it yet). > Now that I look at it again, however, it looks like we haven't been able to > adequately explain why this issue suddenly occurred... First of all, > external-toolchain-package has existed for a year now, so if it is truly random > we'd expect someone to have noticed before. Also, the change in Tumbleweed that > Petr pointed to [1] should actually *fix* the issue, because it removes the > NOSORT. Debian and Ubuntu, conversely, have no such patch so the autobuilders > should see this issue (they use external toolchains a lot)... I've been testing buildroot in Debian unstable for few years and in openSUSE Tumbleweed for about one year. I never noticed this, until now. You can see from [3], that the only thing that changed in openSUSE's make is glob-lstat.patch [4]. It actually changes glob, so it can be relevant. I can test it with and without it this week. In Debian I've been testing it probably on 3.81, 4.0 and 4.1. > /me is confused again... > Regards, > Arnout Kind regards, Petr [1] https://build.opensuse.org/package/view_file/Base:System/make/make-sorted-glob.patch?expand=1 [2] https://git.savannah.gnu.org/cgit/make.git/commit/?id=4fd56724ad281498d3c8b27a4b25b4070f6e4e65 [3] https://build.opensuse.org/package/view_file/Base:System/make/make.changes?expand=1 [4] https://build.opensuse.org/package/view_file/Base:System/make/glob-lstat.patch?expand=1