projects / packages / rpm.git / blame
| Commit | Line | Data |
|---|---|---|
| 8c28fc02 ER |
1 | http://lists.pld-linux.org/mailman/pipermail/pld-devel-en/2016-May/024910.html |
| 2 | ||
| 3 | patch from jbj answer: | |
| 4 | http://lists.pld-linux.org/mailman/pipermail/pld-devel-en/2016-May/024913.html | |
| 5 | ||
| 6 | On May 31, 2016, at 2:52 AM, Elan Ruusamäe wrote: | |
| 7 | ||
| 8 | > On 31.05.2016 09:45, Elan Ruusamäe wrote: | |
| 9 | >> [root@57c8cccdc671 bin]# poldek -u procps | |
| 10 | > oh, and it's 100% reproducible: | |
| 11 | > | |
| 12 | ||
| 13 | Yes. And it has been discussed many times over the last 15y, | |
| 14 | several times on this list. | |
| 15 | ||
| 16 | Short answer: | |
| 17 | Disable the mechanism with attached patch to lib/psm.c | |
| 18 | - OR -- | |
| 19 | Don't use | |
| 20 | %scriptlet -p /sbin/ldconfig | |
| 21 | in packages that are known to invoke an executable that depends on | |
| 22 | a new library. | |
| 23 | ||
| 24 | Longer answer: | |
| 25 | Libraries often have scriptlets like this: | |
| 26 | %post -p /sbin/ldconfig | |
| 27 | %postun -p /sbin/ldconfig | |
| 28 | (or the equivalent within a scriptlet body). | |
| 29 | ||
| 30 | On upgrade, with install before erase, this results in 2 executions of /sbin/ldconfig. | |
| 31 | ||
| 32 | One of those invocations is redundant when a library is being replaced, the | |
| 33 | install %post is sufficient to make the library avaialble for loading. | |
| 34 | ||
| 35 | The redundant invocation of /sbin/ldconfig was deemed a performance flaw @redhat, | |
| 36 | where I was asked to implement the current optimization. | |
| 37 | ||
| 38 | So the RPM optimization detects when /sbin/ldconfig has just been run, and skips the next | |
| 39 | %scriptlet -p /sbin/ldconfig | |
| 40 | (when written exactly like that) if /sbin/ldconfig has been executed. | |
| 41 | ||
| 42 | Issues are only seen when a scriptlet contains an executable that needs | |
| 43 | a changed API: this is usually a very rare occurrence | |
| 44 | ||
| 45 | There are other flaws with the optimization when dependency loops break the sorted order | |
| 46 | (and/or --noorder is used). , and the fix of running /sbin/ldconfig manually is rather simple | |
| 47 | (even if the problem is obscure). | |
| 48 | ||
| 49 | There is another flaw (as in your report), when the very last package | |
| 50 | has a pending skipped /sbin/ldconfig invocation. | |
| 51 | ||
| 52 | The fix of running /sbin/ldconfig manually is rather simple | |
| 53 | (even if the problem is obscure). | |
| 54 | ||
| 55 | hth | |
| 56 | ||
| 57 | 73 de Jeff | |
| 58 | =================================================================== | |
| 59 | RCS file: /v/rpm/cvs/rpm/lib/psm.c,v | |
| 60 | retrieving revision 2.399.2.20 | |
| 61 | diff -p -u -w -r2.399.2.20 psm.c | |
| 62 | --- ./lib/psm.c 11 May 2015 21:10:16 -0000 2.399.2.20 | |
| 63 | +++ ./lib/psm.c 31 May 2016 15:19:20 -0000 | |
| 64 | @@ -800,7 +800,7 @@ static rpmRC runEmbeddedScript(rpmpsm ps | |
| 65 | static int ldconfig_done = 0; | |
| 66 | ||
| 67 | /*@unchecked@*/ /*@observer@*/ /*@null@*/ | |
| 68 | -static const char * ldconfig_path = "/sbin/ldconfig"; | |
| 69 | +static const char * ldconfig_path = NULL; | |
| 70 | ||
| 71 | /** | |
| 72 | * Run scriptlet with args. | |
| 73 | ||
| 74 |