Messages in this thread | | | Date | Tue, 6 Sep 2016 14:25:08 -0400 | From | Paul Gortmaker <> | Subject | [ANNOUNCE] 4.8-rc5-rt1 beta |
| |
This is a carry forward of the -rt patches off v4.6.7-rt11 through mainline commits of v4.7 and up to the current v4.8-rc5.
https://git.kernel.org/cgit/linux/kernel/git/paulg/4.8-rt-patches.git
The repository only contains patches; this is aimed at being a stepping stone for developers who are experimenting with -rt and/or patching and building their own kernel trees. End users who have not created their own trees before really should stick with v4.6-rt.
To be explicitly clear -- this does NOT imply any level of support, or any commitment to release an "official" -rt on any particular kernel version by anyone, myself included.
This repo is meant to give more visibility into how the mainline commits since 4.6 will impact the existing 4.6-rt patches. This is achieved by updating the -rt patches in lock step with the new content as it was/is merged by Linus. The branches exist as fast forward work off of Sebastian's 4.6 patch history on the v4.6-rt.
Patch conflicts/issues of interest ----------------------------------
What follows is probably only of interest to a select handful of people who are routinely hands on with the patches. It is a summary of the changes of interest; all of which are part of the above repo history, i.e.
git whatchanged v4.7-rt..v4.8-rt
but ignoring the more basic patch refresh steps with just simple context changes or obvious fixups.
Patch conflicts/issues of interest 4.6 --> 4.7 ----------------------------------------------
-v4.7 introduced a lot of users of down_write_killable which wasn't implemented for -rt (rwsem_rt.h), so I created one. (see new patch rt-create-down_write_killable.patch -- local commit 14625e96ff085f), it probably warrants a look see from other people smarter than me.
-Al Viro added a 2nd user of down_read_non_owner into NFS unlink. Previously it was just bcache and Sebastian just made that depend on !RT_FULL (md-disable-bcache.patch). So we'll need to decide what to with that; I just ignored NFS for now, since it is horrible anyway. (see upstream commit 884be175351e73c)
-there was considerable rummaging around in mm/page_alloc.c that caused the two -rt patches in there to need more hands on refresh work than normal, but everything seems OK AFAICT.
-task_rq_lock was expanded to take a struct rq_flags instead of the conventional lock flags; updated the patches to match (see upstream eb58075149b7f030 by peterz)
-dcache.h gained an implicit include issue; probably can upstream the trivial fix right away; didn't investigate whether there are configs on mainline where it also breaks. (fixes d9171b9345261e ?)
-about eight patches appeared upstream and were droppred from -rt.
-I initially renamed a couple patches to match the default name as given by "git format-patch" but since that prevents ease of viewing the "diff of a patch" as a double check on the patch refresh, I stopped doing that after the 1st two or three patches.
Patch conflicts/issues of interest 4.7 --> 4.8 ----------------------------------------------
-all the timer wheel rework being present in mainline meant we got to drop those patches from the -rt queue; about 21 patches in all. (11b68271256f28999 in this repo.)
-the amount of stuff living in thread_info was being scaled back by luto (see upstream 13d4ea097d18b419) and as part of that the GET_THREAD_INFO macro got nuked. I temporarily reintroduced it close to the use case of x86 preempt lazy (local ID 822cf4179e6f) while we possibly reconsider where TIF_LAZY ends up living.
-another implicit include issue popped up, this time in the header include/linux/rbtree_augmented.h (local patch added in commit 97cd8b0eeed6, fixes upstream commit c1adf20052d80); needs sending.
-the net patch that replaced some per-cpu stuff with a variable in current had macros in dev.c for the RT vs. non-RT case; now filter.c needs those same macros, so they moved to netdevice.h where the other shared macro was already (local ID 8ff5e6ee96b file net-move-xmit_recursion-to-per-task-variable-on-RT.patch)
-I dropped the two scsi-fcoe-* patches; it appears that what they were trying to fix on -rt will be fixed biy upstream 4b9bc86d5a99 but it probably warrants a second look by someone else.
-the ARM at91 clksrc change has been temporarily dropped, after upstream got re-written to incorporate return values etc. It had already been dragged along kicking and screaming from 3.x days and probably needs love from someone with real hardware.
-another 3 patches made it upstream, (not counting ones mentioned above) including the CR3 preempt issue fixed by Sebastian
Process details: ---------------- Rather than jump them forward all at once and then try and deal with all the new issues at once, I walked them forward step by step through the merges performed by Linus in each 2wk window leading up to the rc1 tag (where all the ~12k new commits appear). I've detailed this exact same process in more detail in the past here:
https://lwn.net/Articles/431364/
This 4.8 was done the same way. So when you see two tags like these:
rt-v4.7-3404-ge831101a73fb rt-v4.7-5222-g468fc7ed5537
you can checkout the patches on tag rt-v4.7-3404-ge831101a73fb and know they will apply to any merge authored by Linus that lies between merge v4.7-3404-ge831101a73fb and any Linus merge prior to the merge at v4.7-5222-g468fc7ed5537. This enables one to actually bisect the complete uprev without having to deal with patch failures.
The list of merges is trivially created with something like: git log --oneline --reverse --merges --author=Torvalds v4.7..v4.8-rc1
Sanity boot tested on x86-64 SMP with what is largely a defconfig, but with RT_FULL=y and NFS=n.
| |