On Thu, Nov 07, 2019 at 08:29:40AM +0100, Romain Perier wrote: > Le mer. 30 oct. 2019 à 09:21, Allen <allen.lkml@xxxxxxxxx> a écrit : > > > > Romain, > > > > > > > First of all Romain, nice work. I started working on this > > set a few months back, but could only carve out limited time. > > > > I sent out RFC for this sometime in May[1]. And my approach > > was a little different when compared to what you have sent on the > > list. > > > > Well, I have pushed my work to github[2], only thing I could > > think of as an improvement in your patch set it to break it down > > into smaller chunk so that it's easier to review. I have made each > > occurrence of tasklet_init() into a commit[3] which I thought would > > make it easier to review. I'll leave that decision to you and kees. > > > > Let me know if I could help in any way. > > > > [1] https://www.openwall.com/lists/kernel-hardening/2019/05/06/1 > > [2] https://github.com/allenpais/tasklet > > [3] Sample list of patches: > > 5d0b728649b6 atm/solos-pci: Convert tasklets to use new tasklet_init API > > e5144c3c16d8 atm: Convert tasklets to use new tasklet_init API > > 71028976d3ed arch/um: Convert tasklets to use new tasklet_init API > > c9a39c23b78c xfrm: Convert tasklets to use new tasklet_init API > > 91d93fe12bbc mac80211: Convert tasklets to use new tasklet_init API > > d68f1e9e4531 ipv4: Convert tasklets to use new tasklet_init API > > 4f9379dcd8ad sound/timer: Convert tasklets to use new tasklet_init API > > b4519111b75e drivers/usb: Convert tasklets to use new tasklet_init API > > 52f04bf54a5a drivers:vt/keyboard: Convert tasklets to use new tasklet_init API > > 295de7c9812c dma/virt-dma: Convert tasklets to use new tasklet_init API > > 6c713c83b58f dma/dw: Convert tasklets to use new tasklet_init API > > eaaaaba8a4a7 debug:Convert tasklets to use new tasklet_init API > > b23f4ff5021b tasklet: prepare to change tasklet API > > From experience, this is better to group bunch of commits like we > currently do with Kees on this series, instead to have one commit per > change (I mean for huge patchset) > Mainly because you have too much replacements with this API change, > and it will be really complicated to merge. > > Last time I have proposed an API change for removing "pci_pool" , it > was a patchset of 20 commits (something like this), it tooks 6 months > to be merged :) (with a fine grain granularity on each commit) > > This is better to be the more atomic as possible. If we split the "one > massive tasklet_init replacement" commit into many commit, I am sure > that we find old tasklet API for months in the kernel... it is not > something we want , imho. + treewide commits are common in the kernel > tree, for important API changes :) > > @Kees: agreed ? > > I think that the timer_list approach is good. You can help by > providing feedbacks and by testing if you want. It worked well the last time. :) I think splitting the non-mechanical changes and landing those first is the right approach. Then we can land a massive treewide for all the "easy" cases without trickling them in over months. -Kees -- Kees Cook
