Hi, David Miller wrote: [Mon Sep 22 2014, 12:22:46AM EDT] > From: Bob Picco <bpicco@xxxxxxxxxx> > Date: Fri, 19 Sep 2014 15:57:31 -0400 > > > Hi, > > David Miller wrote: [Fri Sep 19 2014, 03:28:46PM EDT] > >> From: David Miller <davem@xxxxxxxxxxxxx> > >> Date: Thu, 18 Sep 2014 19:16:37 -0400 (EDT) > >> > >> > I'm half-way done with the "use kernel page tables for everything" > >> > patch, and will post once I have something that at least work for > >> > me. > >> > >> Bob, I ran into a major roadblock which is going to delay this work > >> a little bit. > > Okay. > >> > >> In fact it pointed out a bug we introduced when increasing max phys > >> bits to 47. We broke DEBUG_PAGEALLOC completely. The issue is that > >> the page table layout only supports up to 43-bits. > > Oh bugger yes. I recall testing this when first introducting four level > > page table layout in late 2012 and early 2013. It just isn't an everyday > > thought but VERY nice to have. > >> > >> I'll brain storm about this over the weekend. > > Me too. > > Ok, here is what I have so far and I just got it working. I thought about this issue some over the weekend and PRIQ. Most variants considered for this issue were not appealing. > > The long and short of it is that we have a 4-level page table > but only for the PAGE_OFFSET linear kernel mappings. This is the variant that seems good to me. It should prevent future issues for at least another core revision - me hopes. > > To the rest of the kernel this is transparent, because all > vmalloc et al. mappings have virtual addresses which index > to swapper_pg_dir. Yes, I reviewed most of the code before attempting to use it :) Jalap?no and T5-2 seem happy. From T5-2: [ 0.000000] Kernel: Using 3 locked TLB entries for main kernel image. > > The only thing I have to do is bring back 256MB et al. mapping usage > for suitable aligned PAGE_OFFSET linear areas. It's only using 4MB > mappings at the moment. Okay. I'll review more. Oh and boot with DEBUG_PAGEALLOC at some point. Alright, need to chase a DEBUG_PAGEALLOC issue with prom_halt() blink: [bpicco@zareason ~]$ cat /tmp/t5-blink.txt [23233.728105] IPv6: ADDRCONF(NETDEV_CHANGE): eth1: link becomes ready [23233.753380] SUN4V-DTLB: Error at TPC[446930], tl 1 [23233.762729] SUN4V-DTLB: TPC<dma_4v_map_page+0x10/0x220> [23233.773124] SUN4V-DTLB: O7[55206c] [23233.779910] SUN4V-DTLB: O7<cache_grow+0x12c/0x200> [23233.789455] SUN4V-DTLB: vaddr[1103f23e000] ctx[0] pte[0] error[2] . We were nearly at the console login prompt. Hm, maybe turn off LDOMS config :) thanx and nice! -- To unsubscribe from this list: send the line "unsubscribe sparclinux" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
