From: Grant Grundler Date: Sat, 22 Oct 2005 02:37:20 +0000 (-0400) Subject: [PARISC] Update sba_iommu from parisc tree X-Git-Tag: v2.6.15-rc1~732^2^2~9^2~50 X-Git-Url: http://git.openpandora.org/cgi-bin/gitweb.cgi?a=commitdiff_plain;h=64908ad95c34f25849412d6d4735ac10f8fb6575;p=pandora-kernel.git [PARISC] Update sba_iommu from parisc tree revert use of %%sr0 in fdc asm. Thanks to Joel Soete for pointing out this oversight. Signed-off-by: Grant Grundler 2.6.14-rc2-pa3 move "sync" outside the main loop that fills IO Pdir. Signed-off-by: Grant Grundler remove explicit use of sr0 in fdc ops. Thanks to Joel Soete for reminding me were I added those... Signed-off-by: Grant Grundler 2.6.14-rc2-pa2 - make SBA more anal about invalidating pdir entries Previous code cleared the valid flag a pdir entry but it did NOT guarantee this change was visible to the PDIR before writing the PCOM register. Ie the SBA could pick up a stale entry if the write happened to hit the SBA before the cacheline was flushed from the cache. Long term, I think I want to make this a compile time flag. Developement tree should enable anal pdir checking by default and Debian can disable it with either a CONFIG option or one-line patch. fdc/sync options can only negatively affect performance though I haven't measure how much yet. If someone can run netperf TCP_RR across gige and compare -pa1 and -pa2, that would be sufficient. Cleaned up the use of "fdc" to make sure it's using "kernel" space id (specify sr0 but maps to sr4-7). It seems a bit fragile to assume "sr1" gets loaded with KERNEL_SPACE which is how the code works today. Tested on 32 and 64-bit SMP kernels on j6k. Signed-off-by: Grant Grundler remove PDC_NARROW from SBA and document history of PDC_NARROW a bit. It will still show up in an older kernel's .config file. Signed-off-by: Grant Grundler if/ifdef cleanups from Joel Soete. Signed-off-by: Grant Grundler 2.6.12-rc4-pa2 fix 32-bit support for Astro platforms o Since my last SBA code change, SBA could allocate more than 1GB of IOVA space on Astro boxes with more than 1GB of RAM when running 32-bit kernel. This is bad since IOMMU can only talk to the first 1GB at most. Kudos to jejb for quickly spotting that bug. o jejb also noted SBA should *always* reject DMA masks > 32-bits since DMA-mapping.txt indicates caller should try again with 32-bits. o off-by-one error when comparing the mask to IOVA space size. Signed-off-by: Grant Grundler Signed-off-by: Kyle McMartin --- Reading git-diff-tree failed