Skip to contentSkip to conversation

new ia64 kernel patch (relative to 2.5.69)

Linux-ia64 David Mosberger 1 participant 2 messages May 10, 2003 – May 12, 2003

lore
Share & tools
lore
#0 lore raw

The kernel patch linux-2.5.69-ia64-030509.diff.gz is now available at ftp.kernel.org/pub/linux/kernel/ports/ia64/v2.5/. See the bitkeeper pages for a detailed log.

This is a bit of a temporary patch because the ia32 subsystem is currently broken (because of the ioctl changes; Arun, would you mind taking a look?). Other than that, things seem to work quite well for me. Apart from the simulator, I tried Big Sur, zx2000, zx6000, and rx5670 (4-way Itanium 2). The AGP GART stuff seems to be working for both zx1 and i460, but I couldn't get DRM quite to work yet. Some sort of locking problem it appears.

Oh, Keith, you might like the fact that I added NEW_LOCK back again and it's even turned on by default! I talked to Cary Coutant and he thinks it's OK to treat a save of ar.pfs in r0 as a special case. Thus, we can use your scheme for now. Also, I wrote a patch for gcc3.4 which makes it possible to mark ar.pfs clobbered (subsequently, Rich Henderson fixed and improved the patch which is now in the official gcc cvs tree). With this, we can generate an extremely light-weight special call to the spinlock contention routine. I tried all 4 versions (old compiler, new compiler, merced vs mckinley) and they all seemed to work fine.

At the moment, the spinlock contention routines do NOT do any backoff. Someone with access to large machines might want to play with this. I'd be very interested in hearing what kind of performance difference it makes.

I hope to make an updated patch sometime next week (hopefully with the ia32 subsystem working again). So if I goofed somewhere else, send patches.

Oh, and just a friendly reminder that Linus is making noises about 2.6. So if you procrastinated with 2.5 so far, now is the time to get going!

	--david
#1 lore raw

Forgot to mention that I switched the ia64 tree from the PCI-DMA API to the generic DMA API. Doesn't affect device drivers, but all I/O MMU support needs to be updated accordingly. I already did swiotlb and the HP SBA I/O MMU support. I updated that machvec definitions for all platforms, but didn't try to update the SN-specific implementation (since I can't test it anyhow). Updating the SN-files should be straight-forward: you can look at swiotlb.c and hp/common/sba_iommu.c for what needs to be done.

The one caveat is that sba_iommu.c for the moment continues to support only PCI-devices. This shouldn't be a problem for now, but it would be nice if someone could fix it so it can work with any device.

While working on this, I started to wonder whether or not a DMA "sync" operation should imply a memory barrier. The API documentation is definitely ambiguous on this point and I didn't go out to look whether current practice is consistent (I suspect it isn't). Anyone want to argue for or against making "sync" a memory barrier?

	--david