> On May 25, 2018, at 11:23 AM, Jaromír Doleček <[email protected]> > wrote: > > 2018-05-22 12:18 GMT+02:00 Martin Husemann <[email protected]>: >> Here are timing results: > > While 37% speedup is nice, I see there isn't nearly as huge difference > as for my machine. Could you please try with some file which fits > whole in the page cache? > >> Unfortunatley with ubc_direct enabled, it panics quickly: > > There is a comment in uvm_bio.c about some platforms not being able to > write to individual bytes atomically, and hence software having to do > read/modify/write. Could this be somehow causing the machine check on > Alpha?
I don’t think that’s what’s happening here… from the machine check logout frame that Ross Harvey so kindly decoded in the 21272 Core Logic driver, it looks like a bogus bus_space_read_something to a PCI memory region is happening (it’s a Target Abort on a PCI memory read … it’s been a while, but I think I’m interpreting that right…) 32-bit PCI address: address = 0x38d67870, 0x0<No DAC> It would be instructive to see what PCI device has a memory BAR that covers that address. If there isn’t one, then I suspect we’re looking at a memory stomper. > Jaromir -- thorpej
