On Fri, Nov 30, 2018 at 02:44:54PM +0100, Vitaly Kuznetsov wrote: > Roman Kagan <rkagan@xxxxxxxxxxxxx> writes: > > > On Fri, Nov 30, 2018 at 01:15:11PM +0100, Vitaly Kuznetsov wrote: > >> Without 'packed' compiler is free to add optimization paddings and re-order > >> structure fields for randomization/optimization. And structures from > >> hyperv-tlfs.h are used for hypervisor-guest communication, we need to > >> ultimately forbid such practices. > > > > Note that __packed also reduces the structure alignment to 1, which is > > not necessarily what you want. > > > > E.g. some of these structures are passed by pointer to the hypercall, > > which requires its arguments to be 8byte-aligned. > > Hm, > > I thought we always take precautions for Hyper-V hypercall arguments, in > particular > > PV IPI/TLB flush use pre-allocated hyperv_pcpu_input_arg, > hv_post_message() uses pre-allocated message page, other call sites use > fast hypercalls where we use registers. Looks so indeed. > I also checked this patch before sending out, WS2016 guest boots without > issues. Any particular places you're worried about? It's Linux guests on Hyper-V that need to be checked. > > I'm also not sure > > that passing unaligned argument to [rw]msr is ok, need to > > double-check. > > My understanding is that rdmsr/wrmsr instuctions are registers-only. Indeed. > We can, of course, just add __aligned(8) to some structures but I'd like > to find the reason first. I guess you're right, in the current code the relevant structures are made sufficiently aligned by other means. False alarm, sorry. Roman.
