On Thu, Mar 1, 2012 at 7:09 AM, Vivek Goyal <vgoyal at redhat.com> wrote: > On Wed, Feb 29, 2012 at 09:21:23AM -0800, Eugene Surovegin wrote: >> Per-CPU allocations are not guaranteed to be physically contiguous. >> However, kdump kernel and user-space code assumes that per-CPU >> memory, used for saving CPU registers on crash, is. >> This can cause corrupted /proc/vmcore in some cases - the main >> symptom being huge ELF note section. >> >> Force page alignment for note_buf_t to ensure that this assumption holds. >> > > Hi Eugene, > > Where do we make assumption that crash_notes address is page aligned? I never said that. I said that assumption was that note chunks were physically contiguous. [snip] > Are you referring to the fact that note could be stored on two different > physical pages and then kexec-tools does not know about the physical > address of second page, hence second kernel does not know about it > and we never read that data. That sounds like a problem. Yes. This is exactly what happens in some cases. > So it make sense to force the page size alignment and if notes ever > grow beyond 1 page, we need to come up with more complex ways of > communicating more than 1 physical address to second kernel. Yes. Thankfully notes so far are smaller than a page. And if they do grow larger, than kexec-tools have to be modified anyways - currently maximum note size (1K) is hardcoded there. If this happens, I think instead of coming up with a way to communicate phys to virt mapping, it'd be easier just to change th way notes are allocated in the kernel so they are phys contiguous. -- Eugene
