On Mon, Jan 12, 2015 at 11:15:38AM -0500, Vivek Goyal wrote: > On Mon, Jan 12, 2015 at 05:06:46PM +0100, Joerg Roedel wrote: > > On Mon, Jan 12, 2015 at 10:29:19AM -0500, Vivek Goyal wrote: > > > Kdump has the notion of backup region. Where certain parts of old kernels > > > memory can be moved to a different location (first 640K on x86 as of now) > > > and new kernel can make use of this memory now. > > > > > > So we will have to just make sure that no parts of this old page table > > > fall into backup region. > > > > Uuh, looks like the 'iommu-with-kdump-issue' isn't complicated enough > > yet ;) > > Sadly, your above statement is true for all hardware-accessible data > > structures in IOMMU code. I think about how we can solve this, is there > > an easy way to allocate memory that is not in any backup region? > > Hmm..., there does not seem to be any easy way to do this. In fact, as of > now, kernel does not even know where is backup region. All these details are > managed by user space completely (except for new kexec_file_load() syscall). > > That means we are left with ugly options now. > > - Define per arch kexec backup regions in kernel and export it to user > space and let kexec-tools make use of that deinition (instead of > defining its own). That way memory allocation code in kernel can look > at this backup area and skip it for certain allocations. Yes, that makes sense. In fact, I think all allocations for DMA memory need to take this into account to avoid potentially serious data corruption. If any memory for a disk superblock gets allocated in backup memory and a kdump happens, the new kernel might zero out that area and the disk controler then writes the zeroes to disk instead of the superblock. Joerg