On 13/11/19 19:10, Nadav Amit wrote: > >> On Nov 12, 2019, at 1:21 PM, Paolo Bonzini <pbonzini@xxxxxxxxxx> wrote: >> >> CVE-2018-12207 is a microarchitectural implementation issue >> that could allow an unprivileged local attacker to cause system wide >> denial-of-service condition. >> >> Privileged software may change the page size (ex. 4KB, 2MB, 1GB) in the >> paging structures, without following such paging structure changes with >> invalidation of the TLB entries corresponding to the changed pages. In >> this case, the attacker could invoke instruction fetch, which will result >> in the processor hitting multiple TLB entries, reporting a machine check >> error exception, and ultimately hanging the system. >> >> The attached patches mitigate the vulnerability by making huge pages >> non-executable. The processor will not be able to execute an instruction >> residing in a large page (ie. 2MB, 1GB, etc.) without causing a trap into >> the host kernel/hypervisor; KVM will then break the large page into 4KB >> pages and gives executable permission to 4KB pages. > > It sounds that this mitigation will trigger the “page fracturing” problem > I once encountered [1], causing frequent full TLB flushes when invlpg > runs. I wonder if VMs would benefit in performance from changing > /sys/kernel/debug/x86/tlb_single_page_flush_ceiling to zero. > > On a different note - I am not sure I fully understand the exact scenario. > Any chance of getting a kvm-unit-test for this case? No, for now I only have a test that causes lots of 2 MiB pages to be shattered to 4 KiB. But I wanted to get and post a test case in the next week. Paolo > [1] https://patchwork.kernel.org/patch/9099311/ >
