On Wed, Nov 06, 2019 at 03:33:29PM +1100, Paul Mackerras wrote: > On Mon, Nov 04, 2019 at 09:47:53AM +0530, Bharata B Rao wrote: > > KVM PPC module needs ksm_madvise() for supporting secure guests. > > Guest pages that become secure are represented as device private > > pages in the host. Such pages shouldn't participate in KSM merging. > > If we don't do the ksm_madvise call, then as far as I can tell, it > should all still work correctly, but we might have KSM pulling pages > in unnecessarily, causing a reduction in performance. Is that right? I thought so too. When KSM tries to merge a secure page, it should cause a fault resulting in page-out the secure page. However I see the below crash when KSM is enabled and KSM scan tries to kmap and memcmp the device private page. BUG: Unable to handle kernel data access at 0xc007fffe00010000 Faulting instruction address: 0xc0000000000ab5a0 Oops: Kernel access of bad area, sig: 11 [#1] LE PAGE_SIZE=64K MMU=Radix MMU=Hash SMP NR_CPUS=2048 NUMA PowerNV Modules linked in: CPU: 0 PID: 22 Comm: ksmd Not tainted 5.4.0-rc2-00026-g2249c0ae4a53-dirty #376 NIP: c0000000000ab5a0 LR: c0000000003d7c3c CTR: 0000000000000004 REGS: c0000001c85d79b0 TRAP: 0300 Not tainted (5.4.0-rc2-00026-g2249c0ae4a53-dirty) MSR: 900000000280b033 <SF,HV,VEC,VSX,EE,FP,ME,IR,DR,RI,LE> CR: 24002242 XER: 20040000 CFAR: c0000000000ab3d0 DAR: c007fffe00010000 DSISR: 40000000 IRQMASK: 0 GPR00: 0000000000000004 c0000001c85d7c40 c0000000018ce000 c0000001c3880000 GPR04: c007fffe00010000 0000000000010000 0000000000000000 ffffffffffffffff GPR08: c000000001992298 0000603820002138 ffffffffffffffff ffffffff00003a69 GPR12: 0000000024002242 c000000002550000 c0000001c8700000 c00000000179b728 GPR16: c00c01ffff800040 c00000000179b5b8 c00c00000070e200 ffffffffffffffff GPR20: 0000000000000000 0000000000000000 fffffffffffff000 c00000000179b648 GPR24: c0000000024464a0 c00000000249f568 c000000001118918 0000000000000000 GPR28: c0000001c804c590 c00000000249f518 0000000000000000 c0000001c8700000 NIP [c0000000000ab5a0] memcmp+0x320/0x6a0 LR [c0000000003d7c3c] memcmp_pages+0x8c/0xe0 Call Trace: [c0000001c85d7c40] [c0000001c804c590] 0xc0000001c804c590 (unreliable) [c0000001c85d7c70] [c0000000004591d0] ksm_scan_thread+0x960/0x21b0 [c0000001c85d7db0] [c0000000001bf328] kthread+0x198/0x1a0 [c0000001c85d7e20] [c00000000000bfbc] ret_from_kernel_thread+0x5c/0x80 Instruction dump: ebc1fff0 eba1ffe8 eb81ffe0 eb61ffd8 4e800020 38600001 4d810020 3860ffff 4e800020 38000004 7c0903a6 7d201c28 <7d402428> 7c295040 38630008 38840008 In anycase, we wouldn't want secure guests pages to be pulled out due to KSM, hence disabled merging. Regards, Bharata.
