[PATCH v10 1/8] mm: ksm: Export ksm_madvise()

[Bharata B Rao]

On Thu, Nov 07, 2019 at 04:45:35PM +1100, Paul Mackerras wrote:
> On Wed, Nov 06, 2019 at 12:15:42PM +0530, Bharata B Rao wrote:
> > 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 
> 
> Hmmm, that seems like a bug in the ZONE_DEVICE stuff generally.  All
> that ksm is doing as far as I can see is follow_page() and
> kmap_atomic().  I wonder how many other places in the kernel might
> also be prone to crashing if they try to touch device pages?

In the above shown crash, we don't go via follow_page() and hence
I believe we don't hit the fault path. I see that we come here
after getting the page from get_ksm_page() which returns a device
private page which the subsequent memcmp_pages() does kmap_atomic and
tries to access the address resulting in the above crash.

> 
> > In anycase, we wouldn't want secure guests pages to be pulled out due
> > to KSM, hence disabled merging.
> 
> Sure, I don't disagree with that, but I worry that we are papering
> over a bug here.

Looks like yes. May be someone with better understanding of KSM code
can comment here?

Regards,
Bharata.