On Thu, Dec 12, 2024 at 05:00:00PM +0000, David Woodhouse wrote: > > > > Here is the output that I see with that patch applied when rebooting via > > 'systemctl kexec': > > > > RAX=000000010c070001 RBX=0000000000000000 RCX=0000000000000000 RDX=000000047fffb1a0 > > RSI=000000011c444000 RDI=000000011c450002 RBP=ff1cd0424d6e8c00 RSP=ff4178d5c5aebc60 > > R8 =0000000000000000 R9 =000000011c446000 R10=ffffffff909f3e00 R11=0000000000000003 > > R12=0000000000000000 R13=0000000000000001 R14=00000000fee1dead R15=0000000000000000 > > RIP=ff1cd0425c44401c RFL=00010002 [-------] CPL=0 II=0 A20=1 SMM=0 HLT=0 > > ES =0018 0000000000000000 ffffffff 00c09300 DPL=0 DS [-WA] > > CS =0010 0000000000000000 ffffffff 00a09b00 DPL=0 CS64 [-RA] > > SS =0018 0000000000000000 ffffffff 00c09300 DPL=0 DS [-WA] > > DS =0018 0000000000000000 ffffffff 00c09300 DPL=0 DS [-WA] > > FS =0018 0000000000000000 ffffffff 00c09300 DPL=0 DS [-WA] > > GS =0018 0000000000000000 ffffffff 00c09300 DPL=0 DS [-WA] > > LDT=0000 0000000000000000 ffffffff 00c00000 > > TR =0040 fffffe043fedb000 00004087 00008b00 DPL=0 TSS64-busy > > GDT= 0000000000000000 00000000 > > IDT= 0000000000000000 00000000 > > CR0=80050033 CR2=ff1cd0425c4441de CR3=000000011c446000 CR4=00771ef0 > > DR0=0000000000000000 DR1=0000000000000000 DR2=0000000000000000 DR3=0000000000000000 > > DR6=00000000fffe0ff0 DR7=0000000000000400 > > EFER=0000000000000d01 > > Code=41 57 9c 6a 00 9d 0f 20 d8 4c 8b 0d ee 01 00 00 41 0f 22 d9 <48> 89 25 bb 01 00 00 48 89 05 c4 01 00 00 0f 20 c0 48 89 05 b2 01 00 00 41 0f 20 e5 4c 89 > > ... > > > I have attached the output of 'objdump -S'. Please let me know if you > > would like any other information or testing. > > /* Switch to the identity mapped page tables */ > movq %cr3, %rax > e: 0f 20 d8 mov %cr3,%rax > movq kexec_pa_table_page(%rip), %r9 > 11: 4c 8b 0d 00 00 00 00 mov 0x0(%rip),%r9 # 18 <relocate_kernel+0x18> > movq %r9, %cr3 > 18: 41 0f 22 d9 mov %r9,%cr3 > > /* Save %rsp and CRs. */ > movq %rsp, saved_rsp(%rip) > 1c: 48 89 25 00 00 00 00 mov %rsp,0x0(%rip) # 23 <relocate_kernel+0x23> > movq %rax, saved_cr3(%rip) > > > So it's faulting when it tries to write to saved_rsp, which is the > first instruction after it loads the new page tables into %cr3. > > Before loading %cr3, the CPU is running on the original kernel page > tables, It's running from the *virtual* address of the control page, > which in your case is ff1cd0425c444000. In those old page tables, this > control page is read-only. > > But in the newly-loaded page tables, this control page should be read- > write. > > What CPU is this? Shouldn't loading %cr3 have caused the existing TLB > entry to be flushed? The machine I have mainly been reproducing and testing this issue on is an Intel Xeon Gold 6314U but I initially noticed this problem on a few consumer AMD and Intel chips that I have locally. > Can you tell me what happens if we don't write through *that* virtual > mapping of the page, but instead do it a few instructions later through > the identity mapping of the same page (which surely *won't* have an > older, non-writeable TLB entry already...)? Applying that diff on top of 5a82223e0743 makes the issue go away for me. > diff --git a/arch/x86/kernel/relocate_kernel_64.S b/arch/x86/kernel/relocate_kernel_64.S > index 8bc86a1e056a..2763c509e513 100644 > --- a/arch/x86/kernel/relocate_kernel_64.S > +++ b/arch/x86/kernel/relocate_kernel_64.S > @@ -70,6 +70,18 @@ SYM_CODE_START_NOALIGN(relocate_kernel) > movq kexec_pa_table_page(%rip), %r9 > movq %r9, %cr3 > > + /* setup a new stack at the end of the physical control page */ > + lea PAGE_SIZE(%rsi), %rsp > + > + /* jump to identity mapped page */ > + addq $(identity_mapped - relocate_kernel), %rsi > + ANNOTATE_RETPOLINE_SAFE > + jmp *%rsi > + int3 > +SYM_CODE_END(relocate_kernel) > + > +SYM_CODE_START_LOCAL_NOALIGN(identity_mapped) > + UNWIND_HINT_END_OF_STACK > /* Save %rsp and CRs. */ > movq %rsp, saved_rsp(%rip) > movq %rax, saved_cr3(%rip) > @@ -85,19 +97,6 @@ SYM_CODE_START_NOALIGN(relocate_kernel) > /* Save the preserve_context to %r11 as swap_pages clobbers %rcx. */ > movq %rcx, %r11 > > - /* setup a new stack at the end of the physical control page */ > - lea PAGE_SIZE(%rsi), %rsp > - > - /* jump to identity mapped page */ > - addq $(identity_mapped - relocate_kernel), %rsi > - pushq %rsi > - ANNOTATE_UNRET_SAFE > - ret > - int3 > -SYM_CODE_END(relocate_kernel) > - > -SYM_CODE_START_LOCAL_NOALIGN(identity_mapped) > - UNWIND_HINT_END_OF_STACK > /* > * %rdi indirection page > * %rdx start address >
