x86 fixes for v5.5

index f83ca5aa8b77..f07baf0388bc 100644
@@ -172,7 +172,7 @@
 	ALTERNATIVE "jmp .Lend_\@", "", X86_FEATURE_PTI
 	.if \no_user_check == 0
 	/* coming from usermode? */
-	testl	$SEGMENT_RPL_MASK, PT_CS(%esp)
+	testl	$USER_SEGMENT_RPL_MASK, PT_CS(%esp)
 	jz	.Lend_\@
 	.endif
 	/* On user-cr3? */

@@ -205,64 +205,76 @@
 #define CS_FROM_ENTRY_STACK	(1 << 31)
 #define CS_FROM_USER_CR3	(1 << 30)
 #define CS_FROM_KERNEL		(1 << 29)
+#define CS_FROM_ESPFIX		(1 << 28)
 
 .macro FIXUP_FRAME
 	/*
 	 * The high bits of the CS dword (__csh) are used for CS_FROM_*.
 	 * Clear them in case hardware didn't do this for us.
 	 */
-	andl	$0x0000ffff, 3*4(%esp)
+	andl	$0x0000ffff, 4*4(%esp)
 
 #ifdef CONFIG_VM86
-	testl	$X86_EFLAGS_VM, 4*4(%esp)
+	testl	$X86_EFLAGS_VM, 5*4(%esp)
 	jnz	.Lfrom_usermode_no_fixup_\@
 #endif
-	testl	$SEGMENT_RPL_MASK, 3*4(%esp)
+	testl	$USER_SEGMENT_RPL_MASK, 4*4(%esp)
 	jnz	.Lfrom_usermode_no_fixup_\@
 
-	orl	$CS_FROM_KERNEL, 3*4(%esp)
+	orl	$CS_FROM_KERNEL, 4*4(%esp)
 
 	/*
 	 * When we're here from kernel mode; the (exception) stack looks like:
 	 *
-	 *  5*4(%esp) - <previous context>
-	 *  4*4(%esp) - flags
-	 *  3*4(%esp) - cs
-	 *  2*4(%esp) - ip
-	 *  1*4(%esp) - orig_eax
-	 *  0*4(%esp) - gs / function
+	 *  6*4(%esp) - <previous context>
+	 *  5*4(%esp) - flags
+	 *  4*4(%esp) - cs
+	 *  3*4(%esp) - ip
+	 *  2*4(%esp) - orig_eax
+	 *  1*4(%esp) - gs / function
+	 *  0*4(%esp) - fs
 	 *
 	 * Lets build a 5 entry IRET frame after that, such that struct pt_regs
 	 * is complete and in particular regs->sp is correct. This gives us
-	 * the original 5 enties as gap:
+	 * the original 6 enties as gap:
 	 *
-	 * 12*4(%esp) - <previous context>
-	 * 11*4(%esp) - gap / flags
-	 * 10*4(%esp) - gap / cs
-	 *  9*4(%esp) - gap / ip
-	 *  8*4(%esp) - gap / orig_eax
-	 *  7*4(%esp) - gap / gs / function
-	 *  6*4(%esp) - ss
-	 *  5*4(%esp) - sp
-	 *  4*4(%esp) - flags
-	 *  3*4(%esp) - cs
-	 *  2*4(%esp) - ip
-	 *  1*4(%esp) - orig_eax
-	 *  0*4(%esp) - gs / function
+	 * 14*4(%esp) - <previous context>
+	 * 13*4(%esp) - gap / flags
+	 * 12*4(%esp) - gap / cs
+	 * 11*4(%esp) - gap / ip
+	 * 10*4(%esp) - gap / orig_eax
+	 *  9*4(%esp) - gap / gs / function
+	 *  8*4(%esp) - gap / fs
+	 *  7*4(%esp) - ss
+	 *  6*4(%esp) - sp
+	 *  5*4(%esp) - flags
+	 *  4*4(%esp) - cs
+	 *  3*4(%esp) - ip
+	 *  2*4(%esp) - orig_eax
+	 *  1*4(%esp) - gs / function
+	 *  0*4(%esp) - fs
 	 */
 
 	pushl	%ss		# ss
 	pushl	%esp		# sp (points at ss)
-	addl	$6*4, (%esp)	# point sp back at the previous context
-	pushl	6*4(%esp)	# flags
-	pushl	6*4(%esp)	# cs
-	pushl	6*4(%esp)	# ip
-	pushl	6*4(%esp)	# orig_eax
-	pushl	6*4(%esp)	# gs / function
+	addl	$7*4, (%esp)	# point sp back at the previous context
+	pushl	7*4(%esp)	# flags
+	pushl	7*4(%esp)	# cs
+	pushl	7*4(%esp)	# ip
+	pushl	7*4(%esp)	# orig_eax
+	pushl	7*4(%esp)	# gs / function
+	pushl	7*4(%esp)	# fs
 .Lfrom_usermode_no_fixup_\@:
 .endm
 
 .macro IRET_FRAME
+	/*
+	 * We're called with %ds, %es, %fs, and %gs from the interrupted
+	 * frame, so we shouldn't use them.  Also, we may be in ESPFIX
+	 * mode and therefore have a nonzero SS base and an offset ESP,
+	 * so any attempt to access the stack needs to use SS.  (except for
+	 * accesses through %esp, which automatically use SS.)
+	 */
 	testl $CS_FROM_KERNEL, 1*4(%esp)
 	jz .Lfinished_frame_\@
 

@@ -276,31 +288,40 @@
 	movl	5*4(%esp), %eax		# (modified) regs->sp
 
 	movl	4*4(%esp), %ecx		# flags
-	movl	%ecx, -4(%eax)
+	movl	%ecx, %ss:-1*4(%eax)
 
 	movl	3*4(%esp), %ecx		# cs
 	andl	$0x0000ffff, %ecx
-	movl	%ecx, -8(%eax)
+	movl	%ecx, %ss:-2*4(%eax)
 
 	movl	2*4(%esp), %ecx		# ip
-	movl	%ecx, -12(%eax)
+	movl	%ecx, %ss:-3*4(%eax)
 
 	movl	1*4(%esp), %ecx		# eax
-	movl	%ecx, -16(%eax)
+	movl	%ecx, %ss:-4*4(%eax)
 
 	popl	%ecx
-	lea	-16(%eax), %esp
+	lea	-4*4(%eax), %esp
 	popl	%eax
 .Lfinished_frame_\@:
 .endm
 
-.macro SAVE_ALL pt_regs_ax=%eax switch_stacks=0 skip_gs=0
+.macro SAVE_ALL pt_regs_ax=%eax switch_stacks=0 skip_gs=0 unwind_espfix=0
 	cld
 .if \skip_gs == 0
 	PUSH_GS
 .endif
-	FIXUP_FRAME
 	pushl	%fs
+
+	pushl	%eax
+	movl	$(__KERNEL_PERCPU), %eax
+	movl	%eax, %fs
+.if \unwind_espfix > 0
+	UNWIND_ESPFIX_STACK
+.endif
+	popl	%eax
+
+	FIXUP_FRAME
 	pushl	%es
 	pushl	%ds
 	pushl	\pt_regs_ax

@@ -313,8 +334,6 @@
 	movl	$(__USER_DS), %edx
 	movl	%edx, %ds
 	movl	%edx, %es
-	movl	$(__KERNEL_PERCPU), %edx
-	movl	%edx, %fs
 .if \skip_gs == 0
 	SET_KERNEL_GS %edx
 .endif

@@ -324,8 +343,8 @@
 .endif
 .endm
 
-.macro SAVE_ALL_NMI cr3_reg:req
-	SAVE_ALL
+.macro SAVE_ALL_NMI cr3_reg:req unwind_espfix=0
+	SAVE_ALL unwind_espfix=\unwind_espfix
 
 	BUG_IF_WRONG_CR3
 

@@ -357,6 +376,7 @@
 2:	popl	%es
 3:	popl	%fs
 	POP_GS \pop
+	IRET_FRAME
 .pushsection .fixup, "ax"
 4:	movl	$0, (%esp)
 	jmp	1b

@@ -395,7 +415,8 @@
 
 .macro CHECK_AND_APPLY_ESPFIX
 #ifdef CONFIG_X86_ESPFIX32
-#define GDT_ESPFIX_SS PER_CPU_VAR(gdt_page) + (GDT_ENTRY_ESPFIX_SS * 8)
+#define GDT_ESPFIX_OFFSET (GDT_ENTRY_ESPFIX_SS * 8)
+#define GDT_ESPFIX_SS PER_CPU_VAR(gdt_page) + GDT_ESPFIX_OFFSET
 
 	ALTERNATIVE	"jmp .Lend_\@", "", X86_BUG_ESPFIX
 

@@ -1075,7 +1096,6 @@ restore_all:
 	/* Restore user state */
 	RESTORE_REGS pop=4			# skip orig_eax/error_code
 .Lirq_return:
-	IRET_FRAME
 	/*
 	 * ARCH_HAS_MEMBARRIER_SYNC_CORE rely on IRET core serialization
 	 * when returning from IPI handler and when returning from

@@ -1128,30 +1148,43 @@ ENDPROC(entry_INT80_32)
  * We can't call C functions using the ESPFIX stack. This code reads
  * the high word of the segment base from the GDT and swiches to the
  * normal stack and adjusts ESP with the matching offset.
+ *
+ * We might be on user CR3 here, so percpu data is not mapped and we can't
+ * access the GDT through the percpu segment.  Instead, use SGDT to find
+ * the cpu_entry_area alias of the GDT.
  */
 #ifdef CONFIG_X86_ESPFIX32
 	/* fixup the stack */
-	mov	GDT_ESPFIX_SS + 4, %al /* bits 16..23 */
-	mov	GDT_ESPFIX_SS + 7, %ah /* bits 24..31 */
+	pushl	%ecx
+	subl	$2*4, %esp
+	sgdt	(%esp)
+	movl	2(%esp), %ecx				/* GDT address */
+	/*
+	 * Careful: ECX is a linear pointer, so we need to force base
+	 * zero.  %cs is the only known-linear segment we have right now.
+	 */
+	mov	%cs:GDT_ESPFIX_OFFSET + 4(%ecx), %al	/* bits 16..23 */
+	mov	%cs:GDT_ESPFIX_OFFSET + 7(%ecx), %ah	/* bits 24..31 */
 	shl	$16, %eax
+	addl	$2*4, %esp
+	popl	%ecx
 	addl	%esp, %eax			/* the adjusted stack pointer */
 	pushl	$__KERNEL_DS
 	pushl	%eax
 	lss	(%esp), %esp			/* switch to the normal stack segment */
 #endif
 .endm
+
 .macro UNWIND_ESPFIX_STACK
+	/* It's safe to clobber %eax, all other regs need to be preserved */
 #ifdef CONFIG_X86_ESPFIX32
 	movl	%ss, %eax
 	/* see if on espfix stack */
 	cmpw	$__ESPFIX_SS, %ax
-	jne	27f
-	movl	$__KERNEL_DS, %eax
-	movl	%eax, %ds
-	movl	%eax, %es
+	jne	.Lno_fixup_\@
 	/* switch to normal stack */
 	FIXUP_ESPFIX_STACK
-27:
+.Lno_fixup_\@:
 #endif
 .endm
 

@@ -1341,11 +1374,6 @@ END(spurious_interrupt_bug)
 
 #ifdef CONFIG_XEN_PV
 ENTRY(xen_hypervisor_callback)
-	pushl	$-1				/* orig_ax = -1 => not a system call */
-	SAVE_ALL
-	ENCODE_FRAME_POINTER
-	TRACE_IRQS_OFF
-
 	/*
 	 * Check to see if we got the event in the critical
 	 * region in xen_iret_direct, after we've reenabled

@@ -1353,16 +1381,17 @@ ENTRY(xen_hypervisor_callback)
 	 * iret instruction's behaviour where it delivers a
 	 * pending interrupt when enabling interrupts:
 	 */
-	movl	PT_EIP(%esp), %eax
-	cmpl	$xen_iret_start_crit, %eax
+	cmpl	$xen_iret_start_crit, (%esp)
 	jb	1f
-	cmpl	$xen_iret_end_crit, %eax
+	cmpl	$xen_iret_end_crit, (%esp)
 	jae	1f
-
-	jmp	xen_iret_crit_fixup
-
-ENTRY(xen_do_upcall)
-1:	mov	%esp, %eax
+	call	xen_iret_crit_fixup
+1:
+	pushl	$-1				/* orig_ax = -1 => not a system call */
+	SAVE_ALL
+	ENCODE_FRAME_POINTER
+	TRACE_IRQS_OFF
+	mov	%esp, %eax
 	call	xen_evtchn_do_upcall
 #ifndef CONFIG_PREEMPTION
 	call	xen_maybe_preempt_hcall

@@ -1449,10 +1478,9 @@ END(page_fault)
 
 common_exception_read_cr2:
 	/* the function address is in %gs's slot on the stack */
-	SAVE_ALL switch_stacks=1 skip_gs=1
+	SAVE_ALL switch_stacks=1 skip_gs=1 unwind_espfix=1
 
 	ENCODE_FRAME_POINTER
-	UNWIND_ESPFIX_STACK
 
 	/* fixup %gs */
 	GS_TO_REG %ecx

@@ -1474,9 +1502,8 @@ END(common_exception_read_cr2)
 
 common_exception:
 	/* the function address is in %gs's slot on the stack */
-	SAVE_ALL switch_stacks=1 skip_gs=1
+	SAVE_ALL switch_stacks=1 skip_gs=1 unwind_espfix=1
 	ENCODE_FRAME_POINTER
-	UNWIND_ESPFIX_STACK
 
 	/* fixup %gs */
 	GS_TO_REG %ecx

@@ -1515,6 +1542,10 @@ ENTRY(nmi)
 	ASM_CLAC
 
 #ifdef CONFIG_X86_ESPFIX32
+	/*
+	 * ESPFIX_SS is only ever set on the return to user path
+	 * after we've switched to the entry stack.
+	 */
 	pushl	%eax
 	movl	%ss, %eax
 	cmpw	$__ESPFIX_SS, %ax

@@ -1550,6 +1581,11 @@ ENTRY(nmi)
 	movl	%ebx, %esp
 
 .Lnmi_return:
+#ifdef CONFIG_X86_ESPFIX32
+	testl	$CS_FROM_ESPFIX, PT_CS(%esp)
+	jnz	.Lnmi_from_espfix
+#endif
+
 	CHECK_AND_APPLY_ESPFIX
 	RESTORE_ALL_NMI cr3_reg=%edi pop=4
 	jmp	.Lirq_return

@@ -1557,23 +1593,42 @@ ENTRY(nmi)
 #ifdef CONFIG_X86_ESPFIX32
 .Lnmi_espfix_stack:
 	/*
-	 * create the pointer to lss back
+	 * Create the pointer to LSS back
 	 */
 	pushl	%ss
 	pushl	%esp
 	addl	$4, (%esp)
-	/* copy the iret frame of 12 bytes */
-	.rept 3
-	pushl	16(%esp)
-	.endr
-	pushl	%eax
-	SAVE_ALL_NMI cr3_reg=%edi
+
+	/* Copy the (short) IRET frame */
+	pushl	4*4(%esp)	# flags
+	pushl	4*4(%esp)	# cs
+	pushl	4*4(%esp)	# ip
+
+	pushl	%eax		# orig_ax
+
+	SAVE_ALL_NMI cr3_reg=%edi unwind_espfix=1
 	ENCODE_FRAME_POINTER
-	FIXUP_ESPFIX_STACK			# %eax == %esp
+
+	/* clear CS_FROM_KERNEL, set CS_FROM_ESPFIX */
+	xorl	$(CS_FROM_ESPFIX | CS_FROM_KERNEL), PT_CS(%esp)
+
 	xorl	%edx, %edx			# zero error code
-	call	do_nmi
+	movl	%esp, %eax			# pt_regs pointer
+	jmp	.Lnmi_from_sysenter_stack
+
+.Lnmi_from_espfix:
 	RESTORE_ALL_NMI cr3_reg=%edi
-	lss	12+4(%esp), %esp		# back to espfix stack
+	/*
+	 * Because we cleared CS_FROM_KERNEL, IRET_FRAME 'forgot' to
+	 * fix up the gap and long frame:
+	 *
+	 *  3 - original frame	(exception)
+	 *  2 - ESPFIX block	(above)
+	 *  6 - gap		(FIXUP_FRAME)
+	 *  5 - long frame	(FIXUP_FRAME)
+	 *  1 - orig_ax
+	 */
+	lss	(1+5+6)*4(%esp), %esp			# back to espfix stack
 	jmp	.Lirq_return
 #endif
 END(nmi)

index 30f9cb2c0b55..2e6a0676c1f4 100644
@@ -571,6 +571,16 @@ ENTRY(initial_page_table)
 #  error "Kernel PMDs should be 1, 2 or 3"
 # endif
 	.align PAGE_SIZE		/* needs to be page-sized too */
+
+#ifdef CONFIG_PAGE_TABLE_ISOLATION
+	/*
+	 * PTI needs another page so sync_initial_pagetable() works correctly
+	 * and does not scribble over the data which is placed behind the
+	 * actual initial_page_table. See clone_pgd_range().
+	 */
+	.fill 1024, 4, 0
+#endif
+
 #endif
 
 .data

105 lines not on this page · raw message