ARM: vfp: Reimplement VFP exception entry in C code

En/disabling softirqs from asm code turned out to be trickier than
expected, so vfp_support_entry now returns by tail calling
__local_enable_bh_ip() and passing the same arguments that a C call to
local_bh_enable() would pass. However, this is slightly hacky, as we
don't want to carry our own implementation of local_bh_enable().

So let's bite the bullet, and get rid of the asm logic in
vfp_support_entry that reasons about whether or not to save and/or
reload the VFP state, and about whether or not an FP exception is
pending, and only keep the VFP loading logic as a function that is
callable from C.

Replicate the removed logic in vfp_entry(), and use the exact same
reasoning as in the asm code. To emphasize the correspondence, retain
some of the asm comments in the C version as well.

Signed-off-by: Ard Biesheuvel <[e-mail]>
Acked-by: Linus Walleij <[e-mail]>

arch/arm/vfp/entry.S

···
 @  IRQs enabled.
 @
 ENTRY(do_vfp)
-	mov	r1, r10
-	str	lr, [sp, #-8]!
-	add	r3, sp, #4
-	str	r9, [r3]
-	bl	vfp_entry
-	ldr	pc, [sp], #8
+	mov	r1, r0				@ pass trigger opcode via R1
+	mov	r0, sp				@ pass struct pt_regs via R0
+	bl	vfp_support_entry		@ dispatch the VFP exception
+	cmp	r0, #0				@ handled successfully?
+	reteq	r9				@ then use R9 as return address
+	ret	lr				@ pass to undef handler
 ENDPROC(do_vfp)

arch/arm/vfp/vfp.h

···
 };
 
 asmlinkage void vfp_save_state(void *location, u32 fpexc);
+asmlinkage u32 vfp_load_state(const void *location);

arch/arm/vfp/vfphw.S

···
  *
  *  Copyright (C) 2004 ARM Limited.
  *  Written by Deep Blue Solutions Limited.
- *
- * This code is called from the kernel's undefined instruction trap.
- * r1 holds the thread_info pointer
- * r3 holds the return address for successful handling.
- * lr holds the return address for unrecognised instructions.
- * sp points to a struct pt_regs (as defined in include/asm/proc/ptrace.h)
  */
 #include <linux/init.h>
 #include <linux/linkage.h>
···
 #include <linux/kern_levels.h>
 #include <asm/assembler.h>
 #include <asm/asm-offsets.h>
-
-	.macro	DBGSTR, str
-#ifdef DEBUG
-	stmfd	sp!, {r0-r3, ip, lr}
-	ldr	r0, =1f
-	bl	_printk
-	ldmfd	sp!, {r0-r3, ip, lr}
-
-	.pushsection .rodata, "a"
-1:	.ascii	KERN_DEBUG "VFP: \str\n"
-	.byte	0
-	.previous
-#endif
-	.endm
 
 	.macro  DBGSTR1, str, arg
 #ifdef DEBUG
···
 #endif
 	.endm
 
-	.macro  DBGSTR3, str, arg1, arg2, arg3
-#ifdef DEBUG
-	stmfd	sp!, {r0-r3, ip, lr}
-	mov	r3, \arg3
-	mov	r2, \arg2
-	mov	r1, \arg1
-	ldr	r0, =1f
-	bl	_printk
-	ldmfd	sp!, {r0-r3, ip, lr}
-
-	.pushsection .rodata, "a"
-1:	.ascii	KERN_DEBUG "VFP: \str\n"
-	.byte	0
-	.previous
-#endif
-	.endm
-
-
-@ VFP hardware support entry point.
-@
-@  r0  = instruction opcode (32-bit ARM or two 16-bit Thumb)
-@  r1  = thread_info pointer
-@  r2  = PC value to resume execution after successful emulation
-@  r3  = normal "successful" return address
-@  lr  = unrecognised instruction return address
-@  IRQs enabled.
-ENTRY(vfp_support_entry)
-	ldr	r11, [r1, #TI_CPU]	@ CPU number
-	add	r10, r1, #TI_VFPSTATE	@ r10 = workspace
-
-	DBGSTR3	"instr %08x pc %08x state %p", r0, r2, r10
-
-	.fpu	vfpv2
-	VFPFMRX	r1, FPEXC		@ Is the VFP enabled?
-	DBGSTR1	"fpexc %08x", r1
-	tst	r1, #FPEXC_EN
-	bne	look_for_VFP_exceptions	@ VFP is already enabled
-
-	DBGSTR1 "enable %x", r10
-	ldr	r9, vfp_current_hw_state_address
-	orr	r1, r1, #FPEXC_EN	@ user FPEXC has the enable bit set
-	ldr	r4, [r9, r11, lsl #2]	@ vfp_current_hw_state pointer
-	bic	r5, r1, #FPEXC_EX	@ make sure exceptions are disabled
-	cmp	r4, r10			@ this thread owns the hw context?
-#ifndef CONFIG_SMP
-	@ For UP, checking that this thread owns the hw context is
-	@ sufficient to determine that the hardware state is valid.
-	beq	vfp_hw_state_valid
-
-	@ On UP, we lazily save the VFP context.  As a different
-	@ thread wants ownership of the VFP hardware, save the old
-	@ state if there was a previous (valid) owner.
-
-	VFPFMXR	FPEXC, r5		@ enable VFP, disable any pending
-					@ exceptions, so we can get at the
-					@ rest of it
-
-	DBGSTR1	"save old state %p", r4
-	cmp	r4, #0			@ if the vfp_current_hw_state is NULL
-	beq	vfp_reload_hw		@ then the hw state needs reloading
-	VFPFSTMIA r4, r5		@ save the working registers
-	VFPFMRX	r5, FPSCR		@ current status
-	tst	r1, #FPEXC_EX		@ is there additional state to save?
-	beq	1f
-	VFPFMRX	r6, FPINST		@ FPINST (only if FPEXC.EX is set)
-	tst	r1, #FPEXC_FP2V		@ is there an FPINST2 to read?
-	beq	1f
-	VFPFMRX	r8, FPINST2		@ FPINST2 if needed (and present)
-1:
-	stmia	r4, {r1, r5, r6, r8}	@ save FPEXC, FPSCR, FPINST, FPINST2
-vfp_reload_hw:
-
-#else
-	@ For SMP, if this thread does not own the hw context, then we
-	@ need to reload it.  No need to save the old state as on SMP,
-	@ we always save the state when we switch away from a thread.
-	bne	vfp_reload_hw
-
-	@ This thread has ownership of the current hardware context.
-	@ However, it may have been migrated to another CPU, in which
-	@ case the saved state is newer than the hardware context.
-	@ Check this by looking at the CPU number which the state was
-	@ last loaded onto.
-	ldr	ip, [r10, #VFP_CPU]
-	teq	ip, r11
-	beq	vfp_hw_state_valid
-
-vfp_reload_hw:
-	@ We're loading this threads state into the VFP hardware. Update
-	@ the CPU number which contains the most up to date VFP context.
-	str	r11, [r10, #VFP_CPU]
-
-	VFPFMXR	FPEXC, r5		@ enable VFP, disable any pending
-					@ exceptions, so we can get at the
-					@ rest of it
-#endif
-
-	DBGSTR1	"load state %p", r10
-	str	r10, [r9, r11, lsl #2]	@ update the vfp_current_hw_state pointer
+ENTRY(vfp_load_state)
+	@ Load the current VFP state
+	@ r0 - load location
+	@ returns FPEXC
+	DBGSTR1	"load VFP state %p", r0
 					@ Load the saved state back into the VFP
-	VFPFLDMIA r10, r5		@ reload the working registers while
+	VFPFLDMIA r0, r1		@ reload the working registers while
 					@ FPEXC is in a safe state
-	ldmia	r10, {r1, r5, r6, r8}	@ load FPEXC, FPSCR, FPINST, FPINST2
-	tst	r1, #FPEXC_EX		@ is there additional state to restore?
+	ldmia	r0, {r0-r3}		@ load FPEXC, FPSCR, FPINST, FPINST2
+	tst	r0, #FPEXC_EX		@ is there additional state to restore?
 	beq	1f
-	VFPFMXR	FPINST, r6		@ restore FPINST (only if FPEXC.EX is set)
-	tst	r1, #FPEXC_FP2V		@ is there an FPINST2 to write?
+	VFPFMXR	FPINST, r2		@ restore FPINST (only if FPEXC.EX is set)
+	tst	r0, #FPEXC_FP2V		@ is there an FPINST2 to write?
 	beq	1f
-	VFPFMXR	FPINST2, r8		@ FPINST2 if needed (and present)
+	VFPFMXR	FPINST2, r3		@ FPINST2 if needed (and present)
 1:
-	VFPFMXR	FPSCR, r5		@ restore status
-
-@ The context stored in the VFP hardware is up to date with this thread
-vfp_hw_state_valid:
-	tst	r1, #FPEXC_EX
-	bne	process_exception	@ might as well handle the pending
-					@ exception before retrying branch
-					@ out before setting an FPEXC that
-					@ stops us reading stuff
-	VFPFMXR	FPEXC, r1		@ Restore FPEXC last
-	mov	sp, r3			@ we think we have handled things
-	pop	{lr}
-	sub	r2, r2, #4		@ Retry current instruction - if Thumb
-	str	r2, [sp, #S_PC]		@ mode it's two 16-bit instructions,
-					@ else it's one 32-bit instruction, so
-					@ always subtract 4 from the following
-					@ instruction address.
-
-local_bh_enable_and_ret:
-	adr	r0, .
-	mov	r1, #SOFTIRQ_DISABLE_OFFSET
-	b	__local_bh_enable_ip	@ tail call
-
-look_for_VFP_exceptions:
-	@ Check for synchronous or asynchronous exception
-	tst	r1, #FPEXC_EX | FPEXC_DEX
-	bne	process_exception
-	@ On some implementations of the VFP subarch 1, setting FPSCR.IXE
-	@ causes all the CDP instructions to be bounced synchronously without
-	@ setting the FPEXC.EX bit
-	VFPFMRX	r5, FPSCR
-	tst	r5, #FPSCR_IXE
-	bne	process_exception
-
-	tst	r5, #FPSCR_LENGTH_MASK
-	beq	skip
-	orr	r1, r1, #FPEXC_DEX
-	b	process_exception
-skip:
-
-	@ Fall into hand on to next handler - appropriate coproc instr
-	@ not recognised by VFP
-
-	DBGSTR	"not VFP"
-	b	local_bh_enable_and_ret
-
-process_exception:
-	DBGSTR	"bounce"
-	mov	sp, r3			@ setup for a return to the user code.
-	pop	{lr}
-	mov	r2, sp			@ nothing stacked - regdump is at TOS
-
-	@ Now call the C code to package up the bounce to the support code
-	@   r0 holds the trigger instruction
-	@   r1 holds the FPEXC value
-	@   r2 pointer to register dump
-	b	VFP_bounce		@ we have handled this - the support
-					@ code will raise an exception if
-					@ required. If not, the user code will
-					@ retry the faulted instruction
-ENDPROC(vfp_support_entry)
+	VFPFMXR	FPSCR, r1		@ restore status
+	ret	lr
+ENDPROC(vfp_load_state)
 
 ENTRY(vfp_save_state)
 	@ Save the current VFP state
···
 	stmia	r0, {r1, r2, r3, r12}	@ save FPEXC, FPSCR, FPINST, FPINST2
 	ret	lr
 ENDPROC(vfp_save_state)
-
-	.align
-vfp_current_hw_state_address:
-	.word	vfp_current_hw_state
 
 	.macro	tbl_branch, base, tmp, shift
 #ifdef CONFIG_THUMB2_KERNEL

arch/arm/vfp/vfpmodule.c

···
 #include "vfpinstr.h"
 #include "vfp.h"
 
-/*
- * Our undef handlers (in entry.S)
- */
-asmlinkage void vfp_support_entry(u32, void *, u32, u32);
-
 static bool have_vfp __ro_after_init;
 
 /*
···
 /*
  * Package up a bounce condition.
  */
-void VFP_bounce(u32 trigger, u32 fpexc, struct pt_regs *regs)
+static void VFP_bounce(u32 trigger, u32 fpexc, struct pt_regs *regs)
 {
 	u32 fpscr, orig_fpscr, fpsid, exceptions;
 
···
 		 * on VFP subarch 1.
 		 */
 		 vfp_raise_exceptions(VFP_EXCEPTION_ERROR, trigger, fpscr, regs);
-		goto exit;
+		return;
 	}
 
 	/*
···
 	 * the FPEXC.FP2V bit is valid only if FPEXC.EX is 1.
 	 */
 	if ((fpexc & (FPEXC_EX | FPEXC_FP2V)) != (FPEXC_EX | FPEXC_FP2V))
-		goto exit;
+		return;
 
 	/*
 	 * The barrier() here prevents fpinst2 being read
···
 	exceptions = vfp_emulate_instruction(trigger, orig_fpscr, regs);
 	if (exceptions)
 		vfp_raise_exceptions(exceptions, trigger, orig_fpscr, regs);
- exit:
-	local_bh_enable();
 }
 
 static void vfp_enable(void *unused)
···
 }
 
 /*
- * Entered with:
+ * vfp_support_entry - Handle VFP exception from user mode
  *
- *  r0  = instruction opcode (32-bit ARM or two 16-bit Thumb)
- *  r1  = thread_info pointer
- *  r2  = PC value to resume execution after successful emulation
- *  r3  = normal "successful" return address
- *  lr  = unrecognised instruction return address
+ * @regs:	pt_regs structure holding the register state at exception entry
+ * @trigger:	The opcode of the instruction that triggered the exception
+ *
+ * Returns 0 if the exception was handled, or an error code otherwise.
  */
-asmlinkage void vfp_entry(u32 trigger, struct thread_info *ti, u32 resume_pc,
-			  u32 resume_return_address)
+asmlinkage int vfp_support_entry(struct pt_regs *regs, u32 trigger)
 {
+	struct thread_info *ti = current_thread_info();
+	u32 fpexc;
+
 	if (unlikely(!have_vfp))
-		return;
+		return -ENODEV;
 
 	local_bh_disable();
-	vfp_support_entry(trigger, ti, resume_pc, resume_return_address);
+	fpexc = fmrx(FPEXC);
+
+	/*
+	 * If the VFP unit was not enabled yet, we have to check whether the
+	 * VFP state in the CPU's registers is the most recent VFP state
+	 * associated with the process. On UP systems, we don't save the VFP
+	 * state eagerly on a context switch, so we may need to save the
+	 * VFP state to memory first, as it may belong to another process.
+	 */
+	if (!(fpexc & FPEXC_EN)) {
+		/*
+		 * Enable the VFP unit but mask the FP exception flag for the
+		 * time being, so we can access all the registers.
+		 */
+		fpexc |= FPEXC_EN;
+		fmxr(FPEXC, fpexc & ~FPEXC_EX);
+
+		/*
+		 * Check whether or not the VFP state in the CPU's registers is
+		 * the most recent VFP state associated with this task. On SMP,
+		 * migration may result in multiple CPUs holding VFP states
+		 * that belong to the same task, but only the most recent one
+		 * is valid.
+		 */
+		if (!vfp_state_in_hw(ti->cpu, ti)) {
+			if (!IS_ENABLED(CONFIG_SMP) &&
+			    vfp_current_hw_state[ti->cpu] != NULL) {
+				/*
+				 * This CPU is currently holding the most
+				 * recent VFP state associated with another
+				 * task, and we must save that to memory first.
+				 */
+				vfp_save_state(vfp_current_hw_state[ti->cpu],
+					       fpexc);
+			}
+
+			/*
+			 * We can now proceed with loading the task's VFP state
+			 * from memory into the CPU registers.
+			 */
+			fpexc = vfp_load_state(&ti->vfpstate);
+			vfp_current_hw_state[ti->cpu] = &ti->vfpstate;
+#ifdef CONFIG_SMP
+			/*
+			 * Record that this CPU is now the one holding the most
+			 * recent VFP state of the task.
+			 */
+			ti->vfpstate.hard.cpu = ti->cpu;
+#endif
+		}
+
+		if (fpexc & FPEXC_EX)
+			/*
+			 * Might as well handle the pending exception before
+			 * retrying branch out before setting an FPEXC that
+			 * stops us reading stuff.
+			 */
+			goto bounce;
+
+		/*
+		 * No FP exception is pending: just enable the VFP and
+		 * replay the instruction that trapped.
+		 */
+		fmxr(FPEXC, fpexc);
+		regs->ARM_pc -= 4;
+	} else {
+		/* Check for synchronous or asynchronous exceptions */
+		if (!(fpexc & (FPEXC_EX | FPEXC_DEX))) {
+			u32 fpscr = fmrx(FPSCR);
+
+			/*
+			 * On some implementations of the VFP subarch 1,
+			 * setting FPSCR.IXE causes all the CDP instructions to
+			 * be bounced synchronously without setting the
+			 * FPEXC.EX bit
+			 */
+			if (!(fpscr & FPSCR_IXE)) {
+				if (!(fpscr & FPSCR_LENGTH_MASK)) {
+					pr_debug("not VFP\n");
+					local_bh_enable();
+					return -ENOEXEC;
+				}
+				fpexc |= FPEXC_DEX;
+			}
+		}
+bounce:		VFP_bounce(trigger, fpexc, regs);
+	}
+
+	local_bh_enable();
+	return 0;
 }
 
 #ifdef CONFIG_KERNEL_MODE_NEON