Merge tag 'xtensa-20190917' of git://github.com/jcmvbkbc/linux-xtensa

Pull Xtensa updates from Max Filippov:

 - add support for xtensa call0 ABI in userspace

 - update xtensa virt board DTS for PCI root complex in KIO range

 - remove free_initrd_mem

* tag 'xtensa-20190917' of git://github.com/jcmvbkbc/linux-xtensa:
  xtensa: virt: move PCI root complex to KIO range
  xtensa: add support for call0 ABI in userspace
  xtensa: clean up PS_WOE_BIT usage
  xtensa: remove free_initrd_mem

arch/xtensa/Kconfig

@@ -385,6 +385,54 @@ config FAST_SYSCALL_SPILL_REGISTERS
 
 	  If unsure, say N.
 
+config USER_ABI_CALL0
+	bool
+
+choice
+	prompt "Userspace ABI"
+	default USER_ABI_DEFAULT
+	help
+	  Select supported userspace ABI.
+
+	  If unsure, choose the default ABI.
+
+config USER_ABI_DEFAULT
+	bool "Default ABI only"
+	help
+	  Assume default userspace ABI. For XEA2 cores it is windowed ABI.
+	  call0 ABI binaries may be run on such kernel, but signal delivery
+	  will not work correctly for them.
+
+config USER_ABI_CALL0_ONLY
+	bool "Call0 ABI only"
+	select USER_ABI_CALL0
+	help
+	  Select this option to support only call0 ABI in userspace.
+	  Windowed ABI binaries will crash with a segfault caused by
+	  an illegal instruction exception on the first 'entry' opcode.
+
+	  Choose this option if you're planning to run only user code
+	  built with call0 ABI.
+
+config USER_ABI_CALL0_PROBE
+	bool "Support both windowed and call0 ABI by probing"
+	select USER_ABI_CALL0
+	help
+	  Select this option to support both windowed and call0 userspace
+	  ABIs. When enabled all processes are started with PS.WOE disabled
+	  and a fast user exception handler for an illegal instruction is
+	  used to turn on PS.WOE bit on the first 'entry' opcode executed by
+	  the userspace.
+
+	  This option should be enabled for the kernel that must support
+	  both call0 and windowed ABIs in userspace at the same time.
+
+	  Note that Xtensa ISA does not guarantee that entry opcode will
+	  raise an illegal instruction exception on cores with XEA2 when
+	  PS.WOE is disabled, check whether the target core supports it.
+
+endchoice
+
 endmenu
 
 config XTENSA_CALIBRATE_CCOUNT

arch/xtensa/boot/dts/virt.dts

@@ -52,12 +52,12 @@ pci {
 		#size-cells = <2>;
 		#interrupt-cells = <0x1>;
 
-		bus-range = <0x0 0x3f>;
-		reg = <0xc0000000 0x04000000>;
+		bus-range = <0x0 0x3e>;
+		reg = <0xf0100000 0x03f00000>;
 
 		     // BUS_ADDRESS(3)  CPU_PHYSICAL(1)  SIZE(2)
-		ranges = <0x01000000 0x0 0xc4000000  0xc4000000  0x0 0x04000000>,
-			 <0x02000000 0x0 0xc8000000  0xc8000000  0x0 0x18000000>;
+		ranges = <0x01000000 0x0 0xf0000000  0xf0000000  0x0 0x00010000>,
+			 <0x02000000 0x0 0xf4000000  0xf4000000  0x0 0x08000000>;
 
 		     // PCI_DEVICE(3)  INT#(1)  CONTROLLER(PHANDLE)  CONTROLLER_DATA(2)
 		interrupt-map = <

arch/xtensa/include/asm/io.h

@@ -21,6 +21,7 @@
 
 #define IOADDR(x)		(XCHAL_KIO_BYPASS_VADDR + (x))
 #define IO_SPACE_LIMIT ~0
+#define PCI_IOBASE		((void __iomem *)XCHAL_KIO_BYPASS_VADDR)
 
 #ifdef CONFIG_MMU
 

arch/xtensa/include/asm/processor.h

@@ -176,14 +176,21 @@ struct thread_struct {
 
 /*
  * Do necessary setup to start up a newly executed thread.
- * Note: We set-up ps as if we did a call4 to the new pc.
+ * Note: When windowed ABI is used for userspace we set-up ps
+ *       as if we did a call4 to the new pc.
  *       set_thread_state in signal.c depends on it.
  */
-#define USER_PS_VALUE ((1 << PS_WOE_BIT) |				\
+#if IS_ENABLED(CONFIG_USER_ABI_CALL0)
+#define USER_PS_VALUE ((USER_RING << PS_RING_SHIFT) |			\
+		       (1 << PS_UM_BIT) |				\
+		       (1 << PS_EXCM_BIT))
+#else
+#define USER_PS_VALUE (PS_WOE_MASK |					\
 		       (1 << PS_CALLINC_SHIFT) |			\
 		       (USER_RING << PS_RING_SHIFT) |			\
 		       (1 << PS_UM_BIT) |				\
 		       (1 << PS_EXCM_BIT))
+#endif
 
 /* Clearing a0 terminates the backtrace. */
 #define start_thread(regs, new_pc, new_sp) \

arch/xtensa/include/asm/regs.h

@@ -81,6 +81,7 @@
 /*  PS register fields.  */
 
 #define PS_WOE_BIT		18
+#define PS_WOE_MASK		0x00040000
 #define PS_CALLINC_SHIFT	16
 #define PS_CALLINC_MASK		0x00030000
 #define PS_OWB_SHIFT		8

arch/xtensa/kernel/entry.S

@@ -414,7 +414,7 @@ common_exception:
 	movi	a3, LOCKLEVEL
 
 .Lexception:
-	movi	a0, 1 << PS_WOE_BIT
+	movi	a0, PS_WOE_MASK
 	or	a3, a3, a0
 #else
 	addi	a2, a2, -EXCCAUSE_LEVEL1_INTERRUPT
@@ -422,7 +422,7 @@ common_exception:
 	extui	a3, a3, PS_INTLEVEL_SHIFT, PS_INTLEVEL_WIDTH
 					# a3 = PS.INTLEVEL
 	moveqz	a3, a0, a2		# a3 = LOCKLEVEL iff interrupt
-	movi	a2, 1 << PS_WOE_BIT
+	movi	a2, PS_WOE_MASK
 	or	a3, a3, a2
 	rsr	a2, exccause
 #endif
@@ -922,7 +922,7 @@ ENTRY(unrecoverable_exception)
 	wsr	a1, windowbase
 	rsync
 
-	movi	a1, (1 << PS_WOE_BIT) | LOCKLEVEL
+	movi	a1, PS_WOE_MASK | LOCKLEVEL
 	wsr	a1, ps
 	rsync
 
@@ -1003,7 +1003,41 @@ ENTRY(fast_alloca)
 4:	j	_WindowUnderflow4
 ENDPROC(fast_alloca)
 
+#ifdef CONFIG_USER_ABI_CALL0_PROBE
 /*
+ * fast illegal instruction handler.
+ *
+ * This is used to fix up user PS.WOE on the exception caused
+ * by the first opcode related to register window. If PS.WOE is
+ * already set it goes directly to the common user exception handler.
+ *
+ * Entry condition:
+ *
+ *   a0:	trashed, original value saved on stack (PT_AREG0)
+ *   a1:	a1
+ *   a2:	new stack pointer, original in DEPC
+ *   a3:	a3
+ *   depc:	a2, original value saved on stack (PT_DEPC)
+ *   excsave_1:	dispatch table
+ */
+
+ENTRY(fast_illegal_instruction_user)
+
+	rsr	a0, ps
+	bbsi.l	a0, PS_WOE_BIT, user_exception
+	s32i	a3, a2, PT_AREG3
+	movi	a3, PS_WOE_MASK
+	or	a0, a0, a3
+	wsr	a0, ps
+	l32i	a3, a2, PT_AREG3
+	l32i	a0, a2, PT_AREG0
+	rsr	a2, depc
+	rfe
+
+ENDPROC(fast_illegal_instruction_user)
+#endif
+
+	/*
  * fast system calls.
  *
  * WARNING:  The kernel doesn't save the entire user context before
@@ -1359,7 +1393,7 @@ ENTRY(fast_syscall_spill_registers)
 	rsr	a3, excsave1
 	l32i	a1, a3, EXC_TABLE_KSTK
 
-	movi	a4, (1 << PS_WOE_BIT) | LOCKLEVEL
+	movi	a4, PS_WOE_MASK | LOCKLEVEL
 	wsr	a4, ps
 	rsync
 

arch/xtensa/kernel/head.S

@@ -193,7 +193,7 @@ ENTRY(_startup)
 	movi	a1, start_info
 	l32i	a1, a1, 0
 
-	movi	a2, (1 << PS_WOE_BIT) | LOCKLEVEL
+	movi	a2, PS_WOE_MASK | LOCKLEVEL
 					# WOE=1, INTLEVEL=LOCKLEVEL, UM=0
 	wsr	a2, ps			# (enable reg-windows; progmode stack)
 	rsync

arch/xtensa/kernel/setup.c

@@ -61,7 +61,6 @@ struct screen_info screen_info = {
 #ifdef CONFIG_BLK_DEV_INITRD
 extern unsigned long initrd_start;
 extern unsigned long initrd_end;
-int initrd_is_mapped = 0;
 extern int initrd_below_start_ok;
 #endif
 
@@ -332,13 +331,11 @@ void __init setup_arch(char **cmdline_p)
 	/* Reserve some memory regions */
 
 #ifdef CONFIG_BLK_DEV_INITRD
-	if (initrd_start < initrd_end) {
-		initrd_is_mapped = mem_reserve(__pa(initrd_start),
-					       __pa(initrd_end)) == 0;
+	if (initrd_start < initrd_end &&
+	    !mem_reserve(__pa(initrd_start), __pa(initrd_end)))
 		initrd_below_start_ok = 1;
-	} else {
+	else
 		initrd_start = 0;
-	}
 #endif
 
 	mem_reserve(__pa(_stext), __pa(_end));

arch/xtensa/kernel/signal.c

@@ -335,7 +335,8 @@ static int setup_frame(struct ksignal *ksig, sigset_t *set,
 {
 	struct rt_sigframe *frame;
 	int err = 0, sig = ksig->sig;
-	unsigned long sp, ra, tp;
+	unsigned long sp, ra, tp, ps;
+	unsigned int base;
 
 	sp = regs->areg[1];
 
@@ -385,17 +386,26 @@ static int setup_frame(struct ksignal *ksig, sigset_t *set,
 
 	/* Set up registers for signal handler; preserve the threadptr */
 	tp = regs->threadptr;
+	ps = regs->ps;
 	start_thread(regs, (unsigned long) ksig->ka.sa.sa_handler,
 		     (unsigned long) frame);
 
-	/* Set up a stack frame for a call4
-	 * Note: PS.CALLINC is set to one by start_thread
-	 */
-	regs->areg[4] = (((unsigned long) ra) & 0x3fffffff) | 0x40000000;
-	regs->areg[6] = (unsigned long) sig;
-	regs->areg[7] = (unsigned long) &frame->info;
-	regs->areg[8] = (unsigned long) &frame->uc;
+	/* Set up a stack frame for a call4 if userspace uses windowed ABI */
+	if (ps & PS_WOE_MASK) {
+		base = 4;
+		regs->areg[base] =
+			(((unsigned long) ra) & 0x3fffffff) | 0x40000000;
+		ps = (ps & ~(PS_CALLINC_MASK | PS_OWB_MASK)) |
+			(1 << PS_CALLINC_SHIFT);
+	} else {
+		base = 0;
+		regs->areg[base] = (unsigned long) ra;
+	}
+	regs->areg[base + 2] = (unsigned long) sig;
+	regs->areg[base + 3] = (unsigned long) &frame->info;
+	regs->areg[base + 4] = (unsigned long) &frame->uc;
 	regs->threadptr = tp;
+	regs->ps = ps;
 
 	pr_debug("SIG rt deliver (%s:%d): signal=%d sp=%p pc=%08lx\n",
 		 current->comm, current->pid, sig, frame, regs->pc);

arch/xtensa/kernel/stacktrace.c

@@ -44,6 +44,11 @@ void xtensa_backtrace_user(struct pt_regs *regs, unsigned int depth,
 	if (pc == 0 || pc >= TASK_SIZE || ufn(&frame, data))
 		return;
 
+	if (IS_ENABLED(CONFIG_USER_ABI_CALL0_ONLY) ||
+	    (IS_ENABLED(CONFIG_USER_ABI_CALL0_PROBE) &&
+	     !(regs->ps & PS_WOE_MASK)))
+		return;
+
 	/* Two steps:
 	 *
 	 * 1. Look through the register window for the

arch/xtensa/kernel/traps.c

@@ -51,6 +51,7 @@
 extern void kernel_exception(void);
 extern void user_exception(void);
 
+extern void fast_illegal_instruction_user(void);
 extern void fast_syscall_user(void);
 extern void fast_alloca(void);
 extern void fast_unaligned(void);
@@ -87,6 +88,9 @@ typedef struct {
 
 static dispatch_init_table_t __initdata dispatch_init_table[] = {
 
+#ifdef CONFIG_USER_ABI_CALL0_PROBE
+{ EXCCAUSE_ILLEGAL_INSTRUCTION,	USER,	   fast_illegal_instruction_user },
+#endif
 { EXCCAUSE_ILLEGAL_INSTRUCTION,	0,	   do_illegal_instruction},
 { EXCCAUSE_SYSTEM_CALL,		USER,	   fast_syscall_user },
 { EXCCAUSE_SYSTEM_CALL,		0,	   system_call },

arch/xtensa/mm/init.c

@@ -203,16 +203,6 @@ void __init mem_init(void)
 		(unsigned long)(__bss_stop - __bss_start) >> 10);
 }
 
-#ifdef CONFIG_BLK_DEV_INITRD
-extern int initrd_is_mapped;
-
-void free_initrd_mem(unsigned long start, unsigned long end)
-{
-	if (initrd_is_mapped)
-		free_reserved_area((void *)start, (void *)end, -1, "initrd");
-}
-#endif
-
 static void __init parse_memmap_one(char *p)
 {
 	char *oldp;