ftrace: Synchronize the initialization of ftrace_opsgrafted

There's some internal state that ftrace_ops needs to have set, but since
it can be declared outside of the ftrace.c code, it calls
ftrace_ops_init() on the ops in every global function. The issue is that
if two tasks call it on the same ops at the same time it is possible to
have the initialization of one corrupt the initialization of the other
call.

Create a ops_mutex to use to synchronize every initialization of the
ftrace_ops. The mutex is taken within checking the ftrace_ops flag that
states it was initializied but the flag is checked again after the mutex
has been taken. Checking first outside the mutex allows it to shortcut
having to take the mutex. But then the check needs to be done again after
the mute is taken in case of races.

Cc: stable@vger.kernel.org
Link: https://patch.msgid.link/20260902095501.6b59af20@gandalf.local.home
Fixes: f04f24fb7e48d ("ftrace, kprobes: Fix a deadlock on ftrace_regex_lock")
Reported-by: sashiko-bot@kernel.org
Close: https://lore.kernel.org/all/20260829025528.49A831F000E9@smtp.kernel.org/
Signed-off-by: Steven Rostedt <rostedt@goodmis.org>

1 files changed, 199 insertions, 0 deletions

diff --git a/arch/alpha/lib/divide.S b/arch/alpha/lib/divide.S
new file mode 100644
index 000000000..db01840d7
--- /dev/null
+++ b/arch/alpha/lib/divide.S
@@ -0,0 +1,199 @@
+/* SPDX-License-Identifier: GPL-2.0 */
+/*
+ * arch/alpha/lib/divide.S
+ *
+ * (C) 1995 Linus Torvalds
+ *
+ * Alpha division..
+ */
+
+/*
+ * The alpha chip doesn't provide hardware division, so we have to do it
+ * by hand.  The compiler expects the functions
+ *
+ *	__divqu: 64-bit unsigned long divide
+ *	__remqu: 64-bit unsigned long remainder
+ *	__divqs/__remqs: signed 64-bit
+ *	__divlu/__remlu: unsigned 32-bit
+ *	__divls/__remls: signed 32-bit
+ *
+ * These are not normal C functions: instead of the normal
+ * calling sequence, these expect their arguments in registers
+ * $24 and $25, and return the result in $27. Register $28 may
+ * be clobbered (assembly temporary), anything else must be saved. 
+ *
+ * In short: painful.
+ *
+ * This is a rather simple bit-at-a-time algorithm: it's very good
+ * at dividing random 64-bit numbers, but the more usual case where
+ * the divisor is small is handled better by the DEC algorithm
+ * using lookup tables. This uses much less memory, though, and is
+ * nicer on the cache.. Besides, I don't know the copyright status
+ * of the DEC code.
+ */
+
+/*
+ * My temporaries:
+ *	$0 - current bit
+ *	$1 - shifted divisor
+ *	$2 - modulus/quotient
+ *
+ *	$23 - return address
+ *	$24 - dividend
+ *	$25 - divisor
+ *
+ *	$27 - quotient/modulus
+ *	$28 - compare status
+ */
+
+#include <linux/export.h>
+#define halt .long 0
+
+/*
+ * Select function type and registers
+ */
+#define mask	$0
+#define divisor	$1
+#define compare $28
+#define tmp1	$3
+#define tmp2	$4
+
+#ifdef DIV
+#define DIV_ONLY(x,y...) x,##y
+#define MOD_ONLY(x,y...)
+#define func(x) __div##x
+#define modulus $2
+#define quotient $27
+#define GETSIGN(x) xor $24,$25,x
+#define STACK 48
+#else
+#define DIV_ONLY(x,y...)
+#define MOD_ONLY(x,y...) x,##y
+#define func(x) __rem##x
+#define modulus $27
+#define quotient $2
+#define GETSIGN(x) bis $24,$24,x
+#define STACK 32
+#endif
+
+/*
+ * For 32-bit operations, we need to extend to 64-bit
+ */
+#ifdef INTSIZE
+#define ufunction func(lu)
+#define sfunction func(l)
+#define LONGIFY(x) zapnot x,15,x
+#define SLONGIFY(x) addl x,0,x
+#else
+#define ufunction func(qu)
+#define sfunction func(q)
+#define LONGIFY(x)
+#define SLONGIFY(x)
+#endif
+
+.set noat
+.align	3
+.globl	ufunction
+.ent	ufunction
+ufunction:
+	subq	$30,STACK,$30
+	.frame	$30,STACK,$23
+	.prologue 0
+
+7:	stq	$1, 0($30)
+	bis	$25,$25,divisor
+	stq	$2, 8($30)
+	bis	$24,$24,modulus
+	stq	$0,16($30)
+	bis	$31,$31,quotient
+	LONGIFY(divisor)
+	stq	tmp1,24($30)
+	LONGIFY(modulus)
+	bis	$31,1,mask
+	DIV_ONLY(stq tmp2,32($30))
+	beq	divisor, 9f			/* div by zero */
+
+#ifdef INTSIZE
+	/*
+	 * shift divisor left, using 3-bit shifts for
+	 * 32-bit divides as we can't overflow. Three-bit
+	 * shifts will result in looping three times less
+	 * here, but can result in two loops more later.
+	 * Thus using a large shift isn't worth it (and
+	 * s8add pairs better than a sll..)
+	 */
+1:	cmpult	divisor,modulus,compare
+	s8addq	divisor,$31,divisor
+	s8addq	mask,$31,mask
+	bne	compare,1b
+#else
+1:	cmpult	divisor,modulus,compare
+	blt     divisor, 2f
+	addq	divisor,divisor,divisor
+	addq	mask,mask,mask
+	bne	compare,1b
+	unop
+#endif
+
+	/* ok, start to go right again.. */
+2:	DIV_ONLY(addq quotient,mask,tmp2)
+	srl	mask,1,mask
+	cmpule	divisor,modulus,compare
+	subq	modulus,divisor,tmp1
+	DIV_ONLY(cmovne compare,tmp2,quotient)
+	srl	divisor,1,divisor
+	cmovne	compare,tmp1,modulus
+	bne	mask,2b
+
+9:	ldq	$1, 0($30)
+	ldq	$2, 8($30)
+	ldq	$0,16($30)
+	ldq	tmp1,24($30)
+	DIV_ONLY(ldq tmp2,32($30))
+	addq	$30,STACK,$30
+	ret	$31,($23),1
+	.end	ufunction
+EXPORT_SYMBOL(ufunction)
+
+/*
+ * Uhh.. Ugly signed division. I'd rather not have it at all, but
+ * it's needed in some circumstances. There are different ways to
+ * handle this, really. This does:
+ * 	-a / b = a / -b = -(a / b)
+ *	-a % b = -(a % b)
+ *	a % -b = a % b
+ * which is probably not the best solution, but at least should
+ * have the property that (x/y)*y + (x%y) = x.
+ */
+.align 3
+.globl	sfunction
+.ent	sfunction
+sfunction:
+	subq	$30,STACK,$30
+	.frame	$30,STACK,$23
+	.prologue 0
+	bis	$24,$25,$28
+	SLONGIFY($28)
+	bge	$28,7b
+	stq	$24,0($30)
+	subq	$31,$24,$28
+	stq	$25,8($30)
+	cmovlt	$24,$28,$24	/* abs($24) */
+	stq	$23,16($30)
+	subq	$31,$25,$28
+	stq	tmp1,24($30)
+	cmovlt	$25,$28,$25	/* abs($25) */
+	unop
+	bsr	$23,ufunction
+	ldq	$24,0($30)
+	ldq	$25,8($30)
+	GETSIGN($28)
+	subq	$31,$27,tmp1
+	SLONGIFY($28)
+	ldq	$23,16($30)
+	cmovlt	$28,tmp1,$27
+	ldq	tmp1,24($30)
+	addq	$30,STACK,$30
+	ret	$31,($23),1
+	.end	sfunction
+EXPORT_SYMBOL(sfunction)