On Mon, Sep 25, 2023 at 09:53:16AM +0200, Clément Léger wrote:

On 22/09/2023 21:22, Beau Belgrave wrote:

quoted

On Tue, Sep 19, 2023 at 02:59:12PM +0200, Clément Léger wrote:

quoted

On 14/09/2023 19:29, Steven Rostedt wrote:

quoted

On Thu, 14 Sep 2023 13:17:00 -0400
Steven Rostedt [off-list ref] wrote:

quoted

Now lets look at big endian layout:

 uaddr = 0xbeef0004
 enabler = 1;

 memory at 0xbeef0000:  00 00 00 00 00 00 00 02
                                    ^
                                    addr: 0xbeef0004

				(enabler is set )

	bitoffset = uaddr & (sizeof(unsigned long) - 1); bitoffset = 4
	bit_offset *= 8;				 bitoffset = 32
	uaddr &= ~(sizeof(unsigned long) - 1);		 uaddr = 0xbeef0000

	ptr = kaddr + (uaddr & ~PAGE_MASK);

	clear_bit(1 + 32, ptr);

 memory at 0xbeef0000:  00 00 00 00 00 00 00 02
                                  ^
				bit 33 of 0xbeef0000

I don't think that's what you expected!

I believe the above can be fixed with:

	bit_offset = uaddr & (sizeof(unsigned long) - 1);
	if (bit_offset) {
#ifdef CONFIG_CPU_BIG_ENDIAN
		bit_offest = 0;
#else
		bit_offset *= BITS_PER_BYTE;
#endif
		uaddr &= ~(sizeof(unsigned long) - 1);
	}

-- Steve

Actually, after looking more in depth at that, it seems like there are
actually 2 problems that can happen.

First one is atomic access misalignment due to enable_size == 4 and
uaddr not being aligned on a (long) boundary on 64 bits architecture.
This can generate misaligned exceptions on various architectures. This
can be fixed in a more general way according to Masami snippet.

Second one that I can see is on 64 bits, big endian architectures with
enable_size == 4. In that case, the bit provided by the userspace won't
be correctly set since this code kind of assume that the atomic are done
on 32bits value. Since the kernel assume long sized atomic operation, on
big endian 64 bits architecture, the updated bit will actually be in the
next 32 bits word.

Can someone confirm my understanding ?

I have a ppc 64bit BE VM I've been validating this on. If we do the
shifting within user_events (vs a generic set_bit_aligned approach)
64bit BE does not need additional bit manipulation. However, if we were
to blindly pass the address and bit as is to set_bit_aligned() it
assumes the bit number is for a long, not a 32 bit word. So for that
approach we would need to offset the bit in the unaligned case.

Here's a patch I have that I've validated on ppc64 BE, aarch64 LE, and
x86_64 LE. I personally feel more comfortable with this approach than
the generic set_bit_aligned() one.

Thanks,
-Beau

diff --git a/kernel/trace/trace_events_user.c b/kernel/trace/trace_events_user.c
index e3f2b8d72e01..ae854374d0b7 100644
--- a/kernel/trace/trace_events_user.c
+++ b/kernel/trace/trace_events_user.c

@@ -162,6 +162,23 @@ struct user_event_validator {
 	int			flags;
 };
 
+static inline void align_addr_bit(unsigned long *addr, int *bit)
+{
+	if (IS_ALIGNED(*addr, sizeof(long)))
+		return;
+
+	*addr = ALIGN_DOWN(*addr, sizeof(long));
+
+	/*
+	 * We only support 32 and 64 bit values. The only time we need
+	 * to align is a 32 bit value on a 64 bit kernel, which on LE
+	 * is always 32 bits, and on BE requires no change.
+	 */
+#ifdef __LITTLE_ENDIAN
+	*bit += 32;
+#endif

Hi Beau, except the specific alignment that is basically what I ended up
with for the BE 64bit case (ie just bit += 32). Regarding the generic
alignment, depends on what the maintainers wishes (generic of user_event
specific). I also feel like this shoulmd be handle specifically for
user_events which uses set_bit in some non standard way. Any suggestion ?

Looking at this deeper, user_events needs to track some of this
alignment requirements within each enabler. This is needed because when
enablers are disabled/removed they do not have the actual size. The size
is needed to know if we need to update the bits, etc. (IE: If originally
a 32-bit value was used and it's aligned and it's on BE, it needs the
bits shifted.)

My final version that I played around with looked like this for just the
alignment requirements:
+static inline void align_addr_bit(unsigned long *addr, int *bit,
+                                 unsigned long *flags)
+{
+       if (IS_ALIGNED(*addr, sizeof(long))) {
+#ifdef __BIG_ENDIAN
+               if (test_bit(ENABLE_VAL_32_BIT, flags))
+                       *bit += 32;
+#endif
+               return;
+       }
+
+       *addr = ALIGN_DOWN(*addr, sizeof(long));
+
+       /*
+        * We only support 32 and 64 bit values. The only time we need
+        * to align is a 32 bit value on a 64 bit kernel, which on LE
+        * is always 32 bits, and on BE requires no change.
+        */
+#ifdef __LITTLE_ENDIAN
+       *bit += 32;
+#endif
+}

Below is the full patch, which is currently appearing to pass everything
on BE (Please note, the abi_test in selftests needs fixes that I plan to
push out. I'd like to get Steven to take those changes along with yours
together from tracing to ensure we can test BE with these changes
properly).

As you'll see, it requires a bit more work than just a generic unaligned
solution due to the bits having to move for 32-bit on BE and the
tracking requirement on when to do so during delete/clear.

Thanks,
-Beau

diff --git a/kernel/trace/trace_events_user.c b/kernel/trace/trace_events_user.c
index 6f046650e527..b05db15eaea6 100644
--- a/kernel/trace/trace_events_user.c
+++ b/kernel/trace/trace_events_user.c

@@ -127,8 +127,11 @@ struct user_event_enabler {
 /* Bit 7 is for freeing status of enablement */
 #define ENABLE_VAL_FREEING_BIT 7
 
+/* Bit 8 is for marking 32-bit on 64-bit */
+#define ENABLE_VAL_32_BIT 8
+
 /* Only duplicate the bit value */
-#define ENABLE_VAL_DUP_MASK ENABLE_VAL_BIT_MASK
+#define ENABLE_VAL_DUP_MASK (ENABLE_VAL_BIT_MASK | 1 << ENABLE_VAL_32_BIT)
 
 #define ENABLE_BITOPS(e) (&(e)->values)
 

@@ -174,6 +177,29 @@ struct user_event_validator {
 	int			flags;
 };
 
+static inline void align_addr_bit(unsigned long *addr, int *bit,
+				  unsigned long *flags)
+{
+	if (IS_ALIGNED(*addr, sizeof(long))) {
+#ifdef __BIG_ENDIAN
+		if (test_bit(ENABLE_VAL_32_BIT, flags))
+			*bit += 32;
+#endif
+		return;
+	}
+
+	*addr = ALIGN_DOWN(*addr, sizeof(long));
+
+	/*
+	 * We only support 32 and 64 bit values. The only time we need
+	 * to align is a 32 bit value on a 64 bit kernel, which on LE
+	 * is always 32 bits, and on BE requires no change.
+	 */
+#ifdef __LITTLE_ENDIAN
+	*bit += 32;
+#endif
+}
+
 typedef void (*user_event_func_t) (struct user_event *user, struct iov_iter *i,
 				   void *tpdata, bool *faulted);
 

@@ -482,6 +508,7 @@ static int user_event_enabler_write(struct user_event_mm *mm,
 	unsigned long *ptr;
 	struct page *page;
 	void *kaddr;
+	int bit = ENABLE_BIT(enabler);
 	int ret;
 
 	lockdep_assert_held(&event_mutex);

@@ -497,6 +524,8 @@ static int user_event_enabler_write(struct user_event_mm *mm,
 		     test_bit(ENABLE_VAL_FREEING_BIT, ENABLE_BITOPS(enabler))))
 		return -EBUSY;
 
+	align_addr_bit(&uaddr, &bit, ENABLE_BITOPS(enabler));
+
 	ret = pin_user_pages_remote(mm->mm, uaddr, 1, FOLL_WRITE | FOLL_NOFAULT,
 				    &page, NULL);
 

@@ -515,9 +544,9 @@ static int user_event_enabler_write(struct user_event_mm *mm,
 
 	/* Update bit atomically, user tracers must be atomic as well */
 	if (enabler->event && enabler->event->status)
-		set_bit(ENABLE_BIT(enabler), ptr);
+		set_bit(bit, ptr);
 	else
-		clear_bit(ENABLE_BIT(enabler), ptr);
+		clear_bit(bit, ptr);
 
 	kunmap_local(kaddr);
 	unpin_user_pages_dirty_lock(&page, 1, true);

@@ -849,6 +878,12 @@ static struct user_event_enabler
 	enabler->event = user;
 	enabler->addr = uaddr;
 	enabler->values = reg->enable_bit;
+
+#if BITS_PER_LONG >= 64
+	if (reg->enable_size == 4)
+		set_bit(ENABLE_VAL_32_BIT, ENABLE_BITOPS(enabler));
+#endif
+
 retry:
 	/* Prevents state changes from racing with new enablers */
 	mutex_lock(&event_mutex);

@@ -2377,7 +2412,8 @@ static long user_unreg_get(struct user_unreg __user *ureg,
 }
 
 static int user_event_mm_clear_bit(struct user_event_mm *user_mm,
-				   unsigned long uaddr, unsigned char bit)
+				   unsigned long uaddr, unsigned char bit,
+				   unsigned long flags)
 {
 	struct user_event_enabler enabler;
 	int result;

@@ -2385,7 +2421,7 @@ static int user_event_mm_clear_bit(struct user_event_mm *user_mm,
 
 	memset(&enabler, 0, sizeof(enabler));
 	enabler.addr = uaddr;
-	enabler.values = bit;
+	enabler.values = bit | flags;
 retry:
 	/* Prevents state changes from racing with new enablers */
 	mutex_lock(&event_mutex);

@@ -2415,6 +2451,7 @@ static long user_events_ioctl_unreg(unsigned long uarg)
 	struct user_event_mm *mm = current->user_event_mm;
 	struct user_event_enabler *enabler, *next;
 	struct user_unreg reg;
+	unsigned long flags;
 	long ret;
 
 	ret = user_unreg_get(ureg, &reg);

@@ -2425,6 +2462,7 @@ static long user_events_ioctl_unreg(unsigned long uarg)
 	if (!mm)
 		return -ENOENT;
 
+	flags = 0;
 	ret = -ENOENT;
 
 	/*

@@ -2441,6 +2479,10 @@ static long user_events_ioctl_unreg(unsigned long uarg)
 		    ENABLE_BIT(enabler) == reg.disable_bit) {
 			set_bit(ENABLE_VAL_FREEING_BIT, ENABLE_BITOPS(enabler));
 
+			/* We must keep compat flags for the clear */
+			if (test_bit(ENABLE_VAL_32_BIT, ENABLE_BITOPS(enabler)))
+				flags |= 1 << ENABLE_VAL_32_BIT;
+
 			if (!test_bit(ENABLE_VAL_FAULTING_BIT, ENABLE_BITOPS(enabler)))
 				user_event_enabler_destroy(enabler, true);
 

@@ -2454,7 +2496,7 @@ static long user_events_ioctl_unreg(unsigned long uarg)
 	/* Ensure bit is now cleared for user, regardless of event status */
 	if (!ret)
 		ret = user_event_mm_clear_bit(mm, reg.disable_addr,
-					      reg.disable_bit);
+					      reg.disable_bit, flags);
 
 	return ret;
 }