Re: [PATCHv3 1/7] uprobe: Add support for session consumer

On Thu, Sep 12, 2024 at 06:20:29PM +0200, Oleg Nesterov wrote:

On 09/09, Jiri Olsa wrote:

quoted

 static void handler_chain(struct uprobe *uprobe, struct pt_regs *regs)
 {
 	struct uprobe_consumer *uc;
 	int remove = UPROBE_HANDLER_REMOVE;
-	bool need_prep = false; /* prepare return uprobe, when needed */
+	struct return_consumer *ric = NULL;
+	struct return_instance *ri = NULL;
 	bool has_consumers = false;

 	current->utask->auprobe = &uprobe->arch;

 	list_for_each_entry_srcu(uc, &uprobe->consumers, cons_node,
 				 srcu_read_lock_held(&uprobes_srcu)) {
+		__u64 cookie = 0;
 		int rc = 0;

 		if (uc->handler) {
-			rc = uc->handler(uc, regs);
-			WARN(rc & ~UPROBE_HANDLER_MASK,
+			rc = uc->handler(uc, regs, &cookie);
+			WARN(rc < 0 || rc > 2,
 				"bad rc=0x%x from %ps()\n", rc, uc->handler);
 		}

-		if (uc->ret_handler)
-			need_prep = true;
-
+		/*
+		 * The handler can return following values:
+		 * 0 - execute ret_handler (if it's defined)
+		 * 1 - remove uprobe
+		 * 2 - do nothing (ignore ret_handler)
+		 */
 		remove &= rc;
 		has_consumers = true;
+
+		if (rc == 0 && uc->ret_handler) {

should we enter this block if uc->handler == NULL?

yes, consumer can have just ret_handler defined

quoted

+			/*
+			 * Preallocate return_instance object optimistically with
+			 * all possible consumers, so we allocate just once.
+			 */
+			if (!ri) {
+				ri = alloc_return_instance(uprobe->consumers_cnt);

This doesn't look right...

Suppose we have a single consumer C1, so uprobe->consumers_cnt == 1 and
alloc_return_instance() allocates return_instance with for a single consumer,
so that only ri->consumers[0] is valid.

Right after that uprobe_register()->consumer_add() adds another consumer
C2 with ->ret_handler != NULL.

On the next iteration return_consumer_next() will return the invalid addr
== &ri->consumers[1].

perhaps this needs krealloc() ?

damn.. there used to be a lock ;-) ok, for some reason I thought we are safe
in that list iteration and we are not.. I just made selftest that triggers that

I'm not sure the realloc will help, I feel like we need to allocate return
consumer for each called handler separately to be safe

quoted

+				if (!ri)
+					return;

Not sure we should simply return if kzalloc fails... at least it would be better
to clear current->utask->auprobe.

quoted

+	if (ri && !remove)
+		prepare_uretprobe(uprobe, regs, ri); /* put bp at return */
+	else
+		kfree(ri);

Well, if ri != NULL then remove is not possible, afaics... ri != NULL means
that at least one ->handler() returned rc = 0, thus "remove" must be zero.

So it seems you can just do

	if (ri)
		prepare_uretprobe(...);

true, I think that should be enough

thanks,
jirka

Didn't read other parts of your patch yet ;)

Oleg.