modules: catch concurrent module loads, treat them as idempotent

This is the new-and-improved attempt at avoiding huge memory load spikes
when the user space boot sequence tries to load hundreds (or even
thousands) of redundant duplicate modules in parallel.

See commit 9828ed3f695a ("module: error out early on concurrent load of
the same module file") for background and an earlier failed attempt that
was reverted.

That earlier attempt just said "concurrently loading the same module is
silly, just open the module file exclusively and return -ETXTBSY if
somebody else is already loading it".

While it is true that concurrent module loads of the same module is
silly, the reason that earlier attempt then failed was that the
concurrently loaded module would often be a prerequisite for another
module.

Thus failing to load the prerequisite would then cause cascading
failures of the other modules, rather than just short-circuiting that
one unnecessary module load.

At the same time, we still really don't want to load the contents of the
same module file hundreds of times, only to then wait for an eventually
successful load, and have everybody else return -EEXIST.

As a result, this takes another approach, and treats concurrent module
loads from the same file as "idempotent" in the inode.  So if one module
load is ongoing, we don't start a new one, but instead just wait for the
first one to complete and return the same return value as it did.

So unlike the first attempt, this does not return early: the intent is
not to speed up the boot, but to avoid a thundering herd problem in
allocating memory (both physical and virtual) for a module more than
once.

Also note that this does change behavior: it used to be that when you
had concurrent loads, you'd have one "winner" that would return success,
and everybody else would return -EEXIST.

In contrast, this idempotent logic goes all Oprah on the problem, and
says "You are a winner! And you are a winner! We are ALL winners".  But
since there's no possible actual real semantic difference between "you
loaded the module" and "somebody else already loaded the module", this
is more of a feel-good change than an actual honest-to-goodness semantic
change.

Of course, any true Johnny-come-latelies that don't get caught in the
concurrency filter will still return -EEXIST.  It's no different from
not even getting a seat at an Oprah taping.  That's life.

See the long thread on the kernel mailing list about this all, which
includes some numbers for memory use before and after the patch.

Link: https://lore.kernel.org/lkml/20230524213620.3509138-1-mcgrof@kernel.org/
Reviewed-by: Johan Hovold <johan@kernel.org>
Tested-by: Johan Hovold <johan@kernel.org>
Tested-by: Luis Chamberlain <mcgrof@kernel.org>
Tested-by: Dan Williams <dan.j.williams@intel.com>
Tested-by: Rudi Heitbaum <rudi@heitbaum..com>
Tested-by: David Hildenbrand <david@redhat.com>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

1 files changed, 71 insertions, 2 deletions

diff --git a/kernel/module/main.c b/kernel/module/main.c
index 23abfe3e2674..d6de66a6a1ac 100644
--- a/kernel/module/main.c
+++ b/kernel/module/main.c
@@ -3057,15 +3057,82 @@ SYSCALL_DEFINE3(init_module, void __user *, umod,
 	return load_module(&info, uargs, 0);
 }
 
+struct idempotent {
+	const void *cookie;
+	struct hlist_node entry;
+	struct completion complete;
+	int ret;
+};
+
+#define IDEM_HASH_BITS 8
+static struct hlist_head idem_hash[1 << IDEM_HASH_BITS];
+static DEFINE_SPINLOCK(idem_lock);
+
+static bool idempotent(struct idempotent *u, const void *cookie)
+{
+	int hash = hash_ptr(cookie, IDEM_HASH_BITS);
+	struct hlist_head *head = idem_hash + hash;
+	struct idempotent *existing;
+	bool first;
+
+	u->ret = 0;
+	u->cookie = cookie;
+	init_completion(&u->complete);
+
+	spin_lock(&idem_lock);
+	first = true;
+	hlist_for_each_entry(existing, head, entry) {
+		if (existing->cookie != cookie)
+			continue;
+		first = false;
+		break;
+	}
+	hlist_add_head(&u->entry, idem_hash + hash);
+	spin_unlock(&idem_lock);
+
+	return !first;
+}
+
+/*
+ * We were the first one with 'cookie' on the list, and we ended
+ * up completing the operation. We now need to walk the list,
+ * remove everybody - which includes ourselves - fill in the return
+ * value, and then complete the operation.
+ */
+static void idempotent_complete(struct idempotent *u, int ret)
+{
+	const void *cookie = u->cookie;
+	int hash = hash_ptr(cookie, IDEM_HASH_BITS);
+	struct hlist_head *head = idem_hash + hash;
+	struct hlist_node *next;
+	struct idempotent *pos;
+
+	spin_lock(&idem_lock);
+	hlist_for_each_entry_safe(pos, next, head, entry) {
+		if (pos->cookie != cookie)
+			continue;
+		hlist_del(&pos->entry);
+		pos->ret = ret;
+		complete(&pos->complete);
+	}
+	spin_unlock(&idem_lock);
+}
+
 static int init_module_from_file(struct file *f, const char __user * uargs, int flags)
 {
+	struct idempotent idem;
 	struct load_info info = { };
 	void *buf = NULL;
-	int len;
+	int len, ret;
 
 	if (!f || !(f->f_mode & FMODE_READ))
 		return -EBADF;
 
+	if (idempotent(&idem, file_inode(f))) {
+		wait_for_completion(&idem.complete);
+		return idem.ret;
+	}
+
 	len = kernel_read_file(f, 0, &buf, INT_MAX, NULL, READING_MODULE);
 	if (len < 0) {
 		mod_stat_inc(&failed_kreads);
@@ -3086,7 +3153,9 @@ static int init_module_from_file(struct file *f, const char __user * uargs, int
 		info.len = len;
 	}
 
-	return load_module(&info, uargs, flags);
+	ret = load_module(&info, uargs, flags);
+	idempotent_complete(&idem, ret);
+	return ret;
 }
 
 SYSCALL_DEFINE3(finit_module, int, fd, const char __user *, uargs, int, flags)