[PATCH v2 2/2] ftrace: Drop weak function locations of modules when loading them

[Lawrence Lin via B4 Relay <devnull+deduce.gmail.com@kernel.org>]

From: Lawrence Lin <deduce@gmail.com>

Since commit b39181f7c690 ("ftrace: Add FTRACE_MCOUNT_MAX_OFFSET to avoid
adding weak function"), ftrace_module_enable() calls test_for_valid_rec()
for every ftrace record of a module being loaded. For a module address,
kallsyms_lookup() scans the whole symbol table of the module, so loading a
module costs O(records * symbols), all of it under ftrace_lock. amdgpu.ko
has 16821 records and about 67000 symbols, and commit 4099b98203d6
("ftrace: Fix softlockup in ftrace_module_enable") already had to add a
cond_resched() to this loop because of it.

Commit ef378c3b8233 ("scripts/sorttable: Zero out weak functions in
mcount_loc table") fixed vmlinux at build time, noting that the real
solution is to not add a weak function into the ftrace table in the first
place. Modules are not covered by it, and still have such locations: a
weak function in virt/kvm that arch/x86 overrides inside the same kvm.ko
keeps its mcount location but has no symbol. In an x86_64 distribution
build of v7.2.5, kvm.ko has 18 of them; none of the other 6483 modules has
any.

Do the same for modules when they are loaded. In ftrace_process_locs(),
after the locations are sorted, find for each symbol of the module, by
binary search, the locations at most FTRACE_MCOUNT_MAX_OFFSET after it,
and zero the locations no symbol marked. ftrace_process_locs() already
skips zeroed locations, so no record is created for them, and
ftrace_module_enable() no longer has to test every record. This costs one
bit per location, about 2 KB for amdgpu, and O(symbols * log(records))
time.

On a Ryzen 3 3200U (x86_64, v7.2.5, amdgpu loaded from the initramfs,
three boots each), amdgpu finishes initializing 6.19 s into boot without
this patch and 2.00 s with it, and the kernel part of boot reported by
systemd-analyze drops from 6.65 s to 2.46 s. Loading radeon and nouveau,
which have no hardware on that machine, goes from 137 ms to 64 ms and from
457 ms to 118 ms. available_filter_functions loses the 18
__ftrace_invalid_address___ entries of kvm; its module entries are
otherwise unchanged.

Fixes: b39181f7c690 ("ftrace: Add FTRACE_MCOUNT_MAX_OFFSET to avoid adding weak function")
Assisted-by: Claude:claude-opus-5-5
Signed-off-by: Lawrence Lin <deduce@gmail.com>
---
 kernel/trace/ftrace.c | 82 +++++++++++++++++++++++++++++++++++++++++++--------
 1 file changed, 70 insertions(+), 12 deletions(-)

diff --git a/kernel/trace/ftrace.c b/kernel/trace/ftrace.c
index 673a54fdf392..2e1a237ce901 100644
--- a/kernel/trace/ftrace.c
+++ b/kernel/trace/ftrace.c
@@ -4445,11 +4445,6 @@ static int print_rec(struct seq_file *m, unsigned long ip)
 	return ret == NULL ? -1 : 0;
 }
 #else
-static inline int test_for_valid_rec(struct dyn_ftrace *rec)
-{
-	return 1;
-}
-
 static inline int print_rec(struct seq_file *m, unsigned long ip)
 {
 	seq_printf(m, "%ps", (void *)ip);
@@ -7611,6 +7606,73 @@ static void test_is_sorted(unsigned long *start, unsigned long count)
 }
 #endif
 
+#ifdef FTRACE_MCOUNT_MAX_OFFSET
+struct ftrace_mod_locs {
+	unsigned long *start;
+	unsigned long count;
+	unsigned long *valid;
+};
+
+/* Mark the locations that lie at most FTRACE_MCOUNT_MAX_OFFSET after @addr. */
+static void ftrace_mark_valid_locs(void *data, unsigned long addr)
+{
+	struct ftrace_mod_locs *locs = data;
+	unsigned long lo = 0, hi = locs->count, mid, ip;
+
+	/*
+	 * ftrace_call_adjust() moves a location forward by at most
+	 * FTRACE_MCOUNT_MAX_OFFSET, so start looking that far before @addr.
+	 */
+	while (lo < hi) {
+		mid = lo + (hi - lo) / 2;
+		if (locs->start[mid] + FTRACE_MCOUNT_MAX_OFFSET < addr)
+			lo = mid + 1;
+		else
+			hi = mid;
+	}
+
+	for (; lo < locs->count; lo++) {
+		if (locs->start[lo] > addr + FTRACE_MCOUNT_MAX_OFFSET)
+			break;
+		ip = ftrace_call_adjust(locs->start[lo]);
+		if (ip >= addr && ip - addr <= FTRACE_MCOUNT_MAX_OFFSET)
+			__set_bit(lo, locs->valid);
+	}
+}
+
+/*
+ * A weak function overridden within its module keeps its mcount location but
+ * has no symbol. Zero such locations before they become records, as sorttable
+ * does for vmlinux: keep only those with a symbol at most
+ * FTRACE_MCOUNT_MAX_OFFSET before them.
+ */
+static int ftrace_zero_weak_locs(struct module *mod, unsigned long *start,
+				 unsigned long count)
+{
+	struct ftrace_mod_locs locs = { .start = start, .count = count };
+	unsigned long i;
+
+	locs.valid = bitmap_zalloc(count, GFP_KERNEL);
+	if (!locs.valid)
+		return -ENOMEM;
+
+	module_kallsyms_on_each_addr(mod, ftrace_mark_valid_locs, &locs);
+
+	for_each_clear_bit(i, locs.valid, count)
+		start[i] = 0;
+
+	bitmap_free(locs.valid);
+	return 0;
+}
+#else
+static inline int ftrace_zero_weak_locs(struct module *mod,
+					unsigned long *start,
+					unsigned long count)
+{
+	return 0;
+}
+#endif
+
 static int ftrace_process_locs(struct module *mod,
 			       unsigned long *start,
 			       unsigned long *end)
@@ -7644,6 +7706,9 @@ static int ftrace_process_locs(struct module *mod,
 		test_is_sorted(start, count);
 	}
 
+	if (mod && ftrace_zero_weak_locs(mod, start, count))
+		return -ENOMEM;
+
 	start_pg = ftrace_allocate_pages(count, &pages);
 	if (!start_pg)
 		return -ENOMEM;
@@ -8049,13 +8114,6 @@ void ftrace_module_enable(struct module *mod)
 
 		cond_resched();
 
-		/* Weak functions should still be ignored */
-		if (!test_for_valid_rec(rec)) {
-			/* Clear all other flags. Should not be enabled anyway */
-			rec->flags = FTRACE_FL_DISABLED;
-			continue;
-		}
-
 		cnt = 0;
 
 		/*

-- 
2.55.0