Received-SPF: pass (google.com: domain of linux-kernel-owner@vger.kernel.org designates 23.128.96.34 as permitted sender) client-ip=23.128.96.34;
From:   Jim Cromie <jim.cromie@gmail.com>
To:     linux-kernel@vger.kernel.org, jbaron@akamai.com,
        gregkh@linuxfoundation.org
Cc:     david@redhat.com, lb@semihalf.com, linux@rasmusvillemoes.dk,
        joe@perches.com, mcgrof@kernel.org, Liam.Howlett@Oracle.com,
        linux-mm@kvack.org, Jim Cromie <jim.cromie@gmail.com>
Subject: [RFC PATCH 10/10] dyndbg: cache the dynamically generated prefixes per callsite
Date:   Thu, 12 Oct 2023 13:48:34 -0600
Message-ID: <20231012194834.3288085-11-jim.cromie@gmail.com>
In-Reply-To: <20231012194834.3288085-1-jim.cromie@gmail.com>
References: <20231012194834.3288085-1-jim.cromie@gmail.com>
MIME-Version: 1.0
Content-Transfer-Encoding: 8bit
Precedence: bulk

Add a private maple-tree to cache any prefix strings required (and
then generated) for any enabled pr_debug callsites.  And delete any
cache entries if any flags are changed afterwards.

This cache is strictly per-callsite, so if a function has 20
pr_debugs, all enabled with the same flags:

  echo function foo +pmfs > /proc/dynamic_debug/control

there will be 20 separate, identical cache items created.
Maybe this can be trivially improved later.
Or lineno could be folded in too, so the %d is rendered once.

NB: +tl flags are added outside the cache; the thread-id doesnt belong
in the cache, the lineno could be added, esp if the 20:1 sharing isnt
trivial enough.

NBB: lineno:18 is enormously "sparse".  At this point its not so far
to pack it into something else.

Signed-off-by: Jim Cromie <jim.cromie@gmail.com>
---
 lib/dynamic_debug.c | 27 ++++++++++++++++++++++++++-
 1 file changed, 26 insertions(+), 1 deletion(-)

diff --git a/lib/dynamic_debug.c b/lib/dynamic_debug.c
index 92ffd70a07de..02df61ab6403 100644
--- a/lib/dynamic_debug.c
+++ b/lib/dynamic_debug.c
@@ -90,6 +90,8 @@ MODULE_PARM_DESC(verbose, " dynamic_debug/control processing "
 static DEFINE_MTREE(mt_funcs);
 static DEFINE_MTREE(mt_files);
 static DEFINE_MTREE(mt_mods);
+/* cache of composed prefixes for enabled pr_debugs */
+static DEFINE_MTREE(pr_prefixes);
 
 static void dd_mt_scan(struct maple_tree *mt, const char *kind);
 static int param_set_do_scan(const char *instr, const struct kernel_param *kp)
@@ -112,6 +114,11 @@ static const struct kernel_param_ops param_ops_do_scan = {
 };
 module_param_cb(do_scan, &param_ops_do_scan, NULL, 0600);
 
+void ddebug_clear_prefix_cache(const struct _ddebug *dp)
+{
+	mtree_erase(&pr_prefixes, (unsigned long)dp);
+}
+
 /* Return the path relative to source root */
 static inline const char *trim_prefix(const char *path)
 {
@@ -350,6 +357,7 @@ static int ddebug_change(const struct ddebug_query *query,
 			newflags = (dp->flags & modifiers->mask) | modifiers->flags;
 			if (newflags == dp->flags)
 				continue;
+			ddebug_clear_prefix_cache(dp);
 #ifdef CONFIG_JUMP_LABEL
 			if (dp->flags & _DPRINTK_FLAGS_PRINT) {
 				if (!(newflags & _DPRINTK_FLAGS_PRINT))
@@ -855,8 +863,18 @@ static int remaining(int wrote)
 	return 0;
 }
 
-static int __dynamic_emit_prefix(const struct _ddebug *desc, char *buf, int pos)
+static int __dynamic_emit_prefix(struct _ddebug *desc, char *buf, int pos)
 {
+	char *prefix, *cpy;
+
+	if (desc->flags & _DPRINTK_FLAGS_PREFIX_CACHED) {
+		prefix = (char*) mtree_load(&pr_prefixes, (unsigned long)desc);
+		if (prefix) {
+			pos += snprintf(buf + pos, remaining(pos), "%s", prefix);
+			v4pr_info("using prefix cache:%px %s", buf, buf + pos);
+			return pos;
+		}
+	}
 	if (desc->flags & _DPRINTK_FLAGS_INCL_MODNAME)
 		pos += snprintf(buf + pos, remaining(pos), "%s:",
 				desc_modname(desc));
@@ -866,6 +884,13 @@ static int __dynamic_emit_prefix(const struct _ddebug *desc, char *buf, int pos)
 	if (desc->flags & _DPRINTK_FLAGS_INCL_SOURCENAME)
 		pos += snprintf(buf + pos, remaining(pos), "%s:",
 				trim_prefix(desc_filename(desc)));
+
+	/* save dup of buf to cache */
+	cpy = kstrdup(buf + pos, GFP_KERNEL);
+	mtree_store(&pr_prefixes, (unsigned long)desc, (void*)cpy, GFP_KERNEL);
+	desc->flags |= _DPRINTK_FLAGS_PREFIX_CACHED;
+	v3pr_info("filling prefix cache:%px %s", desc, cpy);
+
 	return pos;
 }
 
-- 
2.41.0