slub-dump: walk_caches: stop once target caches are found + walk list backwards

[xairy]

With #147 resolved, I see a couple more opportunities to speed up the slub-dump command.

The main current bottleneck is the walk_caches routine. It walks the cache list forward and keeps walking until it goes through the whole list. Even though it only processes list entries that are present in target_names, the list walking itself takes time.

I see the following changes that could be made:

1. Make walk_caches stop walking the list once it finds all target_names caches (can there be more than 1?).

2. Make walk_caches walk the list backwards. The reason for this is that kmalloc caches are at the end of the list (they are the first caches to be created, and new caches get added to the front of the list). And I think it's fair to assume that kmalloc caches are the most frequent ones to be dumped, so optimizing for them seems to make sense.

With the dirty patch below, running slub-dump kmalloc-1k -vv --cpu 0 -n over KGDB completes 3 times faster (but note that this patch breaks --list).

diff --git a/gef.py b/gef.py
index 811c9062..91875a2f 100644
--- a/gef.py
+++ b/gef.py
@@ -74947,7 +74947,8 @@ class SlubDumpCommand(GenericCommand, BufferingOutput):
     def get_next_kmem_cache(self, addr, point_to_base=True):
         if point_to_base:
             addr += self.kmem_cache_offset_list
-        return read_int_from_memory(addr) - self.kmem_cache_offset_list
+        # XXX: speed up by walking backwards
+        return read_int_from_memory(addr + 8) - self.kmem_cache_offset_list
 
     def get_name(self, addr):
         name_addr = read_int_from_memory(addr + self.kmem_cache_offset_name)
@@ -75258,6 +75259,9 @@ class SlubDumpCommand(GenericCommand, BufferingOutput):
             # goto next
             current_kmem_cache = kmem_cache["next"]
             self.quiet_info("! slub-dump: walk_caches: loop 1: 4")
+            # XXX: dirty hack to bail out if target is found
+            if len(target_names) == 1 and kmem_cache["name"] in target_names:
+                break
 
         if self.args.list:
             return parsed_caches # fast return
```

[bata24]

Thanks for letting me know.

I see.
However, rather than changing the existing logic, it might have less of an impact to "add an option" that traverses the list in reverse.
I think I'll move forward in that direction.

[xairy]

Sure, making the traversal direction an option would work. But please consider making the reverse direction the default - I don't see any benefit in walking the list in the normal direction.