ada symbol lookup cache: disable for 7.9?

ada symbol lookup cache: disable for 7.9?

[Doug Evans]

Hi.

Just a heads up that I've filed three bugs related to ada's symbol lookup cache.
I happened on the first one, 17854, while trying to see why my testcase to
exercise the failure I'd expect from 17855 wasn't working.
It turns out that ada-lang.c accidently never uses its symbol lookup
cache (AFAICT!).
Rather, it will create a new one for each lookup, not recording the
cache in pspace_data,
and leaking the memory.

Once I fixed 17854 I verified 17855 is real, though I don't yet have a
testcase that crashes gdb
(we should add some sort of deterministic assert that the cache has
been flushed after symbols have been reloaded).

I then found 17856, which is that now that the cache is being used, I
see ada-lang.c
trying to save back in the cache a symbol it just found in the cache.
I stopped looking at this point.

Anyways, given that until now ada's symbol lookup cache has never been used,
if we turn it on I'm not sure if there aren't more bugs, and to fix
the mem-leak for 7.9
(which happens for every non-local symbol lookup IIUC), maybe it'd be
safer to disable the cache.

https://sourceware.org/bugzilla/show_bug.cgi?id=17854
https://sourceware.org/bugzilla/show_bug.cgi?id=17855
https://sourceware.org/bugzilla/show_bug.cgi?id=17856

btw, I'm totally ok with not including the general symbol lookup cache in 7.9.
I thought it was sufficiently robust, but once I found 17855 I had to
reassess that. :-)

Re: ada symbol lookup cache: disable for 7.9?

[Joel Brobecker]

[-- Attachment #1: Type: text/plain, Size: 1460 bytes --]

> It turns out that ada-lang.c accidently never uses its symbol lookup
> cache (AFAICT!).
> Rather, it will create a new one for each lookup, not recording the
> cache in pspace_data,
> and leaking the memory.

Indeed!

Attached is a patch that fixes this issue. I've tested on x86_64-linux,
after having applied your testscase update
(https://www.sourceware.org/ml/gdb-patches/2015-01/msg00510.html),
as well as the fix you posted in PR 17855
(https://sourceware.org/bugzilla/show_bug.cgi?id=17855).
No regression.

As to whether we should deactivate it entirely for 7.9, I would
personally prefer we tried to fix the bug(s), and see where we get.
It used to work fine until I accidently botched the initialization
of that cache when I tried to make it per-program-space. That's
what I'll do anyway for AdaCore's version. However, if you think
it's better to disable it entirely for the FSF version, I will
accept that.

gdb/ChangeLog:

        PR gdb/17854:
        * ada-lang.c (ada_get_symbol_cache): Set pspace_data->sym_cache
        when allocating a new one.

We need your fix from PR 17855 before we can commit this one, because
your testcase update reveals a regression as soon as the cache gets
activated (I tested your patch on x86_64-linux as part of this testing).

I'll start looking at re-caching entries obtained from cache. It seems
that it's a real waste of memory, but not bad enough to actually cause
a functional issue?

Thanks,
-- 
Joel

[-- Attachment #2: 0001-Ada-pspace_data-sym_cache-is-always-NULL.patch --]
[-- Type: text/x-diff, Size: 1877 bytes --]

From 13e46b9a4e665ccc9ae6d99f63e240f98f67d184 Mon Sep 17 00:00:00 2001
From: Joel Brobecker <brobecker@adacore.com>
Date: Mon, 19 Jan 2015 12:57:44 +0100
Subject: [PATCH] [Ada] pspace_data->sym_cache is always NULL

The Ada symbol cache has been designed to have one instance of that
of that cache per program space, and for each instance to be created
on-demand. ada_get_symbol_cache is the function responsible for both
lookup and creation on demand.

Unfortunately, ada_get_symbol_cache forgot to store the reference
to newly created caches, thus causing it to:
  - Leak old caches;
  - Allocate a new cache each time the cache is being searched or
    a new entry is to be inserted.

This patch fixes the issue by avoiding the use of the local variable,
which indirectly allowed the bug to happen. We manipulate the reference
in the program-space data instead.

gdb/ChangeLog:

        PR gdb/17854:
        * ada-lang.c (ada_get_symbol_cache): Set pspace_data->sym_cache
        when allocating a new one.
---
 gdb/ada-lang.c | 9 ++++-----
 1 file changed, 4 insertions(+), 5 deletions(-)

diff --git a/gdb/ada-lang.c b/gdb/ada-lang.c
index f5753f1..2c98191 100644
--- a/gdb/ada-lang.c
+++ b/gdb/ada-lang.c
@@ -4404,15 +4404,14 @@ static struct ada_symbol_cache *
 ada_get_symbol_cache (struct program_space *pspace)
 {
   struct ada_pspace_data *pspace_data = get_ada_pspace_data (pspace);
-  struct ada_symbol_cache *sym_cache = pspace_data->sym_cache;
 
-  if (sym_cache == NULL)
+  if (pspace_data->sym_cache == NULL)
     {
-      sym_cache = XCNEW (struct ada_symbol_cache);
-      ada_init_symbol_cache (sym_cache);
+      pspace_data->sym_cache = XCNEW (struct ada_symbol_cache);
+      ada_init_symbol_cache (pspace_data->sym_cache);
     }
 
-  return sym_cache;
+  return pspace_data->sym_cache;
 }
 
 /* Clear all entries from the symbol cache.  */
-- 
1.9.1

Re: ada symbol lookup cache: disable for 7.9?

[Doug Evans]

On Mon, Jan 19, 2015 at 7:04 AM, Joel Brobecker <brobecker@adacore.com> wrote:
> As to whether we should deactivate it entirely for 7.9, I would
> personally prefer we tried to fix the bug(s), and see where we get.
> It used to work fine until I accidently botched the initialization
> of that cache when I tried to make it per-program-space. That's
> what I'll do anyway for AdaCore's version. However, if you think
> it's better to disable it entirely for the FSF version, I will
> accept that.

Since this is ada-specific, I don't feel comfortable making such a call.
I was just thinking that at least until now it hasn't really been
tested in the FSF tree and given how late in the day it is ...
But again, it's your call.

pushed: [Ada] pspace_data->sym_cache is always NULL (was: "ada symbol lookup cache: disable for 7.9?")

[Joel Brobecker]

> gdb/ChangeLog:
> 
>         PR gdb/17854:
>         * ada-lang.c (ada_get_symbol_cache): Set pspace_data->sym_cache
>         when allocating a new one.

This one is now in.

-- 
Joel

pushed/7.9: [Ada] pspace_data->sym_cache is always NULL (was: "ada symbol lookup cache: disable for 7.9?")

[Joel Brobecker]

I forgot about pushing them in the gdb-7.9-branch... Now done.

On Mon, Feb 02, 2015 at 07:33:05AM +0400, Joel Brobecker wrote:
> > gdb/ChangeLog:
> > 
> >         PR gdb/17854:
> >         * ada-lang.c (ada_get_symbol_cache): Set pspace_data->sym_cache
> >         when allocating a new one.
> 
> This one is now in.

-- 
Joel

[PATCH] [Ada] Do not re-cache symbol-lookup result found from cache lookup.

[Joel Brobecker]

When ada-lang.c:ada_lookup_symbol_list_worker finds a match in
the symbol cache, it caches the result again, which is unecessary.
This patch fixes the code to avoid that.

gdb/ChangeLog:

        PR 17856:
        * ada-lang.c (ada_lookup_symbol_list_worker): Do not re-cache
        results found in the cache.

Tested on x86_64-linux, no regression.
I'll push in a couple of days if no objection.

---
 gdb/ada-lang.c | 11 +++++------
 1 file changed, 5 insertions(+), 6 deletions(-)

diff --git a/gdb/ada-lang.c b/gdb/ada-lang.c
index ccad0a0..c081c13 100644
--- a/gdb/ada-lang.c
+++ b/gdb/ada-lang.c
@@ -5612,14 +5612,12 @@ ada_lookup_symbol_list_worker (const char *name0, const struct block *block0,
   const struct block *block;
   const char *name;
   const int wild_match_p = should_use_wild_match (name0);
-  int cacheIfUnique;
+  int syms_from_global_search = 0;
   int ndefns;
 
   obstack_free (&symbol_list_obstack, NULL);
   obstack_init (&symbol_list_obstack);
 
-  cacheIfUnique = 0;
-
   /* Search specified block and its superiors.  */
   name = name0;
   block = block0;
@@ -5662,7 +5660,6 @@ ada_lookup_symbol_list_worker (const char *name0, const struct block *block0,
      already performed this search before.  If we have, then return
      the same result.  */
 
-  cacheIfUnique = 1;
   if (lookup_cached_symbol (name0, namespace, &sym, &block))
     {
       if (sym != NULL)
@@ -5670,6 +5667,8 @@ ada_lookup_symbol_list_worker (const char *name0, const struct block *block0,
       goto done;
     }
 
+  syms_from_global_search = 1;
+
   /* Search symbols from all global blocks.  */
  
   add_nonlocal_symbols (&symbol_list_obstack, name, namespace, 1,
@@ -5688,10 +5687,10 @@ done:
 
   ndefns = remove_extra_symbols (*results, ndefns);
 
-  if (ndefns == 0 && full_search)
+  if (ndefns == 0 && full_search && syms_from_global_search)
     cache_symbol (name0, namespace, NULL, NULL);
 
-  if (ndefns == 1 && full_search && cacheIfUnique)
+  if (ndefns == 1 && full_search && syms_from_global_search)
     cache_symbol (name0, namespace, (*results)[0].sym, (*results)[0].block);
 
   ndefns = remove_irrelevant_renamings (*results, ndefns, block0);
-- 
1.9.1

pushed: [Ada] Do not re-cache symbol-lookup result found from cache lookup.

[Joel Brobecker]

> gdb/ChangeLog:
> 
>         PR 17856:
>         * ada-lang.c (ada_lookup_symbol_list_worker): Do not re-cache
>         results found in the cache.
> 
> Tested on x86_64-linux, no regression.

Pushed!

-- 
Joel

pushed/7.9: [Ada] Do not re-cache symbol-lookup result found from cache lookup.

[Joel Brobecker]

Forgot about the gdb-7.9-branch... Now fixed.

On Mon, Feb 02, 2015 at 07:34:17AM +0400, Joel Brobecker wrote:
> > gdb/ChangeLog:
> > 
> >         PR 17856:
> >         * ada-lang.c (ada_lookup_symbol_list_worker): Do not re-cache
> >         results found in the cache.
> > 
> > Tested on x86_64-linux, no regression.
> 
> Pushed!

-- 
Joel