Patchwork [v2,18/24] Multi-target support

login
register
mail settings
Submitter Simon Marchi
Date Jan. 17, 2020, 3:47 a.m.
Message ID <c17d429d-452d-557d-377d-b7f1bd9bf269@simark.ca>
Download mbox | patch
Permalink /patch/37415/
State New
Headers show

Comments

Simon Marchi - Jan. 17, 2020, 3:47 a.m.
On 2019-10-17 6:50 p.m., Pedro Alves wrote:
> This commit adds multi-target support to GDB.  What this means is that
> with this commit, GDB can now be connected to different targets at the
> same time.  E.g., you can debug a live native process and a core dump
> at the same time, connect to multiple gdbservers, etc.

Hi Pedro,

I think I found something odd happening starting with this commit.  When
the test gdb.threads/tid-reuse.exp runs, we are adding a thread_info to the
inferior's thread list when a thread_info object with the same ptid is
already present.  I don't think this is ever supposed to happen, right?

This can be observed by applying the following diff and running
gdb.threads/tid-reuse.exp.  Note that some distributions are starting to bump
pid_max to a really large value, so make sure it's not that on your system.


What I have found so far, by breaking at thread.c:274 and stepping into
the "delete_thread" call, is that `inferior_ptid == ptid` is false
(inferior_ptid is (0,0,0)).  In delete_thread, we are actually not
deleting the thread because ->deletable () returns false, because
the refcount is 2.

Before this patch, we would enter the `if (inferior_ptid == ptid)`.

If it can help, here is the command line I'm using to debug this, which is
mimicking the test:

  ./gdb -nx --data-directory=data-directory testsuite/outputs/gdb.threads/tid-reuse/tid-reuse -iex "set confirm off" -iex "set pagination off" -ex start -ex "set print thread-events off" -ex "b after_count" -ex c -ex "b do_nothing_thread_func" -ex c -ex d -ex "b after_reuse_time"  -ex c

Simon
Simon Marchi - Jan. 17, 2020, 3:49 p.m.
On 2020-01-16 10:47 p.m., Simon Marchi wrote:
> On 2019-10-17 6:50 p.m., Pedro Alves wrote:
>> This commit adds multi-target support to GDB.  What this means is that
>> with this commit, GDB can now be connected to different targets at the
>> same time.  E.g., you can debug a live native process and a core dump
>> at the same time, connect to multiple gdbservers, etc.
> 
> Hi Pedro,
> 
> I think I found something odd happening starting with this commit.  When
> the test gdb.threads/tid-reuse.exp runs, we are adding a thread_info to the
> inferior's thread list when a thread_info object with the same ptid is
> already present.  I don't think this is ever supposed to happen, right?
> 
> This can be observed by applying the following diff and running
> gdb.threads/tid-reuse.exp.  Note that some distributions are starting to bump
> pid_max to a really large value, so make sure it's not that on your system.
> 
> diff --git a/gdb/thread.c b/gdb/thread.c
> index 4959f938c7f3..8909e371e2ce 100644
> --- a/gdb/thread.c
> +++ b/gdb/thread.c
> @@ -254,7 +254,10 @@ new_thread (struct inferior *inf, ptid_t ptid)
>        struct thread_info *last;
> 
>        for (last = inf->thread_list; last->next != NULL; last = last->next)
> -	;
> +	gdb_assert (ptid != last->ptid);
> +
> +      gdb_assert (ptid != last->ptid);
> +
>        last->next = tp;
>      }
> 
> 
> What I have found so far, by breaking at thread.c:274 and stepping into
> the "delete_thread" call, is that `inferior_ptid == ptid` is false
> (inferior_ptid is (0,0,0)).  In delete_thread, we are actually not
> deleting the thread because ->deletable () returns false, because
> the refcount is 2.
> 
> Before this patch, we would enter the `if (inferior_ptid == ptid)`.
> 
> If it can help, here is the command line I'm using to debug this, which is
> mimicking the test:
> 
>   ./gdb -nx --data-directory=data-directory testsuite/outputs/gdb.threads/tid-reuse/tid-reuse -iex "set confirm off" -iex "set pagination off" -ex start -ex "set print thread-events off" -ex "b after_count" -ex c -ex "b do_nothing_thread_func" -ex c -ex d -ex "b after_reuse_time"  -ex c
> 
> Simon
> 

I filed this in bugzilla, so we can track it better:

https://sourceware.org/bugzilla/show_bug.cgi?id=25412

Simon

Patch

diff --git a/gdb/thread.c b/gdb/thread.c
index 4959f938c7f3..8909e371e2ce 100644
--- a/gdb/thread.c
+++ b/gdb/thread.c
@@ -254,7 +254,10 @@  new_thread (struct inferior *inf, ptid_t ptid)
       struct thread_info *last;

       for (last = inf->thread_list; last->next != NULL; last = last->next)
-	;
+	gdb_assert (ptid != last->ptid);
+
+      gdb_assert (ptid != last->ptid);
+
       last->next = tp;
     }