[pushed] Fix double-free corruption

[pushed] Fix double-free corruption

[Pedro Alves]

Fixes a double-free regression introduced by commit b7b030adc405
("Return unique_xmalloc_ptr from target_read_stralloc"):

gdb.sum:
  Running src/gdb/testsuite/gdb.base/catch-syscall.exp ...
  ERROR: Process no longer exists

Valgrind shows:

  (gdb) catch syscall
  ==3687== Thread 1:
  ==3687== Invalid free() / delete / delete[] / realloc()
  ==3687==    at 0x4C29CF0: free (vg_replace_malloc.c:530)
  ==3687==    by 0x610862: xfree(void*) (common-utils.c:101)
  ==3687==    by 0x440D5D: gdb::xfree_deleter<char>::operator()(char*) const (gdb_unique_ptr.h:34)
  ==3687==    by 0x446CC6: std::unique_ptr<char, gdb::xfree_deleter<char> >::reset(char*) (unique_ptr.h:344)
  ==3687==    by 0x81BE50: xml_fetch_content_from_file(char const*, void*) (xml-support.c:1042)
  ==3687==    by 0x81DA86: xml_init_syscalls_info(char const*) (xml-syscall.c:366)
  ==3687==    by 0x81DBDD: init_syscalls_info(gdbarch*) (xml-syscall.c:398)
  ==3687==    by 0x81E131: get_syscall_by_number(gdbarch*, int, syscall*) (xml-syscall.c:599)
  ==3687==    by 0x5BE86F: catch_syscall_command_1(char*, int, cmd_list_element*) (break-catch-syscall.c:481)
  ==3687==    by 0x4B46B1: do_sfunc(cmd_list_element*, char*, int) (cli-decode.c:138)
  ==3687==    by 0x4B76B8: cmd_func(cmd_list_element*, char*, int) (cli-decode.c:1952)
  ==3687==    by 0x7E91C7: execute_command(char*, int) (top.c:615)
  ==3687==  Address 0x14332ae0 is 0 bytes inside a block of size 4,096 free'd
  ==3687==    at 0x4C2AB8B: realloc (vg_replace_malloc.c:785)
  ==3687==    by 0x610792: xrealloc (common-utils.c:62)
  ==3687==    by 0x81BE3E: xml_fetch_content_from_file(char const*, void*) (xml-support.c:1042)
  ==3687==    by 0x81DA86: xml_init_syscalls_info(char const*) (xml-syscall.c:366)
  ==3687==    by 0x81DBDD: init_syscalls_info(gdbarch*) (xml-syscall.c:398)
  ==3687==    by 0x81E131: get_syscall_by_number(gdbarch*, int, syscall*) (xml-syscall.c:599)
  ==3687==    by 0x5BE86F: catch_syscall_command_1(char*, int, cmd_list_element*) (break-catch-syscall.c:481)
  ==3687==    by 0x4B46B1: do_sfunc(cmd_list_element*, char*, int) (cli-decode.c:138)
  ==3687==    by 0x4B76B8: cmd_func(cmd_list_element*, char*, int) (cli-decode.c:1952)
  ==3687==    by 0x7E91C7: execute_command(char*, int) (top.c:615)
  ==3687==    by 0x6A422D: command_handler(char*) (event-top.c:583)
  ==3687==    by 0x6A45F2: command_line_handler(char*) (event-top.c:773)
  [...]

The problem is that if xrealloc decides it needs a new memory block,
it frees the previous block/pointer, and then text.reset() frees it
again.

gdb/ChangeLog:
2017-10-17  Pedro Alves  <palves@redhat.com>

	* xml-support.c (xml_fetch_content_from_file): Call
	unique_ptr::release() instead unique_ptr::get() when passing
	through xrealloc.
---
 gdb/ChangeLog     | 6 ++++++
 gdb/xml-support.c | 2 +-
 2 files changed, 7 insertions(+), 1 deletion(-)

diff --git a/gdb/ChangeLog b/gdb/ChangeLog
index 6a8d16f..65952c4 100644
--- a/gdb/ChangeLog
+++ b/gdb/ChangeLog
@@ -1,3 +1,9 @@
+2017-10-17  Pedro Alves  <palves@redhat.com>
+
+	* xml-support.c (xml_fetch_content_from_file): Call
+	unique_ptr::release() instead unique_ptr::get() when passing
+	through xrealloc.
+
 2017-10-17  Yao Qi  <yao.qi@linaro.org>
 
 	* regcache.c (regcache::xfer_part): Remove parameters read and
diff --git a/gdb/xml-support.c b/gdb/xml-support.c
index 76d03b9..42a4c91 100644
--- a/gdb/xml-support.c
+++ b/gdb/xml-support.c
@@ -1039,7 +1039,7 @@ xml_fetch_content_from_file (const char *filename, void *baton)
 	break;
 
       len = len * 2;
-      text.reset ((char *) xrealloc (text.get (), len));
+      text.reset ((char *) xrealloc (text.release (), len));
     }
 
   text.get ()[offset] = '\0';
-- 
2.5.5

[PATCH] xml_fetch_content_from_file: Read in whole file in one go (Re: [pushed] Fix double-free corruption)

[Pedro Alves]

On 10/17/2017 12:41 PM, Pedro Alves wrote:
> Fixes a double-free regression introduced by commit b7b030adc405
> ("Return unique_xmalloc_ptr from target_read_stralloc"):
> 
> gdb.sum:
>   Running src/gdb/testsuite/gdb.base/catch-syscall.exp ...
>   ERROR: Process no longer exists
> 


...


> The problem is that if xrealloc decides it needs a new memory block,
> it frees the previous block/pointer, and then text.reset() frees it
> again.

Looking a bit deeper, I can't seem to find a reason this code is
reading in chunks in the first place?  Why not read it all in one
go?  Like patch below.

From a08ecd67cf5bd87c7d2e64ca443017b78d76aa04 Mon Sep 17 00:00:00 2001
From: Pedro Alves <palves@redhat.com>
Date: Tue, 17 Oct 2017 13:02:13 +0100
Subject: [PATCH] xml_fetch_content_from_file: Read in whole file in one go

There doesn't seem to be a good reason we're reading the file one
chunk at a time.

gdb/ChangeLog:
2017-10-17  Pedro Alves  <palves@redhat.com>

	* xml-support.c (xml_fetch_content_from_file): Don't read in
	chunks.  Instead use fseek to determine the file's size, and read
	it in one go.
---
 gdb/xml-support.c | 36 +++++++++++++-----------------------
 1 file changed, 13 insertions(+), 23 deletions(-)

diff --git a/gdb/xml-support.c b/gdb/xml-support.c
index 42a4c91..69aa9db 100644
--- a/gdb/xml-support.c
+++ b/gdb/xml-support.c
@@ -998,7 +998,6 @@ xml_fetch_content_from_file (const char *filename, void *baton)
 {
   const char *dirname = (const char *) baton;
   gdb_file_up file;
-  size_t len, offset;
 
   if (dirname && *dirname)
     {
@@ -1015,34 +1014,25 @@ xml_fetch_content_from_file (const char *filename, void *baton)
   if (file == NULL)
     return NULL;
 
-  /* Read in the whole file, one chunk at a time.  */
-  len = 4096;
-  offset = 0;
-  gdb::unique_xmalloc_ptr<char> text ((char *) xmalloc (len));
-  while (1)
-    {
-      size_t bytes_read;
+  /* Read in the whole file.  */
 
-      /* Continue reading where the last read left off.  Leave at least
-	 one byte so that we can NUL-terminate the result.  */
-      bytes_read = fread (text.get () + offset, 1, len - offset - 1,
-			  file.get ());
-      if (ferror (file.get ()))
-	{
-	  warning (_("Read error from \"%s\""), filename);
-	  return NULL;
-	}
+  size_t len;
 
-      offset += bytes_read;
+  if (fseek (file.get (), 0, SEEK_END) == -1)
+    perror_with_name (_("seek to end of file"));
+  len = ftell (file.get ());
+  rewind (file.get ());
 
-      if (feof (file.get ()))
-	break;
+  gdb::unique_xmalloc_ptr<char> text ((char *) xmalloc (len + 1));
 
-      len = len * 2;
-      text.reset ((char *) xrealloc (text.release (), len));
+  fread (text.get (), 1, len, file.get ());
+  if (ferror (file.get ()))
+    {
+      warning (_("Read error from \"%s\""), filename);
+      return {};
     }
 
-  text.get ()[offset] = '\0';
+  text.get ()[len] = '\0';
   return text;
 }
 
-- 
2.5.5

Re: [PATCH] xml_fetch_content_from_file: Read in whole file in one go (Re: [pushed] Fix double-free corruption)

[John Baldwin]

On Tuesday, October 17, 2017 01:20:14 PM Pedro Alves wrote:
> On 10/17/2017 12:41 PM, Pedro Alves wrote:
> > Fixes a double-free regression introduced by commit b7b030adc405
> > ("Return unique_xmalloc_ptr from target_read_stralloc"):
> > 
> > gdb.sum:
> >   Running src/gdb/testsuite/gdb.base/catch-syscall.exp ...
> >   ERROR: Process no longer exists
> > 
> 
> 
> ...
> 
> 
> > The problem is that if xrealloc decides it needs a new memory block,
> > it frees the previous block/pointer, and then text.reset() frees it
> > again.
> 
> Looking a bit deeper, I can't seem to find a reason this code is
> reading in chunks in the first place?  Why not read it all in one
> go?  Like patch below.

Seems sensible.  It's unlikely that an XML file will be stored directly
on a tape such that seek() is non-optimal. :)

-- 
John Baldwin

Re: [PATCH] xml_fetch_content_from_file: Read in whole file in one go (Re: [pushed] Fix double-free corruption)

[Pedro Alves]

On 10/18/2017 06:43 PM, John Baldwin wrote:
> On Tuesday, October 17, 2017 01:20:14 PM Pedro Alves wrote:

>> Looking a bit deeper, I can't seem to find a reason this code is
>> reading in chunks in the first place?  Why not read it all in one
>> go?  Like patch below.
> 
> Seems sensible.  It's unlikely that an XML file will be stored directly
> on a tape such that seek() is non-optimal. :)
> 

:-)

Thanks much for the review.  I pushed this in now.

-- 
Pedro Alves