This is the mail archive of the gdb-patches@sourceware.org mailing list for the GDB project.


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]
Other format: [Raw text]

Re: [PATCH] Assertion 'xfered>0' in target.c for remote connection


(Resending, since my previous response lost threading information in the mailing list)


On 11/2/2017 10:11 PM, Sergio Durigan Junior wrote:
On Thursday, November 02 2017, pcarroll@codesourcery.com wrote:

We have a customer who is using a Corelis gdb server to connect to gdb.
Occasionally, the gdb server will send a 0-byte block of memory for a read.
When this happens, gdb gives an assertion from target.c:

internal-error: target_xfer_partial: Assertion `*xfered_len > 0' failed.

This problem is almost identical to that fixed in https://sourceware.org/ml/gdb-patches/2014-02/msg00636.html

In this case, remote.c needs to be modified to return TARGET_XFER_EOF instead of TARGET_XFER_OK or TARGET_XFER_UNAVAILABLE when 0 bytes are transferred.
Thanks for the patch.  A few comments below.  BTW, your mail client
seems to have messed up with the formatting of the patch, removing TABs
and replacing them with whitespaces.  Could you please send the patch
again, but using a client that handles patches better?
Yep, resending here.  I don't think I've sent email to the mailing lists for years, so I needed to use a different mail client.

The proposed fix would be:

diff -rup fsf/gdb/ChangeLog fix/gdb/ChangeLog
--- fsf/gdb/ChangeLog   2017-11-02 16:13:19.188615000 -0500
+++ fix/gdb/ChangeLog   2017-11-02 16:13:21.626754500 -0500
@@ -1,3 +1,10 @@
+2017-11-02  Paul Carroll  <pcarroll@codesourcery.com>
+
+       PR gdb/22388
+       * remote.c (remote_write_bytes_aux, remote_read_bytes_1,
+       remote_read_bytes, remote_write_qxfer, remote_xfer_partial):
+       Return TARGET_XFER_EOF if size of returned data is 0.
+
The ChangeLog entries need to be formatted with TABs instead of
whitespaces.  But the text looks great.
Taken care of in the new patch
2017-11-02  Yao Qi  <yao.qi@linaro.org>

         * frame.c (do_frame_register_read): Remove aspace.
diff -rup fsf/gdb/remote.c fix/gdb/remote.c
--- fsf/gdb/remote.c    2017-11-02 16:06:15.979408800 -0500
+++ fix/gdb/remote.c    2017-11-02 15:24:35.536391700 -0500
@@ -8264,7 +8264,7 @@ remote_write_bytes_aux (const char *head
    /* Return UNITS_WRITTEN, not TODO_UNITS, in case escape chars caused us to
       send fewer units than we'd planned.  */
    *xfered_len_units = (ULONGEST) units_written;
-  return TARGET_XFER_OK;
+  return (*xfered_len_units != 0) ? TARGET_XFER_OK : TARGET_XFER_EOF;
}
This one looks correct to me.

/* Write memory data directly to the remote machine.
@@ -8358,7 +8358,7 @@ remote_read_bytes_1 (CORE_ADDR memaddr,
    decoded_bytes = hex2bin (p, myaddr, todo_units * unit_size);
    /* Return what we have.  Let higher layers handle partial reads.  */
    *xfered_len_units = (ULONGEST) (decoded_bytes / unit_size);
-  return TARGET_XFER_OK;
+  return (*xfered_len_units != 0) ? TARGET_XFER_OK : TARGET_XFER_EOF;
}
Likewise.

/* Using the set of read-only target sections of remote, read live
@@ -8455,13 +8455,14 @@ remote_read_bytes (struct target_ops *op
               res = remote_xfer_live_readonly_partial (ops, myaddr, memaddr,
                                                        len, unit_size, xfered_len);
               if (res == TARGET_XFER_OK)
-               return TARGET_XFER_OK;
+               return (*xfered_len != 0) ? TARGET_XFER_OK : TARGET_XFER_EOF;
remote_xfer_live_readonly_partial can only return TARGET_XFER_EOF or
what remote_read_bytes_1 return (which already takes care of the EOF
part), so I don't think you need this part here, although keeping it
wouldn't hurt.
I got rid of this in the new patch, since it doesn't add anything, as you noted.
               else
                 {
                   /* No use trying further, we know some memory starting
                      at MEMADDR isn't available.  */
                   *xfered_len = len;
-                 return TARGET_XFER_UNAVAILABLE;
+                 return (*xfered_len != 0) ?
+                   TARGET_XFER_UNAVAILABLE : TARGET_XFER_EOF;
                 }
             }

@@ -10386,7 +10387,7 @@ remote_write_qxfer (struct target_ops *o
    unpack_varlen_hex (rs->buf, &n);

    *xfered_len = n;
-  return TARGET_XFER_OK;
+  return (*xfered_len != 0) ? TARGET_XFER_OK : TARGET_XFER_EOF;
}

/* Read OBJECT_NAME/ANNEX from the remote target using a qXfer packet.
@@ -10687,7 +10688,7 @@ remote_xfer_partial (struct target_ops *
    strcpy ((char *) readbuf, rs->buf);

    *xfered_len = strlen ((char *) readbuf);
-  return TARGET_XFER_OK;
+  return (*xfered_len != 0) ? TARGET_XFER_OK : TARGET_XFER_EOF;
}

/* Implementation of to_get_memory_xfer_limit.  */
The rest looks sane to me, but I'm not a global maintainer and cannot
approve your patch.  Meanwhile I'd recommend resending the patch to fix
the formatting errors.

Cheers,


Attachment: gdb_0_mem_read_patch
Description: Text document


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]