diff mbox

[v2] Optimize memory_xfer_partial for remote

Message ID 1467058970-62136-1-git-send-email-donb@codesourcery.com
State New
Headers show

Commit Message

Don Breazeal June 27, 2016, 8:22 p.m. UTC
On 6/24/2016 3:23 PM, Pedro Alves wrote:
> On 06/24/2016 10:21 PM, Don Breazeal wrote:
>>
>> and with commit hash: 67c059c29e1fb0cdeacdd2005f955514d8d1fb34
>>
> 
> Write:
> 
>  ... with commit 67c059c29e1f ("Improve performance of large restore
>  commands") ...
> 
> so the reader has a clue what the commit is about without having
> to check.

Hi Pedro,
Thanks for your comments; sorry about the coding convention and clarity
issues. The one above and the others are fixed in the attached patch below.

> 
> 
>> gdb/ChangeLog:
>> 2016-06-24  Don Breazeal  <donb@codesourcery.com>
>>
>> 	* remote.c (remote_get_memory_xfer_limit): New function.
>> 	* target-delegates.c (delegate_get_memory_xfer_limit,
>> 	debug_get_memory_xfer_limit, install_delegators,
>> 	install_dummy_methods, init_debug_target): New functions
>> 	and target_ops initialization from regenerating the file.
> 
> The standard practice is to just say:
> 
>  	* target-delegates.c: Regenerate.

Fixed in the attached patch.

> 
>> 	* target.c (default_get_memory_xfer_limit): New function and
>> 	forward declaration.
>> 	(memory_xfer_partial): Call target_ops.to_get_memory_xfer_limit.
>> 	* target.h (struct target_ops)<to_get_memory_xfer_limit>: New
>> 	member.
> 
> Space between ")<".

Fixed in the attached patch.

> 
>>
>> diff --git a/gdb/remote.c b/gdb/remote.c
>> index 501f3c6..03c7ab7 100644
>> --- a/gdb/remote.c
>> +++ b/gdb/remote.c
>> @@ -10160,6 +10160,12 @@ remote_xfer_partial (struct target_ops *ops, enum target_object object,
>>    return TARGET_XFER_OK;
>>  }
>>  
>> +static ULONGEST
>> +remote_get_memory_xfer_limit (struct target_ops *ops)
> 
> Intro comment, something like "Implementation of ... method.".

Fixed in the attached patch.

> 
>>  
>> +/* The default implementation for the to_get_memory_xfer_limit method.
>> +   The hard-coded limit here was determined to be a reasonable default
>> +   that eliminated exponential slowdown on very large transfers without
>> +   unduly compromising performance on smaller transfers.  */
> 
> Where's this coming from?  Is this new experimentation you did,
> or are you talking about Anton's patch?

Both.  I did some experimentation to verify that things were significantly
slower without any memory transfer limit, which they were, although I never
reproduced the extreme scenario Anton had reported.  Presumably the
performance differences were due to hardware and environment differences.
Regarding the comment, I thought some explanation of the hard-coded number
was appropriate.  Is there a better or more preferable way to do this, e.g.
refer to the commit hash, or does it just seem superfluous?

> 
>> @@ -1301,8 +1314,9 @@ memory_xfer_partial (struct target_ops *ops, enum target_object object,
>>  	 by memory_xfer_partial_1.  We will continually malloc
>>  	 and free a copy of the entire write request for breakpoint
>>  	 shadow handling even though we only end up writing a small
>> -	 subset of it.  Cap writes to 4KB to mitigate this.  */
>> -      len = min (4096, len);
>> +	 subset of it.  Cap writes to a limit specified by the target
>> +	 to mitigate this.  */
>> +      len = min (ops->to_get_memory_xfer_limit (ops), len);
>>  
> 
> Does this still work if remote is not the top-most target?

Yes, see next comment.

> 
> E.g., what happens if you do "record" to push a record_statum
> target on top?  Do we get the 4KB default limit, or the
> remote limit?

I ran this experiment and verified that the target delegation mechanism
worked as expected in this case. It walks down the target stack, skipping
targets that don't implement to_get_memory_xfer_limit, and returning after
calling the function in the first target that does implement it. I also
verified that it walks all the way down the target stack and calls the
default function if no target has defined its own version of
to_get_memory_xfer_limit.

Output of the experiment with record and remote is below, followed by the
revised patch.

Thanks
--Don

[build5-trusty-cs(5.68) 694] bin$ gdb ./gdb
GNU gdb (Ubuntu 7.7.1-0ubuntu5~14.04.2) 7.7.1
Copyright (C) 2014 Free Software Foundation, Inc.
... GDB banner stuff...
Reading symbols from ./gdb...done.
(gdb) set prompt (top)
(top) b target.c:1319
Breakpoint 1 at 0x65985b: file ../../binutils-gdb/gdb/target.c, line 1319.
(top) run ~/test/big
Starting program: /scratch/dbreazea/gdb-5301/install/bin/gdb ~/test/big
[Thread debugging using libthread_db enabled]
Using host libthread_db library "/lib/x86_64-linux-gnu/libthread_db.so.1".
GNU gdb (GDB) 7.11.50.20160624-git
...GDB banner stuff...
Reading symbols from /home/dbreazea/test/big...done.
(gdb) tar rem localhost:51111
Remote debugging using localhost:51111
...Reading symbols msgs...
0x00007ffff7ddb2d0 in ?? () from target:/lib64/ld-linux-x86-64.so.2
(gdb) b main
Breakpoint 1 at 0x40051e: file /home/dbreazea/test/big.c, line 9.
(gdb) c
Continuing.
Reading /lib/x86_64-linux-gnu/libc.so.6 from remote target...
Reading /lib/x86_64-linux-gnu/libc-2.19.so from remote target...
Reading /lib/x86_64-linux-gnu/.debug/libc-2.19.so from remote target...

Breakpoint 1, main () at /home/dbreazea/test/big.c:9
9 printf ("starting...\n");
(gdb) record full <---<<< push the target on top of the remote target
(gdb) b 12
Breakpoint 2 at 0x40054d: file /home/dbreazea/test/big.c, line 12.
(gdb) c
Continuing.

Breakpoint 2, main () at /home/dbreazea/test/big.c:12
12 printf ("done, a[%d] is %d\n", 42, a[42]);
(gdb) restore ../../build/gdb/a.srec <---<<< 'restore' for big writes
Restoring section .sec1 (0x6009a0 to 0x6089a0)

Breakpoint 1, memory_xfer_partial (ops=0xe07e20 <record_full_ops>,
object=TARGET_OBJECT_MEMORY, readbuf=0x0, writebuf=0x212c480 "",
memaddr=6293920, len=32768, xfered_len=0x7fffffffdea8)
at ../../binutils-gdb/gdb/target.c:1319
1319 len = min (ops->to_get_memory_xfer_limit (ops), len);
(top) p ops->to_longname
$1 = 0x9ddc00 "Process record and replay target"
(top) s
delegate_get_memory_xfer_limit (self=0xe07e20 <record_full_ops>)
at ../../binutils-gdb/gdb/target-delegates.c:2070
2070 self = self->beneath;
(top) display self->to_longname
1: self->to_longname = 0x9ddc00 "Process record and replay target"
(top) s
2071 return self->to_get_memory_xfer_limit (self);
1: self->to_longname = 0x8fcd00 "Remote serial target in gdb-specific protocol"
(top) s
remote_get_memory_xfer_limit (ops=0xdf0360 <remote_ops>)
at ../../binutils-gdb/gdb/remote.c:10166
10166 return get_memory_write_packet_size ();
(top) n
10167 }
(top) n
delegate_get_memory_xfer_limit (self=0xdf0360 <remote_ops>)
at ../../binutils-gdb/gdb/target-delegates.c:2072
2072 }
1: self->to_longname = 0x8fcd00 "Remote serial target in gdb-specific protocol"
(top) n
memory_xfer_partial (ops=0xe07e20 <record_full_ops>,
object=TARGET_OBJECT_MEMORY, readbuf=0x0, writebuf=0x212c480 "",
memaddr=6293920, len=16383, xfered_len=0x7fffffffdea8)
at ../../binutils-gdb/gdb/target.c:1321
1321 buf = (gdb_byte *) xmalloc (len);
(top) p len
$2 = 16383 <---<<< 16K (approx) len from remote target function

---------- revised patch -----------
Some analysis we did here showed that increasing the cap on the
transfer size in target.c:memory_xfer_partial could give 20% or more
improvement in remote load across JTAG.  Transfer sizes are capped
to 4K bytes because of performance problems encountered with the
restore command, documented here:

https://sourceware.org/ml/gdb-patches/2013-07/msg00611.html

and in commit 67c059c29e1f ("Improve performance of large restore
commands").

The 4K cap was introduced because in a case where the restore command
requested a 100MB transfer, memory_xfer_partial would repeatedy
allocate and copy an entire 100MB buffer in order to properly handle
breakpoint shadow instructions, even though memory_xfer_partial would
actually only write a small portion of the buffer contents.

A couple of alternative solutions were suggested:
* change the algorithm for handling the breakpoint shadow instructions
* throttle the transfer size up or down based on the previous actual
  transfer size

I tried implementing the throttling approach, and my implementation
reduced the performance in some cases.

This patch implements a new target function that returns that target's
limit on memory transfer size.  It defaults to 4K bytes as before, but
for remote it uses remote.c:get_memory_write_packet_size.

The performance differences that I saw were as follows (in seconds),
using an artificially large application and a 100MB srec file:

USB  load:     orig   53.2 patch  18.9
USB  restore:  orig 1522.4 patch 543.6
Enet load:     orig   12.2 patch  10.0
Enet restore:  orig  368.0 patch 294.3

Tested on x86_64 Linux with native and native-gdbserver, and manually
tested 'load' and 'restore' on a Windows 7 host with a bare-metal ARM
board.

gdb/ChangeLog:
2016-06-27  Don Breazeal  <donb@codesourcery.com>

	* remote.c (remote_get_memory_xfer_limit): New function.
	* target-delegates.c: Regenerate.
	* target.c (default_get_memory_xfer_limit): New function and
	forward declaration.
	(memory_xfer_partial): Call target_ops.to_get_memory_xfer_limit.
	* target.h (struct target_ops) <to_get_memory_xfer_limit>: New
	member.

---
 gdb/remote.c           |  9 +++++++++
 gdb/target-delegates.c | 25 +++++++++++++++++++++++++
 gdb/target.c           | 18 ++++++++++++++++--
 gdb/target.h           |  6 ++++++
 4 files changed, 56 insertions(+), 2 deletions(-)

Comments

Pedro Alves June 30, 2016, 5:06 p.m. UTC | #1
On 06/27/2016 09:22 PM, Don Breazeal wrote:

>>> +/* The default implementation for the to_get_memory_xfer_limit method.
>>> +   The hard-coded limit here was determined to be a reasonable default
>>> +   that eliminated exponential slowdown on very large transfers without
>>> +   unduly compromising performance on smaller transfers.  */
>>
>> Where's this coming from?  Is this new experimentation you did,
>> or are you talking about Anton's patch?
> 
> Both.  I did some experimentation to verify that things were significantly
> slower without any memory transfer limit, which they were, although I never
> reproduced the extreme scenario Anton had reported.  Presumably the
> performance differences were due to hardware and environment differences.
> Regarding the comment, I thought some explanation of the hard-coded number
> was appropriate.  Is there a better or more preferable way to do this, e.g.
> refer to the commit hash, or does it just seem superfluous?

OK, you didn't mention this experimentation, which left me wondering.
Particularly, the mention of "exponential" is what most made me pause,
as it's a qualifier not mentioned elsewhere.

I guess my main problem with the comment is that by reading it in
isolation, one has no clue of how what would cause the slowdown (normally
transferring more at a time is faster!), and thus how to reevaluate
the default in the future.  How about extending to something like:

/* The default implementation for the to_get_memory_xfer_limit method.
   The hard-coded limit here was determined to be a reasonable default
   that eliminated exponential slowdown on very large transfers without
   unduly compromising performance on smaller transfers.  
   This slowdown is mostly caused by memory writing routines doing
   unnecessary work upfront when large requests end up usually
   only partially satisfied.  See memory_xfer_partial's handling of
   breakpoint shadows.  */

Actually, I was going to approve this with that change, but another
another thought crossed my mind, sorry...

I assume you did this experimentation with remote targets?  But this default
will never be used with those, so that experimentation is meaningless for
native targets?  Actually, the whole capping is probably pointless with
native targets, since there's really no marshalling and thus no limit.
That'd suggest making the target method return "-1" or some such
to indicate there's no limit.  WDTY?

Thanks,
Pedro Alves
Don Breazeal June 30, 2016, 5:45 p.m. UTC | #2
On 6/30/2016 10:06 AM, Pedro Alves wrote:
> On 06/27/2016 09:22 PM, Don Breazeal wrote:
> 
>>>> +/* The default implementation for the to_get_memory_xfer_limit method.
>>>> +   The hard-coded limit here was determined to be a reasonable default
>>>> +   that eliminated exponential slowdown on very large transfers without
>>>> +   unduly compromising performance on smaller transfers.  */
>>>
>>> Where's this coming from?  Is this new experimentation you did,
>>> or are you talking about Anton's patch?
>>
>> Both.  I did some experimentation to verify that things were significantly
>> slower without any memory transfer limit, which they were, although I never
>> reproduced the extreme scenario Anton had reported.  Presumably the
>> performance differences were due to hardware and environment differences.
>> Regarding the comment, I thought some explanation of the hard-coded number
>> was appropriate.  Is there a better or more preferable way to do this, e.g.
>> refer to the commit hash, or does it just seem superfluous?
> 
> OK, you didn't mention this experimentation, which left me wondering.
> Particularly, the mention of "exponential" is what most made me pause,
> as it's a qualifier not mentioned elsewhere.

I should have used something like "significant" or "extreme" instead of
exponential.

> 
> I guess my main problem with the comment is that by reading it in
> isolation, one has no clue of how what would cause the slowdown (normally
> transferring more at a time is faster!), and thus how to reevaluate
> the default in the future.  How about extending to something like:
> 
> /* The default implementation for the to_get_memory_xfer_limit method.
>    The hard-coded limit here was determined to be a reasonable default
>    that eliminated exponential slowdown on very large transfers without
>    unduly compromising performance on smaller transfers.  
>    This slowdown is mostly caused by memory writing routines doing
>    unnecessary work upfront when large requests end up usually
>    only partially satisfied.  See memory_xfer_partial's handling of
>    breakpoint shadows.  */
> 
> Actually, I was going to approve this with that change, but another
> another thought crossed my mind, sorry...
> 
> I assume you did this experimentation with remote targets?  But this default
> will never be used with those, so that experimentation is meaningless for
> native targets?  Actually, the whole capping is probably pointless with
> native targets, since there's really no marshalling and thus no limit.
> That'd suggest making the target method return "-1" or some such
> to indicate there's no limit.  WDTY?

That makes sense to me.  If it returns ULONGEST_MAX then the rest of the
patch can stay as-is.  Something like this?

+/* The default implementation for the to_get_memory_xfer_limit method.
+   The default limit is essentially "no limit".  */
+
+static ULONGEST
+default_get_memory_xfer_limit (struct target_ops *self)
+{
+  return ULONGEST_MAX;
+}
+

Thanks
--Don
Pedro Alves June 30, 2016, 6:40 p.m. UTC | #3
On 06/30/2016 06:45 PM, Don Breazeal wrote:

> That makes sense to me.  If it returns ULONGEST_MAX then the rest of the
> patch can stay as-is.  Something like this?
> 
> +/* The default implementation for the to_get_memory_xfer_limit method.
> +   The default limit is essentially "no limit".  */
> +
> +static ULONGEST
> +default_get_memory_xfer_limit (struct target_ops *self)
> +{
> +  return ULONGEST_MAX;
> +}

Agreed.  Though if you use TARGET_DEFAULT_RETURN, then you don't
even need that function:

    /* Return the limit on the size of any single memory transfer
       for the target.  The default limit is essentially "no limit".  */

    ULONGEST (*to_get_memory_xfer_limit) (struct target_ops *)
      TARGET_DEFAULT_RETURN (ULONGEST_MAX);

Thanks,
Pedro Alves
diff mbox

Patch

diff --git a/gdb/remote.c b/gdb/remote.c
index 501f3c6..dfa41ef 100644
--- a/gdb/remote.c
+++ b/gdb/remote.c
@@ -10160,6 +10160,14 @@  remote_xfer_partial (struct target_ops *ops, enum target_object object,
   return TARGET_XFER_OK;
 }
 
+/* Implementation of to_get_memory_xfer_limit.  */
+
+static ULONGEST
+remote_get_memory_xfer_limit (struct target_ops *ops)
+{
+  return get_memory_write_packet_size ();
+}
+
 static int
 remote_search_memory (struct target_ops* ops,
 		      CORE_ADDR start_addr, ULONGEST search_space_len,
@@ -13073,6 +13081,7 @@  Specify the serial device it is connected to\n\
   remote_ops.to_interrupt = remote_interrupt;
   remote_ops.to_pass_ctrlc = remote_pass_ctrlc;
   remote_ops.to_xfer_partial = remote_xfer_partial;
+  remote_ops.to_get_memory_xfer_limit = remote_get_memory_xfer_limit;
   remote_ops.to_rcmd = remote_rcmd;
   remote_ops.to_pid_to_exec_file = remote_pid_to_exec_file;
   remote_ops.to_log_command = serial_log_command;
diff --git a/gdb/target-delegates.c b/gdb/target-delegates.c
index 03aa2cc..18f22b5 100644
--- a/gdb/target-delegates.c
+++ b/gdb/target-delegates.c
@@ -2064,6 +2064,27 @@  debug_xfer_partial (struct target_ops *self, enum target_object arg1, const char
   return result;
 }
 
+static ULONGEST
+delegate_get_memory_xfer_limit (struct target_ops *self)
+{
+  self = self->beneath;
+  return self->to_get_memory_xfer_limit (self);
+}
+
+static ULONGEST
+debug_get_memory_xfer_limit (struct target_ops *self)
+{
+  ULONGEST result;
+  fprintf_unfiltered (gdb_stdlog, "-> %s->to_get_memory_xfer_limit (...)\n", debug_target.to_shortname);
+  result = debug_target.to_get_memory_xfer_limit (&debug_target);
+  fprintf_unfiltered (gdb_stdlog, "<- %s->to_get_memory_xfer_limit (", debug_target.to_shortname);
+  target_debug_print_struct_target_ops_p (&debug_target);
+  fputs_unfiltered (") = ", gdb_stdlog);
+  target_debug_print_ULONGEST (result);
+  fputs_unfiltered ("\n", gdb_stdlog);
+  return result;
+}
+
 static VEC(mem_region_s) *
 delegate_memory_map (struct target_ops *self)
 {
@@ -4223,6 +4244,8 @@  install_delegators (struct target_ops *ops)
     ops->to_get_thread_local_address = delegate_get_thread_local_address;
   if (ops->to_xfer_partial == NULL)
     ops->to_xfer_partial = delegate_xfer_partial;
+  if (ops->to_get_memory_xfer_limit == NULL)
+    ops->to_get_memory_xfer_limit = delegate_get_memory_xfer_limit;
   if (ops->to_memory_map == NULL)
     ops->to_memory_map = delegate_memory_map;
   if (ops->to_flash_erase == NULL)
@@ -4454,6 +4477,7 @@  install_dummy_methods (struct target_ops *ops)
   ops->to_goto_bookmark = tdefault_goto_bookmark;
   ops->to_get_thread_local_address = tdefault_get_thread_local_address;
   ops->to_xfer_partial = tdefault_xfer_partial;
+  ops->to_get_memory_xfer_limit = default_get_memory_xfer_limit;
   ops->to_memory_map = tdefault_memory_map;
   ops->to_flash_erase = tdefault_flash_erase;
   ops->to_flash_done = tdefault_flash_done;
@@ -4610,6 +4634,7 @@  init_debug_target (struct target_ops *ops)
   ops->to_goto_bookmark = debug_goto_bookmark;
   ops->to_get_thread_local_address = debug_get_thread_local_address;
   ops->to_xfer_partial = debug_xfer_partial;
+  ops->to_get_memory_xfer_limit = debug_get_memory_xfer_limit;
   ops->to_memory_map = debug_memory_map;
   ops->to_flash_erase = debug_flash_erase;
   ops->to_flash_done = debug_flash_done;
diff --git a/gdb/target.c b/gdb/target.c
index 6f69ac3..57202b4 100644
--- a/gdb/target.c
+++ b/gdb/target.c
@@ -60,6 +60,8 @@  static int default_region_ok_for_hw_watchpoint (struct target_ops *,
 
 static void default_rcmd (struct target_ops *, const char *, struct ui_file *);
 
+static ULONGEST default_get_memory_xfer_limit (struct target_ops *self);
+
 static ptid_t default_get_ada_task_ptid (struct target_ops *self,
 					 long lwp, long tid);
 
@@ -623,6 +625,17 @@  default_terminal_info (struct target_ops *self, const char *args, int from_tty)
   printf_unfiltered (_("No saved terminal information.\n"));
 }
 
+/* The default implementation for the to_get_memory_xfer_limit method.
+   The hard-coded limit here was determined to be a reasonable default
+   that eliminated exponential slowdown on very large transfers without
+   unduly compromising performance on smaller transfers.  */
+
+static ULONGEST
+default_get_memory_xfer_limit (struct target_ops *self)
+{
+  return 4096;
+}
+
 /* A default implementation for the to_get_ada_task_ptid target method.
 
    This function builds the PTID by using both LWP and TID as part of
@@ -1301,8 +1314,9 @@  memory_xfer_partial (struct target_ops *ops, enum target_object object,
 	 by memory_xfer_partial_1.  We will continually malloc
 	 and free a copy of the entire write request for breakpoint
 	 shadow handling even though we only end up writing a small
-	 subset of it.  Cap writes to 4KB to mitigate this.  */
-      len = min (4096, len);
+	 subset of it.  Cap writes to a limit specified by the target
+	 to mitigate this.  */
+      len = min (ops->to_get_memory_xfer_limit (ops), len);
 
       buf = (gdb_byte *) xmalloc (len);
       old_chain = make_cleanup (xfree, buf);
diff --git a/gdb/target.h b/gdb/target.h
index 6b5b6e0..84f12a9 100644
--- a/gdb/target.h
+++ b/gdb/target.h
@@ -745,6 +745,12 @@  struct target_ops
 						ULONGEST *xfered_len)
       TARGET_DEFAULT_RETURN (TARGET_XFER_E_IO);
 
+    /* Return the limit on the size of any single memory transfer
+       for the target.  */
+
+    ULONGEST (*to_get_memory_xfer_limit) (struct target_ops *)
+      TARGET_DEFAULT_FUNC (default_get_memory_xfer_limit);
+
     /* Returns the memory map for the target.  A return value of NULL
        means that no memory map is available.  If a memory address
        does not fall within any returned regions, it's assumed to be