This is the mail archive of the
gdb-patches@sourceware.org
mailing list for the GDB project.
[PATCH v3] Optimize memory_xfer_partial for remote
- From: Don Breazeal <donb at codesourcery dot com>
- To: <gdb-patches at sourceware dot org>, <palves at redhat dot com>
- Date: Thu, 30 Jun 2016 15:40:24 -0700
- Subject: [PATCH v3] Optimize memory_xfer_partial for remote
- Authentication-results: sourceware.org; auth=none
Hi Pedro,
Here is v3 of the patch. Differences from v2 include returning
ULONGEST_MAX instead of 4096 from the default target function
and using TARGET_DEFAULT_RETURN instead of TARGET_DEFAULT_FUNCTION
in the definition of to_get_memory_xfer_limit.
I re-ran the testsuite with native and native-gdbserver and saw no
regressions. I also verified that the behavior of the memory transfers
was as expected for the remote and native targets.
Does this one look good-to-go?
thanks
--Don
---------- 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 were 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 ULONGEST_MAX bytes,
because for native targets there is no marshaling and thus no limit is
needed.
The performance differences that I saw were as follows (in seconds),
using a remote target with 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:
2016-06-30 Don Breazeal <donb@codesourcery.com>
* remote.c (remote_get_memory_xfer_limit): New function.
* target-delegates.c: Regenerate.
* target.c (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 | 31 +++++++++++++++++++++++++++++++
gdb/target.c | 5 +++--
gdb/target.h | 6 ++++++
4 files changed, 49 insertions(+), 2 deletions(-)
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..2887033 100644
--- a/gdb/target-delegates.c
+++ b/gdb/target-delegates.c
@@ -2064,6 +2064,33 @@ 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
+tdefault_get_memory_xfer_limit (struct target_ops *self)
+{
+ return ULONGEST_MAX;
+}
+
+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 +4250,8 @@ 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 +4483,7 @@ 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 = tdefault_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 +4640,7 @@ 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 bb86adf..c88607e 100644
--- a/gdb/target.c
+++ b/gdb/target.c
@@ -1301,8 +1301,9 @@ 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..68ef470 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_RETURN (ULONGEST_MAX);
+
/* 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
--
2.8.1