Hi Tom,
On Mon, May 13, 2013 at 2:36 PM, Tom Rini trini@ti.com wrote:
On Sat, Apr 27, 2013 at 07:17:37PM -0700, Simon Glass wrote:
This series adds a tracing feature to U-Boot which is useful for profiling boot time and other purposes.
The core trace library relies on standard gcc function instrumentation and a microsecond timer which should work correctly on almost any architecture. Generic board is used to avoid the need to add the same code in multiple places (CONFIG_SYS_GENERIC_BOARD).
Tracing must be built into the U-Boot image at build time, but can be paused and resumed while running. A trace buffer is used to collect trace information. This buffer can then be transmitted to a host for processing. A host-based processing tool is provided which converts the data to the same format used by Linux, and that can be read by the pytimechart GUI tool.
A U-Boot 'trace' command provides access to the trace information, including support for writing it to memory in a few forms.
Support is provided for stopping tracing at the last possible moment in the bootm process (just before U-Boot jumps to the OS). This is done with a new 'fake go' command, which allows bootm to go through the motions of an OS boot without actually committing to it. Once the 'fake go' is complete, U-Boot can stop tracing and continue execution to transmit the trace information to the host, before jumping to the OS for real.
The system is tested on sandbox and has previously worked on ARM Tegra and Exynos machines. A simple test script is provided for sandbox, along with documentation on how to use the feature.
This feature has been successfully used alongside bootstage to locate bottlenecks and reduce overall boot time.
Because of the bootm changes this series sits on top of the image improvements that were done for verified boot. This series is available at:
http://git.denx.de/u-boot-x86.git
in branch 'trace'.
I like the concept, implementations seems reasonable. How far back does gcc support -finstrument-functions? Any chance of getting tegra support at least resurrected (since it uses the generic board framework)?
OK thanks, I will take a look at these points in the next week or two, and make sure it works on Tegra, Exynos and x86 (as well as the current sandbox). There might be a few problems but it should not be too far away.
Thanks!
-- Tom
Regards, Simon