On 07/03/2013 05:20 AM, Wolfgang Denk wrote:
Dear Jim Lin,
In message 1372847667-31928-1-git-send-email-jilin@nvidia.com you wrote:
TFTP booting is slow when a USB keyboard is installed and CONFIG_USB_KEYBOARD is defined. The fix is to change Ctrl-C polling to every second when NET transfer is running.
I'm not sure if we can accept this implementation.
+#ifdef CONFIG_USB_KEYBOARD
/** Reduce ctrl-c checking to 1 second once* to improve TFTP boot performance.*/ctrlc_t = get_timer(kbd_ctrlc_tms);if (ctrlc_t > CONFIG_SYS_HZ) {ctrlc_result = ctrlc();kbd_ctrlc_tms = get_timer(0);} else {ctrlc_result = 0;}if (ctrlc_result) {+#else
get_timer() is used by a number of network related services. For information, just grep for it in the net/ and drivers/net/ directories. The "get_timer(0)" used in your code resets a global resource, and has thus the potential of messing up a number of timeouts running elsewhere in the network code. I wonder to which extend this has actually been considered (and tested) ?
I recall you mentioning this before, but can you expand on this a bit please?
For the two platforms I'm familiar with (Tegra and BCM2835), the implementation of get_timer() is simply:
unlong get_timer(ulong base) { ulong time = read_hw_register() time -= base; return time; }
There's no global state involved. Is this implementation of get_timer() wrong somehow? I'm having a hard time envisaging what kind of global state it's supposed to maintain.
I always thought that every user of get_timer() was supposed to do something like:
ulong base = get_timer(0); ... work to be timed ulong time_diff = get_timer(base);
in other words, every user maintains their own base variable, and hence get_timer(0) doesn't affect any other users.