aboutsummaryrefslogtreecommitdiffstats
path: root/tools/perf
diff options
context:
space:
mode:
authorJohan Hovold <jhovold@gmail.com>2010-05-05 23:57:37 +0200
committerGreg Kroah-Hartman <gregkh@suse.de>2010-05-20 13:21:42 -0700
commit27c7acf22047fbe4ec4cc36b7c2610dba227697c (patch)
treec0a8f217fc2d7a302b4d2e084bb126e825006ca4 /tools/perf
parent4272568b3dd8dbad36014a107c0fbbef6400c917 (diff)
downloadkernel_samsung_smdk4412-27c7acf22047fbe4ec4cc36b7c2610dba227697c.zip
kernel_samsung_smdk4412-27c7acf22047fbe4ec4cc36b7c2610dba227697c.tar.gz
kernel_samsung_smdk4412-27c7acf22047fbe4ec4cc36b7c2610dba227697c.tar.bz2
USB: serial: reimplement generic fifo-based writes
Reimplement fifo-based writes in the generic driver using a multiple pre-allocated urb scheme. In contrast to multi-urb writes, no allocations (of urbs or buffers) are made during run-time and there is less pressure on the host stack queues as currently only two urbs are used (implementation is generic and can handle more than two urbs as well, though). Initial tests using ftdi_sio show that the implementation achieves the same (maximum) throughput at high baudrates as multi-urb writes. The CPU usage is much lower than for multi-urb writes for small write requests and only slightly higher for large (e.g. 2k) requests (due to extra copy via fifo?). Also outperforms multi-urb writes for small write requests on an embedded arm-9 system, where multi-urb writes are CPU-bound at high baudrates (perf reveals that a lot of time is spent in the host stack enqueue function -- could perhaps be a bug as well). Keeping the original write_urb, buffer and flag for now as there are other drivers depending on them. Signed-off-by: Johan Hovold <jhovold@gmail.com> Signed-off-by: Greg Kroah-Hartman <gregkh@suse.de>
Diffstat (limited to 'tools/perf')
0 files changed, 0 insertions, 0 deletions