On Wed, Jun 05, 2002 at 02:42:18PM -0700, Andrew Morton wrote: > A number of PCI video cards hang onto the PCI bus > for outrageous amounts of time. It gives better > benchmarks in Windows magazines. > > You may be getting bitten by this; dunno. Some xfree drivers > have an option to not use this (mis)feature. >
I was starting to fear some kind of hardware problem like this, but fortunately Matrox seems to be of a different kind... I have just received a message from Fernando Pablo Lopez-Lezcano including a patch to the drm driver which fixes it: BEFORE the patch: > fragment latency = 1.451247 ms > cpu latency = 1.160998 ms > 6.5ms (350)| > 1MS num_time_samples=44425 num_times_within_1ms=44075 factor=99.212155 > 2MS num_time_samples=44425 num_times_within_2ms=44075 factor=99.212155 AFTER the patch: > fragment latency = 1.451247 ms > cpu latency = 1.160998 ms > 1.7ms ( 0)| > 1MS num_time_samples=45925 num_times_within_1ms=45925 factor=100.000000 > 2MS num_time_samples=45925 num_times_within_2ms=45925 factor=100.000000 Fernando has posted to the list, but it seems his message is being delayed. This is a summary of it: On Wed, Jun 05, 2002 at 02:25:17PM -0700, Fernando Pablo Lopez-Lezcano wrote: > I did find some problems with the dri interface and found a > patch inside Jussi Laako's patchset ([EMAIL PROTECTED]) > that fixed it (I was testing with latencytest 0.42). See: > http://www.pp.song.fi/~visitor/linux/ And below is the part of the patch I have applied (into a 2.4.17 kernel source). Thank you all! Enrique. ------- diff -urN linux-2.4.19-pre7/drivers/char/drm/mga_dma.c linux-2.4.19-pre7-jl1-ll/drivers/char/drm/mga_dma.c --- linux-2.4.19-pre7/drivers/char/drm/mga_dma.c Wed Aug 8 19:42:15 2001 +++ linux-2.4.19-pre7-jl1-ll/drivers/char/drm/mga_dma.c Mon Apr 29 01:23:18 2002 @@ -61,6 +61,7 @@ MGA_WRITE8( MGA_CRTC_INDEX, 0 ); return 0; } + conditional_schedule(); udelay( 1 ); } @@ -80,6 +81,7 @@ for ( i = 0 ; i < dev_priv->usec_timeout ; i++ ) { status = MGA_READ( MGA_STATUS ) & MGA_DMA_IDLE_MASK; if ( status == MGA_ENDPRDMASTS ) return 0; + conditional_schedule(); udelay( 1 ); } @@ -121,6 +123,7 @@ * How about we clean up after ourselves? */ MGA_WRITE( MGA_RST, MGA_SOFTRESET ); + conditional_schedule(); /* We shouldn't get here anyway... */ udelay( 15 ); /* Wait at least 10 usecs */ MGA_WRITE( MGA_RST, 0 );
