palm_bk3710: UDMA performance fix
PATCH 2/2 palm_bk3710: UDMA performance fix
From: David Brownell <[email protected]>
Fix UDMA throughput bug: tCYC averages t2CYCTYP/2, but the code previously assumed it was the same as t2CYCTYP. (That is, it was using just one clock edge, not both.) Move the table's type declaration so it's adjacent to the table, making it more clear what those numbers mean.
On one system this change increased throughput by almost 4x: UDMA/66 sometimes topped 23 MB/sec (on a drive known to do much better). On another system it was around a 10% win (UDMA/66 up to 7+ MB/sec).
The difference might be caused by the ratio between memory and IDE clocks. In the system with large speedup, this was exactly 2 (as a workaround for a rev 1.1 silicon bug). The other system used a more standard ratio of 1.63 (and rev 2.1 silicon) ... clock domain synch might have some issues, they're not unheard-of.
--- drivers/ide/palm_bk3710.c | 26 ++++++++++++-------------- 1 file changed, 12 insertions(+), 14 deletions(-)
drivers/ide/palm_bk3710.c +12 -14
@@ -39,14 +39,6 @@/* Primary Control Offset */ #define IDE_PALM_ATA_PRI_CTL_OFFSET 0x3F6 -/* -* PalmChip 3710 IDE Controller UDMA timing structure Definition -*/ -struct palm_bk3710_udmatiming { -unsigned int rptime;/* Ready to pause time */ -unsigned int cycletime;/* Cycle Time */ -}; -#define BK3710_BMICP 0x00#define BK3710_BMISP 0x02#define BK3710_BMIDTP 0x04
@@ -67,13 +59,19 @@ struct palm_bk3710_udmatiming {static unsigned ideclk_period;/* in nanoseconds */ +struct palm_bk3710_udmatiming { +unsigned int rptime;/* tRP -- Ready to pause time (nsec) */ +unsigned int cycletime;/* tCYCTYP2/2 -- avg Cycle Time (nsec) */ +/* tENV is always a minimum of 20 nsec */ +}; +static const struct palm_bk3710_udmatiming palm_bk3710_udmatimings[6 ] = { - {160 ,240 },/* UDMA Mode 0 */ - {125 ,160 },/* UDMA Mode 1 */ - {100 ,120 },/* UDMA Mode 2 */ - {100 ,90 },/* UDMA Mode 3 */ - {100 ,60 },/* UDMA Mode 4 */ - {85 ,40 },/* UDMA Mode 5 */ + {160 ,240 /2 ,},/* UDMA Mode 0 */ + {125 ,160 /2 ,},/* UDMA Mode 1 */ + {100 ,120 /2 ,},/* UDMA Mode 2 */ + {100 ,90 /2 ,},/* UDMA Mode 3 */ + {100 ,60 /2 ,},/* UDMA Mode 4 */ + {85 ,40 /2 ,},/* UDMA Mode 5 */ }; static void palm_bk3710_setudmamode(void __iomem *base, unsigned int dev,
Quoted 2 lines · David Brownell
Fix UDMA throughput bug: tCYC averages t2CYCTYP/2, but the code previously assumed it was the same as t2CYCTYP.
Wow, thanks for finding it!
IIUC however, this should have only affected UDMA writes, not reads
because on reads the device controls the strobe timing.
Quoted 1 line
(That is, it was using just one clock edge, not both.)
There's no way only one clock edge could have been used since it would
have resulted in CRC errors, so this comment is not to the point.
Quoted 3 lines
On one system this change increased throughput by almost 4x: UDMA/66 sometimes topped 23 MB/sec (on a drive known to do much better). On another system it was around a 10% win (UDMA/66 up to 7+ MB/sec).
It's interesting that on my DM6467 EVM UDMA/66 reads topped at about 29-30
MB/s even without this patch (measuread with hdparm), and on DM6446 EVM they were only slightly slower...
Quoted 5 lines
The difference might be caused by the ratio between memory and IDE clocks. In the system with large speedup, this was exactly 2 (as a workaround for a rev 1.1 silicon bug). The other system used a more standard ratio of 1.63 (and rev 2.1 silicon) ... clock domain synch might have some issues, they're not unheard-of.
Interesting...
Quoted 1 line
Signed-off-by David Brownell <[email protected]>
The patch itself is:
MBR, Sergei
A message that is not in the archive (sent off-list or not yet received): [email protected]
Quoted 8 lines · Sergei Shtylyov
Quoted 2 lines · David Brownell
Fix UDMA throughput bug: tCYC averages t2CYCTYP/2, but the code previously assumed it was the same as t2CYCTYP.
Wow, thanks for finding it!
IIUC however, this should have only affected UDMA writes, not reads
because on reads the device controls the strobe timing.
Odd, I certainly noticed UDMA *read* performance changes...
Quoted 4 lines
Quoted 1 line
(That is, it was using just one clock edge, not both.)
There's no way only one clock edge could have been used since it would
have resulted in CRC errors, so this comment is not to the point.
That's correct of course ... better to say instead that the throughput is exactly half what it should have been, since the clock was using the wrong period.
Quoted 7 lines
Quoted 3 lines
On one system this change increased throughput by almost 4x: UDMA/66 sometimes topped 23 MB/sec (on a drive known to do much better). On another system it was around a 10% win (UDMA/66 up to 7+ MB/sec).
It's interesting that on my DM6467 EVM UDMA/66 reads topped at about 29-30
MB/s even without this patch (measuread with hdparm), and on DM6446 EVM they were only slightly slower...
I've yet to see transfer rates that fast. Even drives that have topped 40 MB/sec on a PC don't get that speed.
Quoted 7 lines
Quoted 5 lines
The difference might be caused by the ratio between memory and IDE clocks. In the system with large speedup, this was exactly 2 (as a workaround for a rev 1.1 silicon bug). The other system used a more standard ratio of 1.63 (and rev 2.1 silicon) ... clock domain synch might have some issues, they're not unheard-of.
Interesting...
Yeah, but not what I'd call conclusive. The two drives in question were both TravelStars ... one older 20 MByte, and a less old 60 MByte. So i have no reason to think either would have trouble going "fast".
Quoted 9 lines
Quoted 1 line
Signed-off-by David Brownell <[email protected]>The patch itself is:
Acked-by Sergei Shtylyov <[email protected]>MBR, Sergei
Bartlomiej Zolnierkiewicz
AppliedQuoted 28 lines · On Tuesday 21 April 2009 12:02:22 Sergei Shtylyov wrote:
Quoted 2 lines · David Brownell
Fix UDMA throughput bug: tCYC averages t2CYCTYP/2, but the code previously assumed it was the same as t2CYCTYP.
Wow, thanks for finding it!
IIUC however, this should have only affected UDMA writes, not reads
because on reads the device controls the strobe timing.Quoted 1 line
(That is, it was using just one clock edge, not both.)
There's no way only one clock edge could have been used since it would
have resulted in CRC errors, so this comment is not to the point.Quoted 3 lines
On one system this change increased throughput by almost 4x: UDMA/66 sometimes topped 23 MB/sec (on a drive known to do much better). On another system it was around a 10% win (UDMA/66 up to 7+ MB/sec).
It's interesting that on my DM6467 EVM UDMA/66 reads topped at about 29-30
MB/s even without this patch (measuread with hdparm), and on DM6446 EVM they were only slightly slower...Quoted 5 lines
The difference might be caused by the ratio between memory and IDE clocks. In the system with large speedup, this was exactly 2 (as a workaround for a rev 1.1 silicon bug). The other system used a more standard ratio of 1.63 (and rev 2.1 silicon) ... clock domain synch might have some issues, they're not unheard-of.
Quoted 7 lines
Interesting...
Quoted 1 line
Signed-off-by David Brownell <[email protected]>The patch itself is:
Acked-by Sergei Shtylyov <[email protected]>
applied
Quoted 9 lines · Bartlomiej Zolnierkiewicz
Quoted 8 lines · On Tuesday 21 April 2009 12:02:22 Sergei Shtylyov wrote:
Quoted 2 lines · David Brownell
Fix UDMA throughput bug: tCYC averages t2CYCTYP/2, but the code previously assumed it was the same as t2CYCTYP.
Wow, thanks for finding it!
....
Quoted 7 lines
Quoted 5 lines
Quoted 1 line
Signed-off-by David Brownell <[email protected]>The patch itself is:
Acked-by Sergei Shtylyov <[email protected]>applied
Thanks. Further experimentation suggests (a) doing timing computations in picoseconds, rather than nanoseconds, gives better answers and makes that "slow" system faster by 25% or more; and (b) increasing all the UDMA timings by one clock cycle gives another few percent.
Now (a) makes obvious sense; but (b) seems odd ...
- Dave
A message that is not in the archive (sent off-list or not yet received): 200904222026.11495.bzolnier-Re5JQEeQqe8AvxtiuMwx3w@public.gmane.org
Quoted 1 line · Bartlomiej Zolnierkiewicz
applied
By the way ... what about the first patch,
which removed accesses to all those non-existent
registers and bitfields?
Quoted 6 lines · On Wednesday 22 April 2009 22:43:14 David Brownell wrote:
Quoted 1 line · Bartlomiej Zolnierkiewicz
applied
By the way ... what about the first patch,
which removed accesses to all those non-existent
registers and bitfields?
I didn't see any discussion on it and it looked less urgent / more risky (it is not uncommon for documentation to lack some data) than patch #2.
Should it also go upstream for 2.6.30?
Sergei?
A message that is not in the archive (sent off-list or not yet received): 200904222305.54820.bzolnier-Re5JQEeQqe8AvxtiuMwx3w@public.gmane.org
Quoted 7 lines · Bartlomiej Zolnierkiewicz
Quoted 4 lines
By the way ... what about the first patch,
which removed accesses to all those non-existent
registers and bitfields?I didn't see any discussion on it and it looked less urgent / more risky (it is not uncommon for documentation to lack some data) than patch #2.
It's uncommon for TI's documentation to be that far off, for that long, though. "Lacking" docs for 50% of the registers, for several years ... doesn't make sense.
The current reset handling is clearly broken: the docs are quite explicit that the controller doesn't drive the reset signal, it's got to be done through a GPIO. The board designs match that part of the docs. The code is thus contrary to *all* other documentation.
Early DaVinci drivers sometimes exhibited a flagrant disregard for chip documentation. This driver is from about that era. Maybe it started from prototypes using the a different controller design, for example.
Quoted 1 line
Should it also go upstream for 2.6.30?
I can't say either patch would be urgent for 2.6.30, but of course it's good that bugfixes merge ASAP.
Since we're only at RC3, I'd be inclined to push both up right now. If we were at RC6 or so, I'd hold off till the next merge window.
- Dave
Bartlomiej Zolnierkiewicz
AppliedQuoted 25 lines · On Thursday 23 April 2009 00:29:23 David Brownell wrote:
Quoted 7 lines · Bartlomiej Zolnierkiewicz
Quoted 4 lines
By the way ... what about the first patch,
which removed accesses to all those non-existent
registers and bitfields?I didn't see any discussion on it and it looked less urgent / more risky (it is not uncommon for documentation to lack some data) than patch #2.
It's uncommon for TI's documentation to be that far off, for that long, though. "Lacking" docs for 50% of the registers, for several years ... doesn't make sense.
The current reset handling is clearly broken: the docs are quite explicit that the controller doesn't drive the reset signal, it's got to be done through a GPIO. The board designs match that part of the docs. The code is thus contrary to *all* other documentation.
Early DaVinci drivers sometimes exhibited a flagrant disregard for chip documentation. This driver is from about that era. Maybe it started from prototypes using the a different controller design, for example.
Quoted 8 lines
Quoted 1 line
Should it also go upstream for 2.6.30?
I can't say either patch would be urgent for 2.6.30, but of course it's good that bugfixes merge ASAP.
Since we're only at RC3, I'd be inclined to push both up right now. If we were at RC6 or so, I'd hold off till the next merge window.
Thanks for the detailed explanation.
I applied the patch to ide-2.6.git/for-linus.