PATCH: Fix broken PIO with libata
The revaldiation in 2.6.17-rc has broken support for PIO only devices. This is fairly unusual in the SATA world but showed up rather more promptly with the added PATA drivers.
The patch fixes two specific problem cases
#1 If you issue a DMA command via pass through the libata core blindly
issues a DMA command and calls DMA methods that are not present on PIO only controllers causing an Oops. The patch does a simple check and reject of a DMA command in PIO only cases.
#2 The core sets ATA_DFLAG_PIO to indicate PIO commands should be used
on this channel. This same information is available in dev->dma_mode but for some reason we get two sources of the info. The ATA_DFLAG_PIO is set once during setup and then cleared but not re-computed by the revalidate function. This causes DMA commands to be issued when PIO would be and usually an Oops or hang
Also contains a related bracketing fix
drivers/scsi/libata-core.c +4 -0
@@ -1757,6 +1763,10 @@return rc; } +/* This is cleared by the revalidation */ +if (dev->xfer_shift ==ATA_SHIFT_PIO ) + dev->flags |=ATA_DFLAG_PIO ; +DPRINTK ("xfer_shift=%u, xfer_mode=0x%x\n" , dev->xfer_shift, (int )dev->xfer_mode);
drivers/scsi/libata-scsi.c +7 -1
@@ -1944,7 +1944,7 @@return 0 ; dpofua =0 ; -if (ata_dev_supports_fua (args->id) && dev->flags &ATA_DFLAG_LBA48 && +if (ata_dev_supports_fua (args->id) && (dev->flags &ATA_DFLAG_LBA48 ) && (!(dev->flags &ATA_DFLAG_PIO ) || dev->multi_count)) dpofua =1 <<4 ;
@@ -2414,9 +2414,15 @@ {struct ata_taskfile *tf = &(qc->tf);struct scsi_cmnd *cmd = qc->scsicmd; +struct ata_device *dev = qc->dev; +struct ata_port *ap = qc->ap;if ((tf->protocol =ata_scsi_map_proto (scsicmd[1 ])) ==ATA_PROT_UNKNOWN ) goto invalid_fld; + +/* We may not issue DMA commands if no DMA mode is set */ +if (tf->protocol ==ATA_PROT_DMA && dev->dma_mode ==0 ) + goto invalid_fld;if (scsicmd[1 ] &0xe0 )/* PIO multi not supported yet */
Quoted 3 lines · Alan Cox
The revaldiation in 2.6.17-rc has broken support for PIO only devices. This is fairly unusual in the SATA world but showed up rather more promptly with the added PATA drivers.
Excellent; this seems to have solved my oops on CF card insertion problem. :)
However, I still have a problem with pata_pcmcia (that I actually experienced also with the ide-cs driver) where sustained reading or writing to the CF card spikes the CPU with nearly 100% system time. Here's a few seconds of 'vmstat 5' with a 'cat /dev/sdb > /dev/null' in the middle of it:
procs -----------memory---------- ---swap-- -----io---- --system-- ----cpu----
r b swpd free buff cache si so bi bo in cs us sy id wa 1 1 0 502608 13304 254740 0 0 164 42 1089 618 7 2 85 5 1 1 0 495440 20472 254740 0 0 1434 7 1078 1079 1 95 0 4 1 1 0 487264 28664 254740 0 0 1638 8 1073 1063 1 95 0 4 1 1 0 479940 35832 254740 0 0 1434 50 1077 1087 1 95 0 4 1 1 0 472320 43000 254740 0 0 1434 15 1078 1093 10 86 0 4 1 1 0 465104 50168 254740 0 0 1434 56 1077 1053 6 90 0 4 1 1 0 456944 58360 254740 0 0 1638 8 1072 1063 1 94 0 5 1 1 0 449744 65528 254740 0 0 1434 3 1073 1069 1 94 0 5 1 1 0 441600 73720 254740 0 0 1638 5 1072 1071 1 95 0 4 1 1 0 434020 80888 254740 0 0 1434 0 1074 1087 11 85 0 4 0 0 0 428896 86008 254740 0 0 1024 3 1184 1313 4 83 11 3 0 0 0 428896 86008 254740 0 0 0 5 1105 1057 1 0 98 0
Signature · 3 lines
Kevin 'radsaq' Radloff [email protected] http://thesaq.com/
Quoted 3 lines · Kevin Radloff
However, I still have a problem with
pata_pcmcia(that I actually experienced also with the ide-cs driver) where sustained reading or writing to the CF card spikes the CPU with nearly 100% system time.
That is normal. The PCMCIA devices don't support DMA. As a result of this the processor has to fetch each byte itself over the ISA speed PCMCIA bus link.
Alan
Quoted 8 lines · Alan Cox
Quoted 3 lines · Kevin Radloff
However, I still have a problem with
pata_pcmcia(that I actually experienced also with the ide-cs driver) where sustained reading or writing to the CF card spikes the CPU with nearly 100% system time.That is normal. The PCMCIA devices don't support DMA. As a result of this the processor has to fetch each byte itself over the ISA speed PCMCIA bus link.
Hrm, as I recall that only started happening with ide-cs sometime in the single digits of 2.6.x.. And note that it's only maxing out at about 1.5MB/s. Should that saturate my laptop's 1.1GHz Pentium M processor?
Signature · 3 lines
Kevin 'radsaq' Radloff [email protected] http://thesaq.com/
Quoted 6 lines · Kevin Radloff
Quoted 5 lines · Alan Cox
Quoted 3 lines · Kevin Radloff
However, I still have a problem with
pata_pcmcia(that I actually experienced also with the ide-cs driver) where sustained reading or writing to the CF card spikes the CPU with nearly 100% system time.
Quoted 8 lines
Quoted 3 lines
That is normal. The PCMCIA devices don't support DMA. As a result of this the processor has to fetch each byte itself over the ISA speed PCMCIA bus link.
Hrm, as I recall that only started happening with ide-cs sometime in the single digits of 2.6.x.. And note that it's only maxing out at about 1.5MB/s. Should that saturate my laptop's 1.1GHz Pentium M processor?
Doing data xfer using PIO rather than DMA definitely eats tons of CPU cycles.
Jeff
Quoted 11 lines · Jeff Garzik
Quoted 10 lines · Kevin Radloff
Quoted 8 lines · Alan Cox
Quoted 3 lines · Kevin Radloff
However, I still have a problem with
pata_pcmcia(that I actually experienced also with the ide-cs driver) where sustained reading or writing to the CF card spikes the CPU with nearly 100% system time.That is normal. The PCMCIA devices don't support DMA. As a result of this the processor has to fetch each byte itself over the ISA speed PCMCIA bus link.
Quoted 6 lines
Quoted 4 lines
Hrm, as I recall that only started happening with ide-cs sometime in the single digits of 2.6.x.. And note that it's only maxing out at about 1.5MB/s. Should that saturate my laptop's 1.1GHz Pentium M processor?
Doing data xfer using PIO rather than DMA definitely eats tons of CPU cycles.
Yeap, in addition, if doing real PIO (unbuffered by the HBA), the time it takes is soley determined by what PIO mode is in use. It doesn't matter how fast the CPU is. Faster CPUs only end up wasting more cycles. :-(
Signature · 1 line
tejun
Quoted 17 lines · Tejun Heo
Quoted 16 lines · Jeff Garzik
Quoted 14 lines · Kevin Radloff
Quoted 8 lines · Alan Cox
Quoted 3 lines · Kevin Radloff
However, I still have a problem with
pata_pcmcia(that I actually experienced also with the ide-cs driver) where sustained reading or writing to the CF card spikes the CPU with nearly 100% system time.That is normal. The PCMCIA devices don't support DMA. As a result of this the processor has to fetch each byte itself over the ISA speed PCMCIA bus link.
Hrm, as I recall that only started happening with ide-cs sometime in the single digits of 2.6.x.. And note that it's only maxing out at about 1.5MB/s. Should that saturate my laptop's 1.1GHz Pentium M processor?
Quoted 7 lines
Quoted 2 lines
Doing data xfer using PIO rather than DMA definitely eats tons of CPU cycles.
Yeap, in addition, if doing real PIO (unbuffered by the HBA), the time it takes is soley determined by what PIO mode is in use. It doesn't matter how fast the CPU is. Faster CPUs only end up wasting more cycles. :-(
(oops, hit 'reply', but given the incredible importance of my response... ;P)
Ah, well then never mind. ;) I just have a dim memory of it being different a long time ago. At least it works now. :D
Signature · 3 lines
Kevin 'radsaq' Radloff [email protected] http://thesaq.com/
Quoted 4 lines · Kevin Radloff
Hrm, as I recall that only started happening with ide-cs sometime in the single digits of 2.6.x.. And note that it's only maxing out at about 1.5MB/s. Should that saturate my laptop's 1.1GHz Pentium M processor?
Yes. It is possible that adding 32bit I/O support will boost this a little but not much, and it may not work on all cards. The PCMCIA bus is ISA speed, 1.5MB/sec is pretty much flat out
Alan
Quoted 6 lines · Alan Cox
#2 The core sets
ATA_DFLAG_PIOto indicate PIO commands should be used
on this channel. This same information is available indev->dma_modebut for some reason we get two sources of the info. TheATA_DFLAG_PIOis set once during setup and then cleared but not re-computed by the revalidate function. This causes DMA commands to be issued when PIO would be and usually an Oops or hang
Hmmm... I tried to fix this problem in the following commit. With it, ATA_DFLAG_PIO isn't cleared over ata_dev_configure(). Only ata_dev_set_mode() is allowed to diddle with it and does about the same thing as your patch does.
diff-tree ea1dd4e13010eb9dd5ffb4bfabbb472bc238bebb (from 198e0fed9e59461fc1890dd
Author: Tejun Heo <[email protected]> Date: Sun Apr 2 18:51:53 2006 +0900
[PATCH] libata: clear only affected flags during ata_dev_configure()
ata_dev_configure() should not clear dynamic device flags determined elsewhere. Lower eight bits are reserved for feature flags, define ATA_DFLAG_CFG_MASK and clear only those bits before configuring device. Without this patch, ATA_DFLAG_PIO gets turned off during revalidation making PIO mode unuseable.
Quoted 1 line
Also contains a related bracketing fix
Is this agreed upon? I tend to omit almost all unnecessary (by operator precedence) parenthesis, so in new EH and all other stuff, the "a && b & c" sort of lines are abundant. If this is something that's agreed upon, I can do a clean sweep over those.
Signature · 1 line
tejun
Quoted 4 lines · Tejun Heo
Is this agreed upon? I tend to omit almost all unnecessary (by operator precedence) parenthesis, so in new EH and all other stuff, the "a && b & c" sort of lines are abundant. If this is something that's agreed upon, I can do a clean sweep over those.
More parens == easier to review. So
a && b & c should be a && (b & c)
to clearly delineate the separate expressions to the human eye, and also make it clear to the reader that the '&' is intended, and not a typo that should have been '&&'.
Anytime you see a long string of 'if' conditions, and the operators vary, add parents for readability.
Jeff
Quoted 11 lines · Jeff Garzik
Quoted 4 lines · Tejun Heo
Is this agreed upon? I tend to omit almost all unnecessary (by operator precedence) parenthesis, so in new EH and all other stuff, the "a && b & c" sort of lines are abundant. If this is something that's agreed upon, I can do a clean sweep over those.
More parens == easier to review. So
a && b & c should be a && (b & c)
Understood. Usually, my rule is something like doing the least maximizes consistency (style-wise) thus increasing readability in the end, but rules usually suck, don't they? What fits the most eyes is the best, I guess.
Quoted 7 lines
to clearly delineate the separate expressions to the human eye, and also make it clear to the reader that the '&' is intended, and not a typo that should have been '&&'.
Anytime you see a long string of 'if' conditions, and the operators vary, add parents for readability.
Yeap, will do, from now on.
Signature · 1 line
tejun
Quoted 1 line · Jeff Garzik
More parens == easier to review. So
Well... mostly, but not always. I agree about a && (b & c), but I find the parens in ((a == b) && (c != d)) distracting. And once we get to three or more parens in a row, I have to start counting carefully to see where they match up.
Paul.
Quoted 8 lines · Tejun Heo
Quoted 6 lines · Alan Cox
#2 The core sets
ATA_DFLAG_PIOto indicate PIO commands should be used
on this channel. This same information is available indev->dma_modebut for some reason we get two sources of the info. TheATA_DFLAG_PIOis set once during setup and then cleared but not re-computed by the revalidate function. This causes DMA commands to be issued when PIO would be and usually an Oops or hang
Quoted 4 lines
Hmmm... I tried to fix this problem in the following commit. With it,
ATA_DFLAG_PIOisn't cleared overata_dev_configure(). Onlyata_dev_set_mode()is allowed to diddle with it and does about the same thing as your patch does.
I presume he's looking at what users will hit when 2.6.17 is released...
Jeff