Organized summary of the pacthes
A message that is not in the archive (sent off-list or not yet received): [email protected]
Quoted 5 lines · Albert Lee
Summary of the pacthes for your review (as of 6/28/2005):
1.atapi_pio_bytes()fixes
=> Three patches pending:#1-2. __atapi_pio_bytes(): if condition fix http://marc.theaimsgroup.com/?l=linux-ide&m=111838913812842&w=2
I'll probably apply this
Quoted 7 lines
#1-3.
__atapi_pio_bytes(): trailing data handling fix
http://marc.theaimsgroup.com/?l=linux-ide&m=111838955410270&w=2#1-4.
__atapi_pio_bytes(): odd-length data handling fix
http://marc.theaimsgroup.com/?l=linux-ide&m=111838987630460&w=2
Patch #1-4 will overrun the odd-length buffer by one byte. :( Maybe we can just ignore the last byte of odd-length buffer?
These two patches make me REALLY nervous. This overrun business needs to be handled in a different way.
For DMA, we will want to copy 0-3 odd bytes into a 4-byte buffer, and then make that 4-byte buffer than final DMA segment.
For PIO, we might as well use the same buffer, and copy the last 0-3 (or 0-1?) odd bytes.
Never overrun.
Quoted 8 lines
2.
atapi_packet_task()assertion failed fix
=> Revised patch submitted:http://marc.theaimsgroup.com/?l=linux-ide&m=111830130223323&w=2=> Need more revision to remove the 'case
ATA_PROT_ATAPI' handling in it.
(It conflicts with patch #3-2 below:
http://marc.theaimsgroup.com/?l=linux-ide&m=111951880730921&w=2)
Quoted 4 lines
3.
ata_pio_task()race condition fixes.
=> Two patches submitted. Waiting for review:#3-1 http://marc.theaimsgroup.com/?l=linux-ide&m=111958527005103&w=2 #3-2 http://marc.theaimsgroup.com/?l=linux-ide&m=111951880730921&w=2
These are probably OK.
However, I am pondering scrapping all the polling code, since on SATA, interrupt-driven mode is much more desirable.
If I continue to use polling, I will (after further review) apply your patches.
Luckily PIO data paths are not used by users yet (disabled by default), so these problems are not urgent.
Quoted 5 lines
=======
4.ata_qc_complete()race condition fix.
=> Inata_qc_complete(),"qc->flags &= ~ATA_QCFLAG_ACTIVE;" should be _before_ "rc = qc->complete_fn(qc, drv_stat);".
I agree.
Quoted 1 line
=> Otherwise
ata_qc_complete()might race withatapi_request_sense().
If ata_qc_complete() is racing with atapi_request_sense(), more synchronization is needed. ATAPI is violating the host state machine, that wants fixing.
Quoted 3 lines
5.
pata_pdc2027x0.62 update
=> http://marc.theaimsgroup.com/?l=linux-ide&m=111474486330692&w=2
=> Need more revision according to libata-dev tree in git.
I think this patch is not needed, once ATAPI DMA/PIO is fixed properly as described above.
Quoted 2 lines
6. IDENTIFY DEVICE detects phantom drive problem
http://marc.theaimsgroup.com/?l=linux-ide&m=111622547309019&w=2
To answer the question in your email here, we should separate hardware Status register value from a new 'force_error' parameter. This would require a new test in the code
int have_error = forced_error || (tf->status & ATA_ERR);
Eventually we need to separate out errors into separate classes (hi Alan, hi Tejun), such as dma-error, bus-error, device-error, etc. When that is done, the code paths described in your email must change anyway.
Jeff
Quoted 5 lines · Jeff Garzik
These two patches make me REALLY nervous. This overrun business needs to be handled in a different way.
For DMA, we will want to copy 0-3 odd bytes into a 4-byte buffer, and then make that 4-byte buffer than final DMA segment.
You may need 16 or 32 byte alignment for some hardware on the host side.
Quoted 2 lines
However, I am pondering scrapping all the polling code, since on SATA, interrupt-driven mode is much more desirable.
You need both interrupt and polled command issue. ACPI needs to be able to issue taskfiles to the IDE drive fairly early during resume. Right now this isn't done at all. If polled is there it might also prove a useful way to reissue a command when you think the IRQ routing is hosed
8)
There is a really nasty case here that the old IDE layer totally screwed
up: You may need to sequence speed changes into error recovery (BAD CRC) which is coming from a timeout and irq context. In the sata layer at least its from an eh thread.
Alan
Quoted 6 lines · Alan Cox
Quoted 5 lines · Jeff Garzik
These two patches make me REALLY nervous. This overrun business needs to be handled in a different way.
For DMA, we will want to copy 0-3 odd bytes into a 4-byte buffer, and then make that 4-byte buffer than final DMA segment.
Indeed, all chipsets I have seen require that the dma be aligned on 4 byte boundaries, and the transfer be a multiple of 4 bytes. Maybe I have not seen that many, but all I have seen do seem to conform to the original Intel specs (or what was SFF-8038i). Normally this is not an issue all the buffers are powers of 2, 512 bytes or larger.
Also, we do need to honor the 64k boundary, as at least in this case, the underlying device is indeed PATA.
Quoted 2 lines
You may need 16 or 32 byte alignment for some hardware on the host side.
Where would one specify this, at the block layer? Then bounce the portions that extend outside this region as a separate transfer to device (including the odd bytes above)?
Or are you referring to internal buffer creation only?
Quoted 9 lines
Quoted 3 lines
However, I am pondering scrapping all the polling code, since on SATA, interrupt-driven mode is much more desirable.
You need both interrupt and polled command issue. ACPI needs to be able to issue taskfiles to the IDE drive fairly early during resume. Right now this isn't done at all. If polled is there it might also prove a useful way to reissue a command when you think the IRQ routing is hosed
8)
This would be an enhancement, rather than something required currently, correct?
Quoted 3 lines
There is a really nasty case here that the old IDE layer totally screwed
up: You may need to sequence speed changes into error recovery (BAD CRC)
Likewise? Dynamic fallback would be great, and should print a big fat "your cables or drive are failing" message when the fallback was initiated.
Quoted 3 lines
which is coming from a timeout and irq context. In the sata layer at least its from an eh thread.
++doug
Quoted 11 lines · Doug Maxey
Quoted 8 lines
Quoted 2 lines
However, I am pondering scrapping all the polling code, since on SATA, interrupt-driven mode is much more desirable.
You need both interrupt and polled command issue. ACPI needs to be able to issue taskfiles to the IDE drive fairly early during resume. Right now this isn't done at all. If polled is there it might also prove a useful way to reissue a command when you think the IRQ routing is hosed
8)This would be an enhancement, rather than something required currently, correct?
Switching over to interrupt-driven for SATA is basically a requirement.
Since the polling code is what exists now, it MAY be easier to leave it in, which puts "leave in polling code" in the enhancement category.
Jeff
Jeff,
Quoted 21 lines
Quoted 5 lines · Albert Lee
Summary of the pacthes for your review (as of 6/28/2005):
1.atapi_pio_bytes()fixes
=> Three patches pending:#1-2. __atapi_pio_bytes(): if condition fix http://marc.theaimsgroup.com/?l=linux-ide&m=111838913812842&w=2I'll probably apply this
Quoted 7 lines
#1-3.
__atapi_pio_bytes(): trailing data handling fix
http://marc.theaimsgroup.com/?l=linux-ide&m=111838955410270&w=2#1-4.
__atapi_pio_bytes(): odd-length data handling fix
http://marc.theaimsgroup.com/?l=linux-ide&m=111838987630460&w=2
Patch #1-4 will overrun the odd-length buffer by one byte. :( Maybe we can just ignore the last byte of odd-length buffer?
Quoted 8 lines
These two patches make me REALLY nervous. This overrun business needs to be handled in a different way.
For DMA, we will want to copy 0-3 odd bytes into a 4-byte buffer, and then make that 4-byte buffer than final DMA segment.
For PIO, we might as well use the same buffer, and copy the last 0-3 (or 0-1?) odd bytes.
Never overrun.
Thanks for the review and tolerance for the newbie blindness. I'll revise/resubmit patch 1-3 and 1-4 per your advice.
I will also study the DMA padding/alignment advised by you and Alan.
Quoted 16 lines
Quoted 10 lines
2.
atapi_packet_task()assertion failed fix
=> Revised patch submitted:http://marc.theaimsgroup.com/?l=linux-ide&m=111830130223323&w=2=> Need more revision to remove the 'case
ATA_PROT_ATAPI' handling
in it.
(It conflicts with patch #3-2 below:
http://marc.theaimsgroup.com/?l=linux-ide&m=111951880730921&w=2)Quoted 4 lines
3.
ata_pio_task()race condition fixes.
=> Two patches submitted. Waiting for review:#3-1 http://marc.theaimsgroup.com/?l=linux-ide&m=111958527005103&w=2 #3-2 http://marc.theaimsgroup.com/?l=linux-ide&m=111951880730921&w=2
Quoted 3 lines
These are probably OK.
However, I am pondering scrapping all the polling code, since on SATA, interrupt-driven mode is much more desirable.
Is there any draft plan or design for the interrupt driven PIO code available? Also as Alan said, PIO polling would be useful for some situation. Any plan for libata to support both PIO polling mode and the interrupt driven mode?
If permitted, I would like to take this opportunity to participate in the interrupt driven PIO task, under your design guideline.
Quoted 5 lines
If I continue to use polling, I will (after further review) apply your patches.
Luckily PIO data paths are not used by users yet (disabled by default), so these problems are not urgent.
The rare race conditions only occur under hours of heavy stress test.
Based on the testing result, the libata PIO polling code is quite robust. It has been tested with harddisks, cd-rom drives and even an old ATAPI Zip 100 floppy drive, which supports PIO only; and it works fine. (Except high CPU usage due to the nature of polling.)
Quoted 9 lines
Quoted 6 lines
=======
4.ata_qc_complete()race condition fix.
=> Inata_qc_complete(),"qc->flags &= ~ATA_QCFLAG_ACTIVE;" should be _before_ "rc = qc->complete_fn(qc, drv_stat);".I agree.
Thanks, will submit the patch.
Quoted 6 lines
Quoted 3 lines
=> Otherwise
ata_qc_complete()might race withatapi_request_sense().If
ata_qc_complete()is racing withatapi_request_sense(), more synchronization is needed. ATAPI is violating the host state machine, that wants fixing.
It's false alarm.
After moving "qc->flags &= ~ATA_QCFLAG_ACTIVE" described above, no race between ata_qc_complete() and atapi_request_sense() is ever seen. Please ignore my false alarm.
Quoted 8 lines
Quoted 5 lines
5.
pata_pdc2027x0.62 update
=> http://marc.theaimsgroup.com/?l=linux-ide&m=111474486330692&w=2
=> Need more revision according to libata-dev tree in git.I think this patch is not needed, once ATAPI DMA/PIO is fixed properly as described above.
Mmm, quite possibly. Will defer this patch and re-test after ATAPI DMA/PIO fixes are ready.
Quoted 11 lines
Quoted 4 lines
6. IDENTIFY DEVICE detects phantom drive problem
http://marc.theaimsgroup.com/?l=linux-ide&m=111622547309019&w=2To answer the question in your email here, we should separate hardware Status register value from a new '
force_error' parameter. This would require a new test in the codeint have_error = forced_error || (tf->status & ATA_ERR);Eventually we need to separate out errors into separate classes (hi Alan, hi Tejun), such as dma-error, bus-error, device-error, etc. When that is done, the code paths described in your email must change anyway.
Understood. Please discard this patch.
Albert
By the way,
Why is the ide-driver in so slow?
Hdparm -tT gives 28mb/s in 2.6.12
Cat /dev/hda > /dev/null gives 2% user 25% sys 0% idle 73% IOWAIT
Hdparm -tT gives 38mb/s in 2.4.31
Cat /dev/hda > /dev/null gives 2% user 33% sys 65% idle
It feels like DMA is not being applied properly in 2.6.12.
Your comments please.
Thanks!
Quoted 61 lines · Albert Lee
Jeff,
Quoted 23 lines
Quoted 5 lines · Albert Lee
Summary of the pacthes for your review (as of 6/28/2005):
1.atapi_pio_bytes()fixes
=> Three patches pending:#1-2. __atapi_pio_bytes(): if condition fix http://marc.theaimsgroup.com/?l=linux-ide&m=111838913812842&w=2I'll probably apply this
Quoted 7 lines
#1-3.
__atapi_pio_bytes(): trailing data handling fix
http://marc.theaimsgroup.com/?l=linux-ide&m=111838955410270&w=2#1-4.
__atapi_pio_bytes(): odd-length data handling fix
http://marc.theaimsgroup.com/?l=linux-ide&m=111838987630460&w=2
Patch #1-4 will overrun the odd-length buffer by one byte. :( Maybe we can just ignore the last byte of odd-length buffer?Quoted 8 lines
These two patches make me REALLY nervous. This overrun business needs to be handled in a different way.
For DMA, we will want to copy 0-3 odd bytes into a 4-byte buffer, and then make that 4-byte buffer than final DMA segment.
For PIO, we might as well use the same buffer, and copy the last 0-3 (or 0-1?) odd bytes.
Never overrun.
Thanks for the review and tolerance for the newbie blindness. I'll revise/resubmit patch 1-3 and 1-4 per your advice.
I will also study the DMA padding/alignment advised by you and Alan.Quoted 17 lines
Quoted 10 lines
2.
atapi_packet_task()assertion failed fix
=> Revised patch submitted:http://marc.theaimsgroup.com/?l=linux-ide&m=111830130223323&w=2=> Need more revision to remove the 'case
ATA_PROT_ATAPI' handling
in it.
(It conflicts with patch #3-2 below:
http://marc.theaimsgroup.com/?l=linux-ide&m=111951880730921&w=2)Quoted 4 lines
3.
ata_pio_task()race condition fixes.
=> Two patches submitted. Waiting for review:#3-1 http://marc.theaimsgroup.com/?l=linux-ide&m=111958527005103&w=2 #3-2 http://marc.theaimsgroup.com/?l=linux-ide&m=111951880730921&w=2Quoted 3 lines
These are probably OK.
However, I am pondering scrapping all the polling code, since on SATA, interrupt-driven mode is much more desirable.
Quoted 8 lines
Is there any draft plan or design for the interrupt driven PIO code available? Also as Alan said, PIO polling would be useful for some situation. Any plan for libata to support both PIO polling mode and the interrupt driven mode?
If permitted, I would like to take this opportunity to participate in the interrupt driven PIO task, under your design guideline.
Hi, Albert.
For interrupt-driven PIO, IDE driver should be an example albeit quite
complex. Once interrupt-driven PIO is implemented, IMHO, generic polling IO can easily be added by calling interrupt handler periodically while the driver is expecting notification (any command is in-flight). Most low-level SATA driver interrupt handlers are already ready to handle spurious interrupts, and maybe we can add a flag to disable polling for those unable to swallow spurious interrupts if any such device/driver exsits.
Signature · 1 line
tejun
Quoted 7 lines · Jeff Garzik
Quoted 5 lines
=======
4.ata_qc_complete()race condition fix.
=> Inata_qc_complete(),"qc->flags &= ~ATA_QCFLAG_ACTIVE;" should be _before_ "rc = qc->complete_fn(qc, drv_stat);".I agree.
Does it need a membarrier as well?