VIA SATA Raid needs a long time to recover from suspend
I have been debugging a power management problem for a few days now, and I believe I have finally solved the problem. Because it involved patching the kernel, I felt I should share the fix here in hopes that it can be improved and/or integrated into future kernels. Right now I am running 2.6.14.2 on amd64, compiled myself, with the ubuntu breezy amd64 distribution.
First I'll state the fix. It involved changing two lines in include/linux/libata.h:
static inline u8 ata_busy_wait(struct ata_port *ap, unsigned int bits,
unsigned int max)
{
u8 status;
do {
udelay(100); <-- changed to 100
from 10
status = ata_chk_status(ap);
max--;} while ((status & bits) && (max > 0));
return status;
}
and:
static inline u8 ata_wait_idle(struct ata_port *ap) {
u8 status = ata_busy_wait(ap, ATA_BUSY | ATA_DRQ,
10000); <-- changed to 10,000 from 1,000
if (status & (ATA_BUSY | ATA_DRQ)) {
unsigned long l = ap->ioaddr.status_addr;
printk(KERN_WARNING
"ATA: abnormal status 0x%X on port 0x%lX\n",
status, l);
} return status;
}
The problem seems to be that my VIA SATA raid controller requires more time to recover from being suspended. It looks like the code in sata_via.c restores the task file after a resume, then calls ata_wait_idle to wait for the busy bit to clear. The problem was that this function timed out before the busy bit cleared, resulting in messages like this:
ATA: abnormal status 0x80 on port 0xE007
Then if there was an IO request made immediately after resuming, it would timeout and fail, because it was issued before the hardware was ready. Changing the timeout resolved this. I tried changing both the udelay and ata_busy_wait lines to increase the timeout, and it did not seem to matter which I changed, as long as the total timeout was increased by a factor of 100.
Since increasing the maximum timeout, suspend and hibernate work great for me. While experiencing this bug, it may have exposed another bug, which I will mention now in passing. As I said before, after a resume, if there was an IO request made immediately ( before the busy bit finally did clear ) it would timeout and fail. It seemed the kernel filled the buffer cache for the requested block with garbage rather than retry the read. It seems to me that at some point, the read should have been retried. The symptoms of this were:
1) When suspend.sh called resume.sh immediately after the echo mem > /sys/power/state line, then on resume, the read would fail in a block in the resierfs tree that was required to lookup the resume.sh file. This caused reiserfs to complain about errors in the node, and the script failed to execute. Further attempts to touch the script, even with ls -al /etc/acpi/resume.sh failed with EPERM. I would think that at worst, this should fail with EIO or something, not EPERM.
2) At one point I tried running echo mem > /sys/power/state ; df. After the resume, the IO read failed when trying to load df, and I got an error message saying the kernel could not execute the binary file. Further attempts to run df failed also. Other IO at this point was fine.
This leads me to think that when the IO failed, rather than inform the calling code of the failure, for example, with an EIO status, the buffer cache got filled with junk, and this should not happen. Either the operation should succeed, and the correct data be returned, or it should fail, and the caller should be informed of the failure, and not given incorrect data.
When the first IO immediately following the suspend failed, I got these messages:
[ 32.013538] ata1: command 0x35 timeout, stat 0x50 host_stat 0x1 [ 32.045510] ata2: command 0x35 timeout, stat 0x50 host_stat 0x1
As long as no IO was immediately requested after the resume ( i.e. if I echo mem > /sys/power/state on an otherwise idle system, rather than using suspend.sh ) then these errors did not happen, only the abnormal status messages did.
For reference, my system is configured as follows:
Motherboard: Asus K8V Deluxe
CPU: AMD Athlon 64 3200+
RAM: 1 GB of Corsair low latency pc3200 ddr sdram
Video: ATI Radeon 9800 Pro with a Samsung 930B 19 inch LCD display
Disks: 2 WD 36 gig SATA 10,000 rpm raptors in a raid0 configuration on the via sata raid controller Partitions:
/dev/mapper/via_hfciifae1: 40 gig winxp NTFS partition /dev/mapper/via_hfciifae3: 10 gig experimental partition /dev/mapper/via_hfciifae5: 50 meg ext2 /boot partition /dev/mapper/via_hfciifae6: 1 gig swap partition /dev/mapper/via_hfciifae7: 22 gig reiserfs root partition
If anyone has any suggestions of further tests I can perform to narrow down the problem, or a better solution for it, you have my full cooperation. If this fix seems acceptable, then I hope it can be merged in the next kernel release.
PS> Please CC me on any replies, as I am not subscribed to this list
Quoted 33 lines · Phillip Susi
I have been debugging a power management problem for a few days now, and I believe I have finally solved the problem. Because it involved patching the kernel, I felt I should share the fix here in hopes that it can be improved and/or integrated into future kernels. Right now I am running 2.6.14.2 on amd64, compiled myself, with the ubuntu breezy amd64 distribution.
First I'll state the fix. It involved changing two lines in
include/linux/libata.h:static inline u8 ata_busy_wait(struct
ata_port*ap, unsigned int bits,
unsigned int max){ u8 status; do { udelay(100); <-- changed to 100 from 10 status = ata_chk_status(ap); max--; } while ((status & bits) && (max > 0)); return status; }and:
static inline u8
ata_wait_idle(struct ata_port *ap){
u8 status = ata_busy_wait(ap,ATA_BUSY|ATA_DRQ,
10000); <-- changed to 10,000 from 1,000
Quoted 9 lines
if (status & (ATA_BUSY | ATA_DRQ)) {
unsigned long l = ap->ioaddr.status_addr;
printk(KERN_WARNING
"ATA: abnormal status 0x%X on port 0x%lX\n",
status, l);
}
return status;
}This change will increase the minimum delay in both ata_wait_idle() and ata_busy_wait() from 10 usec to 100 usec, which is not a good change.
It would be less damaging to increase the delay in ata_wait_idle() from 1000 to 100,000. A one second spin is a bit sad, but the hardware's bust, so that's the least of the user's worries.
The best fix would be to identify the specific _callers_ of these functions
which need their delay increased. You can do that by adding:
if (max == 0) dump_stack();
at the end of ata_busy_wait(). The resulting stack dumps will tell you where the offending callsites are. With luck, it's just one. If we know which one, that might point us at some chipset-level delay which we're forgetting to do, or something like that.
Quoted 16 lines
The problem seems to be that my VIA SATA raid controller requires more time to recover from being suspended. It looks like the code in
sata_via.crestores the task file after a resume, then callsata_wait_idleto wait for the busy bit to clear. The problem was that this function timed out before the busy bit cleared, resulting in messages like this:ATA: abnormal status
0x80on port0xE007Then if there was an IO request made immediately after resuming, it would timeout and fail, because it was issued before the hardware was ready. Changing the timeout resolved this. I tried changing both the udelay and
ata_busy_waitlines to increase the timeout, and it did not seem to matter which I changed, as long as the total timeout was increased by a factor of 100.
Quoted 8 lines
Since increasing the maximum timeout, suspend and hibernate work great for me. While experiencing this bug, it may have exposed another bug, which I will mention now in passing. As I said before, after a resume, if there was an IO request made immediately ( before the busy bit finally did clear ) it would timeout and fail. It seemed the kernel filled the buffer cache for the requested block with garbage rather than retry the read. It seems to me that at some point, the read should have been retried.
Or should have returned an IO error.
Yes, this might be a driver bug. Again, if we know precisely which callsite is experiencing the timeout then we're better situated to fix it.
Quoted 10 lines · Andrew Morton
This change will increase the minimum delay in both
ata_wait_idle()andata_busy_wait()from 10 usec to 100 usec, which is not a good change.It would be less damaging to increase the delay in
ata_wait_idle()from 1000 to 100,000. A one second spin is a bit sad, but the hardware's bust, so that's the least of the user's worries.The best fix would be to identify the specific _callers_ of these functions
which need their delay increased. You can do that by adding:
Quoted 10 lines
if (max == 0) dump_stack();at the end of
ata_busy_wait(). The resulting stack dumps will tell you where the offending callsites are. With luck, it's just one. If we know which one, that might point us at some chipset-level delay which we're forgetting to do, or something like that.
Right, changing ata_wait_idle's maximum to 100,000 would be a better fix as it would not effect the minimum wait time. Thanks for the dump_stack() tip, I put that in and this is the call stack dumped for the two timeouts following a resume ( though without the stack dump, there were 6 timeouts reported, 3 on each channel ):
[ 2.650217] Call Trace:<ffffffff88048782>{:libata:ata_dev_select+102} <ffffffff880488fa>{:libata:ata_qc_issue_prot+31}
[ 2.650244] <ffffffff8804ba48>{:libata:ata_scsi_translate+172} <ffffffff880244bd>{:scsi_mod:scsi_done+0}
[ 2.650269] <ffffffff880244bd>{:scsi_mod:scsi_done+0} <ffffffff8804bd06>{:libata:ata_scsi_queuecmd+219}
[ 2.650291] <ffffffff88024686>{:scsi_mod:scsi_dispatch_cmd+433}
[ 2.650305] <ffffffff8802978e>{:scsi_mod:scsi_request_fn+611} <ffffffff80148ea3>{sync_page+0}
[ 2.650326] <ffffffff80224858>{generic_unplug_device+7} <ffffffff80224851>{generic_unplug_device+0}
[ 2.650334] <ffffffff8807c118>{:dm_mod:dm_table_unplug_all+49} <ffffffff8807a0f1>{:dm_mod:dm_unplug_all+29}
[ 2.650354] <ffffffff8807c118>{:dm_mod:dm_table_unplug_all+49} <ffffffff8807a0f1>{:dm_mod:dm_unplug_all+29}
[ 2.650372] <ffffffff80164a08>{block_sync_page+58} <ffffffff80148ed6>{sync_page+51}
[ 2.650381] <ffffffff8028eb11>{__wait_on_bit_lock+55} <ffffffff801490fd>{__lock_page+94}
[ 2.650392] <ffffffff8013d37f>{wake_bit_function+0} <ffffffff80149374>{find_get_page+12}
[ 2.650402] <ffffffff8014a4de>{do_generic_mapping_read+472} <ffffffff8014947f>{file_read_actor+0}
[ 2.650410] <ffffffff8014af0a>{__generic_file_aio_read+333} <ffffffff8014b0c0>{generic_file_read+187}
[ 2.650433] <ffffffff8013d274>{autoremove_wake_function+0} <ffffffff80163898>{vfs_read+200}
[ 2.650453] <ffffffff8016ba1f>{kernel_read+65} <ffffffff8016d4c7>{do_execve+291}
[ 2.650464] <ffffffff8010cc31>{sys_execve+51} <ffffffff80139bd8>{worker_thread+0}
[ 2.650475] <ffffffff8010ef34>{execve+100} <ffffffff80139bd8>{worker_thread+0}
[ 2.650482] <ffffffff801397bb>{__call_usermodehelper+0} <ffffffff80139abc>{____call_usermodehelper+134}
[ 2.650496] <ffffffff8010eec6>{child_rip+8} <ffffffff80139bd8>{worker_thread+0}
[ 2.650504] <ffffffff801397bb>{__call_usermodehelper+0} <ffffffff80139a36>{____call_usermodehelper+0}
[ 2.650514] <ffffffff8010eebe>{child_rip+0}
[ 2.693624] ATA: abnormal status 0x80 on port 0xE007
[ 2.693626]
[ 2.693627] Call Trace:<ffffffff880488fa>{:libata:ata_qc_issue_prot+31} <ffffffff8804ba48>{:libata:ata_scsi_translate+172}
[ 2.693646] <ffffffff880244bd>{:scsi_mod:scsi_done+0} <ffffffff880244bd>{:scsi_mod:scsi_done+0}
[ 2.693670] <ffffffff8804bd06>{:libata:ata_scsi_queuecmd+219} <ffffffff88024686>{:scsi_mod:scsi_dispatch_cmd+433}
[ 2.693694] <ffffffff8802978e>{:scsi_mod:scsi_request_fn+611} <ffffffff80148ea3>{sync_page+0}
[ 2.693712] <ffffffff80224858>{generic_unplug_device+7} <ffffffff80224851>{generic_unplug_device+0}
[ 2.693720] <ffffffff8807c118>{:dm_mod:dm_table_unplug_all+49} <ffffffff8807a0f1>{:dm_mod:dm_unplug_all+29}
[ 2.693738] <ffffffff8807c118>{:dm_mod:dm_table_unplug_all+49} <ffffffff8807a0f1>{:dm_mod:dm_unplug_all+29}
[ 2.693756] <ffffffff80164a08>{block_sync_page+58} <ffffffff80148ed6>{sync_page+51}
[ 2.693765] <ffffffff8028eb11>{__wait_on_bit_lock+55} <ffffffff801490fd>{__lock_page+94}
[ 2.693775] <ffffffff8013d37f>{wake_bit_function+0} <ffffffff80149374>{find_get_page+12}
[ 2.693785] <ffffffff8014a4de>{do_generic_mapping_read+472} <ffffffff8014947f>{file_read_actor+0}
[ 2.693793] <ffffffff8014af0a>{__generic_file_aio_read+333} <ffffffff8014b0c0>{generic_file_read+187}
[ 2.693816] <ffffffff8013d274>{autoremove_wake_function+0} <ffffffff80163898>{vfs_read+200}
[ 2.693835] <ffffffff8016ba1f>{kernel_read+65} <ffffffff8016d4c7>{do_execve+291}
[ 2.693846] <ffffffff8010cc31>{sys_execve+51} <ffffffff80139bd8>{worker_thread+0}
[ 2.693857] <ffffffff8010ef34>{execve+100} <ffffffff80139bd8>{worker_thread+0}
[ 2.693864] <ffffffff801397bb>{__call_usermodehelper+0} <ffffffff80139abc>{____call_usermodehelper+134}
[ 2.693878] <ffffffff8010eec6>{child_rip+8} <ffffffff80139bd8>{worker_thread+0}
[ 2.693885] <ffffffff801397bb>{__call_usermodehelper+0} <ffffffff80139a36>{____call_usermodehelper+0}
[ 2.693896] <ffffffff8010eebe>{child_rip+0}Maybe a better fix would be to insert a delay after the resume but before the driver even tries to access the hardware. Maybe somewhere in a method in the DSDT?
Quoted 8 lines
Or should have returned an IO error.
Yes, this might be a driver bug. Again, if we know precisely which callsite is experiencing the timeout then we're better situated to fix it.
I made the timeouts happen again by returning the timeouts to the way they were before, and issuing the command:
echo mem > /sys/power/state ; df
After the abnormal status messages, around 30 seconds went by, then this:
[ 14.596710] ata1: command 0x25 timeout, stat 0x50 host_stat 0x1
[ 14.596715]
[ 14.596716] Call Trace:<ffffffff880498e2>{:libata:ata_eng_timeout+496} <ffffffff88027690>{:scsi_mod:scsi_error_handler+0}
[ 14.596752] <ffffffff8804af99>{:libata:ata_scsi_error+21} <ffffffff88027706>{:scsi_mod:scsi_error_handler+118}
[ 14.596775] <ffffffff88027690>{:scsi_mod:scsi_error_handler+0} <ffffffff88027690>{:scsi_mod:scsi_error_handler+0}
[ 14.596803] <ffffffff8013d0d8>{kthread+191} <ffffffff80139bd8>{worker_thread+0}
[ 14.596814] <ffffffff8010eec6>{child_rip+8} <ffffffff80139bd8>{worker_thread+0}
[ 14.596822] <ffffffff8013cf04>{keventd_create_kthread+0} <ffffffff8013d019>{kthread+0}
[ 14.596833] <ffffffff8010eebe>{child_rip+0}
[ 14.597403] ReiserFS: warning: is_tree_node: node level 2295 does not match to the expected one 1
[ 14.597408] ReiserFS: dm-5: warning: vs-5150: search_by_key: invalid format found in block 32828. Fsck?
[ 14.597414] ReiserFS: dm-5: warning: vs-13070: reiserfs_read_locked_inode: i/o failure occurred trying to find stat data of [2 1455 0x0 SD]Bash reports command not found when trying to run df even now as I write this.
I also noticed that during the 30 seconds or so that the system was trying to load df before libata timed out the request, the cpu was pegged at 100% utilization.
Quoted 7 lines · Phillip Susi
Quoted 5 lines · Andrew Morton
This change will increase the minimum delay in both
ata_wait_idle()andata_busy_wait()from 10 usec to 100 usec, which is not a good change.It would be less damaging to increase the delay in
ata_wait_idle()from 1000 to 100,000. A one second spin is a bit sad, but the hardware's bust,
I wonder if this the same problem that prevents resume-from-ram from working on my system when I use an older hard drive, rather than the newer model that came installed (notebook)..
Whenever resume fails, the hard drive light is on solid and the system is unresponsive. And the backlight is off so no debug info available (no serial ports, either).
Could be.. could be.
Cheers
Hi!
Quoted 7 lines
Quoted 6 lines
Quoted 6 lines
This change will increase the minimum delay in both
ata_wait_idle()andata_busy_wait()from 10 usec to 100 usec, which is not a good change.It would be less damaging to increase the delay in
ata_wait_idle()from 1000 to 100,000. A one second spin is a bit sad, but the hardware's bust,
Quoted 7 lines
I wonder if this the same problem that prevents resume-from-ram from working on my system when I use an older hard drive, rather than the newer model that came installed (notebook)..
Whenever resume fails, the hard drive light is on solid and the system is unresponsive. And the backlight is off so no debug info available (no serial ports, either).
Debugging this sucks, sorry. Usefull methods are keyboard leds and system beeper. printk-over-morse comes to mind :-).
Signature · 1 line
64 bytes from 195.113.31.123: icmp_seq=28 ttl=51 time=448769.1 ms