Possible explanation for mptsas ATA pass-through hangs
Subsystems (2)
I may have an explanation for the LSI 1068 HBA hangs provoked by ATA pass-through commands, in particular by smartctl.
First, my version of the symptoms. On an LSI SAS1068E HBA with SATA disks and with smartd running, I'm seeing occasional task, bus, and host resets, some of which lead to hard faults of the HBA requiring a reboot. Abusively looping the smartctl command,
# while true; do smartctl -a /dev/sdb > /dev/null; done
dramatically increases the frequency of these failures to nearly one per minute. A high IO load through the HBA while looping smartctl seems to improve the chance of a full scsi host reset or a non-recoverable hang.
I reduced what smartctl was doing down to a simple test case which causes the hang with a single IO when pointed at the sd interface. See the code at the bottom of this e-mail. It uses an SG_IO ioctl to issue a single pass-through ATA identify device command. If the buffer userspace gives for the read data has certain alignments straddling a page boundary, the task is issued to the HBA but the HBA fails to respond. If run against the sg interface, neither the test code nor smartctl causes a hang.
sd and sg handle the SG_IO ioctl slightly differently. Unless you specifically set a flag to do direct IO, sg passes a buffer of its own, which is page-aligned, to the block layer and later copies the result into the userspace buffer regardless of its alignment. sd, on the other hand, always does direct IO unless the userspace buffer fails an alignment test at block/blk-map.c line 57, in which case a page-aligned buffer is created for the transfer.
The alignment test currently checks for word-alignment, the default setup by scsi_lib.c; therefore, userspace buffers of almost any alignment are given directly to the HBA as DMA targets. The hardware doesn't seem to like at least a couple of the alignments which cross a page boundary (see the test code below). Curiously, many page-boundary-crossing alignments do work just fine.
So, either the hardware has an bug handling certain alignments or the hardware has a stricter alignment requirement than the driver is advertising. If stricter alignment is required, then in no case should misaligned buffers from userspace be allowed through without being bounced or at least causing an error to be returned.
It seems the mptsas driver or its friends could use blk_queue_dma_alignment() to advertise a stricter alignment requirement. If so, sd does the right thing and bounces misaligned buffers (see block/blk-map.c line 57). I gave the following patch to Linus's tree from last night a quick try and it seemed to work. I'm sure this is the wrong place for this call, but it gets the idea across.
drivers/message/fusion/mptscsih.c +2 -0
@@ -2450,6 +2450,8 @@ mptscsih_slave_configure(struct scsi_device *sdev) ioc->name,sdev->tagged_supported, sdev->simple_tags, sdev->ordered_tags)); +blk_queue_dma_alignment (sdev->request_queue,512 -1 ); +return 0 ; }
I look forward to hearing from you guys who know this hardware and code better. Is the hardware at fault, or should the driver be shielding the hardware better?
Does this `fix' the problem for anyone besides me?
Regards,
-- Ryan
Here is a minimal bit of test code which causes the error. BEWARE: this will hose the HBA at which you point it, of course. If that's controlling your root disk...
/*
* sg_bomb -- send SG_IO ioctl which causes HBA to hang *
* usage: sg_bomb <device>
* e.g.: sg_bomb /dev/sdb
* e.g.: sg_bomb /dev/sg1
*
* Modify offset_into_page to adjust the degree of buffer misalignment.
*/
#include <unistd.h>
#include <scsi/sg.h>
#include <sys/ioctl.h>
#include <fcntl.h>
#include <stdlib.h>
int main(int argc, char* argv[])
{
char* filename = argv[1];
unsigned int offset_into_page = 0xe40;
// works: unsigned int offset_into_page = 0x0;
// hangs: unsigned int offset_into_page = 0xf00;
// works: unsigned int offset_into_page = 0xf04;
unsigned char ata_identify_cmd[] = {0x85, 0x08, 0x0e, 0, 0, 0, 0x01,
0, 0, 0, 0, 0, 0, 0, 0xec, 0};
unsigned char sense[32];
unsigned char* data = valloc(0x2000) + offset_into_page;
struct sg_io_hdr hdr = {
.interface_id = 'S',
.dxfer_direction = SG_DXFER_FROM_DEV,
.cmdp = ata_identify_cmd,
.cmd_len = 16,
.dxferp = data,
.dxfer_len = 512,
.sbp = sense,
.mx_sb_len = sizeof(sense),
.timeout = 5000,
};
int fd;
if ((fd = open(filename, O_RDWR|O_NONBLOCK)) < 0)
perror();
return ioctl(fd, SG_IO, &hdr);
}Ryan,
Sorry for delay in response. I am able to see the behavior you have mentioned using your sample program. I am also learning your proposed patch too. I will update my thoughts on your findings ASAP.
Thanks,
Kashyap
Quoted 145 lines
-----Original Message-----
From: Ryan Kuester [mailto:[email protected]]
Sent: Saturday, April 10, 2010 3:06 PM
To: Desai, Kashyap; [email protected]
Subject: Possible explanation for mptsas ATA pass-through hangsI may have an explanation for the LSI 1068 HBA hangs provoked by ATA pass-through commands, in particular by smartctl.
First, my version of the symptoms. On an LSI SAS1068E HBA with SATA disks and with smartd running, I'm seeing occasional task, bus, and host
resets, some of which lead to hard faults of the HBA requiring a reboot.
Abusively looping the smartctl command,# while true; do smartctl -a /dev/sdb > /dev/null; donedramatically increases the frequency of these failures to nearly one per
minute. A high IO load through the HBA while looping smartctl seems to improve the chance of a full scsi host reset or a non-recoverable hang.I reduced what smartctl was doing down to a simple test case which causes
the hang with a single IO when pointed at the sd interface. See the code at the bottom of this e-mail. It uses anSG_IOioctl to issue a single pass-through ATA identify device command. If the buffer userspace
gives for the read data has certain alignments straddling a page boundary,
the task is issued to the HBA but the HBA fails to respond. If run against
the sg interface, neither the test code nor smartctl causes a hang.sd and sg handle the
SG_IOioctl slightly differently. Unless you specifically set a flag to do direct IO, sg passes a buffer of its own, which is page-aligned, to the block layer and later copies the result into the userspace buffer regardless of its alignment. sd, on the other
hand, always does direct IO unless the userspace buffer fails an alignment test atblock/blk-map.cline 57, in which case a page-aligned buffer is created for the transfer.The alignment test currently checks for word-alignment, the default setup by
scsi_lib.c; therefore, userspace buffers of almost any alignment are given directly to the HBA as DMA targets. The hardware doesn't seem to like at least a couple of the alignments which cross a page boundary (see the test code below). Curiously, many page-boundary-crossing alignments do work just fine.So, either the hardware has an bug handling certain alignments or the hardware has a stricter alignment requirement than the driver is advertising. If stricter alignment is required, then in no case should misaligned buffers from userspace be allowed through without being bounced or at least causing an error to be returned.
It seems the mptsas driver or its friends could use
blk_queue_dma_alignment()to advertise a stricter alignment requirement.
If so, sd does the right thing and bounces misaligned buffers (seeblock/blk-map.cline 57). I gave the following patch to Linus's tree from last night a quick try and it seemed to work. I'm sure this is the
wrong place for this call, but it gets the idea across.drivers/message/fusion/mptscsih.c +2
-0@@ -2450,6 +2450,8 @@ mptscsih_slave_configure(struct scsi_device *sdev) ioc->name,sdev->tagged_supported, sdev->simple_tags, sdev->ordered_tags)); +blk_queue_dma_alignment (sdev->request_queue,512 -1 ); +return 0 ; }I look forward to hearing from you guys who know this hardware and code better. Is the hardware at fault, or should the driver be shielding the
hardware better?Does this `fix' the problem for anyone besides me?
Regards,
-- RyanHere is a minimal bit of test code which causes the error. BEWARE:
this
will hose the HBA at which you point it, of course. If that's controlling your root disk.../*
*sg_bomb-- sendSG_IOioctl which causes HBA to hang *
* usage:sg_bomb<device>
* e.g.:sg_bomb/dev/sdb
* e.g.:sg_bomb/dev/sg1
*
* Modifyoffset_into_pageto adjust the degree of buffer
misalignment.
*/#include <
unistd.h>
#include <scsi/sg.h>
#include <sys/ioctl.h>
#include <fcntl.h>
#include <stdlib.h>int
main(int argc, char* argv[]){ char* filename = argv[1]; unsigned int offset_into_page = 0xe40; // works: unsigned int offset_into_page = 0x0; // hangs: unsigned int offset_into_page = 0xf00; // works: unsigned int offset_into_page = 0xf04; unsigned char ata_identify_cmd[] = {0x85, 0x08, 0x0e, 0, 0, 0, 0x01, 0, 0, 0, 0, 0, 0, 0, 0xec, 0}; unsigned char sense[32]; unsigned char* data = valloc(0x2000) + offset_into_page; struct sg_io_hdr hdr = { .interface_id = 'S', .dxfer_direction = SG_DXFER_FROM_DEV, .cmdp = ata_identify_cmd, .cmd_len = 16, .dxferp = data, .dxfer_len = 512, .sbp = sense, .mx_sb_len = sizeof(sense), .timeout = 5000, }; int fd; if ((fd = open(filename, O_RDWR|O_NONBLOCK)) < 0) perror(); return ioctl(fd, SG_IO, &hdr); }