Folks,
I have made a patch (for 2.4) which adds fast failover capabilities to
the RAID 1 system (only tested with SCSI targets). Without this patch,
failover can take as much as 3 minutes during failure. With this patch,
failover can reach subsecond ranges (with appropriate configuration).
The patch does the following:
adds kernel tuneable for bus reset settle time
adds kernel tuneable for host reset settle time
adds kernel tuneable for timeout on normal I/O requests
adds kernel tuneable for timeout on fastfail I/O requests
adds fastfail flag to RAID/SCSI code
I/O requests from a RAID 1 device operating in redundant operation will
request FASTFAILs. I/O requests from a RAID 1 device operating in
degraded mode will request normal I/Os.
adds ability to not unjam the SCSI host on fastfail I/O requests. This
is useful for FibreChannel devices, where an unjam may be unwarranted
form of error recovery because it introduces LIP resets which may slow
performance when a failure has already happened and can't be recovered.
Anyone interested in a 2.6 patch of the same flavor (the fastfail flag
is already in 2.6, but needs a little work on timing tuning)? Any other
comments?
Thanks
-steve
--- linux/drivers/md/raid1.c 2003-10-08 15:59:39.000000000 -0700
+++ linux-mdfix/drivers/md/raid1.c 2003-10-08 16:07:28.000000000 -0700
@@ -563,6 +563,7 @@
struct raid1_bh * r1_bh;
int disks = MD_SB_DISKS;
int i, sum_bhs = 0;
+ int operational_mirrors;
struct mirror_info *mirror;
if (!buffer_locked(bh))
@@ -592,6 +593,13 @@
set_bit(R1BH_SyncPhase, &r1_bh->state);
}
spin_unlock_irq(&conf->segment_lock);
+
+ for (operational_mirrors = 0, i = 0; i < disks; i++) {
+ if (conf->mirrors[i].operational) {
+ operational_mirrors++;
+ }
+ }
+
/*
* i think the read and write branch should be separated completely,
@@ -614,6 +622,7 @@
bh_req->b_blocknr = bh->b_rsector;
bh_req->b_dev = mirror->dev;
bh_req->b_rdev = mirror->dev;
+ bh_req->b_state |= operational_mirrors > 1 ? (1<<BH_FastFail) : 0;
/* bh_req->b_rsector = bh->n_rsector; */
bh_req->b_end_io = raid1_end_request;
bh_req->b_private = r1_bh;
@@ -661,6 +670,7 @@
mbh->b_rsector = bh->b_rsector;
mbh->b_state = (1<<BH_Req) | (1<<BH_Dirty) |
(1<<BH_Mapped) | (1<<BH_Lock);
+ mbh->b_state |= operational_mirrors > 1 ? (1<<BH_FastFail) : 0;
atomic_set(&mbh->b_count, 1);
mbh->b_size = bh->b_size;
--- linux/drivers/scsi/Config.in 2003-10-08 15:59:39.000000000 -0700
+++ linux-mdfix/drivers/scsi/Config.in 2003-10-08 15:27:15.000000000 -0700
@@ -28,6 +28,13 @@
bool ' Verbose SCSI error reporting (kernel size +=12K)' CONFIG_SCSI_CONSTANTS
bool ' SCSI logging facility' CONFIG_SCSI_LOGGING
+bool ' Execute error recovery on I/Os marked FastFail' CONFIG_SCSI_ERROR_RECOVERY_ON_FASTFAIL
+int ' SCSI disk retries for normal requests' CONFIG_SCSI_SD_MAX_RETRIES 5
+int ' SCSI disk I/Os command timeout in jiffies for normal requests' CONFIG_SCSI_SD_TIMEOUT 3000
+int ' SCSI disk I/Os command timeout in jiffies for FastFail requests' CONFIG_SCSI_SD_TIMEOUT_FASTFAIL 25
+int ' SCSI bus reset settle time during error recovery in jiffies' CONFIG_SCSI_BUS_RESET_SETTLE_TIME 500
+int ' SCSI host reset settle time during error recovery in jiffies' CONFIG_SCSI_HOST_RESET_SETTLE_TIME 1000
+
mainmenu_option next_comment
comment 'SCSI low-level drivers'
--- linux/drivers/scsi/scsi_error.c 2003-10-08 15:59:41.000000000 -0700
+++ linux-mdfix/drivers/scsi/scsi_error.c 2003-10-08 15:20:10.000000000 -0700
@@ -66,9 +66,8 @@
* These should *probably* be handled by the host itself.
* Since it is allowed to sleep, it probably should.
*/
-#define BUS_RESET_SETTLE_TIME 5*HZ
-#define HOST_RESET_SETTLE_TIME 10*HZ
-
+#define BUS_RESET_SETTLE_TIME CONFIG_SCSI_BUS_RESET_SETTLE_TIME
+#define HOST_RESET_SETTLE_TIME CONFIG_SCSI_HOST_RESET_SETTLE_TIME
static const char RCSid[] = "$Header: /cvsdev/hhl-kernel-campbell/linux/drivers/scsi/scsi_error.c,v 1.2 2002/04/01 00:59:20 jpuhlman Exp $";
@@ -236,9 +235,23 @@
panic("Error handler thread not present at %p %p %s %d",
SCpnt, SCpnt->host, __FILE__, __LINE__);
}
+
+/*
+ * It is desireable on FibreChannel not to execute
+ * error recovery (unjam_host) on I/O failures because
+ * of the time expensive LIP Reset.
+ */
+#ifdef CONFIG_SCSI_ERROR_RECOVERY_ON_FASTFAIL
if (SCpnt->host->host_busy == SCpnt->host->host_failed) {
up(SCpnt->host->eh_wait);
}
+#else /* CONFIG_SCSI_ERROR_RECOVERY_ON_FASTFAIL */
+ if ((SCpnt->request.bh->b_state & (1<<BH_FastFail)) == 0) {
+ if (SCpnt->host->host_busy == SCpnt->host->host_failed) {
+ up(SCpnt->host->eh_wait);
+ }
+ }
+#endif /* CONFIG_SCSI_ERROR_RECOVERY_ON_FASTFAIL */
}
/*
--- linux/drivers/scsi/sd.c 2003-10-08 15:59:42.000000000 -0700
+++ linux-mdfix/drivers/scsi/sd.c 2003-10-08 15:26:57.000000000 -0700
@@ -76,13 +76,14 @@
#define N_USED_SCSI_DISKS (sd_template.dev_max + SCSI_DISKS_PER_MAJOR - 1)
#define N_USED_SD_MAJORS (N_USED_SCSI_DISKS / SCSI_DISKS_PER_MAJOR)
-#define MAX_RETRIES 5
+#define MAX_RETRIES CONFIG_SCSI_SD_MAX_RETRIES
/*
- * Time out in seconds for disks and Magneto-opticals (which are slower).
+ * Time out in jiffies for disks and Magneto-opticals (which are slower).
*/
-#define SD_TIMEOUT (30 * HZ)
+#define SD_TIMEOUT_FASTFAIL CONFIG_SCSI_SD_TIMEOUT_FASTFAIL
+#define SD_TIMEOUT CONFIG_SCSI_SD_TIMEOUT
#define SD_MOD_TIMEOUT (75 * HZ)
struct hd_struct *sd;
@@ -416,9 +417,14 @@
SCpnt->transfersize = dpnt->device->sector_size;
SCpnt->underflow = this_count << 9;
- SCpnt->allowed = MAX_RETRIES;
- SCpnt->timeout_per_command = (SCpnt->device->type == TYPE_DISK ?
- SD_TIMEOUT : SD_MOD_TIMEOUT);
+ if (SCpnt->request.bh->b_state & (1<<BH_FastFail)) {
+ SCpnt->allowed = 0;
+ SCpnt->timeout_per_command = SD_TIMEOUT_FASTFAIL;
+ } else {
+ SCpnt->allowed = MAX_RETRIES;
+ SCpnt->timeout_per_command = (SCpnt->device->type == TYPE_DISK ?
+ SD_TIMEOUT : SD_MOD_TIMEOUT);
+ }
/*
* This is the completion routine we use. This is matched in terms
--- linux/include/linux/fs.h 2003-10-08 15:59:43.000000000 -0700
+++ linux-mdfix/include/linux/fs.h 2003-10-08 10:44:08.000000000 -0700
@@ -236,6 +236,7 @@
BH_PrivateStart,/* not a state bit, but the first bit available
* for private allocation by other entities
*/
+ BH_FastFail,
};
/*
--- linux/Documentation/Configure.help 2003-08-08 15:56:10.000000000 -0700
+++ linux-mdfix/Documentation/Configure.help 2003-10-08 18:07:29.000000000 -0700
@@ -2446,6 +2446,55 @@
CONFIG_MIPS_GT96100ETH
Say Y here to support the Ethernet subsystem on your GT96100 card.
+Execute error recovery on I/O's marked FastFail
+CONFIG_SCSI_ERROR_RECOVERY_ON_FASTFAIL
+ The MD layer has been enhanced to support a FASTFAIL flag. The
+ fastfail flag is a hint from the MD layer to the SCSI layer not
+ to retry I/O requests and shorten I/O timeouts when an array
+ is operating in a redundant mode.
+
+ The default behavior avoids error recovery on FASTFAIL I/Os
+ because the error recovery time would introduce I/O latency
+ when a redundant I/O path is available.
+
+ Configuring this option to execute error recovery on FASTFAIL
+ I/O failure by resetting the device, the bus, and the host adaptor.
+
+ I/Os that are not marked FASTFAIL will use the standard error
+ recovery mechanisms.
+
+SCSI disk retries for normal requests
+CONFIG_SCSI_SD_MAX_RETRIES
+ Normal requests will retry I/O's when an I/O fails in an adaptor.
+ This value sets the number of retries tried before giving up and
+ returning an error from the SCSI layer to the higher level I/O
+ layer.
+
+SCSI disk I/Os command timeout in jiffies for normal requests
+CONFIG_SCSI_SD_TIMEOUT
+ Normal requests will time out requests according to this value.
+ Once the request is timed out, it will be retried according
+ to the retry count.
+
+SCSI disk I/Os command timeout in jiffies for FastFail requests
+CONFIG_SCSI_SD_TIMEOUT_FASTFAIL
+ FASTFAIL requests are those requests issued by the MD layer
+ from a redundant array. FASTFAIL requests will time out at
+ a rate configured by this value. Also, when a FASTFAIL request
+ is issued, no retries will be attempted.
+
+SCSI bus reset settle time during error recovery in jiffies
+CONFIG_SCSI_BUS_RESET_SETTLE_TIME
+ After a SCSI bus reset during SCSI error handling, the bus
+ is unavailable for the settle time specified by this value. This
+ provides time for the bus to enter a known good state.
+
+SCSI host reset settle time during error recovery in jiffies
+CONFIG_SCSI_HOST_RESET_SETTLE_TIME
+ After a SCSI host reset during SCSI error handling, the adaptor
+ is unavailable for the settle time specified by this value. This
+ provides time for the adaptor to enter a known good state.
+
Zalon SCSI support
CONFIG_SCSI_ZALON
The Zalon is an interface chip that sits between the PA-RISC
