Add the --discard option for Create which will send BLKDISCARD requests to
all disks before assembling the array. This is a fast way to know the
current state of all the disks. If the discard request zeros the data
on the disks (as is common but not universal) then the array is in a
state with correct parity. Thus the initial sync may be skipped.
After issuing each discard request, check if the first page of the
block device is zero. If it is, it is safe to assume the entire
disk should be zero. If it's not report an error.
If all the discard requests are successful and there are no missing
disks thin it is safe to set assume_clean as we know the array is clean.
Signed-off-by: Logan Gunthorpe <logang@xxxxxxxxxxxx>
---
Create.c | 75 ++++++++++++++++++++++++++++++++++++++++++++++++++++++++
ReadMe.c | 1 +
mdadm.c | 4 +++
mdadm.h | 2 ++
4 files changed, 82 insertions(+)
diff --git a/Create.c b/Create.c
index e06ec2ae96a1..db99da1e8571 100644
--- a/Create.c
+++ b/Create.c
@@ -26,6 +26,12 @@
#include "md_u.h"
#include "md_p.h"
#include <ctype.h>
+#include <sys/ioctl.h>
+
+#ifndef BLKDISCARD
+#define BLKDISCARD _IO(0x12,119)
+#endif
+#include <fcntl.h>
static int round_size_and_verify(unsigned long long *size, int chunk)
{
@@ -91,6 +97,63 @@ int default_layout(struct supertype *st, int level, int verbose)
return layout;
}
+static int discard_device(const char *devname, unsigned long long size)
+{
+ uint64_t range[2] = {0, size};
+ unsigned long buf[4096 / sizeof(unsigned long)];
+ unsigned long i;
+ int fd;
+
+ fd = open(devname, O_RDWR | O_EXCL);
+ if (fd < 0) {
+ pr_err("could not open device for discard: %s\n", devname);
+ return 1;
+ }
+
+ if (ioctl(fd, BLKDISCARD, &range)) {
+ pr_err("discard failed on '%s': %m\n", devname);
+ goto out_err;
+ }
+
+ if (read(fd, buf, sizeof(buf)) != sizeof(buf)) {
+ pr_err("failed to readback '%s' after discard: %m\n", devname);
+ goto out_err;
+ }
+
+ for (i = 0; i < ARRAY_SIZE(buf); i++) {
+ if (buf[i]) {
+ pr_err("device did not read back zeros after discard on '%s': %lx\n",
+ devname, buf[i]);
+ goto out_err;
+ }
+ }
+
+ close(fd);
+ return 0;
+
+out_err:
+ close(fd);
+ return 1;
+}
+
+static int discard_devices(struct context *c, struct mddev_dev *devlist,
+ unsigned long long size)
+{
+ struct mddev_dev *dv;
+
+ for (dv = devlist; dv; dv = dv->next) {
+ if (!strcmp(dv->devname, "missing"))
+ continue;
+
+ if (c->verbose)
+ pr_err("discarding all data on: %s\n", dv->devname);
+ if (discard_device(dv->devname, size))
+ return 1;
+ }
+
+ return 0;
+}
+
int Create(struct supertype *st, char *mddev,
char *name, int *uuid,
int subdevs, struct mddev_dev *devlist,
@@ -603,6 +666,18 @@ int Create(struct supertype *st, char *mddev,
}
}
+ if (s->discard) {
+ if (discard_devices(c, devlist, (s->size << 10) +
+ (st->data_offset << 9)))
+ return 1;
+
+ /* All disks are zero so if there are none missing assume
+ * the array is clean
+ */
+ if (first_missing >= s->raiddisks)
+ s->assume_clean = 1;
+ }
+
/* If this is raid4/5, we want to configure the last active slot
* as missing, so that a reconstruct happens (faster than re-parity)
* FIX: Can we do this for raid6 as well?
diff --git a/ReadMe.c b/ReadMe.c
index 7f94847e9769..544a057f83a0 100644
--- a/ReadMe.c
+++ b/ReadMe.c
@@ -138,6 +138,7 @@ struct option long_options[] = {
{"size", 1, 0, 'z'},
{"auto", 1, 0, Auto}, /* also for --assemble */
{"assume-clean",0,0, AssumeClean },
+ {"discard", 0, 0, Discard },
{"metadata", 1, 0, 'e'}, /* superblock format */
{"bitmap", 1, 0, Bitmap},
{"bitmap-chunk", 1, 0, BitmapChunk},
diff --git a/mdadm.c b/mdadm.c
index 972adb524dfb..8dee85a54a6a 100644
--- a/mdadm.c
+++ b/mdadm.c
@@ -602,6 +602,10 @@ int main(int argc, char *argv[])
s.assume_clean = 1;
continue;
+ case O(CREATE, Discard):
+ s.discard = 1;
+ continue;
+
case O(GROW,'n'):
case O(CREATE,'n'):
case O(BUILD,'n'): /* number of raid disks */
diff --git a/mdadm.h b/mdadm.h
index 941a5f3823a0..99769be57ac5 100644
--- a/mdadm.h
+++ b/mdadm.h
@@ -433,6 +433,7 @@ extern char Version[], Usage[], Help[], OptionHelp[],
*/
enum special_options {
AssumeClean = 300,
+ Discard,
BitmapChunk,
WriteBehind,
ReAdd,
@@ -593,6 +594,7 @@ struct shape {
int bitmap_chunk;
char *bitmap_file;
int assume_clean;
+ int discard;
int write_behind;
unsigned long long size;
unsigned long long data_offset;
--
2.30.2
