Use all available CPUs to collect dump

By default, use all available CPUs to collect the dump instead of just
one CPU. This reduces the dump collection time significantly, specially
for the systems with very large memory configuration.

The graph below show the time reduction in dump collection on large
memory system from 4 Hours to 10 Minutes.

                               |
                       14000   | *
                               |
                               |
                               |
                               |
                               |
                               |
                        1600   |      *
                               |
                               |
                        1400   |
 execution Time (Sec)          |
                               |
                        1200   |
                               |
                               |
                        1000   |
                               |            *
                               |
                        800    |
                               |
                               |                   *
                        600    |
                               |                          *     *
                               |
                                ------------------------------------
                                   1    9    17    25     33    41

                                          Number of Threads

System details:
Architecture: PowerPC
/proc/vmcore size: 3.7T
Filter level: 1

Similar tests with different filter levels, 31 and 16, also show significant
reduction in time consumption for dump collection.

Although the above tests were performed only on the PowerPC architecture, but
I think using all CPUs for dump collection will have performance benefits on
other architectures too.

Note: The `--num-threads` command-line argument is added to the `makedumpfile`
core collector only when more than one processing unit is available in the
system. This helps to keep the default case (nr_cpus=1) the same.

Signed-off-by: Sourabh Jain <sourabhjain@linux.ibm.com>

Author: sourabhjains

dracut/99kdumpbase/kdump.sh

@@ -147,6 +147,10 @@ dump_fs() {
     case $CORE_COLLECTOR in
         *makedumpfile*)
             CORE_COLLECTOR=$(echo "$CORE_COLLECTOR" | sed -e "s/-F//g")
+            THREADS=$(nproc)
+            if [ "$THREADS" -gt 1 ]; then
+                CORE_COLLECTOR="$CORE_COLLECTOR --num-threads=$THREADS"
+            fi
             ;;
     esac
 
@@ -384,6 +388,13 @@ dump_raw() {
         /kdumpscripts/monitor_dd_progress.sh $_src_size_mb &
     fi
 
+    if echo "$CORE_COLLECTOR" | grep -q makedumpfile; then
+        THREADS=$(nproc)
+        if [ "$THREADS" -gt 1 ]; then
+            CORE_COLLECTOR="$CORE_COLLECTOR --num-threads=$THREADS"
+        fi
+    fi
+
     dinfo "saving vmcore"
     $CORE_COLLECTOR /proc/vmcore | dd of="$1" bs=$DD_BLKSIZE >> /tmp/dd_progress_file 2>&1 || return 1
     sync

dracut/99kdumpbase/module-setup.sh

@@ -1095,6 +1095,7 @@ install() {
     inst "/usr/bin/printf" "/sbin/printf"
     inst "/usr/bin/logger" "/sbin/logger"
     inst "/usr/bin/chmod" "/sbin/chmod"
+    inst "/usr/bin/nproc" "/sbin/nproc"
     inst "/usr/bin/dirname" "/sbin/dirname"
     inst "/lib/kdump/kdump-lib-initramfs.sh" "/lib/kdump-lib-initramfs.sh"
     inst "/lib/kdump/kdump-logger.sh" "/lib/kdump-logger.sh"