835cc0c847
XFS and UBIFS can pass long lists to list_sort(); this alternative implementation scales better, reaching ~3x performance gain when list length exceeds the L2 cache size. Stand-alone program timings were run on a Core 2 duo L1=32KB L2=4MB, gcc-4.4, with flags extracted from an Ubuntu kernel build. Object size is 581 bytes compared to 455 for Mark J. Roberts' code. Worst case for either implementation is a list length just over a power of two, and to roughly the same degree, so here are timing results for a range of 2^N+1 lengths. List elements were 16 bytes each including malloc overhead; initial order was random. time (msec) Tatham-Roberts | generic-Mullis-v2 loop_count length | | ratio 4000000 2 206 294 1.427 2000000 3 176 227 1.289 1000000 5 199 172 0.864 500000 9 235 178 0.757 250000 17 243 182 0.748 125000 33 261 196 0.750 62500 65 277 209 0.754 31250 129 292 219 0.75 15625 257 317 235 0.741 7812 513 340 252 0.741 3906 1025 362 267 0.737 1953 2049 388 283 0.729 ~ L1 size 976 4097 556 323 0.580 488 8193 678 361 0.532 244 16385 773 395 0.510 122 32769 844 418 0.495 61 65537 917 454 0.495 30 131073 1128 543 0.481 15 262145 2355 869 0.369 ~ L2 size 7 524289 5597 1714 0.306 3 1048577 6218 2022 0.325 Mark's code does not actually implement the usual or generic mergesort, but rather a variant from Simon Tatham described here: http://www.chiark.greenend.org.uk/~sgtatham/algorithms/listsort.html Simon's algorithm performs O(log N) passes over the entire input list, doing merges of sublists that double in size on each pass. The generic algorithm instead merges pairs of equal length lists as early as possible, in recursive order. For either algorithm, the elements that extend the list beyond power-of-two length are a special case, handled as nearly as possible as a "rounding-up" to a full POT. Some intuition for the locality of reference implications of merge order may be gotten by watching this animation: http://www.sorting-algorithms.com/merge-sort Simon's algorithm requires only O(1) extra space rather than the generic algorithm's O(log N), but in my non-recursive implementation the actual O(log N) data is merely a vector of ~20 pointers, which I've put on the stack. Long-running list_sort() calls: If the list passed in may be long, or the client's cmp() callback function is slow, the client's cmp() may periodically invoke cond_resched() to voluntarily yield the CPU. All inner loops of list_sort() call back to cmp(). Stability of the sort: distinct elements that compare equal emerge from the sort in the same order as with Mark's code, for simple test cases. A boot-time test is provided to verify this and other correctness requirements. A kernel that uses drm.ko appears to run normally with this change; I have no suitable hardware to similarly test the use by UBIFS. [akpm@linux-foundation.org: style tweaks, fix comment, make list_sort_test __init] Signed-off-by: Don Mullis <don.mullis@gmail.com> Cc: Dave Airlie <airlied@redhat.com> Cc: Andi Kleen <andi@firstfloor.org> Cc: Dave Chinner <david@fromorbit.com> Cc: Artem Bityutskiy <dedekind@infradead.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org> |
||
---|---|---|
.. | ||
lzo | ||
reed_solomon | ||
zlib_deflate | ||
zlib_inflate | ||
.gitignore | ||
argv_split.c | ||
atomic64.c | ||
audit.c | ||
bcd.c | ||
bitmap.c | ||
bitrev.c | ||
bug.c | ||
bust_spinlocks.c | ||
check_signature.c | ||
checksum.c | ||
cmdline.c | ||
cpumask.c | ||
crc7.c | ||
crc16.c | ||
crc32.c | ||
crc32defs.h | ||
crc-ccitt.c | ||
crc-itu-t.c | ||
crc-t10dif.c | ||
ctype.c | ||
debug_locks.c | ||
debugobjects.c | ||
dec_and_lock.c | ||
decompress_bunzip2.c | ||
decompress_inflate.c | ||
decompress_unlzma.c | ||
decompress_unlzo.c | ||
decompress.c | ||
devres.c | ||
div64.c | ||
dma-debug.c | ||
dump_stack.c | ||
dynamic_debug.c | ||
extable.c | ||
fault-inject.c | ||
find_last_bit.c | ||
find_next_bit.c | ||
flex_array.c | ||
gcd.c | ||
gen_crc32table.c | ||
genalloc.c | ||
halfmd4.c | ||
hexdump.c | ||
hweight.c | ||
idr.c | ||
inflate.c | ||
int_sqrt.c | ||
iomap_copy.c | ||
iomap.c | ||
iommu-helper.c | ||
ioremap.c | ||
irq_regs.c | ||
is_single_threaded.c | ||
kasprintf.c | ||
Kconfig | ||
Kconfig.debug | ||
Kconfig.kgdb | ||
Kconfig.kmemcheck | ||
kernel_lock.c | ||
klist.c | ||
kobject_uevent.c | ||
kobject.c | ||
kref.c | ||
libcrc32c.c | ||
list_debug.c | ||
list_sort.c | ||
lmb.c | ||
locking-selftest-hardirq.h | ||
locking-selftest-mutex.h | ||
locking-selftest-rlock-hardirq.h | ||
locking-selftest-rlock-softirq.h | ||
locking-selftest-rlock.h | ||
locking-selftest-rsem.h | ||
locking-selftest-softirq.h | ||
locking-selftest-spin-hardirq.h | ||
locking-selftest-spin-softirq.h | ||
locking-selftest-spin.h | ||
locking-selftest-wlock-hardirq.h | ||
locking-selftest-wlock-softirq.h | ||
locking-selftest-wlock.h | ||
locking-selftest-wsem.h | ||
locking-selftest.c | ||
lru_cache.c | ||
Makefile | ||
nlattr.c | ||
parser.c | ||
percpu_counter.c | ||
plist.c | ||
prio_heap.c | ||
prio_tree.c | ||
proportions.c | ||
radix-tree.c | ||
random32.c | ||
ratelimit.c | ||
rational.c | ||
rbtree.c | ||
reciprocal_div.c | ||
rwsem-spinlock.c | ||
rwsem.c | ||
scatterlist.c | ||
sha1.c | ||
show_mem.c | ||
smp_processor_id.c | ||
sort.c | ||
spinlock_debug.c | ||
string_helpers.c | ||
string.c | ||
swiotlb.c | ||
syscall.c | ||
textsearch.c | ||
ts_bm.c | ||
ts_fsm.c | ||
ts_kmp.c | ||
vsprintf.c |