linux/lib/list_sort.c

218 lines
5.3 KiB
C
Raw Normal View History

#include <linux/kernel.h>
#include <linux/module.h>
#include <linux/list_sort.h>
#include <linux/slab.h>
#include <linux/list.h>
lib: more scalable list_sort() 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>
2010-03-06 05:43:15 +08:00
#define MAX_LIST_LENGTH_BITS 20
/*
* Returns a list organized in an intermediate format suited
* to chaining of merge() calls: null-terminated, no reserved or
* sentinel head node, "prev" links not maintained.
*/
static struct list_head *merge(void *priv,
int (*cmp)(void *priv, struct list_head *a,
struct list_head *b),
struct list_head *a, struct list_head *b)
{
struct list_head head, *tail = &head;
while (a && b) {
/* if equal, take 'a' -- important for sort stability */
if ((*cmp)(priv, a, b) <= 0) {
tail->next = a;
a = a->next;
} else {
tail->next = b;
b = b->next;
}
tail = tail->next;
}
tail->next = a?:b;
return head.next;
}
/*
* Combine final list merge with restoration of standard doubly-linked
* list structure. This approach duplicates code from merge(), but
* runs faster than the tidier alternatives of either a separate final
* prev-link restoration pass, or maintaining the prev links
* throughout.
*/
static void merge_and_restore_back_links(void *priv,
int (*cmp)(void *priv, struct list_head *a,
struct list_head *b),
struct list_head *head,
struct list_head *a, struct list_head *b)
{
struct list_head *tail = head;
while (a && b) {
/* if equal, take 'a' -- important for sort stability */
if ((*cmp)(priv, a, b) <= 0) {
tail->next = a;
a->prev = tail;
a = a->next;
} else {
tail->next = b;
b->prev = tail;
b = b->next;
}
tail = tail->next;
}
tail->next = a ? : b;
do {
/*
* In worst cases this loop may run many iterations.
* Continue callbacks to the client even though no
* element comparison is needed, so the client's cmp()
* routine can invoke cond_resched() periodically.
*/
(*cmp)(priv, tail->next, tail->next);
lib: more scalable list_sort() 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>
2010-03-06 05:43:15 +08:00
tail->next->prev = tail;
tail = tail->next;
} while (tail->next);
tail->next = head;
head->prev = tail;
}
/**
* list_sort - sort a list
* @priv: private data, opaque to list_sort(), passed to @cmp
* @head: the list to sort
* @cmp: the elements comparison function
*
* This function implements "merge sort", which has O(nlog(n))
* complexity.
*
* The comparison function @cmp must return a negative value if @a
* should sort before @b, and a positive value if @a should sort after
* @b. If @a and @b are equivalent, and their original relative
* ordering is to be preserved, @cmp must return 0.
*/
void list_sort(void *priv, struct list_head *head,
lib: more scalable list_sort() 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>
2010-03-06 05:43:15 +08:00
int (*cmp)(void *priv, struct list_head *a,
struct list_head *b))
{
lib: more scalable list_sort() 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>
2010-03-06 05:43:15 +08:00
struct list_head *part[MAX_LIST_LENGTH_BITS+1]; /* sorted partial lists
-- last slot is a sentinel */
int lev; /* index into part[] */
int max_lev = 0;
struct list_head *list;
if (list_empty(head))
return;
lib: more scalable list_sort() 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>
2010-03-06 05:43:15 +08:00
memset(part, 0, sizeof(part));
head->prev->next = NULL;
list = head->next;
lib: more scalable list_sort() 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>
2010-03-06 05:43:15 +08:00
while (list) {
struct list_head *cur = list;
list = list->next;
cur->next = NULL;
for (lev = 0; part[lev]; lev++) {
cur = merge(priv, cmp, part[lev], cur);
part[lev] = NULL;
}
if (lev > max_lev) {
if (unlikely(lev >= ARRAY_SIZE(part)-1)) {
printk_once(KERN_DEBUG "list passed to"
" list_sort() too long for"
" efficiency\n");
lev--;
}
lib: more scalable list_sort() 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>
2010-03-06 05:43:15 +08:00
max_lev = lev;
}
lib: more scalable list_sort() 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>
2010-03-06 05:43:15 +08:00
part[lev] = cur;
}
lib: more scalable list_sort() 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>
2010-03-06 05:43:15 +08:00
for (lev = 0; lev < max_lev; lev++)
if (part[lev])
list = merge(priv, cmp, part[lev], list);
lib: more scalable list_sort() 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>
2010-03-06 05:43:15 +08:00
merge_and_restore_back_links(priv, cmp, head, part[max_lev], list);
}
EXPORT_SYMBOL(list_sort);
lib: more scalable list_sort() 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>
2010-03-06 05:43:15 +08:00
#ifdef DEBUG_LIST_SORT
struct debug_el {
struct list_head l_h;
int value;
unsigned serial;
};
lib: more scalable list_sort() 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>
2010-03-06 05:43:15 +08:00
static int cmp(void *priv, struct list_head *a, struct list_head *b)
{
return container_of(a, struct debug_el, l_h)->value
- container_of(b, struct debug_el, l_h)->value;
}
lib: more scalable list_sort() 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>
2010-03-06 05:43:15 +08:00
/*
* The pattern of set bits in the list length determines which cases
* are hit in list_sort().
*/
#define LIST_SORT_TEST_LENGTH (512+128+2) /* not including head */
static int __init list_sort_test(void)
{
int i, r = 1, count;
struct list_head *head = kmalloc(sizeof(*head), GFP_KERNEL);
struct list_head *cur;
printk(KERN_WARNING "testing list_sort()\n");
cur = head;
for (i = 0; i < LIST_SORT_TEST_LENGTH; i++) {
struct debug_el *el = kmalloc(sizeof(*el), GFP_KERNEL);
BUG_ON(!el);
/* force some equivalencies */
el->value = (r = (r * 725861) % 6599) % (LIST_SORT_TEST_LENGTH/3);
el->serial = i;
el->l_h.prev = cur;
cur->next = &el->l_h;
cur = cur->next;
}
head->prev = cur;
list_sort(NULL, head, cmp);
count = 1;
for (cur = head->next; cur->next != head; cur = cur->next) {
struct debug_el *el = container_of(cur, struct debug_el, l_h);
int cmp_result = cmp(NULL, cur, cur->next);
if (cur->next->prev != cur) {
printk(KERN_EMERG "list_sort() returned "
"a corrupted list!\n");
return 1;
} else if (cmp_result > 0) {
printk(KERN_EMERG "list_sort() failed to sort!\n");
return 1;
} else if (cmp_result == 0 &&
el->serial >= container_of(cur->next,
struct debug_el, l_h)->serial) {
printk(KERN_EMERG "list_sort() failed to preserve order"
" of equivalent elements!\n");
return 1;
}
kfree(cur->prev);
count++;
}
kfree(cur);
if (count != LIST_SORT_TEST_LENGTH) {
printk(KERN_EMERG "list_sort() returned list of"
"different length!\n");
return 1;
}
return 0;
}
module_init(list_sort_test);
#endif