xfs: improve xfarray quicksort pivot

Now that we have the means to do insertion sorts of small in-memory
subsets of an xfarray, use it to improve the quicksort pivot algorithm
by reading 7 records into memory and finding the median of that.  This
should prevent bad partitioning when a[lo] and a[hi] end up next to each
other in the final sort, which can happen when sorting for cntbt repair
when the free space is extremely fragmented (e.g. generic/176).

This doesn't speed up the average quicksort run by much, but it will
(hopefully) avoid the quadratic time collapse for which quicksort is
famous.

Signed-off-by: Darrick J. Wong <djwong@kernel.org>
Reviewed-by: Kent Overstreet <kent.overstreet@linux.dev>
Reviewed-by: Dave Chinner <dchinner@redhat.com>
This commit is contained in:
Darrick J. Wong 2023-08-10 07:48:07 -07:00
parent cf36f4f64c
commit 764018caa9
2 changed files with 154 additions and 75 deletions

View File

@ -427,6 +427,14 @@ static inline xfarray_idx_t *xfarray_sortinfo_hi(struct xfarray_sortinfo *si)
return xfarray_sortinfo_lo(si) + si->max_stack_depth;
}
/* Size of each element in the quicksort pivot array. */
static inline size_t
xfarray_pivot_rec_sz(
struct xfarray *array)
{
return round_up(array->obj_size, 8) + sizeof(xfarray_idx_t);
}
/* Allocate memory to handle the sort. */
static inline int
xfarray_sortinfo_alloc(
@ -437,8 +445,16 @@ xfarray_sortinfo_alloc(
{
struct xfarray_sortinfo *si;
size_t nr_bytes = sizeof(struct xfarray_sortinfo);
size_t pivot_rec_sz = xfarray_pivot_rec_sz(array);
int max_stack_depth;
/*
* The median-of-nine pivot algorithm doesn't work if a subset has
* fewer than 9 items. Make sure the in-memory sort will always take
* over for subsets where this wouldn't be the case.
*/
BUILD_BUG_ON(XFARRAY_QSORT_PIVOT_NR >= XFARRAY_ISORT_NR);
/*
* Tail-call recursion during the partitioning phase means that
* quicksort will never recurse more than log2(nr) times. We need one
@ -453,8 +469,10 @@ xfarray_sortinfo_alloc(
/* Each level of quicksort uses a lo and a hi index */
nr_bytes += max_stack_depth * sizeof(xfarray_idx_t) * 2;
/* Scratchpad for in-memory sort, or one record for the pivot */
nr_bytes += (XFARRAY_ISORT_NR * array->obj_size);
/* Scratchpad for in-memory sort, or finding the pivot */
nr_bytes += max_t(size_t,
(XFARRAY_QSORT_PIVOT_NR + 1) * pivot_rec_sz,
XFARRAY_ISORT_NR * array->obj_size);
si = kvzalloc(nr_bytes, XCHK_GFP_FLAGS);
if (!si)
@ -632,14 +650,43 @@ static inline void *xfarray_sortinfo_pivot(struct xfarray_sortinfo *si)
return xfarray_sortinfo_hi(si) + si->max_stack_depth;
}
/* Return a pointer to the start of the pivot array. */
static inline void *
xfarray_sortinfo_pivot_array(
struct xfarray_sortinfo *si)
{
return xfarray_sortinfo_pivot(si) + si->array->obj_size;
}
/* The xfarray record is stored at the start of each pivot array element. */
static inline void *
xfarray_pivot_array_rec(
void *pa,
size_t pa_recsz,
unsigned int pa_idx)
{
return pa + (pa_recsz * pa_idx);
}
/* The xfarray index is stored at the end of each pivot array element. */
static inline xfarray_idx_t *
xfarray_pivot_array_idx(
void *pa,
size_t pa_recsz,
unsigned int pa_idx)
{
return xfarray_pivot_array_rec(pa, pa_recsz, pa_idx + 1) -
sizeof(xfarray_idx_t);
}
/*
* Find a pivot value for quicksort partitioning, swap it with a[lo], and save
* the cached pivot record for the next step.
*
* Select the median value from a[lo], a[mid], and a[hi]. Put the median in
* a[lo], the lowest in a[mid], and the highest in a[hi]. Using the median of
* the three reduces the chances that we pick the worst case pivot value, since
* it's likely that our array values are nearly sorted.
* Load evenly-spaced records within the given range into memory, sort them,
* and choose the pivot from the median record. Using multiple points will
* improve the quality of the pivot selection, and hopefully avoid the worst
* quicksort behavior, since our array values are nearly always evenly sorted.
*/
STATIC int
xfarray_qsort_pivot(
@ -647,76 +694,99 @@ xfarray_qsort_pivot(
xfarray_idx_t lo,
xfarray_idx_t hi)
{
void *a = xfarray_sortinfo_pivot(si);
void *b = xfarray_scratch(si->array);
xfarray_idx_t mid = lo + ((hi - lo) / 2);
void *pivot = xfarray_sortinfo_pivot(si);
void *parray = xfarray_sortinfo_pivot_array(si);
void *recp;
xfarray_idx_t *idxp;
xfarray_idx_t step = (hi - lo) / (XFARRAY_QSORT_PIVOT_NR - 1);
size_t pivot_rec_sz = xfarray_pivot_rec_sz(si->array);
int i, j;
int error;
/* if a[mid] < a[lo], swap a[mid] and a[lo]. */
error = xfarray_sort_load(si, mid, a);
if (error)
return error;
error = xfarray_sort_load(si, lo, b);
if (error)
return error;
if (xfarray_sort_cmp(si, a, b) < 0) {
error = xfarray_sort_store(si, lo, a);
if (error)
return error;
error = xfarray_sort_store(si, mid, b);
if (error)
return error;
}
ASSERT(step > 0);
/* if a[hi] < a[mid], swap a[mid] and a[hi]. */
error = xfarray_sort_load(si, hi, a);
if (error)
return error;
error = xfarray_sort_load(si, mid, b);
if (error)
return error;
if (xfarray_sort_cmp(si, a, b) < 0) {
error = xfarray_sort_store(si, mid, a);
if (error)
return error;
error = xfarray_sort_store(si, hi, b);
if (error)
return error;
} else {
goto move_front;
}
/* if a[mid] < a[lo], swap a[mid] and a[lo]. */
error = xfarray_sort_load(si, mid, a);
if (error)
return error;
error = xfarray_sort_load(si, lo, b);
if (error)
return error;
if (xfarray_sort_cmp(si, a, b) < 0) {
error = xfarray_sort_store(si, lo, a);
if (error)
return error;
error = xfarray_sort_store(si, mid, b);
if (error)
return error;
}
move_front:
/*
* Move our selected pivot to a[lo]. Recall that a == si->pivot, so
* this leaves us with the pivot cached in the sortinfo structure.
* Load the xfarray indexes of the records we intend to sample into the
* pivot array.
*/
error = xfarray_sort_load(si, lo, b);
idxp = xfarray_pivot_array_idx(parray, pivot_rec_sz, 0);
*idxp = lo;
for (i = 1; i < XFARRAY_QSORT_PIVOT_NR - 1; i++) {
idxp = xfarray_pivot_array_idx(parray, pivot_rec_sz, i);
*idxp = lo + (i * step);
}
idxp = xfarray_pivot_array_idx(parray, pivot_rec_sz,
XFARRAY_QSORT_PIVOT_NR - 1);
*idxp = hi;
/* Load the selected xfarray records into the pivot array. */
for (i = 0; i < XFARRAY_QSORT_PIVOT_NR; i++) {
xfarray_idx_t idx;
recp = xfarray_pivot_array_rec(parray, pivot_rec_sz, i);
idxp = xfarray_pivot_array_idx(parray, pivot_rec_sz, i);
/* No unset records; load directly into the array. */
if (likely(si->array->unset_slots == 0)) {
error = xfarray_sort_load(si, *idxp, recp);
if (error)
return error;
continue;
}
/*
* Load non-null records into the scratchpad without changing
* the xfarray_idx_t in the pivot array.
*/
idx = *idxp;
xfarray_sort_bump_loads(si);
error = xfarray_load_next(si->array, &idx, recp);
if (error)
return error;
}
xfarray_sort_bump_heapsorts(si);
sort(parray, XFARRAY_QSORT_PIVOT_NR, pivot_rec_sz, si->cmp_fn, NULL);
/*
* We sorted the pivot array records (which includes the xfarray
* indices) in xfarray record order. The median element of the pivot
* array contains the xfarray record that we will use as the pivot.
* Copy that xfarray record to the designated space.
*/
recp = xfarray_pivot_array_rec(parray, pivot_rec_sz,
XFARRAY_QSORT_PIVOT_NR / 2);
memcpy(pivot, recp, si->array->obj_size);
/* If the pivot record we chose was already in a[lo] then we're done. */
idxp = xfarray_pivot_array_idx(parray, pivot_rec_sz,
XFARRAY_QSORT_PIVOT_NR / 2);
if (*idxp == lo)
return 0;
/*
* Find the cached copy of a[lo] in the pivot array so that we can swap
* a[lo] and a[pivot].
*/
for (i = 0, j = -1; i < XFARRAY_QSORT_PIVOT_NR; i++) {
idxp = xfarray_pivot_array_idx(parray, pivot_rec_sz, i);
if (*idxp == lo)
j = i;
}
if (j < 0) {
ASSERT(j >= 0);
return -EFSCORRUPTED;
}
/* Swap a[lo] and a[pivot]. */
error = xfarray_sort_store(si, lo, pivot);
if (error)
return error;
error = xfarray_sort_load(si, mid, a);
if (error)
return error;
error = xfarray_sort_store(si, mid, b);
if (error)
return error;
return xfarray_sort_store(si, lo, a);
recp = xfarray_pivot_array_rec(parray, pivot_rec_sz, j);
idxp = xfarray_pivot_array_idx(parray, pivot_rec_sz,
XFARRAY_QSORT_PIVOT_NR / 2);
return xfarray_sort_store(si, *idxp, recp);
}
/*
@ -828,7 +898,7 @@ xfarray_sort_load_cached(
* particularly expensive in the kernel.
*
* 2. For arrays with records in arbitrary or user-controlled order, choose the
* pivot element using a median-of-three decision tree. This reduces the
* pivot element using a median-of-nine decision tree. This reduces the
* probability of selecting a bad pivot value which causes worst case
* behavior (i.e. partition sizes of 1).
*

View File

@ -62,6 +62,9 @@ typedef cmp_func_t xfarray_cmp_fn;
#define XFARRAY_ISORT_SHIFT (4)
#define XFARRAY_ISORT_NR (1U << XFARRAY_ISORT_SHIFT)
/* Evalulate this many points to find the qsort pivot. */
#define XFARRAY_QSORT_PIVOT_NR (9)
struct xfarray_sortinfo {
struct xfarray *array;
@ -91,7 +94,6 @@ struct xfarray_sortinfo {
uint64_t compares;
uint64_t heapsorts;
#endif
/*
* Extra bytes are allocated beyond the end of the structure to store
* quicksort information. C does not permit multiple VLAs per struct,
@ -114,11 +116,18 @@ struct xfarray_sortinfo {
* xfarray_rec_t scratch[ISORT_NR];
*
* Otherwise, we want to partition the records to partition the array.
* We store the chosen pivot record here and use the xfarray scratchpad
* to rearrange the array around the pivot:
*
* xfarray_rec_t pivot;
* We store the chosen pivot record at the start of the scratchpad area
* and use the rest to sample some records to estimate the median.
* The format of the qsort_pivot array enables us to use the kernel
* heapsort function to place the median value in the middle.
*
* struct {
* xfarray_rec_t pivot;
* struct {
* xfarray_rec_t rec; (rounded up to 8 bytes)
* xfarray_idx_t idx;
* } qsort_pivot[QSORT_PIVOT_NR];
* };
* }
*/
};