1 // SPDX-License-Identifier: GPL-2.0-or-later
3 * Copyright (C) 2021-2023 Oracle. All Rights Reserved.
4 * Author: Darrick J. Wong <djwong@kernel.org>
8 #include "xfs_shared.h"
9 #include "xfs_format.h"
10 #include "scrub/xfile.h"
11 #include "scrub/xfarray.h"
12 #include "scrub/scrub.h"
13 #include "scrub/trace.h"
16 * Large Arrays of Fixed-Size Records
17 * ==================================
19 * This memory array uses an xfile (which itself is a memfd "file") to store
20 * large numbers of fixed-size records in memory that can be paged out. This
21 * puts less stress on the memory reclaim algorithms during an online repair
22 * because we don't have to pin so much memory. However, array access is less
23 * direct than would be in a regular memory array. Access to the array is
24 * performed via indexed load and store methods, and an append method is
25 * provided for convenience. Array elements can be unset, which sets them to
26 * all zeroes. Unset entries are skipped during iteration, though direct loads
27 * will return a zeroed buffer. Callers are responsible for concurrency
32 * Pointer to scratch space. Because we can't access the xfile data directly,
33 * we allocate a small amount of memory on the end of the xfarray structure to
34 * buffer array items when we need space to store values temporarily.
36 static inline void *xfarray_scratch(struct xfarray *array)
41 /* Compute array index given an xfile offset. */
44 struct xfarray *array,
47 if (array->obj_size_log >= 0)
48 return (xfarray_idx_t)pos >> array->obj_size_log;
50 return div_u64((xfarray_idx_t)pos, array->obj_size);
53 /* Compute xfile offset of array element. */
54 static inline loff_t xfarray_pos(struct xfarray *array, xfarray_idx_t idx)
56 if (array->obj_size_log >= 0)
57 return idx << array->obj_size_log;
59 return idx * array->obj_size;
63 * Initialize a big memory array. Array records cannot be larger than a
64 * page, and the array cannot span more bytes than the page cache supports.
65 * If @required_capacity is nonzero, the maximum array size will be set to this
66 * quantity and the array creation will fail if the underlying storage cannot
67 * support that many records.
71 const char *description,
72 unsigned long long required_capacity,
74 struct xfarray **arrayp)
76 struct xfarray *array;
80 ASSERT(obj_size < PAGE_SIZE);
82 error = xfile_create(description, 0, &xfile);
87 array = kzalloc(sizeof(struct xfarray) + obj_size, XCHK_GFP_FLAGS);
92 array->obj_size = obj_size;
94 if (is_power_of_2(obj_size))
95 array->obj_size_log = ilog2(obj_size);
97 array->obj_size_log = -1;
99 array->max_nr = xfarray_idx(array, MAX_LFS_FILESIZE);
100 trace_xfarray_create(array, required_capacity);
102 if (required_capacity > 0) {
103 if (array->max_nr < required_capacity) {
107 array->max_nr = required_capacity;
116 xfile_destroy(xfile);
120 /* Destroy the array. */
123 struct xfarray *array)
125 xfile_destroy(array->xfile);
129 /* Load an element from the array. */
132 struct xfarray *array,
136 if (idx >= array->nr)
139 return xfile_obj_load(array->xfile, ptr, array->obj_size,
140 xfarray_pos(array, idx));
143 /* Is this array element potentially unset? */
146 struct xfarray *array,
149 void *temp = xfarray_scratch(array);
152 if (array->unset_slots == 0)
155 error = xfile_obj_load(array->xfile, temp, array->obj_size, pos);
156 if (!error && xfarray_element_is_null(array, temp))
163 * Unset an array element. If @idx is the last element in the array, the
164 * array will be truncated. Otherwise, the entry will be zeroed.
168 struct xfarray *array,
171 void *temp = xfarray_scratch(array);
172 loff_t pos = xfarray_pos(array, idx);
175 if (idx >= array->nr)
178 if (idx == array->nr - 1) {
183 if (xfarray_is_unset(array, pos))
186 memset(temp, 0, array->obj_size);
187 error = xfile_obj_store(array->xfile, temp, array->obj_size, pos);
191 array->unset_slots++;
196 * Store an element in the array. The element must not be completely zeroed,
197 * because those are considered unset sparse elements.
201 struct xfarray *array,
207 if (idx >= array->max_nr)
210 ASSERT(!xfarray_element_is_null(array, ptr));
212 ret = xfile_obj_store(array->xfile, ptr, array->obj_size,
213 xfarray_pos(array, idx));
217 array->nr = max(array->nr, idx + 1);
221 /* Is this array element NULL? */
223 xfarray_element_is_null(
224 struct xfarray *array,
227 return !memchr_inv(ptr, 0, array->obj_size);
231 * Store an element anywhere in the array that is unset. If there are no
232 * unset slots, append the element to the array.
235 xfarray_store_anywhere(
236 struct xfarray *array,
239 void *temp = xfarray_scratch(array);
240 loff_t endpos = xfarray_pos(array, array->nr);
244 /* Find an unset slot to put it in. */
246 pos < endpos && array->unset_slots > 0;
247 pos += array->obj_size) {
248 error = xfile_obj_load(array->xfile, temp, array->obj_size,
250 if (error || !xfarray_element_is_null(array, temp))
253 error = xfile_obj_store(array->xfile, ptr, array->obj_size,
258 array->unset_slots--;
262 /* No unset slots found; attach it on the end. */
263 array->unset_slots = 0;
264 return xfarray_append(array, ptr);
267 /* Return length of array. */
270 struct xfarray *array)
276 * Decide which array item we're going to read as part of an _iter_get.
277 * @cur is the array index, and @pos is the file offset of that array index in
278 * the backing xfile. Returns ENODATA if we reach the end of the records.
280 * Reading from a hole in a sparse xfile causes page instantiation, so for
281 * iterating a (possibly sparse) array we need to figure out if the cursor is
282 * pointing at a totally uninitialized hole and move the cursor up if
287 struct xfarray *array,
291 unsigned int pgoff = offset_in_page(*pos);
292 loff_t end_pos = *pos + array->obj_size - 1;
296 * If the current array record is not adjacent to a page boundary, we
297 * are in the middle of the page. We do not need to move the cursor.
299 if (pgoff != 0 && pgoff + array->obj_size - 1 < PAGE_SIZE)
303 * Call SEEK_DATA on the last byte in the record we're about to read.
304 * If the record ends at (or crosses) the end of a page then we know
305 * that the first byte of the record is backed by pages and don't need
306 * to query it. If instead the record begins at the start of the page
307 * then we know that querying the last byte is just as good as querying
308 * the first byte, since records cannot be larger than a page.
310 * If the call returns the same file offset, we know this record is
311 * backed by real pages. We do not need to move the cursor.
313 new_pos = xfile_seek_data(array->xfile, end_pos);
314 if (new_pos == -ENXIO)
318 if (new_pos == end_pos)
322 * Otherwise, SEEK_DATA told us how far up to move the file pointer to
323 * find more data. Move the array index to the first record past the
324 * byte offset we were given.
326 new_pos = roundup_64(new_pos, array->obj_size);
327 *cur = xfarray_idx(array, new_pos);
328 *pos = xfarray_pos(array, *cur);
333 * Starting at *idx, fetch the next non-null array entry and advance the index
334 * to set up the next _load_next call. Returns ENODATA if we reach the end of
335 * the array. Callers must set @*idx to XFARRAY_CURSOR_INIT before the first
336 * call to this function.
340 struct xfarray *array,
344 xfarray_idx_t cur = *idx;
345 loff_t pos = xfarray_pos(array, cur);
349 if (cur >= array->nr)
353 * Ask the backing store for the location of next possible
354 * written record, then retrieve that record.
356 error = xfarray_find_data(array, &cur, &pos);
359 error = xfarray_load(array, cur, rec);
364 pos += array->obj_size;
365 } while (xfarray_element_is_null(array, rec));
371 /* Sorting functions */
374 # define xfarray_sort_bump_loads(si) do { (si)->loads++; } while (0)
375 # define xfarray_sort_bump_stores(si) do { (si)->stores++; } while (0)
376 # define xfarray_sort_bump_compares(si) do { (si)->compares++; } while (0)
378 # define xfarray_sort_bump_loads(si)
379 # define xfarray_sort_bump_stores(si)
380 # define xfarray_sort_bump_compares(si)
383 /* Load an array element for sorting. */
386 struct xfarray_sortinfo *si,
390 xfarray_sort_bump_loads(si);
391 return xfarray_load(si->array, idx, ptr);
394 /* Store an array element for sorting. */
397 struct xfarray_sortinfo *si,
401 xfarray_sort_bump_stores(si);
402 return xfarray_store(si->array, idx, ptr);
405 /* Compare an array element for sorting. */
408 struct xfarray_sortinfo *si,
412 xfarray_sort_bump_compares(si);
413 return si->cmp_fn(a, b);
416 /* Return a pointer to the low index stack for quicksort partitioning. */
417 static inline xfarray_idx_t *xfarray_sortinfo_lo(struct xfarray_sortinfo *si)
419 return (xfarray_idx_t *)(si + 1);
422 /* Return a pointer to the high index stack for quicksort partitioning. */
423 static inline xfarray_idx_t *xfarray_sortinfo_hi(struct xfarray_sortinfo *si)
425 return xfarray_sortinfo_lo(si) + si->max_stack_depth;
428 /* Allocate memory to handle the sort. */
430 xfarray_sortinfo_alloc(
431 struct xfarray *array,
432 xfarray_cmp_fn cmp_fn,
434 struct xfarray_sortinfo **infop)
436 struct xfarray_sortinfo *si;
437 size_t nr_bytes = sizeof(struct xfarray_sortinfo);
441 * Tail-call recursion during the partitioning phase means that
442 * quicksort will never recurse more than log2(nr) times. We need one
443 * extra level of stack to hold the initial parameters.
445 max_stack_depth = ilog2(array->nr) + 1;
447 /* Each level of quicksort uses a lo and a hi index */
448 nr_bytes += max_stack_depth * sizeof(xfarray_idx_t) * 2;
450 /* One record for the pivot */
451 nr_bytes += array->obj_size;
453 si = kvzalloc(nr_bytes, XCHK_GFP_FLAGS);
460 si->max_stack_depth = max_stack_depth;
461 si->max_stack_used = 1;
463 xfarray_sortinfo_lo(si)[0] = 0;
464 xfarray_sortinfo_hi(si)[0] = array->nr - 1;
466 trace_xfarray_sort(si, nr_bytes);
471 /* Should this sort be terminated by a fatal signal? */
473 xfarray_sort_terminated(
474 struct xfarray_sortinfo *si,
478 * If preemption is disabled, we need to yield to the scheduler every
479 * few seconds so that we don't run afoul of the soft lockup watchdog
480 * or RCU stall detector.
484 if ((si->flags & XFARRAY_SORT_KILLABLE) &&
485 fatal_signal_pending(current)) {
493 /* Do we want an insertion sort? */
496 struct xfarray_sortinfo *si,
501 * For array subsets smaller than 8 elements, it's slightly faster to
502 * use insertion sort than quicksort's stack machine.
504 return (end - start) < 8;
507 /* Return the scratch space within the sortinfo structure. */
508 static inline void *xfarray_sortinfo_isort_scratch(struct xfarray_sortinfo *si)
510 return xfarray_sortinfo_hi(si) + si->max_stack_depth;
514 * Perform an insertion sort on a subset of the array.
515 * Though insertion sort is an O(n^2) algorithm, for small set sizes it's
516 * faster than quicksort's stack machine, so we let it take over for that.
517 * This ought to be replaced with something more efficient.
521 struct xfarray_sortinfo *si,
525 void *a = xfarray_sortinfo_isort_scratch(si);
526 void *b = xfarray_scratch(si->array);
532 trace_xfarray_isort(si, lo, hi);
535 * Move the smallest element in a[lo..hi] to a[lo]. This
536 * simplifies the loop control logic below.
539 error = xfarray_sort_load(si, tmp, b);
542 for (run = lo + 1; run <= hi; run++) {
543 /* if a[run] < a[tmp], tmp = run */
544 error = xfarray_sort_load(si, run, a);
547 if (xfarray_sort_cmp(si, a, b) < 0) {
549 memcpy(b, a, si->array->obj_size);
552 if (xfarray_sort_terminated(si, &error))
557 * The smallest element is a[tmp]; swap with a[lo] if tmp != lo.
558 * Recall that a[tmp] is already in *b.
561 error = xfarray_sort_load(si, lo, a);
564 error = xfarray_sort_store(si, tmp, a);
567 error = xfarray_sort_store(si, lo, b);
573 * Perform an insertion sort on a[lo+1..hi]. We already made sure
574 * that the smallest value in the original range is now in a[lo],
575 * so the inner loop should never underflow.
577 * For each a[lo+2..hi], make sure it's in the correct position
578 * with respect to the elements that came before it.
580 for (run = lo + 2; run <= hi; run++) {
581 error = xfarray_sort_load(si, run, a);
586 * Find the correct place for a[run] by walking leftwards
587 * towards the start of the range until a[tmp] is no longer
588 * greater than a[run].
591 error = xfarray_sort_load(si, tmp, b);
594 while (xfarray_sort_cmp(si, a, b) < 0) {
596 error = xfarray_sort_load(si, tmp, b);
600 if (xfarray_sort_terminated(si, &error))
606 * If tmp != run, then a[tmp..run-1] are all less than a[run],
607 * so right barrel roll a[tmp..run] to get this range in
613 for (i = run; i >= tmp; i--) {
614 error = xfarray_sort_load(si, i - 1, b);
617 error = xfarray_sort_store(si, i, b);
621 if (xfarray_sort_terminated(si, &error))
624 error = xfarray_sort_store(si, tmp, a);
628 if (xfarray_sort_terminated(si, &error))
635 /* Return a pointer to the xfarray pivot record within the sortinfo struct. */
636 static inline void *xfarray_sortinfo_pivot(struct xfarray_sortinfo *si)
638 return xfarray_sortinfo_hi(si) + si->max_stack_depth;
642 * Find a pivot value for quicksort partitioning, swap it with a[lo], and save
643 * the cached pivot record for the next step.
645 * Select the median value from a[lo], a[mid], and a[hi]. Put the median in
646 * a[lo], the lowest in a[mid], and the highest in a[hi]. Using the median of
647 * the three reduces the chances that we pick the worst case pivot value, since
648 * it's likely that our array values are nearly sorted.
652 struct xfarray_sortinfo *si,
656 void *a = xfarray_sortinfo_pivot(si);
657 void *b = xfarray_scratch(si->array);
658 xfarray_idx_t mid = lo + ((hi - lo) / 2);
661 /* if a[mid] < a[lo], swap a[mid] and a[lo]. */
662 error = xfarray_sort_load(si, mid, a);
665 error = xfarray_sort_load(si, lo, b);
668 if (xfarray_sort_cmp(si, a, b) < 0) {
669 error = xfarray_sort_store(si, lo, a);
672 error = xfarray_sort_store(si, mid, b);
677 /* if a[hi] < a[mid], swap a[mid] and a[hi]. */
678 error = xfarray_sort_load(si, hi, a);
681 error = xfarray_sort_load(si, mid, b);
684 if (xfarray_sort_cmp(si, a, b) < 0) {
685 error = xfarray_sort_store(si, mid, a);
688 error = xfarray_sort_store(si, hi, b);
695 /* if a[mid] < a[lo], swap a[mid] and a[lo]. */
696 error = xfarray_sort_load(si, mid, a);
699 error = xfarray_sort_load(si, lo, b);
702 if (xfarray_sort_cmp(si, a, b) < 0) {
703 error = xfarray_sort_store(si, lo, a);
706 error = xfarray_sort_store(si, mid, b);
713 * Move our selected pivot to a[lo]. Recall that a == si->pivot, so
714 * this leaves us with the pivot cached in the sortinfo structure.
716 error = xfarray_sort_load(si, lo, b);
719 error = xfarray_sort_load(si, mid, a);
722 error = xfarray_sort_store(si, mid, b);
725 return xfarray_sort_store(si, lo, a);
729 * Set up the pointers for the next iteration. We push onto the stack all of
730 * the unsorted values between a[lo + 1] and a[end[i]], and we tweak the
731 * current stack frame to point to the unsorted values between a[beg[i]] and
732 * a[lo] so that those values will be sorted when we pop the stack.
736 struct xfarray_sortinfo *si,
737 xfarray_idx_t *si_lo,
738 xfarray_idx_t *si_hi,
742 /* Check for stack overflows */
743 if (si->stack_depth >= si->max_stack_depth - 1) {
744 ASSERT(si->stack_depth < si->max_stack_depth - 1);
745 return -EFSCORRUPTED;
748 si->max_stack_used = max_t(uint8_t, si->max_stack_used,
749 si->stack_depth + 2);
751 si_lo[si->stack_depth + 1] = lo + 1;
752 si_hi[si->stack_depth + 1] = si_hi[si->stack_depth];
753 si_hi[si->stack_depth++] = lo - 1;
756 * Always start with the smaller of the two partitions to keep the
757 * amount of recursion in check.
759 if (si_hi[si->stack_depth] - si_lo[si->stack_depth] >
760 si_hi[si->stack_depth - 1] - si_lo[si->stack_depth - 1]) {
761 swap(si_lo[si->stack_depth], si_lo[si->stack_depth - 1]);
762 swap(si_hi[si->stack_depth], si_hi[si->stack_depth - 1]);
769 * Sort the array elements via quicksort. This implementation incorporates
770 * four optimizations discussed in Sedgewick:
772 * 1. Use an explicit stack of array indices to store the next array partition
773 * to sort. This helps us to avoid recursion in the call stack, which is
774 * particularly expensive in the kernel.
776 * 2. For arrays with records in arbitrary or user-controlled order, choose the
777 * pivot element using a median-of-three decision tree. This reduces the
778 * probability of selecting a bad pivot value which causes worst case
779 * behavior (i.e. partition sizes of 1).
781 * 3. The smaller of the two sub-partitions is pushed onto the stack to start
782 * the next level of recursion, and the larger sub-partition replaces the
783 * current stack frame. This guarantees that we won't need more than
784 * log2(nr) stack space.
786 * 4. Use insertion sort for small sets since since insertion sort is faster
787 * for small, mostly sorted array segments. In the author's experience,
788 * substituting insertion sort for arrays smaller than 8 elements yields
789 * a ~10% reduction in runtime.
793 * Due to the use of signed indices, we can only support up to 2^63 records.
794 * Files can only grow to 2^63 bytes, so this is not much of a limitation.
796 #define QSORT_MAX_RECS (1ULL << 63)
800 struct xfarray *array,
801 xfarray_cmp_fn cmp_fn,
804 struct xfarray_sortinfo *si;
805 xfarray_idx_t *si_lo, *si_hi;
807 void *scratch = xfarray_scratch(array);
808 xfarray_idx_t lo, hi;
813 if (array->nr >= QSORT_MAX_RECS)
816 error = xfarray_sortinfo_alloc(array, cmp_fn, flags, &si);
819 si_lo = xfarray_sortinfo_lo(si);
820 si_hi = xfarray_sortinfo_hi(si);
821 pivot = xfarray_sortinfo_pivot(si);
823 while (si->stack_depth >= 0) {
824 lo = si_lo[si->stack_depth];
825 hi = si_hi[si->stack_depth];
827 trace_xfarray_qsort(si, lo, hi);
829 /* Nothing left in this partition to sort; pop stack. */
835 /* If insertion sort can solve our problems, we're done. */
836 if (xfarray_want_isort(si, lo, hi)) {
837 error = xfarray_isort(si, lo, hi);
844 /* Pick a pivot, move it to a[lo] and stash it. */
845 error = xfarray_qsort_pivot(si, lo, hi);
850 * Rearrange a[lo..hi] such that everything smaller than the
851 * pivot is on the left side of the range and everything larger
852 * than the pivot is on the right side of the range.
856 * Decrement hi until it finds an a[hi] less than the
859 error = xfarray_sort_load(si, hi, scratch);
862 while (xfarray_sort_cmp(si, scratch, pivot) >= 0 &&
864 if (xfarray_sort_terminated(si, &error))
868 error = xfarray_sort_load(si, hi, scratch);
873 if (xfarray_sort_terminated(si, &error))
876 /* Copy that item (a[hi]) to a[lo]. */
878 error = xfarray_sort_store(si, lo++, scratch);
884 * Increment lo until it finds an a[lo] greater than
887 error = xfarray_sort_load(si, lo, scratch);
890 while (xfarray_sort_cmp(si, scratch, pivot) <= 0 &&
892 if (xfarray_sort_terminated(si, &error))
896 error = xfarray_sort_load(si, lo, scratch);
901 if (xfarray_sort_terminated(si, &error))
904 /* Copy that item (a[lo]) to a[hi]. */
906 error = xfarray_sort_store(si, hi--, scratch);
911 if (xfarray_sort_terminated(si, &error))
916 * Put our pivot value in the correct place at a[lo]. All
917 * values between a[beg[i]] and a[lo - 1] should be less than
918 * the pivot; and all values between a[lo + 1] and a[end[i]-1]
919 * should be greater than the pivot.
921 error = xfarray_sort_store(si, lo, pivot);
925 /* Set up the stack frame to process the two partitions. */
926 error = xfarray_qsort_push(si, si_lo, si_hi, lo, hi);
930 if (xfarray_sort_terminated(si, &error))
935 trace_xfarray_sort_stats(si, error);