xref: /openbmc/linux/fs/xfs/scrub/xfarray.c (revision 22b6e7f3)
1 // SPDX-License-Identifier: GPL-2.0-or-later
2 /*
3  * Copyright (C) 2021-2023 Oracle.  All Rights Reserved.
4  * Author: Darrick J. Wong <djwong@kernel.org>
5  */
6 #include "xfs.h"
7 #include "xfs_fs.h"
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"
14 
15 /*
16  * Large Arrays of Fixed-Size Records
17  * ==================================
18  *
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
28  * control.
29  */
30 
31 /*
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.
35  */
36 static inline void *xfarray_scratch(struct xfarray *array)
37 {
38 	return (array + 1);
39 }
40 
41 /* Compute array index given an xfile offset. */
42 static xfarray_idx_t
43 xfarray_idx(
44 	struct xfarray	*array,
45 	loff_t		pos)
46 {
47 	if (array->obj_size_log >= 0)
48 		return (xfarray_idx_t)pos >> array->obj_size_log;
49 
50 	return div_u64((xfarray_idx_t)pos, array->obj_size);
51 }
52 
53 /* Compute xfile offset of array element. */
54 static inline loff_t xfarray_pos(struct xfarray *array, xfarray_idx_t idx)
55 {
56 	if (array->obj_size_log >= 0)
57 		return idx << array->obj_size_log;
58 
59 	return idx * array->obj_size;
60 }
61 
62 /*
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.
68  */
69 int
70 xfarray_create(
71 	const char		*description,
72 	unsigned long long	required_capacity,
73 	size_t			obj_size,
74 	struct xfarray		**arrayp)
75 {
76 	struct xfarray		*array;
77 	struct xfile		*xfile;
78 	int			error;
79 
80 	ASSERT(obj_size < PAGE_SIZE);
81 
82 	error = xfile_create(description, 0, &xfile);
83 	if (error)
84 		return error;
85 
86 	error = -ENOMEM;
87 	array = kzalloc(sizeof(struct xfarray) + obj_size, XCHK_GFP_FLAGS);
88 	if (!array)
89 		goto out_xfile;
90 
91 	array->xfile = xfile;
92 	array->obj_size = obj_size;
93 
94 	if (is_power_of_2(obj_size))
95 		array->obj_size_log = ilog2(obj_size);
96 	else
97 		array->obj_size_log = -1;
98 
99 	array->max_nr = xfarray_idx(array, MAX_LFS_FILESIZE);
100 	trace_xfarray_create(array, required_capacity);
101 
102 	if (required_capacity > 0) {
103 		if (array->max_nr < required_capacity) {
104 			error = -ENOMEM;
105 			goto out_xfarray;
106 		}
107 		array->max_nr = required_capacity;
108 	}
109 
110 	*arrayp = array;
111 	return 0;
112 
113 out_xfarray:
114 	kfree(array);
115 out_xfile:
116 	xfile_destroy(xfile);
117 	return error;
118 }
119 
120 /* Destroy the array. */
121 void
122 xfarray_destroy(
123 	struct xfarray	*array)
124 {
125 	xfile_destroy(array->xfile);
126 	kfree(array);
127 }
128 
129 /* Load an element from the array. */
130 int
131 xfarray_load(
132 	struct xfarray	*array,
133 	xfarray_idx_t	idx,
134 	void		*ptr)
135 {
136 	if (idx >= array->nr)
137 		return -ENODATA;
138 
139 	return xfile_obj_load(array->xfile, ptr, array->obj_size,
140 			xfarray_pos(array, idx));
141 }
142 
143 /* Is this array element potentially unset? */
144 static inline bool
145 xfarray_is_unset(
146 	struct xfarray	*array,
147 	loff_t		pos)
148 {
149 	void		*temp = xfarray_scratch(array);
150 	int		error;
151 
152 	if (array->unset_slots == 0)
153 		return false;
154 
155 	error = xfile_obj_load(array->xfile, temp, array->obj_size, pos);
156 	if (!error && xfarray_element_is_null(array, temp))
157 		return true;
158 
159 	return false;
160 }
161 
162 /*
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.
165  */
166 int
167 xfarray_unset(
168 	struct xfarray	*array,
169 	xfarray_idx_t	idx)
170 {
171 	void		*temp = xfarray_scratch(array);
172 	loff_t		pos = xfarray_pos(array, idx);
173 	int		error;
174 
175 	if (idx >= array->nr)
176 		return -ENODATA;
177 
178 	if (idx == array->nr - 1) {
179 		array->nr--;
180 		return 0;
181 	}
182 
183 	if (xfarray_is_unset(array, pos))
184 		return 0;
185 
186 	memset(temp, 0, array->obj_size);
187 	error = xfile_obj_store(array->xfile, temp, array->obj_size, pos);
188 	if (error)
189 		return error;
190 
191 	array->unset_slots++;
192 	return 0;
193 }
194 
195 /*
196  * Store an element in the array.  The element must not be completely zeroed,
197  * because those are considered unset sparse elements.
198  */
199 int
200 xfarray_store(
201 	struct xfarray	*array,
202 	xfarray_idx_t	idx,
203 	const void	*ptr)
204 {
205 	int		ret;
206 
207 	if (idx >= array->max_nr)
208 		return -EFBIG;
209 
210 	ASSERT(!xfarray_element_is_null(array, ptr));
211 
212 	ret = xfile_obj_store(array->xfile, ptr, array->obj_size,
213 			xfarray_pos(array, idx));
214 	if (ret)
215 		return ret;
216 
217 	array->nr = max(array->nr, idx + 1);
218 	return 0;
219 }
220 
221 /* Is this array element NULL? */
222 bool
223 xfarray_element_is_null(
224 	struct xfarray	*array,
225 	const void	*ptr)
226 {
227 	return !memchr_inv(ptr, 0, array->obj_size);
228 }
229 
230 /*
231  * Store an element anywhere in the array that is unset.  If there are no
232  * unset slots, append the element to the array.
233  */
234 int
235 xfarray_store_anywhere(
236 	struct xfarray	*array,
237 	const void	*ptr)
238 {
239 	void		*temp = xfarray_scratch(array);
240 	loff_t		endpos = xfarray_pos(array, array->nr);
241 	loff_t		pos;
242 	int		error;
243 
244 	/* Find an unset slot to put it in. */
245 	for (pos = 0;
246 	     pos < endpos && array->unset_slots > 0;
247 	     pos += array->obj_size) {
248 		error = xfile_obj_load(array->xfile, temp, array->obj_size,
249 				pos);
250 		if (error || !xfarray_element_is_null(array, temp))
251 			continue;
252 
253 		error = xfile_obj_store(array->xfile, ptr, array->obj_size,
254 				pos);
255 		if (error)
256 			return error;
257 
258 		array->unset_slots--;
259 		return 0;
260 	}
261 
262 	/* No unset slots found; attach it on the end. */
263 	array->unset_slots = 0;
264 	return xfarray_append(array, ptr);
265 }
266 
267 /* Return length of array. */
268 uint64_t
269 xfarray_length(
270 	struct xfarray	*array)
271 {
272 	return array->nr;
273 }
274 
275 /*
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.
279  *
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
283  * necessary.
284  */
285 static inline int
286 xfarray_find_data(
287 	struct xfarray	*array,
288 	xfarray_idx_t	*cur,
289 	loff_t		*pos)
290 {
291 	unsigned int	pgoff = offset_in_page(*pos);
292 	loff_t		end_pos = *pos + array->obj_size - 1;
293 	loff_t		new_pos;
294 
295 	/*
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.
298 	 */
299 	if (pgoff != 0 && pgoff + array->obj_size - 1 < PAGE_SIZE)
300 		return 0;
301 
302 	/*
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.
309 	 *
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.
312 	 */
313 	new_pos = xfile_seek_data(array->xfile, end_pos);
314 	if (new_pos == -ENXIO)
315 		return -ENODATA;
316 	if (new_pos < 0)
317 		return new_pos;
318 	if (new_pos == end_pos)
319 		return 0;
320 
321 	/*
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.
325 	 */
326 	new_pos = roundup_64(new_pos, array->obj_size);
327 	*cur = xfarray_idx(array, new_pos);
328 	*pos = xfarray_pos(array, *cur);
329 	return 0;
330 }
331 
332 /*
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.
337  */
338 int
339 xfarray_load_next(
340 	struct xfarray	*array,
341 	xfarray_idx_t	*idx,
342 	void		*rec)
343 {
344 	xfarray_idx_t	cur = *idx;
345 	loff_t		pos = xfarray_pos(array, cur);
346 	int		error;
347 
348 	do {
349 		if (cur >= array->nr)
350 			return -ENODATA;
351 
352 		/*
353 		 * Ask the backing store for the location of next possible
354 		 * written record, then retrieve that record.
355 		 */
356 		error = xfarray_find_data(array, &cur, &pos);
357 		if (error)
358 			return error;
359 		error = xfarray_load(array, cur, rec);
360 		if (error)
361 			return error;
362 
363 		cur++;
364 		pos += array->obj_size;
365 	} while (xfarray_element_is_null(array, rec));
366 
367 	*idx = cur;
368 	return 0;
369 }
370 
371 /* Sorting functions */
372 
373 #ifdef DEBUG
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)
377 # define xfarray_sort_bump_heapsorts(si) do { (si)->heapsorts++; } while (0)
378 #else
379 # define xfarray_sort_bump_loads(si)
380 # define xfarray_sort_bump_stores(si)
381 # define xfarray_sort_bump_compares(si)
382 # define xfarray_sort_bump_heapsorts(si)
383 #endif /* DEBUG */
384 
385 /* Load an array element for sorting. */
386 static inline int
387 xfarray_sort_load(
388 	struct xfarray_sortinfo	*si,
389 	xfarray_idx_t		idx,
390 	void			*ptr)
391 {
392 	xfarray_sort_bump_loads(si);
393 	return xfarray_load(si->array, idx, ptr);
394 }
395 
396 /* Store an array element for sorting. */
397 static inline int
398 xfarray_sort_store(
399 	struct xfarray_sortinfo	*si,
400 	xfarray_idx_t		idx,
401 	void			*ptr)
402 {
403 	xfarray_sort_bump_stores(si);
404 	return xfarray_store(si->array, idx, ptr);
405 }
406 
407 /* Compare an array element for sorting. */
408 static inline int
409 xfarray_sort_cmp(
410 	struct xfarray_sortinfo	*si,
411 	const void		*a,
412 	const void		*b)
413 {
414 	xfarray_sort_bump_compares(si);
415 	return si->cmp_fn(a, b);
416 }
417 
418 /* Return a pointer to the low index stack for quicksort partitioning. */
419 static inline xfarray_idx_t *xfarray_sortinfo_lo(struct xfarray_sortinfo *si)
420 {
421 	return (xfarray_idx_t *)(si + 1);
422 }
423 
424 /* Return a pointer to the high index stack for quicksort partitioning. */
425 static inline xfarray_idx_t *xfarray_sortinfo_hi(struct xfarray_sortinfo *si)
426 {
427 	return xfarray_sortinfo_lo(si) + si->max_stack_depth;
428 }
429 
430 /* Size of each element in the quicksort pivot array. */
431 static inline size_t
432 xfarray_pivot_rec_sz(
433 	struct xfarray		*array)
434 {
435 	return round_up(array->obj_size, 8) + sizeof(xfarray_idx_t);
436 }
437 
438 /* Allocate memory to handle the sort. */
439 static inline int
440 xfarray_sortinfo_alloc(
441 	struct xfarray		*array,
442 	xfarray_cmp_fn		cmp_fn,
443 	unsigned int		flags,
444 	struct xfarray_sortinfo	**infop)
445 {
446 	struct xfarray_sortinfo	*si;
447 	size_t			nr_bytes = sizeof(struct xfarray_sortinfo);
448 	size_t			pivot_rec_sz = xfarray_pivot_rec_sz(array);
449 	int			max_stack_depth;
450 
451 	/*
452 	 * The median-of-nine pivot algorithm doesn't work if a subset has
453 	 * fewer than 9 items.  Make sure the in-memory sort will always take
454 	 * over for subsets where this wouldn't be the case.
455 	 */
456 	BUILD_BUG_ON(XFARRAY_QSORT_PIVOT_NR >= XFARRAY_ISORT_NR);
457 
458 	/*
459 	 * Tail-call recursion during the partitioning phase means that
460 	 * quicksort will never recurse more than log2(nr) times.  We need one
461 	 * extra level of stack to hold the initial parameters.  In-memory
462 	 * sort will always take care of the last few levels of recursion for
463 	 * us, so we can reduce the stack depth by that much.
464 	 */
465 	max_stack_depth = ilog2(array->nr) + 1 - (XFARRAY_ISORT_SHIFT - 1);
466 	if (max_stack_depth < 1)
467 		max_stack_depth = 1;
468 
469 	/* Each level of quicksort uses a lo and a hi index */
470 	nr_bytes += max_stack_depth * sizeof(xfarray_idx_t) * 2;
471 
472 	/* Scratchpad for in-memory sort, or finding the pivot */
473 	nr_bytes += max_t(size_t,
474 			(XFARRAY_QSORT_PIVOT_NR + 1) * pivot_rec_sz,
475 			XFARRAY_ISORT_NR * array->obj_size);
476 
477 	si = kvzalloc(nr_bytes, XCHK_GFP_FLAGS);
478 	if (!si)
479 		return -ENOMEM;
480 
481 	si->array = array;
482 	si->cmp_fn = cmp_fn;
483 	si->flags = flags;
484 	si->max_stack_depth = max_stack_depth;
485 	si->max_stack_used = 1;
486 
487 	xfarray_sortinfo_lo(si)[0] = 0;
488 	xfarray_sortinfo_hi(si)[0] = array->nr - 1;
489 
490 	trace_xfarray_sort(si, nr_bytes);
491 	*infop = si;
492 	return 0;
493 }
494 
495 /* Should this sort be terminated by a fatal signal? */
496 static inline bool
497 xfarray_sort_terminated(
498 	struct xfarray_sortinfo	*si,
499 	int			*error)
500 {
501 	/*
502 	 * If preemption is disabled, we need to yield to the scheduler every
503 	 * few seconds so that we don't run afoul of the soft lockup watchdog
504 	 * or RCU stall detector.
505 	 */
506 	cond_resched();
507 
508 	if ((si->flags & XFARRAY_SORT_KILLABLE) &&
509 	    fatal_signal_pending(current)) {
510 		if (*error == 0)
511 			*error = -EINTR;
512 		return true;
513 	}
514 	return false;
515 }
516 
517 /* Do we want an in-memory sort? */
518 static inline bool
519 xfarray_want_isort(
520 	struct xfarray_sortinfo *si,
521 	xfarray_idx_t		start,
522 	xfarray_idx_t		end)
523 {
524 	/*
525 	 * For array subsets that fit in the scratchpad, it's much faster to
526 	 * use the kernel's heapsort than quicksort's stack machine.
527 	 */
528 	return (end - start) < XFARRAY_ISORT_NR;
529 }
530 
531 /* Return the scratch space within the sortinfo structure. */
532 static inline void *xfarray_sortinfo_isort_scratch(struct xfarray_sortinfo *si)
533 {
534 	return xfarray_sortinfo_hi(si) + si->max_stack_depth;
535 }
536 
537 /*
538  * Sort a small number of array records using scratchpad memory.  The records
539  * need not be contiguous in the xfile's memory pages.
540  */
541 STATIC int
542 xfarray_isort(
543 	struct xfarray_sortinfo	*si,
544 	xfarray_idx_t		lo,
545 	xfarray_idx_t		hi)
546 {
547 	void			*scratch = xfarray_sortinfo_isort_scratch(si);
548 	loff_t			lo_pos = xfarray_pos(si->array, lo);
549 	loff_t			len = xfarray_pos(si->array, hi - lo + 1);
550 	int			error;
551 
552 	trace_xfarray_isort(si, lo, hi);
553 
554 	xfarray_sort_bump_loads(si);
555 	error = xfile_obj_load(si->array->xfile, scratch, len, lo_pos);
556 	if (error)
557 		return error;
558 
559 	xfarray_sort_bump_heapsorts(si);
560 	sort(scratch, hi - lo + 1, si->array->obj_size, si->cmp_fn, NULL);
561 
562 	xfarray_sort_bump_stores(si);
563 	return xfile_obj_store(si->array->xfile, scratch, len, lo_pos);
564 }
565 
566 /* Grab a page for sorting records. */
567 static inline int
568 xfarray_sort_get_page(
569 	struct xfarray_sortinfo	*si,
570 	loff_t			pos,
571 	uint64_t		len)
572 {
573 	int			error;
574 
575 	error = xfile_get_page(si->array->xfile, pos, len, &si->xfpage);
576 	if (error)
577 		return error;
578 
579 	/*
580 	 * xfile pages must never be mapped into userspace, so we skip the
581 	 * dcache flush when mapping the page.
582 	 */
583 	si->page_kaddr = kmap_local_page(si->xfpage.page);
584 	return 0;
585 }
586 
587 /* Release a page we grabbed for sorting records. */
588 static inline int
589 xfarray_sort_put_page(
590 	struct xfarray_sortinfo	*si)
591 {
592 	if (!si->page_kaddr)
593 		return 0;
594 
595 	kunmap_local(si->page_kaddr);
596 	si->page_kaddr = NULL;
597 
598 	return xfile_put_page(si->array->xfile, &si->xfpage);
599 }
600 
601 /* Decide if these records are eligible for in-page sorting. */
602 static inline bool
603 xfarray_want_pagesort(
604 	struct xfarray_sortinfo	*si,
605 	xfarray_idx_t		lo,
606 	xfarray_idx_t		hi)
607 {
608 	pgoff_t			lo_page;
609 	pgoff_t			hi_page;
610 	loff_t			end_pos;
611 
612 	/* We can only map one page at a time. */
613 	lo_page = xfarray_pos(si->array, lo) >> PAGE_SHIFT;
614 	end_pos = xfarray_pos(si->array, hi) + si->array->obj_size - 1;
615 	hi_page = end_pos >> PAGE_SHIFT;
616 
617 	return lo_page == hi_page;
618 }
619 
620 /* Sort a bunch of records that all live in the same memory page. */
621 STATIC int
622 xfarray_pagesort(
623 	struct xfarray_sortinfo	*si,
624 	xfarray_idx_t		lo,
625 	xfarray_idx_t		hi)
626 {
627 	void			*startp;
628 	loff_t			lo_pos = xfarray_pos(si->array, lo);
629 	uint64_t		len = xfarray_pos(si->array, hi - lo);
630 	int			error = 0;
631 
632 	trace_xfarray_pagesort(si, lo, hi);
633 
634 	xfarray_sort_bump_loads(si);
635 	error = xfarray_sort_get_page(si, lo_pos, len);
636 	if (error)
637 		return error;
638 
639 	xfarray_sort_bump_heapsorts(si);
640 	startp = si->page_kaddr + offset_in_page(lo_pos);
641 	sort(startp, hi - lo + 1, si->array->obj_size, si->cmp_fn, NULL);
642 
643 	xfarray_sort_bump_stores(si);
644 	return xfarray_sort_put_page(si);
645 }
646 
647 /* Return a pointer to the xfarray pivot record within the sortinfo struct. */
648 static inline void *xfarray_sortinfo_pivot(struct xfarray_sortinfo *si)
649 {
650 	return xfarray_sortinfo_hi(si) + si->max_stack_depth;
651 }
652 
653 /* Return a pointer to the start of the pivot array. */
654 static inline void *
655 xfarray_sortinfo_pivot_array(
656 	struct xfarray_sortinfo	*si)
657 {
658 	return xfarray_sortinfo_pivot(si) + si->array->obj_size;
659 }
660 
661 /* The xfarray record is stored at the start of each pivot array element. */
662 static inline void *
663 xfarray_pivot_array_rec(
664 	void			*pa,
665 	size_t			pa_recsz,
666 	unsigned int		pa_idx)
667 {
668 	return pa + (pa_recsz * pa_idx);
669 }
670 
671 /* The xfarray index is stored at the end of each pivot array element. */
672 static inline xfarray_idx_t *
673 xfarray_pivot_array_idx(
674 	void			*pa,
675 	size_t			pa_recsz,
676 	unsigned int		pa_idx)
677 {
678 	return xfarray_pivot_array_rec(pa, pa_recsz, pa_idx + 1) -
679 			sizeof(xfarray_idx_t);
680 }
681 
682 /*
683  * Find a pivot value for quicksort partitioning, swap it with a[lo], and save
684  * the cached pivot record for the next step.
685  *
686  * Load evenly-spaced records within the given range into memory, sort them,
687  * and choose the pivot from the median record.  Using multiple points will
688  * improve the quality of the pivot selection, and hopefully avoid the worst
689  * quicksort behavior, since our array values are nearly always evenly sorted.
690  */
691 STATIC int
692 xfarray_qsort_pivot(
693 	struct xfarray_sortinfo	*si,
694 	xfarray_idx_t		lo,
695 	xfarray_idx_t		hi)
696 {
697 	void			*pivot = xfarray_sortinfo_pivot(si);
698 	void			*parray = xfarray_sortinfo_pivot_array(si);
699 	void			*recp;
700 	xfarray_idx_t		*idxp;
701 	xfarray_idx_t		step = (hi - lo) / (XFARRAY_QSORT_PIVOT_NR - 1);
702 	size_t			pivot_rec_sz = xfarray_pivot_rec_sz(si->array);
703 	int			i, j;
704 	int			error;
705 
706 	ASSERT(step > 0);
707 
708 	/*
709 	 * Load the xfarray indexes of the records we intend to sample into the
710 	 * pivot array.
711 	 */
712 	idxp = xfarray_pivot_array_idx(parray, pivot_rec_sz, 0);
713 	*idxp = lo;
714 	for (i = 1; i < XFARRAY_QSORT_PIVOT_NR - 1; i++) {
715 		idxp = xfarray_pivot_array_idx(parray, pivot_rec_sz, i);
716 		*idxp = lo + (i * step);
717 	}
718 	idxp = xfarray_pivot_array_idx(parray, pivot_rec_sz,
719 			XFARRAY_QSORT_PIVOT_NR - 1);
720 	*idxp = hi;
721 
722 	/* Load the selected xfarray records into the pivot array. */
723 	for (i = 0; i < XFARRAY_QSORT_PIVOT_NR; i++) {
724 		xfarray_idx_t	idx;
725 
726 		recp = xfarray_pivot_array_rec(parray, pivot_rec_sz, i);
727 		idxp = xfarray_pivot_array_idx(parray, pivot_rec_sz, i);
728 
729 		/* No unset records; load directly into the array. */
730 		if (likely(si->array->unset_slots == 0)) {
731 			error = xfarray_sort_load(si, *idxp, recp);
732 			if (error)
733 				return error;
734 			continue;
735 		}
736 
737 		/*
738 		 * Load non-null records into the scratchpad without changing
739 		 * the xfarray_idx_t in the pivot array.
740 		 */
741 		idx = *idxp;
742 		xfarray_sort_bump_loads(si);
743 		error = xfarray_load_next(si->array, &idx, recp);
744 		if (error)
745 			return error;
746 	}
747 
748 	xfarray_sort_bump_heapsorts(si);
749 	sort(parray, XFARRAY_QSORT_PIVOT_NR, pivot_rec_sz, si->cmp_fn, NULL);
750 
751 	/*
752 	 * We sorted the pivot array records (which includes the xfarray
753 	 * indices) in xfarray record order.  The median element of the pivot
754 	 * array contains the xfarray record that we will use as the pivot.
755 	 * Copy that xfarray record to the designated space.
756 	 */
757 	recp = xfarray_pivot_array_rec(parray, pivot_rec_sz,
758 			XFARRAY_QSORT_PIVOT_NR / 2);
759 	memcpy(pivot, recp, si->array->obj_size);
760 
761 	/* If the pivot record we chose was already in a[lo] then we're done. */
762 	idxp = xfarray_pivot_array_idx(parray, pivot_rec_sz,
763 			XFARRAY_QSORT_PIVOT_NR / 2);
764 	if (*idxp == lo)
765 		return 0;
766 
767 	/*
768 	 * Find the cached copy of a[lo] in the pivot array so that we can swap
769 	 * a[lo] and a[pivot].
770 	 */
771 	for (i = 0, j = -1; i < XFARRAY_QSORT_PIVOT_NR; i++) {
772 		idxp = xfarray_pivot_array_idx(parray, pivot_rec_sz, i);
773 		if (*idxp == lo)
774 			j = i;
775 	}
776 	if (j < 0) {
777 		ASSERT(j >= 0);
778 		return -EFSCORRUPTED;
779 	}
780 
781 	/* Swap a[lo] and a[pivot]. */
782 	error = xfarray_sort_store(si, lo, pivot);
783 	if (error)
784 		return error;
785 
786 	recp = xfarray_pivot_array_rec(parray, pivot_rec_sz, j);
787 	idxp = xfarray_pivot_array_idx(parray, pivot_rec_sz,
788 			XFARRAY_QSORT_PIVOT_NR / 2);
789 	return xfarray_sort_store(si, *idxp, recp);
790 }
791 
792 /*
793  * Set up the pointers for the next iteration.  We push onto the stack all of
794  * the unsorted values between a[lo + 1] and a[end[i]], and we tweak the
795  * current stack frame to point to the unsorted values between a[beg[i]] and
796  * a[lo] so that those values will be sorted when we pop the stack.
797  */
798 static inline int
799 xfarray_qsort_push(
800 	struct xfarray_sortinfo	*si,
801 	xfarray_idx_t		*si_lo,
802 	xfarray_idx_t		*si_hi,
803 	xfarray_idx_t		lo,
804 	xfarray_idx_t		hi)
805 {
806 	/* Check for stack overflows */
807 	if (si->stack_depth >= si->max_stack_depth - 1) {
808 		ASSERT(si->stack_depth < si->max_stack_depth - 1);
809 		return -EFSCORRUPTED;
810 	}
811 
812 	si->max_stack_used = max_t(uint8_t, si->max_stack_used,
813 					    si->stack_depth + 2);
814 
815 	si_lo[si->stack_depth + 1] = lo + 1;
816 	si_hi[si->stack_depth + 1] = si_hi[si->stack_depth];
817 	si_hi[si->stack_depth++] = lo - 1;
818 
819 	/*
820 	 * Always start with the smaller of the two partitions to keep the
821 	 * amount of recursion in check.
822 	 */
823 	if (si_hi[si->stack_depth]     - si_lo[si->stack_depth] >
824 	    si_hi[si->stack_depth - 1] - si_lo[si->stack_depth - 1]) {
825 		swap(si_lo[si->stack_depth], si_lo[si->stack_depth - 1]);
826 		swap(si_hi[si->stack_depth], si_hi[si->stack_depth - 1]);
827 	}
828 
829 	return 0;
830 }
831 
832 /*
833  * Load an element from the array into the first scratchpad and cache the page,
834  * if possible.
835  */
836 static inline int
837 xfarray_sort_load_cached(
838 	struct xfarray_sortinfo	*si,
839 	xfarray_idx_t		idx,
840 	void			*ptr)
841 {
842 	loff_t			idx_pos = xfarray_pos(si->array, idx);
843 	pgoff_t			startpage;
844 	pgoff_t			endpage;
845 	int			error = 0;
846 
847 	/*
848 	 * If this load would split a page, release the cached page, if any,
849 	 * and perform a traditional read.
850 	 */
851 	startpage = idx_pos >> PAGE_SHIFT;
852 	endpage = (idx_pos + si->array->obj_size - 1) >> PAGE_SHIFT;
853 	if (startpage != endpage) {
854 		error = xfarray_sort_put_page(si);
855 		if (error)
856 			return error;
857 
858 		if (xfarray_sort_terminated(si, &error))
859 			return error;
860 
861 		return xfile_obj_load(si->array->xfile, ptr,
862 				si->array->obj_size, idx_pos);
863 	}
864 
865 	/* If the cached page is not the one we want, release it. */
866 	if (xfile_page_cached(&si->xfpage) &&
867 	    xfile_page_index(&si->xfpage) != startpage) {
868 		error = xfarray_sort_put_page(si);
869 		if (error)
870 			return error;
871 	}
872 
873 	/*
874 	 * If we don't have a cached page (and we know the load is contained
875 	 * in a single page) then grab it.
876 	 */
877 	if (!xfile_page_cached(&si->xfpage)) {
878 		if (xfarray_sort_terminated(si, &error))
879 			return error;
880 
881 		error = xfarray_sort_get_page(si, startpage << PAGE_SHIFT,
882 				PAGE_SIZE);
883 		if (error)
884 			return error;
885 	}
886 
887 	memcpy(ptr, si->page_kaddr + offset_in_page(idx_pos),
888 			si->array->obj_size);
889 	return 0;
890 }
891 
892 /*
893  * Sort the array elements via quicksort.  This implementation incorporates
894  * four optimizations discussed in Sedgewick:
895  *
896  * 1. Use an explicit stack of array indices to store the next array partition
897  *    to sort.  This helps us to avoid recursion in the call stack, which is
898  *    particularly expensive in the kernel.
899  *
900  * 2. For arrays with records in arbitrary or user-controlled order, choose the
901  *    pivot element using a median-of-nine decision tree.  This reduces the
902  *    probability of selecting a bad pivot value which causes worst case
903  *    behavior (i.e. partition sizes of 1).
904  *
905  * 3. The smaller of the two sub-partitions is pushed onto the stack to start
906  *    the next level of recursion, and the larger sub-partition replaces the
907  *    current stack frame.  This guarantees that we won't need more than
908  *    log2(nr) stack space.
909  *
910  * 4. For small sets, load the records into the scratchpad and run heapsort on
911  *    them because that is very fast.  In the author's experience, this yields
912  *    a ~10% reduction in runtime.
913  *
914  *    If a small set is contained entirely within a single xfile memory page,
915  *    map the page directly and run heap sort directly on the xfile page
916  *    instead of using the load/store interface.  This halves the runtime.
917  *
918  * 5. This optimization is specific to the implementation.  When converging lo
919  *    and hi after selecting a pivot, we will try to retain the xfile memory
920  *    page between load calls, which reduces run time by 50%.
921  */
922 
923 /*
924  * Due to the use of signed indices, we can only support up to 2^63 records.
925  * Files can only grow to 2^63 bytes, so this is not much of a limitation.
926  */
927 #define QSORT_MAX_RECS		(1ULL << 63)
928 
929 int
930 xfarray_sort(
931 	struct xfarray		*array,
932 	xfarray_cmp_fn		cmp_fn,
933 	unsigned int		flags)
934 {
935 	struct xfarray_sortinfo	*si;
936 	xfarray_idx_t		*si_lo, *si_hi;
937 	void			*pivot;
938 	void			*scratch = xfarray_scratch(array);
939 	xfarray_idx_t		lo, hi;
940 	int			error = 0;
941 
942 	if (array->nr < 2)
943 		return 0;
944 	if (array->nr >= QSORT_MAX_RECS)
945 		return -E2BIG;
946 
947 	error = xfarray_sortinfo_alloc(array, cmp_fn, flags, &si);
948 	if (error)
949 		return error;
950 	si_lo = xfarray_sortinfo_lo(si);
951 	si_hi = xfarray_sortinfo_hi(si);
952 	pivot = xfarray_sortinfo_pivot(si);
953 
954 	while (si->stack_depth >= 0) {
955 		lo = si_lo[si->stack_depth];
956 		hi = si_hi[si->stack_depth];
957 
958 		trace_xfarray_qsort(si, lo, hi);
959 
960 		/* Nothing left in this partition to sort; pop stack. */
961 		if (lo >= hi) {
962 			si->stack_depth--;
963 			continue;
964 		}
965 
966 		/*
967 		 * If directly mapping the page and sorting can solve our
968 		 * problems, we're done.
969 		 */
970 		if (xfarray_want_pagesort(si, lo, hi)) {
971 			error = xfarray_pagesort(si, lo, hi);
972 			if (error)
973 				goto out_free;
974 			si->stack_depth--;
975 			continue;
976 		}
977 
978 		/* If insertion sort can solve our problems, we're done. */
979 		if (xfarray_want_isort(si, lo, hi)) {
980 			error = xfarray_isort(si, lo, hi);
981 			if (error)
982 				goto out_free;
983 			si->stack_depth--;
984 			continue;
985 		}
986 
987 		/* Pick a pivot, move it to a[lo] and stash it. */
988 		error = xfarray_qsort_pivot(si, lo, hi);
989 		if (error)
990 			goto out_free;
991 
992 		/*
993 		 * Rearrange a[lo..hi] such that everything smaller than the
994 		 * pivot is on the left side of the range and everything larger
995 		 * than the pivot is on the right side of the range.
996 		 */
997 		while (lo < hi) {
998 			/*
999 			 * Decrement hi until it finds an a[hi] less than the
1000 			 * pivot value.
1001 			 */
1002 			error = xfarray_sort_load_cached(si, hi, scratch);
1003 			if (error)
1004 				goto out_free;
1005 			while (xfarray_sort_cmp(si, scratch, pivot) >= 0 &&
1006 								lo < hi) {
1007 				hi--;
1008 				error = xfarray_sort_load_cached(si, hi,
1009 						scratch);
1010 				if (error)
1011 					goto out_free;
1012 			}
1013 			error = xfarray_sort_put_page(si);
1014 			if (error)
1015 				goto out_free;
1016 
1017 			if (xfarray_sort_terminated(si, &error))
1018 				goto out_free;
1019 
1020 			/* Copy that item (a[hi]) to a[lo]. */
1021 			if (lo < hi) {
1022 				error = xfarray_sort_store(si, lo++, scratch);
1023 				if (error)
1024 					goto out_free;
1025 			}
1026 
1027 			/*
1028 			 * Increment lo until it finds an a[lo] greater than
1029 			 * the pivot value.
1030 			 */
1031 			error = xfarray_sort_load_cached(si, lo, scratch);
1032 			if (error)
1033 				goto out_free;
1034 			while (xfarray_sort_cmp(si, scratch, pivot) <= 0 &&
1035 								lo < hi) {
1036 				lo++;
1037 				error = xfarray_sort_load_cached(si, lo,
1038 						scratch);
1039 				if (error)
1040 					goto out_free;
1041 			}
1042 			error = xfarray_sort_put_page(si);
1043 			if (error)
1044 				goto out_free;
1045 
1046 			if (xfarray_sort_terminated(si, &error))
1047 				goto out_free;
1048 
1049 			/* Copy that item (a[lo]) to a[hi]. */
1050 			if (lo < hi) {
1051 				error = xfarray_sort_store(si, hi--, scratch);
1052 				if (error)
1053 					goto out_free;
1054 			}
1055 
1056 			if (xfarray_sort_terminated(si, &error))
1057 				goto out_free;
1058 		}
1059 
1060 		/*
1061 		 * Put our pivot value in the correct place at a[lo].  All
1062 		 * values between a[beg[i]] and a[lo - 1] should be less than
1063 		 * the pivot; and all values between a[lo + 1] and a[end[i]-1]
1064 		 * should be greater than the pivot.
1065 		 */
1066 		error = xfarray_sort_store(si, lo, pivot);
1067 		if (error)
1068 			goto out_free;
1069 
1070 		/* Set up the stack frame to process the two partitions. */
1071 		error = xfarray_qsort_push(si, si_lo, si_hi, lo, hi);
1072 		if (error)
1073 			goto out_free;
1074 
1075 		if (xfarray_sort_terminated(si, &error))
1076 			goto out_free;
1077 	}
1078 
1079 out_free:
1080 	trace_xfarray_sort_stats(si, error);
1081 	kvfree(si);
1082 	return error;
1083 }
1084