@@ -33,7 +33,6 @@
#include <linux/bit_spinlock.h>
#include <linux/slab.h>
#include <linux/sched/mm.h>
-#include <linux/sort.h>
#include <linux/log2.h>
#include "ctree.h"
#include "disk-io.h"
@@ -752,6 +751,8 @@ struct heuristic_ws {
u32 sample_size;
/* Buckets store counters for each byte value */
struct bucket_item *bucket;
+ /* Sorting buffer */
+ struct bucket_item *bucket_b;
struct list_head list;
};
@@ -763,6 +764,7 @@ static void free_heuristic_ws(struct list_head *ws)
kvfree(workspace->sample);
kfree(workspace->bucket);
+ kfree(workspace->bucket_b);
kfree(workspace);
}
@@ -782,6 +784,10 @@ static struct list_head *alloc_heuristic_ws(void)
if (!ws->bucket)
goto fail;
+ ws->bucket_b = kcalloc(BUCKET_SIZE, sizeof(*ws->bucket_b), GFP_KERNEL);
+ if (!ws->bucket_b)
+ goto fail;
+
INIT_LIST_HEAD(&ws->list);
return &ws->list;
fail:
@@ -1278,13 +1284,127 @@ static u32 shannon_entropy(struct heuristic_ws *ws)
return entropy_sum * 100 / entropy_max;
}
-/* Compare buckets by size, ascending */
-static int bucket_comp_rev(const void *lv, const void *rv)
+#define RADIX_BASE 4
+#define COUNTERS_SIZE (1 << RADIX_BASE)
+
+static inline u8 get4bits(u64 num, u32 shift) {
+ u8 low4bits;
+ num = num >> shift;
+ /* Reverse order */
+ low4bits = (COUNTERS_SIZE - 1) - (num % COUNTERS_SIZE);
+ return low4bits;
+}
+
+static inline void copy_cell(void *dst, u32 dest_i, void *src, u32 src_i)
+{
+ struct bucket_item *dstv = (struct bucket_item *) dst;
+ struct bucket_item *srcv = (struct bucket_item *) src;
+ dstv[dest_i] = srcv[src_i];
+}
+
+static inline u64 get_num(const void *a, u32 i)
+{
+ struct bucket_item *av = (struct bucket_item *) a;
+ return av[i].count;
+}
+
+/*
+ * Use 4 bits as radix base
+ * Use 16 u32 counters for calculating new possition in buf array
+ *
+ * @array - array that will be sorted
+ * @array_buf - buffer array to store sorting results
+ * must be equal in size to @array
+ * @num - array size
+ * @get_num - function to extract number from array
+ * @copy_cell - function to copy data from array to array_buf
+ * and vise versa
+ * @get4bits - function to get 4 bits from number at specified offset
+ */
+
+static void radix_sort(void *array, void *array_buf, u32 num,
+ u64 (*get_num)(const void *, u32 i),
+ void (*copy_cell)(void *dest, u32 dest_i,
+ void* src, u32 src_i),
+ u8 (*get4bits)(u64 num, u32 shift))
{
- const struct bucket_item *l = (const struct bucket_item *)lv;
- const struct bucket_item *r = (const struct bucket_item *)rv;
+ u64 max_num;
+ u64 buf_num;
+ u32 counters[COUNTERS_SIZE];
+ u32 new_addr;
+ u32 addr;
+ u32 bitlen;
+ u32 shift;
+ int i;
+
+ /*
+ * Try avoid useless loop iterations
+ * For small numbers stored in big counters
+ * example: 48 33 4 ... in 64bit array
+ */
+ max_num = get_num(array, 0);
+ for (i = 1; i < num; i++) {
+ buf_num = get_num(array, i);
+ if (buf_num > max_num)
+ max_num = buf_num;
+ }
- return r->count - l->count;
+ buf_num = ilog2(max_num);
+ bitlen = ALIGN(buf_num, RADIX_BASE*2);
+
+ shift = 0;
+ while (shift < bitlen) {
+ memset(counters, 0, sizeof(counters));
+
+ for (i = 0; i < num; i++) {
+ buf_num = get_num(array, i);
+ addr = get4bits(buf_num, shift);
+ counters[addr]++;
+ }
+
+ for (i = 1; i < COUNTERS_SIZE; i++) {
+ counters[i] += counters[i-1];
+ }
+
+ for (i = (num - 1); i >= 0; i --) {
+ buf_num = get_num(array, i);
+ addr = get4bits(buf_num, shift);
+ counters[addr]--;
+ new_addr = counters[addr];
+ copy_cell(array_buf, new_addr, array, i);
+ }
+
+ shift += RADIX_BASE;
+
+ /*
+ * For normal radix, that expected to
+ * move data from tmp array, to main.
+ * But that require some CPU time
+ * Avoid that by doing another sort iteration
+ * to origin array instead of memcpy()
+ */
+ memset(counters, 0, sizeof(counters));
+
+ for (i = 0; i < num; i ++) {
+ buf_num = get_num(array_buf, i);
+ addr = get4bits(buf_num, shift);
+ counters[addr]++;
+ }
+
+ for (i = 1; i < COUNTERS_SIZE; i++) {
+ counters[i] += counters[i-1];
+ }
+
+ for (i = (num - 1); i >= 0; i--) {
+ buf_num = get_num(array_buf, i);
+ addr = get4bits(buf_num, shift);
+ counters[addr]--;
+ new_addr = counters[addr];
+ copy_cell(array, new_addr, array_buf, i);
+ }
+
+ shift += RADIX_BASE;
+ }
}
/*
@@ -1314,7 +1434,8 @@ static int byte_core_set_size(struct heuristic_ws *ws)
struct bucket_item *bucket = ws->bucket;
/* Sort in reverse order */
- sort(bucket, BUCKET_SIZE, sizeof(*bucket), &bucket_comp_rev, NULL);
+ radix_sort(ws->bucket, ws->bucket_b,
+ BUCKET_SIZE, get_num, copy_cell, get4bits);
for (i = 0; i < BYTE_CORE_SET_LOW; i++)
coreset_sum += bucket[i].count;
Slowest part of heuristic for now is kernel heap sort() It's can take up to 55% of runtime on sorting bucket items. As sorting will always call on most data sets to get correctly byte_core_set_size, the only way to speed up heuristic, is to speed up sort on bucket. Add a general radix_sort function. Radix sort require 2 buffers, one full size of input array and one for store counters (jump addresses). That increase usage per heuristic workspace +1KiB 8KiB + 1KiB -> 8KiB + 2KiB That is LSD Radix, i use 4 bit as a base for calculating, to make counters array acceptable small (16 elements * 8 byte). That Radix sort implementation have several points to adjust, I added him to make radix sort general usable in kernel, like heap sort, if needed. Performance tested in userspace copy of heuristic code, throughput: - average <-> random data: ~3500 MiB/s - heap sort - average <-> random data: ~6000 MiB/s - radix sort Changes: v1 -> v2: - Tested on Big Endian - Drop most of multiply operations - Separately allocate sort buffer Changes: v2 -> v3: - Fix uint -> u conversion - Reduce stack size, by reduce vars sizes to u32, restrict input array size to u32 Assume that kernel will never try sorting arrays > 2^32 - Drop max_cell arg (precheck - correctly find max value by it self) Signed-off-by: Timofey Titovets <nefelim4ag@gmail.com> --- fs/btrfs/compression.c | 135 ++++++++++++++++++++++++++++++++++++++++++++++--- 1 file changed, 128 insertions(+), 7 deletions(-)