[sheepdog] [PATCH v2] http: add object allocator
Liu Yuan
namei.unix at gmail.com
Fri Nov 29 07:18:49 CET 2013
This implements a per-vdi allocator for http extent-based container.
We use meta object tracks the free information of data vdi for the object
allocation in a free list. Free list is a redundant structure for bitmap for
faster allocation.
future enhancements:
- Add allocation group for scalability and solve the meta size limitation
- use aio to speed up discard of objects
- add range support for inode update
Signed-off-by: Liu Yuan <namei.unix at gmail.com>
---
v2:
- use uint64_t for i,j,k
sheep/Makefile.am | 3 +-
sheep/http/http.h | 6 +
sheep/http/oalloc.c | 306 +++++++++++++++++++++++++++++++++++++++++++++++++++
3 files changed, 314 insertions(+), 1 deletion(-)
create mode 100644 sheep/http/oalloc.c
diff --git a/sheep/Makefile.am b/sheep/Makefile.am
index 552e86a..af1087f 100644
--- a/sheep/Makefile.am
+++ b/sheep/Makefile.am
@@ -30,7 +30,8 @@ sheep_SOURCES = sheep.c group.c request.c gateway.c store.c vdi.c \
plain_store.c config.c migrate.c md.c
if BUILD_HTTP
-sheep_SOURCES += http/http.c http/kv.c http/s3.c http/swift.c
+sheep_SOURCES += http/http.c http/kv.c http/s3.c http/swift.c \
+ http/oalloc.c
endif
if BUILD_COROSYNC
sheep_SOURCES += cluster/corosync.c
diff --git a/sheep/http/http.h b/sheep/http/http.h
index 495a66c..046d412 100644
--- a/sheep/http/http.h
+++ b/sheep/http/http.h
@@ -108,4 +108,10 @@ int http_request_writes(struct http_request *req, const char *str);
__printf(2, 3)
int http_request_writef(struct http_request *req, const char *fmt, ...);
+/* object_allocator.c */
+int oalloc_new_prepare(uint32_t vid, uint64_t *start, uint64_t count);
+int oalloc_new_finish(uint32_t vid, uint64_t start, uint64_t count);
+int oalloc_free(uint32_t vid, uint64_t start, uint64_t count);
+int oalloc_init(uint32_t vid);
+
#endif /* __SHEEP_HTTP_H__ */
diff --git a/sheep/http/oalloc.c b/sheep/http/oalloc.c
new file mode 100644
index 0000000..56362fa
--- /dev/null
+++ b/sheep/http/oalloc.c
@@ -0,0 +1,306 @@
+/*
+ * Copyright (C) 2013 Taobao Inc.
+ *
+ * Liu Yuan <namei.unix at gmail.com>
+ *
+ * This program is free software; you can redistribute it and/or
+ * modify it under the terms of the GNU General Public License version
+ * 2 as published by the Free Software Foundation.
+ *
+ * You should have received a copy of the GNU General Public License
+ * along with this program. If not, see <http://www.gnu.org/licenses/>.
+ */
+
+#include "sheep_priv.h"
+#include "http.h"
+
+/*
+ * Meta Object tracks the free information of data vdi for the object allocation
+ * in a free list. Free list is a redundant structure for bitmap for faster
+ * allocation.
+ *
+ * +-------------------------------+
+ * | |
+ * | sorted list v------v
+ * +--------------------------------+-----------------------+ +--------+
+ * | Header | fd1 | fd2 | ... | fdN | .... object data .... | <-- | bitmap |
+ * +--------------------------------+-----------------------+ +---------
+ * |<-- 4M -->|
+ *
+ * Best-fit algorithm for allocation and merge and sort the free list at
+ * deallocation. One simple sorted list is effecient enough for extent based
+ * invariable user object.
+ *
+ * XXX: Add allocation group for scalability and solve the meta size limitation
+ */
+
+struct header {
+ uint64_t used;
+ uint64_t nr_free;
+};
+
+struct free_desc {
+ uint64_t start;
+ uint64_t count;
+};
+
+static inline uint32_t oalloc_meta_length(struct header *hd)
+{
+ return sizeof(struct header) + sizeof(struct free_desc) * hd->nr_free;
+}
+
+#define HEADER_TO_FREE_DESC(hd) ((struct free_desc *) \
+ ((char *)hd + sizeof(struct header)))
+
+#define MAX_FREE_DESC ((SD_DATA_OBJ_SIZE - sizeof(struct header)) / \
+ sizeof(struct free_desc))
+
+/*
+ * Initialize the data vdi
+ *
+ * @vid: the vdi where the allocator resides
+ */
+int oalloc_init(uint32_t vid)
+{
+ struct strbuf buf = STRBUF_INIT;
+ struct sd_inode *inode = xmalloc(sizeof(struct sd_inode));
+ struct header hd = {
+ .nr_free = 1,
+ };
+ struct free_desc fd = {
+ .start = 1, /* Use first object as the meta object */
+ .count = MAX_DATA_OBJS - 1,
+ };
+ int ret;
+
+ strbuf_add(&buf, &hd, sizeof(hd));
+ strbuf_add(&buf, &fd, sizeof(fd));
+
+ ret = sd_read_object(vid_to_vdi_oid(vid), (char *)inode,
+ sizeof(*inode), 0);
+ if (ret != SD_RES_SUCCESS) {
+ sd_err("failed to read inode, %" PRIx32", %s", vid,
+ sd_strerror(ret));
+ goto out;
+ }
+ ret = sd_write_object(vid_to_data_oid(vid, 0), buf.buf,
+ buf.len, 0, true);
+ if (ret != SD_RES_SUCCESS) {
+ sd_err("failed to create meta object for %" PRIx32", %s", vid,
+ sd_strerror(ret));
+ goto out;
+ }
+ INODE_SET_VID(inode, 0, vid);
+ ret = sd_inode_write_vid(sheep_bnode_writer, inode, 0,
+ vid, vid, 0, false, false);
+ if (ret != SD_RES_SUCCESS) {
+ sd_err("failed to update inode, %" PRIx32", %s", vid,
+ sd_strerror(ret));
+ goto out;
+ }
+out:
+ strbuf_release(&buf);
+ free(inode);
+ return ret;
+}
+
+/*
+ * Allocate the objects and upate the free list.
+ *
+ * Callers are expected to call oalloc_new_finish() to update the inode bitmap
+ * after filling up the data.
+ *
+ * @vid: the vdi where the allocator resides
+ * @start: start index of the objects to allocate
+ * @count: number of the objects to allocate
+ */
+int oalloc_new_prepare(uint32_t vid, uint64_t *start, uint64_t count)
+{
+ char *meta = xvalloc(SD_DATA_OBJ_SIZE);
+ struct header *hd;
+ struct free_desc *fd;
+ uint64_t oid = vid_to_data_oid(vid, 0), i;
+ int ret;
+
+ ret = sd_read_object(oid, meta, SD_DATA_OBJ_SIZE, 0);
+ if (ret != SD_RES_SUCCESS) {
+ sd_err("failed to read meta %" PRIx64 ", %s", oid,
+ sd_strerror(ret));
+ goto out;
+ }
+
+ hd = (struct header *)meta;
+ fd = (struct free_desc *)(meta + oalloc_meta_length(hd)) - 1;
+ sd_debug("used %"PRIu64", nr_free %"PRIu64, hd->used, hd->nr_free);
+ for (i = 0; i < hd->nr_free; i++, fd--) {
+ sd_debug("start %"PRIu64", count %"PRIu64, fd->start,
+ fd->count);
+ if (fd->count > count)
+ break;
+ }
+ if (i == hd->nr_free) {
+ ret = SD_RES_NO_SPACE;
+ goto out;
+ }
+
+ *start = fd->start;
+ fd->start += count;
+ fd->count -= count;
+ hd->used += count;
+
+ /* Update the meta object */
+ ret = sd_write_object(oid, meta, oalloc_meta_length(hd), 0, false);
+ if (ret != SD_RES_SUCCESS)
+ sd_err("failed to update meta %"PRIx64 ", %s", oid,
+ sd_strerror(ret));
+out:
+ free(meta);
+ return ret;
+}
+
+/*
+ * Update the inode map of the vid
+ *
+ * @vid: the vdi where the allocator resides
+ * @start: start index of the objects to update
+ * @count: number of the objects to update
+ */
+int oalloc_new_finish(uint32_t vid, uint64_t start, uint64_t count)
+{
+ struct sd_inode *inode = xmalloc(sizeof(struct sd_inode));
+ int ret;
+
+ ret = sd_read_object(vid_to_vdi_oid(vid), (char *)inode,
+ sizeof(*inode), 0);
+ if (ret != SD_RES_SUCCESS) {
+ sd_err("failed to read inode, %" PRIx64 ", %s",
+ vid_to_vdi_oid(vid), sd_strerror(ret));
+ goto out;
+ }
+
+ /* TODO: add range support for inode update */
+ for (uint64_t i = 0; i < count; i++) {
+ INODE_SET_VID(inode, start + i, vid);
+
+ ret = sd_inode_write_vid(sheep_bnode_writer, inode, start + i,
+ vid, vid, 0, false, false);
+ if (ret != SD_RES_SUCCESS) {
+ sd_err("failed to update inode, %" PRIx64", %s",
+ vid_to_vdi_oid(vid), sd_strerror(ret));
+ goto out;
+ }
+ }
+out:
+ free(inode);
+ return ret;
+}
+
+static int free_desc_cmp(struct free_desc *a, struct free_desc *b)
+{
+ return -intcmp(a->start, b->start);
+}
+
+static inline int update_and_merge_free_desc(char *meta, uint64_t start,
+ uint64_t count, uint32_t vid)
+{
+ struct header *hd = (struct header *)meta;
+ struct free_desc *tail, *fd = HEADER_TO_FREE_DESC(hd);
+ uint64_t i, j;
+
+ /* Try our best to merge it in place, or append it to tail */
+ for (i = 0; i < hd->nr_free; i++) {
+ if (start + count == fd->start) {
+ fd->start = start;
+ fd->count += count;
+ break;
+ } else if(fd->start + fd->count == start) {
+ fd->count +=count;
+ break;
+ }
+ fd++;
+ }
+
+ if (i == hd->nr_free) {
+ if (hd->nr_free >= MAX_FREE_DESC)
+ return SD_RES_NO_SPACE;
+
+ tail = (struct free_desc *)(meta + oalloc_meta_length(hd));
+ tail->start = start;
+ tail->count = count;
+ hd->nr_free++;
+ }
+
+ hd->used -= count;
+ xqsort(HEADER_TO_FREE_DESC(hd), hd->nr_free, free_desc_cmp);
+
+ /* Merge as hard as we can */
+ j = hd->nr_free - 1;
+ tail = (struct free_desc *)(meta + oalloc_meta_length(hd)) - 1;
+ for (i = 0; i < j; i++, tail--) {
+ struct free_desc *front = tail - 1;
+
+ sd_debug("start %"PRIu64", count %"PRIu64, tail->start,
+ tail->count);
+ if (tail->start + tail->count > front->start)
+ sd_emerg("bad free descriptor found at %"PRIx32, vid);
+ if (tail->start + tail->count == front->start) {
+ front->start = tail->start;
+ front->count += tail->count;
+ memmove(tail, tail + 1, sizeof(*tail) * i);
+ hd->nr_free--;
+ }
+ }
+
+ return SD_RES_SUCCESS;
+}
+
+/*
+ * Discard the allocted objects and update the free list of the allocator
+ *
+ * Caller should check the return value since it might fail.
+ *
+ * @vid: the vdi where the allocator resides
+ * @start: start index of the objects to free
+ * @count: number of the objects to free
+ */
+int oalloc_free(uint32_t vid, uint64_t start, uint64_t count)
+{
+ char *meta = xvalloc(SD_DATA_OBJ_SIZE);
+ struct header *hd;
+ uint64_t oid = vid_to_data_oid(vid, 0), i;
+ int ret;
+
+ ret = sd_read_object(oid, meta, SD_DATA_OBJ_SIZE, 0);
+ if (ret != SD_RES_SUCCESS) {
+ sd_err("failed to read meta %" PRIx64 ", %s", oid,
+ sd_strerror(ret));
+ goto out;
+ }
+
+ ret = update_and_merge_free_desc(meta, start, count, vid);
+ if (ret != SD_RES_SUCCESS)
+ goto out;
+
+ /* XXX use aio to speed up discard of objects */
+ for (i = 0; i < count; i++) {
+ struct sd_req hdr;
+
+ sd_init_req(&hdr, SD_OP_DISCARD_OBJ);
+ hdr.obj.oid = vid_to_data_oid(vid, start + i);
+ ret = exec_local_req(&hdr, NULL);
+ if (ret != SD_RES_SUCCESS)
+ goto out;
+ }
+
+ hd = (struct header *)meta;
+ ret = sd_write_object(oid, meta, oalloc_meta_length(hd), 0, false);
+ if (ret != SD_RES_SUCCESS) {
+ sd_err("failed to update meta %"PRIx64 ", %s", oid,
+ sd_strerror(ret));
+ goto out;
+ }
+ sd_debug("used %"PRIu64", nr_free %"PRIu64, hd->used, hd->nr_free);
+out:
+ free(meta);
+ return ret;
+}
--
1.7.9.5
More information about the sheepdog
mailing list