* Copyright (c) 2002, 2010, Oracle and/or its affiliates. All rights reserved.
* Use is subject to license terms.
*
- * Copyright (c) 2011, 2016, Intel Corporation.
+ * Copyright (c) 2011, 2017, Intel Corporation.
*/
/*
* This file is part of Lustre, http://www.lustre.org/
#include <asm/uaccess.h>
#include <linux/fs.h>
+#include <linux/file.h>
#include <linux/stat.h>
#include <asm/uaccess.h>
#include <linux/mm.h>
#include <linux/pagemap.h>
/* current_is_kswapd() */
#include <linux/swap.h>
+#include <linux/task_io_accounting_ops.h>
#define DEBUG_SUBSYSTEM S_LLITE
static unsigned long ll_ra_count_get(struct ll_sb_info *sbi,
struct ra_io_arg *ria,
- unsigned long pages, unsigned long min)
+ unsigned long pages,
+ unsigned long pages_min)
{
struct ll_ra_info *ra = &sbi->ll_ra_info;
long ret;
}
out:
- if (ret < min) {
+ if (ret < pages_min) {
/* override ra limit for maximum performance */
- atomic_add(min - ret, &ra->ra_cur_pages);
- ret = min;
+ atomic_add(pages_min - ret, &ra->ra_cur_pages);
+ ret = pages_min;
}
RETURN(ret);
}
-void ll_ra_count_put(struct ll_sb_info *sbi, unsigned long len)
+void ll_ra_count_put(struct ll_sb_info *sbi, unsigned long pages)
{
struct ll_ra_info *ra = &sbi->ll_ra_info;
- atomic_sub(len, &ra->ra_cur_pages);
+ atomic_sub(pages, &ra->ra_cur_pages);
}
static void ll_ra_stats_inc_sbi(struct ll_sb_info *sbi, enum ra_stat which)
}
#define RAS_CDEBUG(ras) \
- CDEBUG(D_READA, \
- "lrp %lu cr %lu cp %lu ws %lu wl %lu nra %lu rpc %lu " \
- "r %lu ri %lu csr %lu sf %lu sp %lu sl %lu\n", \
- ras->ras_last_readpage, ras->ras_consecutive_requests, \
- ras->ras_consecutive_pages, ras->ras_window_start, \
- ras->ras_window_len, ras->ras_next_readahead, \
- ras->ras_rpc_size, \
- ras->ras_requests, ras->ras_request_index, \
+ CDEBUG(D_READA, \
+ "lre %llu cr %lu cb %llu wsi %lu wp %lu nra %lu rpc %lu " \
+ "r %lu csr %lu so %llu sb %llu sl %llu lr %lu\n", \
+ ras->ras_last_read_end_bytes, ras->ras_consecutive_requests, \
+ ras->ras_consecutive_bytes, ras->ras_window_start_idx, \
+ ras->ras_window_pages, ras->ras_next_readahead_idx, \
+ ras->ras_rpc_pages, ras->ras_requests, \
ras->ras_consecutive_stride_requests, ras->ras_stride_offset, \
- ras->ras_stride_pages, ras->ras_stride_length)
+ ras->ras_stride_bytes, ras->ras_stride_length, \
+ ras->ras_async_last_readpage_idx)
-static int index_in_window(unsigned long index, unsigned long point,
- unsigned long before, unsigned long after)
+static bool pos_in_window(loff_t pos, loff_t point,
+ unsigned long before, unsigned long after)
{
- unsigned long start = point - before, end = point + after;
+ loff_t start = point - before;
+ loff_t end = point + after;
- if (start > point)
- start = 0;
- if (end < point)
- end = ~0;
+ if (start > point)
+ start = 0;
+ if (end < point)
+ end = ~0;
- return start <= index && index <= end;
-}
-
-void ll_ras_enter(struct file *f)
-{
- struct ll_file_data *fd = LUSTRE_FPRIVATE(f);
- struct ll_readahead_state *ras = &fd->fd_ras;
-
- spin_lock(&ras->ras_lock);
- ras->ras_requests++;
- ras->ras_request_index = 0;
- ras->ras_consecutive_requests++;
- spin_unlock(&ras->ras_lock);
+ return start <= pos && pos <= end;
}
/**
RETURN(rc);
}
-#define RIA_DEBUG(ria) \
- CDEBUG(D_READA, "rs %lu re %lu ro %lu rl %lu rp %lu\n", \
- ria->ria_start, ria->ria_end, ria->ria_stoff, ria->ria_length,\
- ria->ria_pages)
+#define RIA_DEBUG(ria) \
+ CDEBUG(D_READA, "rs %lu re %lu ro %llu rl %llu rb %llu\n", \
+ ria->ria_start_idx, ria->ria_end_idx, ria->ria_stoff, \
+ ria->ria_length, ria->ria_bytes)
static inline int stride_io_mode(struct ll_readahead_state *ras)
{
return ras->ras_consecutive_stride_requests > 1;
}
-/* The function calculates how much pages will be read in
+/* The function calculates how many bytes will be read in
* [off, off + length], in such stride IO area,
* stride_offset = st_off, stride_lengh = st_len,
- * stride_pages = st_pgs
+ * stride_bytes = st_bytes
*
* |------------------|*****|------------------|*****|------------|*****|....
* st_off
- * |--- st_pgs ---|
+ * |--- st_bytes ---|
* |----- st_len -----|
*
- * How many pages it should read in such pattern
+ * How many bytes it should read in such pattern
* |-------------------------------------------------------------|
* off
* |<------ length ------->|
*
* = |<----->| + |-------------------------------------| + |---|
- * start_left st_pgs * i end_left
+ * start_left st_bytes * i end_left
*/
-static unsigned long
-stride_pg_count(pgoff_t st_off, unsigned long st_len, unsigned long st_pgs,
- unsigned long off, unsigned long length)
+static loff_t stride_byte_count(loff_t st_off, loff_t st_len, loff_t st_bytes,
+ loff_t off, loff_t length)
{
- __u64 start = off > st_off ? off - st_off : 0;
- __u64 end = off + length > st_off ? off + length - st_off : 0;
- unsigned long start_left = 0;
- unsigned long end_left = 0;
- unsigned long pg_count;
-
- if (st_len == 0 || length == 0 || end == 0)
- return length;
-
- start_left = do_div(start, st_len);
- if (start_left < st_pgs)
- start_left = st_pgs - start_left;
- else
- start_left = 0;
+ u64 start = off > st_off ? off - st_off : 0;
+ u64 end = off + length > st_off ? off + length - st_off : 0;
+ u64 start_left;
+ u64 end_left;
+ u64 bytes_count;
+
+ if (st_len == 0 || length == 0 || end == 0)
+ return length;
+
+ start = div64_u64_rem(start, st_len, &start_left);
+ if (start_left < st_bytes)
+ start_left = st_bytes - start_left;
+ else
+ start_left = 0;
- end_left = do_div(end, st_len);
- if (end_left > st_pgs)
- end_left = st_pgs;
+ end = div64_u64_rem(end, st_len, &end_left);
+ if (end_left > st_bytes)
+ end_left = st_bytes;
- CDEBUG(D_READA, "start %llu, end %llu start_left %lu end_left %lu\n",
- start, end, start_left, end_left);
+ CDEBUG(D_READA, "start %llu, end %llu start_left %llu end_left %llu\n",
+ start, end, start_left, end_left);
- if (start == end)
- pg_count = end_left - (st_pgs - start_left);
- else
- pg_count = start_left + st_pgs * (end - start - 1) + end_left;
+ if (start == end)
+ bytes_count = end_left - (st_bytes - start_left);
+ else
+ bytes_count = start_left +
+ st_bytes * (end - start - 1) + end_left;
- CDEBUG(D_READA, "st_off %lu, st_len %lu st_pgs %lu off %lu length %lu"
- "pgcount %lu\n", st_off, st_len, st_pgs, off, length, pg_count);
+ CDEBUG(D_READA,
+ "st_off %llu, st_len %llu st_bytes %llu off %llu length %llu bytescount %llu\n",
+ st_off, st_len, st_bytes, off, length, bytes_count);
- return pg_count;
+ return bytes_count;
}
-static int ria_page_count(struct ra_io_arg *ria)
+static unsigned long ria_page_count(struct ra_io_arg *ria)
{
- __u64 length = ria->ria_end >= ria->ria_start ?
- ria->ria_end - ria->ria_start + 1 : 0;
-
- return stride_pg_count(ria->ria_stoff, ria->ria_length,
- ria->ria_pages, ria->ria_start,
- length);
+ loff_t length_bytes = ria->ria_end_idx >= ria->ria_start_idx ?
+ (loff_t)(ria->ria_end_idx -
+ ria->ria_start_idx + 1) << PAGE_SHIFT : 0;
+ loff_t bytes_count;
+
+ if (ria->ria_length > ria->ria_bytes && ria->ria_bytes &&
+ (ria->ria_length & ~PAGE_MASK || ria->ria_bytes & ~PAGE_MASK ||
+ ria->ria_stoff & ~PAGE_MASK)) {
+ /* Over-estimate un-aligned page stride read */
+ unsigned long pg_count = ((ria->ria_bytes +
+ PAGE_SIZE - 1) >> PAGE_SHIFT) + 1;
+ pg_count *= length_bytes / ria->ria_length + 1;
+
+ return pg_count;
+ }
+ bytes_count = stride_byte_count(ria->ria_stoff, ria->ria_length,
+ ria->ria_bytes,
+ (loff_t)ria->ria_start_idx<<PAGE_SHIFT,
+ length_bytes);
+ return (bytes_count + PAGE_SIZE - 1) >> PAGE_SHIFT;
}
-static unsigned long ras_align(struct ll_readahead_state *ras,
- unsigned long index,
- unsigned long *remainder)
+static pgoff_t ras_align(struct ll_readahead_state *ras, pgoff_t index)
{
- unsigned long rem = index % ras->ras_rpc_size;
- if (remainder != NULL)
- *remainder = rem;
- return index - rem;
+ return index - (index % ras->ras_rpc_pages);
}
-/*Check whether the index is in the defined ra-window */
-static int ras_inside_ra_window(unsigned long idx, struct ra_io_arg *ria)
+/* Check whether the index is in the defined ra-window */
+static bool ras_inside_ra_window(pgoff_t idx, struct ra_io_arg *ria)
{
- /* If ria_length == ria_pages, it means non-stride I/O mode,
+ loff_t pos = (loff_t)idx << PAGE_SHIFT;
+
+ /* If ria_length == ria_bytes, it means non-stride I/O mode,
* idx should always inside read-ahead window in this case
* For stride I/O mode, just check whether the idx is inside
- * the ria_pages. */
- return ria->ria_length == 0 || ria->ria_length == ria->ria_pages ||
- (idx >= ria->ria_stoff && (idx - ria->ria_stoff) %
- ria->ria_length < ria->ria_pages);
+ * the ria_bytes.
+ */
+ if (ria->ria_length == 0 || ria->ria_length == ria->ria_bytes)
+ return true;
+
+ if (pos >= ria->ria_stoff) {
+ u64 offset;
+
+ div64_u64_rem(pos - ria->ria_stoff, ria->ria_length, &offset);
+
+ if (offset < ria->ria_bytes ||
+ (ria->ria_length - offset) < PAGE_SIZE)
+ return true;
+ } else if (pos + PAGE_SIZE > ria->ria_stoff) {
+ return true;
+ }
+
+ return false;
}
static unsigned long
struct ra_io_arg *ria, pgoff_t *ra_end)
{
struct cl_read_ahead ra = { 0 };
- int rc = 0, count = 0;
- bool stride_ria;
+ /* busy page count is per stride */
+ int rc = 0, count = 0, busy_page_count = 0;
pgoff_t page_idx;
LASSERT(ria != NULL);
RIA_DEBUG(ria);
- stride_ria = ria->ria_length > ria->ria_pages && ria->ria_pages > 0;
- for (page_idx = ria->ria_start;
- page_idx <= ria->ria_end && ria->ria_reserved > 0; page_idx++) {
+ for (page_idx = ria->ria_start_idx;
+ page_idx <= ria->ria_end_idx && ria->ria_reserved > 0;
+ page_idx++) {
if (ras_inside_ra_window(page_idx, ria)) {
- if (ra.cra_end == 0 || ra.cra_end < page_idx) {
- unsigned long end;
+ if (ra.cra_end_idx == 0 || ra.cra_end_idx < page_idx) {
+ pgoff_t end_idx;
cl_read_ahead_release(env, &ra);
if (rc < 0)
break;
+ /* Do not shrink ria_end_idx at any case until
+ * the minimum end of current read is covered.
+ * And only shrink ria_end_idx if the matched
+ * LDLM lock doesn't cover more. */
+ if (page_idx > ra.cra_end_idx ||
+ (ra.cra_contention &&
+ page_idx > ria->ria_end_idx_min)) {
+ ria->ria_end_idx = ra.cra_end_idx;
+ break;
+ }
+
CDEBUG(D_READA, "idx: %lu, ra: %lu, rpc: %lu\n",
- page_idx, ra.cra_end, ra.cra_rpc_size);
- LASSERTF(ra.cra_end >= page_idx,
+ page_idx, ra.cra_end_idx,
+ ra.cra_rpc_pages);
+ LASSERTF(ra.cra_end_idx >= page_idx,
"object: %p, indcies %lu / %lu\n",
- io->ci_obj, ra.cra_end, page_idx);
+ io->ci_obj, ra.cra_end_idx, page_idx);
/* update read ahead RPC size.
* NB: it's racy but doesn't matter */
- if (ras->ras_rpc_size > ra.cra_rpc_size &&
- ra.cra_rpc_size > 0)
- ras->ras_rpc_size = ra.cra_rpc_size;
+ if (ras->ras_rpc_pages != ra.cra_rpc_pages &&
+ ra.cra_rpc_pages > 0)
+ ras->ras_rpc_pages = ra.cra_rpc_pages;
/* trim it to align with optimal RPC size */
- end = ras_align(ras, ria->ria_end + 1, NULL);
- if (end > 0 && !ria->ria_eof)
- ria->ria_end = end - 1;
- if (ria->ria_end < ria->ria_end_min)
- ria->ria_end = ria->ria_end_min;
- if (ria->ria_end > ra.cra_end)
- ria->ria_end = ra.cra_end;
+ end_idx = ras_align(ras, ria->ria_end_idx + 1);
+ if (end_idx > 0 && !ria->ria_eof)
+ ria->ria_end_idx = end_idx - 1;
+ if (ria->ria_end_idx < ria->ria_end_idx_min)
+ ria->ria_end_idx = ria->ria_end_idx_min;
}
- if (page_idx > ria->ria_end)
+ if (page_idx > ria->ria_end_idx)
break;
/* If the page is inside the read-ahead window */
rc = ll_read_ahead_page(env, io, queue, page_idx);
- if (rc < 0)
+ if (rc < 0 && rc != -EBUSY)
break;
+ if (rc == -EBUSY) {
+ busy_page_count++;
+ CDEBUG(D_READA,
+ "skip busy page: %lu\n", page_idx);
+ /* For page unaligned readahead the first
+ * last pages of each region can be read by
+ * another reader on the same node, and so
+ * may be busy. So only stop for > 2 busy
+ * pages. */
+ if (busy_page_count > 2)
+ break;
+ }
*ra_end = page_idx;
/* Only subtract from reserve & count the page if we
ria->ria_reserved--;
count++;
}
- } else if (stride_ria) {
- /* If it is not in the read-ahead window, and it is
- * read-ahead mode, then check whether it should skip
- * the stride gap */
- pgoff_t offset;
- /* FIXME: This assertion only is valid when it is for
- * forward read-ahead, it will be fixed when backward
- * read-ahead is implemented */
- LASSERTF(page_idx >= ria->ria_stoff,
- "Invalid page_idx %lu rs %lu re %lu ro %lu "
- "rl %lu rp %lu\n", page_idx,
- ria->ria_start, ria->ria_end, ria->ria_stoff,
- ria->ria_length, ria->ria_pages);
- offset = page_idx - ria->ria_stoff;
- offset = offset % (ria->ria_length);
- if (offset > ria->ria_pages) {
- page_idx += ria->ria_length - offset;
- CDEBUG(D_READA, "i %lu skip %lu \n", page_idx,
- ria->ria_length - offset);
- continue;
- }
- }
- }
+ } else if (stride_io_mode(ras)) {
+ /* If it is not in the read-ahead window, and it is
+ * read-ahead mode, then check whether it should skip
+ * the stride gap.
+ */
+ loff_t pos = (loff_t)page_idx << PAGE_SHIFT;
+ u64 offset;
+
+ div64_u64_rem(pos - ria->ria_stoff, ria->ria_length,
+ &offset);
+ if (offset >= ria->ria_bytes) {
+ pos += (ria->ria_length - offset);
+ if ((pos >> PAGE_SHIFT) >= page_idx + 1)
+ page_idx = (pos >> PAGE_SHIFT) - 1;
+ busy_page_count = 0;
+ CDEBUG(D_READA,
+ "Stride: jump %llu pages to %lu\n",
+ ria->ria_length - offset, page_idx);
+ continue;
+ }
+ }
+ }
cl_read_ahead_release(env, &ra);
return count;
}
+static void ll_readahead_work_free(struct ll_readahead_work *work)
+{
+ fput(work->lrw_file);
+ OBD_FREE_PTR(work);
+}
+
+static void ll_readahead_handle_work(struct work_struct *wq);
+static void ll_readahead_work_add(struct inode *inode,
+ struct ll_readahead_work *work)
+{
+ INIT_WORK(&work->lrw_readahead_work, ll_readahead_handle_work);
+ queue_work(ll_i2sbi(inode)->ll_ra_info.ll_readahead_wq,
+ &work->lrw_readahead_work);
+}
+
+static int ll_readahead_file_kms(const struct lu_env *env,
+ struct cl_io *io, __u64 *kms)
+{
+ struct cl_object *clob;
+ struct inode *inode;
+ struct cl_attr *attr = vvp_env_thread_attr(env);
+ int ret;
+
+ clob = io->ci_obj;
+ inode = vvp_object_inode(clob);
+
+ cl_object_attr_lock(clob);
+ ret = cl_object_attr_get(env, clob, attr);
+ cl_object_attr_unlock(clob);
+
+ if (ret != 0)
+ RETURN(ret);
+
+ *kms = attr->cat_kms;
+ return 0;
+}
+
+static void ll_readahead_handle_work(struct work_struct *wq)
+{
+ struct ll_readahead_work *work;
+ struct lu_env *env;
+ __u16 refcheck;
+ struct ra_io_arg *ria;
+ struct inode *inode;
+ struct ll_file_data *fd;
+ struct ll_readahead_state *ras;
+ struct cl_io *io;
+ struct cl_2queue *queue;
+ pgoff_t ra_end_idx = 0;
+ unsigned long pages, pages_min = 0;
+ struct file *file;
+ __u64 kms;
+ int rc;
+ pgoff_t eof_index;
+
+ work = container_of(wq, struct ll_readahead_work,
+ lrw_readahead_work);
+ fd = LUSTRE_FPRIVATE(work->lrw_file);
+ ras = &fd->fd_ras;
+ file = work->lrw_file;
+ inode = file_inode(file);
+
+ env = cl_env_alloc(&refcheck, LCT_NOREF);
+ if (IS_ERR(env))
+ GOTO(out_free_work, rc = PTR_ERR(env));
+
+ io = vvp_env_thread_io(env);
+ ll_io_init(io, file, CIT_READ, NULL);
+
+ rc = ll_readahead_file_kms(env, io, &kms);
+ if (rc != 0)
+ GOTO(out_put_env, rc);
+
+ if (kms == 0) {
+ ll_ra_stats_inc(inode, RA_STAT_ZERO_LEN);
+ GOTO(out_put_env, rc = 0);
+ }
+
+ ria = &ll_env_info(env)->lti_ria;
+ memset(ria, 0, sizeof(*ria));
+
+ ria->ria_start_idx = work->lrw_start_idx;
+ /* Truncate RA window to end of file */
+ eof_index = (pgoff_t)(kms - 1) >> PAGE_SHIFT;
+ if (eof_index <= work->lrw_end_idx) {
+ work->lrw_end_idx = eof_index;
+ ria->ria_eof = true;
+ }
+ if (work->lrw_end_idx <= work->lrw_start_idx)
+ GOTO(out_put_env, rc = 0);
+
+ ria->ria_end_idx = work->lrw_end_idx;
+ pages = ria->ria_end_idx - ria->ria_start_idx + 1;
+ ria->ria_reserved = ll_ra_count_get(ll_i2sbi(inode), ria,
+ ria_page_count(ria), pages_min);
+
+ CDEBUG(D_READA,
+ "async reserved pages: %lu/%lu/%lu, ra_cur %d, ra_max %lu\n",
+ ria->ria_reserved, pages, pages_min,
+ atomic_read(&ll_i2sbi(inode)->ll_ra_info.ra_cur_pages),
+ ll_i2sbi(inode)->ll_ra_info.ra_max_pages);
+
+ if (ria->ria_reserved < pages) {
+ ll_ra_stats_inc(inode, RA_STAT_MAX_IN_FLIGHT);
+ if (PAGES_TO_MiB(ria->ria_reserved) < 1) {
+ ll_ra_count_put(ll_i2sbi(inode), ria->ria_reserved);
+ GOTO(out_put_env, rc = 0);
+ }
+ }
+
+ rc = cl_io_rw_init(env, io, CIT_READ, ria->ria_start_idx, pages);
+ if (rc)
+ GOTO(out_put_env, rc);
+
+ vvp_env_io(env)->vui_io_subtype = IO_NORMAL;
+ vvp_env_io(env)->vui_fd = fd;
+ io->ci_state = CIS_LOCKED;
+ io->ci_async_readahead = true;
+ rc = cl_io_start(env, io);
+ if (rc)
+ GOTO(out_io_fini, rc);
+
+ queue = &io->ci_queue;
+ cl_2queue_init(queue);
+
+ rc = ll_read_ahead_pages(env, io, &queue->c2_qin, ras, ria,
+ &ra_end_idx);
+ if (ria->ria_reserved != 0)
+ ll_ra_count_put(ll_i2sbi(inode), ria->ria_reserved);
+ if (queue->c2_qin.pl_nr > 0) {
+ int count = queue->c2_qin.pl_nr;
+
+ rc = cl_io_submit_rw(env, io, CRT_READ, queue);
+ if (rc == 0)
+ task_io_account_read(PAGE_SIZE * count);
+ }
+ if (ria->ria_end_idx == ra_end_idx && ra_end_idx == (kms >> PAGE_SHIFT))
+ ll_ra_stats_inc(inode, RA_STAT_EOF);
+
+ if (ra_end_idx != ria->ria_end_idx)
+ ll_ra_stats_inc(inode, RA_STAT_FAILED_REACH_END);
+
+ /* TODO: discard all pages until page reinit route is implemented */
+ cl_page_list_discard(env, io, &queue->c2_qin);
+
+ /* Unlock unsent read pages in case of error. */
+ cl_page_list_disown(env, io, &queue->c2_qin);
+
+ cl_2queue_fini(env, queue);
+out_io_fini:
+ cl_io_end(env, io);
+ cl_io_fini(env, io);
+out_put_env:
+ cl_env_put(env, &refcheck);
+out_free_work:
+ if (ra_end_idx > 0)
+ ll_ra_stats_inc_sbi(ll_i2sbi(inode), RA_STAT_ASYNC);
+ ll_readahead_work_free(work);
+}
+
static int ll_readahead(const struct lu_env *env, struct cl_io *io,
struct cl_page_list *queue,
- struct ll_readahead_state *ras, bool hit)
+ struct ll_readahead_state *ras, bool hit,
+ struct file *file)
{
struct vvp_io *vio = vvp_env_io(env);
struct ll_thread_info *lti = ll_env_info(env);
- struct cl_attr *attr = vvp_env_thread_attr(env);
- unsigned long len, mlen = 0;
- pgoff_t ra_end = 0, start = 0, end = 0;
+ unsigned long pages, pages_min = 0;
+ pgoff_t ra_end_idx = 0, start_idx = 0, end_idx = 0;
struct inode *inode;
struct ra_io_arg *ria = <i->lti_ria;
struct cl_object *clob;
clob = io->ci_obj;
inode = vvp_object_inode(clob);
- memset(ria, 0, sizeof *ria);
-
- cl_object_attr_lock(clob);
- ret = cl_object_attr_get(env, clob, attr);
- cl_object_attr_unlock(clob);
-
+ memset(ria, 0, sizeof(*ria));
+ ret = ll_readahead_file_kms(env, io, &kms);
if (ret != 0)
RETURN(ret);
- kms = attr->cat_kms;
+
if (kms == 0) {
ll_ra_stats_inc(inode, RA_STAT_ZERO_LEN);
RETURN(0);
spin_lock(&ras->ras_lock);
/**
- * Note: other thread might rollback the ras_next_readahead,
+ * Note: other thread might rollback the ras_next_readahead_idx,
* if it can not get the full size of prepared pages, see the
* end of this function. For stride read ahead, it needs to
* make sure the offset is no less than ras_stride_offset,
* so that stride read ahead can work correctly.
*/
if (stride_io_mode(ras))
- start = max(ras->ras_next_readahead, ras->ras_stride_offset);
+ start_idx = max_t(pgoff_t, ras->ras_next_readahead_idx,
+ ras->ras_stride_offset >> PAGE_SHIFT);
else
- start = ras->ras_next_readahead;
+ start_idx = ras->ras_next_readahead_idx;
- if (ras->ras_window_len > 0)
- end = ras->ras_window_start + ras->ras_window_len - 1;
+ if (ras->ras_window_pages > 0)
+ end_idx = ras->ras_window_start_idx + ras->ras_window_pages - 1;
/* Enlarge the RA window to encompass the full read */
if (vio->vui_ra_valid &&
- end < vio->vui_ra_start + vio->vui_ra_count - 1)
- end = vio->vui_ra_start + vio->vui_ra_count - 1;
+ end_idx < vio->vui_ra_start_idx + vio->vui_ra_pages - 1)
+ end_idx = vio->vui_ra_start_idx + vio->vui_ra_pages - 1;
- if (end != 0) {
- unsigned long end_index;
+ if (end_idx != 0) {
+ pgoff_t eof_index;
/* Truncate RA window to end of file */
- end_index = (unsigned long)((kms - 1) >> PAGE_SHIFT);
- if (end_index <= end) {
- end = end_index;
+ eof_index = (pgoff_t)((kms - 1) >> PAGE_SHIFT);
+ if (eof_index <= end_idx) {
+ end_idx = eof_index;
ria->ria_eof = true;
}
- }
- ria->ria_start = start;
- ria->ria_end = end;
- /* If stride I/O mode is detected, get stride window*/
- if (stride_io_mode(ras)) {
- ria->ria_stoff = ras->ras_stride_offset;
- ria->ria_length = ras->ras_stride_length;
- ria->ria_pages = ras->ras_stride_pages;
- }
+ }
+ ria->ria_start_idx = start_idx;
+ ria->ria_end_idx = end_idx;
+ /* If stride I/O mode is detected, get stride window*/
+ if (stride_io_mode(ras)) {
+ ria->ria_stoff = ras->ras_stride_offset;
+ ria->ria_length = ras->ras_stride_length;
+ ria->ria_bytes = ras->ras_stride_bytes;
+ }
spin_unlock(&ras->ras_lock);
- if (end == 0) {
+ if (end_idx == 0) {
ll_ra_stats_inc(inode, RA_STAT_ZERO_WINDOW);
RETURN(0);
}
- len = ria_page_count(ria);
- if (len == 0) {
+ pages = ria_page_count(ria);
+ if (pages == 0) {
ll_ra_stats_inc(inode, RA_STAT_ZERO_WINDOW);
RETURN(0);
}
RAS_CDEBUG(ras);
CDEBUG(D_READA, DFID": ria: %lu/%lu, bead: %lu/%lu, hit: %d\n",
PFID(lu_object_fid(&clob->co_lu)),
- ria->ria_start, ria->ria_end,
- vio->vui_ra_valid ? vio->vui_ra_start : 0,
- vio->vui_ra_valid ? vio->vui_ra_count : 0,
+ ria->ria_start_idx, ria->ria_end_idx,
+ vio->vui_ra_valid ? vio->vui_ra_start_idx : 0,
+ vio->vui_ra_valid ? vio->vui_ra_pages : 0,
hit);
/* at least to extend the readahead window to cover current read */
if (!hit && vio->vui_ra_valid &&
- vio->vui_ra_start + vio->vui_ra_count > ria->ria_start) {
- unsigned long remainder;
-
- /* to the end of current read window. */
- mlen = vio->vui_ra_start + vio->vui_ra_count - ria->ria_start;
- /* trim to RPC boundary */
- ras_align(ras, ria->ria_start, &remainder);
- mlen = min(mlen, ras->ras_rpc_size - remainder);
- ria->ria_end_min = ria->ria_start + mlen;
- }
+ vio->vui_ra_start_idx + vio->vui_ra_pages > ria->ria_start_idx)
+ ria->ria_end_idx_min =
+ vio->vui_ra_start_idx + vio->vui_ra_pages - 1;
- ria->ria_reserved = ll_ra_count_get(ll_i2sbi(inode), ria, len, mlen);
- if (ria->ria_reserved < len)
+ ria->ria_reserved = ll_ra_count_get(ll_i2sbi(inode), ria, pages,
+ pages_min);
+ if (ria->ria_reserved < pages)
ll_ra_stats_inc(inode, RA_STAT_MAX_IN_FLIGHT);
CDEBUG(D_READA, "reserved pages: %lu/%lu/%lu, ra_cur %d, ra_max %lu\n",
- ria->ria_reserved, len, mlen,
+ ria->ria_reserved, pages, pages_min,
atomic_read(&ll_i2sbi(inode)->ll_ra_info.ra_cur_pages),
ll_i2sbi(inode)->ll_ra_info.ra_max_pages);
- ret = ll_read_ahead_pages(env, io, queue, ras, ria, &ra_end);
+ ret = ll_read_ahead_pages(env, io, queue, ras, ria, &ra_end_idx);
if (ria->ria_reserved != 0)
ll_ra_count_put(ll_i2sbi(inode), ria->ria_reserved);
- if (ra_end == end && ra_end == (kms >> PAGE_SHIFT))
+ if (ra_end_idx == end_idx && ra_end_idx == (kms >> PAGE_SHIFT))
ll_ra_stats_inc(inode, RA_STAT_EOF);
- CDEBUG(D_READA, "ra_end = %lu end = %lu stride end = %lu pages = %d\n",
- ra_end, end, ria->ria_end, ret);
+ CDEBUG(D_READA,
+ "ra_end_idx = %lu end_idx = %lu stride end = %lu pages = %d\n",
+ ra_end_idx, end_idx, ria->ria_end_idx, ret);
- if (ra_end != end)
+ if (ra_end_idx != end_idx)
ll_ra_stats_inc(inode, RA_STAT_FAILED_REACH_END);
- if (ra_end > 0) {
+ if (ra_end_idx > 0) {
/* update the ras so that the next read-ahead tries from
* where we left off. */
spin_lock(&ras->ras_lock);
- ras->ras_next_readahead = ra_end + 1;
+ ras->ras_next_readahead_idx = ra_end_idx + 1;
spin_unlock(&ras->ras_lock);
RAS_CDEBUG(ras);
}
RETURN(ret);
}
-static void ras_set_start(struct inode *inode, struct ll_readahead_state *ras,
- unsigned long index)
+static void ras_set_start(struct ll_readahead_state *ras, pgoff_t index)
{
- ras->ras_window_start = ras_align(ras, index, NULL);
+ ras->ras_window_start_idx = ras_align(ras, index);
}
/* called with the ras_lock held or from places where it doesn't matter */
-static void ras_reset(struct inode *inode, struct ll_readahead_state *ras,
- unsigned long index)
+static void ras_reset(struct ll_readahead_state *ras, pgoff_t index)
{
- ras->ras_last_readpage = index;
ras->ras_consecutive_requests = 0;
- ras->ras_consecutive_pages = 0;
- ras->ras_window_len = 0;
- ras_set_start(inode, ras, index);
- ras->ras_next_readahead = max(ras->ras_window_start, index + 1);
+ ras->ras_consecutive_bytes = 0;
+ ras->ras_window_pages = 0;
+ ras_set_start(ras, index);
+ ras->ras_next_readahead_idx = max(ras->ras_window_start_idx, index + 1);
RAS_CDEBUG(ras);
}
{
ras->ras_consecutive_stride_requests = 0;
ras->ras_stride_length = 0;
- ras->ras_stride_pages = 0;
+ ras->ras_stride_bytes = 0;
RAS_CDEBUG(ras);
}
void ll_readahead_init(struct inode *inode, struct ll_readahead_state *ras)
{
spin_lock_init(&ras->ras_lock);
- ras->ras_rpc_size = PTLRPC_MAX_BRW_PAGES;
- ras_reset(inode, ras, 0);
+ ras->ras_rpc_pages = PTLRPC_MAX_BRW_PAGES;
+ ras_reset(ras, 0);
+ ras->ras_last_read_end_bytes = 0;
ras->ras_requests = 0;
}
/*
* Check whether the read request is in the stride window.
- * If it is in the stride window, return 1, otherwise return 0.
+ * If it is in the stride window, return true, otherwise return false.
*/
-static int index_in_stride_window(struct ll_readahead_state *ras,
- unsigned long index)
+static bool read_in_stride_window(struct ll_readahead_state *ras,
+ loff_t pos, loff_t count)
{
- unsigned long stride_gap;
+ loff_t stride_gap;
- if (ras->ras_stride_length == 0 || ras->ras_stride_pages == 0 ||
- ras->ras_stride_pages == ras->ras_stride_length)
- return 0;
+ if (ras->ras_stride_length == 0 || ras->ras_stride_bytes == 0 ||
+ ras->ras_stride_bytes == ras->ras_stride_length)
+ return false;
- stride_gap = index - ras->ras_last_readpage - 1;
+ stride_gap = pos - ras->ras_last_read_end_bytes - 1;
/* If it is contiguous read */
if (stride_gap == 0)
- return ras->ras_consecutive_pages + 1 <= ras->ras_stride_pages;
+ return ras->ras_consecutive_bytes + count <=
+ ras->ras_stride_bytes;
/* Otherwise check the stride by itself */
- return (ras->ras_stride_length - ras->ras_stride_pages) == stride_gap &&
- ras->ras_consecutive_pages == ras->ras_stride_pages;
+ return (ras->ras_stride_length - ras->ras_stride_bytes) == stride_gap &&
+ ras->ras_consecutive_bytes == ras->ras_stride_bytes &&
+ count <= ras->ras_stride_bytes;
}
-static void ras_update_stride_detector(struct ll_readahead_state *ras,
- unsigned long index)
+static void ras_init_stride_detector(struct ll_readahead_state *ras,
+ loff_t pos, loff_t count)
{
- unsigned long stride_gap = index - ras->ras_last_readpage - 1;
+ loff_t stride_gap = pos - ras->ras_last_read_end_bytes - 1;
- if (!stride_io_mode(ras) && (stride_gap != 0 ||
- ras->ras_consecutive_stride_requests == 0)) {
- ras->ras_stride_pages = ras->ras_consecutive_pages;
- ras->ras_stride_length = stride_gap +ras->ras_consecutive_pages;
- }
- LASSERT(ras->ras_request_index == 0);
LASSERT(ras->ras_consecutive_stride_requests == 0);
- if (index <= ras->ras_last_readpage) {
+ if (pos <= ras->ras_last_read_end_bytes) {
/*Reset stride window for forward read*/
ras_stride_reset(ras);
return;
}
- ras->ras_stride_pages = ras->ras_consecutive_pages;
- ras->ras_stride_length = stride_gap +ras->ras_consecutive_pages;
+ ras->ras_stride_bytes = ras->ras_consecutive_bytes;
+ ras->ras_stride_length = stride_gap + ras->ras_consecutive_bytes;
+ ras->ras_consecutive_stride_requests++;
+ ras->ras_stride_offset = pos;
RAS_CDEBUG(ras);
- return;
}
static unsigned long
-stride_page_count(struct ll_readahead_state *ras, unsigned long len)
+stride_page_count(struct ll_readahead_state *ras, loff_t len)
{
- return stride_pg_count(ras->ras_stride_offset, ras->ras_stride_length,
- ras->ras_stride_pages, ras->ras_stride_offset,
- len);
+ loff_t bytes_count =
+ stride_byte_count(ras->ras_stride_offset,
+ ras->ras_stride_length, ras->ras_stride_bytes,
+ ras->ras_stride_offset, len);
+
+ return (bytes_count + PAGE_SIZE - 1) >> PAGE_SHIFT;
}
/* Stride Read-ahead window will be increased inc_len according to
* stride I/O pattern */
static void ras_stride_increase_window(struct ll_readahead_state *ras,
- struct ll_ra_info *ra,
- unsigned long inc_len)
+ struct ll_ra_info *ra, loff_t inc_bytes)
{
- unsigned long left, step, window_len;
- unsigned long stride_len;
+ loff_t window_bytes, stride_bytes;
+ u64 left_bytes;
+ u64 step;
+ loff_t end;
- LASSERT(ras->ras_stride_length > 0);
- LASSERTF(ras->ras_window_start + ras->ras_window_len
- >= ras->ras_stride_offset, "window_start %lu, window_len %lu"
- " stride_offset %lu\n", ras->ras_window_start,
- ras->ras_window_len, ras->ras_stride_offset);
+ /* temporarily store in page units to reduce LASSERT() cost below */
+ end = ras->ras_window_start_idx + ras->ras_window_pages;
- stride_len = ras->ras_window_start + ras->ras_window_len -
- ras->ras_stride_offset;
+ LASSERT(ras->ras_stride_length > 0);
+ LASSERTF(end >= (ras->ras_stride_offset >> PAGE_SHIFT),
+ "window_start_idx %lu, window_pages %lu stride_offset %llu\n",
+ ras->ras_window_start_idx, ras->ras_window_pages,
+ ras->ras_stride_offset);
- left = stride_len % ras->ras_stride_length;
- window_len = ras->ras_window_len - left;
+ end <<= PAGE_SHIFT;
+ if (end <= ras->ras_stride_offset)
+ stride_bytes = 0;
+ else
+ stride_bytes = end - ras->ras_stride_offset;
- if (left < ras->ras_stride_pages)
- left += inc_len;
- else
- left = ras->ras_stride_pages + inc_len;
+ div64_u64_rem(stride_bytes, ras->ras_stride_length, &left_bytes);
+ window_bytes = ((loff_t)ras->ras_window_pages << PAGE_SHIFT) -
+ left_bytes;
- LASSERT(ras->ras_stride_pages != 0);
+ if (left_bytes < ras->ras_stride_bytes)
+ left_bytes += inc_bytes;
+ else
+ left_bytes = ras->ras_stride_bytes + inc_bytes;
- step = left / ras->ras_stride_pages;
- left %= ras->ras_stride_pages;
+ LASSERT(ras->ras_stride_bytes != 0);
- window_len += step * ras->ras_stride_length + left;
+ step = div64_u64_rem(left_bytes, ras->ras_stride_bytes, &left_bytes);
- if (stride_page_count(ras, window_len) <= ra->ra_max_pages_per_file)
- ras->ras_window_len = window_len;
+ window_bytes += step * ras->ras_stride_length + left_bytes;
- RAS_CDEBUG(ras);
+ if (stride_page_count(ras, window_bytes) <= ra->ra_max_pages_per_file)
+ ras->ras_window_pages = (window_bytes >> PAGE_SHIFT);
+
+ RAS_CDEBUG(ras);
}
static void ras_increase_window(struct inode *inode,
* information from lower layer. FIXME later
*/
if (stride_io_mode(ras)) {
- ras_stride_increase_window(ras, ra, ras->ras_rpc_size);
+ ras_stride_increase_window(ras, ra,
+ (loff_t)ras->ras_rpc_pages << PAGE_SHIFT);
} else {
- unsigned long wlen;
+ pgoff_t window_pages;
- wlen = min(ras->ras_window_len + ras->ras_rpc_size,
- ra->ra_max_pages_per_file);
- if (wlen < ras->ras_rpc_size)
- ras->ras_window_len = wlen;
+ window_pages = min(ras->ras_window_pages + ras->ras_rpc_pages,
+ ra->ra_max_pages_per_file);
+ if (window_pages < ras->ras_rpc_pages)
+ ras->ras_window_pages = window_pages;
else
- ras->ras_window_len = ras_align(ras, wlen, NULL);
+ ras->ras_window_pages = ras_align(ras, window_pages);
}
}
+/**
+ * Seek within 8 pages are considered as sequential read for now.
+ */
+static inline bool is_loose_seq_read(struct ll_readahead_state *ras, loff_t pos)
+{
+ return pos_in_window(pos, ras->ras_last_read_end_bytes,
+ 8UL << PAGE_SHIFT, 8UL << PAGE_SHIFT);
+}
+
+static void ras_detect_read_pattern(struct ll_readahead_state *ras,
+ struct ll_sb_info *sbi,
+ loff_t pos, size_t count, bool mmap)
+{
+ bool stride_detect = false;
+ pgoff_t index = pos >> PAGE_SHIFT;
+
+ /*
+ * Reset the read-ahead window in two cases. First when the app seeks
+ * or reads to some other part of the file. Secondly if we get a
+ * read-ahead miss that we think we've previously issued. This can
+ * be a symptom of there being so many read-ahead pages that the VM
+ * is reclaiming it before we get to it.
+ */
+ if (!is_loose_seq_read(ras, pos)) {
+ /* Check whether it is in stride I/O mode */
+ if (!read_in_stride_window(ras, pos, count)) {
+ if (ras->ras_consecutive_stride_requests == 0)
+ ras_init_stride_detector(ras, pos, count);
+ else
+ ras_stride_reset(ras);
+ ras->ras_consecutive_bytes = 0;
+ ras_reset(ras, index);
+ } else {
+ ras->ras_consecutive_bytes = 0;
+ ras->ras_consecutive_requests = 0;
+ if (++ras->ras_consecutive_stride_requests > 1)
+ stride_detect = true;
+ RAS_CDEBUG(ras);
+ }
+ ll_ra_stats_inc_sbi(sbi, RA_STAT_DISTANT_READPAGE);
+ } else if (stride_io_mode(ras)) {
+ /*
+ * If this is contiguous read but in stride I/O mode
+ * currently, check whether stride step still is valid,
+ * if invalid, it will reset the stride ra window to
+ * be zero.
+ */
+ if (!read_in_stride_window(ras, pos, count)) {
+ ras_stride_reset(ras);
+ ras->ras_window_pages = 0;
+ ras->ras_next_readahead_idx = index;
+ }
+ }
+
+ ras->ras_consecutive_bytes += count;
+ if (mmap) {
+ pgoff_t idx = ras->ras_consecutive_bytes >> PAGE_SHIFT;
+
+ if ((idx >= 4 && (idx & 3UL) == 0) || stride_detect)
+ ras->ras_need_increase_window = true;
+ } else if ((ras->ras_consecutive_requests > 1 || stride_detect)) {
+ ras->ras_need_increase_window = true;
+ }
+
+ ras->ras_last_read_end_bytes = pos + count - 1;
+}
+
+void ll_ras_enter(struct file *f, loff_t pos, size_t count)
+{
+ struct ll_file_data *fd = LUSTRE_FPRIVATE(f);
+ struct ll_readahead_state *ras = &fd->fd_ras;
+ struct inode *inode = file_inode(f);
+ unsigned long index = pos >> PAGE_SHIFT;
+ struct ll_sb_info *sbi = ll_i2sbi(inode);
+
+ spin_lock(&ras->ras_lock);
+ ras->ras_requests++;
+ ras->ras_consecutive_requests++;
+ ras->ras_need_increase_window = false;
+ ras->ras_no_miss_check = false;
+ /*
+ * On the second access to a file smaller than the tunable
+ * ra_max_read_ahead_whole_pages trigger RA on all pages in the
+ * file up to ra_max_pages_per_file. This is simply a best effort
+ * and only occurs once per open file. Normal RA behavior is reverted
+ * to for subsequent IO.
+ */
+ if (ras->ras_requests >= 2) {
+ __u64 kms_pages;
+ struct ll_ra_info *ra = &sbi->ll_ra_info;
+
+ kms_pages = (i_size_read(inode) + PAGE_SIZE - 1) >>
+ PAGE_SHIFT;
+
+ CDEBUG(D_READA, "kmsp %llu mwp %lu mp %lu\n", kms_pages,
+ ra->ra_max_read_ahead_whole_pages,
+ ra->ra_max_pages_per_file);
+
+ if (kms_pages &&
+ kms_pages <= ra->ra_max_read_ahead_whole_pages) {
+ ras->ras_window_start_idx = 0;
+ ras->ras_next_readahead_idx = index + 1;
+ ras->ras_window_pages = min(ra->ra_max_pages_per_file,
+ ra->ra_max_read_ahead_whole_pages);
+ ras->ras_no_miss_check = true;
+ GOTO(out_unlock, 0);
+ }
+ }
+ ras_detect_read_pattern(ras, sbi, pos, count, false);
+out_unlock:
+ spin_unlock(&ras->ras_lock);
+}
+
+static bool index_in_stride_window(struct ll_readahead_state *ras,
+ pgoff_t index)
+{
+ loff_t pos = (loff_t)index << PAGE_SHIFT;
+
+ if (ras->ras_stride_length == 0 || ras->ras_stride_bytes == 0 ||
+ ras->ras_stride_bytes == ras->ras_stride_length)
+ return false;
+
+ if (pos >= ras->ras_stride_offset) {
+ u64 offset;
+
+ div64_u64_rem(pos - ras->ras_stride_offset,
+ ras->ras_stride_length, &offset);
+ if (offset < ras->ras_stride_bytes ||
+ ras->ras_stride_length - offset < PAGE_SIZE)
+ return true;
+ } else if (ras->ras_stride_offset - pos < PAGE_SIZE) {
+ return true;
+ }
+
+ return false;
+}
+
+/*
+ * ll_ras_enter() is used to detect read pattern according to pos and count.
+ *
+ * ras_update() is used to detect cache miss and
+ * reset window or increase window accordingly
+ */
static void ras_update(struct ll_sb_info *sbi, struct inode *inode,
- struct ll_readahead_state *ras, unsigned long index,
+ struct ll_readahead_state *ras, pgoff_t index,
enum ras_update_flags flags)
{
struct ll_ra_info *ra = &sbi->ll_ra_info;
bool hit = flags & LL_RAS_HIT;
- int zero = 0, stride_detect = 0, ra_miss = 0;
- ENTRY;
+ ENTRY;
spin_lock(&ras->ras_lock);
if (!hit)
CDEBUG(D_READA, DFID " pages at %lu miss.\n",
PFID(ll_inode2fid(inode)), index);
- ll_ra_stats_inc_sbi(sbi, hit ? RA_STAT_HIT : RA_STAT_MISS);
-
- /* reset the read-ahead window in two cases. First when the app seeks
- * or reads to some other part of the file. Secondly if we get a
- * read-ahead miss that we think we've previously issued. This can
- * be a symptom of there being so many read-ahead pages that the VM is
- * reclaiming it before we get to it. */
- if (!index_in_window(index, ras->ras_last_readpage, 8, 8)) {
- zero = 1;
- ll_ra_stats_inc_sbi(sbi, RA_STAT_DISTANT_READPAGE);
- } else if (!hit && ras->ras_window_len &&
- index < ras->ras_next_readahead &&
- index_in_window(index, ras->ras_window_start, 0,
- ras->ras_window_len)) {
- ra_miss = 1;
- ll_ra_stats_inc_sbi(sbi, RA_STAT_MISS_IN_WINDOW);
- }
+ ll_ra_stats_inc_sbi(sbi, hit ? RA_STAT_HIT : RA_STAT_MISS);
- /* On the second access to a file smaller than the tunable
- * ra_max_read_ahead_whole_pages trigger RA on all pages in the
- * file up to ra_max_pages_per_file. This is simply a best effort
- * and only occurs once per open file. Normal RA behavior is reverted
- * to for subsequent IO. The mmap case does not increment
- * ras_requests and thus can never trigger this behavior. */
- if (ras->ras_requests >= 2 && !ras->ras_request_index) {
- __u64 kms_pages;
-
- kms_pages = (i_size_read(inode) + PAGE_SIZE - 1) >>
- PAGE_SHIFT;
+ /*
+ * The readahead window has been expanded to cover whole
+ * file size, we don't care whether ra miss happen or not.
+ * Because we will read whole file to page cache even if
+ * some pages missed.
+ */
+ if (ras->ras_no_miss_check)
+ GOTO(out_unlock, 0);
- CDEBUG(D_READA, "kmsp %llu mwp %lu mp %lu\n", kms_pages,
- ra->ra_max_read_ahead_whole_pages, ra->ra_max_pages_per_file);
-
- if (kms_pages &&
- kms_pages <= ra->ra_max_read_ahead_whole_pages) {
- ras->ras_window_start = 0;
- ras->ras_next_readahead = index + 1;
- ras->ras_window_len = min(ra->ra_max_pages_per_file,
- ra->ra_max_read_ahead_whole_pages);
- GOTO(out_unlock, 0);
- }
- }
- if (zero) {
- /* check whether it is in stride I/O mode*/
- if (!index_in_stride_window(ras, index)) {
- if (ras->ras_consecutive_stride_requests == 0 &&
- ras->ras_request_index == 0) {
- ras_update_stride_detector(ras, index);
- ras->ras_consecutive_stride_requests++;
- } else {
+ if (flags & LL_RAS_MMAP)
+ ras_detect_read_pattern(ras, sbi, (loff_t)index << PAGE_SHIFT,
+ PAGE_SIZE, true);
+
+ if (!hit && ras->ras_window_pages &&
+ index < ras->ras_next_readahead_idx &&
+ pos_in_window(index, ras->ras_window_start_idx, 0,
+ ras->ras_window_pages)) {
+ ll_ra_stats_inc_sbi(sbi, RA_STAT_MISS_IN_WINDOW);
+ ras->ras_need_increase_window = false;
+
+ if (index_in_stride_window(ras, index) &&
+ stride_io_mode(ras)) {
+ /*
+ * if (index != ras->ras_last_readpage + 1)
+ * ras->ras_consecutive_pages = 0;
+ */
+ ras_reset(ras, index);
+
+ /*
+ * If stride-RA hit cache miss, the stride
+ * detector will not be reset to avoid the
+ * overhead of redetecting read-ahead mode,
+ * but on the condition that the stride window
+ * is still intersect with normal sequential
+ * read-ahead window.
+ */
+ if (ras->ras_window_start_idx < ras->ras_stride_offset)
ras_stride_reset(ras);
- }
- ras_reset(inode, ras, index);
- ras->ras_consecutive_pages++;
- GOTO(out_unlock, 0);
- } else {
- ras->ras_consecutive_pages = 0;
- ras->ras_consecutive_requests = 0;
- if (++ras->ras_consecutive_stride_requests > 1)
- stride_detect = 1;
RAS_CDEBUG(ras);
- }
- } else {
- if (ra_miss) {
- if (index_in_stride_window(ras, index) &&
- stride_io_mode(ras)) {
- if (index != ras->ras_last_readpage + 1)
- ras->ras_consecutive_pages = 0;
- ras_reset(inode, ras, index);
-
- /* If stride-RA hit cache miss, the stride
- * detector will not be reset to avoid the
- * overhead of redetecting read-ahead mode,
- * but on the condition that the stride window
- * is still intersect with normal sequential
- * read-ahead window. */
- if (ras->ras_window_start <
- ras->ras_stride_offset)
- ras_stride_reset(ras);
- RAS_CDEBUG(ras);
- } else {
- /* Reset both stride window and normal RA
- * window */
- ras_reset(inode, ras, index);
- ras->ras_consecutive_pages++;
- ras_stride_reset(ras);
- GOTO(out_unlock, 0);
- }
- } else if (stride_io_mode(ras)) {
- /* If this is contiguous read but in stride I/O mode
- * currently, check whether stride step still is valid,
- * if invalid, it will reset the stride ra window*/
- if (!index_in_stride_window(ras, index)) {
- /* Shrink stride read-ahead window to be zero */
- ras_stride_reset(ras);
- ras->ras_window_len = 0;
- ras->ras_next_readahead = index;
- }
+ } else {
+ /*
+ * Reset both stride window and normal RA
+ * window.
+ */
+ ras_reset(ras, index);
+ /* ras->ras_consecutive_pages++; */
+ ras->ras_consecutive_bytes = 0;
+ ras_stride_reset(ras);
+ GOTO(out_unlock, 0);
}
}
- ras->ras_consecutive_pages++;
- ras->ras_last_readpage = index;
- ras_set_start(inode, ras, index);
+ ras_set_start(ras, index);
if (stride_io_mode(ras)) {
/* Since stride readahead is sentivite to the offset
* of read-ahead, so we use original offset here,
- * instead of ras_window_start, which is RPC aligned */
- ras->ras_next_readahead = max(index + 1,
- ras->ras_next_readahead);
- ras->ras_window_start = max(ras->ras_stride_offset,
- ras->ras_window_start);
+ * instead of ras_window_start_idx, which is RPC aligned.
+ */
+ ras->ras_next_readahead_idx = max(index + 1,
+ ras->ras_next_readahead_idx);
+ ras->ras_window_start_idx =
+ max_t(pgoff_t, ras->ras_window_start_idx,
+ ras->ras_stride_offset >> PAGE_SHIFT);
} else {
- if (ras->ras_next_readahead < ras->ras_window_start)
- ras->ras_next_readahead = ras->ras_window_start;
+ if (ras->ras_next_readahead_idx < ras->ras_window_start_idx)
+ ras->ras_next_readahead_idx = ras->ras_window_start_idx;
if (!hit)
- ras->ras_next_readahead = index + 1;
+ ras->ras_next_readahead_idx = index + 1;
}
- RAS_CDEBUG(ras);
- /* Trigger RA in the mmap case where ras_consecutive_requests
- * is not incremented and thus can't be used to trigger RA */
- if (ras->ras_consecutive_pages >= 4 && flags & LL_RAS_MMAP) {
+ if (ras->ras_need_increase_window) {
ras_increase_window(inode, ras, ra);
- /* reset consecutive pages so that the readahead window can
- * grow gradually. */
- ras->ras_consecutive_pages = 0;
- GOTO(out_unlock, 0);
+ ras->ras_need_increase_window = false;
}
- /* Initially reset the stride window offset to next_readahead*/
- if (ras->ras_consecutive_stride_requests == 2 && stride_detect) {
- /**
- * Once stride IO mode is detected, next_readahead should be
- * reset to make sure next_readahead > stride offset
- */
- ras->ras_next_readahead = max(index, ras->ras_next_readahead);
- ras->ras_stride_offset = index;
- ras->ras_window_start = max(index, ras->ras_window_start);
- }
-
- /* The initial ras_window_len is set to the request size. To avoid
- * uselessly reading and discarding pages for random IO the window is
- * only increased once per consecutive request received. */
- if ((ras->ras_consecutive_requests > 1 || stride_detect) &&
- !ras->ras_request_index)
- ras_increase_window(inode, ras, ra);
EXIT;
out_unlock:
- RAS_CDEBUG(ras);
- ras->ras_request_index++;
spin_unlock(&ras->ras_lock);
- return;
}
int ll_writepage(struct page *vmpage, struct writeback_control *wbc)
ENTRY;
if (wbc->range_cyclic) {
- start = mapping->writeback_index << PAGE_SHIFT;
+ start = (loff_t)mapping->writeback_index << PAGE_SHIFT;
end = OBD_OBJECT_EOF;
} else {
start = wbc->range_start;
write_unlock(&fd->fd_lock);
}
-static int ll_io_read_page(const struct lu_env *env, struct cl_io *io,
+int ll_io_read_page(const struct lu_env *env, struct cl_io *io,
struct cl_page *page, struct file *file)
{
struct inode *inode = vvp_object_inode(page->cp_obj);
struct ll_file_data *fd = LUSTRE_FPRIVATE(file);
struct ll_readahead_state *ras = &fd->fd_ras;
struct cl_2queue *queue = &io->ci_queue;
+ struct cl_sync_io *anchor = NULL;
struct vvp_page *vpg;
int rc = 0;
bool uptodate;
cl_page_export(env, page, 1);
cl_page_disown(env, io, page);
} else {
+ anchor = &vvp_env_info(env)->vti_anchor;
+ cl_sync_io_init(anchor, 1);
+ page->cp_sync_io = anchor;
+
cl_2queue_add(queue, page);
}
int rc2;
rc2 = ll_readahead(env, io, &queue->c2_qin, ras,
- uptodate);
+ uptodate, file);
CDEBUG(D_READA, DFID "%d pages read ahead at %lu\n",
PFID(ll_inode2fid(inode)), rc2, vvp_index(vpg));
}
- if (queue->c2_qin.pl_nr > 0)
+ if (queue->c2_qin.pl_nr > 0) {
+ int count = queue->c2_qin.pl_nr;
rc = cl_io_submit_rw(env, io, CRT_READ, queue);
+ if (rc == 0)
+ task_io_account_read(PAGE_SIZE * count);
+ }
- /*
- * Unlock unsent pages in case of error.
- */
+
+ if (anchor != NULL && !cl_page_is_owned(page, io)) { /* have sent */
+ rc = cl_sync_io_wait(env, anchor, 0);
+
+ cl_page_assume(env, io, page);
+ cl_page_list_del(env, &queue->c2_qout, page);
+
+ if (!PageUptodate(cl_page_vmpage(page))) {
+ /* Failed to read a mirror, discard this page so that
+ * new page can be created with new mirror.
+ *
+ * TODO: this is not needed after page reinit
+ * route is implemented */
+ cl_page_discard(env, io, page);
+ }
+ cl_page_disown(env, io, page);
+ }
+
+ /* TODO: discard all pages until page reinit route is implemented */
+ cl_page_list_discard(env, io, &queue->c2_qin);
+
+ /* Unlock unsent read pages in case of error. */
cl_page_list_disown(env, io, &queue->c2_qin);
+
cl_2queue_fini(env, queue);
RETURN(rc);
}
+/*
+ * Possible return value:
+ * 0 no async readahead triggered and fast read could not be used.
+ * 1 no async readahead, but fast read could be used.
+ * 2 async readahead triggered and fast read could be used too.
+ * < 0 on error.
+ */
+static int kickoff_async_readahead(struct file *file, unsigned long pages)
+{
+ struct ll_readahead_work *lrw;
+ struct inode *inode = file_inode(file);
+ struct ll_sb_info *sbi = ll_i2sbi(inode);
+ struct ll_file_data *fd = LUSTRE_FPRIVATE(file);
+ struct ll_readahead_state *ras = &fd->fd_ras;
+ struct ll_ra_info *ra = &sbi->ll_ra_info;
+ unsigned long throttle;
+ pgoff_t start_idx = ras_align(ras, ras->ras_next_readahead_idx);
+ pgoff_t end_idx = start_idx + pages - 1;
+
+ throttle = min(ra->ra_async_pages_per_file_threshold,
+ ra->ra_max_pages_per_file);
+ /*
+ * If this is strided i/o or the window is smaller than the
+ * throttle limit, we do not do async readahead. Otherwise,
+ * we do async readahead, allowing the user thread to do fast i/o.
+ */
+ if (stride_io_mode(ras) || !throttle ||
+ ras->ras_window_pages < throttle)
+ return 0;
+
+ if ((atomic_read(&ra->ra_cur_pages) + pages) > ra->ra_max_pages)
+ return 0;
+
+ if (ras->ras_async_last_readpage_idx == start_idx)
+ return 1;
+
+ /* ll_readahead_work_free() free it */
+ OBD_ALLOC_PTR(lrw);
+ if (lrw) {
+ lrw->lrw_file = get_file(file);
+ lrw->lrw_start_idx = start_idx;
+ lrw->lrw_end_idx = end_idx;
+ spin_lock(&ras->ras_lock);
+ ras->ras_next_readahead_idx = end_idx + 1;
+ ras->ras_async_last_readpage_idx = start_idx;
+ spin_unlock(&ras->ras_lock);
+ ll_readahead_work_add(inode, lrw);
+ } else {
+ return -ENOMEM;
+ }
+
+ return 2;
+}
+
int ll_readpage(struct file *file, struct page *vmpage)
{
struct inode *inode = file_inode(file);
const struct lu_env *env = NULL;
struct cl_io *io = NULL;
struct cl_page *page;
+ struct ll_sb_info *sbi = ll_i2sbi(inode);
int result;
ENTRY;
struct ll_file_data *fd = LUSTRE_FPRIVATE(file);
struct ll_readahead_state *ras = &fd->fd_ras;
struct lu_env *local_env = NULL;
+ unsigned long fast_read_pages =
+ max(RA_REMAIN_WINDOW_MIN, ras->ras_rpc_pages);
struct vvp_page *vpg;
result = -ENODATA;
page = cl_vmpage_page(vmpage, clob);
if (page == NULL) {
unlock_page(vmpage);
+ ll_ra_stats_inc_sbi(sbi, RA_STAT_FAILED_FAST_READ);
RETURN(result);
}
- if (!env) {
- local_env = cl_env_percpu_get();
- env = local_env;
- }
-
vpg = cl2vvp_page(cl_object_page_slice(page->cp_obj, page));
if (vpg->vpg_defer_uptodate) {
enum ras_update_flags flags = LL_RAS_HIT;
/* For fast read, it updates read ahead state only
* if the page is hit in cache because non cache page
* case will be handled by slow read later. */
- ras_update(ll_i2sbi(inode), inode, ras, vvp_index(vpg),
- flags);
+ ras_update(sbi, inode, ras, vvp_index(vpg), flags);
/* avoid duplicate ras_update() call */
vpg->vpg_ra_updated = 1;
/* Check if we can issue a readahead RPC, if that is
* the case, we can't do fast IO because we will need
* a cl_io to issue the RPC. */
- if (ras->ras_window_start + ras->ras_window_len <
- ras->ras_next_readahead + PTLRPC_MAX_BRW_PAGES) {
- /* export the page and skip io stack */
- vpg->vpg_ra_used = 1;
- cl_page_export(env, page, 1);
+ if (ras->ras_window_start_idx + ras->ras_window_pages <
+ ras->ras_next_readahead_idx + fast_read_pages ||
+ kickoff_async_readahead(file, fast_read_pages) > 0)
result = 0;
- }
}
+ if (!env) {
+ local_env = cl_env_percpu_get();
+ env = local_env;
+ }
+
+ /* export the page and skip io stack */
+ if (result == 0) {
+ vpg->vpg_ra_used = 1;
+ cl_page_export(env, page, 1);
+ } else {
+ ll_ra_stats_inc_sbi(sbi, RA_STAT_FAILED_FAST_READ);
+ }
/* release page refcount before unlocking the page to ensure
* the object won't be destroyed in the calling path of
* cl_page_put(). Please see comment in ll_releasepage(). */
LASSERT(page->cp_type == CPT_CACHEABLE);
if (likely(!PageUptodate(vmpage))) {
cl_page_assume(env, io, page);
+
result = ll_io_read_page(env, io, page, file);
} else {
/* Page from a non-object file. */
}
RETURN(result);
}
-
-int ll_page_sync_io(const struct lu_env *env, struct cl_io *io,
- struct cl_page *page, enum cl_req_type crt)
-{
- struct cl_2queue *queue;
- int result;
-
- LASSERT(io->ci_type == CIT_READ || io->ci_type == CIT_WRITE);
-
- queue = &io->ci_queue;
- cl_2queue_init_page(queue, page);
-
- result = cl_io_submit_sync(env, io, crt, queue, 0);
- LASSERT(cl_page_is_owned(page, io));
-
- if (crt == CRT_READ)
- /*
- * in CRT_WRITE case page is left locked even in case of
- * error.
- */
- cl_page_list_disown(env, io, &queue->c2_qin);
- cl_2queue_fini(env, queue);
-
- return result;
-}
git://git.whamcloud.com / fs / lustre-release.git / blobdiff