[PATCH 06/60] staging: lustre: clio: revise readahead to support 16MB IO

From: James Simmons
Date: Sat Jan 28 2017 - 19:16:48 EST


From: Jinshan Xiong <jinshan.xiong@xxxxxxxxx>

Read ahead currently doesn't handle 16MB RPC packets correctly
by assuming the packets are a default size instead of querying
the size. This work adjust the read ahead policy to issue
read ahead RPC by the underlying RPC size.

Signed-off-by: Jinshan Xiong <jinshan.xiong@xxxxxxxxx>
Signed-off-by: Gu Zheng <gzheng@xxxxxxx>
Intel-bug-id: https://jira.hpdd.intel.com/browse/LU-7990
Reviewed-on: http://review.whamcloud.com/19368
Reviewed-by: Andreas Dilger <andreas.dilger@xxxxxxxxx>
Reviewed-by: Li Xi <lixi@xxxxxxx>
Reviewed-by: Oleg Drokin <oleg.drokin@xxxxxxxxx>
Signed-off-by: James Simmons <jsimmons@xxxxxxxxxxxxx>
---
drivers/staging/lustre/lustre/include/cl_object.h | 4 +-
drivers/staging/lustre/lustre/ldlm/ldlm_lib.c | 10 +-
.../staging/lustre/lustre/llite/llite_internal.h | 14 +-
drivers/staging/lustre/lustre/llite/rw.c | 195 ++++++++++-----------
drivers/staging/lustre/lustre/osc/osc_io.c | 3 +-
5 files changed, 114 insertions(+), 112 deletions(-)

diff --git a/drivers/staging/lustre/lustre/include/cl_object.h b/drivers/staging/lustre/lustre/include/cl_object.h
index a1b8301..813e71d 100644
--- a/drivers/staging/lustre/lustre/include/cl_object.h
+++ b/drivers/staging/lustre/lustre/include/cl_object.h
@@ -1452,8 +1452,10 @@ struct cl_read_ahead {
* cra_end is included.
*/
pgoff_t cra_end;
+ /* optimal RPC size for this read, by pages */
+ unsigned long cra_rpc_size;
/*
- * Release routine. If readahead holds resources underneath, this
+ * Release callback. If readahead holds resources underneath, this
* function should be called to release it.
*/
void (*cra_release)(const struct lu_env *env, void *cbdata);
diff --git a/drivers/staging/lustre/lustre/ldlm/ldlm_lib.c b/drivers/staging/lustre/lustre/ldlm/ldlm_lib.c
index 675e25b..95b8c76 100644
--- a/drivers/staging/lustre/lustre/ldlm/ldlm_lib.c
+++ b/drivers/staging/lustre/lustre/ldlm/ldlm_lib.c
@@ -351,13 +351,11 @@ int client_obd_setup(struct obd_device *obddev, struct lustre_cfg *lcfg)
cli->cl_supp_cksum_types = OBD_CKSUM_CRC32;
atomic_set(&cli->cl_resends, OSC_DEFAULT_RESENDS);

- /* This value may be reduced at connect time in
- * ptlrpc_connect_interpret() . We initialize it to only
- * 1MB until we know what the performance looks like.
- * In the future this should likely be increased. LU-1431
+ /*
+ * Set it to possible maximum size. It may be reduced by ocd_brw_size
+ * from OFD after connecting.
*/
- cli->cl_max_pages_per_rpc = min_t(int, PTLRPC_MAX_BRW_PAGES,
- LNET_MTU >> PAGE_SHIFT);
+ cli->cl_max_pages_per_rpc = PTLRPC_MAX_BRW_PAGES;

/*
* set cl_chunkbits default value to PAGE_CACHE_SHIFT,
diff --git a/drivers/staging/lustre/lustre/llite/llite_internal.h b/drivers/staging/lustre/lustre/llite/llite_internal.h
index 2c72177..501957c 100644
--- a/drivers/staging/lustre/lustre/llite/llite_internal.h
+++ b/drivers/staging/lustre/lustre/llite/llite_internal.h
@@ -281,10 +281,8 @@ static inline struct ll_inode_info *ll_i2info(struct inode *inode)
return container_of(inode, struct ll_inode_info, lli_vfs_inode);
}

-/* default to about 40meg of readahead on a given system. That much tied
- * up in 512k readahead requests serviced at 40ms each is about 1GB/s.
- */
-#define SBI_DEFAULT_READAHEAD_MAX (40UL << (20 - PAGE_SHIFT))
+/* default to about 64M of readahead on a given system. */
+#define SBI_DEFAULT_READAHEAD_MAX (64UL << (20 - PAGE_SHIFT))

/* default to read-ahead full files smaller than 2MB on the second read */
#define SBI_DEFAULT_READAHEAD_WHOLE_MAX (2UL << (20 - PAGE_SHIFT))
@@ -321,6 +319,9 @@ struct ll_ra_info {
struct ra_io_arg {
unsigned long ria_start; /* start offset of read-ahead*/
unsigned long ria_end; /* end offset of read-ahead*/
+ unsigned long ria_reserved; /* reserved pages for read-ahead */
+ unsigned long ria_end_min; /* minimum end to cover current read */
+ bool ria_eof; /* reach end of file */
/* If stride read pattern is detected, ria_stoff means where
* stride read is started. Note: for normal read-ahead, the
* value here is meaningless, and also it will not be accessed
@@ -551,6 +552,11 @@ struct ll_readahead_state {
*/
unsigned long ras_window_start, ras_window_len;
/*
+ * Optimal RPC size. It decides how many pages will be sent
+ * for each read-ahead.
+ */
+ unsigned long ras_rpc_size;
+ /*
* Where next read-ahead should start at. This lies within read-ahead
* window. Read-ahead window is read in pieces rather than at once
* because: 1. lustre limits total number of pages under read-ahead by
diff --git a/drivers/staging/lustre/lustre/llite/rw.c b/drivers/staging/lustre/lustre/llite/rw.c
index f10e092..18d3ccb 100644
--- a/drivers/staging/lustre/lustre/llite/rw.c
+++ b/drivers/staging/lustre/lustre/llite/rw.c
@@ -92,25 +92,6 @@ static unsigned long ll_ra_count_get(struct ll_sb_info *sbi,
goto out;
}

- /* If the non-strided (ria_pages == 0) readahead window
- * (ria_start + ret) has grown across an RPC boundary, then trim
- * readahead size by the amount beyond the RPC so it ends on an
- * RPC boundary. If the readahead window is already ending on
- * an RPC boundary (beyond_rpc == 0), or smaller than a full
- * RPC (beyond_rpc < ret) the readahead size is unchanged.
- * The (beyond_rpc != 0) check is skipped since the conditional
- * branch is more expensive than subtracting zero from the result.
- *
- * Strided read is left unaligned to avoid small fragments beyond
- * the RPC boundary from needing an extra read RPC.
- */
- if (ria->ria_pages == 0) {
- long beyond_rpc = (ria->ria_start + ret) % PTLRPC_MAX_BRW_PAGES;
-
- if (/* beyond_rpc != 0 && */ beyond_rpc < ret)
- ret -= beyond_rpc;
- }
-
if (atomic_add_return(ret, &ra->ra_cur_pages) > ra->ra_max_pages) {
atomic_sub(ret, &ra->ra_cur_pages);
ret = 0;
@@ -147,11 +128,12 @@ void ll_ra_stats_inc(struct inode *inode, enum ra_stat which)

#define RAS_CDEBUG(ras) \
CDEBUG(D_READA, \
- "lrp %lu cr %lu cp %lu ws %lu wl %lu nra %lu r %lu ri %lu" \
- "csr %lu sf %lu sp %lu sl %lu\n", \
+ "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, \
ras->ras_consecutive_stride_requests, ras->ras_stride_offset, \
ras->ras_stride_pages, ras->ras_stride_length)
@@ -261,20 +243,6 @@ static int ll_read_ahead_page(const struct lu_env *env, struct cl_io *io,
ria->ria_start, ria->ria_end, ria->ria_stoff, ria->ria_length,\
ria->ria_pages)

-/* Limit this to the blocksize instead of PTLRPC_BRW_MAX_SIZE, since we don't
- * know what the actual RPC size is. If this needs to change, it makes more
- * sense to tune the i_blkbits value for the file based on the OSTs it is
- * striped over, rather than having a constant value for all files here.
- */
-
-/* RAS_INCREASE_STEP should be (1UL << (inode->i_blkbits - PAGE_SHIFT)).
- * Temporarily set RAS_INCREASE_STEP to 1MB. After 4MB RPC is enabled
- * by default, this should be adjusted corresponding with max_read_ahead_mb
- * and max_read_ahead_per_file_mb otherwise the readahead budget can be used
- * up quickly which will affect read performance significantly. See LU-2816
- */
-#define RAS_INCREASE_STEP(inode) (ONE_MB_BRW_SIZE >> PAGE_SHIFT)
-
static inline int stride_io_mode(struct ll_readahead_state *ras)
{
return ras->ras_consecutive_stride_requests > 1;
@@ -345,6 +313,17 @@ static int ria_page_count(struct ra_io_arg *ria)
length);
}

+static unsigned long ras_align(struct ll_readahead_state *ras,
+ unsigned long index,
+ unsigned long *remainder)
+{
+ unsigned long rem = index % ras->ras_rpc_size;
+
+ if (remainder)
+ *remainder = rem;
+ return index - rem;
+}
+
/*Check whether the index is in the defined ra-window */
static int ras_inside_ra_window(unsigned long idx, struct ra_io_arg *ria)
{
@@ -358,42 +337,63 @@ static int ras_inside_ra_window(unsigned long idx, struct ra_io_arg *ria)
ria->ria_length < ria->ria_pages);
}

-static int ll_read_ahead_pages(const struct lu_env *env,
- struct cl_io *io, struct cl_page_list *queue,
- struct ra_io_arg *ria,
- unsigned long *reserved_pages,
- pgoff_t *ra_end)
+static unsigned long
+ll_read_ahead_pages(const struct lu_env *env, struct cl_io *io,
+ struct cl_page_list *queue, struct ll_readahead_state *ras,
+ struct ra_io_arg *ria)
{
struct cl_read_ahead ra = { 0 };
- int rc, count = 0;
+ unsigned long ra_end = 0;
bool stride_ria;
pgoff_t page_idx;
+ int rc;

LASSERT(ria);
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 && *reserved_pages > 0; page_idx++) {
+ page_idx <= ria->ria_end && ria->ria_reserved > 0; page_idx++) {
if (ras_inside_ra_window(page_idx, ria)) {
if (!ra.cra_end || ra.cra_end < page_idx) {
+ unsigned long end;
+
cl_read_ahead_release(env, &ra);

rc = cl_io_read_ahead(env, io, page_idx, &ra);
if (rc < 0)
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,
"object: %p, indcies %lu / %lu\n",
io->ci_obj, ra.cra_end, 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;
+ /* 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;
}

- /* If the page is inside the read-ahead window*/
+ /* If the page is inside the read-ahead window */
rc = ll_read_ahead_page(env, io, queue, page_idx);
- if (!rc) {
- (*reserved_pages)--;
- count++;
- }
+ if (rc < 0)
+ break;
+
+ ra_end = page_idx;
+ if (!rc)
+ ria->ria_reserved--;
} 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
@@ -420,8 +420,7 @@ static int ll_read_ahead_pages(const struct lu_env *env,
}
cl_read_ahead_release(env, &ra);

- *ra_end = page_idx;
- return count;
+ return ra_end;
}

static int ll_readahead(const struct lu_env *env, struct cl_io *io,
@@ -431,7 +430,7 @@ static int ll_readahead(const struct lu_env *env, struct cl_io *io,
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, reserved;
+ unsigned long len, mlen = 0;
pgoff_t ra_end, start = 0, end = 0;
struct inode *inode;
struct ra_io_arg *ria = &lti->lti_ria;
@@ -478,29 +477,15 @@ static int ll_readahead(const struct lu_env *env, struct cl_io *io,
end < vio->vui_ra_start + vio->vui_ra_count - 1)
end = vio->vui_ra_start + vio->vui_ra_count - 1;

- if (end != 0) {
- unsigned long rpc_boundary;
- /*
- * Align RA window to an optimal boundary.
- *
- * XXX This would be better to align to cl_max_pages_per_rpc
- * instead of PTLRPC_MAX_BRW_PAGES, because the RPC size may
- * be aligned to the RAID stripe size in the future and that
- * is more important than the RPC size.
- */
- /* Note: we only trim the RPC, instead of extending the RPC
- * to the boundary, so to avoid reading too much pages during
- * random reading.
- */
- rpc_boundary = (end + 1) & (~(PTLRPC_MAX_BRW_PAGES - 1));
- if (rpc_boundary > 0)
- rpc_boundary--;
-
- if (rpc_boundary > start)
- end = rpc_boundary;
+ if (end) {
+ unsigned long end_index;

/* Truncate RA window to end of file */
- end = min(end, (unsigned long)((kms - 1) >> PAGE_SHIFT));
+ end_index = (unsigned long)((kms - 1) >> PAGE_SHIFT);
+ if (end_index <= end) {
+ end = end_index;
+ ria->ria_eof = true;
+ }

ras->ras_next_readahead = max(end, end + 1);
RAS_CDEBUG(ras);
@@ -535,28 +520,31 @@ static int ll_readahead(const struct lu_env *env, struct cl_io *io,
/* 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 */
- start = ria->ria_start & (PTLRPC_MAX_BRW_PAGES - 1);
- mlen = min(mlen, PTLRPC_MAX_BRW_PAGES - start);
+ ras_align(ras, ria->ria_start, &remainder);
+ mlen = min(mlen, ras->ras_rpc_size - remainder);
+ ria->ria_end_min = ria->ria_start + mlen;
}

- reserved = ll_ra_count_get(ll_i2sbi(inode), ria, len, mlen);
- if (reserved < len)
+ ria->ria_reserved = ll_ra_count_get(ll_i2sbi(inode), ria, len, mlen);
+ if (ria->ria_reserved < len)
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",
- reserved, len, mlen,
+ ria->ria_reserved, len, mlen,
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, ria, &reserved, &ra_end);
+ ra_end = ll_read_ahead_pages(env, io, queue, ras, ria);

- if (reserved != 0)
- ll_ra_count_put(ll_i2sbi(inode), reserved);
+ if (ria->ria_reserved)
+ ll_ra_count_put(ll_i2sbi(inode), ria->ria_reserved);

- if (ra_end == end + 1 && ra_end == (kms >> PAGE_SHIFT))
+ if (ra_end == end && ra_end == (kms >> PAGE_SHIFT))
ll_ra_stats_inc(inode, RA_STAT_EOF);

/* if we didn't get to the end of the region we reserved from
@@ -568,13 +556,13 @@ static int ll_readahead(const struct lu_env *env, struct cl_io *io,
CDEBUG(D_READA, "ra_end = %lu end = %lu stride end = %lu pages = %d\n",
ra_end, end, ria->ria_end, ret);

- if (ra_end != end + 1) {
+ if (ra_end > 0 && ra_end != end) {
ll_ra_stats_inc(inode, RA_STAT_FAILED_REACH_END);
spin_lock(&ras->ras_lock);
- if (ra_end < ras->ras_next_readahead &&
+ if (ra_end <= ras->ras_next_readahead &&
index_in_window(ra_end, ras->ras_window_start, 0,
ras->ras_window_len)) {
- ras->ras_next_readahead = ra_end;
+ ras->ras_next_readahead = ra_end + 1;
RAS_CDEBUG(ras);
}
spin_unlock(&ras->ras_lock);
@@ -586,7 +574,7 @@ static int ll_readahead(const struct lu_env *env, struct cl_io *io,
static void ras_set_start(struct inode *inode, struct ll_readahead_state *ras,
unsigned long index)
{
- ras->ras_window_start = index & (~(RAS_INCREASE_STEP(inode) - 1));
+ ras->ras_window_start = ras_align(ras, index, NULL);
}

/* called with the ras_lock held or from places where it doesn't matter */
@@ -615,6 +603,7 @@ static void ras_stride_reset(struct ll_readahead_state *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_requests = 0;
}
@@ -719,12 +708,15 @@ static void ras_increase_window(struct inode *inode,
* but current clio architecture does not support retrieve such
* information from lower layer. FIXME later
*/
- if (stride_io_mode(ras))
- ras_stride_increase_window(ras, ra, RAS_INCREASE_STEP(inode));
- else
- ras->ras_window_len = min(ras->ras_window_len +
- RAS_INCREASE_STEP(inode),
- ra->ra_max_pages_per_file);
+ if (stride_io_mode(ras)) {
+ ras_stride_increase_window(ras, ra, ras->ras_rpc_size);
+ } else {
+ unsigned long wlen;
+
+ wlen = min(ras->ras_window_len + ras->ras_rpc_size,
+ ra->ra_max_pages_per_file);
+ ras->ras_window_len = ras_align(ras, wlen, NULL);
+ }
}

static void ras_update(struct ll_sb_info *sbi, struct inode *inode,
@@ -852,6 +844,8 @@ static void ras_update(struct ll_sb_info *sbi, struct inode *inode,
* instead of ras_window_start, which is RPC aligned
*/
ras->ras_next_readahead = max(index, ras->ras_next_readahead);
+ ras->ras_window_start = max(ras->ras_stride_offset,
+ ras->ras_window_start);
} else {
if (ras->ras_next_readahead < ras->ras_window_start)
ras->ras_next_readahead = ras->ras_window_start;
@@ -881,7 +875,7 @@ static void ras_update(struct ll_sb_info *sbi, struct inode *inode,
*/
ras->ras_next_readahead = max(index, ras->ras_next_readahead);
ras->ras_stride_offset = index;
- ras->ras_window_len = RAS_INCREASE_STEP(inode);
+ ras->ras_window_start = max(index, ras->ras_window_start);
}

/* The initial ras_window_len is set to the request size. To avoid
@@ -1098,38 +1092,39 @@ static int ll_io_read_page(const struct lu_env *env, struct cl_io *io,
struct cl_2queue *queue = &io->ci_queue;
struct ll_sb_info *sbi = ll_i2sbi(inode);
struct vvp_page *vpg;
+ bool uptodate;
int rc = 0;

vpg = cl2vvp_page(cl_object_page_slice(page->cp_obj, page));
+ uptodate = vpg->vpg_defer_uptodate;
+
if (sbi->ll_ra_info.ra_max_pages_per_file > 0 &&
sbi->ll_ra_info.ra_max_pages > 0) {
struct vvp_io *vio = vvp_env_io(env);
enum ras_update_flags flags = 0;

- if (vpg->vpg_defer_uptodate)
+ if (uptodate)
flags |= LL_RAS_HIT;
if (!vio->vui_ra_valid)
flags |= LL_RAS_MMAP;
ras_update(sbi, inode, ras, vvp_index(vpg), flags);
}

- if (vpg->vpg_defer_uptodate) {
+ cl_2queue_init(queue);
+ if (uptodate) {
vpg->vpg_ra_used = 1;
cl_page_export(env, page, 1);
+ cl_page_disown(env, io, page);
+ } else {
+ cl_page_list_add(&queue->c2_qin, page);
}

- cl_2queue_init(queue);
- /*
- * Add page into the queue even when it is marked uptodate above.
- * this will unlock it automatically as part of cl_page_list_disown().
- */
- cl_page_list_add(&queue->c2_qin, page);
if (sbi->ll_ra_info.ra_max_pages_per_file > 0 &&
sbi->ll_ra_info.ra_max_pages > 0) {
int rc2;

rc2 = ll_readahead(env, io, &queue->c2_qin, ras,
- vpg->vpg_defer_uptodate);
+ uptodate);
CDEBUG(D_READA, DFID "%d pages read ahead at %lu\n",
PFID(ll_inode2fid(inode)), rc2, vvp_index(vpg));
}
diff --git a/drivers/staging/lustre/lustre/osc/osc_io.c b/drivers/staging/lustre/lustre/osc/osc_io.c
index 9402dfc..7e5cd3a 100644
--- a/drivers/staging/lustre/lustre/osc/osc_io.c
+++ b/drivers/staging/lustre/lustre/osc/osc_io.c
@@ -99,6 +99,7 @@ static int osc_io_read_ahead(const struct lu_env *env,
ldlm_lock_decref(&lockh, dlmlock->l_req_mode);
}

+ ra->cra_rpc_size = osc_cli(osc)->cl_max_pages_per_rpc;
ra->cra_end = cl_index(osc2cl(osc),
dlmlock->l_policy_data.l_extent.end);
ra->cra_release = osc_read_ahead_release;
@@ -138,7 +139,7 @@ static int osc_io_submit(const struct lu_env *env,

LASSERT(qin->pl_nr > 0);

- CDEBUG(D_CACHE, "%d %d\n", qin->pl_nr, crt);
+ CDEBUG(D_CACHE | D_READA, "%d %d\n", qin->pl_nr, crt);

osc = cl2osc(ios->cis_obj);
cli = osc_cli(osc);
--
1.8.3.1