Skip to content

Commit

Permalink
DAOS-14484 cart: Implement per-context inflight queue (#13202)
Browse files Browse the repository at this point in the history 
- D_QUOTA_RPCS envariable added. When set, limits the number of RPCs on a wire being sent out by the process.
- RPCs that exceed quota limit (if set), will now be queued by the sender
- Quota support code added to handle and track resources

Signed-off-by: Alexander A Oganezov <alexander.a.oganezov@intel.com>
  • Loading branch information
@frostedcmos @jolivier23
frostedcmos authored and jolivier23 committed Feb 28, 2024
1 parent 0230b28 commit 9cd53a9
Show file tree
Hide file tree
Showing 11 changed files with 318 additions and 48 deletions.
6 changes: 6 additions & 0 deletions src/cart/README.env
View file
Original file line number Diff line number Diff line change
Expand Up @@ -132,6 +132,12 @@ This file lists the environment variables used in CaRT.
It its value exceed 256, then will use 256 for flow control.
Set it to zero means disable the flow control in cart.

. D_QUOTA_RPCS
Set it as the max number of per-context inflight RPCs that a sender will send
onto a wire. Quota on each context is independent of each other.
If it is not set the default value of 64 is used.
Setting it to 0 disables quota

. CRT_CTX_SHARE_ADDR
Set it to non-zero to make all the contexts share one network address, in
this case CaRT will create one SEP and each context maps to one tx/rx
Expand Down
247 changes: 213 additions & 34 deletions src/cart/crt_context.c
View file
Original file line number Diff line number Diff line change
Expand Up @@ -11,6 +11,11 @@
#include "crt_internal.h"

static void crt_epi_destroy(struct crt_ep_inflight *epi);
static int context_quotas_init(crt_context_t crt_ctx);
static int context_quotas_finalize(crt_context_t crt_ctx);

static inline int get_quota_resource(crt_context_t crt_ctx, crt_quota_type_t quota);
static inline void put_quota_resource(crt_context_t crt_ctx, crt_quota_type_t quota);

static struct crt_ep_inflight *
epi_link2ptr(d_list_t *rlink)
Expand Down Expand Up @@ -141,6 +146,13 @@ crt_context_init(crt_context_t crt_ctx)
if (rc != 0)
D_GOTO(out, rc);

rc = D_MUTEX_INIT(&ctx->cc_quotas.mutex, NULL);
if (rc != 0) {
D_MUTEX_DESTROY(&ctx->cc_mutex);
D_GOTO(out, rc);
}

D_INIT_LIST_HEAD(&ctx->cc_quotas.rpc_waitq);
D_INIT_LIST_HEAD(&ctx->cc_link);

/* create timeout binheap */
Expand All @@ -162,6 +174,8 @@ crt_context_init(crt_context_t crt_ctx)
D_GOTO(out_binheap_destroy, rc);
}

rc = context_quotas_init(crt_ctx);

D_GOTO(out, rc);

out_binheap_destroy:
Expand Down Expand Up @@ -684,10 +698,17 @@ crt_context_destroy(crt_context_t crt_ctx, int force)
D_GOTO(out, rc = -DER_UNINIT);
}

rc = context_quotas_finalize(crt_ctx);
if (rc) {
DL_ERROR(rc, "context_quotas_finalize() failed");
if (!force)
D_GOTO(out, rc);
}

ctx = crt_ctx;
rc = crt_grp_ctx_invalid(ctx, false /* locked */);
if (rc) {
D_ERROR("crt_grp_ctx_invalid failed, rc: %d.\n", rc);
DL_ERROR(rc, "crt_grp_ctx_invalid() failed");
if (!force)
D_GOTO(out, rc);
}
Expand Down Expand Up @@ -1167,6 +1188,7 @@ crt_context_req_track(struct crt_rpc_priv *rpc_priv)
d_list_t *rlink;
d_rank_t ep_rank;
int rc = 0;
int quota_rc = 0;
struct crt_grp_priv *grp_priv;

D_ASSERT(crt_ctx != NULL);
Expand All @@ -1177,6 +1199,9 @@ crt_context_req_track(struct crt_rpc_priv *rpc_priv)
D_GOTO(out, rc = CRT_REQ_TRACK_IN_INFLIGHQ);
}

/* check inflight quota. if exceeded, queue this rpc */
quota_rc = get_quota_resource(rpc_priv->crp_pub.cr_ctx, CRT_QUOTA_RPCS);

grp_priv = crt_grp_pub2priv(rpc_priv->crp_pub.cr_ep.ep_grp);
ep_rank = crt_grp_priv_get_primary_rank(grp_priv,
rpc_priv->crp_pub.cr_ep.ep_rank);
Expand Down Expand Up @@ -1228,15 +1253,16 @@ crt_context_req_track(struct crt_rpc_priv *rpc_priv)
rpc_priv->crp_epi = epi;
RPC_ADDREF(rpc_priv);

if (crt_gdata.cg_credit_ep_ctx != 0 &&
if (quota_rc == -DER_QUOTA_LIMIT) {
epi->epi_req_num++;
rpc_priv->crp_state = RPC_STATE_QUEUED;
rc = CRT_REQ_TRACK_IN_WAITQ;
} else if (crt_gdata.cg_credit_ep_ctx != 0 &&
(epi->epi_req_num - epi->epi_reply_num) >= crt_gdata.cg_credit_ep_ctx) {
if (rpc_priv->crp_opc_info->coi_queue_front) {
d_list_add(&rpc_priv->crp_epi_link,
&epi->epi_req_waitq);
} else {
d_list_add_tail(&rpc_priv->crp_epi_link,
&epi->epi_req_waitq);
}
if (rpc_priv->crp_opc_info->coi_queue_front)
d_list_add(&rpc_priv->crp_epi_link, &epi->epi_req_waitq);
else
d_list_add_tail(&rpc_priv->crp_epi_link, &epi->epi_req_waitq);

epi->epi_req_wait_num++;
rpc_priv->crp_state = RPC_STATE_QUEUED;
Expand All @@ -1246,13 +1272,11 @@ crt_context_req_track(struct crt_rpc_priv *rpc_priv)
rc = crt_req_timeout_track(rpc_priv);
D_MUTEX_UNLOCK(&crt_ctx->cc_mutex);
if (rc == 0) {
d_list_add_tail(&rpc_priv->crp_epi_link,
&epi->epi_req_q);
d_list_add_tail(&rpc_priv->crp_epi_link, &epi->epi_req_q);
epi->epi_req_num++;
rc = CRT_REQ_TRACK_IN_INFLIGHQ;
} else {
RPC_ERROR(rpc_priv,
"crt_req_timeout_track failed, rc: %d.\n", rc);
RPC_ERROR(rpc_priv, "crt_req_timeout_track failed, rc: %d.\n", rc);
/* roll back the addref above */
RPC_DECREF(rpc_priv);
}
Expand All @@ -1264,6 +1288,10 @@ crt_context_req_track(struct crt_rpc_priv *rpc_priv)
/* reference taken by d_hash_rec_find or "epi->epi_ref = 1" above */
D_MUTEX_LOCK(&crt_ctx->cc_mutex);
d_hash_rec_decref(&crt_ctx->cc_epi_table, &epi->epi_link);

if (quota_rc == -DER_QUOTA_LIMIT)
d_list_add_tail(&rpc_priv->crp_waitq_link, &crt_ctx->cc_quotas.rpc_waitq);

D_MUTEX_UNLOCK(&crt_ctx->cc_mutex);

out:
Expand All @@ -1280,9 +1308,10 @@ credits_available(struct crt_ep_inflight *epi)
{
int64_t inflight = epi->epi_req_num - epi->epi_reply_num;

D_ASSERTF(inflight >= 0 && inflight <= crt_gdata.cg_credit_ep_ctx,
"req_num=%ld reply_num=%ld credit_ep_ctx=%u\n", epi->epi_req_num,
epi->epi_reply_num, crt_gdata.cg_credit_ep_ctx);
/* TODO: inflight right now includes items queued in quota waitq, and can exceed credit limit */
if (inflight > crt_gdata.cg_credit_ep_ctx)
return 0;

return crt_gdata.cg_credit_ep_ctx - inflight;
}

Expand Down Expand Up @@ -1324,6 +1353,7 @@ crt_context_req_untrack_internal(struct crt_rpc_priv *rpc_priv)
} else {/* RPC_CANCELED or RPC_INITED or RPC_TIMEOUT */
epi->epi_req_num--;
}

D_ASSERT(epi->epi_req_num >= epi->epi_reply_num);

D_MUTEX_UNLOCK(&epi->epi_mutex);
Expand All @@ -1340,6 +1370,27 @@ crt_context_req_untrack_internal(struct crt_rpc_priv *rpc_priv)
RPC_DECREF(rpc_priv);
}

static void
dispatch_rpc(struct crt_rpc_priv *rpc) {
int rc;

D_ASSERTF(rpc != NULL, "rpc is NULL\n");

crt_rpc_lock(rpc);

rc = crt_req_send_internal(rpc);
if (rc == 0) {
crt_rpc_unlock(rpc);
} else {
RPC_ADDREF(rpc);
RPC_ERROR(rpc, "crt_req_send_internal failed, rc: %d\n", rc);
rpc->crp_state = RPC_STATE_INITED;
crt_context_req_untrack_internal(rpc);
/* for error case here */
crt_rpc_complete_and_unlock(rpc, rc);
}
}

void
crt_context_req_untrack(struct crt_rpc_priv *rpc_priv)
{
Expand All @@ -1351,17 +1402,26 @@ crt_context_req_untrack(struct crt_rpc_priv *rpc_priv)

D_ASSERT(crt_ctx != NULL);

if (rpc_priv->crp_pub.cr_opc == CRT_OPC_URI_LOOKUP) {
RPC_TRACE(DB_NET, rpc_priv, "bypass untracking for URI_LOOKUP.\n");
if (rpc_priv->crp_pub.cr_opc == CRT_OPC_URI_LOOKUP)
return;
}

epi = rpc_priv->crp_epi;
D_ASSERT(epi != NULL);

/* Dispatch one rpc from wait_q if any or return resource back */
D_MUTEX_LOCK(&crt_ctx->cc_mutex);
tmp_rpc = d_list_pop_entry(&crt_ctx->cc_quotas.rpc_waitq,
struct crt_rpc_priv, crp_waitq_link);
D_MUTEX_UNLOCK(&crt_ctx->cc_mutex);

if (tmp_rpc != NULL)
dispatch_rpc(tmp_rpc);
else
put_quota_resource(rpc_priv->crp_pub.cr_ctx, CRT_QUOTA_RPCS);

crt_context_req_untrack_internal(rpc_priv);

/* done if flow control disabled */
/* done if ep credit flow control is disabled */
if (crt_gdata.cg_credit_ep_ctx == 0)
return;

Expand Down Expand Up @@ -1408,20 +1468,8 @@ crt_context_req_untrack(struct crt_rpc_priv *rpc_priv)
D_MUTEX_UNLOCK(&epi->epi_mutex);

/* re-submit the rpc req */
while ((tmp_rpc = d_list_pop_entry(&submit_list, struct crt_rpc_priv, crp_tmp_link))) {
crt_rpc_lock(tmp_rpc);
rc = crt_req_send_internal(tmp_rpc);
if (rc == 0) {
crt_rpc_unlock(tmp_rpc);
} else {
RPC_ADDREF(tmp_rpc);
RPC_ERROR(tmp_rpc, "crt_req_send_internal failed, rc: %d\n", rc);
tmp_rpc->crp_state = RPC_STATE_INITED;
crt_context_req_untrack_internal(tmp_rpc);
/* for error case here */
crt_rpc_complete_and_unlock(tmp_rpc, rc);
}
}
while ((tmp_rpc = d_list_pop_entry(&submit_list, struct crt_rpc_priv, crp_tmp_link)))
dispatch_rpc(tmp_rpc);
}

/* TODO: Need per-provider call */
Expand Down Expand Up @@ -1910,3 +1958,134 @@ crt_req_force_completion(struct crt_rpc_priv *rpc_priv)
crt_req_timeout_track(rpc_priv);
D_MUTEX_UNLOCK(&crt_ctx->cc_mutex);
}

static int
context_quotas_init(crt_context_t crt_ctx)
{
struct crt_context *ctx = crt_ctx;
struct crt_quotas *quotas;
int rc = 0;

if (ctx == NULL) {
D_ERROR("NULL context\n");
D_GOTO(out, rc = -DER_INVAL);
}

quotas = &ctx->cc_quotas;

quotas->limit[CRT_QUOTA_RPCS] = crt_gdata.cg_rpc_quota;
quotas->current[CRT_QUOTA_RPCS] = 0;
quotas->enabled[CRT_QUOTA_RPCS] = crt_gdata.cg_rpc_quota > 0 ? true : false;
out:
return rc;
}

static int
context_quotas_finalize(crt_context_t crt_ctx)
{
struct crt_context *ctx = crt_ctx;

if (ctx == NULL) {
D_ERROR("NULL context\n");
return -DER_INVAL;
}

for (int i = 0; i < CRT_QUOTA_COUNT; i++)
ctx->cc_quotas.enabled[i] = false;

return DER_SUCCESS;
}

int
crt_context_quota_limit_set(crt_context_t crt_ctx, crt_quota_type_t quota, int value)
{
struct crt_context *ctx = crt_ctx;
int rc = 0;

if (ctx == NULL) {
D_ERROR("NULL context\n");
D_GOTO(out, rc = -DER_INVAL);
}

if (quota < 0 || quota >= CRT_QUOTA_COUNT) {
D_ERROR("Invalid quota %d passed\n", quota);
D_GOTO(out, rc = -DER_INVAL);
}

D_MUTEX_LOCK(&ctx->cc_quotas.mutex);
ctx->cc_quotas.limit[quota] = value;
D_MUTEX_UNLOCK(&ctx->cc_quotas.mutex);

out:
return rc;
}

int
crt_context_quota_limit_get(crt_context_t crt_ctx, crt_quota_type_t quota, int *value)
{
struct crt_context *ctx = crt_ctx;
int rc = 0;

if (ctx == NULL) {
D_ERROR("NULL context\n");
D_GOTO(out, rc = -DER_INVAL);
}

if (quota < 0 || quota >= CRT_QUOTA_COUNT) {
D_ERROR("Invalid quota %d passed\n", quota);
D_GOTO(out, rc = -DER_INVAL);
}

if (value == NULL) {
D_ERROR("NULL value\n");
D_GOTO(out, rc = -DER_INVAL);
}

*value = ctx->cc_quotas.limit[quota];

out:
return rc;
}

static inline int
get_quota_resource(crt_context_t crt_ctx, crt_quota_type_t quota)
{
struct crt_context *ctx = crt_ctx;
int rc = 0;

D_ASSERTF(ctx != NULL, "NULL context\n");
D_ASSERTF(quota >= 0 && quota < CRT_QUOTA_COUNT, "Invalid quota\n");

/* If quotas not enabled or unlimited quota */
if (!ctx->cc_quotas.enabled[quota] || ctx->cc_quotas.limit[quota] == 0)
return 0;

/* It's ok if we go slightly above quota in a corner case, but avoid locks */
if (ctx->cc_quotas.current[quota] < ctx->cc_quotas.limit[quota]) {
atomic_fetch_add(&ctx->cc_quotas.current[quota], 1);
} else {
D_DEBUG(DB_TRACE, "Quota limit (%d) reached for quota_type=%d\n",
ctx->cc_quotas.limit[quota], quota);
rc = -DER_QUOTA_LIMIT;
}

return rc;
}

static inline void
put_quota_resource(crt_context_t crt_ctx, crt_quota_type_t quota)
{
struct crt_context *ctx = crt_ctx;

D_ASSERTF(ctx != NULL, "NULL context\n");
D_ASSERTF(quota >= 0 && quota < CRT_QUOTA_COUNT, "Invalid quota\n");

/* If quotas not enabled or unlimited quota */
if (!ctx->cc_quotas.enabled[quota] || ctx->cc_quotas.limit[quota] == 0)
return;

D_ASSERTF(ctx->cc_quotas.current[quota] > 0, "Invalid current limit");
atomic_fetch_sub(&ctx->cc_quotas.current[quota], 1);

return;
}
2 changes: 1 addition & 1 deletion src/cart/crt_hg.c
View file
Original file line number Diff line number Diff line change
Expand Up @@ -1392,7 +1392,7 @@ crt_hg_req_send_cb(const struct hg_cb_info *hg_cbinfo)
void
crt_hg_req_send(struct crt_rpc_priv *rpc_priv)
{
hg_return_t hg_ret;
hg_return_t hg_ret;

D_ASSERT(rpc_priv != NULL);

Expand Down
Loading

0 comments on commit 9cd53a9

Please sign in to comment.