From: Prudvi Deti <prudvi.deti@intel.com>
To: dev@dpdk.org
Cc: bruce.richardson@intel.com, orika@nvidia.com,
nirint.shah@intel.com, prudvi.deti@intel.com
Subject: [RFC 4/7] regex/hs: add enqueue and dequeue burst paths
Date: Fri, 28 Aug 2026 11:05:53 +0530 [thread overview]
Message-ID: <20260828053629.1224611-5-prudvi.deti@intel.com> (raw)
In-Reply-To: <20260828053629.1224611-1-prudvi.deti@intel.com>
Implement fast-path enqueue and dequeue burst functions.
Enqueue calls hs_scan() synchronously per operation using per-QP
scratch space. Multi-segment mbufs are linearized before scanning.
Completed ops are stored in a bounded ring for dequeue.
Thread-safety: single-producer/single-consumer per queue pair.
Each QP must be used by exactly one lcore.
Signed-off-by: Prudvi Deti <prudvi.deti@intel.com>
---
drivers/regex/hs/hs_regex.c | 247 ++++++++++++++++++++++++++++++++++--
1 file changed, 234 insertions(+), 13 deletions(-)
diff --git a/drivers/regex/hs/hs_regex.c b/drivers/regex/hs/hs_regex.c
index 116aff9..e84259e 100644
--- a/drivers/regex/hs/hs_regex.c
+++ b/drivers/regex/hs/hs_regex.c
@@ -39,15 +39,60 @@ RTE_LOG_REGISTER_DEFAULT(hs_regex_logtype, NOTICE);
#define HS_LOG(level, ...) \
RTE_LOG_LINE(level, HS_REGEX, __VA_ARGS__)
+/* Match callback context */
+struct hs_match_ctx {
+ struct rte_regex_ops *op;
+ uint16_t max_matches;
+ uint64_t total_matches; /* 64-bit counter for accurate tracking */
+};
+
static int
hs_regex_rule_db_import(struct rte_regexdev *dev, const char *rule_db,
uint32_t rule_db_len);
+static int
+hs_match_cb(unsigned int id, unsigned long long from,
+ unsigned long long to, unsigned int flags __rte_unused,
+ void *context)
+{
+ struct hs_match_ctx *ctx = (struct hs_match_ctx *)context;
+ struct rte_regex_ops *op;
+
+ if (unlikely(ctx == NULL))
+ return 1;
+
+ op = ctx->op;
+ if (unlikely(op == NULL))
+ return 1;
+
+ ctx->total_matches++;
+ op->nb_actual_matches++;
+
+ if (op->nb_matches < ctx->max_matches) {
+ struct rte_regexdev_match *m = &op->matches[op->nb_matches];
+
+ m->rule_id = id;
+ m->start_offset = (uint16_t)from;
+ m->len = (uint16_t)(to - from);
+ op->nb_matches++;
+
+ } else {
+ /* Match list full; report truncation while keeping actual count. */
+ op->rsp_flags |= RTE_REGEX_OPS_RSP_MAX_MATCH_F;
+ }
+
+ return 0;
+}
+
/* Device Info */
static int
-hs_regex_info_get(struct rte_regexdev *dev __rte_unused,
- struct rte_regexdev_info *info)
+hs_regex_info_get(struct rte_regexdev *dev, struct rte_regexdev_info *info)
{
+ (void)dev;
+
+ if (info == NULL)
+ return -EINVAL;
+
info->driver_name = HS_REGEX_DRIVER_NAME;
info->dev = NULL;
info->max_matches = UINT16_MAX;
@@ -79,6 +124,7 @@ hs_regex_configure(struct rte_regexdev *dev,
if (priv == NULL)
return -EINVAL;
+ /* Reconfigure is not allowed while running. */
if (priv->dev_state == HS_REGEX_DEV_STARTED) {
HS_LOG(ERR, "Cannot configure while device is started");
return -EBUSY;
@@ -95,6 +141,7 @@ hs_regex_configure(struct rte_regexdev *dev,
UINT16_MAX;
priv->nb_groups = cfg->nb_groups ? cfg->nb_groups : 1;
+ /* Reconfigure replaces rules, database, and queue resources. */
if (priv->rules) {
uint32_t i;
for (i = 0; i < priv->nb_rules; i++)
@@ -133,6 +180,7 @@ hs_regex_configure(struct rte_regexdev *dev,
HS_LOG(INFO, "Configured: %u queue pairs, max_matches=%u",
priv->nb_queue_pairs, priv->max_matches);
+ /* Configuration complete. */
priv->dev_state = HS_REGEX_DEV_CONFIGURED;
if (cfg->rule_db != NULL && cfg->rule_db_len > 0) {
@@ -140,9 +188,11 @@ hs_regex_configure(struct rte_regexdev *dev,
cfg->rule_db_len);
if (ret < 0) {
HS_LOG(ERR, "Failed to import rule DB in configure");
+ /* Roll back QP allocation on import failure. */
rte_free(priv->qps);
priv->qps = NULL;
priv->nb_queue_pairs = 0;
+ /* Revert state since configure failed */
priv->dev_state = HS_REGEX_DEV_CREATED;
return ret;
}
@@ -184,6 +234,10 @@ hs_regex_qp_setup(struct rte_regexdev *dev, uint16_t qp_id,
nb_desc = (qp_conf && qp_conf->nb_desc) ? qp_conf->nb_desc :
HS_REGEX_DEFAULT_NB_DESC;
+ /*
+ * The ring uses modulo arithmetic on head/tail.
+ * Keep descriptor count as power-of-two for predictable wrap behavior.
+ */
if (nb_desc == 0 || (nb_desc & (nb_desc - 1)) != 0) {
uint16_t orig = nb_desc;
uint32_t aligned = rte_align32pow2(nb_desc ? nb_desc : 1);
@@ -195,11 +249,13 @@ hs_regex_qp_setup(struct rte_regexdev *dev, uint16_t qp_id,
qp_id, orig, nb_desc);
}
+ /* Re-setup replaces previous ring allocation. */
if (qp->ops) {
rte_free(qp->ops);
qp->ops = NULL;
}
+ /* Scratch is recreated on re-setup. */
if (qp->scratch) {
hs_free_scratch(qp->scratch);
qp->scratch = NULL;
@@ -253,6 +309,7 @@ hs_regex_rule_db_update(struct rte_regexdev *dev,
if (nb_rules == 0)
return 0;
+ /* Lazy-init hash table for O(1) duplicate rule_id detection. */
if (!priv->rule_id_hash) {
struct rte_hash_parameters hp = {
.name = "hs_rule_ids",
@@ -271,6 +328,7 @@ hs_regex_rule_db_update(struct rte_regexdev *dev,
if (rules[i].op == RTE_REGEX_RULE_OP_ADD) {
uint32_t idx;
+ /* Reject empty or NULL patterns. */
if (!rules[i].pcre_rule || rules[i].pcre_rule_len == 0) {
HS_LOG(ERR, "Rule %u: NULL or empty pattern",
rules[i].rule_id);
@@ -278,6 +336,7 @@ hs_regex_rule_db_update(struct rte_regexdev *dev,
return i;
}
+ /* Keep rule_id unique for deterministic match reporting. */
if (priv->rule_id_hash &&
rte_hash_lookup(priv->rule_id_hash,
&rules[i].rule_id) >= 0) {
@@ -437,6 +496,7 @@ hs_regex_rule_db_compile_activate(struct rte_regexdev *dev)
if (priv->rules[i].rule_flags & RTE_REGEX_PCRE_RULE_UTF_F)
flags[i] |= HS_FLAG_UTF8;
+ /* Extended parameters */
ext[i].flags = 0;
if (priv->rules[i].min_offset) {
ext[i].flags |= HS_EXT_FLAG_MIN_OFFSET;
@@ -472,6 +532,7 @@ hs_regex_rule_db_compile_activate(struct rte_regexdev *dev)
return -EINVAL;
}
+ /* Allocate scratch per queue pair for scanning. */
for (i = 0; i < priv->nb_queue_pairs; i++) {
struct hs_regex_qp *qp = &priv->qps[i];
@@ -484,6 +545,7 @@ hs_regex_rule_db_compile_activate(struct rte_regexdev *dev)
uint32_t j;
HS_LOG(ERR, "Scratch alloc failed for qp %u", i);
+ /* Partial failure: unwind previous scratch allocations. */
for (j = 0; j < i; j++) {
if (priv->qps[j].scratch) {
hs_free_scratch(priv->qps[j].scratch);
@@ -502,6 +564,10 @@ hs_regex_rule_db_compile_activate(struct rte_regexdev *dev)
return 0;
}
+/*
+ * Import a prebuilt serialized Hyperscan database.
+ * The buffer must be produced by hs_serialize_database().
+ */
static int
hs_regex_rule_db_import(struct rte_regexdev *dev, const char *rule_db,
uint32_t rule_db_len)
@@ -527,17 +593,20 @@ hs_regex_rule_db_import(struct rte_regexdev *dev, const char *rule_db,
return -EINVAL;
}
+ /* Free existing database */
if (priv->db) {
hs_free_database(priv->db);
priv->db = NULL;
}
+ /* Deserialize the precompiled database */
err = hs_deserialize_database(rule_db, (size_t)rule_db_len, &priv->db);
if (err != HS_SUCCESS) {
HS_LOG(ERR, "hs_deserialize_database failed (error %d)", err);
return -EINVAL;
}
+ /* Imported DB also requires per-QP scratch. */
for (i = 0; i < priv->nb_queue_pairs; i++) {
struct hs_regex_qp *qp = &priv->qps[i];
@@ -551,6 +620,7 @@ hs_regex_rule_db_import(struct rte_regexdev *dev, const char *rule_db,
HS_LOG(ERR, "Scratch alloc failed for qp %u"
" after import", i);
+ /* Clean up already allocated scratches */
for (j = 0; j < i; j++) {
if (priv->qps[j].scratch) {
hs_free_scratch(priv->qps[j].scratch);
@@ -569,6 +639,10 @@ hs_regex_rule_db_import(struct rte_regexdev *dev, const char *rule_db,
return 0;
}
+/*
+ * Export the compiled Hyperscan database as a serialized blob.
+ * If rule_db is NULL, returns the required buffer size.
+ */
static int
hs_regex_rule_db_export(struct rte_regexdev *dev, char *rule_db)
{
@@ -596,6 +670,7 @@ hs_regex_rule_db_export(struct rte_regexdev *dev, char *rule_db)
}
if (rule_db == NULL) {
+ /* buf allocated by Hyperscan's malloc, not rte_malloc. */
free(buf);
if (len > INT_MAX) {
HS_LOG(ERR, "Serialized DB too large (%zu bytes)", len);
@@ -605,30 +680,176 @@ hs_regex_rule_db_export(struct rte_regexdev *dev, char *rule_db)
}
memcpy(rule_db, buf, len);
+ /* Hyperscan allocates buf internally via malloc, not rte_malloc. */
free(buf);
return 0;
}
-/* Fast path stubs replaced by real implementations in later patches. */
+/*
+ * Fast Path
+ *
+ * Thread-safety model: single-producer / single-consumer per queue
+ * pair. Each QP must be used by exactly one thread. No locking is
+ * performed on ring operations (head/tail/count). Using the same QP
+ * from multiple threads concurrently causes data races.
+ */
static uint16_t
-hs_regex_enqueue_burst(struct rte_regexdev *dev __rte_unused,
- uint16_t qp_id __rte_unused,
- struct rte_regex_ops **ops __rte_unused,
- uint16_t nb_ops __rte_unused)
+hs_regex_enqueue_burst(struct rte_regexdev *dev, uint16_t qp_id,
+ struct rte_regex_ops **ops, uint16_t nb_ops)
{
- return 0;
+ struct hs_regex_priv *priv;
+ struct hs_regex_qp *qp;
+ uint16_t i;
+ uint16_t free_space;
+
+ if (unlikely(dev == NULL || ops == NULL))
+ return 0;
+
+ priv = dev->data->dev_private;
+ if (unlikely(priv == NULL))
+ return 0;
+
+ /* Validate queue pair index. */
+ if (unlikely(qp_id >= priv->nb_queue_pairs)) {
+ HS_LOG(ERR, "enqueue: invalid qp_id %u (max %u)",
+ qp_id, priv->nb_queue_pairs);
+ return 0;
+ }
+
+ if (unlikely(priv->dev_state != HS_REGEX_DEV_STARTED)) {
+ HS_LOG(ERR, "enqueue: device not started");
+ return 0;
+ }
+
+ if (unlikely(priv->db == NULL)) {
+ HS_LOG(ERR, "enqueue: no compiled database, dropping burst");
+ return 0;
+ }
+
+ qp = &priv->qps[qp_id];
+
+ if (unlikely(qp->scratch == NULL)) {
+ HS_LOG(ERR, "enqueue: qp %u has no scratch, dropping burst",
+ qp_id);
+ return 0;
+ }
+
+ /* Bounded ring: accept only free entries. */
+ free_space = qp->nb_desc - qp->count;
+ if (nb_ops > free_space)
+ nb_ops = free_space;
+
+ for (i = 0; i < nb_ops; i++) {
+ struct rte_regex_ops *op = ops[i];
+ struct rte_mbuf *mbuf;
+ const char *data;
+ uint32_t data_len;
+ struct hs_match_ctx ctx = { .total_matches = 0 };
+ hs_error_t err;
+
+ if (unlikely(op == NULL))
+ break;
+
+ mbuf = op->mbuf;
+ if (unlikely(mbuf == NULL)) {
+ op->nb_matches = 0;
+ op->nb_actual_matches = 0;
+ op->rsp_flags = RTE_REGEX_OPS_RSP_RESOURCE_LIMIT_REACHED_F;
+ goto enqueue_op;
+ }
+
+ /* hs_scan requires contiguous data. */
+ if (rte_pktmbuf_linearize(mbuf) != 0) {
+ op->nb_matches = 0;
+ op->nb_actual_matches = 0;
+ op->rsp_flags = RTE_REGEX_OPS_RSP_RESOURCE_LIMIT_REACHED_F;
+ goto enqueue_op;
+ }
+ data = rte_pktmbuf_mtod(mbuf, const char *);
+ data_len = rte_pktmbuf_pkt_len(mbuf);
+
+ if (unlikely(data_len == 0)) {
+ op->nb_matches = 0;
+ op->nb_actual_matches = 0;
+ op->rsp_flags = 0;
+ goto enqueue_op;
+ }
+
+ op->nb_matches = 0;
+ op->nb_actual_matches = 0;
+ op->rsp_flags = 0;
+
+ ctx.op = op;
+ ctx.max_matches = priv->max_matches;
+ ctx.total_matches = 0;
+
+ err = hs_scan(priv->db, data, data_len, 0,
+ qp->scratch, hs_match_cb, &ctx);
+
+ if (unlikely(err != HS_SUCCESS &&
+ err != HS_SCAN_TERMINATED))
+ op->rsp_flags |=
+ RTE_REGEX_OPS_RSP_RESOURCE_LIMIT_REACHED_F;
+
+enqueue_op:
+ /* Keep completed op for dequeue_burst(). */
+ qp->ops[qp->tail] = op;
+ qp->tail = (qp->tail + 1) % qp->nb_desc;
+ qp->count++;
+
+ qp->qp_matches += ctx.total_matches;
+ }
+
+ qp->qp_enqueued += i;
+ return i;
}
static uint16_t
-hs_regex_dequeue_burst(struct rte_regexdev *dev __rte_unused,
- uint16_t qp_id __rte_unused,
- struct rte_regex_ops **ops __rte_unused,
- uint16_t nb_ops __rte_unused)
+hs_regex_dequeue_burst(struct rte_regexdev *dev, uint16_t qp_id,
+ struct rte_regex_ops **ops, uint16_t nb_ops)
{
- return 0;
+ struct hs_regex_priv *priv;
+ struct hs_regex_qp *qp;
+ uint16_t i;
+ uint16_t avail;
+
+ if (unlikely(dev == NULL || ops == NULL))
+ return 0;
+
+ priv = dev->data->dev_private;
+ if (unlikely(priv == NULL))
+ return 0;
+
+ /* Validate queue pair index. */
+ if (unlikely(qp_id >= priv->nb_queue_pairs)) {
+ HS_LOG(ERR, "dequeue: invalid qp_id %u (max %u)",
+ qp_id, priv->nb_queue_pairs);
+ return 0;
+ }
+
+ if (unlikely(priv->qps == NULL))
+ return 0;
+
+ qp = &priv->qps[qp_id];
+
+ /* Return completed ops currently available in the ring. */
+ avail = qp->count;
+ if (nb_ops > avail)
+ nb_ops = avail;
+
+ for (i = 0; i < nb_ops; i++) {
+ ops[i] = qp->ops[qp->head];
+ qp->head = (qp->head + 1) % qp->nb_desc;
+ qp->count--;
+ }
+
+ /* Device-level stats. */
+ qp->qp_dequeued += i;
+ return i;
}
+/* Operations table */
static const struct rte_regexdev_ops hs_regexdev_ops = {
.dev_info_get = hs_regex_info_get,
.dev_configure = hs_regex_configure,
--
2.43.0
next prev parent reply other threads:[~2026-08-28 5:37 UTC|newest]
Thread overview: 9+ messages / expand[flat|nested] mbox.gz Atom feed top
2026-08-28 5:35 [RFC 0/7] regex/hs: introduce Hyperscan regex PMD Prudvi Deti
2026-08-28 5:35 ` [RFC 1/7] regex/hs: add driver skeleton and build integration Prudvi Deti
2026-08-28 5:35 ` [RFC 2/7] regex/hs: add device configure and queue pair setup Prudvi Deti
2026-08-28 5:35 ` [RFC 3/7] regex/hs: add rule database update and compilation Prudvi Deti
2026-08-28 5:35 ` Prudvi Deti [this message]
2026-08-28 5:35 ` [RFC 5/7] regex/hs: add per-queue-pair extended statistics Prudvi Deti
2026-08-28 5:35 ` [RFC 6/7] regex/hs: add start stop close and device dump Prudvi Deti
2026-08-28 5:35 ` [RFC 7/7] regex/hs: add Hyperscan compile flag support Prudvi Deti
2026-08-28 16:34 ` [RFC 0/7] regex/hs: introduce Hyperscan regex PMD Stephen Hemminger
Reply instructions:
You may reply publicly to this message via plain-text email
using any one of the following methods:
* Save the following mbox file, import it into your mail client,
and reply-to-all from there: mbox
Avoid top-posting and favor interleaved quoting:
https://en.wikipedia.org/wiki/Posting_style#Interleaved_style
* Reply using the --to, --cc, and --in-reply-to
switches of git-send-email(1):
git send-email \
--in-reply-to=20260828053629.1224611-5-prudvi.deti@intel.com \
--to=prudvi.deti@intel.com \
--cc=bruce.richardson@intel.com \
--cc=dev@dpdk.org \
--cc=nirint.shah@intel.com \
--cc=orika@nvidia.com \
/path/to/YOUR_REPLY
https://kernel.org/pub/software/scm/git/docs/git-send-email.html
* If your mail client supports setting the In-Reply-To header
via mailto: links, try the mailto: link
Be sure your reply has a Subject: header at the top and a blank line
before the message body.
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox