Scatter gather mode feature support is added to compress or decompress
the larger data in a single compression or decompression operation.
Signed-off-by: Mahipal Challa <mcha...@marvell.com>
---
v2:
- Checkpatch warning is resolved.
- PMD debug logs are used for debug prints.
- Documentation is updated.
---
---
doc/guides/compressdevs/features/octeontx.ini | 3 +
drivers/compress/octeontx/otx_zip.h | 143 +++++++++++++++---
drivers/compress/octeontx/otx_zip_pmd.c | 84 +++++++---
3 files changed, 194 insertions(+), 36 deletions(-)
diff --git a/doc/guides/compressdevs/features/octeontx.ini
b/doc/guides/compressdevs/features/octeontx.ini
index cc8b025682..8a9b2906d8 100644
--- a/doc/guides/compressdevs/features/octeontx.ini
+++ b/doc/guides/compressdevs/features/octeontx.ini
@@ -8,3 +8,6 @@ HW Accelerated = Y
Deflate = Y
Fixed = Y
Dynamic = Y
+OOP SGL In SGL Out = Y
+OOP SGL In LB Out = Y
+OOP LB In SGL Out = Y
diff --git a/drivers/compress/octeontx/otx_zip.h
b/drivers/compress/octeontx/otx_zip.h
index ee14deb03d..7391360925 100644
--- a/drivers/compress/octeontx/otx_zip.h
+++ b/drivers/compress/octeontx/otx_zip.h
@@ -55,10 +55,13 @@ extern int octtx_zip_logtype_driver;
ZIP_MAX_NCBP_SIZE)/* ~8072ull */
#define ZIP_BUF_SIZE 256
+#define ZIP_SGBUF_SIZE (5 * 1024)
#define ZIP_BURST_SIZE 64
#define ZIP_MAXSEG_SIZE 59460
#define ZIP_EXTRABUF_SIZE 4096
+#define ZIP_MAX_SEGS 300
+#define ZIP_MAX_DATA_SIZE (16*1024*1024)
#define ZIP_SGPTR_ALIGN 16
#define ZIP_CMDQ_ALIGN 128
@@ -102,6 +105,12 @@ struct zipvf_qp;
typedef int (*comp_func_t)(struct rte_comp_op *op, struct zipvf_qp *qp,
struct zip_stream *zstrm, int num);
+/* Scatter gather list */
+struct zipvf_sginfo {
+ union zip_zptr_addr_s sg_addr;
+ union zip_zptr_ctl_s sg_ctl;
+} __rte_aligned(16);
+
/**
* ZIP private stream structure
*/
@@ -144,6 +153,11 @@ struct zipvf_qp {
/* Unique Queue Pair Name */
struct zip_vf *vf;
/* pointer to device, queue belongs to */
+ struct zipvf_sginfo *g_info;
+ struct zipvf_sginfo *s_info;
+ /* SGL pointers */
+ uint64_t num_sgbuf;
+ uint64_t enqed;
} __rte_cache_aligned;
/**
@@ -161,50 +175,134 @@ struct zip_vf {
uint32_t max_nb_queue_pairs;
/* pointer to device qps */
struct rte_mempool *zip_mp;
+ struct rte_mempool *sg_mp;
/* pointer to pools */
} __rte_cache_aligned;
-static inline void
-zipvf_prepare_in_buf(union zip_inst_s *inst, struct rte_comp_op *op)
+static inline int
+zipvf_prepare_sgl(struct rte_mbuf *buf, int64_t offset, struct zipvf_sginfo
*sg_list,
+ uint32_t data_len, const uint16_t max_segs, struct zipvf_qp
*qp)
+{
+ struct zipvf_sginfo *sginfo = (struct zipvf_sginfo *)sg_list;
+ uint32_t tot_buf_len, sgidx;
+ int ret = -EINVAL;
+
+ for (sgidx = tot_buf_len = 0; buf && sgidx < max_segs; buf = buf->next)
{
+ if (offset >= rte_pktmbuf_data_len(buf)) {
+ offset -= rte_pktmbuf_data_len(buf);
+ continue;
+ }
+
+ sginfo[sgidx].sg_ctl.s.length =
(uint16_t)(rte_pktmbuf_data_len(buf) - offset);
+ sginfo[sgidx].sg_addr.s.addr = rte_pktmbuf_iova_offset(buf,
offset);
+
+ offset = 0;
+ tot_buf_len += sginfo[sgidx].sg_ctl.s.length;
+
+ if (tot_buf_len >= data_len) {
+ sginfo[sgidx].sg_ctl.s.length -= tot_buf_len - data_len;
+ ret = 0;
+ break;
+ }
+
+ ZIP_PMD_LOG(DEBUG, "ZIP SGL buf[%d], len = %d, iova =
0x%"PRIx64"\n",
+ sgidx, sginfo[sgidx].sg_ctl.s.length,
sginfo[sgidx].sg_addr.s.addr);
+ ++sgidx;
+ }
+
+ if (unlikely(ret != 0)) {
+ if (sgidx == max_segs)
+ ZIP_PMD_ERR("Exceeded max segments in ZIP SGL (%u)",
max_segs);
+ else
+ ZIP_PMD_ERR("Mbuf chain is too short");
+ }
+ qp->num_sgbuf = ++sgidx;
+
+ ZIP_PMD_LOG(DEBUG, "Tot_buf_len:%d max_segs:%"PRIx64"\n", tot_buf_len,
+ qp->num_sgbuf);
+ return ret;
+}
+
+static inline int
+zipvf_prepare_in_buf(union zip_inst_s *inst, struct zipvf_qp *qp, struct
rte_comp_op *op)
{
uint32_t offset, inlen;
struct rte_mbuf *m_src;
+ int ret = 0;
inlen = op->src.length;
offset = op->src.offset;
m_src = op->m_src;
+ /* Gather input */
+ if (op->m_src->next != NULL && inlen > ZIP_MAXSEG_SIZE) {
+ inst->s.dg = 1;
+
+ ret = zipvf_prepare_sgl(m_src, offset, qp->g_info, inlen,
+ op->m_src->nb_segs, qp);
+
+ inst->s.inp_ptr_addr.s.addr = rte_mem_virt2iova(qp->g_info);
+ inst->s.inp_ptr_ctl.s.length = qp->num_sgbuf;
+ inst->s.inp_ptr_ctl.s.fw = 0;
+
+ ZIP_PMD_LOG(DEBUG, "Gather(input): len(nb_segs):%d, iova:
0x%"PRIx64"\n",
+ inst->s.inp_ptr_ctl.s.length,
inst->s.inp_ptr_addr.s.addr);
+ return ret;
+ }
+
/* Prepare direct input data pointer */
inst->s.dg = 0;
- inst->s.inp_ptr_addr.s.addr =
- rte_pktmbuf_iova_offset(m_src, offset);
+ inst->s.inp_ptr_addr.s.addr = rte_pktmbuf_iova_offset(m_src, offset);
inst->s.inp_ptr_ctl.s.length = inlen;
+
+ ZIP_PMD_LOG(DEBUG, "Direct input - inlen:%d\n", inlen);
+ return ret;
}
-static inline void
-zipvf_prepare_out_buf(union zip_inst_s *inst, struct rte_comp_op *op)
+static inline int
+zipvf_prepare_out_buf(union zip_inst_s *inst, struct zipvf_qp *qp, struct
rte_comp_op *op)
{
- uint32_t offset;
+ uint32_t offset, outlen;
struct rte_mbuf *m_dst;
+ int ret = 0;
offset = op->dst.offset;
m_dst = op->m_dst;
+ outlen = rte_pktmbuf_pkt_len(m_dst) - op->dst.offset;
- /* Prepare direct input data pointer */
+ /* Scatter output */
+ if (op->m_dst->next != NULL && outlen > ZIP_MAXSEG_SIZE) {
+ inst->s.ds = 1;
+ inst->s.totaloutputlength = outlen;
+
+ ret = zipvf_prepare_sgl(m_dst, offset, qp->s_info,
inst->s.totaloutputlength,
+ m_dst->nb_segs, qp);
+
+ inst->s.out_ptr_addr.s.addr = rte_mem_virt2iova(qp->s_info);
+ inst->s.out_ptr_ctl.s.length = qp->num_sgbuf;
+
+ ZIP_PMD_LOG(DEBUG, "Scatter(output): nb_segs:%d,
iova:0x%"PRIx64"\n",
+ inst->s.out_ptr_ctl.s.length,
inst->s.out_ptr_addr.s.addr);
+ return ret;
+ }
+
+ /* Prepare direct output data pointer */
inst->s.ds = 0;
- inst->s.out_ptr_addr.s.addr =
- rte_pktmbuf_iova_offset(m_dst, offset);
- inst->s.totaloutputlength = rte_pktmbuf_pkt_len(m_dst) -
- op->dst.offset;
+ inst->s.out_ptr_addr.s.addr = rte_pktmbuf_iova_offset(m_dst, offset);
+ inst->s.totaloutputlength = rte_pktmbuf_pkt_len(m_dst) - op->dst.offset;
if (inst->s.totaloutputlength == ZIP_MAXSEG_SIZE)
inst->s.totaloutputlength += ZIP_EXTRABUF_SIZE; /* DSTOP */
inst->s.out_ptr_ctl.s.length = inst->s.totaloutputlength;
+
+ ZIP_PMD_LOG(DEBUG, "Direct output - outlen:%d\n",
inst->s.totaloutputlength);
+ return ret;
}
-static inline void
-zipvf_prepare_cmd_stateless(struct rte_comp_op *op, union zip_inst_s *inst)
+static inline int
+zipvf_prepare_cmd_stateless(struct rte_comp_op *op, struct zipvf_qp *qp,
+ union zip_inst_s *inst)
{
/* set flush flag to always 1*/
inst->s.ef = 1;
@@ -217,9 +315,18 @@ zipvf_prepare_cmd_stateless(struct rte_comp_op *op, union
zip_inst_s *inst)
/* Set input checksum */
inst->s.adlercrc32 = op->input_chksum;
- /* Prepare gather buffers */
- zipvf_prepare_in_buf(inst, op);
- zipvf_prepare_out_buf(inst, op);
+ /* Prepare input/output buffers */
+ if (zipvf_prepare_in_buf(inst, qp, op)) {
+ ZIP_PMD_ERR("Con't fill input SGL ");
+ return -EINVAL;
+ }
+
+ if (zipvf_prepare_out_buf(inst, qp, op)) {
+ ZIP_PMD_ERR("Con't fill output SGL ");
+ return -EINVAL;
+ }
+
+ return 0;
}
#ifdef ZIP_DBG
@@ -229,6 +336,7 @@ zip_dump_instruction(void *inst)
union zip_inst_s *cmd83 = (union zip_inst_s *)inst;
printf("####### START ########\n");
+ printf("ZIP Instr:0x%"PRIx64"\n", cmd83);
printf("doneint:%d totaloutputlength:%d\n", cmd83->s.doneint,
cmd83->s.totaloutputlength);
printf("exnum:%d iv:%d exbits:%d hmif:%d halg:%d\n", cmd83->s.exn,
@@ -248,6 +356,7 @@ zip_dump_instruction(void *inst)
printf("inp_ptr.len:%d\n", cmd83->s.inp_ptr_ctl.s.length);
printf("out_ptr.addr:0x%"PRIx64"\n", cmd83->s.out_ptr_addr.s.addr);
printf("out_ptr.len:%d\n", cmd83->s.out_ptr_ctl.s.length);
+ printf("result_ptr.addr:0x%"PRIx64"\n", cmd83->s.res_ptr_addr.s.addr);
printf("result_ptr.len:%d\n", cmd83->s.res_ptr_ctl.s.length);
printf("####### END ########\n");
}
diff --git a/drivers/compress/octeontx/otx_zip_pmd.c
b/drivers/compress/octeontx/otx_zip_pmd.c
index 2d856b19bb..fd20139da6 100644
--- a/drivers/compress/octeontx/otx_zip_pmd.c
+++ b/drivers/compress/octeontx/otx_zip_pmd.c
@@ -16,7 +16,10 @@ static const struct rte_compressdev_capabilities
{ .algo = RTE_COMP_ALGO_DEFLATE,
/* Deflate */
.comp_feature_flags = RTE_COMP_FF_HUFFMAN_FIXED |
- RTE_COMP_FF_HUFFMAN_DYNAMIC,
+ RTE_COMP_FF_HUFFMAN_DYNAMIC |
+ RTE_COMP_FF_OOP_SGL_IN_SGL_OUT |
+ RTE_COMP_FF_OOP_SGL_IN_LB_OUT |
+ RTE_COMP_FF_OOP_LB_IN_SGL_OUT,
/* Non sharable Priv XFORM and Stateless */
.window_size = {
.min = 1,
@@ -46,15 +49,27 @@ zip_process_op(struct rte_comp_op *op,
union zip_inst_s *inst = zstrm->inst[num];
volatile union zip_zres_s *zresult = NULL;
- if ((op->m_src->nb_segs > 1) || (op->m_dst->nb_segs > 1) ||
- (op->src.offset > rte_pktmbuf_pkt_len(op->m_src)) ||
- (op->dst.offset > rte_pktmbuf_pkt_len(op->m_dst))) {
+ if (op->m_src->nb_segs > 1)
+ if (rte_mempool_get(qp->vf->sg_mp, (void *)&qp->g_info) < 0) {
+ ZIP_PMD_ERR("Can't1 allocate object from SG pool");
+ return (-ENOMEM);
+ }
+
+ if (op->m_dst->nb_segs > 1)
+ if (rte_mempool_get(qp->vf->sg_mp, (void *)&qp->s_info) < 0) {
+ rte_mempool_put(qp->vf->sg_mp, qp->g_info);
+ ZIP_PMD_ERR("Can't allocate object from SG pool");
+ return (-ENOMEM);
+ }
+
+ if (zipvf_prepare_cmd_stateless(op, qp, inst)) {
op->status = RTE_COMP_OP_STATUS_INVALID_ARGS;
- ZIP_PMD_ERR("Segmented packet is not supported\n");
- return 0;
- }
+ rte_mempool_put(qp->vf->sg_mp, qp->g_info);
+ rte_mempool_put(qp->vf->sg_mp, qp->s_info);
- zipvf_prepare_cmd_stateless(op, inst);
+ ZIP_PMD_ERR("Can't fill SGL buffers");
+ return -EINVAL;
+ }
zresult = (union zip_zres_s *)zstrm->bufs[RES_BUF + num];
zresult->s.compcode = 0;
@@ -174,10 +189,12 @@ static int
zip_pmd_config(struct rte_compressdev *dev,
struct rte_compressdev_config *config)
{
- int nb_streams;
char res_pool[RTE_MEMZONE_NAMESIZE];
- struct zip_vf *vf;
+ char sg_pool[RTE_MEMZONE_NAMESIZE];
struct rte_mempool *zip_buf_mp;
+ struct rte_mempool *zip_sg_mp;
+ struct zip_vf *vf;
+ int nb_streams;
if (!config || !dev)
return -EIO;
@@ -194,6 +211,9 @@ zip_pmd_config(struct rte_compressdev *dev,
snprintf(res_pool, RTE_MEMZONE_NAMESIZE, "octtx_zip_res_pool%u",
dev->data->dev_id);
+ snprintf(sg_pool, RTE_MEMZONE_NAMESIZE, "octtx_zip_sg_pool%u",
+ dev->data->dev_id);
+
/** TBD Should we use the per core object cache for stream resources */
zip_buf_mp = rte_mempool_create(
res_pool,
@@ -215,7 +235,30 @@ zip_pmd_config(struct rte_compressdev *dev,
return -1;
}
+ /* Scatter gather buffer pool */
+ zip_sg_mp = rte_mempool_create(
+ sg_pool,
+ (2 * nb_streams * ZIP_BURST_SIZE * ZIP_MAX_SEGS),
+ ZIP_SGBUF_SIZE,
+ 0,
+ 0,
+ NULL,
+ NULL,
+ NULL,
+ NULL,
+ SOCKET_ID_ANY,
+ MEMPOOL_F_NO_SPREAD);
+
+ if (zip_sg_mp == NULL) {
+ ZIP_PMD_ERR("Failed to create SG buf mempool
octtx_zip_sg_pool%u",
+ dev->data->dev_id);
+
+ rte_mempool_free(vf->zip_mp);
+ return -1;
+ }
+
vf->zip_mp = zip_buf_mp;
+ vf->sg_mp = zip_sg_mp;
return 0;
}
@@ -243,6 +286,7 @@ zip_pmd_close(struct rte_compressdev *dev)
struct zip_vf *vf = (struct zip_vf *)dev->data->dev_private;
rte_mempool_free(vf->zip_mp);
+ rte_mempool_free(vf->sg_mp);
return 0;
}
@@ -482,9 +526,9 @@ zip_pmd_enqueue_burst(void *queue_pair,
}
}
-#ifdef ZIP_DBG
- ZIP_PMD_INFO("ops_enqd[nb_ops:%d]:%d\n", nb_ops, enqd);
-#endif
+ qp->enqed = enqd;
+ ZIP_PMD_LOG(DEBUG, "ops_enqd[nb_ops:%d]:%d\n", nb_ops, enqd);
+
return enqd;
}
@@ -527,9 +571,7 @@ zip_pmd_dequeue_burst(void *queue_pair,
op->status = RTE_COMP_OP_STATUS_ERROR;
}
- #ifdef ZIP_DBG
- ZIP_PMD_INFO("written %d\n", zresult->s.totalbyteswritten);
- #endif
+ ZIP_PMD_LOG(DEBUG, "written %d\n",
zresult->s.totalbyteswritten);
/* Update op stats */
switch (op->status) {
@@ -548,11 +590,15 @@ zip_pmd_dequeue_burst(void *queue_pair,
/* zstream is reset irrespective of result */
reset_stream(zstrm->inst[i]);
zresult->s.compcode = ZIP_COMP_E_NOTDONE;
+
+ if (op->m_src->nb_segs > 1)
+ rte_mempool_put(qp->vf->sg_mp, qp->g_info);
+
+ if (op->m_dst->nb_segs > 1)
+ rte_mempool_put(qp->vf->sg_mp, qp->s_info);
}
-#ifdef ZIP_DBG
- ZIP_PMD_INFO("ops_deqd[nb_ops:%d]: %d\n", nb_ops, nb_dequeued);
-#endif
+ ZIP_PMD_LOG(DEBUG, "ops_deqd[nb_ops:%d]: %d\n", nb_ops, nb_dequeued);
return nb_dequeued;
}
--
2.25.1
