Skip to contentSkip to conversation

add mlx5 compress PMD

PATCH Superseded Matan Azrad 5 participants 41 messages Jan 11, 2021 – Jan 27, 2021 Arm/Amlogic Meson SoC support

lore
Share & tools
lore
#0 lore raw

cover add mlx5 compress PMD

Add a new compress PMD for Mellanox devices.

The MLX5 compress driver library provides support for Mellanox BlueField 2 families of 25/50/100/200 Gb/s adapters.

Using the BlueField 2 device, the compress class operations can be run in parallel to the net, vdpa, and regex class operations.

This driver is depending on rdma-core like the other mlx5 PMDs, also it is going to use mlx5 DevX to create HW objects directly by the FW.

Add the probing functions, PCI bus connectivity, HW capabilities checks and some basic objects preparations.

Documentations are not ready yet, will send later.

Matan Azrad (10):
common/mlx5: add DevX attributes for compress
drivers: introduce mlx5 compress PMD
compress/mlx5: support basic control operations
common/mlx5: add compress primitives
compress/mlx5: support queue pair operations
compress/mlx5: add transformation operations
compress/mlx5: add memory region management
compress/mlx5: add data-path functions
compress/mlx5: add statistics operations
compress/mlx5: add the supported capabilities

 MAINTAINERS                                 |   4 +
 drivers/common/mlx5/mlx5_common.h           |   1 +
 drivers/common/mlx5/mlx5_common_pci.c       |   7 +
 drivers/common/mlx5/mlx5_common_pci.h       |  36 +-
 drivers/common/mlx5/mlx5_devx_cmds.c        |  10 +
 drivers/common/mlx5/mlx5_devx_cmds.h        |   7 +
 drivers/common/mlx5/mlx5_prm.h              |  59 +-
 drivers/compress/meson.build                |   2 +-
 drivers/compress/mlx5/meson.build           |  26 +
 drivers/compress/mlx5/mlx5_compress.c       | 863 ++++++++++++++++++++++++++++
 drivers/compress/mlx5/mlx5_compress_utils.h |  20 +
 drivers/compress/mlx5/version.map           |   3 +
 12 files changed, 1015 insertions(+), 23 deletions(-)
 create mode 100644 drivers/compress/mlx5/meson.build
 create mode 100644 drivers/compress/mlx5/mlx5_compress.c
 create mode 100644 drivers/compress/mlx5/mlx5_compress_utils.h
 create mode 100644 drivers/compress/mlx5/version.map
Signature · 1 line
1.8.3.1
#1 lore raw

PATCH 1/10 common/mlx5: add DevX attributes for compress

Add the DevX attributes for compress related engiens:
compress
decompress
dma

---
 drivers/common/mlx5/mlx5_devx_cmds.c | 10 ++++++++++
 drivers/common/mlx5/mlx5_devx_cmds.h |  7 +++++++
 drivers/common/mlx5/mlx5_prm.h       | 18 ++++++++++++++++--
 3 files changed, 33 insertions(+), 2 deletions(-)
drivers/common/mlx5/mlx5_devx_cmds.c +10 -0
index 790a701..84a835d 100644
@@ -725,6 +725,16 @@ struct mlx5_devx_obj *
 	attr->log_max_pd = MLX5_GET(cmd_hca_cap, hcattr, log_max_pd);
 	attr->log_max_srq = MLX5_GET(cmd_hca_cap, hcattr, log_max_srq);
 	attr->log_max_srq_sz = MLX5_GET(cmd_hca_cap, hcattr, log_max_srq_sz);
+	attr->mmo_dma_en = MLX5_GET(cmd_hca_cap, hcattr, dma_mmo);
+	attr->mmo_compress_en = MLX5_GET(cmd_hca_cap, hcattr, compress);
+	attr->mmo_decompress_en = MLX5_GET(cmd_hca_cap, hcattr, decompress);
+	attr->compress_min_block_size = MLX5_GET(cmd_hca_cap, hcattr,
+						 compress_min_block_size);
+	attr->log_max_mmo_dma = MLX5_GET(cmd_hca_cap, hcattr, log_dma_mmo_size);
+	attr->log_max_mmo_compress = MLX5_GET(cmd_hca_cap, hcattr,
+					      log_compress_mmo_size);
+	attr->log_max_mmo_decompress = MLX5_GET(cmd_hca_cap, hcattr,
+						log_decompress_mmo_size);
 	if (attr->qos.sup) {
 		MLX5_SET(query_hca_cap_in, in, op_mod,
 			 MLX5_GET_HCA_CAP_OP_MOD_QOS_CAP |
drivers/common/mlx5/mlx5_devx_cmds.h +7 -0
index a14f3bf..a279422 100644
@@ -127,6 +127,13 @@ struct mlx5_hca_attr {
 	uint32_t log_max_srq;
 	uint32_t log_max_srq_sz;
 	uint32_t rss_ind_tbl_cap;
+	uint32_t mmo_dma_en:1;
+	uint32_t mmo_compress_en:1;
+	uint32_t mmo_decompress_en:1;
+	uint32_t compress_min_block_size:4;
+	uint32_t log_max_mmo_dma:5;
+	uint32_t log_max_mmo_compress:5;
+	uint32_t log_max_mmo_decompress:5;
 };
 
 struct mlx5_devx_wq_attr {
drivers/common/mlx5/mlx5_prm.h +16 -2
index 8c9b53c..6f5a6cf 100644
@@ -1120,7 +1120,15 @@ enum {
 struct mlx5_ifc_cmd_hca_cap_bits {
 	u8 reserved_at_0[0x30];
 	u8 vhca_id[0x10];
-	u8 reserved_at_40[0x40];
+	u8 reserved_at_40[0x20];
+	u8 reserved_at_60[0x3];
+	u8 log_regexp_scatter_gather_size[0x5];
+	u8 reserved_at_68[0x3];
+	u8 log_dma_mmo_size[5];
+	u8 reserved_at_70[0x3];
+	u8 log_compress_mmo_size[5];
+	u8 reserved_at_78[0x3];
+	u8 log_decompress_mmo_size[5];
 	u8 log_max_srq_sz[0x8];
 	u8 log_max_qp_sz[0x8];
 	u8 reserved_at_90[0x9];
@@ -1168,7 +1176,13 @@ struct mlx5_ifc_cmd_hca_cap_bits {
 	u8 log_max_ra_res_dc[0x6];
 	u8 reserved_at_140[0xa];
 	u8 log_max_ra_req_qp[0x6];
-	u8 reserved_at_150[0xa];
+	u8 rtr2rts_qp_counters_set_id[1];
+	u8 rts2rts_udp_sport[1];
+	u8 rts2rts_lag_tx_port_affinity[1];
+	u8 dma_mmo[1];
+	u8 compress_min_block_size[4];
+	u8 compress[1];
+	u8 decompress[1];
 	u8 log_max_ra_res_qp[0x6];
 	u8 end_pad[0x1];
 	u8 cc_query_allowed[0x1];
Signature · 1 line
1.8.3.1
#2 lore raw

PATCH 2/10 drivers: introduce mlx5 compress PMD

In reply to Matan Azrad

Add a new compress PMD for Mellanox devices.

The MLX5 compress driver library provides support for Mellanox BlueField 2 families of 25/50/100/200 Gb/s adapters.

Using the BlueField 2 device, the compress class operations can be run in parallel to the net, vdpa, and regex class operations.

This driver is depending on rdma-core like the other mlx5 PMDs, also it is going to use mlx5 DevX to create HW objects directly by the FW.

Add the probing functions, PCI bus connectivity, HW capabilities checks and some basic objects preparations.

---
 MAINTAINERS                                 |   4 +
 drivers/common/mlx5/mlx5_common.h           |   1 +
 drivers/common/mlx5/mlx5_common_pci.c       |   7 +
 drivers/common/mlx5/mlx5_common_pci.h       |  36 ++--
 drivers/compress/meson.build                |   2 +-
 drivers/compress/mlx5/meson.build           |  26 +++
 drivers/compress/mlx5/mlx5_compress.c       | 297 ++++++++++++++++++++++++++++
 drivers/compress/mlx5/mlx5_compress_utils.h |  20 ++
 drivers/compress/mlx5/version.map           |   3 +
 9 files changed, 377 insertions(+), 19 deletions(-)
 create mode 100644 drivers/compress/mlx5/meson.build
 create mode 100644 drivers/compress/mlx5/mlx5_compress.c
 create mode 100644 drivers/compress/mlx5/mlx5_compress_utils.h
 create mode 100644 drivers/compress/mlx5/version.map
MAINTAINERS +4 -0
index eafe9f8..f7527f0 100644
@@ -1132,6 +1132,10 @@ F: drivers/compress/zlib/
 F: doc/guides/compressdevs/zlib.rst
 F: doc/guides/compressdevs/features/zlib.ini
 
+Mellanox mlx5
+M: Matan Azrad <[email protected]>
+F: drivers/compress/mlx5/
+
 
 RegEx Drivers
 -------------
drivers/common/mlx5/mlx5_common.h +1 -0
index e35188d..3855582 100644
@@ -217,6 +217,7 @@ enum mlx5_class {
 	MLX5_CLASS_NET = RTE_BIT64(0),
 	MLX5_CLASS_VDPA = RTE_BIT64(1),
 	MLX5_CLASS_REGEX = RTE_BIT64(2),
+	MLX5_CLASS_COMPRESS = RTE_BIT64(3),
 };
 
 #define MLX5_DBR_SIZE RTE_CACHE_LINE_SIZE
drivers/common/mlx5/mlx5_common_pci.c +7 -0
index 5208972..2b65768 100644
@@ -28,14 +28,21 @@ static TAILQ_HEAD(mlx5_pci_devices_head, mlx5_pci_device) devices_list =
 	{ .name = "vdpa", .driver_class = MLX5_CLASS_VDPA },
 	{ .name = "net", .driver_class = MLX5_CLASS_NET },
 	{ .name = "regex", .driver_class = MLX5_CLASS_REGEX },
+	{ .name = "compress", .driver_class = MLX5_CLASS_COMPRESS },
 };
 
 static const unsigned int mlx5_class_combinations[] = {
 	MLX5_CLASS_NET,
 	MLX5_CLASS_VDPA,
 	MLX5_CLASS_REGEX,
+	MLX5_CLASS_COMPRESS,
 	MLX5_CLASS_NET | MLX5_CLASS_REGEX,
 	MLX5_CLASS_VDPA | MLX5_CLASS_REGEX,
+	MLX5_CLASS_NET | MLX5_CLASS_COMPRESS,
+	MLX5_CLASS_VDPA | MLX5_CLASS_COMPRESS,
+	MLX5_CLASS_REGEX | MLX5_CLASS_COMPRESS,
+	MLX5_CLASS_NET | MLX5_CLASS_REGEX | MLX5_CLASS_COMPRESS,
+	MLX5_CLASS_VDPA | MLX5_CLASS_REGEX | MLX5_CLASS_COMPRESS,
 	/* New class combination should be added here. */
 };
 
drivers/common/mlx5/mlx5_common_pci.h +18 -18
index 41b73e1..de89bb9 100644
@@ -9,26 +9,26 @@
  * @file
  *
  * RTE Mellanox PCI Driver Interface
- * Mellanox ConnectX PCI device supports multiple class (net/vdpa/regex)
- * devices. This layer enables creating such multiple class of devices on a
- * single PCI device by allowing to bind multiple class specific device
+ * Mellanox ConnectX PCI device supports multiple class: net,vdpa,regex and
+ * compress devices. This layer enables creating such multiple class of devices
+ * on a single PCI device by allowing to bind multiple class specific device
  * driver to attach to mlx5_pci driver.
  *
- * -----------    ------------    -------------
- * |   mlx5  |    |   mlx5   |    |   mlx5    |
- * | net pmd |    | vdpa pmd |    | regex pmd |
- * -----------    ------------    -------------
- *      \              |                 /
- *       \             |                /
- *        \       --------------       /
- *         \______|   mlx5     |_____ /
- *                | pci common |
- *                --------------
- *                     |
- *                 -----------
- *                 |   mlx5  |
- *                 | pci dev |
- *                 -----------
+ * -----------    ------------    -------------    ----------------
+ * |   mlx5  |    |   mlx5   |    |   mlx5    |    |     mlx5     |
+ * | net pmd |    | vdpa pmd |    | regex pmd |    | compress pmd |
+ * -----------    ------------    -------------    ----------------
+ *      \              \                    /              /
+ *       \              \                  /              /
+ *        \              \_--------------_/              /
+ *         \_______________|   mlx5     |_______________/
+ *                         | pci common |
+ *                         --------------
+ *                               |
+ *                           -----------
+ *                           |   mlx5  |
+ *                           | pci dev |
+ *                           -----------
  *
  * - mlx5 pci driver binds to mlx5 PCI devices defined by PCI
  *   ID table of all related mlx5 PCI devices.
drivers/compress/meson.build +1 -1
index 33f5e33..d8f3ddb 100644
@@ -5,7 +5,7 @@ if is_windows
 	subdir_done()
 endif
 
-drivers = ['isal', 'octeontx', 'zlib']
+drivers = ['isal', 'octeontx', 'zlib', 'mlx5']
 
 std_deps = ['compressdev'] # compressdev pulls in all other needed deps
 config_flag_fmt = 'RTE_LIBRTE_PMD_@0@'
drivers/compress/mlx5/meson.build new file +26 -0
new file mode 100644
index 0000000..2a6dc3f
@@ -0,0 +1,26 @@
+# SPDX-License-Identifier: BSD-3-Clause
+# Copyright 2021 Mellanox Technologies, Ltd
+
+if not is_linux
+	build = false
+	reason = 'only supported on Linux'
+	subdir_done()
+endif
+
+fmt_name = 'mlx5_compress'
+deps += ['common_mlx5', 'eal', 'compressdev']
+sources = files(
+	'mlx5_compress.c',
+)
+cflags_options = [
+	'-std=c11',
+	'-Wno-strict-prototypes',
+	'-D_BSD_SOURCE',
+	'-D_DEFAULT_SOURCE',
+	'-D_XOPEN_SOURCE=600'
+]
+foreach option:cflags_options
+	if cc.has_argument(option)
+		cflags += option
+	endif
+endforeach
drivers/compress/mlx5/mlx5_compress.c new file +297 -0

297 lines not on this page · raw message

drivers/compress/mlx5/mlx5_compress_utils.h new file +20 -0
new file mode 100644
index 0000000..f93244f
@@ -0,0 +1,20 @@
+/* SPDX-License-Identifier: BSD-3-Clause
+ * Copyright 2021 Mellanox Technologies, Ltd
+ */
+
+#ifndef RTE_PMD_MLX5_COMPRESS_UTILS_H_
+#define RTE_PMD_MLX5_COMPRESS_UTILS_H_
+
+#include <mlx5_common.h>
+
+
+extern int mlx5_compress_logtype;
+
+#define MLX5_COMPRESS_LOG_PREFIX "mlx5_compress"
+/* Generic printf()-like logging macro with automatic line feed. */
+#define DRV_LOG(level, ...) \
+	PMD_DRV_LOG_(level, mlx5_compress_logtype, MLX5_COMPRESS_LOG_PREFIX, \
+		__VA_ARGS__ PMD_DRV_LOG_STRIP PMD_DRV_LOG_OPAREN, \
+		PMD_DRV_LOG_CPAREN)
+
+#endif /* RTE_PMD_MLX5_COMPRESS_UTILS_H_ */
drivers/compress/mlx5/version.map new file +3 -0
new file mode 100644
index 0000000..4a76d1d
@@ -0,0 +1,3 @@
+DPDK_21 {
+	local: *;
+};
Signature · 1 line
1.8.3.1
#3 lore raw

PATCH 3/10 compress/mlx5: support basic control operations

In reply to Matan Azrad

Add initial support for the next operations:
dev_configure
dev_close
dev_infos_get

---
 drivers/compress/mlx5/mlx5_compress.c | 41 ++++++++++++++++++++++++++++++++---
 1 file changed, 38 insertions(+), 3 deletions(-)
drivers/compress/mlx5/mlx5_compress.c +38 -3
index 639dd61..7148798 100644
@@ -32,20 +32,55 @@ struct mlx5_compress_priv {
 	uint8_t min_block_size;
 	/* Minimum huffman block size supported by the device. */
 	struct ibv_pd *pd;
+	struct rte_compressdev_config dev_config;
 };
 
+#define MLX5_COMPRESS_MAX_QPS 1024
+
 TAILQ_HEAD(mlx5_compress_privs, mlx5_compress_priv) mlx5_compress_priv_list =
 				TAILQ_HEAD_INITIALIZER(mlx5_compress_priv_list);
 static pthread_mutex_t priv_list_lock = PTHREAD_MUTEX_INITIALIZER;
 
 int mlx5_compress_logtype;
 
+const struct rte_compressdev_capabilities mlx5_caps[RTE_COMP_ALGO_LIST_END];
+
+
+static void
+mlx5_compress_dev_info_get(struct rte_compressdev *dev,
+			   struct rte_compressdev_info *info)
+{
+	RTE_SET_USED(dev);
+	if (info != NULL) {
+		info->max_nb_queue_pairs = MLX5_COMPRESS_MAX_QPS;
+		info->feature_flags = RTE_COMPDEV_FF_HW_ACCELERATED;
+		info->capabilities = mlx5_caps;
+	}
+}
+
+static int
+mlx5_compress_dev_configure(struct rte_compressdev *dev,
+			    struct rte_compressdev_config *config)
+{
+	struct mlx5_compress_priv *priv = dev->data->dev_private;
+
+	priv->dev_config = *config;
+	return 0;
+}
+
+static int
+mlx5_compress_dev_close(struct rte_compressdev *dev)
+{
+	RTE_SET_USED(dev);
+	return 0;
+}
+
 static struct rte_compressdev_ops mlx5_compress_ops = {
-	.dev_configure		= NULL,
+	.dev_configure		= mlx5_compress_dev_configure,
 	.dev_start		= NULL,
 	.dev_stop		= NULL,
-	.dev_close		= NULL,
-	.dev_infos_get		= NULL,
+	.dev_close		= mlx5_compress_dev_close,
+	.dev_infos_get		= mlx5_compress_dev_info_get,
 	.stats_get		= NULL,
 	.stats_reset		= NULL,
 	.queue_pair_setup	= NULL,
Signature · 1 line
1.8.3.1
#4 lore raw

PATCH 4/10 common/mlx5: add compress primitives

In reply to Matan Azrad

Add the GGA compress WQE related structures and definitions.

---
 drivers/common/mlx5/mlx5_prm.h | 41 +++++++++++++++++++++++++++++++++++++++--
 1 file changed, 39 insertions(+), 2 deletions(-)
drivers/common/mlx5/mlx5_prm.h +39 -2
index 6f5a6cf..00587e8 100644
@@ -412,10 +412,23 @@ struct mlx5_cqe_ts {
 	uint8_t op_own;
 };
 
+/* GGA */
 /* MMO metadata segment */
 
-#define	MLX5_OPCODE_MMO	0x2f
-#define	MLX5_OPC_MOD_MMO_REGEX 0x4
+#define	MLX5_OPCODE_MMO	0x2fu
+#define	MLX5_OPC_MOD_MMO_REGEX 0x4u
+#define	MLX5_OPC_MOD_MMO_COMP 0x2u
+#define	MLX5_OPC_MOD_MMO_DECOMP 0x3u
+#define	MLX5_OPC_MOD_MMO_DMA 0x1u
+
+#define WQE_GGA_COMP_WIN_SIZE_OFFSET 12u
+#define WQE_GGA_COMP_BLOCK_SIZE_OFFSET 16u
+#define WQE_GGA_COMP_DYNAMIC_SIZE_OFFSET 20u
+#define MLX5_GGA_COMP_WIN_SIZE_UNITS 1024u
+#define MLX5_GGA_COMP_WIN_SIZE_MAX (32u * MLX5_GGA_COMP_WIN_SIZE_UNITS)
+#define MLX5_GGA_COMP_LOG_BLOCK_SIZE_MAX 15u
+#define MLX5_GGA_COMP_LOG_DYNAMIC_SIZE_MAX 15u
+#define MLX5_GGA_COMP_LOG_DYNAMIC_SIZE_MIN 0u
 
 struct mlx5_wqe_metadata_seg {
 	uint32_t mmo_control_31_0; /* mmo_control_63_32 is in ctrl_seg.imm */
@@ -423,6 +436,30 @@ struct mlx5_wqe_metadata_seg {
 	uint64_t addr;
 };
 
+struct mlx5_gga_wqe {
+	uint32_t opcode;
+	uint32_t sq_ds;
+	uint32_t flags;
+	uint32_t gga_ctrl1;  /* ws 12-15, bs 16-19, dyns 20-23. */
+	uint32_t gga_ctrl2;
+	uint32_t opaque_lkey;
+	uint64_t opaque_vaddr;
+	struct mlx5_wqe_dseg gather;
+	struct mlx5_wqe_dseg scatter;
+} __rte_packed;
+
+struct mlx5_gga_compress_opaque {
+	uint32_t syndrom;
+	uint32_t reserved0;
+	uint32_t scattered_length;
+	uint32_t gathered_length;
+	uint64_t scatter_crc;
+	uint64_t gather_crc;
+	uint32_t crc32;
+	uint32_t adler32;
+	uint8_t reserved1[216];
+} __rte_packed;
+
 struct mlx5_ifc_regexp_mmo_control_bits {
 	uint8_t reserved_at_31[0x2];
 	uint8_t le[0x1];
Signature · 1 line
1.8.3.1
#5 lore raw

PATCH 5/10 compress/mlx5: support queue pair operations

In reply to Matan Azrad

Add support for the next operations:
queue_pair_setup
queue_pair_release

Create and initialize DevX SQ and CQ for each compress API queue-pair.

---
 drivers/compress/mlx5/mlx5_compress.c | 144 +++++++++++++++++++++++++++++++++-
 1 file changed, 142 insertions(+), 2 deletions(-)
drivers/compress/mlx5/mlx5_compress.c +142 -2
index 7148798..ffd866a 100644
@@ -15,6 +15,8 @@
 #include <mlx5_common_pci.h>
 #include <mlx5_devx_cmds.h>
 #include <mlx5_common_os.h>
+#include <mlx5_common_devx.h>
+#include <mlx5_common_mr.h>
 #include <mlx5_prm.h>
 
 #include "mlx5_compress_utils.h"
@@ -35,6 +37,20 @@ struct mlx5_compress_priv {
 	struct rte_compressdev_config dev_config;
 };
 
+struct mlx5_compress_qp {
+	uint16_t qp_id;
+	uint16_t entries_n;
+	uint16_t pi;
+	uint16_t ci;
+	volatile uint64_t *uar_addr;
+	int socket_id;
+	struct mlx5_devx_cq cq;
+	struct mlx5_devx_sq sq;
+	struct mlx5_pmd_mr opaque_mr;
+	struct rte_comp_op **ops;
+	struct mlx5_compress_priv *priv;
+};
+
 #define MLX5_COMPRESS_MAX_QPS 1024
 
 TAILQ_HEAD(mlx5_compress_privs, mlx5_compress_priv) mlx5_compress_priv_list =
@@ -75,6 +91,130 @@ struct mlx5_compress_priv {
 	return 0;
 }
 
+static int
+mlx5_compress_qp_release(struct rte_compressdev *dev, uint16_t qp_id)
+{
+	struct mlx5_compress_qp *qp = dev->data->queue_pairs[qp_id];
+
+	if (qp->sq.sq != NULL)
+		mlx5_devx_sq_destroy(&qp->sq);
+	if (qp->cq.cq != NULL)
+		mlx5_devx_cq_destroy(&qp->cq);
+	if (qp->opaque_mr.obj != NULL) {
+		void *opaq = qp->opaque_mr.addr;
+
+		mlx5_common_verbs_dereg_mr(&qp->opaque_mr);
+		if (opaq != NULL)
+			rte_free(opaq);
+	}
+	rte_free(qp);
+	dev->data->queue_pairs[qp_id] = NULL;
+	return 0;
+}
+
+static void
+mlx5_compress_init_sq(struct mlx5_compress_qp *qp)
+{
+	volatile struct mlx5_gga_wqe *restrict wqe =
+				    (volatile struct mlx5_gga_wqe *)qp->sq.wqes;
+	volatile struct mlx5_gga_compress_opaque *opaq = qp->opaque_mr.addr;
+	int i;
+
+	/* All the next fields state should stay constant. */
+	for (i = 0; i < qp->entries_n; ++i, ++wqe) {
+		wqe->sq_ds = rte_cpu_to_be_32((qp->sq.sq->id << 8) | 4u);
+		wqe->flags = RTE_BE32(MLX5_COMP_ALWAYS <<
+				      MLX5_COMP_MODE_OFFSET);
+		wqe->opaque_lkey = rte_cpu_to_be_32(qp->opaque_mr.lkey);
+		wqe->opaque_vaddr = rte_cpu_to_be_64
+						((uint64_t)(uintptr_t)&opaq[i]);
+	}
+}
+
+static int
+mlx5_compress_qp_setup(struct rte_compressdev *dev, uint16_t qp_id,
+		       uint32_t max_inflight_ops, int socket_id)
+{
+	struct mlx5_compress_priv *priv = dev->data->dev_private;
+	struct mlx5_compress_qp *qp;
+	struct mlx5_devx_cq_attr cq_attr = {
+		.uar_page_id = mlx5_os_get_devx_uar_page_id(priv->uar),
+	};
+	struct mlx5_devx_create_sq_attr sq_attr = {
+		.user_index = qp_id,
+		.wq_attr = (struct mlx5_devx_wq_attr){
+			.pd = priv->pdn,
+			.uar_page = mlx5_os_get_devx_uar_page_id(priv->uar),
+		},
+	};
+	struct mlx5_devx_modify_sq_attr modify_attr = {
+		.state = MLX5_SQC_STATE_RDY,
+	};
+	uint32_t log_ops_n = rte_log2_u32(max_inflight_ops);
+	uint32_t alloc_size = sizeof(*qp);
+	void *opaq_buf;
+	int ret;
+
+	alloc_size = RTE_ALIGN(alloc_size, RTE_CACHE_LINE_SIZE);
+	alloc_size += sizeof(struct rte_comp_op *) * (1u << log_ops_n);
+	qp = rte_zmalloc_socket(__func__, alloc_size, RTE_CACHE_LINE_SIZE,
+				socket_id);
+	if (qp == NULL) {
+		DRV_LOG(ERR, "Failed to allocate qp memory.");
+		rte_errno = ENOMEM;
+		return -rte_errno;
+	}
+	dev->data->queue_pairs[qp_id] = qp;
+	opaq_buf = rte_calloc(__func__, 1u << log_ops_n,
+			      sizeof(struct mlx5_gga_compress_opaque),
+			      sizeof(struct mlx5_gga_compress_opaque));
+	if (opaq_buf == NULL) {
+		DRV_LOG(ERR, "Failed to allocate opaque memory.");
+		rte_errno = ENOMEM;
+		goto err;
+	}
+	qp->entries_n = 1 << log_ops_n;
+	qp->socket_id = socket_id;
+	qp->qp_id = qp_id;
+	qp->priv = priv;
+	qp->ops = (struct rte_comp_op **)RTE_ALIGN((uintptr_t)(qp + 1),
+						   RTE_CACHE_LINE_SIZE);
+	qp->uar_addr = mlx5_os_get_devx_uar_reg_addr(priv->uar);
+	MLX5_ASSERT(qp->uar_addr);
+	if (mlx5_common_verbs_reg_mr(priv->pd, opaq_buf, qp->entries_n *
+					sizeof(struct mlx5_gga_compress_opaque),
+							 &qp->opaque_mr) != 0) {
+		DRV_LOG(ERR, "Failed to register opaque MR.");
+		rte_errno = ENOMEM;
+		goto err;
+	}
+	ret = mlx5_devx_cq_create(priv->ctx, &qp->cq, log_ops_n, &cq_attr,
+				  socket_id);
+	if (ret != 0) {
+		DRV_LOG(ERR, "Failed to create CQ.");
+		goto err;
+	}
+	sq_attr.cqn = qp->cq.cq->id;
+	ret = mlx5_devx_sq_create(priv->ctx, &qp->sq, log_ops_n, &sq_attr,
+				  socket_id);
+	if (ret != 0) {
+		DRV_LOG(ERR, "Failed to create SQ.");
+		goto err;
+	}
+	mlx5_compress_init_sq(qp);
+	ret = mlx5_devx_cmd_modify_sq(qp->sq.sq, &modify_attr);
+	if (ret != 0) {
+		DRV_LOG(ERR, "Can't change SQ state to ready.");
+		goto err;
+	}
+	DRV_LOG(INFO, "QP %u: SQN=0x%X CQN=0x%X entries num = %u\n",
+		(uint32_t)qp_id, qp->sq.sq->id, qp->cq.cq->id, qp->entries_n);
+	return 0;
+err:
+	mlx5_compress_qp_release(dev, qp_id);
+	return -1;
+}
+
 static struct rte_compressdev_ops mlx5_compress_ops = {
 	.dev_configure		= mlx5_compress_dev_configure,
 	.dev_start		= NULL,
@@ -83,8 +223,8 @@ struct mlx5_compress_priv {
 	.dev_infos_get		= mlx5_compress_dev_info_get,
 	.stats_get		= NULL,
 	.stats_reset		= NULL,
-	.queue_pair_setup	= NULL,
-	.queue_pair_release	= NULL,
+	.queue_pair_setup	= mlx5_compress_qp_setup,
+	.queue_pair_release	= mlx5_compress_qp_release,
 	.private_xform_create	= NULL,
 	.private_xform_free	= NULL,
 	.stream_create		= NULL,
Signature · 1 line
1.8.3.1
#6 lore raw

PATCH 6/10 compress/mlx5: add transformation operations

In reply to Matan Azrad

Add support for the next operations:
private_xform_create
private_xform_free

The driver transformation structure includes preparations for the next GGA WQE fields used by the enqueue function: opcode. compress specific fields checksum type and compress type.

---
 drivers/compress/mlx5/mlx5_compress.c | 122 +++++++++++++++++++++++++++++++++-
 1 file changed, 120 insertions(+), 2 deletions(-)
drivers/compress/mlx5/mlx5_compress.c +120 -2
index ffd866a..132837e 100644
@@ -6,6 +6,7 @@
 #include <rte_log.h>
 #include <rte_errno.h>
 #include <rte_pci.h>
+#include <rte_spinlock.h>
 #include <rte_comp.h>
 #include <rte_compressdev.h>
 #include <rte_compressdev_pmd.h>
@@ -24,6 +25,14 @@
 #define MLX5_COMPRESS_DRIVER_NAME mlx5_compress
 #define MLX5_COMPRESS_LOG_NAME    pmd.compress.mlx5
 
+struct mlx5_compress_xform {
+	SLIST_ENTRY(mlx5_compress_xform) next;
+	enum rte_comp_xform_type type;
+	enum rte_comp_checksum_type csum_type;
+	uint32_t opcode;
+	uint32_t gga_ctrl1; /* BE. */
+};
+
 struct mlx5_compress_priv {
 	TAILQ_ENTRY(mlx5_compress_priv) next;
 	struct ibv_context *ctx; /* Device context. */
@@ -35,6 +44,8 @@ struct mlx5_compress_priv {
 	/* Minimum huffman block size supported by the device. */
 	struct ibv_pd *pd;
 	struct rte_compressdev_config dev_config;
+	SLIST_HEAD(xform_list, mlx5_compress_xform) xform_list;
+	rte_spinlock_t xform_sl;
 };
 
 struct mlx5_compress_qp {
@@ -215,6 +226,113 @@ struct mlx5_compress_qp {
 	return -1;
 }
 
+static int
+mlx5_compress_xform_free(struct rte_compressdev *dev, void *xform)
+{
+	struct mlx5_compress_priv *priv = dev->data->dev_private;
+
+	rte_spinlock_lock(&priv->xform_sl);
+	SLIST_REMOVE(&priv->xform_list, xform, mlx5_compress_xform, next);
+	rte_spinlock_unlock(&priv->xform_sl);
+	rte_free(xform);
+	return 0;
+}
+
+#define MLX5_COMP_MAX_WIN_SIZE_CONF 6u
+
+static int
+mlx5_compress_xform_create(struct rte_compressdev *dev,
+			    const struct rte_comp_xform *xform,
+			    void **private_xform)
+{
+	struct mlx5_compress_priv *priv = dev->data->dev_private;
+	struct mlx5_compress_xform *xfrm;
+	uint32_t size;
+
+	if (xform->type == RTE_COMP_COMPRESS && xform->compress.level ==
+							  RTE_COMP_LEVEL_NONE) {
+		DRV_LOG(ERR, "Non-compressed block is not supported.");
+		return -ENOTSUP;
+	}
+	if ((xform->type == RTE_COMP_COMPRESS && xform->compress.hash_algo !=
+	     RTE_COMP_HASH_ALGO_NONE) || (xform->type == RTE_COMP_DECOMPRESS &&
+		      xform->decompress.hash_algo != RTE_COMP_HASH_ALGO_NONE)) {
+		DRV_LOG(ERR, "SHA is not supported.");
+		return -ENOTSUP;
+	}
+	xfrm = rte_zmalloc_socket(__func__, sizeof(*xfrm), 0,
+						    priv->dev_config.socket_id);
+	if (xfrm == NULL)
+		return -ENOMEM;
+	xfrm->opcode = MLX5_OPCODE_MMO;
+	xfrm->type = xform->type;
+	switch (xform->type) {
+	case RTE_COMP_COMPRESS:
+		switch (xform->compress.algo) {
+		case RTE_COMP_ALGO_NULL:
+			xfrm->opcode += MLX5_OPC_MOD_MMO_DMA <<
+							WQE_CSEG_OPC_MOD_OFFSET;
+			break;
+		case RTE_COMP_ALGO_DEFLATE:
+			size = 1 << xform->compress.window_size;
+			size /= MLX5_GGA_COMP_WIN_SIZE_UNITS;
+			xfrm->gga_ctrl1 += RTE_MIN(rte_log2_u32(size),
+					 MLX5_COMP_MAX_WIN_SIZE_CONF) <<
+					   WQE_GGA_COMP_WIN_SIZE_OFFSET;
+			if (xform->compress.level == RTE_COMP_LEVEL_PMD_DEFAULT)
+				size = MLX5_GGA_COMP_LOG_BLOCK_SIZE_MAX;
+			else
+				size = priv->min_block_size - 1 +
+							  xform->compress.level;
+			xfrm->gga_ctrl1 += RTE_MIN(size,
+					    MLX5_GGA_COMP_LOG_BLOCK_SIZE_MAX) <<
+						 WQE_GGA_COMP_BLOCK_SIZE_OFFSET;
+			xfrm->opcode += MLX5_OPC_MOD_MMO_COMP <<
+							WQE_CSEG_OPC_MOD_OFFSET;
+			size = xform->compress.deflate.huffman ==
+						      RTE_COMP_HUFFMAN_DYNAMIC ?
+					    MLX5_GGA_COMP_LOG_DYNAMIC_SIZE_MAX :
+					     MLX5_GGA_COMP_LOG_DYNAMIC_SIZE_MIN;
+			xfrm->gga_ctrl1 += size <<
+					       WQE_GGA_COMP_DYNAMIC_SIZE_OFFSET;
+			break;
+		default:
+			goto err;
+		}
+		xfrm->csum_type = xform->compress.chksum;
+		break;
+	case RTE_COMP_DECOMPRESS:
+		switch (xform->decompress.algo) {
+		case RTE_COMP_ALGO_NULL:
+			xfrm->opcode += MLX5_OPC_MOD_MMO_DMA <<
+							WQE_CSEG_OPC_MOD_OFFSET;
+			break;
+		case RTE_COMP_ALGO_DEFLATE:
+			xfrm->opcode += MLX5_OPC_MOD_MMO_DECOMP <<
+							WQE_CSEG_OPC_MOD_OFFSET;
+			break;
+		default:
+			goto err;
+		}
+		xfrm->csum_type = xform->decompress.chksum;
+		break;
+	default:
+		DRV_LOG(ERR, "Algorithm %u is not supported.", xform->type);
+		goto err;
+	}
+	DRV_LOG(DEBUG, "New xform: gga ctrl1 = 0x%08X opcode = 0x%08X csum "
+		"type = %d.", xfrm->gga_ctrl1, xfrm->opcode, xfrm->csum_type);
+	xfrm->gga_ctrl1 = rte_cpu_to_be_32(xfrm->gga_ctrl1);
+	rte_spinlock_lock(&priv->xform_sl);
+	SLIST_INSERT_HEAD(&priv->xform_list, xfrm, next);
+	rte_spinlock_unlock(&priv->xform_sl);
+	*private_xform = xfrm;
+	return 0;
+err:
+	rte_free(xfrm);
+	return -ENOTSUP;
+}
+
 static struct rte_compressdev_ops mlx5_compress_ops = {
 	.dev_configure		= mlx5_compress_dev_configure,
 	.dev_start		= NULL,
@@ -225,8 +343,8 @@ struct mlx5_compress_qp {
 	.stats_reset		= NULL,
 	.queue_pair_setup	= mlx5_compress_qp_setup,
 	.queue_pair_release	= mlx5_compress_qp_release,
-	.private_xform_create	= NULL,
-	.private_xform_free	= NULL,
+	.private_xform_create	= mlx5_compress_xform_create,
+	.private_xform_free	= mlx5_compress_xform_free,
 	.stream_create		= NULL,
 	.stream_free		= NULL,
 };
Signature · 1 line
1.8.3.1
#7 lore raw

PATCH 7/10 compress/mlx5: add memory region management

In reply to Matan Azrad

Mellanox user space drivers don't deal with physical addresses, that's why any mbuf virtual address moved directly to the HW descriptor(WQE).

The mapping between the virtual address to the physical address is saved in MR configured by the kernel to the HW.

Each MR has a key that should also be moved to the WQE by the SW.

When the SW see address which is not mapped, it extends the address range and creates a MR using a system call.

Add memory region cache management:
2 level cache per queue-pair - no locks.
1 shared cache between all the queues using a lock.

Using this way, the MR key search per data-path address is optimized.

---
 drivers/compress/mlx5/mlx5_compress.c | 22 ++++++++++++++++++++++
 1 file changed, 22 insertions(+)
drivers/compress/mlx5/mlx5_compress.c +22 -0
index 132837e..ab24a84 100644
@@ -46,6 +46,7 @@ struct mlx5_compress_priv {
 	struct rte_compressdev_config dev_config;
 	SLIST_HEAD(xform_list, mlx5_compress_xform) xform_list;
 	rte_spinlock_t xform_sl;
+	struct mlx5_mr_share_cache mr_scache; /* Global shared MR cache. */
 };
 
 struct mlx5_compress_qp {
@@ -54,6 +55,7 @@ struct mlx5_compress_qp {
 	uint16_t pi;
 	uint16_t ci;
 	volatile uint64_t *uar_addr;
+	struct mlx5_mr_ctrl mr_ctrl;
 	int socket_id;
 	struct mlx5_devx_cq cq;
 	struct mlx5_devx_sq sq;
@@ -118,6 +120,7 @@ struct mlx5_compress_qp {
 		if (opaq != NULL)
 			rte_free(opaq);
 	}
+	mlx5_mr_btree_free(&qp->mr_ctrl.cache_bh);
 	rte_free(qp);
 	dev->data->queue_pairs[qp_id] = NULL;
 	return 0;
@@ -184,6 +187,13 @@ struct mlx5_compress_qp {
 		rte_errno = ENOMEM;
 		goto err;
 	}
+	if (mlx5_mr_btree_init(&qp->mr_ctrl.cache_bh, MLX5_MR_BTREE_CACHE_N,
+			       priv->dev_config.socket_id)) {
+		DRV_LOG(ERR, "Cannot allocate MR Btree for qp %u.",
+			(uint32_t)qp_id);
+		rte_errno = ENOMEM;
+		goto err;
+	}
 	qp->entries_n = 1 << log_ops_n;
 	qp->socket_id = socket_id;
 	qp->qp_id = qp_id;
@@ -513,6 +523,17 @@ struct mlx5_compress_qp {
 		claim_zero(mlx5_glue->close_device(priv->ctx));
 		return -1;
 	}
+	if (mlx5_mr_btree_init(&priv->mr_scache.cache,
+			     MLX5_MR_BTREE_CACHE_N * 2, rte_socket_id()) != 0) {
+		DRV_LOG(ERR, "Failed to allocate shared cache MR memory.");
+		mlx5_compress_hw_global_release(priv);
+		rte_compressdev_pmd_destroy(priv->cdev);
+		claim_zero(mlx5_glue->close_device(priv->ctx));
+		rte_errno = ENOMEM;
+		return -rte_errno;
+	}
+	priv->mr_scache.reg_mr_cb = mlx5_common_verbs_reg_mr;
+	priv->mr_scache.dereg_mr_cb = mlx5_common_verbs_dereg_mr;
 	pthread_mutex_lock(&priv_list_lock);
 	TAILQ_INSERT_TAIL(&mlx5_compress_priv_list, priv, next);
 	pthread_mutex_unlock(&priv_list_lock);
@@ -547,6 +568,7 @@ struct mlx5_compress_qp {
 		TAILQ_REMOVE(&mlx5_compress_priv_list, priv, next);
 	pthread_mutex_unlock(&priv_list_lock);
 	if (found != 0) {
+		mlx5_mr_release_cache(&priv->mr_scache);
 		mlx5_compress_hw_global_release(priv);
 		rte_compressdev_pmd_destroy(priv->cdev);
 		claim_zero(mlx5_glue->close_device(priv->ctx));
Signature · 1 line
1.8.3.1
#8 lore raw

PATCH 8/10 compress/mlx5: add data-path functions

In reply to Matan Azrad

Add implementation for the next compress data-path functions:
dequeue_burst
enqueue_burst

Add the next operation for starting \ stopping data-path:
dev_stop
dev_close

Each compress API enqueued operation is translated to a WQE. Once WQE is done, the HW sends CQE to the CQ, when SW see the CQE the operation will be updated and dequeued.

---
 drivers/compress/mlx5/mlx5_compress.c | 207 +++++++++++++++++++++++++++++++++-
 1 file changed, 203 insertions(+), 4 deletions(-)
drivers/compress/mlx5/mlx5_compress.c +203 -4
index ab24a84..719def2 100644
@@ -18,6 +18,7 @@
 #include <mlx5_common_os.h>
 #include <mlx5_common_devx.h>
 #include <mlx5_common_mr.h>
+#include <mlx5_common_os.h>
 #include <mlx5_prm.h>
 
 #include "mlx5_compress_utils.h"
@@ -343,10 +344,23 @@ struct mlx5_compress_qp {
 	return -ENOTSUP;
 }
 
+static void
+mlx5_compress_dev_stop(struct rte_compressdev *dev)
+{
+	RTE_SET_USED(dev);
+}
+
+static int
+mlx5_compress_dev_start(struct rte_compressdev *dev)
+{
+	RTE_SET_USED(dev);
+	return 0;
+}
+
 static struct rte_compressdev_ops mlx5_compress_ops = {
 	.dev_configure		= mlx5_compress_dev_configure,
-	.dev_start		= NULL,
-	.dev_stop		= NULL,
+	.dev_start		= mlx5_compress_dev_start,
+	.dev_stop		= mlx5_compress_dev_stop,
 	.dev_close		= mlx5_compress_dev_close,
 	.dev_infos_get		= mlx5_compress_dev_info_get,
 	.stats_get		= NULL,
@@ -359,6 +373,191 @@ struct mlx5_compress_qp {
 	.stream_free		= NULL,
 };
 
+static __rte_always_inline uint32_t
+mlx5_compress_dseg_set(struct mlx5_compress_qp *qp,
+		       volatile struct mlx5_wqe_dseg *restrict dseg,
+		       struct rte_mbuf *restrict mbuf,
+		       uint32_t offset, uint32_t len)
+
+{
+	uintptr_t addr = rte_pktmbuf_mtod_offset(mbuf, uintptr_t, offset);
+
+	dseg->bcount = rte_cpu_to_be_32(len);
+	dseg->lkey = mlx5_mr_addr2mr_bh(qp->priv->pd, 0, &qp->priv->mr_scache,
+					&qp->mr_ctrl, addr,
+					!!(mbuf->ol_flags & EXT_ATTACHED_MBUF));
+	dseg->pbuf = rte_cpu_to_be_64(addr);
+	return dseg->lkey;
+}
+
+static uint16_t
+mlx5_compress_enqueue_burst(void *queue_pair, struct rte_comp_op **ops,
+			    uint16_t nb_ops)
+{
+	struct mlx5_compress_qp *qp = queue_pair;
+	volatile struct mlx5_gga_wqe *wqes = (volatile struct mlx5_gga_wqe *)
+							      qp->sq.wqes, *wqe;
+	struct mlx5_compress_xform *xform;
+	struct rte_comp_op *op;
+	uint16_t mask = qp->entries_n - 1;
+	uint16_t remain = qp->entries_n - (qp->pi - qp->ci);
+
+	if (remain < nb_ops) {
+		ops[remain]->status = RTE_COMP_OP_STATUS_NOT_PROCESSED;
+		nb_ops = remain;
+	} else {
+		remain = nb_ops;
+	}
+	if (unlikely(remain == 0))
+		return 0;
+	do {
+		wqe = &wqes[qp->pi & mask];
+		rte_prefetch0(&wqes[(qp->pi + 1) & mask]);
+		op = *ops;
+		xform = op->private_xform;
+		if (unlikely(op->op_type != RTE_COMP_OP_STATELESS ||
+		    (xform->type == RTE_COMP_COMPRESS && op->flush_flag <
+		     RTE_COMP_FLUSH_FULL) || (mlx5_compress_dseg_set(qp,
+					&wqe->gather, op->m_src, op->src.offset,
+					       op->src.length) == UINT32_MAX) ||
+		    (mlx5_compress_dseg_set(qp, &wqe->scatter, op->m_dst,
+					    op->dst.offset,
+					    rte_pktmbuf_pkt_len(op->m_dst) -
+					    op->dst.offset) == UINT32_MAX))) {
+			op->status = RTE_COMP_OP_STATUS_INVALID_ARGS;
+			nb_ops -= remain;
+			if (unlikely(nb_ops == 0))
+				return 0;
+			break;
+		}
+		wqe->gga_ctrl1 = xform->gga_ctrl1;
+		wqe->opcode = rte_cpu_to_be_32(xform->opcode + (qp->pi << 8));
+		qp->ops[qp->pi & mask] = op;
+		++ops;
+		qp->pi++;
+	} while (--remain);
+	rte_io_wmb();
+	qp->sq.db_rec[MLX5_SND_DBR] = rte_cpu_to_be_32(qp->pi);
+	rte_wmb();
+	*qp->uar_addr = *(volatile uint64_t *)wqe; /* Assume 64 bit ARCH.*/
+	rte_wmb();
+	return nb_ops;
+}
+
+static void
+mlx5_compress_dump_err_objs(volatile uint32_t *cqe, volatile uint32_t *wqe,
+			     volatile uint32_t *opaq)
+{
+	size_t i;
+
+	DRV_LOG(ERR, "Error cqe:");
+	for (i = 0; i < sizeof(struct mlx5_err_cqe) >> 2; i += 4)
+		DRV_LOG(ERR, "%08X %08X %08X %08X", cqe[i], cqe[i + 1],
+			cqe[i + 2], cqe[i + 3]);
+	DRV_LOG(ERR, "\nError wqe:");
+	for (i = 0; i < sizeof(struct mlx5_gga_wqe) >> 2; i += 4)
+		DRV_LOG(ERR, "%08X %08X %08X %08X", wqe[i], wqe[i + 1],
+			wqe[i + 2], wqe[i + 3]);
+	DRV_LOG(ERR, "\nError opaq:");
+	for (i = 0; i < sizeof(struct mlx5_gga_compress_opaque) >> 2; i += 4)
+		DRV_LOG(ERR, "%08X %08X %08X %08X", opaq[i], opaq[i + 1],
+			opaq[i + 2], opaq[i + 3]);
+}
+
+static void
+mlx5_compress_cqe_err_handle(struct mlx5_compress_qp *qp,
+			     struct rte_comp_op *op)
+{
+	const uint32_t idx = qp->ci & (qp->entries_n - 1);
+	volatile struct mlx5_err_cqe *cqe = (volatile struct mlx5_err_cqe *)
+							      &qp->cq.cqes[idx];
+	volatile struct mlx5_gga_wqe *wqes = (volatile struct mlx5_gga_wqe *)
+								    qp->sq.wqes;
+	volatile struct mlx5_gga_compress_opaque *opaq = qp->opaque_mr.addr;
+
+	op->status = RTE_COMP_OP_STATUS_ERROR;
+	op->consumed = 0;
+	op->produced = 0;
+	op->output_chksum = 0;
+	op->debug_status = rte_be_to_cpu_32(opaq[idx].syndrom) |
+			      ((uint64_t)rte_be_to_cpu_32(cqe->syndrome) << 32);
+	mlx5_compress_dump_err_objs((volatile uint32_t *)cqe,
+				 (volatile uint32_t *)&wqes[idx],
+				 (volatile uint32_t *)&opaq[idx]);
+}
+
+static uint16_t
+mlx5_compress_dequeue_burst(void *queue_pair, struct rte_comp_op **ops,
+			    uint16_t nb_ops)
+{
+	struct mlx5_compress_qp *qp = queue_pair;
+	volatile struct mlx5_compress_xform *restrict xform;
+	volatile struct mlx5_cqe *restrict cqe;
+	volatile struct mlx5_gga_compress_opaque *opaq = qp->opaque_mr.addr;
+	struct rte_comp_op *restrict op;
+	const unsigned int cq_size = qp->entries_n;
+	const unsigned int mask = cq_size - 1;
+	uint32_t idx;
+	uint32_t next_idx = qp->ci & mask;
+	const uint16_t max = RTE_MIN((uint16_t)(qp->pi - qp->ci), nb_ops);
+	uint16_t i = 0;
+	int ret;
+
+	if (unlikely(max == 0))
+		return 0;
+	do {
+		idx = next_idx;
+		next_idx = (qp->ci + 1) & mask;
+		rte_prefetch0(&qp->cq.cqes[next_idx]);
+		rte_prefetch0(qp->ops[next_idx]);
+		op = qp->ops[idx];
+		cqe = &qp->cq.cqes[idx];
+		ret = check_cqe(cqe, cq_size, qp->ci);
+		/*
+		 * Be sure owner read is done before any other cookie field or
+		 * opaque field.
+		 */
+		rte_io_rmb();
+		if (unlikely(ret != MLX5_CQE_STATUS_SW_OWN)) {
+			if (likely(ret == MLX5_CQE_STATUS_HW_OWN))
+				break;
+			mlx5_compress_cqe_err_handle(qp, op);
+		} else {
+			xform = op->private_xform;
+			op->status = RTE_COMP_OP_STATUS_SUCCESS;
+			op->consumed = op->src.length;
+			op->produced = rte_be_to_cpu_32(cqe->byte_cnt);
+			MLX5_ASSERT(cqe->byte_cnt ==
+				    qp->opaque_buf[idx].scattered_length);
+			switch (xform->csum_type) {
+			case RTE_COMP_CHECKSUM_CRC32:
+				op->output_chksum = (uint64_t)rte_be_to_cpu_32
+						    (opaq[idx].crc32);
+				break;
+			case RTE_COMP_CHECKSUM_ADLER32:
+				op->output_chksum = (uint64_t)rte_be_to_cpu_32
+					    (opaq[idx].adler32) << 32;
+				break;
+			case RTE_COMP_CHECKSUM_CRC32_ADLER32:
+				op->output_chksum = (uint64_t)rte_be_to_cpu_32
+							     (opaq[idx].crc32) |
+						     ((uint64_t)rte_be_to_cpu_32
+						     (opaq[idx].adler32) << 32);
+				break;
+			default:
+				break;
+			}
+		}
+		ops[i++] = op;
+		qp->ci++;
+	} while (i < max);
+	if (likely(i != 0)) {
+		rte_io_wmb();
+		qp->cq.db_rec[0] = rte_cpu_to_be_32(qp->ci);
+	}
+	return i;
+}
+
 static struct ibv_device *
 mlx5_compress_get_ib_device_match(struct rte_pci_addr *addr)
 {
@@ -510,8 +709,8 @@ struct mlx5_compress_qp {
 	DRV_LOG(INFO,
 		"Compress device %s was created successfully.", ibv->name);
 	cdev->dev_ops = &mlx5_compress_ops;
-	cdev->dequeue_burst = NULL;
-	cdev->enqueue_burst = NULL;
+	cdev->dequeue_burst = mlx5_compress_dequeue_burst;
+	cdev->enqueue_burst = mlx5_compress_enqueue_burst;
 	cdev->feature_flags = RTE_COMPDEV_FF_HW_ACCELERATED;
 	priv = cdev->data->dev_private;
 	priv->ctx = ctx;
Signature · 1 line
1.8.3.1
#9 lore raw

PATCH 9/10 compress/mlx5: add statistics operations

In reply to Matan Azrad

Add support for the next statistics operations:
stats_get
stats_reset

These statistics are counted by the SW data-path.

---
 drivers/compress/mlx5/mlx5_compress.c | 36 +++++++++++++++++++++++++++++++++--
 1 file changed, 34 insertions(+), 2 deletions(-)
drivers/compress/mlx5/mlx5_compress.c +34 -2
index 719def2..d768453 100644
@@ -63,6 +63,7 @@ struct mlx5_compress_qp {
 	struct mlx5_pmd_mr opaque_mr;
 	struct rte_comp_op **ops;
 	struct mlx5_compress_priv *priv;
+	struct rte_compressdev_stats stats;
 };
 
 #define MLX5_COMPRESS_MAX_QPS 1024
@@ -357,14 +358,42 @@ struct mlx5_compress_qp {
 	return 0;
 }
 
+static void
+mlx5_compress_stats_get(struct rte_compressdev *dev,
+		struct rte_compressdev_stats *stats)
+{
+	int qp_id;
+
+	for (qp_id = 0; qp_id < dev->data->nb_queue_pairs; qp_id++) {
+		struct mlx5_compress_qp *qp = dev->data->queue_pairs[qp_id];
+
+		stats->enqueued_count += qp->stats.enqueued_count;
+		stats->dequeued_count += qp->stats.dequeued_count;
+		stats->enqueue_err_count += qp->stats.enqueue_err_count;
+		stats->dequeue_err_count += qp->stats.dequeue_err_count;
+	}
+}
+
+static void
+mlx5_compress_stats_reset(struct rte_compressdev *dev)
+{
+	int qp_id;
+
+	for (qp_id = 0; qp_id < dev->data->nb_queue_pairs; qp_id++) {
+		struct mlx5_compress_qp *qp = dev->data->queue_pairs[qp_id];
+
+		memset(&qp->stats, 0, sizeof(qp->stats));
+	}
+}
+
 static struct rte_compressdev_ops mlx5_compress_ops = {
 	.dev_configure		= mlx5_compress_dev_configure,
 	.dev_start		= mlx5_compress_dev_start,
 	.dev_stop		= mlx5_compress_dev_stop,
 	.dev_close		= mlx5_compress_dev_close,
 	.dev_infos_get		= mlx5_compress_dev_info_get,
-	.stats_get		= NULL,
-	.stats_reset		= NULL,
+	.stats_get		= mlx5_compress_stats_get,
+	.stats_reset		= mlx5_compress_stats_reset,
 	.queue_pair_setup	= mlx5_compress_qp_setup,
 	.queue_pair_release	= mlx5_compress_qp_release,
 	.private_xform_create	= mlx5_compress_xform_create,
@@ -436,6 +465,7 @@ struct mlx5_compress_qp {
 		++ops;
 		qp->pi++;
 	} while (--remain);
+	qp->stats.enqueued_count += nb_ops;
 	rte_io_wmb();
 	qp->sq.db_rec[MLX5_SND_DBR] = rte_cpu_to_be_32(qp->pi);
 	rte_wmb();
@@ -484,6 +514,7 @@ struct mlx5_compress_qp {
 	mlx5_compress_dump_err_objs((volatile uint32_t *)cqe,
 				 (volatile uint32_t *)&wqes[idx],
 				 (volatile uint32_t *)&opaq[idx]);
+	qp->stats.dequeue_err_count++;
 }
 
 static uint16_t
@@ -554,6 +585,7 @@ struct mlx5_compress_qp {
 	if (likely(i != 0)) {
 		rte_io_wmb();
 		qp->cq.db_rec[0] = rte_cpu_to_be_32(qp->ci);
+		qp->stats.dequeued_count += i;
 	}
 	return i;
 }
Signature · 1 line
1.8.3.1
#10 lore raw

PATCH 10/10 compress/mlx5: add the supported capabilities

In reply to Matan Azrad

Add all the capabilities supported by the device.

---
 drivers/compress/mlx5/mlx5_compress.c | 24 ++++++++++++++++++++++--
 1 file changed, 22 insertions(+), 2 deletions(-)
drivers/compress/mlx5/mlx5_compress.c +22 -2
index d768453..7384351 100644
@@ -74,8 +74,28 @@ struct mlx5_compress_qp {
 
 int mlx5_compress_logtype;
 
-const struct rte_compressdev_capabilities mlx5_caps[RTE_COMP_ALGO_LIST_END];
-
+static const struct rte_compressdev_capabilities mlx5_caps[] = {
+	{
+		.algo = RTE_COMP_ALGO_NULL,
+		.comp_feature_flags = RTE_COMP_FF_ADLER32_CHECKSUM |
+				      RTE_COMP_FF_CRC32_CHECKSUM |
+				      RTE_COMP_FF_CRC32_ADLER32_CHECKSUM |
+				      RTE_COMP_FF_SHAREABLE_PRIV_XFORM,
+	},
+	{
+		.algo = RTE_COMP_ALGO_DEFLATE,
+		.comp_feature_flags = RTE_COMP_FF_ADLER32_CHECKSUM |
+				      RTE_COMP_FF_CRC32_CHECKSUM |
+				      RTE_COMP_FF_CRC32_ADLER32_CHECKSUM |
+				      RTE_COMP_FF_SHAREABLE_PRIV_XFORM |
+				      RTE_COMP_FF_HUFFMAN_FIXED |
+				      RTE_COMP_FF_HUFFMAN_DYNAMIC,
+		.window_size = {.min = 10, .max = 15, .increment = 1},
+	},
+	{
+		.algo = RTE_COMP_ALGO_LIST_END,
+	}
+};
 
 static void
 mlx5_compress_dev_info_get(struct rte_compressdev *dev,
Signature · 1 line
1.8.3.1
#11 lore raw
In reply to Matan Azrad

+ Akhil

Regards,
Asaf Penso

Quoted 57 lines

-----Original Message-----
From: dev <[email protected]> On Behalf Of Matan Azrad
Sent: Monday, January 11, 2021 3:59 PM
To: [email protected]
Cc: NBU-Contact-Thomas Monjalon <[email protected]>; Ashish Gupta <[email protected]>; Fiona Trahe <[email protected]>
Subject: [dpdk-dev] [PATCH 00/10] add mlx5 compress PMD

Add a new compress PMD for Mellanox devices.

The MLX5 compress driver library provides support for Mellanox BlueField 2 families of 25/50/100/200 Gb/s adapters.

Using the BlueField 2 device, the compress class operations can be run in parallel to the net, vdpa, and regex class operations.

This driver is depending on rdma-core like the other mlx5 PMDs, also it is going to use mlx5 DevX to create HW objects directly by the FW.

Add the probing functions, PCI bus connectivity, HW capabilities checks and some basic objects preparations.

Documentations are not ready yet, will send later.

Matan Azrad (10):
common/mlx5: add DevX attributes for compress
drivers: introduce mlx5 compress PMD
compress/mlx5: support basic control operations
common/mlx5: add compress primitives
compress/mlx5: support queue pair operations
compress/mlx5: add transformation operations
compress/mlx5: add memory region management
compress/mlx5: add data-path functions
compress/mlx5: add statistics operations
compress/mlx5: add the supported capabilities

MAINTAINERS                                 |   4 +
drivers/common/mlx5/mlx5_common.h           |   1 +
drivers/common/mlx5/mlx5_common_pci.c       |   7 +
drivers/common/mlx5/mlx5_common_pci.h       |  36 +-
drivers/common/mlx5/mlx5_devx_cmds.c        |  10 +
drivers/common/mlx5/mlx5_devx_cmds.h        |   7 +
drivers/common/mlx5/mlx5_prm.h              |  59 +-
drivers/compress/meson.build                |   2 +-
drivers/compress/mlx5/meson.build           |  26 +
drivers/compress/mlx5/mlx5_compress.c       | 863
++++++++++++++++++++++++++++
drivers/compress/mlx5/mlx5_compress_utils.h |  20 +
drivers/compress/mlx5/version.map           |   3 +
12 files changed, 1015 insertions(+), 23 deletions(-)  create mode 100644

drivers/compress/mlx5/meson.build create mode 100644 drivers/compress/mlx5/mlx5_compress.c
create mode 100644 drivers/compress/mlx5/mlx5_compress_utils.h create mode 100644 drivers/compress/mlx5/version.map

--
1.8.3.1

#12 lore raw

v2 cover add mlx5 compress PMD

In reply to Matan Azrad

Add a new compress PMD for Mellanox devices.

The MLX5 compress driver library provides support for Mellanox BlueField 2 families of 25/50/100/200 Gb/s adapters.

Using the BlueField 2 device, the compress class operations can be run in parallel to the net, vdpa, and regex class operations.

This driver is depending on rdma-core like the other mlx5 PMDs, also it is going to use mlx5 DevX to create HW objects directly by the FW.

v2:
Add documentation.

Matan Azrad (10):
common/mlx5: add DevX attributes for compress
drivers: introduce mlx5 compress PMD
compress/mlx5: support basic control operations
common/mlx5: add compress primitives
compress/mlx5: support queue pair operations
compress/mlx5: add transformation operations
compress/mlx5: add memory region management
compress/mlx5: add data-path functions
compress/mlx5: add statistics operations
compress/mlx5: add the supported capabilities

 MAINTAINERS                                 |   4 +
 doc/guides/compressdevs/features/mlx5.ini   |  13 +
 doc/guides/compressdevs/index.rst           |   1 +
 doc/guides/compressdevs/mlx5.rst            |  84 +++
 doc/guides/rel_notes/release_21_02.rst      |   6 +
 drivers/common/mlx5/mlx5_common.h           |   1 +
 drivers/common/mlx5/mlx5_common_pci.c       |   7 +
 drivers/common/mlx5/mlx5_common_pci.h       |  36 +-
 drivers/common/mlx5/mlx5_devx_cmds.c        |  10 +
 drivers/common/mlx5/mlx5_devx_cmds.h        |   7 +
 drivers/common/mlx5/mlx5_prm.h              |  59 +-
 drivers/compress/meson.build                |   2 +-
 drivers/compress/mlx5/meson.build           |  26 +
 drivers/compress/mlx5/mlx5_compress.c       | 863 ++++++++++++++++++++++++++++
 drivers/compress/mlx5/mlx5_compress_utils.h |  20 +
 drivers/compress/mlx5/version.map           |   3 +
 16 files changed, 1119 insertions(+), 23 deletions(-)
 create mode 100644 doc/guides/compressdevs/features/mlx5.ini
 create mode 100644 doc/guides/compressdevs/mlx5.rst
 create mode 100644 drivers/compress/mlx5/meson.build
 create mode 100644 drivers/compress/mlx5/mlx5_compress.c
 create mode 100644 drivers/compress/mlx5/mlx5_compress_utils.h
 create mode 100644 drivers/compress/mlx5/version.map
Signature · 1 line
1.8.3.1
#13 lore raw

PATCH v2 1/10 common/mlx5: add DevX attributes for compress

Add the DevX attributes for compress related engiens:
compress
decompress
dma

---
 drivers/common/mlx5/mlx5_devx_cmds.c | 10 ++++++++++
 drivers/common/mlx5/mlx5_devx_cmds.h |  7 +++++++
 drivers/common/mlx5/mlx5_prm.h       | 18 ++++++++++++++++--
 3 files changed, 33 insertions(+), 2 deletions(-)
drivers/common/mlx5/mlx5_devx_cmds.c +10 -0
index 4d01f52..f3ed789 100644
@@ -725,6 +725,16 @@ struct mlx5_devx_obj *
 	attr->log_max_pd = MLX5_GET(cmd_hca_cap, hcattr, log_max_pd);
 	attr->log_max_srq = MLX5_GET(cmd_hca_cap, hcattr, log_max_srq);
 	attr->log_max_srq_sz = MLX5_GET(cmd_hca_cap, hcattr, log_max_srq_sz);
+	attr->mmo_dma_en = MLX5_GET(cmd_hca_cap, hcattr, dma_mmo);
+	attr->mmo_compress_en = MLX5_GET(cmd_hca_cap, hcattr, compress);
+	attr->mmo_decompress_en = MLX5_GET(cmd_hca_cap, hcattr, decompress);
+	attr->compress_min_block_size = MLX5_GET(cmd_hca_cap, hcattr,
+						 compress_min_block_size);
+	attr->log_max_mmo_dma = MLX5_GET(cmd_hca_cap, hcattr, log_dma_mmo_size);
+	attr->log_max_mmo_compress = MLX5_GET(cmd_hca_cap, hcattr,
+					      log_compress_mmo_size);
+	attr->log_max_mmo_decompress = MLX5_GET(cmd_hca_cap, hcattr,
+						log_decompress_mmo_size);
 	if (attr->qos.sup) {
 		MLX5_SET(query_hca_cap_in, in, op_mod,
 			 MLX5_GET_HCA_CAP_OP_MOD_QOS_CAP |
drivers/common/mlx5/mlx5_devx_cmds.h +7 -0
index 8d993df..31ad18a 100644
@@ -127,6 +127,13 @@ struct mlx5_hca_attr {
 	uint32_t log_max_srq;
 	uint32_t log_max_srq_sz;
 	uint32_t rss_ind_tbl_cap;
+	uint32_t mmo_dma_en:1;
+	uint32_t mmo_compress_en:1;
+	uint32_t mmo_decompress_en:1;
+	uint32_t compress_min_block_size:4;
+	uint32_t log_max_mmo_dma:5;
+	uint32_t log_max_mmo_compress:5;
+	uint32_t log_max_mmo_decompress:5;
 };
 
 struct mlx5_devx_wq_attr {
drivers/common/mlx5/mlx5_prm.h +16 -2
index 9e2d1d0..e489a0a 100644
@@ -1120,7 +1120,15 @@ enum {
 struct mlx5_ifc_cmd_hca_cap_bits {
 	u8 reserved_at_0[0x30];
 	u8 vhca_id[0x10];
-	u8 reserved_at_40[0x40];
+	u8 reserved_at_40[0x20];
+	u8 reserved_at_60[0x3];
+	u8 log_regexp_scatter_gather_size[0x5];
+	u8 reserved_at_68[0x3];
+	u8 log_dma_mmo_size[5];
+	u8 reserved_at_70[0x3];
+	u8 log_compress_mmo_size[5];
+	u8 reserved_at_78[0x3];
+	u8 log_decompress_mmo_size[5];
 	u8 log_max_srq_sz[0x8];
 	u8 log_max_qp_sz[0x8];
 	u8 reserved_at_90[0x9];
@@ -1168,7 +1176,13 @@ struct mlx5_ifc_cmd_hca_cap_bits {
 	u8 log_max_ra_res_dc[0x6];
 	u8 reserved_at_140[0xa];
 	u8 log_max_ra_req_qp[0x6];
-	u8 reserved_at_150[0xa];
+	u8 rtr2rts_qp_counters_set_id[1];
+	u8 rts2rts_udp_sport[1];
+	u8 rts2rts_lag_tx_port_affinity[1];
+	u8 dma_mmo[1];
+	u8 compress_min_block_size[4];
+	u8 compress[1];
+	u8 decompress[1];
 	u8 log_max_ra_res_qp[0x6];
 	u8 end_pad[0x1];
 	u8 cc_query_allowed[0x1];
Signature · 1 line
1.8.3.1
Acked-by on v2 1/10
#14 lore raw

Re: [dpdk-dev] [PATCH v2 01/10] common/mlx5: add DevX attributes for compress

Quoted 10 lines

-----Original Message-----
From: dev <[email protected]> On Behalf Of Matan Azrad
Sent: Wednesday, January 13, 2021 18:18
To: [email protected]
Cc: NBU-Contact-Thomas Monjalon <[email protected]>; Ashish Gupta <[email protected]>; Fiona Trahe <[email protected]>; [email protected]
Subject: [dpdk-dev] [PATCH v2 01/10] common/mlx5: add DevX attributes for compress

Quoted 6 lines

Add the DevX attributes for compress related engiens:
compress
decompress
dma

Typo: engiens -> engines
It would be nice to have hyphens:
- compress
- decompress
- dma

Beside this,

#15 lore raw

PATCH v2 2/10 drivers: introduce mlx5 compress PMD

In reply to Matan Azrad

Add a new compress PMD for Mellanox devices.

The MLX5 compress driver library provides support for Mellanox BlueField 2 families of 25/50/100/200 Gb/s adapters.

Using the BlueField 2 device, the compress class operations can be run in parallel to the net, vdpa, and regex class operations.

This driver is depending on rdma-core like the other mlx5 PMDs, also it is going to use mlx5 DevX to create HW objects directly by the FW.

Add the probing functions, PCI bus connectivity, HW capabilities checks and some basic objects preparations.

---
 MAINTAINERS                                 |   4 +
 drivers/common/mlx5/mlx5_common.h           |   1 +
 drivers/common/mlx5/mlx5_common_pci.c       |   7 +
 drivers/common/mlx5/mlx5_common_pci.h       |  36 ++--
 drivers/compress/meson.build                |   2 +-
 drivers/compress/mlx5/meson.build           |  26 +++
 drivers/compress/mlx5/mlx5_compress.c       | 297 ++++++++++++++++++++++++++++
 drivers/compress/mlx5/mlx5_compress_utils.h |  20 ++
 drivers/compress/mlx5/version.map           |   3 +
 9 files changed, 377 insertions(+), 19 deletions(-)
 create mode 100644 drivers/compress/mlx5/meson.build
 create mode 100644 drivers/compress/mlx5/mlx5_compress.c
 create mode 100644 drivers/compress/mlx5/mlx5_compress_utils.h
 create mode 100644 drivers/compress/mlx5/version.map
MAINTAINERS +4 -0
index 6787b15..f2badd9 100644
@@ -1133,6 +1133,10 @@ F: drivers/compress/zlib/
 F: doc/guides/compressdevs/zlib.rst
 F: doc/guides/compressdevs/features/zlib.ini
 
+Mellanox mlx5
+M: Matan Azrad <[email protected]>
+F: drivers/compress/mlx5/
+
 
 RegEx Drivers
 -------------
drivers/common/mlx5/mlx5_common.h +1 -0
index e35188d..3855582 100644
@@ -217,6 +217,7 @@ enum mlx5_class {
 	MLX5_CLASS_NET = RTE_BIT64(0),
 	MLX5_CLASS_VDPA = RTE_BIT64(1),
 	MLX5_CLASS_REGEX = RTE_BIT64(2),
+	MLX5_CLASS_COMPRESS = RTE_BIT64(3),
 };
 
 #define MLX5_DBR_SIZE RTE_CACHE_LINE_SIZE
drivers/common/mlx5/mlx5_common_pci.c +7 -0
index 5208972..2b65768 100644
@@ -28,14 +28,21 @@ static TAILQ_HEAD(mlx5_pci_devices_head, mlx5_pci_device) devices_list =
 	{ .name = "vdpa", .driver_class = MLX5_CLASS_VDPA },
 	{ .name = "net", .driver_class = MLX5_CLASS_NET },
 	{ .name = "regex", .driver_class = MLX5_CLASS_REGEX },
+	{ .name = "compress", .driver_class = MLX5_CLASS_COMPRESS },
 };
 
 static const unsigned int mlx5_class_combinations[] = {
 	MLX5_CLASS_NET,
 	MLX5_CLASS_VDPA,
 	MLX5_CLASS_REGEX,
+	MLX5_CLASS_COMPRESS,
 	MLX5_CLASS_NET | MLX5_CLASS_REGEX,
 	MLX5_CLASS_VDPA | MLX5_CLASS_REGEX,
+	MLX5_CLASS_NET | MLX5_CLASS_COMPRESS,
+	MLX5_CLASS_VDPA | MLX5_CLASS_COMPRESS,
+	MLX5_CLASS_REGEX | MLX5_CLASS_COMPRESS,
+	MLX5_CLASS_NET | MLX5_CLASS_REGEX | MLX5_CLASS_COMPRESS,
+	MLX5_CLASS_VDPA | MLX5_CLASS_REGEX | MLX5_CLASS_COMPRESS,
 	/* New class combination should be added here. */
 };
 
drivers/common/mlx5/mlx5_common_pci.h +18 -18
index 41b73e1..de89bb9 100644
@@ -9,26 +9,26 @@
  * @file
  *
  * RTE Mellanox PCI Driver Interface
- * Mellanox ConnectX PCI device supports multiple class (net/vdpa/regex)
- * devices. This layer enables creating such multiple class of devices on a
- * single PCI device by allowing to bind multiple class specific device
+ * Mellanox ConnectX PCI device supports multiple class: net,vdpa,regex and
+ * compress devices. This layer enables creating such multiple class of devices
+ * on a single PCI device by allowing to bind multiple class specific device
  * driver to attach to mlx5_pci driver.
  *
- * -----------    ------------    -------------
- * |   mlx5  |    |   mlx5   |    |   mlx5    |
- * | net pmd |    | vdpa pmd |    | regex pmd |
- * -----------    ------------    -------------
- *      \              |                 /
- *       \             |                /
- *        \       --------------       /
- *         \______|   mlx5     |_____ /
- *                | pci common |
- *                --------------
- *                     |
- *                 -----------
- *                 |   mlx5  |
- *                 | pci dev |
- *                 -----------
+ * -----------    ------------    -------------    ----------------
+ * |   mlx5  |    |   mlx5   |    |   mlx5    |    |     mlx5     |
+ * | net pmd |    | vdpa pmd |    | regex pmd |    | compress pmd |
+ * -----------    ------------    -------------    ----------------
+ *      \              \                    /              /
+ *       \              \                  /              /
+ *        \              \_--------------_/              /
+ *         \_______________|   mlx5     |_______________/
+ *                         | pci common |
+ *                         --------------
+ *                               |
+ *                           -----------
+ *                           |   mlx5  |
+ *                           | pci dev |
+ *                           -----------
  *
  * - mlx5 pci driver binds to mlx5 PCI devices defined by PCI
  *   ID table of all related mlx5 PCI devices.
drivers/compress/meson.build +1 -1
index 33f5e33..d8f3ddb 100644
@@ -5,7 +5,7 @@ if is_windows
 	subdir_done()
 endif
 
-drivers = ['isal', 'octeontx', 'zlib']
+drivers = ['isal', 'octeontx', 'zlib', 'mlx5']
 
 std_deps = ['compressdev'] # compressdev pulls in all other needed deps
 config_flag_fmt = 'RTE_LIBRTE_PMD_@0@'
drivers/compress/mlx5/meson.build new file +26 -0
new file mode 100644
index 0000000..2a6dc3f
@@ -0,0 +1,26 @@
+# SPDX-License-Identifier: BSD-3-Clause
+# Copyright 2021 Mellanox Technologies, Ltd
+
+if not is_linux
+	build = false
+	reason = 'only supported on Linux'
+	subdir_done()
+endif
+
+fmt_name = 'mlx5_compress'
+deps += ['common_mlx5', 'eal', 'compressdev']
+sources = files(
+	'mlx5_compress.c',
+)
+cflags_options = [
+	'-std=c11',
+	'-Wno-strict-prototypes',
+	'-D_BSD_SOURCE',
+	'-D_DEFAULT_SOURCE',
+	'-D_XOPEN_SOURCE=600'
+]
+foreach option:cflags_options
+	if cc.has_argument(option)
+		cflags += option
+	endif
+endforeach
drivers/compress/mlx5/mlx5_compress.c new file +297 -0

297 lines not on this page · raw message

drivers/compress/mlx5/mlx5_compress_utils.h new file +20 -0
new file mode 100644
index 0000000..f93244f
@@ -0,0 +1,20 @@
+/* SPDX-License-Identifier: BSD-3-Clause
+ * Copyright 2021 Mellanox Technologies, Ltd
+ */
+
+#ifndef RTE_PMD_MLX5_COMPRESS_UTILS_H_
+#define RTE_PMD_MLX5_COMPRESS_UTILS_H_
+
+#include <mlx5_common.h>
+
+
+extern int mlx5_compress_logtype;
+
+#define MLX5_COMPRESS_LOG_PREFIX "mlx5_compress"
+/* Generic printf()-like logging macro with automatic line feed. */
+#define DRV_LOG(level, ...) \
+	PMD_DRV_LOG_(level, mlx5_compress_logtype, MLX5_COMPRESS_LOG_PREFIX, \
+		__VA_ARGS__ PMD_DRV_LOG_STRIP PMD_DRV_LOG_OPAREN, \
+		PMD_DRV_LOG_CPAREN)
+
+#endif /* RTE_PMD_MLX5_COMPRESS_UTILS_H_ */
drivers/compress/mlx5/version.map new file +3 -0
new file mode 100644
index 0000000..4a76d1d
@@ -0,0 +1,3 @@
+DPDK_21 {
+	local: *;
+};
Signature · 1 line
1.8.3.1
#16 lore raw

PATCH v2 3/10 compress/mlx5: support basic control operations

In reply to Matan Azrad

Add initial support for the next operations:
dev_configure
dev_close
dev_infos_get

---
 drivers/compress/mlx5/mlx5_compress.c | 41 ++++++++++++++++++++++++++++++++---
 1 file changed, 38 insertions(+), 3 deletions(-)
drivers/compress/mlx5/mlx5_compress.c +38 -3
index 639dd61..7148798 100644
@@ -32,20 +32,55 @@ struct mlx5_compress_priv {
 	uint8_t min_block_size;
 	/* Minimum huffman block size supported by the device. */
 	struct ibv_pd *pd;
+	struct rte_compressdev_config dev_config;
 };
 
+#define MLX5_COMPRESS_MAX_QPS 1024
+
 TAILQ_HEAD(mlx5_compress_privs, mlx5_compress_priv) mlx5_compress_priv_list =
 				TAILQ_HEAD_INITIALIZER(mlx5_compress_priv_list);
 static pthread_mutex_t priv_list_lock = PTHREAD_MUTEX_INITIALIZER;
 
 int mlx5_compress_logtype;
 
+const struct rte_compressdev_capabilities mlx5_caps[RTE_COMP_ALGO_LIST_END];
+
+
+static void
+mlx5_compress_dev_info_get(struct rte_compressdev *dev,
+			   struct rte_compressdev_info *info)
+{
+	RTE_SET_USED(dev);
+	if (info != NULL) {
+		info->max_nb_queue_pairs = MLX5_COMPRESS_MAX_QPS;
+		info->feature_flags = RTE_COMPDEV_FF_HW_ACCELERATED;
+		info->capabilities = mlx5_caps;
+	}
+}
+
+static int
+mlx5_compress_dev_configure(struct rte_compressdev *dev,
+			    struct rte_compressdev_config *config)
+{
+	struct mlx5_compress_priv *priv = dev->data->dev_private;
+
+	priv->dev_config = *config;
+	return 0;
+}
+
+static int
+mlx5_compress_dev_close(struct rte_compressdev *dev)
+{
+	RTE_SET_USED(dev);
+	return 0;
+}
+
 static struct rte_compressdev_ops mlx5_compress_ops = {
-	.dev_configure		= NULL,
+	.dev_configure		= mlx5_compress_dev_configure,
 	.dev_start		= NULL,
 	.dev_stop		= NULL,
-	.dev_close		= NULL,
-	.dev_infos_get		= NULL,
+	.dev_close		= mlx5_compress_dev_close,
+	.dev_infos_get		= mlx5_compress_dev_info_get,
 	.stats_get		= NULL,
 	.stats_reset		= NULL,
 	.queue_pair_setup	= NULL,
Signature · 1 line
1.8.3.1
#17 lore raw

PATCH v2 4/10 common/mlx5: add compress primitives

In reply to Matan Azrad

Add the GGA compress WQE related structures and definitions.

---
 drivers/common/mlx5/mlx5_prm.h | 41 +++++++++++++++++++++++++++++++++++++++--
 1 file changed, 39 insertions(+), 2 deletions(-)
drivers/common/mlx5/mlx5_prm.h +39 -2
index e489a0a..15ef0dc 100644
@@ -412,10 +412,23 @@ struct mlx5_cqe_ts {
 	uint8_t op_own;
 };
 
+/* GGA */
 /* MMO metadata segment */
 
-#define	MLX5_OPCODE_MMO	0x2f
-#define	MLX5_OPC_MOD_MMO_REGEX 0x4
+#define	MLX5_OPCODE_MMO	0x2fu
+#define	MLX5_OPC_MOD_MMO_REGEX 0x4u
+#define	MLX5_OPC_MOD_MMO_COMP 0x2u
+#define	MLX5_OPC_MOD_MMO_DECOMP 0x3u
+#define	MLX5_OPC_MOD_MMO_DMA 0x1u
+
+#define WQE_GGA_COMP_WIN_SIZE_OFFSET 12u
+#define WQE_GGA_COMP_BLOCK_SIZE_OFFSET 16u
+#define WQE_GGA_COMP_DYNAMIC_SIZE_OFFSET 20u
+#define MLX5_GGA_COMP_WIN_SIZE_UNITS 1024u
+#define MLX5_GGA_COMP_WIN_SIZE_MAX (32u * MLX5_GGA_COMP_WIN_SIZE_UNITS)
+#define MLX5_GGA_COMP_LOG_BLOCK_SIZE_MAX 15u
+#define MLX5_GGA_COMP_LOG_DYNAMIC_SIZE_MAX 15u
+#define MLX5_GGA_COMP_LOG_DYNAMIC_SIZE_MIN 0u
 
 struct mlx5_wqe_metadata_seg {
 	uint32_t mmo_control_31_0; /* mmo_control_63_32 is in ctrl_seg.imm */
@@ -423,6 +436,30 @@ struct mlx5_wqe_metadata_seg {
 	uint64_t addr;
 };
 
+struct mlx5_gga_wqe {
+	uint32_t opcode;
+	uint32_t sq_ds;
+	uint32_t flags;
+	uint32_t gga_ctrl1;  /* ws 12-15, bs 16-19, dyns 20-23. */
+	uint32_t gga_ctrl2;
+	uint32_t opaque_lkey;
+	uint64_t opaque_vaddr;
+	struct mlx5_wqe_dseg gather;
+	struct mlx5_wqe_dseg scatter;
+} __rte_packed;
+
+struct mlx5_gga_compress_opaque {
+	uint32_t syndrom;
+	uint32_t reserved0;
+	uint32_t scattered_length;
+	uint32_t gathered_length;
+	uint64_t scatter_crc;
+	uint64_t gather_crc;
+	uint32_t crc32;
+	uint32_t adler32;
+	uint8_t reserved1[216];
+} __rte_packed;
+
 struct mlx5_ifc_regexp_mmo_control_bits {
 	uint8_t reserved_at_31[0x2];
 	uint8_t le[0x1];
Signature · 1 line
1.8.3.1
#18 lore raw

PATCH v2 5/10 compress/mlx5: support queue pair operations

In reply to Matan Azrad

Add support for the next operations:
queue_pair_setup
queue_pair_release

Create and initialize DevX SQ and CQ for each compress API queue-pair.

---
 drivers/compress/mlx5/mlx5_compress.c | 144 +++++++++++++++++++++++++++++++++-
 1 file changed, 142 insertions(+), 2 deletions(-)
drivers/compress/mlx5/mlx5_compress.c +142 -2
index 7148798..ffd866a 100644
@@ -15,6 +15,8 @@
 #include <mlx5_common_pci.h>
 #include <mlx5_devx_cmds.h>
 #include <mlx5_common_os.h>
+#include <mlx5_common_devx.h>
+#include <mlx5_common_mr.h>
 #include <mlx5_prm.h>
 
 #include "mlx5_compress_utils.h"
@@ -35,6 +37,20 @@ struct mlx5_compress_priv {
 	struct rte_compressdev_config dev_config;
 };
 
+struct mlx5_compress_qp {
+	uint16_t qp_id;
+	uint16_t entries_n;
+	uint16_t pi;
+	uint16_t ci;
+	volatile uint64_t *uar_addr;
+	int socket_id;
+	struct mlx5_devx_cq cq;
+	struct mlx5_devx_sq sq;
+	struct mlx5_pmd_mr opaque_mr;
+	struct rte_comp_op **ops;
+	struct mlx5_compress_priv *priv;
+};
+
 #define MLX5_COMPRESS_MAX_QPS 1024
 
 TAILQ_HEAD(mlx5_compress_privs, mlx5_compress_priv) mlx5_compress_priv_list =
@@ -75,6 +91,130 @@ struct mlx5_compress_priv {
 	return 0;
 }
 
+static int
+mlx5_compress_qp_release(struct rte_compressdev *dev, uint16_t qp_id)
+{
+	struct mlx5_compress_qp *qp = dev->data->queue_pairs[qp_id];
+
+	if (qp->sq.sq != NULL)
+		mlx5_devx_sq_destroy(&qp->sq);
+	if (qp->cq.cq != NULL)
+		mlx5_devx_cq_destroy(&qp->cq);
+	if (qp->opaque_mr.obj != NULL) {
+		void *opaq = qp->opaque_mr.addr;
+
+		mlx5_common_verbs_dereg_mr(&qp->opaque_mr);
+		if (opaq != NULL)
+			rte_free(opaq);
+	}
+	rte_free(qp);
+	dev->data->queue_pairs[qp_id] = NULL;
+	return 0;
+}
+
+static void
+mlx5_compress_init_sq(struct mlx5_compress_qp *qp)
+{
+	volatile struct mlx5_gga_wqe *restrict wqe =
+				    (volatile struct mlx5_gga_wqe *)qp->sq.wqes;
+	volatile struct mlx5_gga_compress_opaque *opaq = qp->opaque_mr.addr;
+	int i;
+
+	/* All the next fields state should stay constant. */
+	for (i = 0; i < qp->entries_n; ++i, ++wqe) {
+		wqe->sq_ds = rte_cpu_to_be_32((qp->sq.sq->id << 8) | 4u);
+		wqe->flags = RTE_BE32(MLX5_COMP_ALWAYS <<
+				      MLX5_COMP_MODE_OFFSET);
+		wqe->opaque_lkey = rte_cpu_to_be_32(qp->opaque_mr.lkey);
+		wqe->opaque_vaddr = rte_cpu_to_be_64
+						((uint64_t)(uintptr_t)&opaq[i]);
+	}
+}
+
+static int
+mlx5_compress_qp_setup(struct rte_compressdev *dev, uint16_t qp_id,
+		       uint32_t max_inflight_ops, int socket_id)
+{
+	struct mlx5_compress_priv *priv = dev->data->dev_private;
+	struct mlx5_compress_qp *qp;
+	struct mlx5_devx_cq_attr cq_attr = {
+		.uar_page_id = mlx5_os_get_devx_uar_page_id(priv->uar),
+	};
+	struct mlx5_devx_create_sq_attr sq_attr = {
+		.user_index = qp_id,
+		.wq_attr = (struct mlx5_devx_wq_attr){
+			.pd = priv->pdn,
+			.uar_page = mlx5_os_get_devx_uar_page_id(priv->uar),
+		},
+	};
+	struct mlx5_devx_modify_sq_attr modify_attr = {
+		.state = MLX5_SQC_STATE_RDY,
+	};
+	uint32_t log_ops_n = rte_log2_u32(max_inflight_ops);
+	uint32_t alloc_size = sizeof(*qp);
+	void *opaq_buf;
+	int ret;
+
+	alloc_size = RTE_ALIGN(alloc_size, RTE_CACHE_LINE_SIZE);
+	alloc_size += sizeof(struct rte_comp_op *) * (1u << log_ops_n);
+	qp = rte_zmalloc_socket(__func__, alloc_size, RTE_CACHE_LINE_SIZE,
+				socket_id);
+	if (qp == NULL) {
+		DRV_LOG(ERR, "Failed to allocate qp memory.");
+		rte_errno = ENOMEM;
+		return -rte_errno;
+	}
+	dev->data->queue_pairs[qp_id] = qp;
+	opaq_buf = rte_calloc(__func__, 1u << log_ops_n,
+			      sizeof(struct mlx5_gga_compress_opaque),
+			      sizeof(struct mlx5_gga_compress_opaque));
+	if (opaq_buf == NULL) {
+		DRV_LOG(ERR, "Failed to allocate opaque memory.");
+		rte_errno = ENOMEM;
+		goto err;
+	}
+	qp->entries_n = 1 << log_ops_n;
+	qp->socket_id = socket_id;
+	qp->qp_id = qp_id;
+	qp->priv = priv;
+	qp->ops = (struct rte_comp_op **)RTE_ALIGN((uintptr_t)(qp + 1),
+						   RTE_CACHE_LINE_SIZE);
+	qp->uar_addr = mlx5_os_get_devx_uar_reg_addr(priv->uar);
+	MLX5_ASSERT(qp->uar_addr);
+	if (mlx5_common_verbs_reg_mr(priv->pd, opaq_buf, qp->entries_n *
+					sizeof(struct mlx5_gga_compress_opaque),
+							 &qp->opaque_mr) != 0) {
+		DRV_LOG(ERR, "Failed to register opaque MR.");
+		rte_errno = ENOMEM;
+		goto err;
+	}
+	ret = mlx5_devx_cq_create(priv->ctx, &qp->cq, log_ops_n, &cq_attr,
+				  socket_id);
+	if (ret != 0) {
+		DRV_LOG(ERR, "Failed to create CQ.");
+		goto err;
+	}
+	sq_attr.cqn = qp->cq.cq->id;
+	ret = mlx5_devx_sq_create(priv->ctx, &qp->sq, log_ops_n, &sq_attr,
+				  socket_id);
+	if (ret != 0) {
+		DRV_LOG(ERR, "Failed to create SQ.");
+		goto err;
+	}
+	mlx5_compress_init_sq(qp);
+	ret = mlx5_devx_cmd_modify_sq(qp->sq.sq, &modify_attr);
+	if (ret != 0) {
+		DRV_LOG(ERR, "Can't change SQ state to ready.");
+		goto err;
+	}
+	DRV_LOG(INFO, "QP %u: SQN=0x%X CQN=0x%X entries num = %u\n",
+		(uint32_t)qp_id, qp->sq.sq->id, qp->cq.cq->id, qp->entries_n);
+	return 0;
+err:
+	mlx5_compress_qp_release(dev, qp_id);
+	return -1;
+}
+
 static struct rte_compressdev_ops mlx5_compress_ops = {
 	.dev_configure		= mlx5_compress_dev_configure,
 	.dev_start		= NULL,
@@ -83,8 +223,8 @@ struct mlx5_compress_priv {
 	.dev_infos_get		= mlx5_compress_dev_info_get,
 	.stats_get		= NULL,
 	.stats_reset		= NULL,
-	.queue_pair_setup	= NULL,
-	.queue_pair_release	= NULL,
+	.queue_pair_setup	= mlx5_compress_qp_setup,
+	.queue_pair_release	= mlx5_compress_qp_release,
 	.private_xform_create	= NULL,
 	.private_xform_free	= NULL,
 	.stream_create		= NULL,
Signature · 1 line
1.8.3.1
#19 lore raw

PATCH v2 6/10 compress/mlx5: add transformation operations

In reply to Matan Azrad

Add support for the next operations:
private_xform_create
private_xform_free

The driver transformation structure includes preparations for the next GGA WQE fields used by the enqueue function: opcode. compress specific fields checksum type and compress type.

---
 drivers/compress/mlx5/mlx5_compress.c | 122 +++++++++++++++++++++++++++++++++-
 1 file changed, 120 insertions(+), 2 deletions(-)
drivers/compress/mlx5/mlx5_compress.c +120 -2
index ffd866a..132837e 100644
@@ -6,6 +6,7 @@
 #include <rte_log.h>
 #include <rte_errno.h>
 #include <rte_pci.h>
+#include <rte_spinlock.h>
 #include <rte_comp.h>
 #include <rte_compressdev.h>
 #include <rte_compressdev_pmd.h>
@@ -24,6 +25,14 @@
 #define MLX5_COMPRESS_DRIVER_NAME mlx5_compress
 #define MLX5_COMPRESS_LOG_NAME    pmd.compress.mlx5
 
+struct mlx5_compress_xform {
+	SLIST_ENTRY(mlx5_compress_xform) next;
+	enum rte_comp_xform_type type;
+	enum rte_comp_checksum_type csum_type;
+	uint32_t opcode;
+	uint32_t gga_ctrl1; /* BE. */
+};
+
 struct mlx5_compress_priv {
 	TAILQ_ENTRY(mlx5_compress_priv) next;
 	struct ibv_context *ctx; /* Device context. */
@@ -35,6 +44,8 @@ struct mlx5_compress_priv {
 	/* Minimum huffman block size supported by the device. */
 	struct ibv_pd *pd;
 	struct rte_compressdev_config dev_config;
+	SLIST_HEAD(xform_list, mlx5_compress_xform) xform_list;
+	rte_spinlock_t xform_sl;
 };
 
 struct mlx5_compress_qp {
@@ -215,6 +226,113 @@ struct mlx5_compress_qp {
 	return -1;
 }
 
+static int
+mlx5_compress_xform_free(struct rte_compressdev *dev, void *xform)
+{
+	struct mlx5_compress_priv *priv = dev->data->dev_private;
+
+	rte_spinlock_lock(&priv->xform_sl);
+	SLIST_REMOVE(&priv->xform_list, xform, mlx5_compress_xform, next);
+	rte_spinlock_unlock(&priv->xform_sl);
+	rte_free(xform);
+	return 0;
+}
+
+#define MLX5_COMP_MAX_WIN_SIZE_CONF 6u
+
+static int
+mlx5_compress_xform_create(struct rte_compressdev *dev,
+			    const struct rte_comp_xform *xform,
+			    void **private_xform)
+{
+	struct mlx5_compress_priv *priv = dev->data->dev_private;
+	struct mlx5_compress_xform *xfrm;
+	uint32_t size;
+
+	if (xform->type == RTE_COMP_COMPRESS && xform->compress.level ==
+							  RTE_COMP_LEVEL_NONE) {
+		DRV_LOG(ERR, "Non-compressed block is not supported.");
+		return -ENOTSUP;
+	}
+	if ((xform->type == RTE_COMP_COMPRESS && xform->compress.hash_algo !=
+	     RTE_COMP_HASH_ALGO_NONE) || (xform->type == RTE_COMP_DECOMPRESS &&
+		      xform->decompress.hash_algo != RTE_COMP_HASH_ALGO_NONE)) {
+		DRV_LOG(ERR, "SHA is not supported.");
+		return -ENOTSUP;
+	}
+	xfrm = rte_zmalloc_socket(__func__, sizeof(*xfrm), 0,
+						    priv->dev_config.socket_id);
+	if (xfrm == NULL)
+		return -ENOMEM;
+	xfrm->opcode = MLX5_OPCODE_MMO;
+	xfrm->type = xform->type;
+	switch (xform->type) {
+	case RTE_COMP_COMPRESS:
+		switch (xform->compress.algo) {
+		case RTE_COMP_ALGO_NULL:
+			xfrm->opcode += MLX5_OPC_MOD_MMO_DMA <<
+							WQE_CSEG_OPC_MOD_OFFSET;
+			break;
+		case RTE_COMP_ALGO_DEFLATE:
+			size = 1 << xform->compress.window_size;
+			size /= MLX5_GGA_COMP_WIN_SIZE_UNITS;
+			xfrm->gga_ctrl1 += RTE_MIN(rte_log2_u32(size),
+					 MLX5_COMP_MAX_WIN_SIZE_CONF) <<
+					   WQE_GGA_COMP_WIN_SIZE_OFFSET;
+			if (xform->compress.level == RTE_COMP_LEVEL_PMD_DEFAULT)
+				size = MLX5_GGA_COMP_LOG_BLOCK_SIZE_MAX;
+			else
+				size = priv->min_block_size - 1 +
+							  xform->compress.level;
+			xfrm->gga_ctrl1 += RTE_MIN(size,
+					    MLX5_GGA_COMP_LOG_BLOCK_SIZE_MAX) <<
+						 WQE_GGA_COMP_BLOCK_SIZE_OFFSET;
+			xfrm->opcode += MLX5_OPC_MOD_MMO_COMP <<
+							WQE_CSEG_OPC_MOD_OFFSET;
+			size = xform->compress.deflate.huffman ==
+						      RTE_COMP_HUFFMAN_DYNAMIC ?
+					    MLX5_GGA_COMP_LOG_DYNAMIC_SIZE_MAX :
+					     MLX5_GGA_COMP_LOG_DYNAMIC_SIZE_MIN;
+			xfrm->gga_ctrl1 += size <<
+					       WQE_GGA_COMP_DYNAMIC_SIZE_OFFSET;
+			break;
+		default:
+			goto err;
+		}
+		xfrm->csum_type = xform->compress.chksum;
+		break;
+	case RTE_COMP_DECOMPRESS:
+		switch (xform->decompress.algo) {
+		case RTE_COMP_ALGO_NULL:
+			xfrm->opcode += MLX5_OPC_MOD_MMO_DMA <<
+							WQE_CSEG_OPC_MOD_OFFSET;
+			break;
+		case RTE_COMP_ALGO_DEFLATE:
+			xfrm->opcode += MLX5_OPC_MOD_MMO_DECOMP <<
+							WQE_CSEG_OPC_MOD_OFFSET;
+			break;
+		default:
+			goto err;
+		}
+		xfrm->csum_type = xform->decompress.chksum;
+		break;
+	default:
+		DRV_LOG(ERR, "Algorithm %u is not supported.", xform->type);
+		goto err;
+	}
+	DRV_LOG(DEBUG, "New xform: gga ctrl1 = 0x%08X opcode = 0x%08X csum "
+		"type = %d.", xfrm->gga_ctrl1, xfrm->opcode, xfrm->csum_type);
+	xfrm->gga_ctrl1 = rte_cpu_to_be_32(xfrm->gga_ctrl1);
+	rte_spinlock_lock(&priv->xform_sl);
+	SLIST_INSERT_HEAD(&priv->xform_list, xfrm, next);
+	rte_spinlock_unlock(&priv->xform_sl);
+	*private_xform = xfrm;
+	return 0;
+err:
+	rte_free(xfrm);
+	return -ENOTSUP;
+}
+
 static struct rte_compressdev_ops mlx5_compress_ops = {
 	.dev_configure		= mlx5_compress_dev_configure,
 	.dev_start		= NULL,
@@ -225,8 +343,8 @@ struct mlx5_compress_qp {
 	.stats_reset		= NULL,
 	.queue_pair_setup	= mlx5_compress_qp_setup,
 	.queue_pair_release	= mlx5_compress_qp_release,
-	.private_xform_create	= NULL,
-	.private_xform_free	= NULL,
+	.private_xform_create	= mlx5_compress_xform_create,
+	.private_xform_free	= mlx5_compress_xform_free,
 	.stream_create		= NULL,
 	.stream_free		= NULL,
 };
Signature · 1 line
1.8.3.1
#20 lore raw

PATCH v2 7/10 compress/mlx5: add memory region management

In reply to Matan Azrad

Mellanox user space drivers don't deal with physical addresses, that's why any mbuf virtual address moved directly to the HW descriptor(WQE).

The mapping between the virtual address to the physical address is saved in MR configured by the kernel to the HW.

Each MR has a key that should also be moved to the WQE by the SW.

When the SW see address which is not mapped, it extends the address range and creates a MR using a system call.

Add memory region cache management:
2 level cache per queue-pair - no locks.
1 shared cache between all the queues using a lock.

Using this way, the MR key search per data-path address is optimized.

---
 drivers/compress/mlx5/mlx5_compress.c | 22 ++++++++++++++++++++++
 1 file changed, 22 insertions(+)
drivers/compress/mlx5/mlx5_compress.c +22 -0
index 132837e..ab24a84 100644
@@ -46,6 +46,7 @@ struct mlx5_compress_priv {
 	struct rte_compressdev_config dev_config;
 	SLIST_HEAD(xform_list, mlx5_compress_xform) xform_list;
 	rte_spinlock_t xform_sl;
+	struct mlx5_mr_share_cache mr_scache; /* Global shared MR cache. */
 };
 
 struct mlx5_compress_qp {
@@ -54,6 +55,7 @@ struct mlx5_compress_qp {
 	uint16_t pi;
 	uint16_t ci;
 	volatile uint64_t *uar_addr;
+	struct mlx5_mr_ctrl mr_ctrl;
 	int socket_id;
 	struct mlx5_devx_cq cq;
 	struct mlx5_devx_sq sq;
@@ -118,6 +120,7 @@ struct mlx5_compress_qp {
 		if (opaq != NULL)
 			rte_free(opaq);
 	}
+	mlx5_mr_btree_free(&qp->mr_ctrl.cache_bh);
 	rte_free(qp);
 	dev->data->queue_pairs[qp_id] = NULL;
 	return 0;
@@ -184,6 +187,13 @@ struct mlx5_compress_qp {
 		rte_errno = ENOMEM;
 		goto err;
 	}
+	if (mlx5_mr_btree_init(&qp->mr_ctrl.cache_bh, MLX5_MR_BTREE_CACHE_N,
+			       priv->dev_config.socket_id)) {
+		DRV_LOG(ERR, "Cannot allocate MR Btree for qp %u.",
+			(uint32_t)qp_id);
+		rte_errno = ENOMEM;
+		goto err;
+	}
 	qp->entries_n = 1 << log_ops_n;
 	qp->socket_id = socket_id;
 	qp->qp_id = qp_id;
@@ -513,6 +523,17 @@ struct mlx5_compress_qp {
 		claim_zero(mlx5_glue->close_device(priv->ctx));
 		return -1;
 	}
+	if (mlx5_mr_btree_init(&priv->mr_scache.cache,
+			     MLX5_MR_BTREE_CACHE_N * 2, rte_socket_id()) != 0) {
+		DRV_LOG(ERR, "Failed to allocate shared cache MR memory.");
+		mlx5_compress_hw_global_release(priv);
+		rte_compressdev_pmd_destroy(priv->cdev);
+		claim_zero(mlx5_glue->close_device(priv->ctx));
+		rte_errno = ENOMEM;
+		return -rte_errno;
+	}
+	priv->mr_scache.reg_mr_cb = mlx5_common_verbs_reg_mr;
+	priv->mr_scache.dereg_mr_cb = mlx5_common_verbs_dereg_mr;
 	pthread_mutex_lock(&priv_list_lock);
 	TAILQ_INSERT_TAIL(&mlx5_compress_priv_list, priv, next);
 	pthread_mutex_unlock(&priv_list_lock);
@@ -547,6 +568,7 @@ struct mlx5_compress_qp {
 		TAILQ_REMOVE(&mlx5_compress_priv_list, priv, next);
 	pthread_mutex_unlock(&priv_list_lock);
 	if (found != 0) {
+		mlx5_mr_release_cache(&priv->mr_scache);
 		mlx5_compress_hw_global_release(priv);
 		rte_compressdev_pmd_destroy(priv->cdev);
 		claim_zero(mlx5_glue->close_device(priv->ctx));
Signature · 1 line
1.8.3.1
#21 lore raw

PATCH v2 8/10 compress/mlx5: add data-path functions

In reply to Matan Azrad

Add implementation for the next compress data-path functions:
dequeue_burst
enqueue_burst

Add the next operation for starting \ stopping data-path:
dev_stop
dev_close

Each compress API enqueued operation is translated to a WQE. Once WQE is done, the HW sends CQE to the CQ, when SW see the CQE the operation will be updated and dequeued.

---
 drivers/compress/mlx5/mlx5_compress.c | 207 +++++++++++++++++++++++++++++++++-
 1 file changed, 203 insertions(+), 4 deletions(-)
drivers/compress/mlx5/mlx5_compress.c +203 -4
index ab24a84..719def2 100644
@@ -18,6 +18,7 @@
 #include <mlx5_common_os.h>
 #include <mlx5_common_devx.h>
 #include <mlx5_common_mr.h>
+#include <mlx5_common_os.h>
 #include <mlx5_prm.h>
 
 #include "mlx5_compress_utils.h"
@@ -343,10 +344,23 @@ struct mlx5_compress_qp {
 	return -ENOTSUP;
 }
 
+static void
+mlx5_compress_dev_stop(struct rte_compressdev *dev)
+{
+	RTE_SET_USED(dev);
+}
+
+static int
+mlx5_compress_dev_start(struct rte_compressdev *dev)
+{
+	RTE_SET_USED(dev);
+	return 0;
+}
+
 static struct rte_compressdev_ops mlx5_compress_ops = {
 	.dev_configure		= mlx5_compress_dev_configure,
-	.dev_start		= NULL,
-	.dev_stop		= NULL,
+	.dev_start		= mlx5_compress_dev_start,
+	.dev_stop		= mlx5_compress_dev_stop,
 	.dev_close		= mlx5_compress_dev_close,
 	.dev_infos_get		= mlx5_compress_dev_info_get,
 	.stats_get		= NULL,
@@ -359,6 +373,191 @@ struct mlx5_compress_qp {
 	.stream_free		= NULL,
 };
 
+static __rte_always_inline uint32_t
+mlx5_compress_dseg_set(struct mlx5_compress_qp *qp,
+		       volatile struct mlx5_wqe_dseg *restrict dseg,
+		       struct rte_mbuf *restrict mbuf,
+		       uint32_t offset, uint32_t len)
+
+{
+	uintptr_t addr = rte_pktmbuf_mtod_offset(mbuf, uintptr_t, offset);
+
+	dseg->bcount = rte_cpu_to_be_32(len);
+	dseg->lkey = mlx5_mr_addr2mr_bh(qp->priv->pd, 0, &qp->priv->mr_scache,
+					&qp->mr_ctrl, addr,
+					!!(mbuf->ol_flags & EXT_ATTACHED_MBUF));
+	dseg->pbuf = rte_cpu_to_be_64(addr);
+	return dseg->lkey;
+}
+
+static uint16_t
+mlx5_compress_enqueue_burst(void *queue_pair, struct rte_comp_op **ops,
+			    uint16_t nb_ops)
+{
+	struct mlx5_compress_qp *qp = queue_pair;
+	volatile struct mlx5_gga_wqe *wqes = (volatile struct mlx5_gga_wqe *)
+							      qp->sq.wqes, *wqe;
+	struct mlx5_compress_xform *xform;
+	struct rte_comp_op *op;
+	uint16_t mask = qp->entries_n - 1;
+	uint16_t remain = qp->entries_n - (qp->pi - qp->ci);
+
+	if (remain < nb_ops) {
+		ops[remain]->status = RTE_COMP_OP_STATUS_NOT_PROCESSED;
+		nb_ops = remain;
+	} else {
+		remain = nb_ops;
+	}
+	if (unlikely(remain == 0))
+		return 0;
+	do {
+		wqe = &wqes[qp->pi & mask];
+		rte_prefetch0(&wqes[(qp->pi + 1) & mask]);
+		op = *ops;
+		xform = op->private_xform;
+		if (unlikely(op->op_type != RTE_COMP_OP_STATELESS ||
+		    (xform->type == RTE_COMP_COMPRESS && op->flush_flag <
+		     RTE_COMP_FLUSH_FULL) || (mlx5_compress_dseg_set(qp,
+					&wqe->gather, op->m_src, op->src.offset,
+					       op->src.length) == UINT32_MAX) ||
+		    (mlx5_compress_dseg_set(qp, &wqe->scatter, op->m_dst,
+					    op->dst.offset,
+					    rte_pktmbuf_pkt_len(op->m_dst) -
+					    op->dst.offset) == UINT32_MAX))) {
+			op->status = RTE_COMP_OP_STATUS_INVALID_ARGS;
+			nb_ops -= remain;
+			if (unlikely(nb_ops == 0))
+				return 0;
+			break;
+		}
+		wqe->gga_ctrl1 = xform->gga_ctrl1;
+		wqe->opcode = rte_cpu_to_be_32(xform->opcode + (qp->pi << 8));
+		qp->ops[qp->pi & mask] = op;
+		++ops;
+		qp->pi++;
+	} while (--remain);
+	rte_io_wmb();
+	qp->sq.db_rec[MLX5_SND_DBR] = rte_cpu_to_be_32(qp->pi);
+	rte_wmb();
+	*qp->uar_addr = *(volatile uint64_t *)wqe; /* Assume 64 bit ARCH.*/
+	rte_wmb();
+	return nb_ops;
+}
+
+static void
+mlx5_compress_dump_err_objs(volatile uint32_t *cqe, volatile uint32_t *wqe,
+			     volatile uint32_t *opaq)
+{
+	size_t i;
+
+	DRV_LOG(ERR, "Error cqe:");
+	for (i = 0; i < sizeof(struct mlx5_err_cqe) >> 2; i += 4)
+		DRV_LOG(ERR, "%08X %08X %08X %08X", cqe[i], cqe[i + 1],
+			cqe[i + 2], cqe[i + 3]);
+	DRV_LOG(ERR, "\nError wqe:");
+	for (i = 0; i < sizeof(struct mlx5_gga_wqe) >> 2; i += 4)
+		DRV_LOG(ERR, "%08X %08X %08X %08X", wqe[i], wqe[i + 1],
+			wqe[i + 2], wqe[i + 3]);
+	DRV_LOG(ERR, "\nError opaq:");
+	for (i = 0; i < sizeof(struct mlx5_gga_compress_opaque) >> 2; i += 4)
+		DRV_LOG(ERR, "%08X %08X %08X %08X", opaq[i], opaq[i + 1],
+			opaq[i + 2], opaq[i + 3]);
+}
+
+static void
+mlx5_compress_cqe_err_handle(struct mlx5_compress_qp *qp,
+			     struct rte_comp_op *op)
+{
+	const uint32_t idx = qp->ci & (qp->entries_n - 1);
+	volatile struct mlx5_err_cqe *cqe = (volatile struct mlx5_err_cqe *)
+							      &qp->cq.cqes[idx];
+	volatile struct mlx5_gga_wqe *wqes = (volatile struct mlx5_gga_wqe *)
+								    qp->sq.wqes;
+	volatile struct mlx5_gga_compress_opaque *opaq = qp->opaque_mr.addr;
+
+	op->status = RTE_COMP_OP_STATUS_ERROR;
+	op->consumed = 0;
+	op->produced = 0;
+	op->output_chksum = 0;
+	op->debug_status = rte_be_to_cpu_32(opaq[idx].syndrom) |
+			      ((uint64_t)rte_be_to_cpu_32(cqe->syndrome) << 32);
+	mlx5_compress_dump_err_objs((volatile uint32_t *)cqe,
+				 (volatile uint32_t *)&wqes[idx],
+				 (volatile uint32_t *)&opaq[idx]);
+}
+
+static uint16_t
+mlx5_compress_dequeue_burst(void *queue_pair, struct rte_comp_op **ops,
+			    uint16_t nb_ops)
+{
+	struct mlx5_compress_qp *qp = queue_pair;
+	volatile struct mlx5_compress_xform *restrict xform;
+	volatile struct mlx5_cqe *restrict cqe;
+	volatile struct mlx5_gga_compress_opaque *opaq = qp->opaque_mr.addr;
+	struct rte_comp_op *restrict op;
+	const unsigned int cq_size = qp->entries_n;
+	const unsigned int mask = cq_size - 1;
+	uint32_t idx;
+	uint32_t next_idx = qp->ci & mask;
+	const uint16_t max = RTE_MIN((uint16_t)(qp->pi - qp->ci), nb_ops);
+	uint16_t i = 0;
+	int ret;
+
+	if (unlikely(max == 0))
+		return 0;
+	do {
+		idx = next_idx;
+		next_idx = (qp->ci + 1) & mask;
+		rte_prefetch0(&qp->cq.cqes[next_idx]);
+		rte_prefetch0(qp->ops[next_idx]);
+		op = qp->ops[idx];
+		cqe = &qp->cq.cqes[idx];
+		ret = check_cqe(cqe, cq_size, qp->ci);
+		/*
+		 * Be sure owner read is done before any other cookie field or
+		 * opaque field.
+		 */
+		rte_io_rmb();
+		if (unlikely(ret != MLX5_CQE_STATUS_SW_OWN)) {
+			if (likely(ret == MLX5_CQE_STATUS_HW_OWN))
+				break;
+			mlx5_compress_cqe_err_handle(qp, op);
+		} else {
+			xform = op->private_xform;
+			op->status = RTE_COMP_OP_STATUS_SUCCESS;
+			op->consumed = op->src.length;
+			op->produced = rte_be_to_cpu_32(cqe->byte_cnt);
+			MLX5_ASSERT(cqe->byte_cnt ==
+				    qp->opaque_buf[idx].scattered_length);
+			switch (xform->csum_type) {
+			case RTE_COMP_CHECKSUM_CRC32:
+				op->output_chksum = (uint64_t)rte_be_to_cpu_32
+						    (opaq[idx].crc32);
+				break;
+			case RTE_COMP_CHECKSUM_ADLER32:
+				op->output_chksum = (uint64_t)rte_be_to_cpu_32
+					    (opaq[idx].adler32) << 32;
+				break;
+			case RTE_COMP_CHECKSUM_CRC32_ADLER32:
+				op->output_chksum = (uint64_t)rte_be_to_cpu_32
+							     (opaq[idx].crc32) |
+						     ((uint64_t)rte_be_to_cpu_32
+						     (opaq[idx].adler32) << 32);
+				break;
+			default:
+				break;
+			}
+		}
+		ops[i++] = op;
+		qp->ci++;
+	} while (i < max);
+	if (likely(i != 0)) {
+		rte_io_wmb();
+		qp->cq.db_rec[0] = rte_cpu_to_be_32(qp->ci);
+	}
+	return i;
+}
+
 static struct ibv_device *
 mlx5_compress_get_ib_device_match(struct rte_pci_addr *addr)
 {
@@ -510,8 +709,8 @@ struct mlx5_compress_qp {
 	DRV_LOG(INFO,
 		"Compress device %s was created successfully.", ibv->name);
 	cdev->dev_ops = &mlx5_compress_ops;
-	cdev->dequeue_burst = NULL;
-	cdev->enqueue_burst = NULL;
+	cdev->dequeue_burst = mlx5_compress_dequeue_burst;
+	cdev->enqueue_burst = mlx5_compress_enqueue_burst;
 	cdev->feature_flags = RTE_COMPDEV_FF_HW_ACCELERATED;
 	priv = cdev->data->dev_private;
 	priv->ctx = ctx;
Signature · 1 line
1.8.3.1
#22 lore raw

PATCH v2 9/10 compress/mlx5: add statistics operations

In reply to Matan Azrad

Add support for the next statistics operations:
stats_get
stats_reset

These statistics are counted by the SW data-path.

---
 drivers/compress/mlx5/mlx5_compress.c | 36 +++++++++++++++++++++++++++++++++--
 1 file changed, 34 insertions(+), 2 deletions(-)
drivers/compress/mlx5/mlx5_compress.c +34 -2
index 719def2..d768453 100644
@@ -63,6 +63,7 @@ struct mlx5_compress_qp {
 	struct mlx5_pmd_mr opaque_mr;
 	struct rte_comp_op **ops;
 	struct mlx5_compress_priv *priv;
+	struct rte_compressdev_stats stats;
 };
 
 #define MLX5_COMPRESS_MAX_QPS 1024
@@ -357,14 +358,42 @@ struct mlx5_compress_qp {
 	return 0;
 }
 
+static void
+mlx5_compress_stats_get(struct rte_compressdev *dev,
+		struct rte_compressdev_stats *stats)
+{
+	int qp_id;
+
+	for (qp_id = 0; qp_id < dev->data->nb_queue_pairs; qp_id++) {
+		struct mlx5_compress_qp *qp = dev->data->queue_pairs[qp_id];
+
+		stats->enqueued_count += qp->stats.enqueued_count;
+		stats->dequeued_count += qp->stats.dequeued_count;
+		stats->enqueue_err_count += qp->stats.enqueue_err_count;
+		stats->dequeue_err_count += qp->stats.dequeue_err_count;
+	}
+}
+
+static void
+mlx5_compress_stats_reset(struct rte_compressdev *dev)
+{
+	int qp_id;
+
+	for (qp_id = 0; qp_id < dev->data->nb_queue_pairs; qp_id++) {
+		struct mlx5_compress_qp *qp = dev->data->queue_pairs[qp_id];
+
+		memset(&qp->stats, 0, sizeof(qp->stats));
+	}
+}
+
 static struct rte_compressdev_ops mlx5_compress_ops = {
 	.dev_configure		= mlx5_compress_dev_configure,
 	.dev_start		= mlx5_compress_dev_start,
 	.dev_stop		= mlx5_compress_dev_stop,
 	.dev_close		= mlx5_compress_dev_close,
 	.dev_infos_get		= mlx5_compress_dev_info_get,
-	.stats_get		= NULL,
-	.stats_reset		= NULL,
+	.stats_get		= mlx5_compress_stats_get,
+	.stats_reset		= mlx5_compress_stats_reset,
 	.queue_pair_setup	= mlx5_compress_qp_setup,
 	.queue_pair_release	= mlx5_compress_qp_release,
 	.private_xform_create	= mlx5_compress_xform_create,
@@ -436,6 +465,7 @@ struct mlx5_compress_qp {
 		++ops;
 		qp->pi++;
 	} while (--remain);
+	qp->stats.enqueued_count += nb_ops;
 	rte_io_wmb();
 	qp->sq.db_rec[MLX5_SND_DBR] = rte_cpu_to_be_32(qp->pi);
 	rte_wmb();
@@ -484,6 +514,7 @@ struct mlx5_compress_qp {
 	mlx5_compress_dump_err_objs((volatile uint32_t *)cqe,
 				 (volatile uint32_t *)&wqes[idx],
 				 (volatile uint32_t *)&opaq[idx]);
+	qp->stats.dequeue_err_count++;
 }
 
 static uint16_t
@@ -554,6 +585,7 @@ struct mlx5_compress_qp {
 	if (likely(i != 0)) {
 		rte_io_wmb();
 		qp->cq.db_rec[0] = rte_cpu_to_be_32(qp->ci);
+		qp->stats.dequeued_count += i;
 	}
 	return i;
 }
Signature · 1 line
1.8.3.1
#23 lore raw

PATCH v2 10/10 compress/mlx5: add the supported capabilities

In reply to Matan Azrad

Add all the capabilities supported by the device.

Add the driver documentations.

---
 doc/guides/compressdevs/features/mlx5.ini | 13 +++++
 doc/guides/compressdevs/index.rst         |  1 +
 doc/guides/compressdevs/mlx5.rst          | 84 +++++++++++++++++++++++++++++++
 doc/guides/rel_notes/release_21_02.rst    |  6 +++
 drivers/compress/mlx5/mlx5_compress.c     | 24 ++++++++-
 5 files changed, 126 insertions(+), 2 deletions(-)
 create mode 100644 doc/guides/compressdevs/features/mlx5.ini
 create mode 100644 doc/guides/compressdevs/mlx5.rst
doc/guides/compressdevs/features/mlx5.ini new file +13 -0
new file mode 100644
index 0000000..891ce47
@@ -0,0 +1,13 @@
+;
+; Refer to default.ini for the full list of available PMD features.
+;
+; Supported features of 'MLX5' compression driver.
+;
+[Features]
+HW Accelerated = Y
+Deflate        = Y
+Adler32        = Y
+Crc32          = Y
+Adler32&Crc32  = Y
+Fixed          = Y
+Dynamic        = Y
doc/guides/compressdevs/index.rst +1 -0
index 1f37e26..8f9f3a5 100644
@@ -14,3 +14,4 @@ Compression Device Drivers
     octeontx
     qat_comp
     zlib
+    mlx5
doc/guides/compressdevs/mlx5.rst new file +84 -0

85 lines not on this page · raw message

doc/guides/rel_notes/release_21_02.rst +6 -0
index 706cbf8..f672d7f 100644
@@ -51,6 +51,12 @@ New Features
      * Other libs
      * Apps, Examples, Tools (if significant)
 
+* **Added mlx5 compress PMD.**
+
+  Added a new compress PMD driver for Bluefield 2 adapters.
+
+  See the :doc:`../compressdevs/mlx5` for more details.
+
      This section is a comment. Do not overwrite or remove it.
      Also, make sure to start the actual text at the margin.
      =======================================================
drivers/compress/mlx5/mlx5_compress.c +22 -2
index d768453..7384351 100644
@@ -74,8 +74,28 @@ struct mlx5_compress_qp {
 
 int mlx5_compress_logtype;
 
-const struct rte_compressdev_capabilities mlx5_caps[RTE_COMP_ALGO_LIST_END];
-
+static const struct rte_compressdev_capabilities mlx5_caps[] = {
+	{
+		.algo = RTE_COMP_ALGO_NULL,
+		.comp_feature_flags = RTE_COMP_FF_ADLER32_CHECKSUM |
+				      RTE_COMP_FF_CRC32_CHECKSUM |
+				      RTE_COMP_FF_CRC32_ADLER32_CHECKSUM |
+				      RTE_COMP_FF_SHAREABLE_PRIV_XFORM,
+	},
+	{
+		.algo = RTE_COMP_ALGO_DEFLATE,
+		.comp_feature_flags = RTE_COMP_FF_ADLER32_CHECKSUM |
+				      RTE_COMP_FF_CRC32_CHECKSUM |
+				      RTE_COMP_FF_CRC32_ADLER32_CHECKSUM |
+				      RTE_COMP_FF_SHAREABLE_PRIV_XFORM |
+				      RTE_COMP_FF_HUFFMAN_FIXED |
+				      RTE_COMP_FF_HUFFMAN_DYNAMIC,
+		.window_size = {.min = 10, .max = 15, .increment = 1},
+	},
+	{
+		.algo = RTE_COMP_ALGO_LIST_END,
+	}
+};
 
 static void
 mlx5_compress_dev_info_get(struct rte_compressdev *dev,
Signature · 1 line
1.8.3.1
#24 lore raw

Re: [dpdk-dev] [PATCH v2 10/10] compress/mlx5: add the supported capabilities

Quoted 44 lines

Subject: [PATCH v2 10/10] compress/mlx5: add the supported capabilities

Add all the capabilities supported by the device.

Add the driver documentations.

---
 doc/guides/compressdevs/features/mlx5.ini | 13 +++++
 doc/guides/compressdevs/index.rst         |  1 +
 doc/guides/compressdevs/mlx5.rst          | 84
+++++++++++++++++++++++++++++++
 doc/guides/rel_notes/release_21_02.rst    |  6 +++
 drivers/compress/mlx5/mlx5_compress.c     | 24 ++++++++-
 5 files changed, 126 insertions(+), 2 deletions(-)
 create mode 100644 doc/guides/compressdevs/features/mlx5.ini
 create mode 100644 doc/guides/compressdevs/mlx5.rst
doc/guides/compressdevs/features/mlx5.ini new file +13 -0
new file mode 100644
index 0000000..891ce47
@@ -0,0 +1,13 @@
+;
+; Refer to default.ini for the full list of available PMD features.
+;
+; Supported features of 'MLX5' compression driver.
+;
+[Features]
+HW Accelerated = Y
+Deflate        = Y
+Adler32        = Y
+Crc32          = Y
+Adler32&Crc32  = Y
+Fixed          = Y
+Dynamic        = Y
doc/guides/compressdevs/index.rst +0 -0
index 1f37e26..8f9f3a5 100644

doc/guides/compressdevs/index.rst Compression Device Drivers · line 17

Quoted 4 lines
doc/guides/compressdevs/index.rst +1 -0
@@ -14,3 +14,4 @@ Compression Device Drivers
     octeontx
     qat_comp
     zlib
+    mlx5

Maintain alphabetical order.

Also fix compilation issues reported by CI
http://mails.dpdk.org/archives/test-report/2021-January/173975.html

There is no information about the installation and how to test this PMD in the documentation.

Quoted 155 lines
doc/guides/compressdevs/mlx5.rst new file +37 -0
new file mode 100644
index 0000000..4ee26b0
@@ -0,0 +1,84 @@
+.. SPDX-License-Identifier: BSD-3-Clause
+   Copyright 2021 Mellanox Technologies, Ltd
+
+.. include:: <isonum.txt>
+
+MLX5 compress driver
+====================
+
+The MLX5 compress driver library
+(**librte_compress_mlx5**) provides support for **Mellanox BlueField 2**
+families of 25/50/100/200 Gb/s adapters.
+
+Design
+------
+
+This PMD is configuring the compress, decompress amd DMA engines.
+
+For security reasons and robustness, this driver only deals with virtual
+memory addresses. The way resources allocations are handled by the kernel,
+combined with hardware specifications that allow to handle virtual memory
+addresses directly, ensure that DPDK applications cannot access random
+physical memory (or memory that does not belong to the current process).
+
+The PMD uses libibverbs and libmlx5 to access the device firmware
+or directly the hardware components.
+There are different levels of objects and bypassing abilities
+to get the best performances:
+
+- Verbs is a complete high-level generic API.
+- Direct Verbs is a device-specific API.
+- DevX allows to access firmware objects.
+
+Enabling librte_compress_mlx5 causes DPDK applications to be linked against
+libibverbs.
+
+Mellanox mlx5 pci device can be probed by number of different pci devices,
+for example net / vDPA / compress. To select the compress PMD

``class=compress``
+should be specified as device parameter. The compress device can be probed and

doc/guides/compressdevs/mlx5.rst new file +2 -0
+used with other Mellanox classes, by adding more options in the class.
+For example: ``class=net:compress`` will probe both the net PMD and the

compress

doc/guides/compressdevs/mlx5.rst new file +33 -0
+PMD.
+
+Features
+--------
+
+Compress mlx5 PMD has support for:
+
+Compression/Decompression algorithm:
+
+* DEFLATE.
+
+NULL algorithm for DMA operations.
+
+Huffman code type:
+
+* FIXED.
+* DYNAMIC.
+
+Window size support:
+
+1KB, 2KB, 4KB, 8KB, 16KB and 32KB.
+
+Sharable transformation.
+
+Checksum generation:
+
+* CRC32, Adler32 and combined checksum.
+
+Limitations
+-----------
+
+* Scatter-Gather, SHA and Stateful are not supported.
+* Non-compressed block is not supported in compress (supported in

decompress).

doc/guides/compressdevs/mlx5.rst new file +11 -0
+
+Supported NICs
+--------------
+
+* Mellanox\ |reg| BlueField 2 SmartNIC
+
+Prerequisites
+-------------
+
+- Mellanox OFED version: **5.2**
+  see :doc:`../../nics/mlx5` guide for more Mellanox OFED details.
\ No newline at end of file
doc/guides/rel_notes/release_21_02.rst +6 -0
index 706cbf8..f672d7f 100644
@@ -51,6 +51,12 @@ New Features
      * Other libs
      * Apps, Examples, Tools (if significant)
 
+* **Added mlx5 compress PMD.**
+
+  Added a new compress PMD driver for Bluefield 2 adapters.
+
+  See the :doc:`../compressdevs/mlx5` for more details.
+
      This section is a comment. Do not overwrite or remove it.
      Also, make sure to start the actual text at the margin.
      =======================================================
drivers/compress/mlx5/mlx5_compress.c +6 -2
index d768453..7384351 100644
@@ -74,8 +74,28 @@ struct mlx5_compress_qp {
 
 int mlx5_compress_logtype;
 
-const struct rte_compressdev_capabilities
 mlx5_caps[RTE_COMP_ALGO_LIST_END];
-
+static const struct rte_compressdev_capabilities mlx5_caps[] = {
+	{
+		.algo = RTE_COMP_ALGO_NULL,
+		.comp_feature_flags = RTE_COMP_FF_ADLER32_CHECKSUM |
+				      RTE_COMP_FF_CRC32_CHECKSUM |
+

RTE_COMP_FF_CRC32_ADLER32_CHECKSUM |

drivers/compress/mlx5/mlx5_compress.c +7 -0
+				      RTE_COMP_FF_SHAREABLE_PRIV_XFORM,
+	},
+	{
+		.algo = RTE_COMP_ALGO_DEFLATE,
+		.comp_feature_flags = RTE_COMP_FF_ADLER32_CHECKSUM |
+				      RTE_COMP_FF_CRC32_CHECKSUM |
+

RTE_COMP_FF_CRC32_ADLER32_CHECKSUM |

drivers/compress/mlx5/mlx5_compress.c +9 -1
+				      RTE_COMP_FF_SHAREABLE_PRIV_XFORM |
+				      RTE_COMP_FF_HUFFMAN_FIXED |
+				      RTE_COMP_FF_HUFFMAN_DYNAMIC,
+		.window_size = {.min = 10, .max = 15, .increment = 1},
+	},
+	{
+		.algo = RTE_COMP_ALGO_LIST_END,
+	}
+};
 
 static void
 mlx5_compress_dev_info_get(struct rte_compressdev *dev,
--

1.8.3.1

#25 lore raw

Re: [dpdk-dev] [PATCH v2 10/10] compress/mlx5: add the supported capabilities

Hi Akhil

Thanks for the review.

PSB

From: Akhil Goyal

Quoted 40 lines
Quoted 40 lines

Add all the capabilities supported by the device.

Add the driver documentations.

---
 doc/guides/compressdevs/features/mlx5.ini | 13 +++++
 doc/guides/compressdevs/index.rst         |  1 +
 doc/guides/compressdevs/mlx5.rst          | 84
+++++++++++++++++++++++++++++++
 doc/guides/rel_notes/release_21_02.rst    |  6 +++
 drivers/compress/mlx5/mlx5_compress.c     | 24 ++++++++-
 5 files changed, 126 insertions(+), 2 deletions(-)  create mode
100644 doc/guides/compressdevs/features/mlx5.ini
 create mode 100644 doc/guides/compressdevs/mlx5.rst
doc/guides/compressdevs/features/mlx5.ini new file +13 -0
new file mode 100644
index 0000000..891ce47
@@ -0,0 +1,13 @@
+;
+; Refer to default.ini for the full list of available PMD features.
+;
+; Supported features of 'MLX5' compression driver.
+;
+[Features]
+HW Accelerated = Y
+Deflate        = Y
+Adler32        = Y
+Crc32          = Y
+Adler32&Crc32  = Y
+Fixed          = Y
+Dynamic        = Y
doc/guides/compressdevs/index.rst +0 -1
index 1f37e26..8f9f3a5 100644
--- a/doc/guides/compressdevs/index.rst
Quoted 9 lines

doc/guides/compressdevs/index.rst Compression Device Drivers · line 17

Quoted 6 lines

+++ b/doc/guides/compressdevs/index.rst

doc/guides/compressdevs/index.rst +1 -0
@@ -14,3 +14,4 @@ Compression Device Drivers
     octeontx
     qat_comp
     zlib
+    mlx5

Maintain alphabetical order.

sure

Quoted 9 lines

Also fix compilation issues reported by CI
https://nam11.safelinks.protection.outlook.com/?url=http%3A%2F%2Fmails.dpd
k.org%2Farchives%2Ftest-report%2F2021- January%2F173975.html&amp;data=04%7C01%7Cmatan%40nvidia.com%7C58e 8dd9d0a724957b66808d8bca33bd5%7C43083d15727340c1b7db39efd9ccc17a% 7C0%7C0%7C637466756562484499%7CUnknown%7CTWFpbGZsb3d8eyJWIjoiM C4wLjAwMDAiLCJQIjoiV2luMzIiLCJBTiI6Ik1haWwiLCJXVCI6Mn0%3D%7C1000& amp;sdata=npYG53ohqpHZMpRTZX5VFzzeF3sf082HKh3JgS9Zd2k%3D&amp;rese rved=0

No real issue - it is just sent when the dependency series hadn't been integrated.

Quoted 2 lines

There is no information about the installation and how to test this PMD in the documentation.

You can see below that I added a reference to the installation guide in net/mlx5 rst - it is the same.

Quoted 159 lines
Quoted 19 lines
doc/guides/compressdevs/mlx5.rst new file +12 -0
new file mode 100644
index 0000000..4ee26b0
@@ -0,0 +1,84 @@
+.. SPDX-License-Identifier: BSD-3-Clause
+   Copyright 2021 Mellanox Technologies, Ltd
+
+.. include:: <isonum.txt>
+
+MLX5 compress driver
+====================
+
+The MLX5 compress driver library
+(**librte_compress_mlx5**) provides support for **Mellanox BlueField
+2** families of 25/50/100/200 Gb/s adapters.
+

doc/guides/compressdevs/mlx5.rst line 22

Quoted 10 lines
doc/guides/compressdevs/mlx5.rst new file +10 -0
+Design
+------
+
+This PMD is configuring the compress, decompress amd DMA engines.
+
+For security reasons and robustness, this driver only deals with
+virtual memory addresses. The way resources allocations are handled
+by the kernel, combined with hardware specifications that allow to
+handle virtual memory addresses directly, ensure that DPDK
+applications cannot access random physical memory (or memory that does

not belong to the current process).

Quoted 129 lines
doc/guides/compressdevs/mlx5.rst new file +16 -0
+
+The PMD uses libibverbs and libmlx5 to access the device firmware or
+directly the hardware components.
+There are different levels of objects and bypassing abilities to get
+the best performances:
+
+- Verbs is a complete high-level generic API.
+- Direct Verbs is a device-specific API.
+- DevX allows to access firmware objects.
+
+Enabling librte_compress_mlx5 causes DPDK applications to be linked
+against libibverbs.
+
+Mellanox mlx5 pci device can be probed by number of different pci
+devices, for example net / vDPA / compress. To select the compress
+PMD

``class=compress``

doc/guides/compressdevs/mlx5.rst new file +2 -0
+should be specified as device parameter. The compress device can be
+probed

and

doc/guides/compressdevs/mlx5.rst new file +3 -0
+used with other Mellanox classes, by adding more options in the class.
+For example: ``class=net:compress`` will probe both the net PMD and
+the

compress

doc/guides/compressdevs/mlx5.rst new file +33 -0
+PMD.
+
+Features
+--------
+
+Compress mlx5 PMD has support for:
+
+Compression/Decompression algorithm:
+
+* DEFLATE.
+
+NULL algorithm for DMA operations.
+
+Huffman code type:
+
+* FIXED.
+* DYNAMIC.
+
+Window size support:
+
+1KB, 2KB, 4KB, 8KB, 16KB and 32KB.
+
+Sharable transformation.
+
+Checksum generation:
+
+* CRC32, Adler32 and combined checksum.
+
+Limitations
+-----------
+
+* Scatter-Gather, SHA and Stateful are not supported.
+* Non-compressed block is not supported in compress (supported in

decompress).

doc/guides/compressdevs/mlx5.rst new file +11 -0
+
+Supported NICs
+--------------
+
+* Mellanox\ |reg| BlueField 2 SmartNIC
+
+Prerequisites
+-------------
+
+- Mellanox OFED version: **5.2**
+  see :doc:`../../nics/mlx5` guide for more Mellanox OFED details.
\ No newline at end of file
doc/guides/rel_notes/release_21_02.rst +6 -0
index 706cbf8..f672d7f 100644
@@ -51,6 +51,12 @@ New Features
      * Other libs
      * Apps, Examples, Tools (if significant)
 
+* **Added mlx5 compress PMD.**
+
+  Added a new compress PMD driver for Bluefield 2 adapters.
+
+  See the :doc:`../compressdevs/mlx5` for more details.
+
      This section is a comment. Do not overwrite or remove it.
      Also, make sure to start the actual text at the margin.
      =======================================================
drivers/compress/mlx5/mlx5_compress.c +6 -2
index d768453..7384351 100644
@@ -74,8 +74,28 @@ struct mlx5_compress_qp {
 
 int mlx5_compress_logtype;
 
-const struct rte_compressdev_capabilities
 mlx5_caps[RTE_COMP_ALGO_LIST_END];
-
+static const struct rte_compressdev_capabilities mlx5_caps[] = {
+     {
+             .algo = RTE_COMP_ALGO_NULL,
+             .comp_feature_flags = RTE_COMP_FF_ADLER32_CHECKSUM |
+                                   RTE_COMP_FF_CRC32_CHECKSUM |
+

RTE_COMP_FF_CRC32_ADLER32_CHECKSUM |

drivers/compress/mlx5/mlx5_compress.c +7 -0
+                                   RTE_COMP_FF_SHAREABLE_PRIV_XFORM,
+     },
+     {
+             .algo = RTE_COMP_ALGO_DEFLATE,
+             .comp_feature_flags = RTE_COMP_FF_ADLER32_CHECKSUM |
+                                   RTE_COMP_FF_CRC32_CHECKSUM |
+

RTE_COMP_FF_CRC32_ADLER32_CHECKSUM |

drivers/compress/mlx5/mlx5_compress.c +9 -1
+                                   RTE_COMP_FF_SHAREABLE_PRIV_XFORM |
+                                   RTE_COMP_FF_HUFFMAN_FIXED |
+                                   RTE_COMP_FF_HUFFMAN_DYNAMIC,
+             .window_size = {.min = 10, .max = 15, .increment = 1},
+     },
+     {
+             .algo = RTE_COMP_ALGO_LIST_END,
+     }
+};
 
 static void
 mlx5_compress_dev_info_get(struct rte_compressdev *dev,
--

1.8.3.1

#26 lore raw

v3 cover add mlx5 compress PMD

In reply to Matan Azrad

Add a new compress PMD for Mellanox devices.

The MLX5 compress driver library provides support for Mellanox BlueField 2 families of 25/50/100/200 Gb/s adapters.

Using the BlueField 2 device, the compress class operations can be run in parallel to the net, vdpa, and regex class operations.

This driver is depending on rdma-core like the other mlx5 PMDs, also it is going to use mlx5 DevX to create HW objects directly by the FW.

v3:
Address Slava and Akhil comments:
- Maintain alphabetical order in compress driver lists.
- Improve docs \ commit logs with more information.
- Fix checkptch warning.
- Support 32-bit systems.
- Validate primary procces only working.
- fix memory leak in error flow of QP creation.
- Improve coding style.
- Improve operation error report in data-path.
rebased code.

v2:
Add documentation.

Matan Azrad (11):
common/mlx5: add DevX attributes for compress
drivers: introduce mlx5 compress PMD
compress/mlx5: support basic control operations
common/mlx5: add compress primitives
compress/mlx5: support queue pair operations
compress/mlx5: add transformation operations
compress/mlx5: add memory region management
compress/mlx5: add data-path functions
compress/mlx5: add statistics operations
compress/mlx5: support 32-bit systems
compress/mlx5: add the supported capabilities

 MAINTAINERS                                 |   4 +
 doc/guides/compressdevs/features/mlx5.ini   |  13 +
 doc/guides/compressdevs/index.rst           |   1 +
 doc/guides/compressdevs/mlx5.rst            |  94 +++
 doc/guides/rel_notes/release_21_02.rst      |   6 +
 drivers/common/mlx5/mlx5_common.h           |   1 +
 drivers/common/mlx5/mlx5_common_pci.c       |   7 +
 drivers/common/mlx5/mlx5_common_pci.h       |  36 +-
 drivers/common/mlx5/mlx5_devx_cmds.c        |  10 +
 drivers/common/mlx5/mlx5_devx_cmds.h        |   7 +
 drivers/common/mlx5/mlx5_prm.h              |  59 +-
 drivers/compress/meson.build                |   2 +-
 drivers/compress/mlx5/meson.build           |  26 +
 drivers/compress/mlx5/mlx5_compress.c       | 905 ++++++++++++++++++++++++++++
 drivers/compress/mlx5/mlx5_compress_utils.h |  20 +
 drivers/compress/mlx5/version.map           |   3 +
 16 files changed, 1171 insertions(+), 23 deletions(-)
 create mode 100644 doc/guides/compressdevs/features/mlx5.ini
 create mode 100644 doc/guides/compressdevs/mlx5.rst
 create mode 100644 drivers/compress/mlx5/meson.build
 create mode 100644 drivers/compress/mlx5/mlx5_compress.c
 create mode 100644 drivers/compress/mlx5/mlx5_compress_utils.h
 create mode 100644 drivers/compress/mlx5/version.map
Signature · 1 line
1.8.3.1
#27 lore raw

PATCH v3 1/11 common/mlx5: add DevX attributes for compress

Add the DevX attributes for compress related engines:
- compress
- decompress
- dma

---
 drivers/common/mlx5/mlx5_devx_cmds.c | 10 ++++++++++
 drivers/common/mlx5/mlx5_devx_cmds.h |  7 +++++++
 drivers/common/mlx5/mlx5_prm.h       | 18 ++++++++++++++++--
 3 files changed, 33 insertions(+), 2 deletions(-)
drivers/common/mlx5/mlx5_devx_cmds.c +10 -0
index d5859c2..33acd73 100644
@@ -732,6 +732,16 @@ struct mlx5_devx_obj *
 	attr->log_max_pd = MLX5_GET(cmd_hca_cap, hcattr, log_max_pd);
 	attr->log_max_srq = MLX5_GET(cmd_hca_cap, hcattr, log_max_srq);
 	attr->log_max_srq_sz = MLX5_GET(cmd_hca_cap, hcattr, log_max_srq_sz);
+	attr->mmo_dma_en = MLX5_GET(cmd_hca_cap, hcattr, dma_mmo);
+	attr->mmo_compress_en = MLX5_GET(cmd_hca_cap, hcattr, compress);
+	attr->mmo_decompress_en = MLX5_GET(cmd_hca_cap, hcattr, decompress);
+	attr->compress_min_block_size = MLX5_GET(cmd_hca_cap, hcattr,
+						 compress_min_block_size);
+	attr->log_max_mmo_dma = MLX5_GET(cmd_hca_cap, hcattr, log_dma_mmo_size);
+	attr->log_max_mmo_compress = MLX5_GET(cmd_hca_cap, hcattr,
+					      log_compress_mmo_size);
+	attr->log_max_mmo_decompress = MLX5_GET(cmd_hca_cap, hcattr,
+						log_decompress_mmo_size);
 	if (attr->qos.sup) {
 		MLX5_SET(query_hca_cap_in, in, op_mod,
 			 MLX5_GET_HCA_CAP_OP_MOD_QOS_CAP |
drivers/common/mlx5/mlx5_devx_cmds.h +7 -0
index bf83a90..696a981 100644
@@ -130,6 +130,13 @@ struct mlx5_hca_attr {
 	uint32_t log_max_srq;
 	uint32_t log_max_srq_sz;
 	uint32_t rss_ind_tbl_cap;
+	uint32_t mmo_dma_en:1;
+	uint32_t mmo_compress_en:1;
+	uint32_t mmo_decompress_en:1;
+	uint32_t compress_min_block_size:4;
+	uint32_t log_max_mmo_dma:5;
+	uint32_t log_max_mmo_compress:5;
+	uint32_t log_max_mmo_decompress:5;
 };
 
 struct mlx5_devx_wq_attr {
drivers/common/mlx5/mlx5_prm.h +16 -2
index c9eba22..72c843f 100644
@@ -1127,7 +1127,15 @@ enum {
 struct mlx5_ifc_cmd_hca_cap_bits {
 	u8 reserved_at_0[0x30];
 	u8 vhca_id[0x10];
-	u8 reserved_at_40[0x40];
+	u8 reserved_at_40[0x20];
+	u8 reserved_at_60[0x3];
+	u8 log_regexp_scatter_gather_size[0x5];
+	u8 reserved_at_68[0x3];
+	u8 log_dma_mmo_size[5];
+	u8 reserved_at_70[0x3];
+	u8 log_compress_mmo_size[5];
+	u8 reserved_at_78[0x3];
+	u8 log_decompress_mmo_size[5];
 	u8 log_max_srq_sz[0x8];
 	u8 log_max_qp_sz[0x8];
 	u8 reserved_at_90[0x9];
@@ -1175,7 +1183,13 @@ struct mlx5_ifc_cmd_hca_cap_bits {
 	u8 log_max_ra_res_dc[0x6];
 	u8 reserved_at_140[0xa];
 	u8 log_max_ra_req_qp[0x6];
-	u8 reserved_at_150[0xa];
+	u8 rtr2rts_qp_counters_set_id[1];
+	u8 rts2rts_udp_sport[1];
+	u8 rts2rts_lag_tx_port_affinity[1];
+	u8 dma_mmo[1];
+	u8 compress_min_block_size[4];
+	u8 compress[1];
+	u8 decompress[1];
 	u8 log_max_ra_res_qp[0x6];
 	u8 end_pad[0x1];
 	u8 cc_query_allowed[0x1];
Signature · 1 line
1.8.3.1
#28 lore raw

Re: [dpdk-dev] [PATCH v3 01/11] common/mlx5: add DevX attributes for compress

Quoted 70 lines

Subject: [dpdk-dev] [PATCH v3 01/11] common/mlx5: add DevX attributes for compress

Add the DevX attributes for compress related engines:
- compress
- decompress
- dma

---
 drivers/common/mlx5/mlx5_devx_cmds.c | 10 ++++++++++
drivers/common/mlx5/mlx5_devx_cmds.h |  7 +++++++
 drivers/common/mlx5/mlx5_prm.h       | 18 ++++++++++++++++--
 3 files changed, 33 insertions(+), 2 deletions(-)
drivers/common/mlx5/mlx5_devx_cmds.c +6 -0
index d5859c2..33acd73 100644
@@ -732,6 +732,16 @@ struct mlx5_devx_obj *
 	attr->log_max_pd = MLX5_GET(cmd_hca_cap, hcattr, log_max_pd);
 	attr->log_max_srq = MLX5_GET(cmd_hca_cap, hcattr, log_max_srq);
 	attr->log_max_srq_sz = MLX5_GET(cmd_hca_cap, hcattr,
 log_max_srq_sz);
+	attr->mmo_dma_en = MLX5_GET(cmd_hca_cap, hcattr, dma_mmo);
+	attr->mmo_compress_en = MLX5_GET(cmd_hca_cap, hcattr,
 compress);
+	attr->mmo_decompress_en = MLX5_GET(cmd_hca_cap, hcattr,
 decompress);
+	attr->compress_min_block_size = MLX5_GET(cmd_hca_cap, hcattr,
+						 compress_min_block_size);
+	attr->log_max_mmo_dma = MLX5_GET(cmd_hca_cap, hcattr,

log_dma_mmo_size);

drivers/common/mlx5/mlx5_devx_cmds.c +3 -0
+	attr->log_max_mmo_compress = MLX5_GET(cmd_hca_cap, hcattr,
+					      log_compress_mmo_size);
+	attr->log_max_mmo_decompress = MLX5_GET(cmd_hca_cap,

hcattr,

drivers/common/mlx5/mlx5_devx_cmds.c +1 -0
+						log_decompress_mmo_size);
 	if (attr->qos.sup) {
 		MLX5_SET(query_hca_cap_in, in, op_mod,
 			 MLX5_GET_HCA_CAP_OP_MOD_QOS_CAP |
drivers/common/mlx5/mlx5_devx_cmds.h +7 -0
index bf83a90..696a981 100644
@@ -130,6 +130,13 @@ struct mlx5_hca_attr {
 	uint32_t log_max_srq;
 	uint32_t log_max_srq_sz;
 	uint32_t rss_ind_tbl_cap;
+	uint32_t mmo_dma_en:1;
+	uint32_t mmo_compress_en:1;
+	uint32_t mmo_decompress_en:1;
+	uint32_t compress_min_block_size:4;
+	uint32_t log_max_mmo_dma:5;
+	uint32_t log_max_mmo_compress:5;
+	uint32_t log_max_mmo_decompress:5;
 };
 
 struct mlx5_devx_wq_attr {
(quoted diff) +0 -0

b/drivers/common/mlx5/mlx5_prm.h index c9eba22..72c843f 100644

drivers/common/mlx5/mlx5_prm.h +0 -0
@@ -1127,7 +1127,15 @@ enum {
 struct mlx5_ifc_cmd_hca_cap_bits {
 	u8 reserved_at_0[0x30];
 	u8 vhca_id[0x10];

drivers/common/mlx5/mlx5_prm.h enum { · line 1138

Quoted 10 lines
drivers/common/mlx5/mlx5_prm.h +9 -1
-	u8 reserved_at_40[0x40];
+	u8 reserved_at_40[0x20];
+	u8 reserved_at_60[0x3];
+	u8 log_regexp_scatter_gather_size[0x5];
+	u8 reserved_at_68[0x3];
+	u8 log_dma_mmo_size[5];
+	u8 reserved_at_70[0x3];
+	u8 log_compress_mmo_size[5];
+	u8 reserved_at_78[0x3];
+	u8 log_decompress_mmo_size[5];

Small comment, PRM bit array size is always defined in hex format.

Quoted 5 lines
drivers/common/mlx5/mlx5_prm.h +0 -0
 	u8 log_max_srq_sz[0x8];
 	u8 log_max_qp_sz[0x8];
 	u8 reserved_at_90[0x9];
@@ -1175,7 +1183,13 @@ struct mlx5_ifc_cmd_hca_cap_bits {
 	u8 log_max_ra_res_dc[0x6];

drivers/common/mlx5/mlx5_prm.h struct mlx5_ifc_cmd_hca_cap_bits { · line 1192

Quoted 10 lines
drivers/common/mlx5/mlx5_prm.h +7 -1
 	u8 reserved_at_140[0xa];
 	u8 log_max_ra_req_qp[0x6];
-	u8 reserved_at_150[0xa];
+	u8 rtr2rts_qp_counters_set_id[1];
+	u8 rts2rts_udp_sport[1];
+	u8 rts2rts_lag_tx_port_affinity[1];
+	u8 dma_mmo[1];
+	u8 compress_min_block_size[4];
+	u8 compress[1];
+	u8 decompress[1];

Same.

Quoted 5 lines
drivers/common/mlx5/mlx5_prm.h +0 -1
 	u8 log_max_ra_res_qp[0x6];
 	u8 end_pad[0x1];
 	u8 cc_query_allowed[0x1];
--

1.8.3.1

#29 lore raw

Re: [dpdk-dev] [PATCH v3 01/11] common/mlx5: add DevX attributes for compress

From: Tal Shnaiderman

Quoted 100 lines

Sent: Thursday, January 21, 2021 6:52 PM
To: Matan Azrad <[email protected]>; [email protected]
Cc: NBU-Contact-Thomas Monjalon <[email protected]>; Ashish Gupta <[email protected]>; Fiona Trahe <[email protected]>; [email protected]
Subject: RE: [dpdk-dev] [PATCH v3 01/11] common/mlx5: add DevX attributes for compress

Quoted 70 lines

Subject: [dpdk-dev] [PATCH v3 01/11] common/mlx5: add DevX attributes for compress

Add the DevX attributes for compress related engines:
- compress
- decompress
- dma

---
 drivers/common/mlx5/mlx5_devx_cmds.c | 10 ++++++++++
drivers/common/mlx5/mlx5_devx_cmds.h |  7 +++++++
 drivers/common/mlx5/mlx5_prm.h       | 18 ++++++++++++++++--
 3 files changed, 33 insertions(+), 2 deletions(-)
drivers/common/mlx5/mlx5_devx_cmds.c +6 -0
index d5859c2..33acd73 100644
@@ -732,6 +732,16 @@ struct mlx5_devx_obj *
 	attr->log_max_pd = MLX5_GET(cmd_hca_cap, hcattr, log_max_pd);
 	attr->log_max_srq = MLX5_GET(cmd_hca_cap, hcattr, log_max_srq);
 	attr->log_max_srq_sz = MLX5_GET(cmd_hca_cap, hcattr,
 log_max_srq_sz);
+	attr->mmo_dma_en = MLX5_GET(cmd_hca_cap, hcattr, dma_mmo);
+	attr->mmo_compress_en = MLX5_GET(cmd_hca_cap, hcattr,
 compress);
+	attr->mmo_decompress_en = MLX5_GET(cmd_hca_cap, hcattr,
 decompress);
+	attr->compress_min_block_size = MLX5_GET(cmd_hca_cap, hcattr,
+						 compress_min_block_size);
+	attr->log_max_mmo_dma = MLX5_GET(cmd_hca_cap, hcattr,

log_dma_mmo_size);

drivers/common/mlx5/mlx5_devx_cmds.c +3 -0
+	attr->log_max_mmo_compress = MLX5_GET(cmd_hca_cap, hcattr,
+					      log_compress_mmo_size);
+	attr->log_max_mmo_decompress = MLX5_GET(cmd_hca_cap,

hcattr,

drivers/common/mlx5/mlx5_devx_cmds.c +1 -0
+						log_decompress_mmo_size);
 	if (attr->qos.sup) {
 		MLX5_SET(query_hca_cap_in, in, op_mod,
 			 MLX5_GET_HCA_CAP_OP_MOD_QOS_CAP | diff --git

a/drivers/common/mlx5/mlx5_devx_cmds.h
b/drivers/common/mlx5/mlx5_devx_cmds.h
index bf83a90..696a981 100644

drivers/common/mlx5/mlx5_devx_cmds.h +7 -0
@@ -130,6 +130,13 @@ struct mlx5_hca_attr {
 	uint32_t log_max_srq;
 	uint32_t log_max_srq_sz;
 	uint32_t rss_ind_tbl_cap;
+	uint32_t mmo_dma_en:1;
+	uint32_t mmo_compress_en:1;
+	uint32_t mmo_decompress_en:1;
+	uint32_t compress_min_block_size:4;
+	uint32_t log_max_mmo_dma:5;
+	uint32_t log_max_mmo_compress:5;
+	uint32_t log_max_mmo_decompress:5;
 };
 
 struct mlx5_devx_wq_attr {
(quoted diff) +0 -0

b/drivers/common/mlx5/mlx5_prm.h index c9eba22..72c843f 100644

drivers/common/mlx5/mlx5_prm.h +0 -0
@@ -1127,7 +1127,15 @@ enum {
 struct mlx5_ifc_cmd_hca_cap_bits {
 	u8 reserved_at_0[0x30];
 	u8 vhca_id[0x10];

drivers/common/mlx5/mlx5_prm.h enum { · line 1138

Quoted 10 lines
drivers/common/mlx5/mlx5_prm.h +9 -1
-	u8 reserved_at_40[0x40];
+	u8 reserved_at_40[0x20];
+	u8 reserved_at_60[0x3];
+	u8 log_regexp_scatter_gather_size[0x5];
+	u8 reserved_at_68[0x3];
+	u8 log_dma_mmo_size[5];
+	u8 reserved_at_70[0x3];
+	u8 log_compress_mmo_size[5];
+	u8 reserved_at_78[0x3];
+	u8 log_decompress_mmo_size[5];

Small comment, PRM bit array size is always defined in hex format.

Quoted 9 lines
drivers/common/mlx5/mlx5_prm.h +1 -1
 	u8 log_max_srq_sz[0x8];
 	u8 log_max_qp_sz[0x8];
 	u8 reserved_at_90[0x9];
@@ -1175,7 +1183,13 @@ struct mlx5_ifc_cmd_hca_cap_bits {
 	u8 log_max_ra_res_dc[0x6];
 	u8 reserved_at_140[0xa];
 	u8 log_max_ra_req_qp[0x6];
-	u8 reserved_at_150[0xa];
+	u8 rtr2rts_qp_counters_set_id[1];
Quoted 14 lines

drivers/common/mlx5/mlx5_prm.h struct mlx5_ifc_cmd_hca_cap_bits { · line 1192

Quoted 6 lines
drivers/common/mlx5/mlx5_prm.h +6 -0
+	u8 rts2rts_udp_sport[1];
+	u8 rts2rts_lag_tx_port_affinity[1];
+	u8 dma_mmo[1];
+	u8 compress_min_block_size[4];
+	u8 compress[1];
+	u8 decompress[1];

Same.

Quoted 5 lines
drivers/common/mlx5/mlx5_prm.h +0 -1
 	u8 log_max_ra_res_qp[0x6];
 	u8 end_pad[0x1];
 	u8 cc_query_allowed[0x1];
--

1.8.3.1

Yes, not look urgent, Akhil let me know if you can take it in integration or need to send more version for all the series because of it.

#30 lore raw

PATCH v3 2/11 drivers: introduce mlx5 compress PMD

In reply to Matan Azrad

Add a new compress PMD for Mellanox devices.

The MLX5 compress driver library provides support for Mellanox BlueField 2 families of 25/50/100/200 Gb/s adapters.

GGAs (Generic Global Accelerators) are offload engines that can be used to do memory to memory tasks on data.
These engines are part of the ARM complex of the BlueField 2 chip, and as such they do not use NIC related resources (e.g. RX/TX bandwidth). They do share the same PCI and memory bandwidth.

So, using the BlueField 2 device, the compress class operations can be run in parallel to the net, vdpa, and regex class operations.

This driver is depending on rdma-core like the other mlx5 PMDs, also it is going to use mlx5 DevX to create HW objects directly by the FW.

Add the probing functions, PCI bus connectivity, HW capabilities checks and some basic objects preparations.

---
 MAINTAINERS                                 |   4 +
 drivers/common/mlx5/mlx5_common.h           |   1 +
 drivers/common/mlx5/mlx5_common_pci.c       |   7 +
 drivers/common/mlx5/mlx5_common_pci.h       |  36 ++--
 drivers/compress/meson.build                |   2 +-
 drivers/compress/mlx5/meson.build           |  26 +++
 drivers/compress/mlx5/mlx5_compress.c       | 298 ++++++++++++++++++++++++++++
 drivers/compress/mlx5/mlx5_compress_utils.h |  20 ++
 drivers/compress/mlx5/version.map           |   3 +
 9 files changed, 378 insertions(+), 19 deletions(-)
 create mode 100644 drivers/compress/mlx5/meson.build
 create mode 100644 drivers/compress/mlx5/mlx5_compress.c
 create mode 100644 drivers/compress/mlx5/mlx5_compress_utils.h
 create mode 100644 drivers/compress/mlx5/version.map
MAINTAINERS +4 -0
index aa973a3..4b1bdf3 100644
@@ -1132,6 +1132,10 @@ F: drivers/compress/zlib/
 F: doc/guides/compressdevs/zlib.rst
 F: doc/guides/compressdevs/features/zlib.ini
 
+Mellanox mlx5
+M: Matan Azrad <[email protected]>
+F: drivers/compress/mlx5/
+
 
 RegEx Drivers
 -------------
drivers/common/mlx5/mlx5_common.h +1 -0
index e35188d..3855582 100644
@@ -217,6 +217,7 @@ enum mlx5_class {
 	MLX5_CLASS_NET = RTE_BIT64(0),
 	MLX5_CLASS_VDPA = RTE_BIT64(1),
 	MLX5_CLASS_REGEX = RTE_BIT64(2),
+	MLX5_CLASS_COMPRESS = RTE_BIT64(3),
 };
 
 #define MLX5_DBR_SIZE RTE_CACHE_LINE_SIZE
drivers/common/mlx5/mlx5_common_pci.c +7 -0
index 5208972..2b65768 100644
@@ -28,14 +28,21 @@ static TAILQ_HEAD(mlx5_pci_devices_head, mlx5_pci_device) devices_list =
 	{ .name = "vdpa", .driver_class = MLX5_CLASS_VDPA },
 	{ .name = "net", .driver_class = MLX5_CLASS_NET },
 	{ .name = "regex", .driver_class = MLX5_CLASS_REGEX },
+	{ .name = "compress", .driver_class = MLX5_CLASS_COMPRESS },
 };
 
 static const unsigned int mlx5_class_combinations[] = {
 	MLX5_CLASS_NET,
 	MLX5_CLASS_VDPA,
 	MLX5_CLASS_REGEX,
+	MLX5_CLASS_COMPRESS,
 	MLX5_CLASS_NET | MLX5_CLASS_REGEX,
 	MLX5_CLASS_VDPA | MLX5_CLASS_REGEX,
+	MLX5_CLASS_NET | MLX5_CLASS_COMPRESS,
+	MLX5_CLASS_VDPA | MLX5_CLASS_COMPRESS,
+	MLX5_CLASS_REGEX | MLX5_CLASS_COMPRESS,
+	MLX5_CLASS_NET | MLX5_CLASS_REGEX | MLX5_CLASS_COMPRESS,
+	MLX5_CLASS_VDPA | MLX5_CLASS_REGEX | MLX5_CLASS_COMPRESS,
 	/* New class combination should be added here. */
 };
 
drivers/common/mlx5/mlx5_common_pci.h +18 -18
index 41b73e1..de89bb9 100644
@@ -9,26 +9,26 @@
  * @file
  *
  * RTE Mellanox PCI Driver Interface
- * Mellanox ConnectX PCI device supports multiple class (net/vdpa/regex)
- * devices. This layer enables creating such multiple class of devices on a
- * single PCI device by allowing to bind multiple class specific device
+ * Mellanox ConnectX PCI device supports multiple class: net,vdpa,regex and
+ * compress devices. This layer enables creating such multiple class of devices
+ * on a single PCI device by allowing to bind multiple class specific device
  * driver to attach to mlx5_pci driver.
  *
- * -----------    ------------    -------------
- * |   mlx5  |    |   mlx5   |    |   mlx5    |
- * | net pmd |    | vdpa pmd |    | regex pmd |
- * -----------    ------------    -------------
- *      \              |                 /
- *       \             |                /
- *        \       --------------       /
- *         \______|   mlx5     |_____ /
- *                | pci common |
- *                --------------
- *                     |
- *                 -----------
- *                 |   mlx5  |
- *                 | pci dev |
- *                 -----------
+ * -----------    ------------    -------------    ----------------
+ * |   mlx5  |    |   mlx5   |    |   mlx5    |    |     mlx5     |
+ * | net pmd |    | vdpa pmd |    | regex pmd |    | compress pmd |
+ * -----------    ------------    -------------    ----------------
+ *      \              \                    /              /
+ *       \              \                  /              /
+ *        \              \_--------------_/              /
+ *         \_______________|   mlx5     |_______________/
+ *                         | pci common |
+ *                         --------------
+ *                               |
+ *                           -----------
+ *                           |   mlx5  |
+ *                           | pci dev |
+ *                           -----------
  *
  * - mlx5 pci driver binds to mlx5 PCI devices defined by PCI
  *   ID table of all related mlx5 PCI devices.
drivers/compress/meson.build +1 -1
index cd96def..49fa02d 100644
@@ -5,6 +5,6 @@ if is_windows
 	subdir_done()
 endif
 
-drivers = ['isal', 'octeontx', 'zlib']
+drivers = ['isal', 'mlx5', 'octeontx', 'zlib']
 
 std_deps = ['compressdev'] # compressdev pulls in all other needed deps
drivers/compress/mlx5/meson.build new file +26 -0
new file mode 100644
index 0000000..2a6dc3f
@@ -0,0 +1,26 @@
+# SPDX-License-Identifier: BSD-3-Clause
+# Copyright 2021 Mellanox Technologies, Ltd
+
+if not is_linux
+	build = false
+	reason = 'only supported on Linux'
+	subdir_done()
+endif
+
+fmt_name = 'mlx5_compress'
+deps += ['common_mlx5', 'eal', 'compressdev']
+sources = files(
+	'mlx5_compress.c',
+)
+cflags_options = [
+	'-std=c11',
+	'-Wno-strict-prototypes',
+	'-D_BSD_SOURCE',
+	'-D_DEFAULT_SOURCE',
+	'-D_XOPEN_SOURCE=600'
+]
+foreach option:cflags_options
+	if cc.has_argument(option)
+		cflags += option
+	endif
+endforeach
drivers/compress/mlx5/mlx5_compress.c new file +298 -0

298 lines not on this page · raw message

drivers/compress/mlx5/mlx5_compress_utils.h new file +20 -0
new file mode 100644
index 0000000..f93244f
@@ -0,0 +1,20 @@
+/* SPDX-License-Identifier: BSD-3-Clause
+ * Copyright 2021 Mellanox Technologies, Ltd
+ */
+
+#ifndef RTE_PMD_MLX5_COMPRESS_UTILS_H_
+#define RTE_PMD_MLX5_COMPRESS_UTILS_H_
+
+#include <mlx5_common.h>
+
+
+extern int mlx5_compress_logtype;
+
+#define MLX5_COMPRESS_LOG_PREFIX "mlx5_compress"
+/* Generic printf()-like logging macro with automatic line feed. */
+#define DRV_LOG(level, ...) \
+	PMD_DRV_LOG_(level, mlx5_compress_logtype, MLX5_COMPRESS_LOG_PREFIX, \
+		__VA_ARGS__ PMD_DRV_LOG_STRIP PMD_DRV_LOG_OPAREN, \
+		PMD_DRV_LOG_CPAREN)
+
+#endif /* RTE_PMD_MLX5_COMPRESS_UTILS_H_ */
drivers/compress/mlx5/version.map new file +3 -0
new file mode 100644
index 0000000..4a76d1d
@@ -0,0 +1,3 @@
+DPDK_21 {
+	local: *;
+};
Signature · 1 line
1.8.3.1
#31 lore raw

PATCH v3 3/11 compress/mlx5: support basic control operations

In reply to Matan Azrad

Add initial support for the next operations:
- dev_configure
- dev_close
- dev_infos_get

---
 drivers/compress/mlx5/mlx5_compress.c | 43 ++++++++++++++++++++++++++++++++---
 1 file changed, 40 insertions(+), 3 deletions(-)
drivers/compress/mlx5/mlx5_compress.c +40 -3
index 604044c..efb77d0 100644
@@ -21,6 +21,7 @@
 
 #define MLX5_COMPRESS_DRIVER_NAME mlx5_compress
 #define MLX5_COMPRESS_LOG_NAME    pmd.compress.mlx5
+#define MLX5_COMPRESS_MAX_QPS 1024
 
 struct mlx5_compress_priv {
 	TAILQ_ENTRY(mlx5_compress_priv) next;
@@ -32,6 +33,7 @@ struct mlx5_compress_priv {
 	uint8_t min_block_size;
 	/* Minimum huffman block size supported by the device. */
 	struct ibv_pd *pd;
+	struct rte_compressdev_config dev_config;
 };
 
 TAILQ_HEAD(mlx5_compress_privs, mlx5_compress_priv) mlx5_compress_priv_list =
@@ -40,12 +42,47 @@ struct mlx5_compress_priv {
 
 int mlx5_compress_logtype;
 
+const struct rte_compressdev_capabilities mlx5_caps[RTE_COMP_ALGO_LIST_END];
+
+
+static void
+mlx5_compress_dev_info_get(struct rte_compressdev *dev,
+			   struct rte_compressdev_info *info)
+{
+	RTE_SET_USED(dev);
+	if (info != NULL) {
+		info->max_nb_queue_pairs = MLX5_COMPRESS_MAX_QPS;
+		info->feature_flags = RTE_COMPDEV_FF_HW_ACCELERATED;
+		info->capabilities = mlx5_caps;
+	}
+}
+
+static int
+mlx5_compress_dev_configure(struct rte_compressdev *dev,
+			    struct rte_compressdev_config *config)
+{
+	struct mlx5_compress_priv *priv;
+
+	if (dev == NULL || config == NULL)
+		return -EINVAL;
+	priv = dev->data->dev_private;
+	priv->dev_config = *config;
+	return 0;
+}
+
+static int
+mlx5_compress_dev_close(struct rte_compressdev *dev)
+{
+	RTE_SET_USED(dev);
+	return 0;
+}
+
 static struct rte_compressdev_ops mlx5_compress_ops = {
-	.dev_configure		= NULL,
+	.dev_configure		= mlx5_compress_dev_configure,
 	.dev_start		= NULL,
 	.dev_stop		= NULL,
-	.dev_close		= NULL,
-	.dev_infos_get		= NULL,
+	.dev_close		= mlx5_compress_dev_close,
+	.dev_infos_get		= mlx5_compress_dev_info_get,
 	.stats_get		= NULL,
 	.stats_reset		= NULL,
 	.queue_pair_setup	= NULL,
Signature · 1 line
1.8.3.1
#32 lore raw

PATCH v3 4/11 common/mlx5: add compress primitives

In reply to Matan Azrad

Add the GGA compress WQE related structures and definitions.

---
 drivers/common/mlx5/mlx5_prm.h | 41 +++++++++++++++++++++++++++++++++++++++--
 1 file changed, 39 insertions(+), 2 deletions(-)
drivers/common/mlx5/mlx5_prm.h +39 -2
index 72c843f..8809b92 100644
@@ -412,10 +412,23 @@ struct mlx5_cqe_ts {
 	uint8_t op_own;
 };
 
+/* GGA */
 /* MMO metadata segment */
 
-#define	MLX5_OPCODE_MMO	0x2f
-#define	MLX5_OPC_MOD_MMO_REGEX 0x4
+#define	MLX5_OPCODE_MMO	0x2fu
+#define	MLX5_OPC_MOD_MMO_REGEX 0x4u
+#define	MLX5_OPC_MOD_MMO_COMP 0x2u
+#define	MLX5_OPC_MOD_MMO_DECOMP 0x3u
+#define	MLX5_OPC_MOD_MMO_DMA 0x1u
+
+#define WQE_GGA_COMP_WIN_SIZE_OFFSET 12u
+#define WQE_GGA_COMP_BLOCK_SIZE_OFFSET 16u
+#define WQE_GGA_COMP_DYNAMIC_SIZE_OFFSET 20u
+#define MLX5_GGA_COMP_WIN_SIZE_UNITS 1024u
+#define MLX5_GGA_COMP_WIN_SIZE_MAX (32u * MLX5_GGA_COMP_WIN_SIZE_UNITS)
+#define MLX5_GGA_COMP_LOG_BLOCK_SIZE_MAX 15u
+#define MLX5_GGA_COMP_LOG_DYNAMIC_SIZE_MAX 15u
+#define MLX5_GGA_COMP_LOG_DYNAMIC_SIZE_MIN 0u
 
 struct mlx5_wqe_metadata_seg {
 	uint32_t mmo_control_31_0; /* mmo_control_63_32 is in ctrl_seg.imm */
@@ -423,6 +436,30 @@ struct mlx5_wqe_metadata_seg {
 	uint64_t addr;
 };
 
+struct mlx5_gga_wqe {
+	uint32_t opcode;
+	uint32_t sq_ds;
+	uint32_t flags;
+	uint32_t gga_ctrl1;  /* ws 12-15, bs 16-19, dyns 20-23. */
+	uint32_t gga_ctrl2;
+	uint32_t opaque_lkey;
+	uint64_t opaque_vaddr;
+	struct mlx5_wqe_dseg gather;
+	struct mlx5_wqe_dseg scatter;
+} __rte_packed;
+
+struct mlx5_gga_compress_opaque {
+	uint32_t syndrom;
+	uint32_t reserved0;
+	uint32_t scattered_length;
+	uint32_t gathered_length;
+	uint64_t scatter_crc;
+	uint64_t gather_crc;
+	uint32_t crc32;
+	uint32_t adler32;
+	uint8_t reserved1[216];
+} __rte_packed;
+
 struct mlx5_ifc_regexp_mmo_control_bits {
 	uint8_t reserved_at_31[0x2];
 	uint8_t le[0x1];
Signature · 1 line
1.8.3.1
#33 lore raw

PATCH v3 5/11 compress/mlx5: support queue pair operations

In reply to Matan Azrad

Add support for the next operations:
- queue_pair_setup
- queue_pair_release

Create and initialize DevX SQ and CQ for each compress API queue-pair.

---
 drivers/compress/mlx5/mlx5_compress.c | 148 +++++++++++++++++++++++++++++++++-
 1 file changed, 146 insertions(+), 2 deletions(-)
drivers/compress/mlx5/mlx5_compress.c +146 -2
index efb77d0..a2507c2 100644
@@ -15,6 +15,8 @@
 #include <mlx5_common_pci.h>
 #include <mlx5_devx_cmds.h>
 #include <mlx5_common_os.h>
+#include <mlx5_common_devx.h>
+#include <mlx5_common_mr.h>
 #include <mlx5_prm.h>
 
 #include "mlx5_compress_utils.h"
@@ -36,6 +38,20 @@ struct mlx5_compress_priv {
 	struct rte_compressdev_config dev_config;
 };
 
+struct mlx5_compress_qp {
+	uint16_t qp_id;
+	uint16_t entries_n;
+	uint16_t pi;
+	uint16_t ci;
+	volatile uint64_t *uar_addr;
+	int socket_id;
+	struct mlx5_devx_cq cq;
+	struct mlx5_devx_sq sq;
+	struct mlx5_pmd_mr opaque_mr;
+	struct rte_comp_op **ops;
+	struct mlx5_compress_priv *priv;
+};
+
 TAILQ_HEAD(mlx5_compress_privs, mlx5_compress_priv) mlx5_compress_priv_list =
 				TAILQ_HEAD_INITIALIZER(mlx5_compress_priv_list);
 static pthread_mutex_t priv_list_lock = PTHREAD_MUTEX_INITIALIZER;
@@ -77,6 +93,134 @@ struct mlx5_compress_priv {
 	return 0;
 }
 
+static int
+mlx5_compress_qp_release(struct rte_compressdev *dev, uint16_t qp_id)
+{
+	struct mlx5_compress_qp *qp = dev->data->queue_pairs[qp_id];
+
+	if (qp->sq.sq != NULL)
+		mlx5_devx_sq_destroy(&qp->sq);
+	if (qp->cq.cq != NULL)
+		mlx5_devx_cq_destroy(&qp->cq);
+	if (qp->opaque_mr.obj != NULL) {
+		void *opaq = qp->opaque_mr.addr;
+
+		mlx5_common_verbs_dereg_mr(&qp->opaque_mr);
+		if (opaq != NULL)
+			rte_free(opaq);
+	}
+	rte_free(qp);
+	dev->data->queue_pairs[qp_id] = NULL;
+	return 0;
+}
+
+static void
+mlx5_compress_init_sq(struct mlx5_compress_qp *qp)
+{
+	volatile struct mlx5_gga_wqe *restrict wqe =
+				    (volatile struct mlx5_gga_wqe *)qp->sq.wqes;
+	volatile struct mlx5_gga_compress_opaque *opaq = qp->opaque_mr.addr;
+	const uint32_t sq_ds = rte_cpu_to_be_32((qp->sq.sq->id << 8) | 4u);
+	const uint32_t flags = RTE_BE32(MLX5_COMP_ALWAYS <<
+					MLX5_COMP_MODE_OFFSET);
+	const uint32_t opaq_lkey = rte_cpu_to_be_32(qp->opaque_mr.lkey);
+	int i;
+
+	/* All the next fields state should stay constant. */
+	for (i = 0; i < qp->entries_n; ++i, ++wqe) {
+		wqe->sq_ds = sq_ds;
+		wqe->flags = flags;
+		wqe->opaque_lkey = opaq_lkey;
+		wqe->opaque_vaddr = rte_cpu_to_be_64
+						((uint64_t)(uintptr_t)&opaq[i]);
+	}
+}
+
+static int
+mlx5_compress_qp_setup(struct rte_compressdev *dev, uint16_t qp_id,
+		       uint32_t max_inflight_ops, int socket_id)
+{
+	struct mlx5_compress_priv *priv = dev->data->dev_private;
+	struct mlx5_compress_qp *qp;
+	struct mlx5_devx_cq_attr cq_attr = {
+		.uar_page_id = mlx5_os_get_devx_uar_page_id(priv->uar),
+	};
+	struct mlx5_devx_create_sq_attr sq_attr = {
+		.user_index = qp_id,
+		.wq_attr = (struct mlx5_devx_wq_attr){
+			.pd = priv->pdn,
+			.uar_page = mlx5_os_get_devx_uar_page_id(priv->uar),
+		},
+	};
+	struct mlx5_devx_modify_sq_attr modify_attr = {
+		.state = MLX5_SQC_STATE_RDY,
+	};
+	uint32_t log_ops_n = rte_log2_u32(max_inflight_ops);
+	uint32_t alloc_size = sizeof(*qp);
+	void *opaq_buf;
+	int ret;
+
+	alloc_size = RTE_ALIGN(alloc_size, RTE_CACHE_LINE_SIZE);
+	alloc_size += sizeof(struct rte_comp_op *) * (1u << log_ops_n);
+	qp = rte_zmalloc_socket(__func__, alloc_size, RTE_CACHE_LINE_SIZE,
+				socket_id);
+	if (qp == NULL) {
+		DRV_LOG(ERR, "Failed to allocate qp memory.");
+		rte_errno = ENOMEM;
+		return -rte_errno;
+	}
+	dev->data->queue_pairs[qp_id] = qp;
+	opaq_buf = rte_calloc(__func__, 1u << log_ops_n,
+			      sizeof(struct mlx5_gga_compress_opaque),
+			      sizeof(struct mlx5_gga_compress_opaque));
+	if (opaq_buf == NULL) {
+		DRV_LOG(ERR, "Failed to allocate opaque memory.");
+		rte_errno = ENOMEM;
+		goto err;
+	}
+	qp->entries_n = 1 << log_ops_n;
+	qp->socket_id = socket_id;
+	qp->qp_id = qp_id;
+	qp->priv = priv;
+	qp->ops = (struct rte_comp_op **)RTE_ALIGN((uintptr_t)(qp + 1),
+						   RTE_CACHE_LINE_SIZE);
+	qp->uar_addr = mlx5_os_get_devx_uar_reg_addr(priv->uar);
+	MLX5_ASSERT(qp->uar_addr);
+	if (mlx5_common_verbs_reg_mr(priv->pd, opaq_buf, qp->entries_n *
+					sizeof(struct mlx5_gga_compress_opaque),
+							 &qp->opaque_mr) != 0) {
+		rte_free(opaq_buf);
+		DRV_LOG(ERR, "Failed to register opaque MR.");
+		rte_errno = ENOMEM;
+		goto err;
+	}
+	ret = mlx5_devx_cq_create(priv->ctx, &qp->cq, log_ops_n, &cq_attr,
+				  socket_id);
+	if (ret != 0) {
+		DRV_LOG(ERR, "Failed to create CQ.");
+		goto err;
+	}
+	sq_attr.cqn = qp->cq.cq->id;
+	ret = mlx5_devx_sq_create(priv->ctx, &qp->sq, log_ops_n, &sq_attr,
+				  socket_id);
+	if (ret != 0) {
+		DRV_LOG(ERR, "Failed to create SQ.");
+		goto err;
+	}
+	mlx5_compress_init_sq(qp);
+	ret = mlx5_devx_cmd_modify_sq(qp->sq.sq, &modify_attr);
+	if (ret != 0) {
+		DRV_LOG(ERR, "Can't change SQ state to ready.");
+		goto err;
+	}
+	DRV_LOG(INFO, "QP %u: SQN=0x%X CQN=0x%X entries num = %u\n",
+		(uint32_t)qp_id, qp->sq.sq->id, qp->cq.cq->id, qp->entries_n);
+	return 0;
+err:
+	mlx5_compress_qp_release(dev, qp_id);
+	return -1;
+}
+
 static struct rte_compressdev_ops mlx5_compress_ops = {
 	.dev_configure		= mlx5_compress_dev_configure,
 	.dev_start		= NULL,
@@ -85,8 +229,8 @@ struct mlx5_compress_priv {
 	.dev_infos_get		= mlx5_compress_dev_info_get,
 	.stats_get		= NULL,
 	.stats_reset		= NULL,
-	.queue_pair_setup	= NULL,
-	.queue_pair_release	= NULL,
+	.queue_pair_setup	= mlx5_compress_qp_setup,
+	.queue_pair_release	= mlx5_compress_qp_release,
 	.private_xform_create	= NULL,
 	.private_xform_free	= NULL,
 	.stream_create		= NULL,
Signature · 1 line
1.8.3.1
#34 lore raw

PATCH v3 6/11 compress/mlx5: add transformation operations

In reply to Matan Azrad

Add support for the next operations:
- private_xform_create
- private_xform_free

The driver transformation structure includes preparations for the next GGA WQE fields used by the enqueue function: opcode. compress specific fields (window size, block size and dynamic size) checksum type and compress type.

---
 drivers/compress/mlx5/mlx5_compress.c | 121 +++++++++++++++++++++++++++++++++-
 1 file changed, 119 insertions(+), 2 deletions(-)
drivers/compress/mlx5/mlx5_compress.c +119 -2
index a2507c2..cc9be72 100644
@@ -6,6 +6,7 @@
 #include <rte_log.h>
 #include <rte_errno.h>
 #include <rte_pci.h>
+#include <rte_spinlock.h>
 #include <rte_comp.h>
 #include <rte_compressdev.h>
 #include <rte_compressdev_pmd.h>
@@ -24,6 +25,15 @@
 #define MLX5_COMPRESS_DRIVER_NAME mlx5_compress
 #define MLX5_COMPRESS_LOG_NAME    pmd.compress.mlx5
 #define MLX5_COMPRESS_MAX_QPS 1024
+#define MLX5_COMP_MAX_WIN_SIZE_CONF 6u
+
+struct mlx5_compress_xform {
+	LIST_ENTRY(mlx5_compress_xform) next;
+	enum rte_comp_xform_type type;
+	enum rte_comp_checksum_type csum_type;
+	uint32_t opcode;
+	uint32_t gga_ctrl1; /* BE. */
+};
 
 struct mlx5_compress_priv {
 	TAILQ_ENTRY(mlx5_compress_priv) next;
@@ -36,6 +46,8 @@ struct mlx5_compress_priv {
 	/* Minimum huffman block size supported by the device. */
 	struct ibv_pd *pd;
 	struct rte_compressdev_config dev_config;
+	LIST_HEAD(xform_list, mlx5_compress_xform) xform_list;
+	rte_spinlock_t xform_sl;
 };
 
 struct mlx5_compress_qp {
@@ -221,6 +233,111 @@ struct mlx5_compress_qp {
 	return -1;
 }
 
+static int
+mlx5_compress_xform_free(struct rte_compressdev *dev, void *xform)
+{
+	struct mlx5_compress_priv *priv = dev->data->dev_private;
+
+	rte_spinlock_lock(&priv->xform_sl);
+	LIST_REMOVE((struct mlx5_compress_xform *)xform, next);
+	rte_spinlock_unlock(&priv->xform_sl);
+	rte_free(xform);
+	return 0;
+}
+
+static int
+mlx5_compress_xform_create(struct rte_compressdev *dev,
+			   const struct rte_comp_xform *xform,
+			   void **private_xform)
+{
+	struct mlx5_compress_priv *priv = dev->data->dev_private;
+	struct mlx5_compress_xform *xfrm;
+	uint32_t size;
+
+	if (xform->type == RTE_COMP_COMPRESS && xform->compress.level ==
+							  RTE_COMP_LEVEL_NONE) {
+		DRV_LOG(ERR, "Non-compressed block is not supported.");
+		return -ENOTSUP;
+	}
+	if ((xform->type == RTE_COMP_COMPRESS && xform->compress.hash_algo !=
+	     RTE_COMP_HASH_ALGO_NONE) || (xform->type == RTE_COMP_DECOMPRESS &&
+		      xform->decompress.hash_algo != RTE_COMP_HASH_ALGO_NONE)) {
+		DRV_LOG(ERR, "SHA is not supported.");
+		return -ENOTSUP;
+	}
+	xfrm = rte_zmalloc_socket(__func__, sizeof(*xfrm), 0,
+						    priv->dev_config.socket_id);
+	if (xfrm == NULL)
+		return -ENOMEM;
+	xfrm->opcode = MLX5_OPCODE_MMO;
+	xfrm->type = xform->type;
+	switch (xform->type) {
+	case RTE_COMP_COMPRESS:
+		switch (xform->compress.algo) {
+		case RTE_COMP_ALGO_NULL:
+			xfrm->opcode += MLX5_OPC_MOD_MMO_DMA <<
+							WQE_CSEG_OPC_MOD_OFFSET;
+			break;
+		case RTE_COMP_ALGO_DEFLATE:
+			size = 1 << xform->compress.window_size;
+			size /= MLX5_GGA_COMP_WIN_SIZE_UNITS;
+			xfrm->gga_ctrl1 += RTE_MIN(rte_log2_u32(size),
+					 MLX5_COMP_MAX_WIN_SIZE_CONF) <<
+					   WQE_GGA_COMP_WIN_SIZE_OFFSET;
+			if (xform->compress.level == RTE_COMP_LEVEL_PMD_DEFAULT)
+				size = MLX5_GGA_COMP_LOG_BLOCK_SIZE_MAX;
+			else
+				size = priv->min_block_size - 1 +
+							  xform->compress.level;
+			xfrm->gga_ctrl1 += RTE_MIN(size,
+					    MLX5_GGA_COMP_LOG_BLOCK_SIZE_MAX) <<
+						 WQE_GGA_COMP_BLOCK_SIZE_OFFSET;
+			xfrm->opcode += MLX5_OPC_MOD_MMO_COMP <<
+							WQE_CSEG_OPC_MOD_OFFSET;
+			size = xform->compress.deflate.huffman ==
+						      RTE_COMP_HUFFMAN_DYNAMIC ?
+					    MLX5_GGA_COMP_LOG_DYNAMIC_SIZE_MAX :
+					     MLX5_GGA_COMP_LOG_DYNAMIC_SIZE_MIN;
+			xfrm->gga_ctrl1 += size <<
+					       WQE_GGA_COMP_DYNAMIC_SIZE_OFFSET;
+			break;
+		default:
+			goto err;
+		}
+		xfrm->csum_type = xform->compress.chksum;
+		break;
+	case RTE_COMP_DECOMPRESS:
+		switch (xform->decompress.algo) {
+		case RTE_COMP_ALGO_NULL:
+			xfrm->opcode += MLX5_OPC_MOD_MMO_DMA <<
+							WQE_CSEG_OPC_MOD_OFFSET;
+			break;
+		case RTE_COMP_ALGO_DEFLATE:
+			xfrm->opcode += MLX5_OPC_MOD_MMO_DECOMP <<
+							WQE_CSEG_OPC_MOD_OFFSET;
+			break;
+		default:
+			goto err;
+		}
+		xfrm->csum_type = xform->decompress.chksum;
+		break;
+	default:
+		DRV_LOG(ERR, "Algorithm %u is not supported.", xform->type);
+		goto err;
+	}
+	DRV_LOG(DEBUG, "New xform: gga ctrl1 = 0x%08X opcode = 0x%08X csum "
+		"type = %d.", xfrm->gga_ctrl1, xfrm->opcode, xfrm->csum_type);
+	xfrm->gga_ctrl1 = rte_cpu_to_be_32(xfrm->gga_ctrl1);
+	rte_spinlock_lock(&priv->xform_sl);
+	LIST_INSERT_HEAD(&priv->xform_list, xfrm, next);
+	rte_spinlock_unlock(&priv->xform_sl);
+	*private_xform = xfrm;
+	return 0;
+err:
+	rte_free(xfrm);
+	return -ENOTSUP;
+}
+
 static struct rte_compressdev_ops mlx5_compress_ops = {
 	.dev_configure		= mlx5_compress_dev_configure,
 	.dev_start		= NULL,
@@ -231,8 +348,8 @@ struct mlx5_compress_qp {
 	.stats_reset		= NULL,
 	.queue_pair_setup	= mlx5_compress_qp_setup,
 	.queue_pair_release	= mlx5_compress_qp_release,
-	.private_xform_create	= NULL,
-	.private_xform_free	= NULL,
+	.private_xform_create	= mlx5_compress_xform_create,
+	.private_xform_free	= mlx5_compress_xform_free,
 	.stream_create		= NULL,
 	.stream_free		= NULL,
 };
Signature · 1 line
1.8.3.1
#35 lore raw

PATCH v3 7/11 compress/mlx5: add memory region management

In reply to Matan Azrad

Mellanox user space drivers don't deal with physical addresses, that's why any mbuf virtual address moved directly to the HW descriptor(WQE).

The mapping between the virtual address to the physical address is saved in MR configured by the kernel to the HW.

Each MR has a key that should also be moved to the WQE by the SW.

When the SW see address which is not mapped, it extends the address range and creates a MR using a system call.

Add memory region cache management:
- 2 level cache per queue-pair - no locks.
- 1 shared cache between all the queues using a lock.

Using this way, the MR key search per data-path address is optimized.

---
 drivers/compress/mlx5/mlx5_compress.c | 22 ++++++++++++++++++++++
 1 file changed, 22 insertions(+)
drivers/compress/mlx5/mlx5_compress.c +22 -0
index cc9be72..467b80a 100644
@@ -48,6 +48,7 @@ struct mlx5_compress_priv {
 	struct rte_compressdev_config dev_config;
 	LIST_HEAD(xform_list, mlx5_compress_xform) xform_list;
 	rte_spinlock_t xform_sl;
+	struct mlx5_mr_share_cache mr_scache; /* Global shared MR cache. */
 };
 
 struct mlx5_compress_qp {
@@ -56,6 +57,7 @@ struct mlx5_compress_qp {
 	uint16_t pi;
 	uint16_t ci;
 	volatile uint64_t *uar_addr;
+	struct mlx5_mr_ctrl mr_ctrl;
 	int socket_id;
 	struct mlx5_devx_cq cq;
 	struct mlx5_devx_sq sq;
@@ -121,6 +123,7 @@ struct mlx5_compress_qp {
 		if (opaq != NULL)
 			rte_free(opaq);
 	}
+	mlx5_mr_btree_free(&qp->mr_ctrl.cache_bh);
 	rte_free(qp);
 	dev->data->queue_pairs[qp_id] = NULL;
 	return 0;
@@ -190,6 +193,13 @@ struct mlx5_compress_qp {
 		rte_errno = ENOMEM;
 		goto err;
 	}
+	if (mlx5_mr_btree_init(&qp->mr_ctrl.cache_bh, MLX5_MR_BTREE_CACHE_N,
+			       priv->dev_config.socket_id)) {
+		DRV_LOG(ERR, "Cannot allocate MR Btree for qp %u.",
+			(uint32_t)qp_id);
+		rte_errno = ENOMEM;
+		goto err;
+	}
 	qp->entries_n = 1 << log_ops_n;
 	qp->socket_id = socket_id;
 	qp->qp_id = qp_id;
@@ -523,6 +533,17 @@ struct mlx5_compress_qp {
 		claim_zero(mlx5_glue->close_device(priv->ctx));
 		return -1;
 	}
+	if (mlx5_mr_btree_init(&priv->mr_scache.cache,
+			     MLX5_MR_BTREE_CACHE_N * 2, rte_socket_id()) != 0) {
+		DRV_LOG(ERR, "Failed to allocate shared cache MR memory.");
+		mlx5_compress_hw_global_release(priv);
+		rte_compressdev_pmd_destroy(priv->cdev);
+		claim_zero(mlx5_glue->close_device(priv->ctx));
+		rte_errno = ENOMEM;
+		return -rte_errno;
+	}
+	priv->mr_scache.reg_mr_cb = mlx5_common_verbs_reg_mr;
+	priv->mr_scache.dereg_mr_cb = mlx5_common_verbs_dereg_mr;
 	pthread_mutex_lock(&priv_list_lock);
 	TAILQ_INSERT_TAIL(&mlx5_compress_priv_list, priv, next);
 	pthread_mutex_unlock(&priv_list_lock);
@@ -553,6 +574,7 @@ struct mlx5_compress_qp {
 		TAILQ_REMOVE(&mlx5_compress_priv_list, priv, next);
 	pthread_mutex_unlock(&priv_list_lock);
 	if (priv) {
+		mlx5_mr_release_cache(&priv->mr_scache);
 		mlx5_compress_hw_global_release(priv);
 		rte_compressdev_pmd_destroy(priv->cdev);
 		claim_zero(mlx5_glue->close_device(priv->ctx));
Signature · 1 line
1.8.3.1
#36 lore raw

PATCH v3 8/11 compress/mlx5: add data-path functions

In reply to Matan Azrad

Add implementation for the next compress data-path functions:
- dequeue_burst
- enqueue_burst

Add the next operation for starting \ stopping data-path:
- dev_stop
- dev_close

Each compress API enqueued operation is translated to a WQE. Once WQE is done, the HW sends CQE to the CQ, when SW see the CQE the operation will be updated and dequeued.

---
 drivers/compress/mlx5/mlx5_compress.c | 219 +++++++++++++++++++++++++++++++++-
 1 file changed, 215 insertions(+), 4 deletions(-)
drivers/compress/mlx5/mlx5_compress.c +215 -4
index 467b80a..17aa206 100644
@@ -18,6 +18,7 @@
 #include <mlx5_common_os.h>
 #include <mlx5_common_devx.h>
 #include <mlx5_common_mr.h>
+#include <mlx5_common_os.h>
 #include <mlx5_prm.h>
 
 #include "mlx5_compress_utils.h"
@@ -348,10 +349,23 @@ struct mlx5_compress_qp {
 	return -ENOTSUP;
 }
 
+static void
+mlx5_compress_dev_stop(struct rte_compressdev *dev)
+{
+	RTE_SET_USED(dev);
+}
+
+static int
+mlx5_compress_dev_start(struct rte_compressdev *dev)
+{
+	RTE_SET_USED(dev);
+	return 0;
+}
+
 static struct rte_compressdev_ops mlx5_compress_ops = {
 	.dev_configure		= mlx5_compress_dev_configure,
-	.dev_start		= NULL,
-	.dev_stop		= NULL,
+	.dev_start		= mlx5_compress_dev_start,
+	.dev_stop		= mlx5_compress_dev_stop,
 	.dev_close		= mlx5_compress_dev_close,
 	.dev_infos_get		= mlx5_compress_dev_info_get,
 	.stats_get		= NULL,
@@ -364,6 +378,203 @@ struct mlx5_compress_qp {
 	.stream_free		= NULL,
 };
 
+static __rte_always_inline uint32_t
+mlx5_compress_dseg_set(struct mlx5_compress_qp *qp,
+		       volatile struct mlx5_wqe_dseg *restrict dseg,
+		       struct rte_mbuf *restrict mbuf,
+		       uint32_t offset, uint32_t len)
+{
+	uintptr_t addr = rte_pktmbuf_mtod_offset(mbuf, uintptr_t, offset);
+
+	dseg->bcount = rte_cpu_to_be_32(len);
+	dseg->lkey = mlx5_mr_addr2mr_bh(qp->priv->pd, 0, &qp->priv->mr_scache,
+					&qp->mr_ctrl, addr,
+					!!(mbuf->ol_flags & EXT_ATTACHED_MBUF));
+	dseg->pbuf = rte_cpu_to_be_64(addr);
+	return dseg->lkey;
+}
+
+static uint16_t
+mlx5_compress_enqueue_burst(void *queue_pair, struct rte_comp_op **ops,
+			    uint16_t nb_ops)
+{
+	struct mlx5_compress_qp *qp = queue_pair;
+	volatile struct mlx5_gga_wqe *wqes = (volatile struct mlx5_gga_wqe *)
+							      qp->sq.wqes, *wqe;
+	struct mlx5_compress_xform *xform;
+	struct rte_comp_op *op;
+	uint16_t mask = qp->entries_n - 1;
+	uint16_t remain = qp->entries_n - (qp->pi - qp->ci);
+	uint16_t idx;
+	bool invalid;
+
+	if (remain < nb_ops)
+		nb_ops = remain;
+	else
+		remain = nb_ops;
+	if (unlikely(remain == 0))
+		return 0;
+	do {
+		idx = qp->pi & mask;
+		wqe = &wqes[idx];
+		rte_prefetch0(&wqes[(qp->pi + 1) & mask]);
+		op = *ops++;
+		xform = op->private_xform;
+		/*
+		 * Check operation arguments and error cases:
+		 *   - Operation type must be state-less.
+		 *   - Compress operation flush flag must be FULL or FINAL.
+		 *   - Source and destination buffers must be mapped internally.
+		 */
+		invalid = op->op_type != RTE_COMP_OP_STATELESS ||
+					    (xform->type == RTE_COMP_COMPRESS &&
+					  op->flush_flag < RTE_COMP_FLUSH_FULL);
+		if (unlikely(invalid ||
+			     (mlx5_compress_dseg_set(qp, &wqe->gather,
+						     op->m_src,
+						     op->src.offset,
+						     op->src.length) ==
+								  UINT32_MAX) ||
+			     (mlx5_compress_dseg_set(qp, &wqe->scatter,
+						op->m_dst,
+						op->dst.offset,
+						rte_pktmbuf_pkt_len(op->m_dst) -
+							      op->dst.offset) ==
+								 UINT32_MAX))) {
+			op->status = invalid ? RTE_COMP_OP_STATUS_INVALID_ARGS :
+						       RTE_COMP_OP_STATUS_ERROR;
+			nb_ops -= remain;
+			if (unlikely(nb_ops == 0))
+				return 0;
+			break;
+		}
+		wqe->gga_ctrl1 = xform->gga_ctrl1;
+		wqe->opcode = rte_cpu_to_be_32(xform->opcode + (qp->pi << 8));
+		qp->ops[idx] = op;
+		qp->pi++;
+	} while (--remain);
+	rte_io_wmb();
+	qp->sq.db_rec[MLX5_SND_DBR] = rte_cpu_to_be_32(qp->pi);
+	rte_wmb();
+	*qp->uar_addr = *(volatile uint64_t *)wqe; /* Assume 64 bit ARCH.*/
+	rte_wmb();
+	return nb_ops;
+}
+
+static void
+mlx5_compress_dump_err_objs(volatile uint32_t *cqe, volatile uint32_t *wqe,
+			     volatile uint32_t *opaq)
+{
+	size_t i;
+
+	DRV_LOG(ERR, "Error cqe:");
+	for (i = 0; i < sizeof(struct mlx5_err_cqe) >> 2; i += 4)
+		DRV_LOG(ERR, "%08X %08X %08X %08X", cqe[i], cqe[i + 1],
+			cqe[i + 2], cqe[i + 3]);
+	DRV_LOG(ERR, "\nError wqe:");
+	for (i = 0; i < sizeof(struct mlx5_gga_wqe) >> 2; i += 4)
+		DRV_LOG(ERR, "%08X %08X %08X %08X", wqe[i], wqe[i + 1],
+			wqe[i + 2], wqe[i + 3]);
+	DRV_LOG(ERR, "\nError opaq:");
+	for (i = 0; i < sizeof(struct mlx5_gga_compress_opaque) >> 2; i += 4)
+		DRV_LOG(ERR, "%08X %08X %08X %08X", opaq[i], opaq[i + 1],
+			opaq[i + 2], opaq[i + 3]);
+}
+
+static void
+mlx5_compress_cqe_err_handle(struct mlx5_compress_qp *qp,
+			     struct rte_comp_op *op)
+{
+	const uint32_t idx = qp->ci & (qp->entries_n - 1);
+	volatile struct mlx5_err_cqe *cqe = (volatile struct mlx5_err_cqe *)
+							      &qp->cq.cqes[idx];
+	volatile struct mlx5_gga_wqe *wqes = (volatile struct mlx5_gga_wqe *)
+								    qp->sq.wqes;
+	volatile struct mlx5_gga_compress_opaque *opaq = qp->opaque_mr.addr;
+
+	op->status = RTE_COMP_OP_STATUS_ERROR;
+	op->consumed = 0;
+	op->produced = 0;
+	op->output_chksum = 0;
+	op->debug_status = rte_be_to_cpu_32(opaq[idx].syndrom) |
+			      ((uint64_t)rte_be_to_cpu_32(cqe->syndrome) << 32);
+	mlx5_compress_dump_err_objs((volatile uint32_t *)cqe,
+				 (volatile uint32_t *)&wqes[idx],
+				 (volatile uint32_t *)&opaq[idx]);
+}
+
+static uint16_t
+mlx5_compress_dequeue_burst(void *queue_pair, struct rte_comp_op **ops,
+			    uint16_t nb_ops)
+{
+	struct mlx5_compress_qp *qp = queue_pair;
+	volatile struct mlx5_compress_xform *restrict xform;
+	volatile struct mlx5_cqe *restrict cqe;
+	volatile struct mlx5_gga_compress_opaque *opaq = qp->opaque_mr.addr;
+	struct rte_comp_op *restrict op;
+	const unsigned int cq_size = qp->entries_n;
+	const unsigned int mask = cq_size - 1;
+	uint32_t idx;
+	uint32_t next_idx = qp->ci & mask;
+	const uint16_t max = RTE_MIN((uint16_t)(qp->pi - qp->ci), nb_ops);
+	uint16_t i = 0;
+	int ret;
+
+	if (unlikely(max == 0))
+		return 0;
+	do {
+		idx = next_idx;
+		next_idx = (qp->ci + 1) & mask;
+		rte_prefetch0(&qp->cq.cqes[next_idx]);
+		rte_prefetch0(qp->ops[next_idx]);
+		op = qp->ops[idx];
+		cqe = &qp->cq.cqes[idx];
+		ret = check_cqe(cqe, cq_size, qp->ci);
+		/*
+		 * Be sure owner read is done before any other cookie field or
+		 * opaque field.
+		 */
+		rte_io_rmb();
+		if (unlikely(ret != MLX5_CQE_STATUS_SW_OWN)) {
+			if (likely(ret == MLX5_CQE_STATUS_HW_OWN))
+				break;
+			mlx5_compress_cqe_err_handle(qp, op);
+		} else {
+			xform = op->private_xform;
+			op->status = RTE_COMP_OP_STATUS_SUCCESS;
+			op->consumed = op->src.length;
+			op->produced = rte_be_to_cpu_32(cqe->byte_cnt);
+			MLX5_ASSERT(cqe->byte_cnt ==
+				    qp->opaque_buf[idx].scattered_length);
+			switch (xform->csum_type) {
+			case RTE_COMP_CHECKSUM_CRC32:
+				op->output_chksum = (uint64_t)rte_be_to_cpu_32
+						    (opaq[idx].crc32);
+				break;
+			case RTE_COMP_CHECKSUM_ADLER32:
+				op->output_chksum = (uint64_t)rte_be_to_cpu_32
+					    (opaq[idx].adler32) << 32;
+				break;
+			case RTE_COMP_CHECKSUM_CRC32_ADLER32:
+				op->output_chksum = (uint64_t)rte_be_to_cpu_32
+							     (opaq[idx].crc32) |
+						     ((uint64_t)rte_be_to_cpu_32
+						     (opaq[idx].adler32) << 32);
+				break;
+			default:
+				break;
+			}
+		}
+		ops[i++] = op;
+		qp->ci++;
+	} while (i < max);
+	if (likely(i != 0)) {
+		rte_io_wmb();
+		qp->cq.db_rec[0] = rte_cpu_to_be_32(qp->ci);
+	}
+	return i;
+}
+
 static struct ibv_device *
 mlx5_compress_get_ib_device_match(struct rte_pci_addr *addr)
 {
@@ -520,8 +731,8 @@ struct mlx5_compress_qp {
 	DRV_LOG(INFO,
 		"Compress device %s was created successfully.", ibv->name);
 	cdev->dev_ops = &mlx5_compress_ops;
-	cdev->dequeue_burst = NULL;
-	cdev->enqueue_burst = NULL;
+	cdev->dequeue_burst = mlx5_compress_dequeue_burst;
+	cdev->enqueue_burst = mlx5_compress_enqueue_burst;
 	cdev->feature_flags = RTE_COMPDEV_FF_HW_ACCELERATED;
 	priv = cdev->data->dev_private;
 	priv->ctx = ctx;
Signature · 1 line
1.8.3.1
#37 lore raw

PATCH v3 9/11 compress/mlx5: add statistics operations

In reply to Matan Azrad

Add support for the next statistics operations:
- stats_get
- stats_reset

These statistics are counted by the SW data-path.

---
 drivers/compress/mlx5/mlx5_compress.c | 36 +++++++++++++++++++++++++++++++++--
 1 file changed, 34 insertions(+), 2 deletions(-)
drivers/compress/mlx5/mlx5_compress.c +34 -2
index 17aa206..7a43d9e 100644
@@ -65,6 +65,7 @@ struct mlx5_compress_qp {
 	struct mlx5_pmd_mr opaque_mr;
 	struct rte_comp_op **ops;
 	struct mlx5_compress_priv *priv;
+	struct rte_compressdev_stats stats;
 };
 
 TAILQ_HEAD(mlx5_compress_privs, mlx5_compress_priv) mlx5_compress_priv_list =
@@ -362,14 +363,42 @@ struct mlx5_compress_qp {
 	return 0;
 }
 
+static void
+mlx5_compress_stats_get(struct rte_compressdev *dev,
+		struct rte_compressdev_stats *stats)
+{
+	int qp_id;
+
+	for (qp_id = 0; qp_id < dev->data->nb_queue_pairs; qp_id++) {
+		struct mlx5_compress_qp *qp = dev->data->queue_pairs[qp_id];
+
+		stats->enqueued_count += qp->stats.enqueued_count;
+		stats->dequeued_count += qp->stats.dequeued_count;
+		stats->enqueue_err_count += qp->stats.enqueue_err_count;
+		stats->dequeue_err_count += qp->stats.dequeue_err_count;
+	}
+}
+
+static void
+mlx5_compress_stats_reset(struct rte_compressdev *dev)
+{
+	int qp_id;
+
+	for (qp_id = 0; qp_id < dev->data->nb_queue_pairs; qp_id++) {
+		struct mlx5_compress_qp *qp = dev->data->queue_pairs[qp_id];
+
+		memset(&qp->stats, 0, sizeof(qp->stats));
+	}
+}
+
 static struct rte_compressdev_ops mlx5_compress_ops = {
 	.dev_configure		= mlx5_compress_dev_configure,
 	.dev_start		= mlx5_compress_dev_start,
 	.dev_stop		= mlx5_compress_dev_stop,
 	.dev_close		= mlx5_compress_dev_close,
 	.dev_infos_get		= mlx5_compress_dev_info_get,
-	.stats_get		= NULL,
-	.stats_reset		= NULL,
+	.stats_get		= mlx5_compress_stats_get,
+	.stats_reset		= mlx5_compress_stats_reset,
 	.queue_pair_setup	= mlx5_compress_qp_setup,
 	.queue_pair_release	= mlx5_compress_qp_release,
 	.private_xform_create	= mlx5_compress_xform_create,
@@ -453,6 +482,7 @@ struct mlx5_compress_qp {
 		qp->ops[idx] = op;
 		qp->pi++;
 	} while (--remain);
+	qp->stats.enqueued_count += nb_ops;
 	rte_io_wmb();
 	qp->sq.db_rec[MLX5_SND_DBR] = rte_cpu_to_be_32(qp->pi);
 	rte_wmb();
@@ -501,6 +531,7 @@ struct mlx5_compress_qp {
 	mlx5_compress_dump_err_objs((volatile uint32_t *)cqe,
 				 (volatile uint32_t *)&wqes[idx],
 				 (volatile uint32_t *)&opaq[idx]);
+	qp->stats.dequeue_err_count++;
 }
 
 static uint16_t
@@ -571,6 +602,7 @@ struct mlx5_compress_qp {
 	if (likely(i != 0)) {
 		rte_io_wmb();
 		qp->cq.db_rec[0] = rte_cpu_to_be_32(qp->ci);
+		qp->stats.dequeued_count += i;
 	}
 	return i;
 }
Signature · 1 line
1.8.3.1
#38 lore raw

PATCH v3 10/11 compress/mlx5: support 32-bit systems

In reply to Matan Azrad

In order to support 32-bit systems, the 8B doorbell write should be done by 2 4B stores.

The order between the store is important, that's why memory barrier should be used between them.

The doorbell address is shared between all the queues, that's why a lock should wrap the 2 stores.

---
 drivers/compress/mlx5/mlx5_compress.c | 32 ++++++++++++++++++++++++++++----
 1 file changed, 28 insertions(+), 4 deletions(-)
drivers/compress/mlx5/mlx5_compress.c +28 -4
index 7a43d9e..f7ef913 100644
@@ -50,6 +50,10 @@ struct mlx5_compress_priv {
 	LIST_HEAD(xform_list, mlx5_compress_xform) xform_list;
 	rte_spinlock_t xform_sl;
 	struct mlx5_mr_share_cache mr_scache; /* Global shared MR cache. */
+	volatile uint64_t *uar_addr;
+#ifndef RTE_ARCH_64
+	rte_spinlock_t uar32_sl;
+#endif /* RTE_ARCH_64 */
 };
 
 struct mlx5_compress_qp {
@@ -57,7 +61,6 @@ struct mlx5_compress_qp {
 	uint16_t entries_n;
 	uint16_t pi;
 	uint16_t ci;
-	volatile uint64_t *uar_addr;
 	struct mlx5_mr_ctrl mr_ctrl;
 	int socket_id;
 	struct mlx5_devx_cq cq;
@@ -208,8 +211,6 @@ struct mlx5_compress_qp {
 	qp->priv = priv;
 	qp->ops = (struct rte_comp_op **)RTE_ALIGN((uintptr_t)(qp + 1),
 						   RTE_CACHE_LINE_SIZE);
-	qp->uar_addr = mlx5_os_get_devx_uar_reg_addr(priv->uar);
-	MLX5_ASSERT(qp->uar_addr);
 	if (mlx5_common_verbs_reg_mr(priv->pd, opaq_buf, qp->entries_n *
 					sizeof(struct mlx5_gga_compress_opaque),
 							 &qp->opaque_mr) != 0) {
@@ -423,6 +424,24 @@ struct mlx5_compress_qp {
 	return dseg->lkey;
 }
 
+/*
+ * Provide safe 64bit store operation to mlx5 UAR region for both 32bit and
+ * 64bit architectures.
+ */
+static __rte_always_inline void
+mlx5_compress_uar_write(uint64_t val, struct mlx5_compress_priv *priv)
+{
+#ifdef RTE_ARCH_64
+	*priv->uar_addr = val;
+#else /* !RTE_ARCH_64 */
+	rte_spinlock_lock(&priv->uar32_sl);
+	*(volatile uint32_t *)priv->uar_addr = val;
+	rte_io_wmb();
+	*((volatile uint32_t *)priv->uar_addr + 1) = val >> 32;
+	rte_spinlock_unlock(&priv->uar32_sl);
+#endif
+}
+
 static uint16_t
 mlx5_compress_enqueue_burst(void *queue_pair, struct rte_comp_op **ops,
 			    uint16_t nb_ops)
@@ -486,7 +505,7 @@ struct mlx5_compress_qp {
 	rte_io_wmb();
 	qp->sq.db_rec[MLX5_SND_DBR] = rte_cpu_to_be_32(qp->pi);
 	rte_wmb();
-	*qp->uar_addr = *(volatile uint64_t *)wqe; /* Assume 64 bit ARCH.*/
+	mlx5_compress_uar_write(*(volatile uint64_t *)wqe, qp->priv);
 	rte_wmb();
 	return nb_ops;
 }
@@ -692,6 +711,11 @@ struct mlx5_compress_qp {
 		DRV_LOG(ERR, "Failed to allocate UAR.");
 		return -1;
 	}
+	priv->uar_addr = mlx5_os_get_devx_uar_reg_addr(priv->uar);
+	MLX5_ASSERT(qp->uar_addr);
+#ifndef RTE_ARCH_64
+	rte_spinlock_init(&priv->uar32_sl);
+#endif /* RTE_ARCH_64 */
 	return 0;
 }
 
Signature · 1 line
1.8.3.1
#39 lore raw

PATCH v3 11/11 compress/mlx5: add the supported capabilities

In reply to Matan Azrad

Add all the capabilities supported by the device.

Add the driver documentations.

---
 doc/guides/compressdevs/features/mlx5.ini | 13 +++++
 doc/guides/compressdevs/index.rst         |  1 +
 doc/guides/compressdevs/mlx5.rst          | 94 +++++++++++++++++++++++++++++++
 doc/guides/rel_notes/release_21_02.rst    |  6 ++
 drivers/compress/mlx5/mlx5_compress.c     | 24 +++++++-
 5 files changed, 136 insertions(+), 2 deletions(-)
 create mode 100644 doc/guides/compressdevs/features/mlx5.ini
 create mode 100644 doc/guides/compressdevs/mlx5.rst
doc/guides/compressdevs/features/mlx5.ini new file +13 -0
new file mode 100644
index 0000000..891ce47
@@ -0,0 +1,13 @@
+;
+; Refer to default.ini for the full list of available PMD features.
+;
+; Supported features of 'MLX5' compression driver.
+;
+[Features]
+HW Accelerated = Y
+Deflate        = Y
+Adler32        = Y
+Crc32          = Y
+Adler32&Crc32  = Y
+Fixed          = Y
+Dynamic        = Y
doc/guides/compressdevs/index.rst +1 -0
index 1f37e26..54a3ef4 100644
@@ -11,6 +11,7 @@ Compression Device Drivers
 
     overview
     isal
+    mlx5
     octeontx
     qat_comp
     zlib
doc/guides/compressdevs/mlx5.rst new file +94 -0

94 lines not on this page · raw message

doc/guides/rel_notes/release_21_02.rst +6 -0
index ae36b6a..3550f4e 100644
@@ -51,6 +51,12 @@ New Features
      * Other libs
      * Apps, Examples, Tools (if significant)
 
+* **Added mlx5 compress PMD.**
+
+  Added a new compress PMD driver for Bluefield 2 adapters.
+
+  See the :doc:`../compressdevs/mlx5` for more details.
+
      This section is a comment. Do not overwrite or remove it.
      Also, make sure to start the actual text at the margin.
      =======================================================
drivers/compress/mlx5/mlx5_compress.c +22 -2
index f7ef913..a274366 100644
@@ -77,8 +77,28 @@ struct mlx5_compress_qp {
 
 int mlx5_compress_logtype;
 
-const struct rte_compressdev_capabilities mlx5_caps[RTE_COMP_ALGO_LIST_END];
-
+static const struct rte_compressdev_capabilities mlx5_caps[] = {
+	{
+		.algo = RTE_COMP_ALGO_NULL,
+		.comp_feature_flags = RTE_COMP_FF_ADLER32_CHECKSUM |
+				      RTE_COMP_FF_CRC32_CHECKSUM |
+				      RTE_COMP_FF_CRC32_ADLER32_CHECKSUM |
+				      RTE_COMP_FF_SHAREABLE_PRIV_XFORM,
+	},
+	{
+		.algo = RTE_COMP_ALGO_DEFLATE,
+		.comp_feature_flags = RTE_COMP_FF_ADLER32_CHECKSUM |
+				      RTE_COMP_FF_CRC32_CHECKSUM |
+				      RTE_COMP_FF_CRC32_ADLER32_CHECKSUM |
+				      RTE_COMP_FF_SHAREABLE_PRIV_XFORM |
+				      RTE_COMP_FF_HUFFMAN_FIXED |
+				      RTE_COMP_FF_HUFFMAN_DYNAMIC,
+		.window_size = {.min = 10, .max = 15, .increment = 1},
+	},
+	{
+		.algo = RTE_COMP_ALGO_LIST_END,
+	}
+};
 
 static void
 mlx5_compress_dev_info_get(struct rte_compressdev *dev,
Signature · 1 line
1.8.3.1
Applied
#40 lore raw
In reply to Matan Azrad
Quoted 25 lines

Subject: [PATCH v3 00/11] add mlx5 compress PMD

Add a new compress PMD for Mellanox devices.

The MLX5 compress driver library provides support for Mellanox BlueField 2 families of 25/50/100/200 Gb/s adapters.

Using the BlueField 2 device, the compress class operations can be run in parallel to the net, vdpa, and regex class operations.

This driver is depending on rdma-core like the other mlx5 PMDs, also it is going to use mlx5 DevX to create HW objects directly by the FW.

v3:
Address Slava and Akhil comments:
- Maintain alphabetical order in compress driver lists.
- Improve docs \ commit logs with more information.
- Fix checkptch warning.
- Support 32-bit systems.
- Validate primary procces only working.
- fix memory leak in error flow of QP creation.
- Improve coding style.
- Improve operation error report in data-path.
rebased code.

Quoted 3 lines

v2:
Add documentation.

Applied to dpdk-next-crypto

Send a separate patch for whatever changes you suspect in patch 1/11. I will squash it before the tree is pulled in main branch.

Thanks.