[ovs-dev,12/25] netdev-offload-dpdk: Introduce map APIs for table id and miss context

Different vports are differentiated in HW by different tables, as in the
SW model. As such we need to map to table ids. Also, as this offload
involves multiple tables, a miss might occur in the target table. In
such case we need to recover the packet and continue in SW. Introduce a
mapping for a miss context for that.

Signed-off-by: Eli Britstein <elibr@mellanox.com>
Reviewed-by: Roni Bar Yanai <roniba@mellanox.com>
---
 lib/netdev-offload-dpdk.c | 267 ++++++++++++++++++++++++++++++++++++++++++++++
 1 file changed, 267 insertions(+)

On Mon, Jan 20, 2020 at 7:09 AM Eli Britstein <elibr@mellanox.com> wrote:
>
> Different vports are differentiated in HW by different tables, as in the
> SW model. As such we need to map to table ids. Also, as this offload
> involves multiple tables, a miss might occur in the target table. In
> such case we need to recover the packet and continue in SW. Introduce a
> mapping for a miss context for that.
>

Hi Eli,

Can you elaborate a little more on
"Different vports are differentiated in HW by different tables, as in
the SW model."?

Do you mean for each tunnel type in HW, there is a separate table?
ex: a vport such as vxlan_sys_4789 has one table, and gre_sys has another table?

And is this table in HW or in SW?

Thanks
William

> Signed-off-by: Eli Britstein <elibr@mellanox.com>
> Reviewed-by: Roni Bar Yanai <roniba@mellanox.com>
> ---
>  lib/netdev-offload-dpdk.c | 267 ++++++++++++++++++++++++++++++++++++++++++++++
>  1 file changed, 267 insertions(+)
>
> diff --git a/lib/netdev-offload-dpdk.c b/lib/netdev-offload-dpdk.c
> index c80f07e77..fc890b915 100644
> --- a/lib/netdev-offload-dpdk.c
> +++ b/lib/netdev-offload-dpdk.c
> @@ -29,6 +29,8 @@
>  #include "openvswitch/vlog.h"
>  #include "packets.h"
>  #include "uuid.h"
> +#include "id-pool.h"
> +#include "odp-util.h"
>
>  VLOG_DEFINE_THIS_MODULE(netdev_offload_dpdk);
>  static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(100, 5);
> @@ -125,6 +127,271 @@ ufid_to_rte_flow_disassociate(const ovs_u128 *ufid)
>                UUID_ARGS((struct uuid *) ufid));
>  }
>
> +/* A generic data structure used for mapping data to id and id to data. The
> + * elements are reference coutned. As changes are done only from the single
> + * offload thread, no locks are required.
> + * "name" and "dump_context_data" are used for log messages.
> + * "d2i_hmap" is the data-to-id map.
> + * "i2d_hmap" is the id-to-data map.
> + * "id_alloc" is used to allocate an id for a new data.
> + * "id_free" is used to free an id for the last data release.
> + * "data_size" is the size of the data in the elements.
> + */
> +struct context_metadata {
> +    const char *name;
> +    struct ds *(*dump_context_data)(struct ds *s, void *data);
> +    struct hmap d2i_hmap;
> +    struct hmap i2d_hmap;
> +    uint32_t (*id_alloc)(void);
> +    void (*id_free)(uint32_t id);
> +    size_t data_size;
> +};
> +
> +struct context_data {
> +    struct hmap_node d2i_node;
> +    struct hmap_node i2d_node;
> +    void *data;
> +    uint32_t id;
> +    uint32_t refcnt;
> +};
> +
> +static int
> +get_context_data_id_by_data(struct context_metadata *md,
> +                            struct context_data *data_req,
> +                            uint32_t *id)
> +{
> +    struct context_data *data_cur;
> +    size_t dhash, ihash;
> +    struct ds s;
> +
> +    ds_init(&s);
> +    dhash = hash_bytes(data_req->data, md->data_size, 0);
> +    HMAP_FOR_EACH_WITH_HASH (data_cur, d2i_node, dhash, &md->d2i_hmap) {
> +        if (!memcmp(data_req->data, data_cur->data, md->data_size)) {
> +            data_cur->refcnt++;
> +            VLOG_DBG_RL(&rl,
> +                        "%s: %s: '%s', refcnt=%d, id=%d", __func__, md->name,
> +                        ds_cstr(md->dump_context_data(&s, data_cur->data)),
> +                        data_cur->refcnt, data_cur->id);
> +            ds_destroy(&s);
> +            *id = data_cur->id;
> +            return 0;
> +        }
> +    }
> +
> +    data_cur = xzalloc(sizeof *data_cur);
> +    if (!data_cur) {
> +        goto err;
> +    }
> +    data_cur->data = xmalloc(md->data_size);
> +    if (!data_cur->data) {
> +        goto err_data_alloc;
> +    }
> +    memcpy(data_cur->data, data_req->data, md->data_size);
> +    data_cur->refcnt = 1;
> +    data_cur->id = md->id_alloc();
> +    if (data_cur->id == 0) {
> +        goto err_id_alloc;
> +    }
> +    hmap_insert(&md->d2i_hmap, &data_cur->d2i_node, dhash);
> +    ihash = hash_add(0, data_cur->id);
> +    hmap_insert(&md->i2d_hmap, &data_cur->i2d_node, ihash);
> +    VLOG_DBG_RL(&rl, "%s: %s: '%s', refcnt=%d, id=%d", __func__, md->name,
> +                ds_cstr(md->dump_context_data(&s, data_cur->data)),
> +                data_cur->refcnt, data_cur->id);
> +    *id = data_cur->id;
> +    ds_destroy(&s);
> +    return 0;
> +
> +err_id_alloc:
> +    free(data_cur->data);
> +err_data_alloc:
> +    free(data_cur);
> +err:
> +    VLOG_ERR_RL(&rl, "%s: %s: error. '%s'", __func__, md->name,
> +                ds_cstr(md->dump_context_data(&s, data_cur->data)));
> +    ds_destroy(&s);
> +    return -1;
> +}
> +
> +static int
> +get_context_data_by_id(struct context_metadata *md, uint32_t id, void *data)
> +{
> +    size_t ihash = hash_add(0, id);
> +    struct context_data *data_cur;
> +    struct ds s;
> +
> +    ds_init(&s);
> +    HMAP_FOR_EACH_WITH_HASH (data_cur, i2d_node, ihash, &md->i2d_hmap) {
> +        if (data_cur->id == id) {
> +            memcpy(data, data_cur->data, md->data_size);
> +            ds_destroy(&s);
> +            return 0;
> +        }
> +    }
> +
> +    ds_destroy(&s);
> +    return -1;
> +}
> +
> +static void
> +put_context_data_by_id(struct context_metadata *md, uint32_t id)
> +{
> +    struct context_data *data_cur;
> +    size_t ihash;
> +    struct ds s;
> +
> +    if (id == 0) {
> +        return;
> +    }
> +    ihash = hash_add(0, id);
> +    HMAP_FOR_EACH_WITH_HASH (data_cur, i2d_node, ihash, &md->i2d_hmap) {
> +        if (data_cur->id == id) {
> +            data_cur->refcnt--;
> +            ds_init(&s);
> +            VLOG_DBG_RL(&rl,
> +                        "%s: %s: '%s', refcnt=%d, id=%d", __func__, md->name,
> +                        ds_cstr(md->dump_context_data(&s, data_cur->data)),
> +                        data_cur->refcnt, data_cur->id);
> +            ds_destroy(&s);
> +            if (data_cur->refcnt == 0) {
> +                hmap_remove(&md->i2d_hmap, &data_cur->i2d_node);
> +                hmap_remove(&md->d2i_hmap, &data_cur->d2i_node);
> +                free(data_cur);
> +                md->id_free(id);
> +            }
> +            return;
> +        }
> +    }
> +    VLOG_ERR_RL(&rl,
> +                "%s: %s: error. id=%d not found", __func__, md->name, id);
> +}
> +
> +struct table_id_data {
> +    odp_port_t vport;
> +};
> +
> +static struct ds *
> +dump_table_id(struct ds *s, void *data)
> +{
> +    struct table_id_data *table_id_data = data;
> +
> +    ds_put_format(s, "vport=%"PRIu32, table_id_data->vport);
> +    return s;
> +}
> +
> +#define MIN_TABLE_ID     1
> +#define MAX_TABLE_ID     0xFFFF
> +
> +static struct id_pool *table_id_pool = NULL;
> +static uint32_t
> +table_id_alloc(void)
> +{
> +    uint32_t id;
> +
> +    if (!table_id_pool) {
> +        /* Haven't initiated yet, do it here */
> +        table_id_pool = id_pool_create(MIN_TABLE_ID, MAX_TABLE_ID);
> +    }
> +
> +    if (id_pool_alloc_id(table_id_pool, &id)) {
> +        return id;
> +    }
> +
> +    return 0;
> +}
> +
> +static void
> +table_id_free(uint32_t id)
> +{
> +    id_pool_free_id(table_id_pool, id);
> +}
> +
> +static struct context_metadata table_id_md = {
> +    .name = "table_id",
> +    .dump_context_data = dump_table_id,
> +    .d2i_hmap = HMAP_INITIALIZER(&table_id_md.d2i_hmap),
> +    .i2d_hmap = HMAP_INITIALIZER(&table_id_md.i2d_hmap),
> +    .id_alloc = table_id_alloc,
> +    .id_free = table_id_free,
> +    .data_size = sizeof(struct table_id_data),
> +};
> +
> +OVS_UNUSED
> +static int
> +get_table_id(odp_port_t vport, uint32_t *table_id)
> +{
> +    struct table_id_data table_id_data = { .vport = vport };
> +    struct context_data table_id_context = {
> +        .data = &table_id_data,
> +    };
> +
> +    if (vport == ODPP_NONE) {
> +        *table_id = 0;
> +        return 0;
> +    }
> +
> +    return get_context_data_id_by_data(&table_id_md, &table_id_context,
> +                                       table_id);
> +}
> +
> +OVS_UNUSED
> +static void
> +put_table_id(uint32_t table_id)
> +{
> +    put_context_data_by_id(&table_id_md, table_id);
> +}
> +
> +struct flow_miss_ctx {
> +    odp_port_t vport;
> +};
> +
> +static struct ds *
> +dump_flow_ctx_id(struct ds *s, void *data)
> +{
> +    struct flow_miss_ctx *flow_ctx_data = data;
> +
> +    ds_put_format(s, "vport=%"PRIu32, flow_ctx_data->vport);
> +    return s;
> +}
> +
> +static struct context_metadata flow_miss_ctx_md = {
> +    .name = "flow_miss_ctx",
> +    .dump_context_data = dump_flow_ctx_id,
> +    .d2i_hmap = HMAP_INITIALIZER(&flow_miss_ctx_md.d2i_hmap),
> +    .i2d_hmap = HMAP_INITIALIZER(&flow_miss_ctx_md.i2d_hmap),
> +    .id_alloc = netdev_offload_flow_mark_alloc,
> +    .id_free = netdev_offload_flow_mark_free,
> +    .data_size = sizeof(struct flow_miss_ctx),
> +};
> +
> +OVS_UNUSED
> +static int
> +get_flow_miss_ctx_id(struct flow_miss_ctx *flow_ctx_data,
> +                     uint32_t *miss_ctx_id)
> +{
> +    struct context_data flow_ctx = {
> +        .data = flow_ctx_data,
> +    };
> +
> +    return get_context_data_id_by_data(&flow_miss_ctx_md, &flow_ctx,
> +                                       miss_ctx_id);
> +}
> +
> +OVS_UNUSED
> +static void
> +put_flow_miss_ctx_id(uint32_t flow_ctx_id)
> +{
> +    put_context_data_by_id(&flow_miss_ctx_md, flow_ctx_id);
> +}
> +
> +OVS_UNUSED
> +static int
> +find_flow_miss_ctx(int flow_ctx_id, struct flow_miss_ctx *ctx)
> +{
> +    return get_context_data_by_id(&flow_miss_ctx_md, flow_ctx_id, ctx);
> +}
> +
>  /*
>   * To avoid individual xrealloc calls for each new element, a 'curent_max'
>   * is used to keep track of current allocated number of elements. Starts
> --
> 2.14.5
>
> _______________________________________________
> dev mailing list
> dev@openvswitch.org
> https://mail.openvswitch.org/mailman/listinfo/ovs-dev

On 7/11/2020 9:44 PM, William Tu wrote:
> On Mon, Jan 20, 2020 at 7:09 AM Eli Britstein <elibr@mellanox.com> wrote:
>> Different vports are differentiated in HW by different tables, as in the
>> SW model. As such we need to map to table ids. Also, as this offload
>> involves multiple tables, a miss might occur in the target table. In
>> such case we need to recover the packet and continue in SW. Introduce a
>> mapping for a miss context for that.
>>
> Hi Eli,
>
> Can you elaborate a little more on
> "Different vports are differentiated in HW by different tables, as in
> the SW model."?
>
> Do you mean for each tunnel type in HW, there is a separate table?
> ex: a vport such as vxlan_sys_4789 has one table, and gre_sys has another table?
>
> And is this table in HW or in SW?

Yes. In SW there is a match on the in_port (vxlan_sys_4789, gre_sys and 
also non-vports). This is equivalent to different logical tables. In HW, 
in this patch-set this is done by different groups (HW different tables).

Please note that we would like to progress with an approach to enhance 
DPDK and put some of the logic inside the PMD. See:

Link to RFC: http://mails.dpdk.org/archives/dev/2020-June/169656.html
Link to patchset: http://mails.dpdk.org/archives/dev/2020-June/171590.html

> Thanks
> William
>
>> Signed-off-by: Eli Britstein <elibr@mellanox.com>
>> Reviewed-by: Roni Bar Yanai <roniba@mellanox.com>
>> ---
>>   lib/netdev-offload-dpdk.c | 267 ++++++++++++++++++++++++++++++++++++++++++++++
>>   1 file changed, 267 insertions(+)
>>
>> diff --git a/lib/netdev-offload-dpdk.c b/lib/netdev-offload-dpdk.c
>> index c80f07e77..fc890b915 100644
>> --- a/lib/netdev-offload-dpdk.c
>> +++ b/lib/netdev-offload-dpdk.c
>> @@ -29,6 +29,8 @@
>>   #include "openvswitch/vlog.h"
>>   #include "packets.h"
>>   #include "uuid.h"
>> +#include "id-pool.h"
>> +#include "odp-util.h"
>>
>>   VLOG_DEFINE_THIS_MODULE(netdev_offload_dpdk);
>>   static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(100, 5);
>> @@ -125,6 +127,271 @@ ufid_to_rte_flow_disassociate(const ovs_u128 *ufid)
>>                 UUID_ARGS((struct uuid *) ufid));
>>   }
>>
>> +/* A generic data structure used for mapping data to id and id to data. The
>> + * elements are reference coutned. As changes are done only from the single
>> + * offload thread, no locks are required.
>> + * "name" and "dump_context_data" are used for log messages.
>> + * "d2i_hmap" is the data-to-id map.
>> + * "i2d_hmap" is the id-to-data map.
>> + * "id_alloc" is used to allocate an id for a new data.
>> + * "id_free" is used to free an id for the last data release.
>> + * "data_size" is the size of the data in the elements.
>> + */
>> +struct context_metadata {
>> +    const char *name;
>> +    struct ds *(*dump_context_data)(struct ds *s, void *data);
>> +    struct hmap d2i_hmap;
>> +    struct hmap i2d_hmap;
>> +    uint32_t (*id_alloc)(void);
>> +    void (*id_free)(uint32_t id);
>> +    size_t data_size;
>> +};
>> +
>> +struct context_data {
>> +    struct hmap_node d2i_node;
>> +    struct hmap_node i2d_node;
>> +    void *data;
>> +    uint32_t id;
>> +    uint32_t refcnt;
>> +};
>> +
>> +static int
>> +get_context_data_id_by_data(struct context_metadata *md,
>> +                            struct context_data *data_req,
>> +                            uint32_t *id)
>> +{
>> +    struct context_data *data_cur;
>> +    size_t dhash, ihash;
>> +    struct ds s;
>> +
>> +    ds_init(&s);
>> +    dhash = hash_bytes(data_req->data, md->data_size, 0);
>> +    HMAP_FOR_EACH_WITH_HASH (data_cur, d2i_node, dhash, &md->d2i_hmap) {
>> +        if (!memcmp(data_req->data, data_cur->data, md->data_size)) {
>> +            data_cur->refcnt++;
>> +            VLOG_DBG_RL(&rl,
>> +                        "%s: %s: '%s', refcnt=%d, id=%d", __func__, md->name,
>> +                        ds_cstr(md->dump_context_data(&s, data_cur->data)),
>> +                        data_cur->refcnt, data_cur->id);
>> +            ds_destroy(&s);
>> +            *id = data_cur->id;
>> +            return 0;
>> +        }
>> +    }
>> +
>> +    data_cur = xzalloc(sizeof *data_cur);
>> +    if (!data_cur) {
>> +        goto err;
>> +    }
>> +    data_cur->data = xmalloc(md->data_size);
>> +    if (!data_cur->data) {
>> +        goto err_data_alloc;
>> +    }
>> +    memcpy(data_cur->data, data_req->data, md->data_size);
>> +    data_cur->refcnt = 1;
>> +    data_cur->id = md->id_alloc();
>> +    if (data_cur->id == 0) {
>> +        goto err_id_alloc;
>> +    }
>> +    hmap_insert(&md->d2i_hmap, &data_cur->d2i_node, dhash);
>> +    ihash = hash_add(0, data_cur->id);
>> +    hmap_insert(&md->i2d_hmap, &data_cur->i2d_node, ihash);
>> +    VLOG_DBG_RL(&rl, "%s: %s: '%s', refcnt=%d, id=%d", __func__, md->name,
>> +                ds_cstr(md->dump_context_data(&s, data_cur->data)),
>> +                data_cur->refcnt, data_cur->id);
>> +    *id = data_cur->id;
>> +    ds_destroy(&s);
>> +    return 0;
>> +
>> +err_id_alloc:
>> +    free(data_cur->data);
>> +err_data_alloc:
>> +    free(data_cur);
>> +err:
>> +    VLOG_ERR_RL(&rl, "%s: %s: error. '%s'", __func__, md->name,
>> +                ds_cstr(md->dump_context_data(&s, data_cur->data)));
>> +    ds_destroy(&s);
>> +    return -1;
>> +}
>> +
>> +static int
>> +get_context_data_by_id(struct context_metadata *md, uint32_t id, void *data)
>> +{
>> +    size_t ihash = hash_add(0, id);
>> +    struct context_data *data_cur;
>> +    struct ds s;
>> +
>> +    ds_init(&s);
>> +    HMAP_FOR_EACH_WITH_HASH (data_cur, i2d_node, ihash, &md->i2d_hmap) {
>> +        if (data_cur->id == id) {
>> +            memcpy(data, data_cur->data, md->data_size);
>> +            ds_destroy(&s);
>> +            return 0;
>> +        }
>> +    }
>> +
>> +    ds_destroy(&s);
>> +    return -1;
>> +}
>> +
>> +static void
>> +put_context_data_by_id(struct context_metadata *md, uint32_t id)
>> +{
>> +    struct context_data *data_cur;
>> +    size_t ihash;
>> +    struct ds s;
>> +
>> +    if (id == 0) {
>> +        return;
>> +    }
>> +    ihash = hash_add(0, id);
>> +    HMAP_FOR_EACH_WITH_HASH (data_cur, i2d_node, ihash, &md->i2d_hmap) {
>> +        if (data_cur->id == id) {
>> +            data_cur->refcnt--;
>> +            ds_init(&s);
>> +            VLOG_DBG_RL(&rl,
>> +                        "%s: %s: '%s', refcnt=%d, id=%d", __func__, md->name,
>> +                        ds_cstr(md->dump_context_data(&s, data_cur->data)),
>> +                        data_cur->refcnt, data_cur->id);
>> +            ds_destroy(&s);
>> +            if (data_cur->refcnt == 0) {
>> +                hmap_remove(&md->i2d_hmap, &data_cur->i2d_node);
>> +                hmap_remove(&md->d2i_hmap, &data_cur->d2i_node);
>> +                free(data_cur);
>> +                md->id_free(id);
>> +            }
>> +            return;
>> +        }
>> +    }
>> +    VLOG_ERR_RL(&rl,
>> +                "%s: %s: error. id=%d not found", __func__, md->name, id);
>> +}
>> +
>> +struct table_id_data {
>> +    odp_port_t vport;
>> +};
>> +
>> +static struct ds *
>> +dump_table_id(struct ds *s, void *data)
>> +{
>> +    struct table_id_data *table_id_data = data;
>> +
>> +    ds_put_format(s, "vport=%"PRIu32, table_id_data->vport);
>> +    return s;
>> +}
>> +
>> +#define MIN_TABLE_ID     1
>> +#define MAX_TABLE_ID     0xFFFF
>> +
>> +static struct id_pool *table_id_pool = NULL;
>> +static uint32_t
>> +table_id_alloc(void)
>> +{
>> +    uint32_t id;
>> +
>> +    if (!table_id_pool) {
>> +        /* Haven't initiated yet, do it here */
>> +        table_id_pool = id_pool_create(MIN_TABLE_ID, MAX_TABLE_ID);
>> +    }
>> +
>> +    if (id_pool_alloc_id(table_id_pool, &id)) {
>> +        return id;
>> +    }
>> +
>> +    return 0;
>> +}
>> +
>> +static void
>> +table_id_free(uint32_t id)
>> +{
>> +    id_pool_free_id(table_id_pool, id);
>> +}
>> +
>> +static struct context_metadata table_id_md = {
>> +    .name = "table_id",
>> +    .dump_context_data = dump_table_id,
>> +    .d2i_hmap = HMAP_INITIALIZER(&table_id_md.d2i_hmap),
>> +    .i2d_hmap = HMAP_INITIALIZER(&table_id_md.i2d_hmap),
>> +    .id_alloc = table_id_alloc,
>> +    .id_free = table_id_free,
>> +    .data_size = sizeof(struct table_id_data),
>> +};
>> +
>> +OVS_UNUSED
>> +static int
>> +get_table_id(odp_port_t vport, uint32_t *table_id)
>> +{
>> +    struct table_id_data table_id_data = { .vport = vport };
>> +    struct context_data table_id_context = {
>> +        .data = &table_id_data,
>> +    };
>> +
>> +    if (vport == ODPP_NONE) {
>> +        *table_id = 0;
>> +        return 0;
>> +    }
>> +
>> +    return get_context_data_id_by_data(&table_id_md, &table_id_context,
>> +                                       table_id);
>> +}
>> +
>> +OVS_UNUSED
>> +static void
>> +put_table_id(uint32_t table_id)
>> +{
>> +    put_context_data_by_id(&table_id_md, table_id);
>> +}
>> +
>> +struct flow_miss_ctx {
>> +    odp_port_t vport;
>> +};
>> +
>> +static struct ds *
>> +dump_flow_ctx_id(struct ds *s, void *data)
>> +{
>> +    struct flow_miss_ctx *flow_ctx_data = data;
>> +
>> +    ds_put_format(s, "vport=%"PRIu32, flow_ctx_data->vport);
>> +    return s;
>> +}
>> +
>> +static struct context_metadata flow_miss_ctx_md = {
>> +    .name = "flow_miss_ctx",
>> +    .dump_context_data = dump_flow_ctx_id,
>> +    .d2i_hmap = HMAP_INITIALIZER(&flow_miss_ctx_md.d2i_hmap),
>> +    .i2d_hmap = HMAP_INITIALIZER(&flow_miss_ctx_md.i2d_hmap),
>> +    .id_alloc = netdev_offload_flow_mark_alloc,
>> +    .id_free = netdev_offload_flow_mark_free,
>> +    .data_size = sizeof(struct flow_miss_ctx),
>> +};
>> +
>> +OVS_UNUSED
>> +static int
>> +get_flow_miss_ctx_id(struct flow_miss_ctx *flow_ctx_data,
>> +                     uint32_t *miss_ctx_id)
>> +{
>> +    struct context_data flow_ctx = {
>> +        .data = flow_ctx_data,
>> +    };
>> +
>> +    return get_context_data_id_by_data(&flow_miss_ctx_md, &flow_ctx,
>> +                                       miss_ctx_id);
>> +}
>> +
>> +OVS_UNUSED
>> +static void
>> +put_flow_miss_ctx_id(uint32_t flow_ctx_id)
>> +{
>> +    put_context_data_by_id(&flow_miss_ctx_md, flow_ctx_id);
>> +}
>> +
>> +OVS_UNUSED
>> +static int
>> +find_flow_miss_ctx(int flow_ctx_id, struct flow_miss_ctx *ctx)
>> +{
>> +    return get_context_data_by_id(&flow_miss_ctx_md, flow_ctx_id, ctx);
>> +}
>> +
>>   /*
>>    * To avoid individual xrealloc calls for each new element, a 'curent_max'
>>    * is used to keep track of current allocated number of elements. Starts
>> --
>> 2.14.5
>>
>> _______________________________________________
>> dev mailing list
>> dev@openvswitch.org
>> https://eur03.safelinks.protection.outlook.com/?url=https%3A%2F%2Fmail.openvswitch.org%2Fmailman%2Flistinfo%2Fovs-dev&amp;data=02%7C01%7Celibr%40mellanox.com%7Cb16d91817888446a8fc508d825ca92c6%7Ca652971c7d2e4d9ba6a4d149256f461b%7C0%7C0%7C637300899272197221&amp;sdata=y2Ig3QRkmo0I%2B2bYbeJoA7G%2Bl%2BYoNP6bcuHrn1rKfys%3D&amp;reserved=0

On Sat, Jul 11, 2020 at 11:22 PM Eli Britstein <elibr@mellanox.com> wrote:
>
>
> On 7/11/2020 9:44 PM, William Tu wrote:
> > On Mon, Jan 20, 2020 at 7:09 AM Eli Britstein <elibr@mellanox.com> wrote:
> >> Different vports are differentiated in HW by different tables, as in the
> >> SW model. As such we need to map to table ids. Also, as this offload
> >> involves multiple tables, a miss might occur in the target table. In
> >> such case we need to recover the packet and continue in SW. Introduce a
> >> mapping for a miss context for that.
> >>
> > Hi Eli,
> >
> > Can you elaborate a little more on
> > "Different vports are differentiated in HW by different tables, as in
> > the SW model."?
> >
> > Do you mean for each tunnel type in HW, there is a separate table?
> > ex: a vport such as vxlan_sys_4789 has one table, and gre_sys has another table?
> >
> > And is this table in HW or in SW?
>
> Yes. In SW there is a match on the in_port (vxlan_sys_4789, gre_sys and
> also non-vports). This is equivalent to different logical tables. In HW,
> in this patch-set this is done by different groups (HW different tables).
>
> Please note that we would like to progress with an approach to enhance
> DPDK and put some of the logic inside the PMD. See:
>
> Link to RFC: http://mails.dpdk.org/archives/dev/2020-June/169656.html
> Link to patchset: http://mails.dpdk.org/archives/dev/2020-June/171590.html

Thank you!
I will check it out.
William