net/mlx5: Use order-0 allocations for EQs

Currently we are allocating high-order page for EQs. In case of
fragmented system, VF hot remove/add in VMs for example, there isn't
enough contiguous memory for EQs allocation, which results in crashing
of the VM.
Therefore, use order-0 fragments for the EQ allocations instead.

Performance tests:
ConnectX-5 100Gbps, CPU: Intel(R) Xeon(R) CPU E5-2697 v3 @ 2.60GHz
Performance tests show no sensible degradation.

Signed-off-by: Tariq Toukan <tariqt@nvidia.com>
Signed-off-by: Shay Drory <shayd@nvidia.com>
Signed-off-by: Saeed Mahameed <saeedm@nvidia.com>
This commit is contained in:
Tariq Toukan 2021-03-10 23:09:15 -08:00 коммит произвёл Saeed Mahameed
Родитель c3e666f1ad
Коммит 26bf30902c
5 изменённых файлов: 32 добавлений и 22 удалений

Просмотреть файл

@ -128,7 +128,7 @@ int mlx5e_health_eq_diag_fmsg(struct mlx5_eq_comp *eq, struct devlink_fmsg *fmsg
if (err)
return err;
err = devlink_fmsg_u32_pair_put(fmsg, "size", eq->core.nent);
err = devlink_fmsg_u32_pair_put(fmsg, "size", eq_get_size(&eq->core));
if (err)
return err;

Просмотреть файл

@ -271,7 +271,7 @@ static void init_eq_buf(struct mlx5_eq *eq)
struct mlx5_eqe *eqe;
int i;
for (i = 0; i < eq->nent; i++) {
for (i = 0; i < eq_get_size(eq); i++) {
eqe = get_eqe(eq, i);
eqe->owner = MLX5_EQE_OWNER_INIT_VAL;
}
@ -281,8 +281,10 @@ static int
create_map_eq(struct mlx5_core_dev *dev, struct mlx5_eq *eq,
struct mlx5_eq_param *param)
{
u8 log_eq_size = order_base_2(param->nent + MLX5_NUM_SPARE_EQE);
struct mlx5_cq_table *cq_table = &eq->cq_table;
u32 out[MLX5_ST_SZ_DW(create_eq_out)] = {0};
u8 log_eq_stride = ilog2(MLX5_EQE_SIZE);
struct mlx5_priv *priv = &dev->priv;
u8 vecidx = param->irq_index;
__be64 *pas;
@ -297,16 +299,18 @@ create_map_eq(struct mlx5_core_dev *dev, struct mlx5_eq *eq,
spin_lock_init(&cq_table->lock);
INIT_RADIX_TREE(&cq_table->tree, GFP_ATOMIC);
eq->nent = roundup_pow_of_two(param->nent + MLX5_NUM_SPARE_EQE);
eq->cons_index = 0;
err = mlx5_buf_alloc(dev, eq->nent * MLX5_EQE_SIZE, &eq->buf);
err = mlx5_frag_buf_alloc_node(dev, wq_get_byte_sz(log_eq_size, log_eq_stride),
&eq->frag_buf, dev->priv.numa_node);
if (err)
return err;
mlx5_init_fbc(eq->frag_buf.frags, log_eq_stride, log_eq_size, &eq->fbc);
init_eq_buf(eq);
inlen = MLX5_ST_SZ_BYTES(create_eq_in) +
MLX5_FLD_SZ_BYTES(create_eq_in, pas[0]) * eq->buf.npages;
MLX5_FLD_SZ_BYTES(create_eq_in, pas[0]) * eq->frag_buf.npages;
in = kvzalloc(inlen, GFP_KERNEL);
if (!in) {
@ -315,7 +319,7 @@ create_map_eq(struct mlx5_core_dev *dev, struct mlx5_eq *eq,
}
pas = (__be64 *)MLX5_ADDR_OF(create_eq_in, in, pas);
mlx5_fill_page_array(&eq->buf, pas);
mlx5_fill_page_frag_array(&eq->frag_buf, pas);
MLX5_SET(create_eq_in, in, opcode, MLX5_CMD_OP_CREATE_EQ);
if (!param->mask[0] && MLX5_CAP_GEN(dev, log_max_uctx))
@ -326,11 +330,11 @@ create_map_eq(struct mlx5_core_dev *dev, struct mlx5_eq *eq,
param->mask[i]);
eqc = MLX5_ADDR_OF(create_eq_in, in, eq_context_entry);
MLX5_SET(eqc, eqc, log_eq_size, ilog2(eq->nent));
MLX5_SET(eqc, eqc, log_eq_size, eq->fbc.log_sz);
MLX5_SET(eqc, eqc, uar_page, priv->uar->index);
MLX5_SET(eqc, eqc, intr, vecidx);
MLX5_SET(eqc, eqc, log_page_size,
eq->buf.page_shift - MLX5_ADAPTER_PAGE_SHIFT);
eq->frag_buf.page_shift - MLX5_ADAPTER_PAGE_SHIFT);
err = mlx5_cmd_exec(dev, in, inlen, out, sizeof(out));
if (err)
@ -356,7 +360,7 @@ err_in:
kvfree(in);
err_buf:
mlx5_buf_free(dev, &eq->buf);
mlx5_frag_buf_free(dev, &eq->frag_buf);
return err;
}
@ -413,7 +417,7 @@ static int destroy_unmap_eq(struct mlx5_core_dev *dev, struct mlx5_eq *eq)
eq->eqn);
synchronize_irq(eq->irqn);
mlx5_buf_free(dev, &eq->buf);
mlx5_frag_buf_free(dev, &eq->frag_buf);
return err;
}
@ -764,10 +768,11 @@ EXPORT_SYMBOL(mlx5_eq_destroy_generic);
struct mlx5_eqe *mlx5_eq_get_eqe(struct mlx5_eq *eq, u32 cc)
{
u32 ci = eq->cons_index + cc;
u32 nent = eq_get_size(eq);
struct mlx5_eqe *eqe;
eqe = get_eqe(eq, ci & (eq->nent - 1));
eqe = ((eqe->owner & 1) ^ !!(ci & eq->nent)) ? NULL : eqe;
eqe = get_eqe(eq, ci & (nent - 1));
eqe = ((eqe->owner & 1) ^ !!(ci & nent)) ? NULL : eqe;
/* Make sure we read EQ entry contents after we've
* checked the ownership bit.
*/

Просмотреть файл

@ -22,15 +22,15 @@ struct mlx5_cq_table {
};
struct mlx5_eq {
struct mlx5_frag_buf_ctrl fbc;
struct mlx5_frag_buf frag_buf;
struct mlx5_core_dev *dev;
struct mlx5_cq_table cq_table;
__be32 __iomem *doorbell;
u32 cons_index;
struct mlx5_frag_buf buf;
unsigned int vecidx;
unsigned int irqn;
u8 eqn;
int nent;
struct mlx5_rsc_debug *dbg;
};
@ -47,16 +47,21 @@ struct mlx5_eq_comp {
struct list_head list;
};
static inline u32 eq_get_size(struct mlx5_eq *eq)
{
return eq->fbc.sz_m1 + 1;
}
static inline struct mlx5_eqe *get_eqe(struct mlx5_eq *eq, u32 entry)
{
return mlx5_buf_offset(&eq->buf, entry * MLX5_EQE_SIZE);
return mlx5_frag_buf_get_wqe(&eq->fbc, entry);
}
static inline struct mlx5_eqe *next_eqe_sw(struct mlx5_eq *eq)
{
struct mlx5_eqe *eqe = get_eqe(eq, eq->cons_index & (eq->nent - 1));
struct mlx5_eqe *eqe = get_eqe(eq, eq->cons_index & eq->fbc.sz_m1);
return ((eqe->owner & 1) ^ !!(eq->cons_index & eq->nent)) ? NULL : eqe;
return (eqe->owner ^ (eq->cons_index >> eq->fbc.log_sz)) & 1 ? NULL : eqe;
}
static inline void eq_update_ci(struct mlx5_eq *eq, int arm)

Просмотреть файл

@ -34,11 +34,6 @@
#include "wq.h"
#include "mlx5_core.h"
static u32 wq_get_byte_sz(u8 log_sz, u8 log_stride)
{
return ((u32)1 << log_sz) << log_stride;
}
int mlx5_wq_cyc_create(struct mlx5_core_dev *mdev, struct mlx5_wq_param *param,
void *wqc, struct mlx5_wq_cyc *wq,
struct mlx5_wq_ctrl *wq_ctrl)

Просмотреть файл

@ -873,6 +873,11 @@ static inline u32 mlx5_base_mkey(const u32 key)
return key & 0xffffff00u;
}
static inline u32 wq_get_byte_sz(u8 log_sz, u8 log_stride)
{
return ((u32)1 << log_sz) << log_stride;
}
static inline void mlx5_init_fbc_offset(struct mlx5_buf_list *frags,
u8 log_stride, u8 log_sz,
u16 strides_offset,