{
struct mlx5_ib_mr *mmr = to_mmr(ibmr);
- if (ibmr->type == IB_MR_TYPE_INTEGRITY)
- dereg_mr(to_mdev(mmr->pi_mr->ibmr.device), mmr->pi_mr);
+ if (ibmr->type == IB_MR_TYPE_INTEGRITY) {
+ dereg_mr(to_mdev(mmr->mtt_mr->ibmr.device), mmr->mtt_mr);
+ dereg_mr(to_mdev(mmr->klm_mr->ibmr.device), mmr->klm_mr);
+ }
dereg_mr(to_mdev(ibmr->device), mmr);
}
static struct mlx5_ib_mr *mlx5_ib_alloc_pi_mr(struct ib_pd *pd,
- u32 max_num_sg, u32 max_num_meta_sg)
+ u32 max_num_sg, u32 max_num_meta_sg,
+ int desc_size, int access_mode)
{
struct mlx5_ib_dev *dev = to_mdev(pd->device);
int inlen = MLX5_ST_SZ_BYTES(create_mkey_in);
mkc = MLX5_ADDR_OF(create_mkey_in, in, memory_key_mkey_entry);
MLX5_SET(mkc, mkc, free, 1);
MLX5_SET(mkc, mkc, translations_octword_size, ndescs);
+ if (access_mode == MLX5_MKC_ACCESS_MODE_MTT)
+ MLX5_SET(mkc, mkc, log_page_size, PAGE_SHIFT);
MLX5_SET(mkc, mkc, qpn, 0xffffff);
MLX5_SET(mkc, mkc, pd, to_mpd(pd)->pdn);
- mr->access_mode = MLX5_MKC_ACCESS_MODE_KLMS;
+ mr->access_mode = access_mode;
- err = mlx5_alloc_priv_descs(pd->device, mr,
- ndescs, sizeof(struct mlx5_klm));
+ err = mlx5_alloc_priv_descs(pd->device, mr, ndescs, desc_size);
if (err)
goto err_free_in;
- mr->desc_size = sizeof(struct mlx5_klm);
+ mr->desc_size = desc_size;
mr->max_descs = ndescs;
MLX5_SET(mkc, mkc, access_mode_1_0, mr->access_mode & 0x3);
mr->sig->sig_err_exists = false;
/* Next UMR, Arm SIGERR */
++mr->sig->sigerr_count;
- mr->pi_mr = mlx5_ib_alloc_pi_mr(pd, max_num_sg,
- max_num_meta_sg);
- if (IS_ERR(mr->pi_mr)) {
- err = PTR_ERR(mr->pi_mr);
+ mr->klm_mr = mlx5_ib_alloc_pi_mr(pd, max_num_sg,
+ max_num_meta_sg,
+ sizeof(struct mlx5_klm),
+ MLX5_MKC_ACCESS_MODE_KLMS);
+ if (IS_ERR(mr->klm_mr)) {
+ err = PTR_ERR(mr->klm_mr);
goto err_destroy_psv;
}
+ mr->mtt_mr = mlx5_ib_alloc_pi_mr(pd, max_num_sg,
+ max_num_meta_sg,
+ sizeof(struct mlx5_mtt),
+ MLX5_MKC_ACCESS_MODE_MTT);
+ if (IS_ERR(mr->mtt_mr)) {
+ err = PTR_ERR(mr->mtt_mr);
+ goto err_free_klm_mr;
+ }
} else {
mlx5_ib_warn(dev, "Invalid mr type %d\n", mr_type);
err = -EINVAL;
return &mr->ibmr;
err_free_pi_mr:
- if (mr->pi_mr) {
- dereg_mr(to_mdev(mr->pi_mr->ibmr.device), mr->pi_mr);
- mr->pi_mr = NULL;
+ if (mr->mtt_mr) {
+ dereg_mr(to_mdev(mr->mtt_mr->ibmr.device), mr->mtt_mr);
+ mr->mtt_mr = NULL;
+ }
+err_free_klm_mr:
+ if (mr->klm_mr) {
+ dereg_mr(to_mdev(mr->klm_mr->ibmr.device), mr->klm_mr);
+ mr->klm_mr = NULL;
}
err_destroy_psv:
if (mr->sig) {
return 0;
}
-int mlx5_ib_map_mr_sg_pi(struct ib_mr *ibmr, struct scatterlist *data_sg,
+static int mlx5_set_page_pi(struct ib_mr *ibmr, u64 addr)
+{
+ struct mlx5_ib_mr *mr = to_mmr(ibmr);
+ __be64 *descs;
+
+ if (unlikely(mr->ndescs + mr->meta_ndescs == mr->max_descs))
+ return -ENOMEM;
+
+ descs = mr->descs;
+ descs[mr->ndescs + mr->meta_ndescs++] =
+ cpu_to_be64(addr | MLX5_EN_RD | MLX5_EN_WR);
+
+ return 0;
+}
+
+static int
+mlx5_ib_map_mtt_mr_sg_pi(struct ib_mr *ibmr, struct scatterlist *data_sg,
int data_sg_nents, unsigned int *data_sg_offset,
struct scatterlist *meta_sg, int meta_sg_nents,
unsigned int *meta_sg_offset)
{
struct mlx5_ib_mr *mr = to_mmr(ibmr);
- struct mlx5_ib_mr *pi_mr = mr->pi_mr;
+ struct mlx5_ib_mr *pi_mr = mr->mtt_mr;
int n;
+ u64 iova;
- WARN_ON(ibmr->type != IB_MR_TYPE_INTEGRITY);
+ pi_mr->ndescs = 0;
+ pi_mr->meta_ndescs = 0;
+ pi_mr->meta_length = 0;
+
+ ib_dma_sync_single_for_cpu(ibmr->device, pi_mr->desc_map,
+ pi_mr->desc_size * pi_mr->max_descs,
+ DMA_TO_DEVICE);
+
+ pi_mr->ibmr.page_size = ibmr->page_size;
+ n = ib_sg_to_pages(&pi_mr->ibmr, data_sg, data_sg_nents, data_sg_offset,
+ mlx5_set_page);
+ if (n != data_sg_nents)
+ return n;
+
+ iova = pi_mr->ibmr.iova;
+ pi_mr->data_length = pi_mr->ibmr.length;
+ pi_mr->ibmr.length = pi_mr->data_length;
+ ibmr->length = pi_mr->data_length;
+
+ if (meta_sg_nents) {
+ u64 page_mask = ~((u64)ibmr->page_size - 1);
+
+ n += ib_sg_to_pages(&pi_mr->ibmr, meta_sg, meta_sg_nents,
+ meta_sg_offset, mlx5_set_page_pi);
+
+ pi_mr->meta_length = pi_mr->ibmr.length;
+ /*
+ * PI address for the HW is the offset of the metadata address
+ * relative to the first data page address.
+ * It equals to first data page address + size of data pages +
+ * metadata offset at the first metadata page
+ */
+ pi_mr->pi_iova = (iova & page_mask) +
+ pi_mr->ndescs * ibmr->page_size +
+ (pi_mr->ibmr.iova & ~page_mask);
+ /*
+ * In order to use one MTT MR for data and metadata, we register
+ * also the gaps between the end of the data and the start of
+ * the metadata (the sig MR will verify that the HW will access
+ * to right addresses). This mapping is safe because we use
+ * internal mkey for the registration.
+ */
+ pi_mr->ibmr.length = pi_mr->pi_iova + pi_mr->meta_length - iova;
+ pi_mr->ibmr.iova = iova;
+ ibmr->length += pi_mr->meta_length;
+ }
+
+ ib_dma_sync_single_for_device(ibmr->device, pi_mr->desc_map,
+ pi_mr->desc_size * pi_mr->max_descs,
+ DMA_TO_DEVICE);
+
+ return n;
+}
+
+static int
+mlx5_ib_map_klm_mr_sg_pi(struct ib_mr *ibmr, struct scatterlist *data_sg,
+ int data_sg_nents, unsigned int *data_sg_offset,
+ struct scatterlist *meta_sg, int meta_sg_nents,
+ unsigned int *meta_sg_offset)
+{
+ struct mlx5_ib_mr *mr = to_mmr(ibmr);
+ struct mlx5_ib_mr *pi_mr = mr->klm_mr;
+ int n;
pi_mr->ndescs = 0;
pi_mr->meta_ndescs = 0;
n = mlx5_ib_sg_to_klms(pi_mr, data_sg, data_sg_nents, data_sg_offset,
meta_sg, meta_sg_nents, meta_sg_offset);
+ ib_dma_sync_single_for_device(ibmr->device, pi_mr->desc_map,
+ pi_mr->desc_size * pi_mr->max_descs,
+ DMA_TO_DEVICE);
+
/* This is zero-based memory region */
pi_mr->ibmr.iova = 0;
+ pi_mr->pi_iova = pi_mr->data_length;
ibmr->length = pi_mr->ibmr.length;
- ibmr->iova = pi_mr->ibmr.iova;
- ibmr->sig_attrs->meta_length = pi_mr->meta_length;
- ib_dma_sync_single_for_device(ibmr->device, pi_mr->desc_map,
- pi_mr->desc_size * pi_mr->max_descs,
- DMA_TO_DEVICE);
+ return n;
+}
+int mlx5_ib_map_mr_sg_pi(struct ib_mr *ibmr, struct scatterlist *data_sg,
+ int data_sg_nents, unsigned int *data_sg_offset,
+ struct scatterlist *meta_sg, int meta_sg_nents,
+ unsigned int *meta_sg_offset)
+{
+ struct mlx5_ib_mr *mr = to_mmr(ibmr);
+ struct mlx5_ib_mr *pi_mr = mr->mtt_mr;
+ int n;
+
+ WARN_ON(ibmr->type != IB_MR_TYPE_INTEGRITY);
+
+ /*
+ * As a performance optimization, if possible, there is no need to map
+ * the sg lists to KLM descriptors. First try to map the sg lists to MTT
+ * descriptors and fallback to KLM only in case of a failure.
+ * It's more efficient for the HW to work with MTT descriptors
+ * (especially in high load).
+ * Use KLM (indirect access) only if it's mandatory.
+ */
+ n = mlx5_ib_map_mtt_mr_sg_pi(ibmr, data_sg, data_sg_nents,
+ data_sg_offset, meta_sg, meta_sg_nents,
+ meta_sg_offset);
+ if (n == data_sg_nents + meta_sg_nents)
+ goto out;
+
+ pi_mr = mr->klm_mr;
+ n = mlx5_ib_map_klm_mr_sg_pi(ibmr, data_sg, data_sg_nents,
+ data_sg_offset, meta_sg, meta_sg_nents,
+ meta_sg_offset);
if (unlikely(n != data_sg_nents + meta_sg_nents))
return -ENOMEM;
+out:
+ /* This is zero-based memory region */
+ ibmr->iova = 0;
+ mr->pi_mr = pi_mr;
+ ibmr->sig_attrs->meta_length = pi_mr->meta_length;
+
return 0;
}
projects / openwrt / staging / blogic.git / commitdiff