int mthca_alloc_srq_buf(struct ibv_pd *pd, struct ibv_srq_attr *attr,
struct mthca_srq *srq)
{
struct mthca_data_seg *scatter;
void *wqe;
int size;
int i;
srq->wrid = malloc(srq->max * sizeof (uint64_t));
if (!srq->wrid)
return -1;
size = sizeof (struct mthca_next_seg) +
srq->max_gs * sizeof (struct mthca_data_seg);
for (srq->wqe_shift = 6; 1 << srq->wqe_shift < size; ++srq->wqe_shift)
; /* nothing */
srq->buf_size = srq->max << srq->wqe_shift;
if (mthca_alloc_buf(&srq->buf,
align(srq->buf_size, to_mdev(pd->context->device)->page_size),
to_mdev(pd->context->device)->page_size)) {
free(srq->wrid);
return -1;
}
memset(srq->buf.buf, 0, srq->buf_size);
/*
* Now initialize the SRQ buffer so that all of the WQEs are
* linked into the list of free WQEs. In addition, set the
* scatter list L_Keys to the sentry value of 0x100.
*/
for (i = 0; i < srq->max; ++i) {
struct mthca_next_seg *next;
next = wqe = get_wqe(srq, i);
if (i < srq->max - 1) {
*wqe_to_link(wqe) = i + 1;
next->nda_op = htonl(((i + 1) << srq->wqe_shift) | 1);
} else {
*wqe_to_link(wqe) = -1;
next->nda_op = 0;
}
for (scatter = wqe + sizeof (struct mthca_next_seg);
(void *) scatter < wqe + (1 << srq->wqe_shift);
++scatter)
scatter->lkey = htonl(MTHCA_INVAL_LKEY);
}
srq->first_free = 0;
srq->last_free = srq->max - 1;
srq->last = get_wqe(srq, srq->max - 1);
return 0;
}
void mthca_free_srq_wqe(struct mthca_srq *srq, int ind)
{
struct mthca_next_seg *last_free;
pthread_spin_lock(&srq->lock);
last_free = get_wqe(srq, srq->last_free);
*wqe_to_link(last_free) = ind;
last_free->nda_op = htonl((ind << srq->wqe_shift) | 1);
*wqe_to_link(get_wqe(srq, ind)) = -1;
srq->last_free = ind;
pthread_spin_unlock(&srq->lock);
}
int mlx5_ib_post_srq_recv(struct ib_srq *ibsrq, struct ib_recv_wr *wr,
struct ib_recv_wr **bad_wr)
{
struct mlx5_ib_srq *srq = to_msrq(ibsrq);
struct mlx5_wqe_srq_next_seg *next;
struct mlx5_wqe_data_seg *scat;
unsigned long flags;
int err = 0;
int nreq;
int i;
spin_lock_irqsave(&srq->lock, flags);
for (nreq = 0; wr; nreq++, wr = wr->next) {
if (unlikely(wr->num_sge > srq->msrq.max_gs)) {
err = -EINVAL;
*bad_wr = wr;
break;
}
if (unlikely(srq->head == srq->tail)) {
err = -ENOMEM;
*bad_wr = wr;
break;
}
srq->wrid[srq->head] = wr->wr_id;
next = get_wqe(srq, srq->head);
srq->head = be16_to_cpu(next->next_wqe_index);
scat = (struct mlx5_wqe_data_seg *)(next + 1);
for (i = 0; i < wr->num_sge; i++) {
scat[i].byte_count = cpu_to_be32(wr->sg_list[i].length);
scat[i].lkey = cpu_to_be32(wr->sg_list[i].lkey);
scat[i].addr = cpu_to_be64(wr->sg_list[i].addr);
}
if (i < srq->msrq.max_avail_gather) {
scat[i].byte_count = 0;
scat[i].lkey = cpu_to_be32(MLX5_INVALID_LKEY);
scat[i].addr = 0;
}
}
if (likely(nreq)) {
srq->wqe_ctr += nreq;
/* Make sure that descriptors are written before
* doorbell record.
*/
wmb();
*srq->db.db = cpu_to_be32(srq->wqe_ctr);
}
spin_unlock_irqrestore(&srq->lock, flags);
return err;
}
int mlx5_alloc_srq_buf(struct ibv_context *context, struct mlx5_srq *srq)
{
struct mlx5_wqe_srq_next_seg *next;
int size;
int buf_size;
int i;
struct mlx5_context *ctx;
ctx = to_mctx(context);
if (srq->max_gs < 0) {
errno = EINVAL;
return -1;
}
srq->wrid = malloc(srq->max * sizeof *srq->wrid);
if (!srq->wrid)
return -1;
size = sizeof(struct mlx5_wqe_srq_next_seg) +
srq->max_gs * sizeof(struct mlx5_wqe_data_seg);
size = max(32, size);
size = mlx5_round_up_power_of_two(size);
if (size > ctx->max_recv_wr) {
errno = EINVAL;
return -1;
}
srq->max_gs = (size - sizeof(struct mlx5_wqe_srq_next_seg)) /
sizeof(struct mlx5_wqe_data_seg);
srq->wqe_shift = mlx5_ilog2(size);
buf_size = srq->max * size;
if (mlx5_alloc_buf(&srq->buf, buf_size,
to_mdev(context->device)->page_size)) {
free(srq->wrid);
return -1;
}
memset(srq->buf.buf, 0, buf_size);
/*
* Now initialize the SRQ buffer so that all of the WQEs are
* linked into the list of free WQEs.
*/
for (i = 0; i < srq->max; ++i) {
next = get_wqe(srq, i);
next->next_wqe_index = htobe16((i + 1) & (srq->max - 1));
}
srq->head = 0;
srq->tail = srq->max - 1;
return 0;
}
void mlx4_free_srq_wqe(struct mlx4_srq *srq, int ind)
{
struct mlx4_wqe_srq_next_seg *next;
pthread_spin_lock(&srq->lock);
next = get_wqe(srq, srq->tail);
next->next_wqe_index = htobe16(ind);
srq->tail = ind;
pthread_spin_unlock(&srq->lock);
}
void mlx4_ib_free_srq_wqe(struct mlx4_ib_srq *srq, int wqe_index)
{
struct mlx4_wqe_srq_next_seg *next;
spin_lock(&srq->lock);
next = get_wqe(srq, srq->tail);
next->next_wqe_index = cpu_to_be16(wqe_index);
srq->tail = wqe_index;
spin_unlock(&srq->lock);
}
void mlx4_ib_free_srq_wqe(struct mlx4_ib_srq *srq, int wqe_index)
{
struct mlx4_wqe_srq_next_seg *next;
/* always called with interrupts disabled. */
spin_lock(&srq->lock);
next = get_wqe(srq, srq->tail);
next->next_wqe_index = cpu_to_be16(wqe_index);
srq->tail = wqe_index;
spin_unlock(&srq->lock);
}
int mlx4_alloc_srq_buf(struct ibv_pd *pd, struct ibv_srq_attr *attr,
struct mlx4_srq *srq)
{
struct mlx4_wqe_srq_next_seg *next;
struct mlx4_wqe_data_seg *scatter;
int size;
int buf_size;
int i;
srq->wrid = malloc(srq->max * sizeof (uint64_t));
if (!srq->wrid)
return -1;
size = sizeof (struct mlx4_wqe_srq_next_seg) +
srq->max_gs * sizeof (struct mlx4_wqe_data_seg);
for (srq->wqe_shift = 5; 1 << srq->wqe_shift < size; ++srq->wqe_shift)
; /* nothing */
buf_size = srq->max << srq->wqe_shift;
if (mlx4_alloc_buf(&srq->buf, buf_size,
to_mdev(pd->context->device)->page_size)) {
free(srq->wrid);
return -1;
}
memset(srq->buf.buf, 0, buf_size);
/*
* Now initialize the SRQ buffer so that all of the WQEs are
* linked into the list of free WQEs.
*/
for (i = 0; i < srq->max; ++i) {
next = get_wqe(srq, i);
next->next_wqe_index = htobe16((i + 1) & (srq->max - 1));
for (scatter = (void *) (next + 1);
(void *) scatter < (void *) next + (1 << srq->wqe_shift);
++scatter)
scatter->lkey = htobe32(MLX4_INVALID_LKEY);
}
srq->head = 0;
srq->tail = srq->max - 1;
return 0;
}
int mlx5_copy_to_recv_srq(struct mlx5_srq *srq, int idx, void *buf, int size)
{
struct mlx5_wqe_srq_next_seg *next;
struct mlx5_wqe_data_seg *scat;
int copy;
int i;
int max = 1 << (srq->wqe_shift - 4);
next = get_wqe(srq, idx);
scat = (struct mlx5_wqe_data_seg *) (next + 1);
for (i = 0; i < max; ++i) {
copy = min_t(long, size, be32toh(scat->byte_count));
memcpy((void *)(unsigned long)be64toh(scat->addr), buf, copy);
size -= copy;
if (size <= 0)
return IBV_WC_SUCCESS;
buf += copy;
++scat;
}
return IBV_WC_LOC_LEN_ERR;
}
struct ib_srq *mlx4_ib_create_srq(struct ib_pd *pd,
struct ib_srq_init_attr *init_attr,
struct ib_udata *udata)
{
struct mlx4_ib_dev *dev = to_mdev(pd->device);
struct mlx4_ib_srq *srq;
struct mlx4_wqe_srq_next_seg *next;
int desc_size;
int buf_size;
int err;
int i;
if (init_attr->attr.max_wr >= dev->dev->caps.max_srq_wqes ||
init_attr->attr.max_sge > dev->dev->caps.max_srq_sge)
return ERR_PTR(-EINVAL);
srq = kmalloc(sizeof *srq, GFP_KERNEL);
if (!srq)
return ERR_PTR(-ENOMEM);
mutex_init(&srq->mutex);
spin_lock_init(&srq->lock);
srq->msrq.max = roundup_pow_of_two(init_attr->attr.max_wr + 1);
srq->msrq.max_gs = init_attr->attr.max_sge;
desc_size = max(32UL,
roundup_pow_of_two(sizeof (struct mlx4_wqe_srq_next_seg) +
srq->msrq.max_gs *
sizeof (struct mlx4_wqe_data_seg)));
srq->msrq.wqe_shift = ilog2(desc_size);
buf_size = srq->msrq.max * desc_size;
if (pd->uobject) {
struct mlx4_ib_create_srq ucmd;
if (ib_copy_from_udata(&ucmd, udata, sizeof ucmd)) {
err = -EFAULT;
goto err_srq;
}
srq->umem = ib_umem_get(pd->uobject->context, ucmd.buf_addr,
buf_size, 0, 0);
if (IS_ERR(srq->umem)) {
err = PTR_ERR(srq->umem);
goto err_srq;
}
err = mlx4_mtt_init(dev->dev, ib_umem_page_count(srq->umem),
ilog2(srq->umem->page_size), &srq->mtt);
if (err)
goto err_buf;
err = mlx4_ib_umem_write_mtt(dev, &srq->mtt, srq->umem);
if (err)
goto err_mtt;
err = mlx4_ib_db_map_user(to_mucontext(pd->uobject->context),
ucmd.db_addr, &srq->db);
if (err)
goto err_mtt;
} else {
err = mlx4_db_alloc(dev->dev, &srq->db, 0);
if (err)
goto err_srq;
*srq->db.db = 0;
if (mlx4_buf_alloc(dev->dev, buf_size, PAGE_SIZE * 2, &srq->buf)) {
err = -ENOMEM;
goto err_db;
}
srq->head = 0;
srq->tail = srq->msrq.max - 1;
srq->wqe_ctr = 0;
for (i = 0; i < srq->msrq.max; ++i) {
next = get_wqe(srq, i);
next->next_wqe_index =
cpu_to_be16((i + 1) & (srq->msrq.max - 1));
}
err = mlx4_mtt_init(dev->dev, srq->buf.npages, srq->buf.page_shift,
&srq->mtt);
if (err)
goto err_buf;
err = mlx4_buf_write_mtt(dev->dev, &srq->mtt, &srq->buf);
if (err)
goto err_mtt;
srq->wrid = kmalloc(srq->msrq.max * sizeof (u64), GFP_KERNEL);
if (!srq->wrid) {
err = -ENOMEM;
goto err_mtt;
}
}
err = mlx4_srq_alloc(dev->dev, to_mpd(pd)->pdn, &srq->mtt,
srq->db.dma, &srq->msrq);
if (err)
goto err_wrid;
srq->msrq.event = mlx4_ib_srq_event;
if (pd->uobject)
if (ib_copy_to_udata(udata, &srq->msrq.srqn, sizeof (__u32))) {
err = -EFAULT;
goto err_wrid;
}
init_attr->attr.max_wr = srq->msrq.max - 1;
return &srq->ibsrq;
err_wrid:
if (pd->uobject)
mlx4_ib_db_unmap_user(to_mucontext(pd->uobject->context), &srq->db);
else
kfree(srq->wrid);
err_mtt:
mlx4_mtt_cleanup(dev->dev, &srq->mtt);
err_buf:
if (pd->uobject)
ib_umem_release(srq->umem);
else
mlx4_buf_free(dev->dev, buf_size, &srq->buf);
err_db:
if (!pd->uobject)
mlx4_db_free(dev->dev, &srq->db);
err_srq:
kfree(srq);
return ERR_PTR(err);
}
static int create_srq_kernel(struct mlx5_ib_dev *dev, struct mlx5_ib_srq *srq,
struct mlx5_srq_attr *in, int buf_size)
{
int err;
int i;
struct mlx5_wqe_srq_next_seg *next;
int page_shift;
int npages;
err = mlx5_db_alloc(dev->mdev, &srq->db);
if (err) {
mlx5_ib_warn(dev, "alloc dbell rec failed\n");
return err;
}
if (mlx5_buf_alloc(dev->mdev, buf_size, &srq->buf)) {
mlx5_ib_dbg(dev, "buf alloc failed\n");
err = -ENOMEM;
goto err_db;
}
page_shift = srq->buf.page_shift;
srq->head = 0;
srq->tail = srq->msrq.max - 1;
srq->wqe_ctr = 0;
for (i = 0; i < srq->msrq.max; i++) {
next = get_wqe(srq, i);
next->next_wqe_index =
cpu_to_be16((i + 1) & (srq->msrq.max - 1));
}
npages = DIV_ROUND_UP(srq->buf.npages, 1 << (page_shift - PAGE_SHIFT));
mlx5_ib_dbg(dev, "buf_size %d, page_shift %d, npages %d, calc npages %d\n",
buf_size, page_shift, srq->buf.npages, npages);
in->pas = mlx5_vzalloc(sizeof(*in->pas) * npages);
if (!in->pas) {
err = -ENOMEM;
goto err_buf;
}
mlx5_fill_page_array(&srq->buf, in->pas);
srq->wrid = kmalloc(srq->msrq.max * sizeof(u64), GFP_KERNEL);
if (!srq->wrid) {
mlx5_ib_dbg(dev, "kmalloc failed %lu\n",
(unsigned long)(srq->msrq.max * sizeof(u64)));
err = -ENOMEM;
goto err_in;
}
srq->wq_sig = !!srq_signature;
in->log_page_size = page_shift - MLX5_ADAPTER_PAGE_SHIFT;
if (MLX5_CAP_GEN(dev->mdev, cqe_version) == MLX5_CQE_VERSION_V1 &&
in->type == IB_SRQT_XRC)
in->user_index = MLX5_IB_DEFAULT_UIDX;
return 0;
err_in:
kvfree(in->pas);
err_buf:
mlx5_buf_free(dev->mdev, &srq->buf);
err_db:
mlx5_db_free(dev->mdev, &srq->db);
return err;
}
int mthca_tavor_post_srq_recv(struct ibv_srq *ibsrq,
struct ibv_recv_wr *wr,
struct ibv_recv_wr **bad_wr)
{
struct mthca_srq *srq = to_msrq(ibsrq);
uint32_t doorbell[2];
int err = 0;
int first_ind;
int ind;
int next_ind;
int nreq;
int i;
void *wqe;
void *prev_wqe;
pthread_spin_lock(&srq->lock);
first_ind = srq->first_free;
for (nreq = 0; wr; wr = wr->next) {
ind = srq->first_free;
wqe = get_wqe(srq, ind);
next_ind = *wqe_to_link(wqe);
if (next_ind < 0) {
err = -1;
*bad_wr = wr;
break;
}
prev_wqe = srq->last;
srq->last = wqe;
((struct mthca_next_seg *) wqe)->ee_nds = 0;
/* flags field will always remain 0 */
wqe += sizeof (struct mthca_next_seg);
if (wr->num_sge > srq->max_gs) {
err = -1;
*bad_wr = wr;
srq->last = prev_wqe;
break;
}
for (i = 0; i < wr->num_sge; ++i) {
((struct mthca_data_seg *) wqe)->byte_count =
htonl(wr->sg_list[i].length);
((struct mthca_data_seg *) wqe)->lkey =
htonl(wr->sg_list[i].lkey);
((struct mthca_data_seg *) wqe)->addr =
htonll(wr->sg_list[i].addr);
wqe += sizeof (struct mthca_data_seg);
}
if (i < srq->max_gs) {
((struct mthca_data_seg *) wqe)->byte_count = 0;
((struct mthca_data_seg *) wqe)->lkey = htonl(MTHCA_INVAL_LKEY);
((struct mthca_data_seg *) wqe)->addr = 0;
}
((struct mthca_next_seg *) prev_wqe)->ee_nds =
htonl(MTHCA_NEXT_DBD);
srq->wrid[ind] = wr->wr_id;
srq->first_free = next_ind;
if (++nreq == MTHCA_TAVOR_MAX_WQES_PER_RECV_DB) {
nreq = 0;
doorbell[0] = htonl(first_ind << srq->wqe_shift);
doorbell[1] = htonl(srq->srqn << 8);
/*
* Make sure that descriptors are written
* before doorbell is rung.
*/
wmb();
mthca_write64(doorbell, to_mctx(ibsrq->context), MTHCA_RECV_DOORBELL);
first_ind = srq->first_free;
}
}
if (nreq) {
doorbell[0] = htonl(first_ind << srq->wqe_shift);
doorbell[1] = htonl((srq->srqn << 8) | nreq);
/*
* Make sure that descriptors are written before
* doorbell is rung.
*/
wmb();
mthca_write64(doorbell, to_mctx(ibsrq->context), MTHCA_RECV_DOORBELL);
}
pthread_spin_unlock(&srq->lock);
return err;
}
int mlx4_ib_post_srq_recv(struct ib_srq *ibsrq, struct ib_recv_wr *wr,
struct ib_recv_wr **bad_wr)
{
struct mlx4_ib_srq *srq = to_msrq(ibsrq);
struct mlx4_wqe_srq_next_seg *next;
struct mlx4_wqe_data_seg *scat;
unsigned long flags;
int err = 0;
int nreq;
int i;
struct mlx4_ib_dev *mdev = to_mdev(ibsrq->device);
spin_lock_irqsave(&srq->lock, flags);
if (mdev->dev->persist->state & MLX4_DEVICE_STATE_INTERNAL_ERROR) {
err = -EIO;
*bad_wr = wr;
nreq = 0;
goto out;
}
for (nreq = 0; wr; ++nreq, wr = wr->next) {
if (unlikely(wr->num_sge > srq->msrq.max_gs)) {
err = -EINVAL;
*bad_wr = wr;
break;
}
if (unlikely(srq->head == srq->tail)) {
err = -ENOMEM;
*bad_wr = wr;
break;
}
srq->wrid[srq->head] = wr->wr_id;
next = get_wqe(srq, srq->head);
srq->head = be16_to_cpu(next->next_wqe_index);
scat = (struct mlx4_wqe_data_seg *) (next + 1);
for (i = 0; i < wr->num_sge; ++i) {
scat[i].byte_count = cpu_to_be32(wr->sg_list[i].length);
scat[i].lkey = cpu_to_be32(wr->sg_list[i].lkey);
scat[i].addr = cpu_to_be64(wr->sg_list[i].addr);
}
if (i < srq->msrq.max_gs) {
scat[i].byte_count = 0;
scat[i].lkey = cpu_to_be32(MLX4_INVALID_LKEY);
scat[i].addr = 0;
}
}
if (likely(nreq)) {
srq->wqe_ctr += nreq;
/*
* Make sure that descriptors are written before
* doorbell record.
*/
wmb();
*srq->db.db = cpu_to_be32(srq->wqe_ctr);
}
out:
spin_unlock_irqrestore(&srq->lock, flags);
return err;
}
int mthca_arbel_post_srq_recv(struct ibv_srq *ibsrq,
struct ibv_recv_wr *wr,
struct ibv_recv_wr **bad_wr)
{
struct mthca_srq *srq = to_msrq(ibsrq);
int err = 0;
int ind;
int next_ind;
int nreq;
int i;
void *wqe;
pthread_spin_lock(&srq->lock);
for (nreq = 0; wr; ++nreq, wr = wr->next) {
ind = srq->first_free;
wqe = get_wqe(srq, ind);
next_ind = *wqe_to_link(wqe);
if (next_ind < 0) {
err = -1;
*bad_wr = wr;
break;
}
((struct mthca_next_seg *) wqe)->ee_nds = 0;
/* flags field will always remain 0 */
wqe += sizeof (struct mthca_next_seg);
if (wr->num_sge > srq->max_gs) {
err = -1;
*bad_wr = wr;
break;
}
for (i = 0; i < wr->num_sge; ++i) {
((struct mthca_data_seg *) wqe)->byte_count =
htonl(wr->sg_list[i].length);
((struct mthca_data_seg *) wqe)->lkey =
htonl(wr->sg_list[i].lkey);
((struct mthca_data_seg *) wqe)->addr =
htonll(wr->sg_list[i].addr);
wqe += sizeof (struct mthca_data_seg);
}
if (i < srq->max_gs) {
((struct mthca_data_seg *) wqe)->byte_count = 0;
((struct mthca_data_seg *) wqe)->lkey = htonl(MTHCA_INVAL_LKEY);
((struct mthca_data_seg *) wqe)->addr = 0;
}
srq->wrid[ind] = wr->wr_id;
srq->first_free = next_ind;
}
if (nreq) {
srq->counter += nreq;
/*
* Make sure that descriptors are written before
* we write doorbell record.
*/
wmb();
*srq->db = htonl(srq->counter);
}
pthread_spin_unlock(&srq->lock);
return err;
}
int mlx4_post_srq_recv(struct ibv_srq *ibsrq,
struct ibv_recv_wr *wr,
struct ibv_recv_wr **bad_wr)
{
struct mlx4_srq *srq = to_msrq(ibsrq);
struct mlx4_wqe_srq_next_seg *next;
struct mlx4_wqe_data_seg *scat;
int err = 0;
int nreq;
int i;
pthread_spin_lock(&srq->lock);
for (nreq = 0; wr; ++nreq, wr = wr->next) {
if (wr->num_sge > srq->max_gs) {
err = -1;
*bad_wr = wr;
break;
}
if (srq->head == srq->tail) {
/* SRQ is full*/
err = -1;
*bad_wr = wr;
break;
}
srq->wrid[srq->head] = wr->wr_id;
next = get_wqe(srq, srq->head);
srq->head = be16toh(next->next_wqe_index);
scat = (struct mlx4_wqe_data_seg *) (next + 1);
for (i = 0; i < wr->num_sge; ++i) {
scat[i].byte_count = htobe32(wr->sg_list[i].length);
scat[i].lkey = htobe32(wr->sg_list[i].lkey);
scat[i].addr = htobe64(wr->sg_list[i].addr);
}
if (i < srq->max_gs) {
scat[i].byte_count = 0;
scat[i].lkey = htobe32(MLX4_INVALID_LKEY);
scat[i].addr = 0;
}
}
if (nreq) {
srq->counter += nreq;
/*
* Make sure that descriptors are written before
* we write doorbell record.
*/
udma_to_device_barrier();
*srq->db = htobe32(srq->counter);
}
pthread_spin_unlock(&srq->lock);
return err;
}
int rxe_completer(void *arg)
{
struct rxe_qp *qp = (struct rxe_qp *)arg;
struct rxe_dev *rxe = to_rdev(qp->ibqp.device);
struct rxe_send_wqe *wqe = wqe;
struct sk_buff *skb = NULL;
struct rxe_pkt_info *pkt = NULL;
enum comp_state state;
rxe_add_ref(qp);
if (!qp->valid || qp->req.state == QP_STATE_ERROR ||
qp->req.state == QP_STATE_RESET) {
rxe_drain_resp_pkts(qp, qp->valid &&
qp->req.state == QP_STATE_ERROR);
goto exit;
}
if (qp->comp.timeout) {
qp->comp.timeout_retry = 1;
qp->comp.timeout = 0;
} else {
qp->comp.timeout_retry = 0;
}
if (qp->req.need_retry)
goto exit;
state = COMPST_GET_ACK;
while (1) {
pr_debug("qp#%d state = %s\n", qp_num(qp),
comp_state_name[state]);
switch (state) {
case COMPST_GET_ACK:
skb = skb_dequeue(&qp->resp_pkts);
if (skb) {
pkt = SKB_TO_PKT(skb);
qp->comp.timeout_retry = 0;
}
state = COMPST_GET_WQE;
break;
case COMPST_GET_WQE:
state = get_wqe(qp, pkt, &wqe);
break;
case COMPST_CHECK_PSN:
state = check_psn(qp, pkt, wqe);
break;
case COMPST_CHECK_ACK:
state = check_ack(qp, pkt, wqe);
break;
case COMPST_READ:
state = do_read(qp, pkt, wqe);
break;
case COMPST_ATOMIC:
state = do_atomic(qp, pkt, wqe);
break;
case COMPST_WRITE_SEND:
if (wqe->state == wqe_state_pending &&
wqe->last_psn == pkt->psn)
state = COMPST_COMP_ACK;
else
state = COMPST_UPDATE_COMP;
break;
case COMPST_COMP_ACK:
state = complete_ack(qp, pkt, wqe);
break;
case COMPST_COMP_WQE:
state = complete_wqe(qp, pkt, wqe);
break;
case COMPST_UPDATE_COMP:
if (pkt->mask & RXE_END_MASK)
qp->comp.opcode = -1;
else
qp->comp.opcode = pkt->opcode;
if (psn_compare(pkt->psn, qp->comp.psn) >= 0)
qp->comp.psn = (pkt->psn + 1) & BTH_PSN_MASK;
if (qp->req.wait_psn) {
qp->req.wait_psn = 0;
rxe_run_task(&qp->req.task, 1);
}
state = COMPST_DONE;
break;
case COMPST_DONE:
if (pkt) {
rxe_drop_ref(pkt->qp);
kfree_skb(skb);
skb = NULL;
}
goto done;
case COMPST_EXIT:
if (qp->comp.timeout_retry && wqe) {
state = COMPST_ERROR_RETRY;
break;
}
/* re reset the timeout counter if
* (1) QP is type RC
* (2) the QP is alive
* (3) there is a packet sent by the requester that
* might be acked (we still might get spurious
* timeouts but try to keep them as few as possible)
* (4) the timeout parameter is set
*/
if ((qp_type(qp) == IB_QPT_RC) &&
(qp->req.state == QP_STATE_READY) &&
(psn_compare(qp->req.psn, qp->comp.psn) > 0) &&
qp->qp_timeout_jiffies)
mod_timer(&qp->retrans_timer,
jiffies + qp->qp_timeout_jiffies);
WARN_ON_ONCE(skb);
goto exit;
case COMPST_ERROR_RETRY:
/* we come here if the retry timer fired and we did
* not receive a response packet. try to retry the send
* queue if that makes sense and the limits have not
* been exceeded. remember that some timeouts are
* spurious since we do not reset the timer but kick
* it down the road or let it expire
*/
/* there is nothing to retry in this case */
if (!wqe || (wqe->state == wqe_state_posted)) {
WARN_ON_ONCE(skb);
goto exit;
}
if (qp->comp.retry_cnt > 0) {
if (qp->comp.retry_cnt != 7)
qp->comp.retry_cnt--;
/* no point in retrying if we have already
* seen the last ack that the requester could
* have caused
*/
if (psn_compare(qp->req.psn,
qp->comp.psn) > 0) {
/* tell the requester to retry the
* send queue next time around
*/
rxe_counter_inc(rxe,
RXE_CNT_COMP_RETRY);
qp->req.need_retry = 1;
rxe_run_task(&qp->req.task, 1);
}
if (pkt) {
rxe_drop_ref(pkt->qp);
kfree_skb(skb);
skb = NULL;
}
WARN_ON_ONCE(skb);
goto exit;
} else {
rxe_counter_inc(rxe, RXE_CNT_RETRY_EXCEEDED);
wqe->status = IB_WC_RETRY_EXC_ERR;
state = COMPST_ERROR;
}
break;
case COMPST_RNR_RETRY:
if (qp->comp.rnr_retry > 0) {
if (qp->comp.rnr_retry != 7)
qp->comp.rnr_retry--;
qp->req.need_retry = 1;
pr_debug("qp#%d set rnr nak timer\n",
qp_num(qp));
mod_timer(&qp->rnr_nak_timer,
jiffies + rnrnak_jiffies(aeth_syn(pkt)
& ~AETH_TYPE_MASK));
rxe_drop_ref(pkt->qp);
kfree_skb(skb);
skb = NULL;
goto exit;
} else {
rxe_counter_inc(rxe,
RXE_CNT_RNR_RETRY_EXCEEDED);
wqe->status = IB_WC_RNR_RETRY_EXC_ERR;
state = COMPST_ERROR;
}
break;
case COMPST_ERROR:
WARN_ON_ONCE(wqe->status == IB_WC_SUCCESS);
do_complete(qp, wqe);
rxe_qp_error(qp);
if (pkt) {
rxe_drop_ref(pkt->qp);
kfree_skb(skb);
skb = NULL;
}
WARN_ON_ONCE(skb);
goto exit;
}
}
exit:
/* we come here if we are done with processing and want the task to
* exit from the loop calling us
*/
WARN_ON_ONCE(skb);
rxe_drop_ref(qp);
return -EAGAIN;
done:
/* we come here if we have processed a packet we want the task to call
* us again to see if there is anything else to do
*/
WARN_ON_ONCE(skb);
rxe_drop_ref(qp);
return 0;
}
static int create_srq_kernel(struct mlx5_ib_dev *dev, struct mlx5_ib_srq *srq,
struct mlx5_create_srq_mbox_in **in, int buf_size,
int *inlen)
{
int err;
int i;
struct mlx5_wqe_srq_next_seg *next;
int page_shift;
int npages;
err = mlx5_db_alloc(dev->mdev, &srq->db);
if (err) {
mlx5_ib_warn(dev, "alloc dbell rec failed\n");
return err;
}
*srq->db.db = 0;
if (mlx5_buf_alloc(dev->mdev, buf_size, PAGE_SIZE * 2, &srq->buf)) {
mlx5_ib_dbg(dev, "buf alloc failed\n");
err = -ENOMEM;
goto err_db;
}
page_shift = srq->buf.page_shift;
srq->head = 0;
srq->tail = srq->msrq.max - 1;
srq->wqe_ctr = 0;
for (i = 0; i < srq->msrq.max; i++) {
next = get_wqe(srq, i);
next->next_wqe_index =
cpu_to_be16((i + 1) & (srq->msrq.max - 1));
}
npages = DIV_ROUND_UP(srq->buf.npages, 1 << (page_shift - PAGE_SHIFT));
mlx5_ib_dbg(dev, "buf_size %d, page_shift %d, npages %d, calc npages %d\n",
buf_size, page_shift, srq->buf.npages, npages);
*inlen = sizeof(**in) + sizeof(*(*in)->pas) * npages;
*in = mlx5_vzalloc(*inlen);
if (!*in) {
err = -ENOMEM;
goto err_buf;
}
mlx5_fill_page_array(&srq->buf, (*in)->pas);
srq->wrid = kmalloc(srq->msrq.max * sizeof(u64), GFP_KERNEL);
if (!srq->wrid) {
mlx5_ib_dbg(dev, "kmalloc failed %lu\n",
(unsigned long)(srq->msrq.max * sizeof(u64)));
err = -ENOMEM;
goto err_in;
}
srq->wq_sig = !!srq_signature;
(*in)->ctx.log_pg_sz = page_shift - MLX5_ADAPTER_PAGE_SHIFT;
return 0;
err_in:
mlx5_vfree(*in);
err_buf:
mlx5_buf_free(dev->mdev, &srq->buf);
err_db:
mlx5_db_free(dev->mdev, &srq->db);
return err;
}
struct ib_srq *mlx4_ib_create_srq(struct ib_pd *pd,
struct ib_srq_init_attr *init_attr,
struct ib_udata *udata)
{
struct mlx4_ib_dev *dev = to_mdev(pd->device);
struct mlx4_ib_srq *srq;
struct mlx4_wqe_srq_next_seg *next;
struct mlx4_wqe_data_seg *scatter;
u32 cqn;
u16 xrcdn;
int desc_size;
int buf_size;
int err;
int i;
/* Sanity check SRQ size before proceeding */
if (init_attr->attr.max_wr >= dev->dev->caps.max_srq_wqes ||
init_attr->attr.max_sge > dev->dev->caps.max_srq_sge)
return ERR_PTR(-EINVAL);
srq = kmalloc(sizeof *srq, GFP_KERNEL);
if (!srq)
return ERR_PTR(-ENOMEM);
mutex_init(&srq->mutex);
spin_lock_init(&srq->lock);
srq->msrq.max = roundup_pow_of_two(init_attr->attr.max_wr + 1);
srq->msrq.max_gs = init_attr->attr.max_sge;
desc_size = max(32UL,
roundup_pow_of_two(sizeof (struct mlx4_wqe_srq_next_seg) +
srq->msrq.max_gs *
sizeof (struct mlx4_wqe_data_seg)));
srq->msrq.wqe_shift = ilog2(desc_size);
buf_size = srq->msrq.max * desc_size;
if (pd->uobject) {
struct mlx4_ib_create_srq ucmd;
if (ib_copy_from_udata(&ucmd, udata, sizeof ucmd)) {
err = -EFAULT;
goto err_srq;
}
srq->umem = ib_umem_get(pd->uobject->context, ucmd.buf_addr,
buf_size, 0, 0);
if (IS_ERR(srq->umem)) {
err = PTR_ERR(srq->umem);
goto err_srq;
}
err = mlx4_mtt_init(dev->dev, ib_umem_page_count(srq->umem),
ilog2(srq->umem->page_size), &srq->mtt);
if (err)
goto err_buf;
err = mlx4_ib_umem_write_mtt(dev, &srq->mtt, srq->umem);
if (err)
goto err_mtt;
err = mlx4_ib_db_map_user(to_mucontext(pd->uobject->context),
ucmd.db_addr, &srq->db);
if (err)
goto err_mtt;
} else {
err = mlx4_db_alloc(dev->dev, &srq->db, 0, GFP_KERNEL);
if (err)
goto err_srq;
*srq->db.db = 0;
if (mlx4_buf_alloc(dev->dev, buf_size, PAGE_SIZE * 2, &srq->buf,
GFP_KERNEL)) {
err = -ENOMEM;
goto err_db;
}
srq->head = 0;
srq->tail = srq->msrq.max - 1;
srq->wqe_ctr = 0;
for (i = 0; i < srq->msrq.max; ++i) {
next = get_wqe(srq, i);
next->next_wqe_index =
cpu_to_be16((i + 1) & (srq->msrq.max - 1));
for (scatter = (void *) (next + 1);
(void *) scatter < (void *) next + desc_size;
++scatter)
scatter->lkey = cpu_to_be32(MLX4_INVALID_LKEY);
}
err = mlx4_mtt_init(dev->dev, srq->buf.npages, srq->buf.page_shift,
&srq->mtt);
if (err)
goto err_buf;
err = mlx4_buf_write_mtt(dev->dev, &srq->mtt, &srq->buf, GFP_KERNEL);
if (err)
goto err_mtt;
srq->wrid = kmalloc(srq->msrq.max * sizeof (u64), GFP_KERNEL);
if (!srq->wrid) {
err = -ENOMEM;
goto err_mtt;
}
}
cqn = (init_attr->srq_type == IB_SRQT_XRC) ?
to_mcq(init_attr->ext.xrc.cq)->mcq.cqn : 0;
xrcdn = (init_attr->srq_type == IB_SRQT_XRC) ?
to_mxrcd(init_attr->ext.xrc.xrcd)->xrcdn :
(u16) dev->dev->caps.reserved_xrcds;
err = mlx4_srq_alloc(dev->dev, to_mpd(pd)->pdn, cqn, xrcdn, &srq->mtt,
srq->db.dma, &srq->msrq);
if (err)
goto err_wrid;
srq->msrq.event = mlx4_ib_srq_event;
srq->ibsrq.ext.xrc.srq_num = srq->msrq.srqn;
if (pd->uobject)
if (ib_copy_to_udata(udata, &srq->msrq.srqn, sizeof (__u32))) {
err = -EFAULT;
goto err_wrid;
}
init_attr->attr.max_wr = srq->msrq.max - 1;
return &srq->ibsrq;
err_wrid:
if (pd->uobject)
mlx4_ib_db_unmap_user(to_mucontext(pd->uobject->context), &srq->db);
else
kfree(srq->wrid);
err_mtt:
mlx4_mtt_cleanup(dev->dev, &srq->mtt);
err_buf:
if (pd->uobject)
ib_umem_release(srq->umem);
else
mlx4_buf_free(dev->dev, buf_size, &srq->buf);
err_db:
if (!pd->uobject)
mlx4_db_free(dev->dev, &srq->db);
err_srq:
kfree(srq);
return ERR_PTR(err);
}
int mlx4_en_activate_rx_rings(struct mlx4_en_priv *priv)
{
struct mlx4_en_dev *mdev = priv->mdev;
struct mlx4_wqe_srq_next_seg *next;
struct mlx4_en_rx_ring *ring;
int i;
int ring_ind;
int err;
int stride = roundup_pow_of_two(sizeof(struct mlx4_en_rx_desc) +
DS_SIZE * priv->num_frags);
int max_gs = (stride - sizeof(struct mlx4_wqe_srq_next_seg)) / DS_SIZE;
for (ring_ind = 0; ring_ind < priv->rx_ring_num; ring_ind++) {
ring = &priv->rx_ring[ring_ind];
ring->prod = 0;
ring->cons = 0;
ring->actual_size = 0;
ring->cqn = priv->rx_cq[ring_ind].mcq.cqn;
ring->stride = stride;
ring->log_stride = ffs(ring->stride) - 1;
ring->buf_size = ring->size * ring->stride;
memset(ring->rx_info, 0, sizeof(*(ring->rx_info)));
memset(ring->buf, 0, ring->buf_size);
mlx4_en_update_rx_prod_db(ring);
/* Initailize all descriptors */
for (i = 0; i < ring->size; i++)
mlx4_en_init_rx_desc(priv, ring, i);
/* Initialize page allocators */
err = mlx4_en_init_allocator(priv, ring);
if (err) {
mlx4_err(mdev, "Failed initializing ring allocator\n");
goto err_allocator;
}
/* Fill Rx buffers */
ring->full = 0;
}
if (mlx4_en_fill_rx_buffers(priv))
goto err_buffers;
for (ring_ind = 0; ring_ind < priv->rx_ring_num; ring_ind++) {
ring = &priv->rx_ring[ring_ind];
mlx4_en_update_rx_prod_db(ring);
/* Configure SRQ representing the ring */
ring->srq.max = ring->size;
ring->srq.max_gs = max_gs;
ring->srq.wqe_shift = ilog2(ring->stride);
for (i = 0; i < ring->srq.max; ++i) {
next = get_wqe(ring, i);
next->next_wqe_index =
cpu_to_be16((i + 1) & (ring->srq.max - 1));
}
err = mlx4_srq_alloc(mdev->dev, mdev->priv_pdn,
&ring->wqres.mtt, ring->wqres.db.dma,
&ring->srq);
if (err){
mlx4_err(mdev, "Failed to allocate srq\n");
goto err_srq;
}
ring->srq.event = mlx4_en_srq_event;
}
return 0;
err_srq:
while (ring_ind >= 0) {
ring = &priv->rx_ring[ring_ind];
mlx4_srq_free(mdev->dev, &ring->srq);
ring_ind--;
}
err_buffers:
for (ring_ind = 0; ring_ind < priv->rx_ring_num; ring_ind++)
mlx4_en_free_rx_buf(priv, &priv->rx_ring[ring_ind]);
ring_ind = priv->rx_ring_num - 1;
err_allocator:
while (ring_ind >= 0) {
mlx4_en_destroy_allocator(priv, &priv->rx_ring[ring_ind]);
ring_ind--;
}
return err;
}
