On 11/19/19 12:31 PM, Jason Gunthorpe wrote:
On Mon, Nov 18, 2019 at 11:54:38AM -0800, rao Shoaib wrote:
From: Rao Shoaib <rao.shoaib@xxxxxxxxxx>
Introduce maximum WQE size to impose limits on max SGE's and inline data
Signed-off-by: Rao Shoaib <rao.shoaib@xxxxxxxxxx>
drivers/infiniband/sw/rxe/rxe_param.h | 3 ++-
drivers/infiniband/sw/rxe/rxe_qp.c | 7 +++++--
2 files changed, 7 insertions(+), 3 deletions(-)
diff --git a/drivers/infiniband/sw/rxe/rxe_param.h b/drivers/infiniband/sw/rxe/rxe_param.h
index 1b596fb..31fb5c7 100644
+++ b/drivers/infiniband/sw/rxe/rxe_param.h
@@ -68,7 +68,6 @@ enum rxe_device_param {
RXE_HW_VER = 0,
RXE_MAX_QP = 0x10000,
RXE_MAX_QP_WR = 0x4000,
- RXE_MAX_INLINE_DATA = 400,
RXE_DEVICE_CAP_FLAGS = IB_DEVICE_BAD_PKEY_CNTR
| IB_DEVICE_BAD_QKEY_CNTR
| IB_DEVICE_AUTO_PATH_MIG
@@ -79,7 +78,9 @@ enum rxe_device_param {
| IB_DEVICE_RC_RNR_NAK_GEN
| IB_DEVICE_SRQ_RESIZE
| IB_DEVICE_MEM_MGT_EXTENSIONS,
+ RXE_MAX_WQE_SIZE = 0x2d0, /* For RXE_MAX_SGE */
This shouldn't just be a random constant, I think you are trying to
say:
RXE_MAX_WQE_SIZE = sizeof(struct rxe_send_wqe) + sizeof(struct ib_sge)*RXE_MAX_SGE
I thought you wanted this value to be independent of RXE_MAX_SGE, else
why are defining it.
Just say that
RXE_MAX_SGE = 32,
+ RXE_MAX_INLINE_DATA = RXE_MAX_WQE_SIZE,
This is mixed up now, it should be
RXE_MAX_INLINE_DATA = RXE_MAX_WQE_SIZE - sizeof(rxe_send_wqe)
I agree to what you are suggesting, it will make the current patch
better. However, In my previous patch I had
RXE_MAX_INLINE_DATA = RXE_MAX_SGE * sizeof(struct ib_sge)
IMHO that conveys the intent much better. I do not see the reason for
defining RXE_MAX_WQE_SIZE, ib_device_attr does not even have an entry
for it and hence the value is not used anywhere by rxe or by any other
relevant driver.
I will re-submit with the changes you have suggested.
Shoaib
diff --git a/drivers/infiniband/sw/rxe/rxe_qp.c b/drivers/infiniband/sw/rxe/rxe_qp.c
index aeea994..323e43d 100644
+++ b/drivers/infiniband/sw/rxe/rxe_qp.c
@@ -78,9 +78,12 @@ static int rxe_qp_chk_cap(struct rxe_dev *rxe, struct ib_qp_cap *cap,
}
}
- if (cap->max_inline_data > rxe->max_inline_data) {
+ if (cap->max_inline_data >
+ rxe->max_inline_data - sizeof(struct rxe_send_wqe)) {
pr_warn("invalid max inline data = %d > %d\n",
- cap->max_inline_data, rxe->max_inline_data);
+ cap->max_inline_data,
+ rxe->max_inline_data -
+ (u32)sizeof(struct rxe_send_wqe));
Then this isn't needed
And this code in the other patch:
+ wqe_size = max_t(int, init->cap.max_send_sge * sizeof(struct ib_sge),
+ init->cap.max_inline_data);
+ qp->sq.max_sge = wqe_size/sizeof(struct ib_sge);
+ qp->sq.max_inline = wqe_size;
Makes sense as both max_inline_data and 'init->cap.max_send_sge *
sizeof(struct ib_sge)' are exclusive of the wqe header
Jason
