2010-04-22 06:30:06 +08:00
|
|
|
/*
|
|
|
|
* Copyright (c) 2009-2010 Chelsio, Inc. All rights reserved.
|
|
|
|
*
|
|
|
|
* This software is available to you under a choice of one of two
|
|
|
|
* licenses. You may choose to be licensed under the terms of the GNU
|
|
|
|
* General Public License (GPL) Version 2, available from the file
|
|
|
|
* COPYING in the main directory of this source tree, or the
|
|
|
|
* OpenIB.org BSD license below:
|
|
|
|
*
|
|
|
|
* Redistribution and use in source and binary forms, with or
|
|
|
|
* without modification, are permitted provided that the following
|
|
|
|
* conditions are met:
|
|
|
|
*
|
|
|
|
* - Redistributions of source code must retain the above
|
|
|
|
* copyright notice, this list of conditions and the following
|
|
|
|
* disclaimer.
|
|
|
|
* - Redistributions in binary form must reproduce the above
|
|
|
|
* copyright notice, this list of conditions and the following
|
|
|
|
* disclaimer in the documentation and/or other materials
|
|
|
|
* provided with the distribution.
|
|
|
|
*
|
|
|
|
* THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
|
|
|
|
* EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
|
|
|
|
* MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
|
|
|
|
* NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
|
|
|
|
* BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
|
|
|
|
* ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
|
|
|
|
* CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
|
|
|
|
* SOFTWARE.
|
|
|
|
*/
|
|
|
|
#ifndef __T4_H__
|
|
|
|
#define __T4_H__
|
|
|
|
|
|
|
|
#include "t4_hw.h"
|
|
|
|
#include "t4_regs.h"
|
2015-06-09 20:53:12 +08:00
|
|
|
#include "t4_values.h"
|
2010-04-22 06:30:06 +08:00
|
|
|
#include "t4_msg.h"
|
|
|
|
#include "t4fw_ri_api.h"
|
|
|
|
|
2013-08-06 23:34:35 +08:00
|
|
|
#define T4_MAX_NUM_PD 65536
|
2013-08-06 23:34:39 +08:00
|
|
|
#define T4_MAX_MR_SIZE (~0ULL)
|
2010-04-22 06:30:06 +08:00
|
|
|
#define T4_PAGESIZE_MASK 0xffff000 /* 4KB-128MB */
|
|
|
|
#define T4_STAG_UNSET 0xffffffff
|
|
|
|
#define T4_FW_MAJ 0
|
2015-01-16 11:54:47 +08:00
|
|
|
#define PCIE_MA_SYNC_A 0x30b4
|
2010-04-22 06:30:06 +08:00
|
|
|
|
|
|
|
struct t4_status_page {
|
|
|
|
__be32 rsvd1; /* flit 0 - hw owns */
|
|
|
|
__be16 rsvd2;
|
|
|
|
__be16 qid;
|
|
|
|
__be16 cidx;
|
|
|
|
__be16 pidx;
|
|
|
|
u8 qp_err; /* flit 1 - sw owns */
|
|
|
|
u8 db_off;
|
2012-05-18 17:59:30 +08:00
|
|
|
u8 pad;
|
|
|
|
u16 host_wq_pidx;
|
|
|
|
u16 host_cidx;
|
|
|
|
u16 host_pidx;
|
2010-04-22 06:30:06 +08:00
|
|
|
};
|
|
|
|
|
2010-06-11 03:03:00 +08:00
|
|
|
#define T4_EQ_ENTRY_SIZE 64
|
2010-04-22 06:30:06 +08:00
|
|
|
|
2010-09-18 04:40:15 +08:00
|
|
|
#define T4_SQ_NUM_SLOTS 5
|
2010-06-11 03:03:00 +08:00
|
|
|
#define T4_SQ_NUM_BYTES (T4_EQ_ENTRY_SIZE * T4_SQ_NUM_SLOTS)
|
2010-04-22 06:30:06 +08:00
|
|
|
#define T4_MAX_SEND_SGE ((T4_SQ_NUM_BYTES - sizeof(struct fw_ri_send_wr) - \
|
|
|
|
sizeof(struct fw_ri_isgl)) / sizeof(struct fw_ri_sge))
|
|
|
|
#define T4_MAX_SEND_INLINE ((T4_SQ_NUM_BYTES - sizeof(struct fw_ri_send_wr) - \
|
|
|
|
sizeof(struct fw_ri_immd)))
|
|
|
|
#define T4_MAX_WRITE_INLINE ((T4_SQ_NUM_BYTES - \
|
|
|
|
sizeof(struct fw_ri_rdma_write_wr) - \
|
|
|
|
sizeof(struct fw_ri_immd)))
|
|
|
|
#define T4_MAX_WRITE_SGE ((T4_SQ_NUM_BYTES - \
|
|
|
|
sizeof(struct fw_ri_rdma_write_wr) - \
|
|
|
|
sizeof(struct fw_ri_isgl)) / sizeof(struct fw_ri_sge))
|
|
|
|
#define T4_MAX_FR_IMMD ((T4_SQ_NUM_BYTES - sizeof(struct fw_ri_fr_nsmr_wr) - \
|
2010-09-18 04:40:15 +08:00
|
|
|
sizeof(struct fw_ri_immd)) & ~31UL)
|
2014-04-09 22:38:27 +08:00
|
|
|
#define T4_MAX_FR_IMMD_DEPTH (T4_MAX_FR_IMMD / sizeof(u64))
|
|
|
|
#define T4_MAX_FR_DSGL 1024
|
|
|
|
#define T4_MAX_FR_DSGL_DEPTH (T4_MAX_FR_DSGL / sizeof(u64))
|
|
|
|
|
|
|
|
static inline int t4_max_fr_depth(int use_dsgl)
|
|
|
|
{
|
|
|
|
return use_dsgl ? T4_MAX_FR_DSGL_DEPTH : T4_MAX_FR_IMMD_DEPTH;
|
|
|
|
}
|
2010-04-22 06:30:06 +08:00
|
|
|
|
|
|
|
#define T4_RQ_NUM_SLOTS 2
|
2010-06-11 03:03:00 +08:00
|
|
|
#define T4_RQ_NUM_BYTES (T4_EQ_ENTRY_SIZE * T4_RQ_NUM_SLOTS)
|
2010-05-21 05:58:05 +08:00
|
|
|
#define T4_MAX_RECV_SGE 4
|
2010-04-22 06:30:06 +08:00
|
|
|
|
|
|
|
union t4_wr {
|
|
|
|
struct fw_ri_res_wr res;
|
|
|
|
struct fw_ri_wr ri;
|
|
|
|
struct fw_ri_rdma_write_wr write;
|
|
|
|
struct fw_ri_send_wr send;
|
|
|
|
struct fw_ri_rdma_read_wr read;
|
|
|
|
struct fw_ri_bind_mw_wr bind;
|
|
|
|
struct fw_ri_fr_nsmr_wr fr;
|
2016-09-16 22:54:52 +08:00
|
|
|
struct fw_ri_fr_nsmr_tpte_wr fr_tpte;
|
2010-04-22 06:30:06 +08:00
|
|
|
struct fw_ri_inv_lstag_wr inv;
|
|
|
|
struct t4_status_page status;
|
2010-06-11 03:03:00 +08:00
|
|
|
__be64 flits[T4_EQ_ENTRY_SIZE / sizeof(__be64) * T4_SQ_NUM_SLOTS];
|
2010-04-22 06:30:06 +08:00
|
|
|
};
|
|
|
|
|
|
|
|
union t4_recv_wr {
|
|
|
|
struct fw_ri_recv_wr recv;
|
|
|
|
struct t4_status_page status;
|
2010-06-11 03:03:00 +08:00
|
|
|
__be64 flits[T4_EQ_ENTRY_SIZE / sizeof(__be64) * T4_RQ_NUM_SLOTS];
|
2010-04-22 06:30:06 +08:00
|
|
|
};
|
|
|
|
|
|
|
|
static inline void init_wr_hdr(union t4_wr *wqe, u16 wrid,
|
|
|
|
enum fw_wr_opcodes opcode, u8 flags, u8 len16)
|
|
|
|
{
|
|
|
|
wqe->send.opcode = (u8)opcode;
|
|
|
|
wqe->send.flags = flags;
|
|
|
|
wqe->send.wrid = wrid;
|
|
|
|
wqe->send.r1[0] = 0;
|
|
|
|
wqe->send.r1[1] = 0;
|
|
|
|
wqe->send.r1[2] = 0;
|
|
|
|
wqe->send.len16 = len16;
|
|
|
|
}
|
|
|
|
|
|
|
|
/* CQE/AE status codes */
|
|
|
|
#define T4_ERR_SUCCESS 0x0
|
|
|
|
#define T4_ERR_STAG 0x1 /* STAG invalid: either the */
|
|
|
|
/* STAG is offlimt, being 0, */
|
|
|
|
/* or STAG_key mismatch */
|
|
|
|
#define T4_ERR_PDID 0x2 /* PDID mismatch */
|
|
|
|
#define T4_ERR_QPID 0x3 /* QPID mismatch */
|
|
|
|
#define T4_ERR_ACCESS 0x4 /* Invalid access right */
|
|
|
|
#define T4_ERR_WRAP 0x5 /* Wrap error */
|
|
|
|
#define T4_ERR_BOUND 0x6 /* base and bounds voilation */
|
|
|
|
#define T4_ERR_INVALIDATE_SHARED_MR 0x7 /* attempt to invalidate a */
|
|
|
|
/* shared memory region */
|
|
|
|
#define T4_ERR_INVALIDATE_MR_WITH_MW_BOUND 0x8 /* attempt to invalidate a */
|
|
|
|
/* shared memory region */
|
|
|
|
#define T4_ERR_ECC 0x9 /* ECC error detected */
|
|
|
|
#define T4_ERR_ECC_PSTAG 0xA /* ECC error detected when */
|
|
|
|
/* reading PSTAG for a MW */
|
|
|
|
/* Invalidate */
|
|
|
|
#define T4_ERR_PBL_ADDR_BOUND 0xB /* pbl addr out of bounds: */
|
|
|
|
/* software error */
|
|
|
|
#define T4_ERR_SWFLUSH 0xC /* SW FLUSHED */
|
|
|
|
#define T4_ERR_CRC 0x10 /* CRC error */
|
|
|
|
#define T4_ERR_MARKER 0x11 /* Marker error */
|
|
|
|
#define T4_ERR_PDU_LEN_ERR 0x12 /* invalid PDU length */
|
|
|
|
#define T4_ERR_OUT_OF_RQE 0x13 /* out of RQE */
|
|
|
|
#define T4_ERR_DDP_VERSION 0x14 /* wrong DDP version */
|
|
|
|
#define T4_ERR_RDMA_VERSION 0x15 /* wrong RDMA version */
|
|
|
|
#define T4_ERR_OPCODE 0x16 /* invalid rdma opcode */
|
|
|
|
#define T4_ERR_DDP_QUEUE_NUM 0x17 /* invalid ddp queue number */
|
|
|
|
#define T4_ERR_MSN 0x18 /* MSN error */
|
|
|
|
#define T4_ERR_TBIT 0x19 /* tag bit not set correctly */
|
|
|
|
#define T4_ERR_MO 0x1A /* MO not 0 for TERMINATE */
|
|
|
|
/* or READ_REQ */
|
|
|
|
#define T4_ERR_MSN_GAP 0x1B
|
|
|
|
#define T4_ERR_MSN_RANGE 0x1C
|
|
|
|
#define T4_ERR_IRD_OVERFLOW 0x1D
|
|
|
|
#define T4_ERR_RQE_ADDR_BOUND 0x1E /* RQE addr out of bounds: */
|
|
|
|
/* software error */
|
|
|
|
#define T4_ERR_INTERNAL_ERR 0x1F /* internal error (opcode */
|
|
|
|
/* mismatch) */
|
|
|
|
/*
|
|
|
|
* CQE defs
|
|
|
|
*/
|
|
|
|
struct t4_cqe {
|
|
|
|
__be32 header;
|
|
|
|
__be32 len;
|
|
|
|
union {
|
|
|
|
struct {
|
|
|
|
__be32 stag;
|
|
|
|
__be32 msn;
|
|
|
|
} rcqe;
|
|
|
|
struct {
|
2016-09-16 22:54:52 +08:00
|
|
|
u32 stag;
|
2010-04-22 06:30:06 +08:00
|
|
|
u16 nada2;
|
|
|
|
u16 cidx;
|
|
|
|
} scqe;
|
|
|
|
struct {
|
|
|
|
__be32 wrid_hi;
|
|
|
|
__be32 wrid_low;
|
|
|
|
} gen;
|
iw_cxgb4: refactor sq/rq drain logic
With the addition of the IB/Core drain API, iw_cxgb4 supported drain
by watching the CQs when the QP was out of RTS and signalling "drain
complete" when the last CQE is polled. This, however, doesn't fully
support the drain semantics. Namely, the drain logic is supposed to signal
"drain complete" only when the application has _processed_ the last CQE,
not just removed them from the CQ. Thus a small timing hole exists that
can cause touch after free type bugs in applications using the drain API
(nvmf, iSER, for example). So iw_cxgb4 needs a better solution.
The iWARP Verbs spec mandates that "_at some point_ after the QP is
moved to ERROR", the iWARP driver MUST synchronously fail post_send and
post_recv calls. iw_cxgb4 was currently not allowing any posts once the
QP is in ERROR. This was in part due to the fact that the HW queues for
the QP in ERROR state are disabled at this point, so there wasn't much
else to do but fail the post operation synchronously. This restriction
is what drove the first drain implementation in iw_cxgb4 that has the
above mentioned flaw.
This patch changes iw_cxgb4 to allow post_send and post_recv WRs after
the QP is moved to ERROR state for kernel mode users, thus still adhering
to the Verbs spec for user mode users, but allowing flush WRs for kernel
users. Since the HW queues are disabled, we just synthesize a CQE for
this post, queue it to the SW CQ, and then call the CQ event handler.
This enables proper drain operations for the various storage applications.
Signed-off-by: Steve Wise <swise@opengridcomputing.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>
2016-12-22 23:04:59 +08:00
|
|
|
u64 drain_cookie;
|
2010-04-22 06:30:06 +08:00
|
|
|
} u;
|
|
|
|
__be64 reserved;
|
|
|
|
__be64 bits_type_ts;
|
|
|
|
};
|
|
|
|
|
|
|
|
/* macros for flit 0 of the cqe */
|
|
|
|
|
2015-01-16 11:54:47 +08:00
|
|
|
#define CQE_QPID_S 12
|
|
|
|
#define CQE_QPID_M 0xFFFFF
|
|
|
|
#define CQE_QPID_G(x) ((((x) >> CQE_QPID_S)) & CQE_QPID_M)
|
|
|
|
#define CQE_QPID_V(x) ((x)<<CQE_QPID_S)
|
2010-04-22 06:30:06 +08:00
|
|
|
|
2015-01-16 11:54:47 +08:00
|
|
|
#define CQE_SWCQE_S 11
|
|
|
|
#define CQE_SWCQE_M 0x1
|
|
|
|
#define CQE_SWCQE_G(x) ((((x) >> CQE_SWCQE_S)) & CQE_SWCQE_M)
|
|
|
|
#define CQE_SWCQE_V(x) ((x)<<CQE_SWCQE_S)
|
2010-04-22 06:30:06 +08:00
|
|
|
|
2015-01-16 11:54:47 +08:00
|
|
|
#define CQE_STATUS_S 5
|
|
|
|
#define CQE_STATUS_M 0x1F
|
|
|
|
#define CQE_STATUS_G(x) ((((x) >> CQE_STATUS_S)) & CQE_STATUS_M)
|
|
|
|
#define CQE_STATUS_V(x) ((x)<<CQE_STATUS_S)
|
2010-04-22 06:30:06 +08:00
|
|
|
|
2015-01-16 11:54:47 +08:00
|
|
|
#define CQE_TYPE_S 4
|
|
|
|
#define CQE_TYPE_M 0x1
|
|
|
|
#define CQE_TYPE_G(x) ((((x) >> CQE_TYPE_S)) & CQE_TYPE_M)
|
|
|
|
#define CQE_TYPE_V(x) ((x)<<CQE_TYPE_S)
|
2010-04-22 06:30:06 +08:00
|
|
|
|
2015-01-16 11:54:47 +08:00
|
|
|
#define CQE_OPCODE_S 0
|
|
|
|
#define CQE_OPCODE_M 0xF
|
|
|
|
#define CQE_OPCODE_G(x) ((((x) >> CQE_OPCODE_S)) & CQE_OPCODE_M)
|
|
|
|
#define CQE_OPCODE_V(x) ((x)<<CQE_OPCODE_S)
|
2010-04-22 06:30:06 +08:00
|
|
|
|
2015-01-16 11:54:47 +08:00
|
|
|
#define SW_CQE(x) (CQE_SWCQE_G(be32_to_cpu((x)->header)))
|
|
|
|
#define CQE_QPID(x) (CQE_QPID_G(be32_to_cpu((x)->header)))
|
|
|
|
#define CQE_TYPE(x) (CQE_TYPE_G(be32_to_cpu((x)->header)))
|
2010-04-22 06:30:06 +08:00
|
|
|
#define SQ_TYPE(x) (CQE_TYPE((x)))
|
|
|
|
#define RQ_TYPE(x) (!CQE_TYPE((x)))
|
2015-01-16 11:54:47 +08:00
|
|
|
#define CQE_STATUS(x) (CQE_STATUS_G(be32_to_cpu((x)->header)))
|
|
|
|
#define CQE_OPCODE(x) (CQE_OPCODE_G(be32_to_cpu((x)->header)))
|
2010-04-22 06:30:06 +08:00
|
|
|
|
|
|
|
#define CQE_SEND_OPCODE(x)( \
|
2015-01-16 11:54:47 +08:00
|
|
|
(CQE_OPCODE_G(be32_to_cpu((x)->header)) == FW_RI_SEND) || \
|
|
|
|
(CQE_OPCODE_G(be32_to_cpu((x)->header)) == FW_RI_SEND_WITH_SE) || \
|
|
|
|
(CQE_OPCODE_G(be32_to_cpu((x)->header)) == FW_RI_SEND_WITH_INV) || \
|
|
|
|
(CQE_OPCODE_G(be32_to_cpu((x)->header)) == FW_RI_SEND_WITH_SE_INV))
|
2010-04-22 06:30:06 +08:00
|
|
|
|
|
|
|
#define CQE_LEN(x) (be32_to_cpu((x)->len))
|
|
|
|
|
|
|
|
/* used for RQ completion processing */
|
|
|
|
#define CQE_WRID_STAG(x) (be32_to_cpu((x)->u.rcqe.stag))
|
|
|
|
#define CQE_WRID_MSN(x) (be32_to_cpu((x)->u.rcqe.msn))
|
|
|
|
|
|
|
|
/* used for SQ completion processing */
|
|
|
|
#define CQE_WRID_SQ_IDX(x) ((x)->u.scqe.cidx)
|
2016-09-16 22:54:52 +08:00
|
|
|
#define CQE_WRID_FR_STAG(x) (be32_to_cpu((x)->u.scqe.stag))
|
2010-04-22 06:30:06 +08:00
|
|
|
|
|
|
|
/* generic accessor macros */
|
2014-07-15 00:04:53 +08:00
|
|
|
#define CQE_WRID_HI(x) (be32_to_cpu((x)->u.gen.wrid_hi))
|
|
|
|
#define CQE_WRID_LOW(x) (be32_to_cpu((x)->u.gen.wrid_low))
|
iw_cxgb4: refactor sq/rq drain logic
With the addition of the IB/Core drain API, iw_cxgb4 supported drain
by watching the CQs when the QP was out of RTS and signalling "drain
complete" when the last CQE is polled. This, however, doesn't fully
support the drain semantics. Namely, the drain logic is supposed to signal
"drain complete" only when the application has _processed_ the last CQE,
not just removed them from the CQ. Thus a small timing hole exists that
can cause touch after free type bugs in applications using the drain API
(nvmf, iSER, for example). So iw_cxgb4 needs a better solution.
The iWARP Verbs spec mandates that "_at some point_ after the QP is
moved to ERROR", the iWARP driver MUST synchronously fail post_send and
post_recv calls. iw_cxgb4 was currently not allowing any posts once the
QP is in ERROR. This was in part due to the fact that the HW queues for
the QP in ERROR state are disabled at this point, so there wasn't much
else to do but fail the post operation synchronously. This restriction
is what drove the first drain implementation in iw_cxgb4 that has the
above mentioned flaw.
This patch changes iw_cxgb4 to allow post_send and post_recv WRs after
the QP is moved to ERROR state for kernel mode users, thus still adhering
to the Verbs spec for user mode users, but allowing flush WRs for kernel
users. Since the HW queues are disabled, we just synthesize a CQE for
this post, queue it to the SW CQ, and then call the CQ event handler.
This enables proper drain operations for the various storage applications.
Signed-off-by: Steve Wise <swise@opengridcomputing.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>
2016-12-22 23:04:59 +08:00
|
|
|
#define CQE_DRAIN_COOKIE(x) ((x)->u.drain_cookie)
|
2010-04-22 06:30:06 +08:00
|
|
|
|
|
|
|
/* macros for flit 3 of the cqe */
|
2015-01-16 11:54:47 +08:00
|
|
|
#define CQE_GENBIT_S 63
|
|
|
|
#define CQE_GENBIT_M 0x1
|
|
|
|
#define CQE_GENBIT_G(x) (((x) >> CQE_GENBIT_S) & CQE_GENBIT_M)
|
|
|
|
#define CQE_GENBIT_V(x) ((x)<<CQE_GENBIT_S)
|
2010-04-22 06:30:06 +08:00
|
|
|
|
2015-01-16 11:54:47 +08:00
|
|
|
#define CQE_OVFBIT_S 62
|
|
|
|
#define CQE_OVFBIT_M 0x1
|
|
|
|
#define CQE_OVFBIT_G(x) ((((x) >> CQE_OVFBIT_S)) & CQE_OVFBIT_M)
|
2010-04-22 06:30:06 +08:00
|
|
|
|
2015-01-16 11:54:47 +08:00
|
|
|
#define CQE_IQTYPE_S 60
|
|
|
|
#define CQE_IQTYPE_M 0x3
|
|
|
|
#define CQE_IQTYPE_G(x) ((((x) >> CQE_IQTYPE_S)) & CQE_IQTYPE_M)
|
2010-04-22 06:30:06 +08:00
|
|
|
|
2015-01-16 11:54:47 +08:00
|
|
|
#define CQE_TS_M 0x0fffffffffffffffULL
|
|
|
|
#define CQE_TS_G(x) ((x) & CQE_TS_M)
|
2010-04-22 06:30:06 +08:00
|
|
|
|
2015-01-16 11:54:47 +08:00
|
|
|
#define CQE_OVFBIT(x) ((unsigned)CQE_OVFBIT_G(be64_to_cpu((x)->bits_type_ts)))
|
|
|
|
#define CQE_GENBIT(x) ((unsigned)CQE_GENBIT_G(be64_to_cpu((x)->bits_type_ts)))
|
|
|
|
#define CQE_TS(x) (CQE_TS_G(be64_to_cpu((x)->bits_type_ts)))
|
2010-04-22 06:30:06 +08:00
|
|
|
|
|
|
|
struct t4_swsqe {
|
|
|
|
u64 wr_id;
|
|
|
|
struct t4_cqe cqe;
|
|
|
|
int read_len;
|
|
|
|
int opcode;
|
|
|
|
int complete;
|
|
|
|
int signaled;
|
|
|
|
u16 idx;
|
2013-08-06 23:34:35 +08:00
|
|
|
int flushed;
|
2014-07-15 00:04:54 +08:00
|
|
|
struct timespec host_ts;
|
|
|
|
u64 sge_ts;
|
2010-04-22 06:30:06 +08:00
|
|
|
};
|
|
|
|
|
2010-09-14 00:23:57 +08:00
|
|
|
static inline pgprot_t t4_pgprot_wc(pgprot_t prot)
|
|
|
|
{
|
2011-03-14 18:36:11 +08:00
|
|
|
#if defined(__i386__) || defined(__x86_64__) || defined(CONFIG_PPC64)
|
2010-09-14 00:23:57 +08:00
|
|
|
return pgprot_writecombine(prot);
|
|
|
|
#else
|
|
|
|
return pgprot_noncached(prot);
|
|
|
|
#endif
|
|
|
|
}
|
|
|
|
|
|
|
|
enum {
|
|
|
|
T4_SQ_ONCHIP = (1<<0),
|
|
|
|
};
|
|
|
|
|
2010-04-22 06:30:06 +08:00
|
|
|
struct t4_sq {
|
|
|
|
union t4_wr *queue;
|
|
|
|
dma_addr_t dma_addr;
|
2010-06-03 13:37:50 +08:00
|
|
|
DEFINE_DMA_UNMAP_ADDR(mapping);
|
2010-09-14 00:23:57 +08:00
|
|
|
unsigned long phys_addr;
|
2010-04-22 06:30:06 +08:00
|
|
|
struct t4_swsqe *sw_sq;
|
|
|
|
struct t4_swsqe *oldest_read;
|
2015-06-09 20:53:12 +08:00
|
|
|
void __iomem *bar2_va;
|
|
|
|
u64 bar2_pa;
|
2010-04-22 06:30:06 +08:00
|
|
|
size_t memsize;
|
2015-06-09 20:53:12 +08:00
|
|
|
u32 bar2_qid;
|
2010-04-22 06:30:06 +08:00
|
|
|
u32 qid;
|
|
|
|
u16 in_use;
|
|
|
|
u16 size;
|
|
|
|
u16 cidx;
|
|
|
|
u16 pidx;
|
2010-06-11 03:03:00 +08:00
|
|
|
u16 wq_pidx;
|
cxgb4/iw_cxgb4: Doorbell Drop Avoidance Bug Fixes
The current logic suffers from a slow response time to disable user DB
usage, and also fails to avoid DB FIFO drops under heavy load. This commit
fixes these deficiencies and makes the avoidance logic more optimal.
This is done by more efficiently notifying the ULDs of potential DB
problems, and implements a smoother flow control algorithm in iw_cxgb4,
which is the ULD that puts the most load on the DB fifo.
Design:
cxgb4:
Direct ULD callback from the DB FULL/DROP interrupt handler. This allows
the ULD to stop doing user DB writes as quickly as possible.
While user DB usage is disabled, the LLD will accumulate DB write events
for its queues. Then once DB usage is reenabled, a single DB write is
done for each queue with its accumulated write count. This reduces the
load put on the DB fifo when reenabling.
iw_cxgb4:
Instead of marking each qp to indicate DB writes are disabled, we create
a device-global status page that each user process maps. This allows
iw_cxgb4 to only set this single bit to disable all DB writes for all
user QPs vs traversing the idr of all the active QPs. If the libcxgb4
doesn't support this, then we fall back to the old approach of marking
each QP. Thus we allow the new driver to work with an older libcxgb4.
When the LLD upcalls iw_cxgb4 indicating DB FULL, we disable all DB writes
via the status page and transition the DB state to STOPPED. As user
processes see that DB writes are disabled, they call into iw_cxgb4
to submit their DB write events. Since the DB state is in STOPPED,
the QP trying to write gets enqueued on a new DB "flow control" list.
As subsequent DB writes are submitted for this flow controlled QP, the
amount of writes are accumulated for each QP on the flow control list.
So all the user QPs that are actively ringing the DB get put on this
list and the number of writes they request are accumulated.
When the LLD upcalls iw_cxgb4 indicating DB EMPTY, which is in a workq
context, we change the DB state to FLOW_CONTROL, and begin resuming all
the QPs that are on the flow control list. This logic runs on until
the flow control list is empty or we exit FLOW_CONTROL mode (due to
a DB DROP upcall, for example). QPs are removed from this list, and
their accumulated DB write counts written to the DB FIFO. Sets of QPs,
called chunks in the code, are removed at one time. The chunk size is 64.
So 64 QPs are resumed at a time, and before the next chunk is resumed, the
logic waits (blocks) for the DB FIFO to drain. This prevents resuming to
quickly and overflowing the FIFO. Once the flow control list is empty,
the db state transitions back to NORMAL and user QPs are again allowed
to write directly to the user DB register.
The algorithm is designed such that if the DB write load is high enough,
then all the DB writes get submitted by the kernel using this flow
controlled approach to avoid DB drops. As the load lightens though, we
resume to normal DB writes directly by user applications.
Signed-off-by: Steve Wise <swise@opengridcomputing.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
2014-03-15 00:22:08 +08:00
|
|
|
u16 wq_pidx_inc;
|
2010-09-14 00:23:57 +08:00
|
|
|
u16 flags;
|
2013-08-06 23:34:35 +08:00
|
|
|
short flush_cidx;
|
2010-04-22 06:30:06 +08:00
|
|
|
};
|
|
|
|
|
|
|
|
struct t4_swrqe {
|
|
|
|
u64 wr_id;
|
2014-07-15 00:04:54 +08:00
|
|
|
struct timespec host_ts;
|
|
|
|
u64 sge_ts;
|
2010-04-22 06:30:06 +08:00
|
|
|
};
|
|
|
|
|
|
|
|
struct t4_rq {
|
|
|
|
union t4_recv_wr *queue;
|
|
|
|
dma_addr_t dma_addr;
|
2010-06-03 13:37:50 +08:00
|
|
|
DEFINE_DMA_UNMAP_ADDR(mapping);
|
2010-04-22 06:30:06 +08:00
|
|
|
struct t4_swrqe *sw_rq;
|
2015-06-09 20:53:12 +08:00
|
|
|
void __iomem *bar2_va;
|
|
|
|
u64 bar2_pa;
|
2010-04-22 06:30:06 +08:00
|
|
|
size_t memsize;
|
2015-06-09 20:53:12 +08:00
|
|
|
u32 bar2_qid;
|
2010-04-22 06:30:06 +08:00
|
|
|
u32 qid;
|
|
|
|
u32 msn;
|
|
|
|
u32 rqt_hwaddr;
|
|
|
|
u16 rqt_size;
|
|
|
|
u16 in_use;
|
|
|
|
u16 size;
|
|
|
|
u16 cidx;
|
|
|
|
u16 pidx;
|
2010-06-11 03:03:00 +08:00
|
|
|
u16 wq_pidx;
|
cxgb4/iw_cxgb4: Doorbell Drop Avoidance Bug Fixes
The current logic suffers from a slow response time to disable user DB
usage, and also fails to avoid DB FIFO drops under heavy load. This commit
fixes these deficiencies and makes the avoidance logic more optimal.
This is done by more efficiently notifying the ULDs of potential DB
problems, and implements a smoother flow control algorithm in iw_cxgb4,
which is the ULD that puts the most load on the DB fifo.
Design:
cxgb4:
Direct ULD callback from the DB FULL/DROP interrupt handler. This allows
the ULD to stop doing user DB writes as quickly as possible.
While user DB usage is disabled, the LLD will accumulate DB write events
for its queues. Then once DB usage is reenabled, a single DB write is
done for each queue with its accumulated write count. This reduces the
load put on the DB fifo when reenabling.
iw_cxgb4:
Instead of marking each qp to indicate DB writes are disabled, we create
a device-global status page that each user process maps. This allows
iw_cxgb4 to only set this single bit to disable all DB writes for all
user QPs vs traversing the idr of all the active QPs. If the libcxgb4
doesn't support this, then we fall back to the old approach of marking
each QP. Thus we allow the new driver to work with an older libcxgb4.
When the LLD upcalls iw_cxgb4 indicating DB FULL, we disable all DB writes
via the status page and transition the DB state to STOPPED. As user
processes see that DB writes are disabled, they call into iw_cxgb4
to submit their DB write events. Since the DB state is in STOPPED,
the QP trying to write gets enqueued on a new DB "flow control" list.
As subsequent DB writes are submitted for this flow controlled QP, the
amount of writes are accumulated for each QP on the flow control list.
So all the user QPs that are actively ringing the DB get put on this
list and the number of writes they request are accumulated.
When the LLD upcalls iw_cxgb4 indicating DB EMPTY, which is in a workq
context, we change the DB state to FLOW_CONTROL, and begin resuming all
the QPs that are on the flow control list. This logic runs on until
the flow control list is empty or we exit FLOW_CONTROL mode (due to
a DB DROP upcall, for example). QPs are removed from this list, and
their accumulated DB write counts written to the DB FIFO. Sets of QPs,
called chunks in the code, are removed at one time. The chunk size is 64.
So 64 QPs are resumed at a time, and before the next chunk is resumed, the
logic waits (blocks) for the DB FIFO to drain. This prevents resuming to
quickly and overflowing the FIFO. Once the flow control list is empty,
the db state transitions back to NORMAL and user QPs are again allowed
to write directly to the user DB register.
The algorithm is designed such that if the DB write load is high enough,
then all the DB writes get submitted by the kernel using this flow
controlled approach to avoid DB drops. As the load lightens though, we
resume to normal DB writes directly by user applications.
Signed-off-by: Steve Wise <swise@opengridcomputing.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
2014-03-15 00:22:08 +08:00
|
|
|
u16 wq_pidx_inc;
|
2010-04-22 06:30:06 +08:00
|
|
|
};
|
|
|
|
|
|
|
|
struct t4_wq {
|
|
|
|
struct t4_sq sq;
|
|
|
|
struct t4_rq rq;
|
|
|
|
void __iomem *db;
|
|
|
|
struct c4iw_rdev *rdev;
|
2013-08-06 23:34:35 +08:00
|
|
|
int flushed;
|
2010-04-22 06:30:06 +08:00
|
|
|
};
|
|
|
|
|
|
|
|
static inline int t4_rqes_posted(struct t4_wq *wq)
|
|
|
|
{
|
|
|
|
return wq->rq.in_use;
|
|
|
|
}
|
|
|
|
|
|
|
|
static inline int t4_rq_empty(struct t4_wq *wq)
|
|
|
|
{
|
|
|
|
return wq->rq.in_use == 0;
|
|
|
|
}
|
|
|
|
|
|
|
|
static inline int t4_rq_full(struct t4_wq *wq)
|
|
|
|
{
|
|
|
|
return wq->rq.in_use == (wq->rq.size - 1);
|
|
|
|
}
|
|
|
|
|
|
|
|
static inline u32 t4_rq_avail(struct t4_wq *wq)
|
|
|
|
{
|
|
|
|
return wq->rq.size - 1 - wq->rq.in_use;
|
|
|
|
}
|
|
|
|
|
2010-06-11 03:03:00 +08:00
|
|
|
static inline void t4_rq_produce(struct t4_wq *wq, u8 len16)
|
2010-04-22 06:30:06 +08:00
|
|
|
{
|
|
|
|
wq->rq.in_use++;
|
|
|
|
if (++wq->rq.pidx == wq->rq.size)
|
|
|
|
wq->rq.pidx = 0;
|
2010-06-11 03:03:00 +08:00
|
|
|
wq->rq.wq_pidx += DIV_ROUND_UP(len16*16, T4_EQ_ENTRY_SIZE);
|
|
|
|
if (wq->rq.wq_pidx >= wq->rq.size * T4_RQ_NUM_SLOTS)
|
|
|
|
wq->rq.wq_pidx %= wq->rq.size * T4_RQ_NUM_SLOTS;
|
2010-04-22 06:30:06 +08:00
|
|
|
}
|
|
|
|
|
|
|
|
static inline void t4_rq_consume(struct t4_wq *wq)
|
|
|
|
{
|
|
|
|
wq->rq.in_use--;
|
|
|
|
wq->rq.msn++;
|
|
|
|
if (++wq->rq.cidx == wq->rq.size)
|
|
|
|
wq->rq.cidx = 0;
|
|
|
|
}
|
|
|
|
|
2012-05-18 17:59:30 +08:00
|
|
|
static inline u16 t4_rq_host_wq_pidx(struct t4_wq *wq)
|
|
|
|
{
|
|
|
|
return wq->rq.queue[wq->rq.size].status.host_wq_pidx;
|
|
|
|
}
|
|
|
|
|
|
|
|
static inline u16 t4_rq_wq_size(struct t4_wq *wq)
|
|
|
|
{
|
|
|
|
return wq->rq.size * T4_RQ_NUM_SLOTS;
|
|
|
|
}
|
|
|
|
|
2010-09-14 00:23:57 +08:00
|
|
|
static inline int t4_sq_onchip(struct t4_sq *sq)
|
|
|
|
{
|
|
|
|
return sq->flags & T4_SQ_ONCHIP;
|
|
|
|
}
|
|
|
|
|
2010-04-22 06:30:06 +08:00
|
|
|
static inline int t4_sq_empty(struct t4_wq *wq)
|
|
|
|
{
|
|
|
|
return wq->sq.in_use == 0;
|
|
|
|
}
|
|
|
|
|
|
|
|
static inline int t4_sq_full(struct t4_wq *wq)
|
|
|
|
{
|
|
|
|
return wq->sq.in_use == (wq->sq.size - 1);
|
|
|
|
}
|
|
|
|
|
|
|
|
static inline u32 t4_sq_avail(struct t4_wq *wq)
|
|
|
|
{
|
|
|
|
return wq->sq.size - 1 - wq->sq.in_use;
|
|
|
|
}
|
|
|
|
|
2010-06-11 03:03:00 +08:00
|
|
|
static inline void t4_sq_produce(struct t4_wq *wq, u8 len16)
|
2010-04-22 06:30:06 +08:00
|
|
|
{
|
|
|
|
wq->sq.in_use++;
|
|
|
|
if (++wq->sq.pidx == wq->sq.size)
|
|
|
|
wq->sq.pidx = 0;
|
2010-06-11 03:03:00 +08:00
|
|
|
wq->sq.wq_pidx += DIV_ROUND_UP(len16*16, T4_EQ_ENTRY_SIZE);
|
|
|
|
if (wq->sq.wq_pidx >= wq->sq.size * T4_SQ_NUM_SLOTS)
|
|
|
|
wq->sq.wq_pidx %= wq->sq.size * T4_SQ_NUM_SLOTS;
|
2010-04-22 06:30:06 +08:00
|
|
|
}
|
|
|
|
|
|
|
|
static inline void t4_sq_consume(struct t4_wq *wq)
|
|
|
|
{
|
2013-08-06 23:34:35 +08:00
|
|
|
BUG_ON(wq->sq.in_use < 1);
|
|
|
|
if (wq->sq.cidx == wq->sq.flush_cidx)
|
|
|
|
wq->sq.flush_cidx = -1;
|
2010-04-22 06:30:06 +08:00
|
|
|
wq->sq.in_use--;
|
|
|
|
if (++wq->sq.cidx == wq->sq.size)
|
|
|
|
wq->sq.cidx = 0;
|
|
|
|
}
|
|
|
|
|
2012-05-18 17:59:30 +08:00
|
|
|
static inline u16 t4_sq_host_wq_pidx(struct t4_wq *wq)
|
|
|
|
{
|
|
|
|
return wq->sq.queue[wq->sq.size].status.host_wq_pidx;
|
|
|
|
}
|
|
|
|
|
|
|
|
static inline u16 t4_sq_wq_size(struct t4_wq *wq)
|
|
|
|
{
|
|
|
|
return wq->sq.size * T4_SQ_NUM_SLOTS;
|
|
|
|
}
|
|
|
|
|
2014-04-09 22:38:25 +08:00
|
|
|
/* This function copies 64 byte coalesced work request to memory
|
|
|
|
* mapped BAR2 space. For coalesced WRs, the SGE fetches data
|
|
|
|
* from the FIFO instead of from Host.
|
|
|
|
*/
|
|
|
|
static inline void pio_copy(u64 __iomem *dst, u64 *src)
|
|
|
|
{
|
|
|
|
int count = 8;
|
|
|
|
|
|
|
|
while (count) {
|
|
|
|
writeq(*src, dst);
|
|
|
|
src++;
|
|
|
|
dst++;
|
|
|
|
count--;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2015-09-23 19:49:27 +08:00
|
|
|
static inline void t4_ring_sq_db(struct t4_wq *wq, u16 inc, union t4_wr *wqe)
|
2010-04-22 06:30:06 +08:00
|
|
|
{
|
2014-04-09 22:38:25 +08:00
|
|
|
|
|
|
|
/* Flush host queue memory writes. */
|
2010-04-22 06:30:06 +08:00
|
|
|
wmb();
|
2015-06-09 20:53:12 +08:00
|
|
|
if (wq->sq.bar2_va) {
|
|
|
|
if (inc == 1 && wq->sq.bar2_qid == 0 && wqe) {
|
2014-04-09 22:38:25 +08:00
|
|
|
PDBG("%s: WC wq->sq.pidx = %d\n",
|
|
|
|
__func__, wq->sq.pidx);
|
2015-06-09 20:53:12 +08:00
|
|
|
pio_copy((u64 __iomem *)
|
|
|
|
(wq->sq.bar2_va + SGE_UDB_WCDOORBELL),
|
|
|
|
(u64 *)wqe);
|
2014-04-09 22:38:25 +08:00
|
|
|
} else {
|
|
|
|
PDBG("%s: DB wq->sq.pidx = %d\n",
|
|
|
|
__func__, wq->sq.pidx);
|
2015-06-09 20:53:12 +08:00
|
|
|
writel(PIDX_T5_V(inc) | QID_V(wq->sq.bar2_qid),
|
|
|
|
wq->sq.bar2_va + SGE_UDB_KDOORBELL);
|
2014-04-09 22:38:25 +08:00
|
|
|
}
|
|
|
|
|
|
|
|
/* Flush user doorbell area writes. */
|
|
|
|
wmb();
|
|
|
|
return;
|
|
|
|
}
|
2015-01-05 19:00:43 +08:00
|
|
|
writel(QID_V(wq->sq.qid) | PIDX_V(inc), wq->db);
|
2010-04-22 06:30:06 +08:00
|
|
|
}
|
|
|
|
|
2015-09-23 19:49:27 +08:00
|
|
|
static inline void t4_ring_rq_db(struct t4_wq *wq, u16 inc,
|
2014-04-09 22:38:25 +08:00
|
|
|
union t4_recv_wr *wqe)
|
2010-04-22 06:30:06 +08:00
|
|
|
{
|
2014-04-09 22:38:25 +08:00
|
|
|
|
|
|
|
/* Flush host queue memory writes. */
|
2010-04-22 06:30:06 +08:00
|
|
|
wmb();
|
2015-06-09 20:53:12 +08:00
|
|
|
if (wq->rq.bar2_va) {
|
|
|
|
if (inc == 1 && wq->rq.bar2_qid == 0 && wqe) {
|
2014-04-09 22:38:25 +08:00
|
|
|
PDBG("%s: WC wq->rq.pidx = %d\n",
|
|
|
|
__func__, wq->rq.pidx);
|
2015-06-09 20:53:12 +08:00
|
|
|
pio_copy((u64 __iomem *)
|
|
|
|
(wq->rq.bar2_va + SGE_UDB_WCDOORBELL),
|
|
|
|
(void *)wqe);
|
2014-04-09 22:38:25 +08:00
|
|
|
} else {
|
|
|
|
PDBG("%s: DB wq->rq.pidx = %d\n",
|
|
|
|
__func__, wq->rq.pidx);
|
2015-06-09 20:53:12 +08:00
|
|
|
writel(PIDX_T5_V(inc) | QID_V(wq->rq.bar2_qid),
|
|
|
|
wq->rq.bar2_va + SGE_UDB_KDOORBELL);
|
2014-04-09 22:38:25 +08:00
|
|
|
}
|
|
|
|
|
|
|
|
/* Flush user doorbell area writes. */
|
|
|
|
wmb();
|
|
|
|
return;
|
|
|
|
}
|
2015-01-05 19:00:43 +08:00
|
|
|
writel(QID_V(wq->rq.qid) | PIDX_V(inc), wq->db);
|
2010-04-22 06:30:06 +08:00
|
|
|
}
|
|
|
|
|
|
|
|
static inline int t4_wq_in_error(struct t4_wq *wq)
|
|
|
|
{
|
2010-09-14 00:23:57 +08:00
|
|
|
return wq->rq.queue[wq->rq.size].status.qp_err;
|
2010-04-22 06:30:06 +08:00
|
|
|
}
|
|
|
|
|
|
|
|
static inline void t4_set_wq_in_error(struct t4_wq *wq)
|
|
|
|
{
|
|
|
|
wq->rq.queue[wq->rq.size].status.qp_err = 1;
|
|
|
|
}
|
|
|
|
|
|
|
|
static inline void t4_disable_wq_db(struct t4_wq *wq)
|
|
|
|
{
|
|
|
|
wq->rq.queue[wq->rq.size].status.db_off = 1;
|
|
|
|
}
|
|
|
|
|
|
|
|
static inline void t4_enable_wq_db(struct t4_wq *wq)
|
|
|
|
{
|
|
|
|
wq->rq.queue[wq->rq.size].status.db_off = 0;
|
|
|
|
}
|
|
|
|
|
|
|
|
static inline int t4_wq_db_enabled(struct t4_wq *wq)
|
|
|
|
{
|
2010-09-14 00:23:57 +08:00
|
|
|
return !wq->rq.queue[wq->rq.size].status.db_off;
|
2010-04-22 06:30:06 +08:00
|
|
|
}
|
|
|
|
|
2014-08-01 03:35:43 +08:00
|
|
|
enum t4_cq_flags {
|
|
|
|
CQ_ARMED = 1,
|
|
|
|
};
|
|
|
|
|
2010-04-22 06:30:06 +08:00
|
|
|
struct t4_cq {
|
|
|
|
struct t4_cqe *queue;
|
|
|
|
dma_addr_t dma_addr;
|
2010-06-03 13:37:50 +08:00
|
|
|
DEFINE_DMA_UNMAP_ADDR(mapping);
|
2010-04-22 06:30:06 +08:00
|
|
|
struct t4_cqe *sw_queue;
|
|
|
|
void __iomem *gts;
|
2015-06-09 20:53:12 +08:00
|
|
|
void __iomem *bar2_va;
|
|
|
|
u64 bar2_pa;
|
|
|
|
u32 bar2_qid;
|
2010-04-22 06:30:06 +08:00
|
|
|
struct c4iw_rdev *rdev;
|
|
|
|
size_t memsize;
|
2010-05-21 05:57:43 +08:00
|
|
|
__be64 bits_type_ts;
|
2010-04-22 06:30:06 +08:00
|
|
|
u32 cqid;
|
2015-04-22 04:15:00 +08:00
|
|
|
u32 qid_mask;
|
2014-06-07 00:10:42 +08:00
|
|
|
int vector;
|
2010-04-22 06:30:06 +08:00
|
|
|
u16 size; /* including status page */
|
|
|
|
u16 cidx;
|
|
|
|
u16 sw_pidx;
|
|
|
|
u16 sw_cidx;
|
|
|
|
u16 sw_in_use;
|
|
|
|
u16 cidx_inc;
|
|
|
|
u8 gen;
|
|
|
|
u8 error;
|
2014-08-01 03:35:43 +08:00
|
|
|
unsigned long flags;
|
2010-04-22 06:30:06 +08:00
|
|
|
};
|
|
|
|
|
2015-06-09 20:53:12 +08:00
|
|
|
static inline void write_gts(struct t4_cq *cq, u32 val)
|
|
|
|
{
|
|
|
|
if (cq->bar2_va)
|
|
|
|
writel(val | INGRESSQID_V(cq->bar2_qid),
|
|
|
|
cq->bar2_va + SGE_UDB_GTS);
|
|
|
|
else
|
|
|
|
writel(val | INGRESSQID_V(cq->cqid), cq->gts);
|
|
|
|
}
|
|
|
|
|
2014-08-01 03:35:43 +08:00
|
|
|
static inline int t4_clear_cq_armed(struct t4_cq *cq)
|
|
|
|
{
|
|
|
|
return test_and_clear_bit(CQ_ARMED, &cq->flags);
|
|
|
|
}
|
|
|
|
|
2010-04-22 06:30:06 +08:00
|
|
|
static inline int t4_arm_cq(struct t4_cq *cq, int se)
|
|
|
|
{
|
|
|
|
u32 val;
|
2010-05-21 05:57:49 +08:00
|
|
|
|
2014-08-01 03:35:43 +08:00
|
|
|
set_bit(CQ_ARMED, &cq->flags);
|
2015-01-05 19:00:43 +08:00
|
|
|
while (cq->cidx_inc > CIDXINC_M) {
|
2015-06-09 20:53:12 +08:00
|
|
|
val = SEINTARM_V(0) | CIDXINC_V(CIDXINC_M) | TIMERREG_V(7);
|
|
|
|
write_gts(cq, val);
|
2015-01-05 19:00:43 +08:00
|
|
|
cq->cidx_inc -= CIDXINC_M;
|
2010-05-21 05:57:49 +08:00
|
|
|
}
|
2015-06-09 20:53:12 +08:00
|
|
|
val = SEINTARM_V(se) | CIDXINC_V(cq->cidx_inc) | TIMERREG_V(6);
|
|
|
|
write_gts(cq, val);
|
2010-05-21 05:57:49 +08:00
|
|
|
cq->cidx_inc = 0;
|
2010-04-22 06:30:06 +08:00
|
|
|
return 0;
|
|
|
|
}
|
|
|
|
|
|
|
|
static inline void t4_swcq_produce(struct t4_cq *cq)
|
|
|
|
{
|
|
|
|
cq->sw_in_use++;
|
2013-08-06 23:34:35 +08:00
|
|
|
if (cq->sw_in_use == cq->size) {
|
|
|
|
PDBG("%s cxgb4 sw cq overflow cqid %u\n", __func__, cq->cqid);
|
|
|
|
cq->error = 1;
|
|
|
|
BUG_ON(1);
|
|
|
|
}
|
2010-04-22 06:30:06 +08:00
|
|
|
if (++cq->sw_pidx == cq->size)
|
|
|
|
cq->sw_pidx = 0;
|
|
|
|
}
|
|
|
|
|
|
|
|
static inline void t4_swcq_consume(struct t4_cq *cq)
|
|
|
|
{
|
2013-08-06 23:34:35 +08:00
|
|
|
BUG_ON(cq->sw_in_use < 1);
|
2010-04-22 06:30:06 +08:00
|
|
|
cq->sw_in_use--;
|
|
|
|
if (++cq->sw_cidx == cq->size)
|
|
|
|
cq->sw_cidx = 0;
|
|
|
|
}
|
|
|
|
|
|
|
|
static inline void t4_hwcq_consume(struct t4_cq *cq)
|
|
|
|
{
|
2010-05-21 05:57:43 +08:00
|
|
|
cq->bits_type_ts = cq->queue[cq->cidx].bits_type_ts;
|
2015-01-05 19:00:43 +08:00
|
|
|
if (++cq->cidx_inc == (cq->size >> 4) || cq->cidx_inc == CIDXINC_M) {
|
2011-03-12 06:30:42 +08:00
|
|
|
u32 val;
|
|
|
|
|
2015-06-09 20:53:12 +08:00
|
|
|
val = SEINTARM_V(0) | CIDXINC_V(cq->cidx_inc) | TIMERREG_V(7);
|
|
|
|
write_gts(cq, val);
|
2010-05-21 05:57:49 +08:00
|
|
|
cq->cidx_inc = 0;
|
2011-03-12 06:30:42 +08:00
|
|
|
}
|
2010-04-22 06:30:06 +08:00
|
|
|
if (++cq->cidx == cq->size) {
|
|
|
|
cq->cidx = 0;
|
|
|
|
cq->gen ^= 1;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
static inline int t4_valid_cqe(struct t4_cq *cq, struct t4_cqe *cqe)
|
|
|
|
{
|
|
|
|
return (CQE_GENBIT(cqe) == cq->gen);
|
|
|
|
}
|
|
|
|
|
2016-08-23 22:57:33 +08:00
|
|
|
static inline int t4_cq_notempty(struct t4_cq *cq)
|
|
|
|
{
|
|
|
|
return cq->sw_in_use || t4_valid_cqe(cq, &cq->queue[cq->cidx]);
|
|
|
|
}
|
|
|
|
|
2010-04-22 06:30:06 +08:00
|
|
|
static inline int t4_next_hw_cqe(struct t4_cq *cq, struct t4_cqe **cqe)
|
|
|
|
{
|
2010-05-21 05:57:43 +08:00
|
|
|
int ret;
|
|
|
|
u16 prev_cidx;
|
2010-04-22 06:30:06 +08:00
|
|
|
|
2010-05-21 05:57:43 +08:00
|
|
|
if (cq->cidx == 0)
|
|
|
|
prev_cidx = cq->size - 1;
|
2010-04-22 06:30:06 +08:00
|
|
|
else
|
2010-05-21 05:57:43 +08:00
|
|
|
prev_cidx = cq->cidx - 1;
|
|
|
|
|
|
|
|
if (cq->queue[prev_cidx].bits_type_ts != cq->bits_type_ts) {
|
|
|
|
ret = -EOVERFLOW;
|
2010-04-22 06:30:06 +08:00
|
|
|
cq->error = 1;
|
2010-05-21 05:57:43 +08:00
|
|
|
printk(KERN_ERR MOD "cq overflow cqid %u\n", cq->cqid);
|
2013-08-06 23:34:35 +08:00
|
|
|
BUG_ON(1);
|
2010-05-21 05:57:43 +08:00
|
|
|
} else if (t4_valid_cqe(cq, &cq->queue[cq->cidx])) {
|
2014-04-09 22:38:26 +08:00
|
|
|
|
|
|
|
/* Ensure CQE is flushed to memory */
|
|
|
|
rmb();
|
2010-05-21 05:57:43 +08:00
|
|
|
*cqe = &cq->queue[cq->cidx];
|
|
|
|
ret = 0;
|
|
|
|
} else
|
|
|
|
ret = -ENODATA;
|
2010-04-22 06:30:06 +08:00
|
|
|
return ret;
|
|
|
|
}
|
|
|
|
|
|
|
|
static inline struct t4_cqe *t4_next_sw_cqe(struct t4_cq *cq)
|
|
|
|
{
|
2013-08-06 23:34:35 +08:00
|
|
|
if (cq->sw_in_use == cq->size) {
|
|
|
|
PDBG("%s cxgb4 sw cq overflow cqid %u\n", __func__, cq->cqid);
|
|
|
|
cq->error = 1;
|
|
|
|
BUG_ON(1);
|
|
|
|
return NULL;
|
|
|
|
}
|
2010-04-22 06:30:06 +08:00
|
|
|
if (cq->sw_in_use)
|
|
|
|
return &cq->sw_queue[cq->sw_cidx];
|
|
|
|
return NULL;
|
|
|
|
}
|
|
|
|
|
|
|
|
static inline int t4_next_cqe(struct t4_cq *cq, struct t4_cqe **cqe)
|
|
|
|
{
|
|
|
|
int ret = 0;
|
|
|
|
|
|
|
|
if (cq->error)
|
|
|
|
ret = -ENODATA;
|
|
|
|
else if (cq->sw_in_use)
|
|
|
|
*cqe = &cq->sw_queue[cq->sw_cidx];
|
|
|
|
else
|
|
|
|
ret = t4_next_hw_cqe(cq, cqe);
|
|
|
|
return ret;
|
|
|
|
}
|
|
|
|
|
|
|
|
static inline int t4_cq_in_error(struct t4_cq *cq)
|
|
|
|
{
|
|
|
|
return ((struct t4_status_page *)&cq->queue[cq->size])->qp_err;
|
|
|
|
}
|
|
|
|
|
|
|
|
static inline void t4_set_cq_in_error(struct t4_cq *cq)
|
|
|
|
{
|
|
|
|
((struct t4_status_page *)&cq->queue[cq->size])->qp_err = 1;
|
|
|
|
}
|
|
|
|
#endif
|
cxgb4/iw_cxgb4: Doorbell Drop Avoidance Bug Fixes
The current logic suffers from a slow response time to disable user DB
usage, and also fails to avoid DB FIFO drops under heavy load. This commit
fixes these deficiencies and makes the avoidance logic more optimal.
This is done by more efficiently notifying the ULDs of potential DB
problems, and implements a smoother flow control algorithm in iw_cxgb4,
which is the ULD that puts the most load on the DB fifo.
Design:
cxgb4:
Direct ULD callback from the DB FULL/DROP interrupt handler. This allows
the ULD to stop doing user DB writes as quickly as possible.
While user DB usage is disabled, the LLD will accumulate DB write events
for its queues. Then once DB usage is reenabled, a single DB write is
done for each queue with its accumulated write count. This reduces the
load put on the DB fifo when reenabling.
iw_cxgb4:
Instead of marking each qp to indicate DB writes are disabled, we create
a device-global status page that each user process maps. This allows
iw_cxgb4 to only set this single bit to disable all DB writes for all
user QPs vs traversing the idr of all the active QPs. If the libcxgb4
doesn't support this, then we fall back to the old approach of marking
each QP. Thus we allow the new driver to work with an older libcxgb4.
When the LLD upcalls iw_cxgb4 indicating DB FULL, we disable all DB writes
via the status page and transition the DB state to STOPPED. As user
processes see that DB writes are disabled, they call into iw_cxgb4
to submit their DB write events. Since the DB state is in STOPPED,
the QP trying to write gets enqueued on a new DB "flow control" list.
As subsequent DB writes are submitted for this flow controlled QP, the
amount of writes are accumulated for each QP on the flow control list.
So all the user QPs that are actively ringing the DB get put on this
list and the number of writes they request are accumulated.
When the LLD upcalls iw_cxgb4 indicating DB EMPTY, which is in a workq
context, we change the DB state to FLOW_CONTROL, and begin resuming all
the QPs that are on the flow control list. This logic runs on until
the flow control list is empty or we exit FLOW_CONTROL mode (due to
a DB DROP upcall, for example). QPs are removed from this list, and
their accumulated DB write counts written to the DB FIFO. Sets of QPs,
called chunks in the code, are removed at one time. The chunk size is 64.
So 64 QPs are resumed at a time, and before the next chunk is resumed, the
logic waits (blocks) for the DB FIFO to drain. This prevents resuming to
quickly and overflowing the FIFO. Once the flow control list is empty,
the db state transitions back to NORMAL and user QPs are again allowed
to write directly to the user DB register.
The algorithm is designed such that if the DB write load is high enough,
then all the DB writes get submitted by the kernel using this flow
controlled approach to avoid DB drops. As the load lightens though, we
resume to normal DB writes directly by user applications.
Signed-off-by: Steve Wise <swise@opengridcomputing.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
2014-03-15 00:22:08 +08:00
|
|
|
|
|
|
|
struct t4_dev_status_page {
|
|
|
|
u8 db_off;
|
2015-12-11 15:32:01 +08:00
|
|
|
u8 pad1;
|
|
|
|
u16 pad2;
|
|
|
|
u32 pad3;
|
|
|
|
u64 qp_start;
|
|
|
|
u64 qp_size;
|
|
|
|
u64 cq_start;
|
|
|
|
u64 cq_size;
|
cxgb4/iw_cxgb4: Doorbell Drop Avoidance Bug Fixes
The current logic suffers from a slow response time to disable user DB
usage, and also fails to avoid DB FIFO drops under heavy load. This commit
fixes these deficiencies and makes the avoidance logic more optimal.
This is done by more efficiently notifying the ULDs of potential DB
problems, and implements a smoother flow control algorithm in iw_cxgb4,
which is the ULD that puts the most load on the DB fifo.
Design:
cxgb4:
Direct ULD callback from the DB FULL/DROP interrupt handler. This allows
the ULD to stop doing user DB writes as quickly as possible.
While user DB usage is disabled, the LLD will accumulate DB write events
for its queues. Then once DB usage is reenabled, a single DB write is
done for each queue with its accumulated write count. This reduces the
load put on the DB fifo when reenabling.
iw_cxgb4:
Instead of marking each qp to indicate DB writes are disabled, we create
a device-global status page that each user process maps. This allows
iw_cxgb4 to only set this single bit to disable all DB writes for all
user QPs vs traversing the idr of all the active QPs. If the libcxgb4
doesn't support this, then we fall back to the old approach of marking
each QP. Thus we allow the new driver to work with an older libcxgb4.
When the LLD upcalls iw_cxgb4 indicating DB FULL, we disable all DB writes
via the status page and transition the DB state to STOPPED. As user
processes see that DB writes are disabled, they call into iw_cxgb4
to submit their DB write events. Since the DB state is in STOPPED,
the QP trying to write gets enqueued on a new DB "flow control" list.
As subsequent DB writes are submitted for this flow controlled QP, the
amount of writes are accumulated for each QP on the flow control list.
So all the user QPs that are actively ringing the DB get put on this
list and the number of writes they request are accumulated.
When the LLD upcalls iw_cxgb4 indicating DB EMPTY, which is in a workq
context, we change the DB state to FLOW_CONTROL, and begin resuming all
the QPs that are on the flow control list. This logic runs on until
the flow control list is empty or we exit FLOW_CONTROL mode (due to
a DB DROP upcall, for example). QPs are removed from this list, and
their accumulated DB write counts written to the DB FIFO. Sets of QPs,
called chunks in the code, are removed at one time. The chunk size is 64.
So 64 QPs are resumed at a time, and before the next chunk is resumed, the
logic waits (blocks) for the DB FIFO to drain. This prevents resuming to
quickly and overflowing the FIFO. Once the flow control list is empty,
the db state transitions back to NORMAL and user QPs are again allowed
to write directly to the user DB register.
The algorithm is designed such that if the DB write load is high enough,
then all the DB writes get submitted by the kernel using this flow
controlled approach to avoid DB drops. As the load lightens though, we
resume to normal DB writes directly by user applications.
Signed-off-by: Steve Wise <swise@opengridcomputing.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
2014-03-15 00:22:08 +08:00
|
|
|
};
|