IB/qib: Add cache line awareness to qib_qp and qib_devdata structures

This patch reorganizes the QP and devdata files to be more cache line aware.

qib_qp fields in particular are split into read-mostly, send, and receive fields.

qib_devdata fields are split into read-mostly and read/write fields

Testing has show that bidirectional tests improve by as much as 100%
with this patch.

Signed-off-by: Mike Marciniszyn <mike.marciniszyn@intel.com>
Signed-off-by: Roland Dreier <roland@purestorage.com>

7 files changed, 120 insertions, 94 deletions

diff --git a/drivers/infiniband/hw/qib/qib.h b/drivers/infiniband/hw/qib/qib.h
index 2d63887..7e62f41 100644
--- a/drivers/infiniband/hw/qib/qib.h
+++ b/drivers/infiniband/hw/qib/qib.h
@@ -530,8 +530,6 @@ struct qib_pportdata {
 	/* qib_lflags driver is waiting for */
 	u32 state_wanted;
 	spinlock_t lflags_lock;
-	/* number of (port-specific) interrupts for this port -- saturates... */
-	u32 int_counter;
 
 	/* ref count for each pkey */
 	atomic_t pkeyrefs[4];
@@ -543,24 +541,26 @@ struct qib_pportdata {
 	u64 *statusp;
 
 	/* SendDMA related entries */
-	spinlock_t            sdma_lock;
-	struct qib_sdma_state sdma_state;
-	unsigned long         sdma_buf_jiffies;
+
+	/* read mostly */
 	struct qib_sdma_desc *sdma_descq;
+	struct qib_sdma_state sdma_state;
+	dma_addr_t       sdma_descq_phys;
+	volatile __le64 *sdma_head_dma; /* DMA'ed by chip */
+	dma_addr_t       sdma_head_phys;
+	u16                   sdma_descq_cnt;
+
+	/* read/write using lock */
+	spinlock_t            sdma_lock ____cacheline_aligned_in_smp;
+	struct list_head      sdma_activelist;
 	u64                   sdma_descq_added;
 	u64                   sdma_descq_removed;
-	u16                   sdma_descq_cnt;
 	u16                   sdma_descq_tail;
 	u16                   sdma_descq_head;
-	u16                   sdma_next_intr;
-	u16                   sdma_reset_wait;
 	u8                    sdma_generation;
-	struct tasklet_struct sdma_sw_clean_up_task;
-	struct list_head      sdma_activelist;
 
-	dma_addr_t       sdma_descq_phys;
-	volatile __le64 *sdma_head_dma; /* DMA'ed by chip */
-	dma_addr_t       sdma_head_phys;
+	struct tasklet_struct sdma_sw_clean_up_task
+		____cacheline_aligned_in_smp;
 
 	wait_queue_head_t state_wait; /* for state_wanted */
 
diff --git a/drivers/infiniband/hw/qib/qib_qp.c b/drivers/infiniband/hw/qib/qib_qp.c
index 7e7e16f..1ce56b5 100644
--- a/drivers/infiniband/hw/qib/qib_qp.c
+++ b/drivers/infiniband/hw/qib/qib_qp.c
@@ -1038,6 +1038,11 @@ struct ib_qp *qib_create_qp(struct ib_pd *ibpd,
 			goto bail_swq;
 		}
 		RCU_INIT_POINTER(qp->next, NULL);
+		qp->s_hdr = kzalloc(sizeof(*qp->s_hdr), GFP_KERNEL);
+		if (!qp->s_hdr) {
+			ret = ERR_PTR(-ENOMEM);
+			goto bail_qp;
+		}
 		qp->timeout_jiffies =
 			usecs_to_jiffies((4096UL * (1UL << qp->timeout)) /
 				1000UL);
@@ -1159,6 +1164,7 @@ bail_ip:
 		vfree(qp->r_rq.wq);
 	free_qpn(&dev->qpn_table, qp->ibqp.qp_num);
 bail_qp:
+	kfree(qp->s_hdr);
 	kfree(qp);
 bail_swq:
 	vfree(swq);
@@ -1214,6 +1220,7 @@ int qib_destroy_qp(struct ib_qp *ibqp)
 	else
 		vfree(qp->r_rq.wq);
 	vfree(qp->s_wq);
+	kfree(qp->s_hdr);
 	kfree(qp);
 	return 0;
 }
diff --git a/drivers/infiniband/hw/qib/qib_rc.c b/drivers/infiniband/hw/qib/qib_rc.c
index 765b4cb..b641416 100644
--- a/drivers/infiniband/hw/qib/qib_rc.c
+++ b/drivers/infiniband/hw/qib/qib_rc.c
@@ -244,9 +244,9 @@ int qib_make_rc_req(struct qib_qp *qp)
 	int ret = 0;
 	int delta;
 
-	ohdr = &qp->s_hdr.u.oth;
+	ohdr = &qp->s_hdr->u.oth;
 	if (qp->remote_ah_attr.ah_flags & IB_AH_GRH)
-		ohdr = &qp->s_hdr.u.l.oth;
+		ohdr = &qp->s_hdr->u.l.oth;
 
 	/*
 	 * The lock is needed to synchronize between the sending tasklet,
diff --git a/drivers/infiniband/hw/qib/qib_ruc.c b/drivers/infiniband/hw/qib/qib_ruc.c
index b4b37e4..c0ee7e0 100644
--- a/drivers/infiniband/hw/qib/qib_ruc.c
+++ b/drivers/infiniband/hw/qib/qib_ruc.c
@@ -688,17 +688,17 @@ void qib_make_ruc_header(struct qib_qp *qp, struct qib_other_headers *ohdr,
 	nwords = (qp->s_cur_size + extra_bytes) >> 2;
 	lrh0 = QIB_LRH_BTH;
 	if (unlikely(qp->remote_ah_attr.ah_flags & IB_AH_GRH)) {
-		qp->s_hdrwords += qib_make_grh(ibp, &qp->s_hdr.u.l.grh,
+		qp->s_hdrwords += qib_make_grh(ibp, &qp->s_hdr->u.l.grh,
 					       &qp->remote_ah_attr.grh,
 					       qp->s_hdrwords, nwords);
 		lrh0 = QIB_LRH_GRH;
 	}
 	lrh0 |= ibp->sl_to_vl[qp->remote_ah_attr.sl] << 12 |
 		qp->remote_ah_attr.sl << 4;
-	qp->s_hdr.lrh[0] = cpu_to_be16(lrh0);
-	qp->s_hdr.lrh[1] = cpu_to_be16(qp->remote_ah_attr.dlid);
-	qp->s_hdr.lrh[2] = cpu_to_be16(qp->s_hdrwords + nwords + SIZE_OF_CRC);
-	qp->s_hdr.lrh[3] = cpu_to_be16(ppd_from_ibp(ibp)->lid |
+	qp->s_hdr->lrh[0] = cpu_to_be16(lrh0);
+	qp->s_hdr->lrh[1] = cpu_to_be16(qp->remote_ah_attr.dlid);
+	qp->s_hdr->lrh[2] = cpu_to_be16(qp->s_hdrwords + nwords + SIZE_OF_CRC);
+	qp->s_hdr->lrh[3] = cpu_to_be16(ppd_from_ibp(ibp)->lid |
 				       qp->remote_ah_attr.src_path_bits);
 	bth0 |= qib_get_pkey(ibp, qp->s_pkey_index);
 	bth0 |= extra_bytes << 20;
@@ -758,7 +758,7 @@ void qib_do_send(struct work_struct *work)
 			 * If the packet cannot be sent now, return and
 			 * the send tasklet will be woken up later.
 			 */
-			if (qib_verbs_send(qp, &qp->s_hdr, qp->s_hdrwords,
+			if (qib_verbs_send(qp, qp->s_hdr, qp->s_hdrwords,
 					   qp->s_cur_sge, qp->s_cur_size))
 				break;
 			/* Record that s_hdr is empty. */
diff --git a/drivers/infiniband/hw/qib/qib_uc.c b/drivers/infiniband/hw/qib/qib_uc.c
index 7ce2ac2..ce7387f 100644
--- a/drivers/infiniband/hw/qib/qib_uc.c
+++ b/drivers/infiniband/hw/qib/qib_uc.c
@@ -72,9 +72,9 @@ int qib_make_uc_req(struct qib_qp *qp)
 		goto done;
 	}
 
-	ohdr = &qp->s_hdr.u.oth;
+	ohdr = &qp->s_hdr->u.oth;
 	if (qp->remote_ah_attr.ah_flags & IB_AH_GRH)
-		ohdr = &qp->s_hdr.u.l.oth;
+		ohdr = &qp->s_hdr->u.l.oth;
 
 	/* header size in 32-bit words LRH+BTH = (8+12)/4. */
 	hwords = 5;
diff --git a/drivers/infiniband/hw/qib/qib_ud.c b/drivers/infiniband/hw/qib/qib_ud.c
index 828609f..a468bf2 100644
--- a/drivers/infiniband/hw/qib/qib_ud.c
+++ b/drivers/infiniband/hw/qib/qib_ud.c
@@ -321,11 +321,11 @@ int qib_make_ud_req(struct qib_qp *qp)
 
 	if (ah_attr->ah_flags & IB_AH_GRH) {
 		/* Header size in 32-bit words. */
-		qp->s_hdrwords += qib_make_grh(ibp, &qp->s_hdr.u.l.grh,
+		qp->s_hdrwords += qib_make_grh(ibp, &qp->s_hdr->u.l.grh,
 					       &ah_attr->grh,
 					       qp->s_hdrwords, nwords);
 		lrh0 = QIB_LRH_GRH;
-		ohdr = &qp->s_hdr.u.l.oth;
+		ohdr = &qp->s_hdr->u.l.oth;
 		/*
 		 * Don't worry about sending to locally attached multicast
 		 * QPs.  It is unspecified by the spec. what happens.
@@ -333,7 +333,7 @@ int qib_make_ud_req(struct qib_qp *qp)
 	} else {
 		/* Header size in 32-bit words. */
 		lrh0 = QIB_LRH_BTH;
-		ohdr = &qp->s_hdr.u.oth;
+		ohdr = &qp->s_hdr->u.oth;
 	}
 	if (wqe->wr.opcode == IB_WR_SEND_WITH_IMM) {
 		qp->s_hdrwords++;
@@ -346,15 +346,15 @@ int qib_make_ud_req(struct qib_qp *qp)
 		lrh0 |= 0xF000; /* Set VL (see ch. 13.5.3.1) */
 	else
 		lrh0 |= ibp->sl_to_vl[ah_attr->sl] << 12;
-	qp->s_hdr.lrh[0] = cpu_to_be16(lrh0);
-	qp->s_hdr.lrh[1] = cpu_to_be16(ah_attr->dlid);  /* DEST LID */
-	qp->s_hdr.lrh[2] = cpu_to_be16(qp->s_hdrwords + nwords + SIZE_OF_CRC);
+	qp->s_hdr->lrh[0] = cpu_to_be16(lrh0);
+	qp->s_hdr->lrh[1] = cpu_to_be16(ah_attr->dlid);  /* DEST LID */
+	qp->s_hdr->lrh[2] = cpu_to_be16(qp->s_hdrwords + nwords + SIZE_OF_CRC);
 	lid = ppd->lid;
 	if (lid) {
 		lid |= ah_attr->src_path_bits & ((1 << ppd->lmc) - 1);
-		qp->s_hdr.lrh[3] = cpu_to_be16(lid);
+		qp->s_hdr->lrh[3] = cpu_to_be16(lid);
 	} else
-		qp->s_hdr.lrh[3] = IB_LID_PERMISSIVE;
+		qp->s_hdr->lrh[3] = IB_LID_PERMISSIVE;
 	if (wqe->wr.send_flags & IB_SEND_SOLICITED)
 		bth0 |= IB_BTH_SOLICITED;
 	bth0 |= extra_bytes << 20;
diff --git a/drivers/infiniband/hw/qib/qib_verbs.h b/drivers/infiniband/hw/qib/qib_verbs.h
index 0c19ef0..4876060 100644
--- a/drivers/infiniband/hw/qib/qib_verbs.h
+++ b/drivers/infiniband/hw/qib/qib_verbs.h
@@ -367,9 +367,10 @@ struct qib_rwq {
 
 struct qib_rq {
 	struct qib_rwq *wq;
-	spinlock_t lock; /* protect changes in this struct */
 	u32 size;               /* size of RWQE array */
 	u8 max_sge;
+	spinlock_t lock /* protect changes in this struct */
+		____cacheline_aligned_in_smp;
 };
 
 struct qib_srq {
@@ -412,31 +413,75 @@ struct qib_ack_entry {
  */
 struct qib_qp {
 	struct ib_qp ibqp;
-	struct qib_qp *next;            /* link list for QPN hash table */
-	struct qib_qp *timer_next;      /* link list for qib_ib_timer() */
-	struct list_head iowait;        /* link for wait PIO buf */
-	struct list_head rspwait;       /* link for waititing to respond */
+	/* read mostly fields above and below */
 	struct ib_ah_attr remote_ah_attr;
 	struct ib_ah_attr alt_ah_attr;
-	struct qib_ib_header s_hdr;     /* next packet header to send */
-	atomic_t refcount;
-	wait_queue_head_t wait;
-	wait_queue_head_t wait_dma;
-	struct timer_list s_timer;
-	struct work_struct s_work;
+	struct qib_qp *next;            /* link list for QPN hash table */
+	struct qib_swqe *s_wq;  /* send work queue */
 	struct qib_mmap_info *ip;
+	struct qib_ib_header *s_hdr;     /* next packet header to send */
+	unsigned long timeout_jiffies;  /* computed from timeout */
+
+	enum ib_mtu path_mtu;
+	u32 remote_qpn;
+	u32 pmtu;		/* decoded from path_mtu */
+	u32 qkey;               /* QKEY for this QP (for UD or RD) */
+	u32 s_size;             /* send work queue size */
+	u32 s_rnr_timeout;      /* number of milliseconds for RNR timeout */
+
+	u8 state;               /* QP state */
+	u8 qp_access_flags;
+	u8 alt_timeout;         /* Alternate path timeout for this QP */
+	u8 timeout;             /* Timeout for this QP */
+	u8 s_srate;
+	u8 s_mig_state;
+	u8 port_num;
+	u8 s_pkey_index;        /* PKEY index to use */
+	u8 s_alt_pkey_index;    /* Alternate path PKEY index to use */
+	u8 r_max_rd_atomic;     /* max number of RDMA read/atomic to receive */
+	u8 s_max_rd_atomic;     /* max number of RDMA read/atomic to send */
+	u8 s_retry_cnt;         /* number of times to retry */
+	u8 s_rnr_retry_cnt;
+	u8 r_min_rnr_timer;     /* retry timeout value for RNR NAKs */
+	u8 s_max_sge;           /* size of s_wq->sg_list */
+	u8 s_draining;
+
+	/* start of read/write fields */
+
+	atomic_t refcount ____cacheline_aligned_in_smp;
+	wait_queue_head_t wait;
+
+
+	struct qib_ack_entry s_ack_queue[QIB_MAX_RDMA_ATOMIC + 1]
+		____cacheline_aligned_in_smp;
+	struct qib_sge_state s_rdma_read_sge;
+
+	spinlock_t r_lock ____cacheline_aligned_in_smp;      /* used for APM */
+	unsigned long r_aflags;
+	u64 r_wr_id;            /* ID for current receive WQE */
+	u32 r_ack_psn;          /* PSN for next ACK or atomic ACK */
+	u32 r_len;              /* total length of r_sge */
+	u32 r_rcv_len;          /* receive data len processed */
+	u32 r_psn;              /* expected rcv packet sequence number */
+	u32 r_msn;              /* message sequence number */
+
+	u8 r_state;             /* opcode of last packet received */
+	u8 r_flags;
+	u8 r_head_ack_queue;    /* index into s_ack_queue[] */
+
+	struct list_head rspwait;       /* link for waititing to respond */
+
+	struct qib_sge_state r_sge;     /* current receive data */
+	struct qib_rq r_rq;             /* receive work queue */
+
+	spinlock_t s_lock ____cacheline_aligned_in_smp;
 	struct qib_sge_state *s_cur_sge;
+	u32 s_flags;
 	struct qib_verbs_txreq *s_tx;
-	struct qib_mregion *s_rdma_mr;
+	struct qib_swqe *s_wqe;
 	struct qib_sge_state s_sge;     /* current send request data */
-	struct qib_ack_entry s_ack_queue[QIB_MAX_RDMA_ATOMIC + 1];
-	struct qib_sge_state s_ack_rdma_sge;
-	struct qib_sge_state s_rdma_read_sge;
-	struct qib_sge_state r_sge;     /* current receive data */
-	spinlock_t r_lock;      /* used for APM */
-	spinlock_t s_lock;
+	struct qib_mregion *s_rdma_mr;
 	atomic_t s_dma_busy;
-	u32 s_flags;
 	u32 s_cur_size;         /* size of send packet in bytes */
 	u32 s_len;              /* total length of s_sge */
 	u32 s_rdma_read_len;    /* total length of s_rdma_read_sge */
@@ -447,60 +492,34 @@ struct qib_qp {
 	u32 s_psn;              /* current packet sequence number */
 	u32 s_ack_rdma_psn;     /* PSN for sending RDMA read responses */
 	u32 s_ack_psn;          /* PSN for acking sends and RDMA writes */
-	u32 s_rnr_timeout;      /* number of milliseconds for RNR timeout */
-	u32 r_ack_psn;          /* PSN for next ACK or atomic ACK */
-	u64 r_wr_id;            /* ID for current receive WQE */
-	unsigned long r_aflags;
-	u32 r_len;              /* total length of r_sge */
-	u32 r_rcv_len;          /* receive data len processed */
-	u32 r_psn;              /* expected rcv packet sequence number */
-	u32 r_msn;              /* message sequence number */
+	u32 s_head;             /* new entries added here */
+	u32 s_tail;             /* next entry to process */
+	u32 s_cur;              /* current work queue entry */
+	u32 s_acked;            /* last un-ACK'ed entry */
+	u32 s_last;             /* last completed entry */
+	u32 s_ssn;              /* SSN of tail entry */
+	u32 s_lsn;              /* limit sequence number (credit) */
 	u16 s_hdrwords;         /* size of s_hdr in 32 bit words */
 	u16 s_rdma_ack_cnt;
-	u8 state;               /* QP state */
 	u8 s_state;             /* opcode of last packet sent */
 	u8 s_ack_state;         /* opcode of packet to ACK */
 	u8 s_nak_state;         /* non-zero if NAK is pending */
-	u8 r_state;             /* opcode of last packet received */
 	u8 r_nak_state;         /* non-zero if NAK is pending */
-	u8 r_min_rnr_timer;     /* retry timeout value for RNR NAKs */
-	u8 r_flags;
-	u8 r_max_rd_atomic;     /* max number of RDMA read/atomic to receive */
-	u8 r_head_ack_queue;    /* index into s_ack_queue[] */
-	u8 qp_access_flags;
-	u8 s_max_sge;           /* size of s_wq->sg_list */
-	u8 s_retry_cnt;         /* number of times to retry */
-	u8 s_rnr_retry_cnt;
 	u8 s_retry;             /* requester retry counter */
 	u8 s_rnr_retry;         /* requester RNR retry counter */
-	u8 s_pkey_index;        /* PKEY index to use */
-	u8 s_alt_pkey_index;    /* Alternate path PKEY index to use */
-	u8 s_max_rd_atomic;     /* max number of RDMA read/atomic to send */
 	u8 s_num_rd_atomic;     /* number of RDMA read/atomic pending */
 	u8 s_tail_ack_queue;    /* index into s_ack_queue[] */
-	u8 s_srate;
-	u8 s_draining;
-	u8 s_mig_state;
-	u8 timeout;             /* Timeout for this QP */
-	u8 alt_timeout;         /* Alternate path timeout for this QP */
-	u8 port_num;
-	enum ib_mtu path_mtu;
-	u32 pmtu;		/* decoded from path_mtu */
-	u32 remote_qpn;
-	u32 qkey;               /* QKEY for this QP (for UD or RD) */
-	u32 s_size;             /* send work queue size */
-	u32 s_head;             /* new entries added here */
-	u32 s_tail;             /* next entry to process */
-	u32 s_cur;              /* current work queue entry */
-	u32 s_acked;            /* last un-ACK'ed entry */
-	u32 s_last;             /* last completed entry */
-	u32 s_ssn;              /* SSN of tail entry */
-	u32 s_lsn;              /* limit sequence number (credit) */
-	unsigned long timeout_jiffies;  /* computed from timeout */
-	struct qib_swqe *s_wq;  /* send work queue */
-	struct qib_swqe *s_wqe;
-	struct qib_rq r_rq;             /* receive work queue */
-	struct qib_sge r_sg_list[0];    /* verified SGEs */
+
+	struct qib_sge_state s_ack_rdma_sge;
+	struct timer_list s_timer;
+	struct list_head iowait;        /* link for wait PIO buf */
+
+	struct work_struct s_work;
+
+	wait_queue_head_t wait_dma;
+
+	struct qib_sge r_sg_list[0] /* verified SGEs */
+		____cacheline_aligned_in_smp;
 };
 
 /*