#include <linux/dmapool.h>
#include "rds.h"
-#include "rdma.h"
#include "ib.h"
-static void rds_ib_send_rdma_complete(struct rds_message *rm,
- int wc_status)
+/*
+ * Convert IB-specific error message to RDS error message and call core
+ * completion handler.
+ */
+static void rds_ib_send_complete(struct rds_message *rm,
+ int wc_status,
+ void (*complete)(struct rds_message *rm, int status))
{
int notify_status;
notify_status = RDS_RDMA_OTHER_ERROR;
break;
}
- rds_rdma_send_complete(rm, notify_status);
-}
-
-static void rds_ib_send_unmap_rdma(struct rds_ib_connection *ic,
- struct rds_rdma_op *op)
-{
- if (op->r_mapped) {
- ib_dma_unmap_sg(ic->i_cm_id->device,
- op->r_sg, op->r_nents,
- op->r_write ? DMA_TO_DEVICE : DMA_FROM_DEVICE);
- op->r_mapped = 0;
- }
+ complete(rm, notify_status);
}
static void rds_ib_send_unmap_rm(struct rds_ib_connection *ic,
rdsdebug("ic %p send %p rm %p\n", ic, send, rm);
ib_dma_unmap_sg(ic->i_cm_id->device,
- rm->m_sg, rm->m_nents,
- DMA_TO_DEVICE);
+ rm->data.m_sg, rm->data.m_nents,
+ DMA_TO_DEVICE);
- if (rm->m_rdma_op != NULL) {
- rds_ib_send_unmap_rdma(ic, rm->m_rdma_op);
+ if (rm->rdma.m_rdma_op.r_active) {
+ struct rds_rdma_op *op = &rm->rdma.m_rdma_op;
+
+ if (op->r_mapped) {
+ ib_dma_unmap_sg(ic->i_cm_id->device,
+ op->r_sg, op->r_nents,
+ op->r_write ? DMA_TO_DEVICE : DMA_FROM_DEVICE);
+ op->r_mapped = 0;
+ }
/* If the user asked for a completion notification on this
* message, we can implement three different semantics:
* operation itself unmapped the RDMA buffers, which takes care
* of synching.
*/
- rds_ib_send_rdma_complete(rm, wc_status);
+ rds_ib_send_complete(rm, wc_status, rds_rdma_send_complete);
- if (rm->m_rdma_op->r_write)
- rds_stats_add(s_send_rdma_bytes, rm->m_rdma_op->r_bytes);
+ if (rm->rdma.m_rdma_op.r_write)
+ rds_stats_add(s_send_rdma_bytes, rm->rdma.m_rdma_op.r_bytes);
else
- rds_stats_add(s_recv_rdma_bytes, rm->m_rdma_op->r_bytes);
+ rds_stats_add(s_recv_rdma_bytes, rm->rdma.m_rdma_op.r_bytes);
+ }
+
+ if (rm->atomic.op_active) {
+ struct rm_atomic_op *op = &rm->atomic;
+
+ /* unmap atomic recvbuf */
+ if (op->op_mapped) {
+ ib_dma_unmap_sg(ic->i_cm_id->device, op->op_sg, 1,
+ DMA_FROM_DEVICE);
+ op->op_mapped = 0;
+ }
+
+ rds_ib_send_complete(rm, wc_status, rds_atomic_send_complete);
+
+ if (rm->atomic.op_type == RDS_ATOMIC_TYPE_CSWP)
+ rds_stats_inc(s_atomic_cswp);
+ else
+ rds_stats_inc(s_atomic_fadd);
}
/* If anyone waited for this message to get flushed out, wake
send->s_wr.wr_id = i;
send->s_wr.sg_list = send->s_sge;
- send->s_wr.num_sge = 1;
- send->s_wr.opcode = IB_WR_SEND;
- send->s_wr.send_flags = 0;
send->s_wr.ex.imm_data = 0;
- sge = rds_ib_data_sge(ic, send->s_sge);
- sge->lkey = ic->i_mr->lkey;
-
- sge = rds_ib_header_sge(ic, send->s_sge);
+ sge = &send->s_sge[0];
sge->addr = ic->i_send_hdrs_dma + (i * sizeof(struct rds_header));
sge->length = sizeof(struct rds_header);
sge->lkey = ic->i_mr->lkey;
+
+ send->s_sge[1].lkey = ic->i_mr->lkey;
}
}
u32 i;
for (i = 0, send = ic->i_sends; i < ic->i_send_ring.w_nr; i++, send++) {
- if (send->s_wr.opcode == 0xdead)
+ if (!send->s_rm || send->s_wr.opcode == 0xdead)
continue;
- if (send->s_rm)
- rds_ib_send_unmap_rm(ic, send, IB_WC_WR_FLUSH_ERR);
- if (send->s_op)
- rds_ib_send_unmap_rdma(ic, send->s_op);
+ rds_ib_send_unmap_rm(ic, send, IB_WC_WR_FLUSH_ERR);
}
}
break;
case IB_WR_RDMA_WRITE:
case IB_WR_RDMA_READ:
+ case IB_WR_ATOMIC_FETCH_AND_ADD:
+ case IB_WR_ATOMIC_CMP_AND_SWP:
/* Nothing to be done - the SG list will be unmapped
* when the SEND completes. */
break;
rm = rds_send_get_message(conn, send->s_op);
if (rm) {
- if (rm->m_rdma_op)
- rds_ib_send_unmap_rdma(ic, rm->m_rdma_op);
- rds_ib_send_rdma_complete(rm, wc.status);
+ rds_ib_send_unmap_rm(ic, send, wc.status);
+ rds_ib_send_complete(rm, wc.status, rds_rdma_send_complete);
rds_message_put(rm);
}
}
set_bit(IB_ACK_REQUESTED, &ic->i_ack_flags);
}
-static inline void
-rds_ib_xmit_populate_wr(struct rds_ib_connection *ic,
- struct rds_ib_send_work *send, unsigned int pos,
- unsigned long buffer, unsigned int length,
- int send_flags)
-{
- struct ib_sge *sge;
-
- WARN_ON(pos != send - ic->i_sends);
-
- send->s_wr.send_flags = send_flags;
- send->s_wr.opcode = IB_WR_SEND;
- send->s_wr.num_sge = 2;
- send->s_wr.next = NULL;
- send->s_queued = jiffies;
- send->s_op = NULL;
-
- if (length != 0) {
- sge = rds_ib_data_sge(ic, send->s_sge);
- sge->addr = buffer;
- sge->length = length;
- sge->lkey = ic->i_mr->lkey;
-
- sge = rds_ib_header_sge(ic, send->s_sge);
- } else {
- /* We're sending a packet with no payload. There is only
- * one SGE */
- send->s_wr.num_sge = 1;
- sge = &send->s_sge[0];
- }
-
- sge->addr = ic->i_send_hdrs_dma + (pos * sizeof(struct rds_header));
- sge->length = sizeof(struct rds_header);
- sge->lkey = ic->i_mr->lkey;
-}
-
/*
* This can be called multiple times for a given message. The first time
* we see a message we map its scatterlist into the IB device so that
u32 pos;
u32 i;
u32 work_alloc;
- u32 credit_alloc;
+ u32 credit_alloc = 0;
u32 posted;
u32 adv_credits = 0;
int send_flags = 0;
- int sent;
+ int bytes_sent = 0;
int ret;
int flow_controlled = 0;
goto out;
}
- credit_alloc = work_alloc;
if (ic->i_flowctl) {
credit_alloc = rds_ib_send_grab_credits(ic, work_alloc, &posted, 0, RDS_MAX_ADV_CREDIT);
adv_credits += posted;
}
/* map the message the first time we see it */
- if (ic->i_rm == NULL) {
- /*
- printk(KERN_NOTICE "rds_ib_xmit prep msg dport=%u flags=0x%x len=%d\n",
- be16_to_cpu(rm->m_inc.i_hdr.h_dport),
- rm->m_inc.i_hdr.h_flags,
- be32_to_cpu(rm->m_inc.i_hdr.h_len));
- */
- if (rm->m_nents) {
- rm->m_count = ib_dma_map_sg(dev,
- rm->m_sg, rm->m_nents, DMA_TO_DEVICE);
- rdsdebug("ic %p mapping rm %p: %d\n", ic, rm, rm->m_count);
- if (rm->m_count == 0) {
+ if (!ic->i_rm) {
+ if (rm->data.m_nents) {
+ rm->data.m_count = ib_dma_map_sg(dev,
+ rm->data.m_sg,
+ rm->data.m_nents,
+ DMA_TO_DEVICE);
+ rdsdebug("ic %p mapping rm %p: %d\n", ic, rm, rm->data.m_count);
+ if (rm->data.m_count == 0) {
rds_ib_stats_inc(s_ib_tx_sg_mapping_failure);
rds_ib_ring_unalloc(&ic->i_send_ring, work_alloc);
ret = -ENOMEM; /* XXX ? */
goto out;
}
} else {
- rm->m_count = 0;
+ rm->data.m_count = 0;
}
ic->i_unsignaled_wrs = rds_ib_sysctl_max_unsig_wrs;
- ic->i_unsignaled_bytes = rds_ib_sysctl_max_unsig_bytes;
rds_message_addref(rm);
ic->i_rm = rm;
/* If it has a RDMA op, tell the peer we did it. This is
* used by the peer to release use-once RDMA MRs. */
- if (rm->m_rdma_op) {
+ if (rm->rdma.m_rdma_op.r_active) {
struct rds_ext_header_rdma ext_hdr;
- ext_hdr.h_rdma_rkey = cpu_to_be32(rm->m_rdma_op->r_key);
+ ext_hdr.h_rdma_rkey = cpu_to_be32(rm->rdma.m_rdma_op.r_key);
rds_message_add_extension(&rm->m_inc.i_hdr,
RDS_EXTHDR_RDMA, &ext_hdr, sizeof(ext_hdr));
}
BUG_ON(adv_credits > 255);
}
- send = &ic->i_sends[pos];
- first = send;
- prev = NULL;
- scat = &rm->m_sg[sg];
- sent = 0;
- i = 0;
-
/* Sometimes you want to put a fence between an RDMA
* READ and the following SEND.
* We could either do this all the time
* or when requested by the user. Right now, we let
* the application choose.
*/
- if (rm->m_rdma_op && rm->m_rdma_op->r_fence)
+ if (rm->rdma.m_rdma_op.r_active && rm->rdma.m_rdma_op.r_fence)
send_flags = IB_SEND_FENCE;
- /*
- * We could be copying the header into the unused tail of the page.
- * That would need to be changed in the future when those pages might
- * be mapped userspace pages or page cache pages. So instead we always
- * use a second sge and our long-lived ring of mapped headers. We send
- * the header after the data so that the data payload can be aligned on
- * the receiver.
- */
+ /* Each frag gets a header. Msgs may be 0 bytes */
+ send = &ic->i_sends[pos];
+ first = send;
+ prev = NULL;
+ scat = &rm->data.m_sg[sg];
+ i = 0;
+ do {
+ unsigned int len = 0;
- /* handle a 0-len message */
- if (be32_to_cpu(rm->m_inc.i_hdr.h_len) == 0) {
- rds_ib_xmit_populate_wr(ic, send, pos, 0, 0, send_flags);
- goto add_header;
- }
+ /* Set up the header */
+ send->s_wr.send_flags = send_flags;
+ send->s_wr.opcode = IB_WR_SEND;
+ send->s_wr.num_sge = 1;
+ send->s_wr.next = NULL;
+ send->s_queued = jiffies;
+ send->s_op = NULL;
- /* if there's data reference it with a chain of work reqs */
- for (; i < work_alloc && scat != &rm->m_sg[rm->m_count]; i++) {
- unsigned int len;
+ send->s_sge[0].addr = ic->i_send_hdrs_dma
+ + (pos * sizeof(struct rds_header));
+ send->s_sge[0].length = sizeof(struct rds_header);
- send = &ic->i_sends[pos];
+ memcpy(&ic->i_send_hdrs[pos], &rm->m_inc.i_hdr, sizeof(struct rds_header));
+
+ /* Set up the data, if present */
+ if (i < work_alloc
+ && scat != &rm->data.m_sg[rm->data.m_count]) {
+ len = min(RDS_FRAG_SIZE, ib_sg_dma_len(dev, scat) - off);
+ send->s_wr.num_sge = 2;
- len = min(RDS_FRAG_SIZE, ib_sg_dma_len(dev, scat) - off);
- rds_ib_xmit_populate_wr(ic, send, pos,
- ib_sg_dma_address(dev, scat) + off, len,
- send_flags);
+ send->s_sge[1].addr = ib_sg_dma_address(dev, scat) + off;
+ send->s_sge[1].length = len;
+
+ bytes_sent += len;
+ off += len;
+ if (off == ib_sg_dma_len(dev, scat)) {
+ scat++;
+ off = 0;
+ }
+ }
/*
* We want to delay signaling completions just enough to get
send->s_wr.send_flags |= IB_SEND_SIGNALED | IB_SEND_SOLICITED;
}
- ic->i_unsignaled_bytes -= len;
- if (ic->i_unsignaled_bytes <= 0) {
- ic->i_unsignaled_bytes = rds_ib_sysctl_max_unsig_bytes;
- send->s_wr.send_flags |= IB_SEND_SIGNALED | IB_SEND_SOLICITED;
- }
-
/*
* Always signal the last one if we're stopping due to flow control.
*/
rdsdebug("send %p wr %p num_sge %u next %p\n", send,
&send->s_wr, send->s_wr.num_sge, send->s_wr.next);
- sent += len;
- off += len;
- if (off == ib_sg_dma_len(dev, scat)) {
- scat++;
- off = 0;
- }
-
-add_header:
- /* Tack on the header after the data. The header SGE should already
- * have been set up to point to the right header buffer. */
- memcpy(&ic->i_send_hdrs[pos], &rm->m_inc.i_hdr, sizeof(struct rds_header));
-
- if (0) {
- struct rds_header *hdr = &ic->i_send_hdrs[pos];
-
- printk(KERN_NOTICE "send WR dport=%u flags=0x%x len=%d\n",
- be16_to_cpu(hdr->h_dport),
- hdr->h_flags,
- be32_to_cpu(hdr->h_len));
- }
if (adv_credits) {
struct rds_header *hdr = &ic->i_send_hdrs[pos];
prev = send;
pos = (pos + 1) % ic->i_send_ring.w_nr;
- }
+ send = &ic->i_sends[pos];
+ i++;
+
+ } while (i < work_alloc
+ && scat != &rm->data.m_sg[rm->data.m_count]);
/* Account the RDS header in the number of bytes we sent, but just once.
* The caller has no concept of fragmentation. */
if (hdr_off == 0)
- sent += sizeof(struct rds_header);
+ bytes_sent += sizeof(struct rds_header);
/* if we finished the message then send completion owns it */
- if (scat == &rm->m_sg[rm->m_count]) {
+ if (scat == &rm->data.m_sg[rm->data.m_count]) {
prev->s_rm = ic->i_rm;
prev->s_wr.send_flags |= IB_SEND_SIGNALED | IB_SEND_SOLICITED;
ic->i_rm = NULL;
}
+ /* Put back wrs & credits we didn't use */
if (i < work_alloc) {
rds_ib_ring_unalloc(&ic->i_send_ring, work_alloc - i);
work_alloc = i;
goto out;
}
- ret = sent;
+ ret = bytes_sent;
out:
BUG_ON(adv_credits);
return ret;
}
+/*
+ * Issue atomic operation.
+ * A simplified version of the rdma case, we always map 1 SG, and
+ * only 8 bytes, for the return value from the atomic operation.
+ */
+int rds_ib_xmit_atomic(struct rds_connection *conn, struct rm_atomic_op *op)
+{
+ struct rds_ib_connection *ic = conn->c_transport_data;
+ struct rds_ib_send_work *send = NULL;
+ struct ib_send_wr *failed_wr;
+ struct rds_ib_device *rds_ibdev;
+ u32 pos;
+ u32 work_alloc;
+ int ret;
+
+ rds_ibdev = ib_get_client_data(ic->i_cm_id->device, &rds_ib_client);
+
+ work_alloc = rds_ib_ring_alloc(&ic->i_send_ring, 1, &pos);
+ if (work_alloc != 1) {
+ rds_ib_ring_unalloc(&ic->i_send_ring, work_alloc);
+ rds_ib_stats_inc(s_ib_tx_ring_full);
+ ret = -ENOMEM;
+ goto out;
+ }
+
+ /* address of send request in ring */
+ send = &ic->i_sends[pos];
+ send->s_queued = jiffies;
+
+ if (op->op_type == RDS_ATOMIC_TYPE_CSWP) {
+ send->s_wr.opcode = IB_WR_ATOMIC_CMP_AND_SWP;
+ send->s_wr.wr.atomic.compare_add = op->op_compare;
+ send->s_wr.wr.atomic.swap = op->op_swap_add;
+ } else { /* FADD */
+ send->s_wr.opcode = IB_WR_ATOMIC_FETCH_AND_ADD;
+ send->s_wr.wr.atomic.compare_add = op->op_swap_add;
+ send->s_wr.wr.atomic.swap = 0;
+ }
+ send->s_wr.send_flags = IB_SEND_SIGNALED;
+ send->s_wr.num_sge = 1;
+ send->s_wr.next = NULL;
+ send->s_wr.wr.atomic.remote_addr = op->op_remote_addr;
+ send->s_wr.wr.atomic.rkey = op->op_rkey;
+
+ /* map 8 byte retval buffer to the device */
+ ret = ib_dma_map_sg(ic->i_cm_id->device, op->op_sg, 1, DMA_FROM_DEVICE);
+ rdsdebug("ic %p mapping atomic op %p. mapped %d pg\n", ic, op, ret);
+ if (ret != 1) {
+ rds_ib_ring_unalloc(&ic->i_send_ring, work_alloc);
+ rds_ib_stats_inc(s_ib_tx_sg_mapping_failure);
+ ret = -ENOMEM; /* XXX ? */
+ goto out;
+ }
+
+ /* Convert our struct scatterlist to struct ib_sge */
+ send->s_sge[0].addr = ib_sg_dma_address(ic->i_cm_id->device, op->op_sg);
+ send->s_sge[0].length = ib_sg_dma_len(ic->i_cm_id->device, op->op_sg);
+ send->s_sge[0].lkey = ic->i_mr->lkey;
+
+ rdsdebug("rva %Lx rpa %Lx len %u\n", op->op_remote_addr,
+ send->s_sge[0].addr, send->s_sge[0].length);
+
+ failed_wr = &send->s_wr;
+ ret = ib_post_send(ic->i_cm_id->qp, &send->s_wr, &failed_wr);
+ rdsdebug("ic %p send %p (wr %p) ret %d wr %p\n", ic,
+ send, &send->s_wr, ret, failed_wr);
+ BUG_ON(failed_wr != &send->s_wr);
+ if (ret) {
+ printk(KERN_WARNING "RDS/IB: atomic ib_post_send to %pI4 "
+ "returned %d\n", &conn->c_faddr, ret);
+ rds_ib_ring_unalloc(&ic->i_send_ring, work_alloc);
+ goto out;
+ }
+
+ if (unlikely(failed_wr != &send->s_wr)) {
+ printk(KERN_WARNING "RDS/IB: atomic ib_post_send() rc=%d, but failed_wqe updated!\n", ret);
+ BUG_ON(failed_wr != &send->s_wr);
+ }
+
+out:
+ return ret;
+}
+
int rds_ib_xmit_rdma(struct rds_connection *conn, struct rds_rdma_op *op)
{
struct rds_ib_connection *ic = conn->c_transport_data;