2018-05-05 02:34:32 +07:00
|
|
|
/* SPDX-License-Identifier: GPL-2.0 OR BSD-3-Clause */
|
2007-09-11 00:50:12 +07:00
|
|
|
/*
|
2017-10-31 03:22:14 +07:00
|
|
|
* Copyright (c) 2014-2017 Oracle. All rights reserved.
|
2007-09-11 00:50:12 +07:00
|
|
|
* Copyright (c) 2003-2007 Network Appliance, 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 BSD-type
|
|
|
|
* 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.
|
|
|
|
*
|
|
|
|
* Neither the name of the Network Appliance, Inc. nor the names of
|
|
|
|
* its contributors may be used to endorse or promote products
|
|
|
|
* derived from this software without specific prior written
|
|
|
|
* permission.
|
|
|
|
*
|
|
|
|
* THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
|
|
|
|
* "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
|
|
|
|
* LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
|
|
|
|
* A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
|
|
|
|
* OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
|
|
|
|
* SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
|
|
|
|
* LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
|
|
|
|
* DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
|
|
|
|
* THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
|
|
|
|
* (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
|
|
|
|
* OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
|
|
|
|
*/
|
|
|
|
|
|
|
|
#ifndef _LINUX_SUNRPC_XPRT_RDMA_H
|
|
|
|
#define _LINUX_SUNRPC_XPRT_RDMA_H
|
|
|
|
|
|
|
|
#include <linux/wait.h> /* wait_queue_head_t, etc */
|
|
|
|
#include <linux/spinlock.h> /* spinlock_t, etc */
|
2011-07-27 06:09:06 +07:00
|
|
|
#include <linux/atomic.h> /* atomic_t, etc */
|
2014-05-28 21:32:17 +07:00
|
|
|
#include <linux/workqueue.h> /* struct work_struct */
|
2007-09-11 00:50:12 +07:00
|
|
|
|
|
|
|
#include <rdma/rdma_cm.h> /* RDMA connection api */
|
|
|
|
#include <rdma/ib_verbs.h> /* RDMA verbs api */
|
|
|
|
|
|
|
|
#include <linux/sunrpc/clnt.h> /* rpc_xprt */
|
|
|
|
#include <linux/sunrpc/rpc_rdma.h> /* RPC/RDMA protocol */
|
|
|
|
#include <linux/sunrpc/xprtrdma.h> /* xprt parameters */
|
|
|
|
|
2008-10-10 02:01:41 +07:00
|
|
|
#define RDMA_RESOLVE_TIMEOUT (5000) /* 5 seconds */
|
|
|
|
#define RDMA_CONNECT_RETRY_MAX (2) /* retries if no listener backlog */
|
|
|
|
|
2016-01-08 02:50:10 +07:00
|
|
|
#define RPCRDMA_BIND_TO (60U * HZ)
|
|
|
|
#define RPCRDMA_INIT_REEST_TO (5U * HZ)
|
|
|
|
#define RPCRDMA_MAX_REEST_TO (30U * HZ)
|
|
|
|
#define RPCRDMA_IDLE_DISC_TO (5U * 60 * HZ)
|
|
|
|
|
2007-09-11 00:50:12 +07:00
|
|
|
/*
|
|
|
|
* Interface Adapter -- one per transport instance
|
|
|
|
*/
|
|
|
|
struct rpcrdma_ia {
|
2015-03-31 01:34:21 +07:00
|
|
|
const struct rpcrdma_memreg_ops *ri_ops;
|
2015-05-26 22:51:56 +07:00
|
|
|
struct ib_device *ri_device;
|
2007-09-11 00:50:12 +07:00
|
|
|
struct rdma_cm_id *ri_id;
|
|
|
|
struct ib_pd *ri_pd;
|
|
|
|
struct completion ri_done;
|
2017-04-12 00:23:10 +07:00
|
|
|
struct completion ri_remove_done;
|
2007-09-11 00:50:12 +07:00
|
|
|
int ri_async_rc;
|
2016-09-15 21:57:07 +07:00
|
|
|
unsigned int ri_max_segs;
|
2017-12-15 08:57:47 +07:00
|
|
|
unsigned int ri_max_frwr_depth;
|
2016-05-03 01:41:05 +07:00
|
|
|
unsigned int ri_max_inline_write;
|
|
|
|
unsigned int ri_max_inline_read;
|
2017-02-09 05:00:10 +07:00
|
|
|
unsigned int ri_max_send_sges;
|
2017-02-09 04:59:54 +07:00
|
|
|
bool ri_implicit_roundup;
|
xprtrdma: Support for SG_GAP devices
Some devices (such as the Mellanox CX-4) can register, under a
single R_key, a set of memory regions that are not contiguous. When
this is done, all the segments in a Reply list, say, can then be
invalidated in a single LocalInv Work Request (or via Remote
Invalidation, which can invalidate exactly one R_key when completing
a Receive).
This means a single FastReg WR is used to register, and one or zero
LocalInv WRs can invalidate, the memory involved with RDMA transfers
on behalf of an RPC.
In addition, xprtrdma constructs some Reply chunks from three or
more segments. By registering them with SG_GAP, only one segment
is needed for the Reply chunk, allowing the whole chunk to be
invalidated remotely.
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: Anna Schumaker <Anna.Schumaker@Netapp.com>
2016-11-29 22:52:24 +07:00
|
|
|
enum ib_mr_type ri_mrtype;
|
2017-04-12 00:23:10 +07:00
|
|
|
unsigned long ri_flags;
|
2015-01-21 23:03:35 +07:00
|
|
|
struct ib_qp_attr ri_qp_attr;
|
|
|
|
struct ib_qp_init_attr ri_qp_init_attr;
|
2007-09-11 00:50:12 +07:00
|
|
|
};
|
|
|
|
|
2017-04-12 00:23:10 +07:00
|
|
|
enum {
|
|
|
|
RPCRDMA_IAF_REMOVING = 0,
|
|
|
|
};
|
|
|
|
|
2007-09-11 00:50:12 +07:00
|
|
|
/*
|
|
|
|
* RDMA Endpoint -- one per transport instance
|
|
|
|
*/
|
|
|
|
|
|
|
|
struct rpcrdma_ep {
|
xprtrdma: Add data structure to manage RDMA Send arguments
Problem statement:
Recently Sagi Grimberg <sagi@grimberg.me> observed that kernel RDMA-
enabled storage initiators don't handle delayed Send completion
correctly. If Send completion is delayed beyond the end of a ULP
transaction, the ULP may release resources that are still being used
by the HCA to complete a long-running Send operation.
This is a common design trait amongst our initiators. Most Send
operations are faster than the ULP transaction they are part of.
Waiting for a completion for these is typically unnecessary.
Infrequently, a network partition or some other problem crops up
where an ordering problem can occur. In NFS parlance, the RPC Reply
arrives and completes the RPC, but the HCA is still retrying the
Send WR that conveyed the RPC Call. In this case, the HCA can try
to use memory that has been invalidated or DMA unmapped, and the
connection is lost. If that memory has been re-used for something
else (possibly not related to NFS), and the Send retransmission
exposes that data on the wire.
Thus we cannot assume that it is safe to release Send-related
resources just because a ULP reply has arrived.
After some analysis, we have determined that the completion
housekeeping will not be difficult for xprtrdma:
- Inline Send buffers are registered via the local DMA key, and
are already left DMA mapped for the lifetime of a transport
connection, thus no additional handling is necessary for those
- Gathered Sends involving page cache pages _will_ need to
DMA unmap those pages after the Send completes. But like
inline send buffers, they are registered via the local DMA key,
and thus will not need to be invalidated
In addition, RPC completion will need to wait for Send completion
in the latter case. However, nearly always, the Send that conveys
the RPC Call will have completed long before the RPC Reply
arrives, and thus no additional latency will be accrued.
Design notes:
In this patch, the rpcrdma_sendctx object is introduced, and a
lock-free circular queue is added to manage a set of them per
transport.
The RPC client's send path already prevents sending more than one
RPC Call at the same time. This allows us to treat the consumer
side of the queue (rpcrdma_sendctx_get_locked) as if there is a
single consumer thread.
The producer side of the queue (rpcrdma_sendctx_put_locked) is
invoked only from the Send completion handler, which is a single
thread of execution (soft IRQ).
The only care that needs to be taken is with the tail index, which
is shared between the producer and consumer. Only the producer
updates the tail index. The consumer compares the head with the
tail to ensure that the a sendctx that is in use is never handed
out again (or, expressed more conventionally, the queue is empty).
When the sendctx queue empties completely, there are enough Sends
outstanding that posting more Send operations can result in a Send
Queue overflow. In this case, the ULP is told to wait and try again.
This introduces strong Send Queue accounting to xprtrdma.
As a final touch, Jason Gunthorpe <jgunthorpe@obsidianresearch.com>
suggested a mechanism that does not require signaling every Send.
We signal once every N Sends, and perform SGE unmapping of N Send
operations during that one completion.
Reported-by: Sagi Grimberg <sagi@grimberg.me>
Suggested-by: Jason Gunthorpe <jgunthorpe@obsidianresearch.com>
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: Anna Schumaker <Anna.Schumaker@Netapp.com>
2017-10-20 21:48:12 +07:00
|
|
|
unsigned int rep_send_count;
|
|
|
|
unsigned int rep_send_batch;
|
2007-09-11 00:50:12 +07:00
|
|
|
int rep_connected;
|
|
|
|
struct ib_qp_init_attr rep_attr;
|
|
|
|
wait_queue_head_t rep_connect_wait;
|
2016-09-15 21:57:07 +07:00
|
|
|
struct rpcrdma_connect_private rep_cm_private;
|
2007-09-11 00:50:12 +07:00
|
|
|
struct rdma_conn_param rep_remote_cma;
|
2014-05-28 21:32:17 +07:00
|
|
|
struct delayed_work rep_connect_worker;
|
2007-09-11 00:50:12 +07:00
|
|
|
};
|
|
|
|
|
2015-10-25 04:27:51 +07:00
|
|
|
/* Pre-allocate extra Work Requests for handling backward receives
|
|
|
|
* and sends. This is a fixed value because the Work Queues are
|
|
|
|
* allocated when the forward channel is set up.
|
|
|
|
*/
|
|
|
|
#if defined(CONFIG_SUNRPC_BACKCHANNEL)
|
|
|
|
#define RPCRDMA_BACKWARD_WRS (8)
|
|
|
|
#else
|
|
|
|
#define RPCRDMA_BACKWARD_WRS (0)
|
|
|
|
#endif
|
|
|
|
|
2015-01-21 23:04:00 +07:00
|
|
|
/* Registered buffer -- registered kmalloc'd memory for RDMA SEND/RECV
|
|
|
|
*
|
|
|
|
* The below structure appears at the front of a large region of kmalloc'd
|
|
|
|
* memory, which always starts on a good alignment boundary.
|
|
|
|
*/
|
|
|
|
|
|
|
|
struct rpcrdma_regbuf {
|
|
|
|
struct ib_sge rg_iov;
|
xprtrdma: Delay DMA mapping Send and Receive buffers
Currently, each regbuf is allocated and DMA mapped at the same time.
This is done during transport creation.
When a device driver is unloaded, every DMA-mapped buffer in use by
a transport has to be unmapped, and then remapped to the new
device if the driver is loaded again. Remapping will have to be done
_after_ the connect worker has set up the new device.
But there's an ordering problem:
call_allocate, which invokes xprt_rdma_allocate which calls
rpcrdma_alloc_regbuf to allocate Send buffers, happens _before_
the connect worker can run to set up the new device.
Instead, at transport creation, allocate each buffer, but leave it
unmapped. Once the RPC carries these buffers into ->send_request, by
which time a transport connection should have been established,
check to see that the RPC's buffers have been DMA mapped. If not,
map them there.
When device driver unplug support is added, it will simply unmap all
the transport's regbufs, but it doesn't have to deallocate the
underlying memory.
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: Anna Schumaker <Anna.Schumaker@Netapp.com>
2016-09-15 21:56:18 +07:00
|
|
|
struct ib_device *rg_device;
|
2016-09-15 21:56:10 +07:00
|
|
|
enum dma_data_direction rg_direction;
|
2015-01-21 23:04:00 +07:00
|
|
|
__be32 rg_base[0] __attribute__ ((aligned(256)));
|
|
|
|
};
|
|
|
|
|
|
|
|
static inline u64
|
|
|
|
rdmab_addr(struct rpcrdma_regbuf *rb)
|
|
|
|
{
|
|
|
|
return rb->rg_iov.addr;
|
|
|
|
}
|
|
|
|
|
|
|
|
static inline u32
|
|
|
|
rdmab_length(struct rpcrdma_regbuf *rb)
|
|
|
|
{
|
|
|
|
return rb->rg_iov.length;
|
|
|
|
}
|
|
|
|
|
|
|
|
static inline u32
|
|
|
|
rdmab_lkey(struct rpcrdma_regbuf *rb)
|
|
|
|
{
|
|
|
|
return rb->rg_iov.lkey;
|
|
|
|
}
|
|
|
|
|
2017-04-12 00:23:02 +07:00
|
|
|
static inline struct ib_device *
|
|
|
|
rdmab_device(struct rpcrdma_regbuf *rb)
|
|
|
|
{
|
|
|
|
return rb->rg_device;
|
|
|
|
}
|
|
|
|
|
2016-01-08 02:50:10 +07:00
|
|
|
#define RPCRDMA_DEF_GFP (GFP_NOIO | __GFP_NOWARN)
|
|
|
|
|
2016-05-03 01:40:56 +07:00
|
|
|
/* To ensure a transport can always make forward progress,
|
|
|
|
* the number of RDMA segments allowed in header chunk lists
|
|
|
|
* is capped at 8. This prevents less-capable devices and
|
|
|
|
* memory registrations from overrunning the Send buffer
|
|
|
|
* while building chunk lists.
|
|
|
|
*
|
|
|
|
* Elements of the Read list take up more room than the
|
|
|
|
* Write list or Reply chunk. 8 read segments means the Read
|
|
|
|
* list (or Write list or Reply chunk) cannot consume more
|
|
|
|
* than
|
|
|
|
*
|
|
|
|
* ((8 + 2) * read segment size) + 1 XDR words, or 244 bytes.
|
|
|
|
*
|
|
|
|
* And the fixed part of the header is another 24 bytes.
|
|
|
|
*
|
|
|
|
* The smallest inline threshold is 1024 bytes, ensuring that
|
|
|
|
* at least 750 bytes are available for RPC messages.
|
|
|
|
*/
|
2016-09-15 21:56:02 +07:00
|
|
|
enum {
|
|
|
|
RPCRDMA_MAX_HDR_SEGS = 8,
|
|
|
|
RPCRDMA_HDRBUF_SIZE = 256,
|
|
|
|
};
|
2016-05-03 01:40:56 +07:00
|
|
|
|
2007-09-11 00:50:12 +07:00
|
|
|
/*
|
2017-10-17 02:01:22 +07:00
|
|
|
* struct rpcrdma_rep -- this structure encapsulates state required
|
|
|
|
* to receive and complete an RPC Reply, asychronously. It needs
|
|
|
|
* several pieces of state:
|
2007-09-11 00:50:12 +07:00
|
|
|
*
|
2017-10-17 02:01:22 +07:00
|
|
|
* o receive buffer and ib_sge (donated to provider)
|
|
|
|
* o status of receive (success or not, length, inv rkey)
|
|
|
|
* o bookkeeping state to get run by reply handler (XDR stream)
|
2007-09-11 00:50:12 +07:00
|
|
|
*
|
2017-10-17 02:01:22 +07:00
|
|
|
* These structures are allocated during transport initialization.
|
|
|
|
* N of these are associated with a transport instance, managed by
|
|
|
|
* struct rpcrdma_buffer. N is the max number of outstanding RPCs.
|
2007-09-11 00:50:12 +07:00
|
|
|
*/
|
|
|
|
|
|
|
|
struct rpcrdma_rep {
|
2016-03-04 23:28:36 +07:00
|
|
|
struct ib_cqe rr_cqe;
|
2017-10-17 02:01:14 +07:00
|
|
|
__be32 rr_xid;
|
|
|
|
__be32 rr_vers;
|
|
|
|
__be32 rr_proc;
|
2016-09-15 21:57:16 +07:00
|
|
|
int rr_wc_flags;
|
|
|
|
u32 rr_inv_rkey;
|
2018-05-05 02:35:20 +07:00
|
|
|
bool rr_temp;
|
2017-08-04 01:30:52 +07:00
|
|
|
struct rpcrdma_regbuf *rr_rdmabuf;
|
2015-05-26 22:51:37 +07:00
|
|
|
struct rpcrdma_xprt *rr_rxprt;
|
2015-10-25 04:27:10 +07:00
|
|
|
struct work_struct rr_work;
|
2017-08-04 01:30:03 +07:00
|
|
|
struct xdr_buf rr_hdrbuf;
|
|
|
|
struct xdr_stream rr_stream;
|
2017-10-17 02:01:22 +07:00
|
|
|
struct rpc_rqst *rr_rqst;
|
2015-01-21 23:04:25 +07:00
|
|
|
struct list_head rr_list;
|
2016-09-15 21:56:51 +07:00
|
|
|
struct ib_recv_wr rr_recv_wr;
|
2007-09-11 00:50:12 +07:00
|
|
|
};
|
|
|
|
|
xprtrdma: Add data structure to manage RDMA Send arguments
Problem statement:
Recently Sagi Grimberg <sagi@grimberg.me> observed that kernel RDMA-
enabled storage initiators don't handle delayed Send completion
correctly. If Send completion is delayed beyond the end of a ULP
transaction, the ULP may release resources that are still being used
by the HCA to complete a long-running Send operation.
This is a common design trait amongst our initiators. Most Send
operations are faster than the ULP transaction they are part of.
Waiting for a completion for these is typically unnecessary.
Infrequently, a network partition or some other problem crops up
where an ordering problem can occur. In NFS parlance, the RPC Reply
arrives and completes the RPC, but the HCA is still retrying the
Send WR that conveyed the RPC Call. In this case, the HCA can try
to use memory that has been invalidated or DMA unmapped, and the
connection is lost. If that memory has been re-used for something
else (possibly not related to NFS), and the Send retransmission
exposes that data on the wire.
Thus we cannot assume that it is safe to release Send-related
resources just because a ULP reply has arrived.
After some analysis, we have determined that the completion
housekeeping will not be difficult for xprtrdma:
- Inline Send buffers are registered via the local DMA key, and
are already left DMA mapped for the lifetime of a transport
connection, thus no additional handling is necessary for those
- Gathered Sends involving page cache pages _will_ need to
DMA unmap those pages after the Send completes. But like
inline send buffers, they are registered via the local DMA key,
and thus will not need to be invalidated
In addition, RPC completion will need to wait for Send completion
in the latter case. However, nearly always, the Send that conveys
the RPC Call will have completed long before the RPC Reply
arrives, and thus no additional latency will be accrued.
Design notes:
In this patch, the rpcrdma_sendctx object is introduced, and a
lock-free circular queue is added to manage a set of them per
transport.
The RPC client's send path already prevents sending more than one
RPC Call at the same time. This allows us to treat the consumer
side of the queue (rpcrdma_sendctx_get_locked) as if there is a
single consumer thread.
The producer side of the queue (rpcrdma_sendctx_put_locked) is
invoked only from the Send completion handler, which is a single
thread of execution (soft IRQ).
The only care that needs to be taken is with the tail index, which
is shared between the producer and consumer. Only the producer
updates the tail index. The consumer compares the head with the
tail to ensure that the a sendctx that is in use is never handed
out again (or, expressed more conventionally, the queue is empty).
When the sendctx queue empties completely, there are enough Sends
outstanding that posting more Send operations can result in a Send
Queue overflow. In this case, the ULP is told to wait and try again.
This introduces strong Send Queue accounting to xprtrdma.
As a final touch, Jason Gunthorpe <jgunthorpe@obsidianresearch.com>
suggested a mechanism that does not require signaling every Send.
We signal once every N Sends, and perform SGE unmapping of N Send
operations during that one completion.
Reported-by: Sagi Grimberg <sagi@grimberg.me>
Suggested-by: Jason Gunthorpe <jgunthorpe@obsidianresearch.com>
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: Anna Schumaker <Anna.Schumaker@Netapp.com>
2017-10-20 21:48:12 +07:00
|
|
|
/* struct rpcrdma_sendctx - DMA mapped SGEs to unmap after Send completes
|
|
|
|
*/
|
2017-10-20 21:48:36 +07:00
|
|
|
struct rpcrdma_req;
|
xprtrdma: Add data structure to manage RDMA Send arguments
Problem statement:
Recently Sagi Grimberg <sagi@grimberg.me> observed that kernel RDMA-
enabled storage initiators don't handle delayed Send completion
correctly. If Send completion is delayed beyond the end of a ULP
transaction, the ULP may release resources that are still being used
by the HCA to complete a long-running Send operation.
This is a common design trait amongst our initiators. Most Send
operations are faster than the ULP transaction they are part of.
Waiting for a completion for these is typically unnecessary.
Infrequently, a network partition or some other problem crops up
where an ordering problem can occur. In NFS parlance, the RPC Reply
arrives and completes the RPC, but the HCA is still retrying the
Send WR that conveyed the RPC Call. In this case, the HCA can try
to use memory that has been invalidated or DMA unmapped, and the
connection is lost. If that memory has been re-used for something
else (possibly not related to NFS), and the Send retransmission
exposes that data on the wire.
Thus we cannot assume that it is safe to release Send-related
resources just because a ULP reply has arrived.
After some analysis, we have determined that the completion
housekeeping will not be difficult for xprtrdma:
- Inline Send buffers are registered via the local DMA key, and
are already left DMA mapped for the lifetime of a transport
connection, thus no additional handling is necessary for those
- Gathered Sends involving page cache pages _will_ need to
DMA unmap those pages after the Send completes. But like
inline send buffers, they are registered via the local DMA key,
and thus will not need to be invalidated
In addition, RPC completion will need to wait for Send completion
in the latter case. However, nearly always, the Send that conveys
the RPC Call will have completed long before the RPC Reply
arrives, and thus no additional latency will be accrued.
Design notes:
In this patch, the rpcrdma_sendctx object is introduced, and a
lock-free circular queue is added to manage a set of them per
transport.
The RPC client's send path already prevents sending more than one
RPC Call at the same time. This allows us to treat the consumer
side of the queue (rpcrdma_sendctx_get_locked) as if there is a
single consumer thread.
The producer side of the queue (rpcrdma_sendctx_put_locked) is
invoked only from the Send completion handler, which is a single
thread of execution (soft IRQ).
The only care that needs to be taken is with the tail index, which
is shared between the producer and consumer. Only the producer
updates the tail index. The consumer compares the head with the
tail to ensure that the a sendctx that is in use is never handed
out again (or, expressed more conventionally, the queue is empty).
When the sendctx queue empties completely, there are enough Sends
outstanding that posting more Send operations can result in a Send
Queue overflow. In this case, the ULP is told to wait and try again.
This introduces strong Send Queue accounting to xprtrdma.
As a final touch, Jason Gunthorpe <jgunthorpe@obsidianresearch.com>
suggested a mechanism that does not require signaling every Send.
We signal once every N Sends, and perform SGE unmapping of N Send
operations during that one completion.
Reported-by: Sagi Grimberg <sagi@grimberg.me>
Suggested-by: Jason Gunthorpe <jgunthorpe@obsidianresearch.com>
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: Anna Schumaker <Anna.Schumaker@Netapp.com>
2017-10-20 21:48:12 +07:00
|
|
|
struct rpcrdma_xprt;
|
|
|
|
struct rpcrdma_sendctx {
|
|
|
|
struct ib_send_wr sc_wr;
|
|
|
|
struct ib_cqe sc_cqe;
|
|
|
|
struct rpcrdma_xprt *sc_xprt;
|
2017-10-20 21:48:36 +07:00
|
|
|
struct rpcrdma_req *sc_req;
|
xprtrdma: Add data structure to manage RDMA Send arguments
Problem statement:
Recently Sagi Grimberg <sagi@grimberg.me> observed that kernel RDMA-
enabled storage initiators don't handle delayed Send completion
correctly. If Send completion is delayed beyond the end of a ULP
transaction, the ULP may release resources that are still being used
by the HCA to complete a long-running Send operation.
This is a common design trait amongst our initiators. Most Send
operations are faster than the ULP transaction they are part of.
Waiting for a completion for these is typically unnecessary.
Infrequently, a network partition or some other problem crops up
where an ordering problem can occur. In NFS parlance, the RPC Reply
arrives and completes the RPC, but the HCA is still retrying the
Send WR that conveyed the RPC Call. In this case, the HCA can try
to use memory that has been invalidated or DMA unmapped, and the
connection is lost. If that memory has been re-used for something
else (possibly not related to NFS), and the Send retransmission
exposes that data on the wire.
Thus we cannot assume that it is safe to release Send-related
resources just because a ULP reply has arrived.
After some analysis, we have determined that the completion
housekeeping will not be difficult for xprtrdma:
- Inline Send buffers are registered via the local DMA key, and
are already left DMA mapped for the lifetime of a transport
connection, thus no additional handling is necessary for those
- Gathered Sends involving page cache pages _will_ need to
DMA unmap those pages after the Send completes. But like
inline send buffers, they are registered via the local DMA key,
and thus will not need to be invalidated
In addition, RPC completion will need to wait for Send completion
in the latter case. However, nearly always, the Send that conveys
the RPC Call will have completed long before the RPC Reply
arrives, and thus no additional latency will be accrued.
Design notes:
In this patch, the rpcrdma_sendctx object is introduced, and a
lock-free circular queue is added to manage a set of them per
transport.
The RPC client's send path already prevents sending more than one
RPC Call at the same time. This allows us to treat the consumer
side of the queue (rpcrdma_sendctx_get_locked) as if there is a
single consumer thread.
The producer side of the queue (rpcrdma_sendctx_put_locked) is
invoked only from the Send completion handler, which is a single
thread of execution (soft IRQ).
The only care that needs to be taken is with the tail index, which
is shared between the producer and consumer. Only the producer
updates the tail index. The consumer compares the head with the
tail to ensure that the a sendctx that is in use is never handed
out again (or, expressed more conventionally, the queue is empty).
When the sendctx queue empties completely, there are enough Sends
outstanding that posting more Send operations can result in a Send
Queue overflow. In this case, the ULP is told to wait and try again.
This introduces strong Send Queue accounting to xprtrdma.
As a final touch, Jason Gunthorpe <jgunthorpe@obsidianresearch.com>
suggested a mechanism that does not require signaling every Send.
We signal once every N Sends, and perform SGE unmapping of N Send
operations during that one completion.
Reported-by: Sagi Grimberg <sagi@grimberg.me>
Suggested-by: Jason Gunthorpe <jgunthorpe@obsidianresearch.com>
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: Anna Schumaker <Anna.Schumaker@Netapp.com>
2017-10-20 21:48:12 +07:00
|
|
|
unsigned int sc_unmap_count;
|
|
|
|
struct ib_sge sc_sges[];
|
|
|
|
};
|
|
|
|
|
|
|
|
/* Limit the number of SGEs that can be unmapped during one
|
|
|
|
* Send completion. This caps the amount of work a single
|
|
|
|
* completion can do before returning to the provider.
|
|
|
|
*
|
|
|
|
* Setting this to zero disables Send completion batching.
|
|
|
|
*/
|
|
|
|
enum {
|
|
|
|
RPCRDMA_MAX_SEND_BATCH = 7,
|
|
|
|
};
|
|
|
|
|
2014-07-30 04:24:09 +07:00
|
|
|
/*
|
2017-12-15 08:57:55 +07:00
|
|
|
* struct rpcrdma_mr - external memory region metadata
|
2014-07-30 04:24:09 +07:00
|
|
|
*
|
|
|
|
* An external memory region is any buffer or page that is registered
|
|
|
|
* on the fly (ie, not pre-registered).
|
|
|
|
*
|
2017-12-15 08:57:55 +07:00
|
|
|
* Each rpcrdma_buffer has a list of free MWs anchored in rb_mrs. During
|
2014-07-30 04:24:09 +07:00
|
|
|
* call_allocate, rpcrdma_buffer_get() assigns one to each segment in
|
|
|
|
* an rpcrdma_req. Then rpcrdma_register_external() grabs these to keep
|
|
|
|
* track of registration metadata while each RPC is pending.
|
|
|
|
* rpcrdma_deregister_external() uses this metadata to unmap and
|
|
|
|
* release these resources when an RPC is complete.
|
|
|
|
*/
|
2017-12-15 08:57:47 +07:00
|
|
|
enum rpcrdma_frwr_state {
|
|
|
|
FRWR_IS_INVALID, /* ready to be used */
|
|
|
|
FRWR_IS_VALID, /* in use */
|
|
|
|
FRWR_FLUSHED_FR, /* flushed FASTREG WR */
|
|
|
|
FRWR_FLUSHED_LI, /* flushed LOCALINV WR */
|
2014-07-30 04:24:09 +07:00
|
|
|
};
|
|
|
|
|
2017-12-15 08:57:47 +07:00
|
|
|
struct rpcrdma_frwr {
|
2014-07-30 04:24:09 +07:00
|
|
|
struct ib_mr *fr_mr;
|
2016-03-04 23:28:53 +07:00
|
|
|
struct ib_cqe fr_cqe;
|
2017-12-15 08:57:47 +07:00
|
|
|
enum rpcrdma_frwr_state fr_state;
|
2016-03-04 23:28:53 +07:00
|
|
|
struct completion fr_linv_done;
|
2015-12-17 05:22:31 +07:00
|
|
|
union {
|
|
|
|
struct ib_reg_wr fr_regwr;
|
|
|
|
struct ib_send_wr fr_invwr;
|
|
|
|
};
|
2014-07-30 04:24:09 +07:00
|
|
|
};
|
|
|
|
|
2015-05-26 22:53:23 +07:00
|
|
|
struct rpcrdma_fmr {
|
2016-06-30 00:52:37 +07:00
|
|
|
struct ib_fmr *fm_mr;
|
|
|
|
u64 *fm_physaddrs;
|
2014-07-30 04:24:09 +07:00
|
|
|
};
|
|
|
|
|
2017-12-15 08:57:55 +07:00
|
|
|
struct rpcrdma_mr {
|
|
|
|
struct list_head mr_list;
|
|
|
|
struct scatterlist *mr_sg;
|
|
|
|
int mr_nents;
|
|
|
|
enum dma_data_direction mr_dir;
|
2014-07-30 04:24:09 +07:00
|
|
|
union {
|
2015-05-26 22:53:23 +07:00
|
|
|
struct rpcrdma_fmr fmr;
|
2017-12-15 08:57:47 +07:00
|
|
|
struct rpcrdma_frwr frwr;
|
2016-03-04 23:28:45 +07:00
|
|
|
};
|
2017-12-15 08:57:55 +07:00
|
|
|
struct rpcrdma_xprt *mr_xprt;
|
|
|
|
u32 mr_handle;
|
|
|
|
u32 mr_length;
|
|
|
|
u64 mr_offset;
|
|
|
|
struct list_head mr_all;
|
2014-07-30 04:24:09 +07:00
|
|
|
};
|
|
|
|
|
2007-09-11 00:50:12 +07:00
|
|
|
/*
|
|
|
|
* struct rpcrdma_req -- structure central to the request/reply sequence.
|
|
|
|
*
|
|
|
|
* N of these are associated with a transport instance, and stored in
|
|
|
|
* struct rpcrdma_buffer. N is the max number of outstanding requests.
|
|
|
|
*
|
|
|
|
* It includes pre-registered buffer memory for send AND recv.
|
|
|
|
* The recv buffer, however, is not owned by this structure, and
|
|
|
|
* is "donated" to the hardware when a recv is posted. When a
|
|
|
|
* reply is handled, the recv buffer used is given back to the
|
|
|
|
* struct rpcrdma_req associated with the request.
|
|
|
|
*
|
|
|
|
* In addition to the basic memory, this structure includes an array
|
|
|
|
* of iovs for send operations. The reason is that the iovs passed to
|
|
|
|
* ib_post_{send,recv} must not be modified until the work request
|
|
|
|
* completes.
|
|
|
|
*/
|
|
|
|
|
2016-06-30 00:54:25 +07:00
|
|
|
/* Maximum number of page-sized "segments" per chunk list to be
|
|
|
|
* registered or invalidated. Must handle a Reply chunk:
|
|
|
|
*/
|
|
|
|
enum {
|
|
|
|
RPCRDMA_MAX_IOV_SEGS = 3,
|
|
|
|
RPCRDMA_MAX_DATA_SEGS = ((1 * 1024 * 1024) / PAGE_SIZE) + 1,
|
|
|
|
RPCRDMA_MAX_SEGS = RPCRDMA_MAX_DATA_SEGS +
|
|
|
|
RPCRDMA_MAX_IOV_SEGS,
|
|
|
|
};
|
|
|
|
|
2007-09-11 00:50:12 +07:00
|
|
|
struct rpcrdma_mr_seg { /* chunk descriptors */
|
|
|
|
u32 mr_len; /* length of chunk or segment */
|
|
|
|
struct page *mr_page; /* owning page, if any */
|
|
|
|
char *mr_offset; /* kva if no page, else offset */
|
|
|
|
};
|
|
|
|
|
2017-02-09 05:00:18 +07:00
|
|
|
/* The Send SGE array is provisioned to send a maximum size
|
|
|
|
* inline request:
|
xprtrdma: Use gathered Send for large inline messages
An RPC Call message that is sent inline but that has a data payload
(ie, one or more items in rq_snd_buf's page list) must be "pulled
up:"
- call_allocate has to reserve enough RPC Call buffer space to
accommodate the data payload
- call_transmit has to memcopy the rq_snd_buf's page list and tail
into its head iovec before it is sent
As the inline threshold is increased beyond its current 1KB default,
however, this means data payloads of more than a few KB are copied
by the host CPU. For example, if the inline threshold is increased
just to 4KB, then NFS WRITE requests up to 4KB would involve a
memcpy of the NFS WRITE's payload data into the RPC Call buffer.
This is an undesirable amount of participation by the host CPU.
The inline threshold may be much larger than 4KB in the future,
after negotiation with a peer server.
Instead of copying the components of rq_snd_buf into its head iovec,
construct a gather list of these components, and send them all in
place. The same approach is already used in the Linux server's
RPC-over-RDMA reply path.
This mechanism also eliminates the need for rpcrdma_tail_pullup,
which is used to manage the XDR pad and trailing inline content when
a Read list is present.
This requires that the pages in rq_snd_buf's page list be DMA-mapped
during marshaling, and unmapped when a data-bearing RPC is
completed. This is slightly less efficient for very small I/O
payloads, but significantly more efficient as data payload size and
inline threshold increase past a kilobyte.
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: Anna Schumaker <Anna.Schumaker@Netapp.com>
2016-09-15 21:57:24 +07:00
|
|
|
* - RPC-over-RDMA header
|
|
|
|
* - xdr_buf head iovec
|
2017-02-09 05:00:18 +07:00
|
|
|
* - RPCRDMA_MAX_INLINE bytes, in pages
|
xprtrdma: Use gathered Send for large inline messages
An RPC Call message that is sent inline but that has a data payload
(ie, one or more items in rq_snd_buf's page list) must be "pulled
up:"
- call_allocate has to reserve enough RPC Call buffer space to
accommodate the data payload
- call_transmit has to memcopy the rq_snd_buf's page list and tail
into its head iovec before it is sent
As the inline threshold is increased beyond its current 1KB default,
however, this means data payloads of more than a few KB are copied
by the host CPU. For example, if the inline threshold is increased
just to 4KB, then NFS WRITE requests up to 4KB would involve a
memcpy of the NFS WRITE's payload data into the RPC Call buffer.
This is an undesirable amount of participation by the host CPU.
The inline threshold may be much larger than 4KB in the future,
after negotiation with a peer server.
Instead of copying the components of rq_snd_buf into its head iovec,
construct a gather list of these components, and send them all in
place. The same approach is already used in the Linux server's
RPC-over-RDMA reply path.
This mechanism also eliminates the need for rpcrdma_tail_pullup,
which is used to manage the XDR pad and trailing inline content when
a Read list is present.
This requires that the pages in rq_snd_buf's page list be DMA-mapped
during marshaling, and unmapped when a data-bearing RPC is
completed. This is slightly less efficient for very small I/O
payloads, but significantly more efficient as data payload size and
inline threshold increase past a kilobyte.
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: Anna Schumaker <Anna.Schumaker@Netapp.com>
2016-09-15 21:57:24 +07:00
|
|
|
* - xdr_buf tail iovec
|
2017-02-09 05:00:18 +07:00
|
|
|
*
|
|
|
|
* The actual number of array elements consumed by each RPC
|
|
|
|
* depends on the device's max_sge limit.
|
xprtrdma: Use gathered Send for large inline messages
An RPC Call message that is sent inline but that has a data payload
(ie, one or more items in rq_snd_buf's page list) must be "pulled
up:"
- call_allocate has to reserve enough RPC Call buffer space to
accommodate the data payload
- call_transmit has to memcopy the rq_snd_buf's page list and tail
into its head iovec before it is sent
As the inline threshold is increased beyond its current 1KB default,
however, this means data payloads of more than a few KB are copied
by the host CPU. For example, if the inline threshold is increased
just to 4KB, then NFS WRITE requests up to 4KB would involve a
memcpy of the NFS WRITE's payload data into the RPC Call buffer.
This is an undesirable amount of participation by the host CPU.
The inline threshold may be much larger than 4KB in the future,
after negotiation with a peer server.
Instead of copying the components of rq_snd_buf into its head iovec,
construct a gather list of these components, and send them all in
place. The same approach is already used in the Linux server's
RPC-over-RDMA reply path.
This mechanism also eliminates the need for rpcrdma_tail_pullup,
which is used to manage the XDR pad and trailing inline content when
a Read list is present.
This requires that the pages in rq_snd_buf's page list be DMA-mapped
during marshaling, and unmapped when a data-bearing RPC is
completed. This is slightly less efficient for very small I/O
payloads, but significantly more efficient as data payload size and
inline threshold increase past a kilobyte.
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: Anna Schumaker <Anna.Schumaker@Netapp.com>
2016-09-15 21:57:24 +07:00
|
|
|
*/
|
|
|
|
enum {
|
2017-02-09 05:00:10 +07:00
|
|
|
RPCRDMA_MIN_SEND_SGES = 3,
|
2017-02-09 05:00:18 +07:00
|
|
|
RPCRDMA_MAX_PAGE_SGES = RPCRDMA_MAX_INLINE >> PAGE_SHIFT,
|
xprtrdma: Use gathered Send for large inline messages
An RPC Call message that is sent inline but that has a data payload
(ie, one or more items in rq_snd_buf's page list) must be "pulled
up:"
- call_allocate has to reserve enough RPC Call buffer space to
accommodate the data payload
- call_transmit has to memcopy the rq_snd_buf's page list and tail
into its head iovec before it is sent
As the inline threshold is increased beyond its current 1KB default,
however, this means data payloads of more than a few KB are copied
by the host CPU. For example, if the inline threshold is increased
just to 4KB, then NFS WRITE requests up to 4KB would involve a
memcpy of the NFS WRITE's payload data into the RPC Call buffer.
This is an undesirable amount of participation by the host CPU.
The inline threshold may be much larger than 4KB in the future,
after negotiation with a peer server.
Instead of copying the components of rq_snd_buf into its head iovec,
construct a gather list of these components, and send them all in
place. The same approach is already used in the Linux server's
RPC-over-RDMA reply path.
This mechanism also eliminates the need for rpcrdma_tail_pullup,
which is used to manage the XDR pad and trailing inline content when
a Read list is present.
This requires that the pages in rq_snd_buf's page list be DMA-mapped
during marshaling, and unmapped when a data-bearing RPC is
completed. This is slightly less efficient for very small I/O
payloads, but significantly more efficient as data payload size and
inline threshold increase past a kilobyte.
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: Anna Schumaker <Anna.Schumaker@Netapp.com>
2016-09-15 21:57:24 +07:00
|
|
|
RPCRDMA_MAX_SEND_SGES = 1 + 1 + RPCRDMA_MAX_PAGE_SGES + 1,
|
|
|
|
};
|
2015-08-04 00:03:39 +07:00
|
|
|
|
2016-06-30 00:54:25 +07:00
|
|
|
struct rpcrdma_buffer;
|
2007-09-11 00:50:12 +07:00
|
|
|
struct rpcrdma_req {
|
2017-06-08 22:52:12 +07:00
|
|
|
struct list_head rl_list;
|
2018-05-05 02:35:09 +07:00
|
|
|
struct rpc_rqst rl_slot;
|
2015-08-04 00:03:39 +07:00
|
|
|
struct rpcrdma_buffer *rl_buffer;
|
2016-09-15 21:56:43 +07:00
|
|
|
struct rpcrdma_rep *rl_reply;
|
2017-08-10 23:47:28 +07:00
|
|
|
struct xdr_stream rl_stream;
|
|
|
|
struct xdr_buf rl_hdrbuf;
|
xprtrdma: Add data structure to manage RDMA Send arguments
Problem statement:
Recently Sagi Grimberg <sagi@grimberg.me> observed that kernel RDMA-
enabled storage initiators don't handle delayed Send completion
correctly. If Send completion is delayed beyond the end of a ULP
transaction, the ULP may release resources that are still being used
by the HCA to complete a long-running Send operation.
This is a common design trait amongst our initiators. Most Send
operations are faster than the ULP transaction they are part of.
Waiting for a completion for these is typically unnecessary.
Infrequently, a network partition or some other problem crops up
where an ordering problem can occur. In NFS parlance, the RPC Reply
arrives and completes the RPC, but the HCA is still retrying the
Send WR that conveyed the RPC Call. In this case, the HCA can try
to use memory that has been invalidated or DMA unmapped, and the
connection is lost. If that memory has been re-used for something
else (possibly not related to NFS), and the Send retransmission
exposes that data on the wire.
Thus we cannot assume that it is safe to release Send-related
resources just because a ULP reply has arrived.
After some analysis, we have determined that the completion
housekeeping will not be difficult for xprtrdma:
- Inline Send buffers are registered via the local DMA key, and
are already left DMA mapped for the lifetime of a transport
connection, thus no additional handling is necessary for those
- Gathered Sends involving page cache pages _will_ need to
DMA unmap those pages after the Send completes. But like
inline send buffers, they are registered via the local DMA key,
and thus will not need to be invalidated
In addition, RPC completion will need to wait for Send completion
in the latter case. However, nearly always, the Send that conveys
the RPC Call will have completed long before the RPC Reply
arrives, and thus no additional latency will be accrued.
Design notes:
In this patch, the rpcrdma_sendctx object is introduced, and a
lock-free circular queue is added to manage a set of them per
transport.
The RPC client's send path already prevents sending more than one
RPC Call at the same time. This allows us to treat the consumer
side of the queue (rpcrdma_sendctx_get_locked) as if there is a
single consumer thread.
The producer side of the queue (rpcrdma_sendctx_put_locked) is
invoked only from the Send completion handler, which is a single
thread of execution (soft IRQ).
The only care that needs to be taken is with the tail index, which
is shared between the producer and consumer. Only the producer
updates the tail index. The consumer compares the head with the
tail to ensure that the a sendctx that is in use is never handed
out again (or, expressed more conventionally, the queue is empty).
When the sendctx queue empties completely, there are enough Sends
outstanding that posting more Send operations can result in a Send
Queue overflow. In this case, the ULP is told to wait and try again.
This introduces strong Send Queue accounting to xprtrdma.
As a final touch, Jason Gunthorpe <jgunthorpe@obsidianresearch.com>
suggested a mechanism that does not require signaling every Send.
We signal once every N Sends, and perform SGE unmapping of N Send
operations during that one completion.
Reported-by: Sagi Grimberg <sagi@grimberg.me>
Suggested-by: Jason Gunthorpe <jgunthorpe@obsidianresearch.com>
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: Anna Schumaker <Anna.Schumaker@Netapp.com>
2017-10-20 21:48:12 +07:00
|
|
|
struct rpcrdma_sendctx *rl_sendctx;
|
xprtrdma: Initialize separate RPC call and reply buffers
RPC-over-RDMA needs to separate its RPC call and reply buffers.
o When an RPC Call is sent, rq_snd_buf is DMA mapped for an RDMA
Send operation using DMA_TO_DEVICE
o If the client expects a large RPC reply, it DMA maps rq_rcv_buf
as part of a Reply chunk using DMA_FROM_DEVICE
The two mappings are for data movement in opposite directions.
DMA-API.txt suggests that if these mappings share a DMA cacheline,
bad things can happen. This could occur in the final bytes of
rq_snd_buf and the first bytes of rq_rcv_buf if the two buffers
happen to share a DMA cacheline.
On x86_64 the cacheline size is typically 8 bytes, and RPC call
messages are usually much smaller than the send buffer, so this
hasn't been a noticeable problem. But the DMA cacheline size can be
larger on other platforms.
Also, often rq_rcv_buf starts most of the way into a page, thus
an additional RDMA segment is needed to map and register the end of
that buffer. Try to avoid that scenario to reduce the cost of
registering and invalidating Reply chunks.
Instead of carrying a single regbuf that covers both rq_snd_buf and
rq_rcv_buf, each struct rpcrdma_req now carries one regbuf for
rq_snd_buf and one regbuf for rq_rcv_buf.
Some incidental changes worth noting:
- To clear out some spaghetti, refactor xprt_rdma_allocate.
- The value stored in rg_size is the same as the value stored in
the iov.length field, so eliminate rg_size
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: Anna Schumaker <Anna.Schumaker@Netapp.com>
2016-09-15 21:55:53 +07:00
|
|
|
struct rpcrdma_regbuf *rl_rdmabuf; /* xprt header */
|
|
|
|
struct rpcrdma_regbuf *rl_sendbuf; /* rq_snd_buf */
|
|
|
|
struct rpcrdma_regbuf *rl_recvbuf; /* rq_rcv_buf */
|
2015-10-25 04:27:43 +07:00
|
|
|
|
|
|
|
struct list_head rl_all;
|
2017-10-20 21:48:20 +07:00
|
|
|
unsigned long rl_flags;
|
2016-06-30 00:54:25 +07:00
|
|
|
|
|
|
|
struct list_head rl_registered; /* registered segments */
|
|
|
|
struct rpcrdma_mr_seg rl_segments[RPCRDMA_MAX_SEGS];
|
2007-09-11 00:50:12 +07:00
|
|
|
};
|
2015-01-21 23:04:08 +07:00
|
|
|
|
2017-10-20 21:48:20 +07:00
|
|
|
/* rl_flags */
|
|
|
|
enum {
|
2017-12-15 08:57:23 +07:00
|
|
|
RPCRDMA_REQ_F_PENDING = 0,
|
2017-10-20 21:48:36 +07:00
|
|
|
RPCRDMA_REQ_F_TX_RESOURCES,
|
2017-10-20 21:48:20 +07:00
|
|
|
};
|
|
|
|
|
2015-01-21 23:04:08 +07:00
|
|
|
static inline struct rpcrdma_req *
|
2017-12-21 04:31:37 +07:00
|
|
|
rpcr_to_rdmar(const struct rpc_rqst *rqst)
|
2015-01-21 23:04:08 +07:00
|
|
|
{
|
2018-05-05 02:35:09 +07:00
|
|
|
return container_of(rqst, struct rpcrdma_req, rl_slot);
|
2015-01-21 23:04:08 +07:00
|
|
|
}
|
2007-09-11 00:50:12 +07:00
|
|
|
|
2017-02-09 05:00:43 +07:00
|
|
|
static inline void
|
2017-12-15 08:57:55 +07:00
|
|
|
rpcrdma_mr_push(struct rpcrdma_mr *mr, struct list_head *list)
|
2017-02-09 05:00:43 +07:00
|
|
|
{
|
2017-12-15 08:57:55 +07:00
|
|
|
list_add_tail(&mr->mr_list, list);
|
2017-02-09 05:00:43 +07:00
|
|
|
}
|
|
|
|
|
2017-12-15 08:57:55 +07:00
|
|
|
static inline struct rpcrdma_mr *
|
|
|
|
rpcrdma_mr_pop(struct list_head *list)
|
2017-02-09 05:00:43 +07:00
|
|
|
{
|
2017-12-15 08:57:55 +07:00
|
|
|
struct rpcrdma_mr *mr;
|
2017-02-09 05:00:43 +07:00
|
|
|
|
2017-12-15 08:57:55 +07:00
|
|
|
mr = list_first_entry(list, struct rpcrdma_mr, mr_list);
|
xprtrdma: Fix list corruption / DMAR errors during MR recovery
The ro_release_mr methods check whether mr->mr_list is empty.
Therefore, be sure to always use list_del_init when removing an MR
linked into a list using that field. Otherwise, when recovering from
transport failures or device removal, list corruption can result, or
MRs can get mapped or unmapped an odd number of times, resulting in
IOMMU-related failures.
In general this fix is appropriate back to v4.8. However, code
changes since then make it impossible to apply this patch directly
to stable kernels. The fix would have to be applied by hand or
reworked for kernels earlier than v4.16.
Backport guidance -- there are several cases:
- When creating an MR, initialize mr_list so that using list_empty
on an as-yet-unused MR is safe.
- When an MR is being handled by the remote invalidation path,
ensure that mr_list is reinitialized when it is removed from
rl_registered.
- When an MR is being handled by rpcrdma_destroy_mrs, it is removed
from mr_all, but it may still be on an rl_registered list. In
that case, the MR needs to be removed from that list before being
released.
- Other cases are covered by using list_del_init in rpcrdma_mr_pop.
Fixes: 9d6b04097882 ('xprtrdma: Place registered MWs on a ... ')
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: Anna Schumaker <Anna.Schumaker@Netapp.com>
2018-05-01 22:37:14 +07:00
|
|
|
list_del_init(&mr->mr_list);
|
2017-12-15 08:57:55 +07:00
|
|
|
return mr;
|
2017-02-09 05:00:43 +07:00
|
|
|
}
|
|
|
|
|
2007-09-11 00:50:12 +07:00
|
|
|
/*
|
|
|
|
* struct rpcrdma_buffer -- holds list/queue of pre-registered memory for
|
|
|
|
* inline requests/replies, and client/server credits.
|
|
|
|
*
|
|
|
|
* One of these is associated with a transport instance
|
|
|
|
*/
|
|
|
|
struct rpcrdma_buffer {
|
2017-12-15 08:57:55 +07:00
|
|
|
spinlock_t rb_mrlock; /* protect rb_mrs list */
|
|
|
|
struct list_head rb_mrs;
|
2015-05-26 22:53:13 +07:00
|
|
|
struct list_head rb_all;
|
|
|
|
|
xprtrdma: Add data structure to manage RDMA Send arguments
Problem statement:
Recently Sagi Grimberg <sagi@grimberg.me> observed that kernel RDMA-
enabled storage initiators don't handle delayed Send completion
correctly. If Send completion is delayed beyond the end of a ULP
transaction, the ULP may release resources that are still being used
by the HCA to complete a long-running Send operation.
This is a common design trait amongst our initiators. Most Send
operations are faster than the ULP transaction they are part of.
Waiting for a completion for these is typically unnecessary.
Infrequently, a network partition or some other problem crops up
where an ordering problem can occur. In NFS parlance, the RPC Reply
arrives and completes the RPC, but the HCA is still retrying the
Send WR that conveyed the RPC Call. In this case, the HCA can try
to use memory that has been invalidated or DMA unmapped, and the
connection is lost. If that memory has been re-used for something
else (possibly not related to NFS), and the Send retransmission
exposes that data on the wire.
Thus we cannot assume that it is safe to release Send-related
resources just because a ULP reply has arrived.
After some analysis, we have determined that the completion
housekeeping will not be difficult for xprtrdma:
- Inline Send buffers are registered via the local DMA key, and
are already left DMA mapped for the lifetime of a transport
connection, thus no additional handling is necessary for those
- Gathered Sends involving page cache pages _will_ need to
DMA unmap those pages after the Send completes. But like
inline send buffers, they are registered via the local DMA key,
and thus will not need to be invalidated
In addition, RPC completion will need to wait for Send completion
in the latter case. However, nearly always, the Send that conveys
the RPC Call will have completed long before the RPC Reply
arrives, and thus no additional latency will be accrued.
Design notes:
In this patch, the rpcrdma_sendctx object is introduced, and a
lock-free circular queue is added to manage a set of them per
transport.
The RPC client's send path already prevents sending more than one
RPC Call at the same time. This allows us to treat the consumer
side of the queue (rpcrdma_sendctx_get_locked) as if there is a
single consumer thread.
The producer side of the queue (rpcrdma_sendctx_put_locked) is
invoked only from the Send completion handler, which is a single
thread of execution (soft IRQ).
The only care that needs to be taken is with the tail index, which
is shared between the producer and consumer. Only the producer
updates the tail index. The consumer compares the head with the
tail to ensure that the a sendctx that is in use is never handed
out again (or, expressed more conventionally, the queue is empty).
When the sendctx queue empties completely, there are enough Sends
outstanding that posting more Send operations can result in a Send
Queue overflow. In this case, the ULP is told to wait and try again.
This introduces strong Send Queue accounting to xprtrdma.
As a final touch, Jason Gunthorpe <jgunthorpe@obsidianresearch.com>
suggested a mechanism that does not require signaling every Send.
We signal once every N Sends, and perform SGE unmapping of N Send
operations during that one completion.
Reported-by: Sagi Grimberg <sagi@grimberg.me>
Suggested-by: Jason Gunthorpe <jgunthorpe@obsidianresearch.com>
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: Anna Schumaker <Anna.Schumaker@Netapp.com>
2017-10-20 21:48:12 +07:00
|
|
|
unsigned long rb_sc_head;
|
|
|
|
unsigned long rb_sc_tail;
|
|
|
|
unsigned long rb_sc_last;
|
|
|
|
struct rpcrdma_sendctx **rb_sc_ctxs;
|
|
|
|
|
2015-10-25 04:27:02 +07:00
|
|
|
spinlock_t rb_lock; /* protect buf lists */
|
|
|
|
struct list_head rb_send_bufs;
|
|
|
|
struct list_head rb_recv_bufs;
|
2018-05-05 02:35:57 +07:00
|
|
|
unsigned long rb_flags;
|
2015-05-26 22:53:13 +07:00
|
|
|
u32 rb_max_requests;
|
2017-10-17 02:01:39 +07:00
|
|
|
u32 rb_credits; /* most recent credit grant */
|
2018-05-05 02:35:20 +07:00
|
|
|
int rb_posted_receives;
|
2015-10-25 04:27:43 +07:00
|
|
|
|
|
|
|
u32 rb_bc_srv_max_requests;
|
|
|
|
spinlock_t rb_reqslock; /* protect rb_allreqs */
|
|
|
|
struct list_head rb_allreqs;
|
2016-01-08 02:50:10 +07:00
|
|
|
|
|
|
|
u32 rb_bc_max_requests;
|
2016-06-30 00:52:54 +07:00
|
|
|
|
|
|
|
spinlock_t rb_recovery_lock; /* protect rb_stale_mrs */
|
|
|
|
struct list_head rb_stale_mrs;
|
|
|
|
struct delayed_work rb_recovery_worker;
|
2016-06-30 00:54:00 +07:00
|
|
|
struct delayed_work rb_refresh_worker;
|
2007-09-11 00:50:12 +07:00
|
|
|
};
|
|
|
|
#define rdmab_to_ia(b) (&container_of((b), struct rpcrdma_xprt, rx_buf)->rx_ia)
|
|
|
|
|
2018-05-05 02:35:57 +07:00
|
|
|
/* rb_flags */
|
|
|
|
enum {
|
|
|
|
RPCRDMA_BUF_F_EMPTY_SCQ = 0,
|
|
|
|
};
|
|
|
|
|
2007-09-11 00:50:12 +07:00
|
|
|
/*
|
|
|
|
* Internal structure for transport instance creation. This
|
|
|
|
* exists primarily for modularity.
|
|
|
|
*
|
|
|
|
* This data should be set with mount options
|
|
|
|
*/
|
|
|
|
struct rpcrdma_create_data_internal {
|
|
|
|
unsigned int max_requests; /* max requests (slots) in flight */
|
|
|
|
unsigned int rsize; /* mount rsize - max read hdr+data */
|
|
|
|
unsigned int wsize; /* mount wsize - max write hdr+data */
|
|
|
|
unsigned int inline_rsize; /* max non-rdma read data payload */
|
|
|
|
unsigned int inline_wsize; /* max non-rdma write data payload */
|
|
|
|
};
|
|
|
|
|
|
|
|
/*
|
|
|
|
* Statistics for RPCRDMA
|
|
|
|
*/
|
|
|
|
struct rpcrdma_stats {
|
2017-08-22 22:19:35 +07:00
|
|
|
/* accessed when sending a call */
|
2007-09-11 00:50:12 +07:00
|
|
|
unsigned long read_chunk_count;
|
|
|
|
unsigned long write_chunk_count;
|
|
|
|
unsigned long reply_chunk_count;
|
|
|
|
unsigned long long total_rdma_request;
|
|
|
|
|
2017-08-22 22:19:35 +07:00
|
|
|
/* rarely accessed error counters */
|
2007-09-11 00:50:12 +07:00
|
|
|
unsigned long long pullup_copy_count;
|
|
|
|
unsigned long hardway_register_count;
|
|
|
|
unsigned long failed_marshal_count;
|
|
|
|
unsigned long bad_reply_count;
|
2016-06-30 00:52:54 +07:00
|
|
|
unsigned long mrs_recovered;
|
|
|
|
unsigned long mrs_orphaned;
|
2016-06-30 00:54:00 +07:00
|
|
|
unsigned long mrs_allocated;
|
xprtrdma: Add data structure to manage RDMA Send arguments
Problem statement:
Recently Sagi Grimberg <sagi@grimberg.me> observed that kernel RDMA-
enabled storage initiators don't handle delayed Send completion
correctly. If Send completion is delayed beyond the end of a ULP
transaction, the ULP may release resources that are still being used
by the HCA to complete a long-running Send operation.
This is a common design trait amongst our initiators. Most Send
operations are faster than the ULP transaction they are part of.
Waiting for a completion for these is typically unnecessary.
Infrequently, a network partition or some other problem crops up
where an ordering problem can occur. In NFS parlance, the RPC Reply
arrives and completes the RPC, but the HCA is still retrying the
Send WR that conveyed the RPC Call. In this case, the HCA can try
to use memory that has been invalidated or DMA unmapped, and the
connection is lost. If that memory has been re-used for something
else (possibly not related to NFS), and the Send retransmission
exposes that data on the wire.
Thus we cannot assume that it is safe to release Send-related
resources just because a ULP reply has arrived.
After some analysis, we have determined that the completion
housekeeping will not be difficult for xprtrdma:
- Inline Send buffers are registered via the local DMA key, and
are already left DMA mapped for the lifetime of a transport
connection, thus no additional handling is necessary for those
- Gathered Sends involving page cache pages _will_ need to
DMA unmap those pages after the Send completes. But like
inline send buffers, they are registered via the local DMA key,
and thus will not need to be invalidated
In addition, RPC completion will need to wait for Send completion
in the latter case. However, nearly always, the Send that conveys
the RPC Call will have completed long before the RPC Reply
arrives, and thus no additional latency will be accrued.
Design notes:
In this patch, the rpcrdma_sendctx object is introduced, and a
lock-free circular queue is added to manage a set of them per
transport.
The RPC client's send path already prevents sending more than one
RPC Call at the same time. This allows us to treat the consumer
side of the queue (rpcrdma_sendctx_get_locked) as if there is a
single consumer thread.
The producer side of the queue (rpcrdma_sendctx_put_locked) is
invoked only from the Send completion handler, which is a single
thread of execution (soft IRQ).
The only care that needs to be taken is with the tail index, which
is shared between the producer and consumer. Only the producer
updates the tail index. The consumer compares the head with the
tail to ensure that the a sendctx that is in use is never handed
out again (or, expressed more conventionally, the queue is empty).
When the sendctx queue empties completely, there are enough Sends
outstanding that posting more Send operations can result in a Send
Queue overflow. In this case, the ULP is told to wait and try again.
This introduces strong Send Queue accounting to xprtrdma.
As a final touch, Jason Gunthorpe <jgunthorpe@obsidianresearch.com>
suggested a mechanism that does not require signaling every Send.
We signal once every N Sends, and perform SGE unmapping of N Send
operations during that one completion.
Reported-by: Sagi Grimberg <sagi@grimberg.me>
Suggested-by: Jason Gunthorpe <jgunthorpe@obsidianresearch.com>
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: Anna Schumaker <Anna.Schumaker@Netapp.com>
2017-10-20 21:48:12 +07:00
|
|
|
unsigned long empty_sendctx_q;
|
2017-08-22 22:19:35 +07:00
|
|
|
|
|
|
|
/* accessed when receiving a reply */
|
|
|
|
unsigned long long total_rdma_reply;
|
|
|
|
unsigned long long fixup_copy_count;
|
2017-10-20 21:48:36 +07:00
|
|
|
unsigned long reply_waits_for_send;
|
2016-09-15 21:57:16 +07:00
|
|
|
unsigned long local_inv_needed;
|
2017-08-22 22:19:35 +07:00
|
|
|
unsigned long nomsg_call_count;
|
|
|
|
unsigned long bcall_count;
|
2007-09-11 00:50:12 +07:00
|
|
|
};
|
|
|
|
|
2015-03-31 01:34:21 +07:00
|
|
|
/*
|
|
|
|
* Per-registration mode operations
|
|
|
|
*/
|
2015-03-31 01:34:30 +07:00
|
|
|
struct rpcrdma_xprt;
|
2015-03-31 01:34:21 +07:00
|
|
|
struct rpcrdma_memreg_ops {
|
2017-08-15 02:38:30 +07:00
|
|
|
struct rpcrdma_mr_seg *
|
|
|
|
(*ro_map)(struct rpcrdma_xprt *,
|
2016-06-30 00:54:16 +07:00
|
|
|
struct rpcrdma_mr_seg *, int, bool,
|
2017-12-15 08:57:55 +07:00
|
|
|
struct rpcrdma_mr **);
|
2018-03-01 03:30:59 +07:00
|
|
|
int (*ro_send)(struct rpcrdma_ia *ia,
|
|
|
|
struct rpcrdma_req *req);
|
2017-12-15 08:56:26 +07:00
|
|
|
void (*ro_reminv)(struct rpcrdma_rep *rep,
|
2017-12-15 08:57:55 +07:00
|
|
|
struct list_head *mrs);
|
2015-12-17 05:22:39 +07:00
|
|
|
void (*ro_unmap_sync)(struct rpcrdma_xprt *,
|
2017-06-08 22:52:04 +07:00
|
|
|
struct list_head *);
|
2017-12-15 08:57:55 +07:00
|
|
|
void (*ro_recover_mr)(struct rpcrdma_mr *mr);
|
2015-03-31 01:35:26 +07:00
|
|
|
int (*ro_open)(struct rpcrdma_ia *,
|
|
|
|
struct rpcrdma_ep *,
|
|
|
|
struct rpcrdma_create_data_internal *);
|
2015-03-31 01:34:30 +07:00
|
|
|
size_t (*ro_maxpages)(struct rpcrdma_xprt *);
|
2016-06-30 00:54:00 +07:00
|
|
|
int (*ro_init_mr)(struct rpcrdma_ia *,
|
2017-12-15 08:57:55 +07:00
|
|
|
struct rpcrdma_mr *);
|
|
|
|
void (*ro_release_mr)(struct rpcrdma_mr *mr);
|
2015-03-31 01:34:21 +07:00
|
|
|
const char *ro_displayname;
|
2016-09-15 21:57:16 +07:00
|
|
|
const int ro_send_w_inv_ok;
|
2015-03-31 01:34:21 +07:00
|
|
|
};
|
|
|
|
|
|
|
|
extern const struct rpcrdma_memreg_ops rpcrdma_fmr_memreg_ops;
|
|
|
|
extern const struct rpcrdma_memreg_ops rpcrdma_frwr_memreg_ops;
|
|
|
|
|
2007-09-11 00:50:12 +07:00
|
|
|
/*
|
|
|
|
* RPCRDMA transport -- encapsulates the structures above for
|
|
|
|
* integration with RPC.
|
|
|
|
*
|
|
|
|
* The contained structures are embedded, not pointers,
|
|
|
|
* for convenience. This structure need not be visible externally.
|
|
|
|
*
|
|
|
|
* It is allocated and initialized during mount, and released
|
|
|
|
* during unmount.
|
|
|
|
*/
|
|
|
|
struct rpcrdma_xprt {
|
2015-01-21 23:02:37 +07:00
|
|
|
struct rpc_xprt rx_xprt;
|
2007-09-11 00:50:12 +07:00
|
|
|
struct rpcrdma_ia rx_ia;
|
|
|
|
struct rpcrdma_ep rx_ep;
|
|
|
|
struct rpcrdma_buffer rx_buf;
|
|
|
|
struct rpcrdma_create_data_internal rx_data;
|
2015-01-21 23:02:37 +07:00
|
|
|
struct delayed_work rx_connect_worker;
|
2007-09-11 00:50:12 +07:00
|
|
|
struct rpcrdma_stats rx_stats;
|
|
|
|
};
|
|
|
|
|
2015-01-21 23:02:37 +07:00
|
|
|
#define rpcx_to_rdmax(x) container_of(x, struct rpcrdma_xprt, rx_xprt)
|
2007-09-11 00:50:12 +07:00
|
|
|
#define rpcx_to_rdmad(x) (rpcx_to_rdmax(x)->rx_data)
|
|
|
|
|
2017-12-15 08:56:50 +07:00
|
|
|
static inline const char *
|
|
|
|
rpcrdma_addrstr(const struct rpcrdma_xprt *r_xprt)
|
|
|
|
{
|
|
|
|
return r_xprt->rx_xprt.address_strings[RPC_DISPLAY_ADDR];
|
|
|
|
}
|
|
|
|
|
|
|
|
static inline const char *
|
|
|
|
rpcrdma_portstr(const struct rpcrdma_xprt *r_xprt)
|
|
|
|
{
|
|
|
|
return r_xprt->rx_xprt.address_strings[RPC_DISPLAY_PORT];
|
|
|
|
}
|
|
|
|
|
2008-10-10 02:01:11 +07:00
|
|
|
/* Setting this to 0 ensures interoperability with early servers.
|
|
|
|
* Setting this to 1 enhances certain unaligned read/write performance.
|
|
|
|
* Default is 0, see sysctl entry and rpc_rdma.c rpcrdma_convert_iovs() */
|
|
|
|
extern int xprt_rdma_pad_optimize;
|
|
|
|
|
2017-04-12 00:22:54 +07:00
|
|
|
/* This setting controls the hunt for a supported memory
|
|
|
|
* registration strategy.
|
|
|
|
*/
|
|
|
|
extern unsigned int xprt_rdma_memreg_strategy;
|
|
|
|
|
2007-09-11 00:50:12 +07:00
|
|
|
/*
|
|
|
|
* Interface Adapter calls - xprtrdma/verbs.c
|
|
|
|
*/
|
2017-12-15 08:56:58 +07:00
|
|
|
int rpcrdma_ia_open(struct rpcrdma_xprt *xprt);
|
2017-04-12 00:23:10 +07:00
|
|
|
void rpcrdma_ia_remove(struct rpcrdma_ia *ia);
|
2007-09-11 00:50:12 +07:00
|
|
|
void rpcrdma_ia_close(struct rpcrdma_ia *);
|
2016-06-30 00:53:27 +07:00
|
|
|
bool frwr_is_supported(struct rpcrdma_ia *);
|
|
|
|
bool fmr_is_supported(struct rpcrdma_ia *);
|
2007-09-11 00:50:12 +07:00
|
|
|
|
2017-10-17 02:01:30 +07:00
|
|
|
extern struct workqueue_struct *rpcrdma_receive_wq;
|
|
|
|
|
2007-09-11 00:50:12 +07:00
|
|
|
/*
|
|
|
|
* Endpoint calls - xprtrdma/verbs.c
|
|
|
|
*/
|
|
|
|
int rpcrdma_ep_create(struct rpcrdma_ep *, struct rpcrdma_ia *,
|
|
|
|
struct rpcrdma_create_data_internal *);
|
2014-05-28 21:33:16 +07:00
|
|
|
void rpcrdma_ep_destroy(struct rpcrdma_ep *, struct rpcrdma_ia *);
|
2007-09-11 00:50:12 +07:00
|
|
|
int rpcrdma_ep_connect(struct rpcrdma_ep *, struct rpcrdma_ia *);
|
2016-11-29 22:53:37 +07:00
|
|
|
void rpcrdma_conn_func(struct rpcrdma_ep *ep);
|
2014-07-30 04:25:55 +07:00
|
|
|
void rpcrdma_ep_disconnect(struct rpcrdma_ep *, struct rpcrdma_ia *);
|
2007-09-11 00:50:12 +07:00
|
|
|
|
|
|
|
int rpcrdma_ep_post(struct rpcrdma_ia *, struct rpcrdma_ep *,
|
|
|
|
struct rpcrdma_req *);
|
2018-05-05 02:35:20 +07:00
|
|
|
void rpcrdma_post_recvs(struct rpcrdma_xprt *r_xprt, bool temp);
|
2007-09-11 00:50:12 +07:00
|
|
|
|
|
|
|
/*
|
|
|
|
* Buffer calls - xprtrdma/verbs.c
|
|
|
|
*/
|
2015-10-25 04:27:43 +07:00
|
|
|
struct rpcrdma_req *rpcrdma_create_req(struct rpcrdma_xprt *);
|
2016-09-15 21:56:26 +07:00
|
|
|
void rpcrdma_destroy_req(struct rpcrdma_req *);
|
2015-01-21 23:03:44 +07:00
|
|
|
int rpcrdma_buffer_create(struct rpcrdma_xprt *);
|
2007-09-11 00:50:12 +07:00
|
|
|
void rpcrdma_buffer_destroy(struct rpcrdma_buffer *);
|
xprtrdma: Add data structure to manage RDMA Send arguments
Problem statement:
Recently Sagi Grimberg <sagi@grimberg.me> observed that kernel RDMA-
enabled storage initiators don't handle delayed Send completion
correctly. If Send completion is delayed beyond the end of a ULP
transaction, the ULP may release resources that are still being used
by the HCA to complete a long-running Send operation.
This is a common design trait amongst our initiators. Most Send
operations are faster than the ULP transaction they are part of.
Waiting for a completion for these is typically unnecessary.
Infrequently, a network partition or some other problem crops up
where an ordering problem can occur. In NFS parlance, the RPC Reply
arrives and completes the RPC, but the HCA is still retrying the
Send WR that conveyed the RPC Call. In this case, the HCA can try
to use memory that has been invalidated or DMA unmapped, and the
connection is lost. If that memory has been re-used for something
else (possibly not related to NFS), and the Send retransmission
exposes that data on the wire.
Thus we cannot assume that it is safe to release Send-related
resources just because a ULP reply has arrived.
After some analysis, we have determined that the completion
housekeeping will not be difficult for xprtrdma:
- Inline Send buffers are registered via the local DMA key, and
are already left DMA mapped for the lifetime of a transport
connection, thus no additional handling is necessary for those
- Gathered Sends involving page cache pages _will_ need to
DMA unmap those pages after the Send completes. But like
inline send buffers, they are registered via the local DMA key,
and thus will not need to be invalidated
In addition, RPC completion will need to wait for Send completion
in the latter case. However, nearly always, the Send that conveys
the RPC Call will have completed long before the RPC Reply
arrives, and thus no additional latency will be accrued.
Design notes:
In this patch, the rpcrdma_sendctx object is introduced, and a
lock-free circular queue is added to manage a set of them per
transport.
The RPC client's send path already prevents sending more than one
RPC Call at the same time. This allows us to treat the consumer
side of the queue (rpcrdma_sendctx_get_locked) as if there is a
single consumer thread.
The producer side of the queue (rpcrdma_sendctx_put_locked) is
invoked only from the Send completion handler, which is a single
thread of execution (soft IRQ).
The only care that needs to be taken is with the tail index, which
is shared between the producer and consumer. Only the producer
updates the tail index. The consumer compares the head with the
tail to ensure that the a sendctx that is in use is never handed
out again (or, expressed more conventionally, the queue is empty).
When the sendctx queue empties completely, there are enough Sends
outstanding that posting more Send operations can result in a Send
Queue overflow. In this case, the ULP is told to wait and try again.
This introduces strong Send Queue accounting to xprtrdma.
As a final touch, Jason Gunthorpe <jgunthorpe@obsidianresearch.com>
suggested a mechanism that does not require signaling every Send.
We signal once every N Sends, and perform SGE unmapping of N Send
operations during that one completion.
Reported-by: Sagi Grimberg <sagi@grimberg.me>
Suggested-by: Jason Gunthorpe <jgunthorpe@obsidianresearch.com>
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: Anna Schumaker <Anna.Schumaker@Netapp.com>
2017-10-20 21:48:12 +07:00
|
|
|
struct rpcrdma_sendctx *rpcrdma_sendctx_get_locked(struct rpcrdma_buffer *buf);
|
2007-09-11 00:50:12 +07:00
|
|
|
|
2017-12-15 08:57:55 +07:00
|
|
|
struct rpcrdma_mr *rpcrdma_mr_get(struct rpcrdma_xprt *r_xprt);
|
|
|
|
void rpcrdma_mr_put(struct rpcrdma_mr *mr);
|
2017-12-15 08:58:04 +07:00
|
|
|
void rpcrdma_mr_unmap_and_put(struct rpcrdma_mr *mr);
|
2017-12-15 08:57:55 +07:00
|
|
|
void rpcrdma_mr_defer_recovery(struct rpcrdma_mr *mr);
|
|
|
|
|
2007-09-11 00:50:12 +07:00
|
|
|
struct rpcrdma_req *rpcrdma_buffer_get(struct rpcrdma_buffer *);
|
|
|
|
void rpcrdma_buffer_put(struct rpcrdma_req *);
|
|
|
|
void rpcrdma_recv_buffer_put(struct rpcrdma_rep *);
|
|
|
|
|
2016-09-15 21:56:26 +07:00
|
|
|
struct rpcrdma_regbuf *rpcrdma_alloc_regbuf(size_t, enum dma_data_direction,
|
2016-09-15 21:56:10 +07:00
|
|
|
gfp_t);
|
xprtrdma: Delay DMA mapping Send and Receive buffers
Currently, each regbuf is allocated and DMA mapped at the same time.
This is done during transport creation.
When a device driver is unloaded, every DMA-mapped buffer in use by
a transport has to be unmapped, and then remapped to the new
device if the driver is loaded again. Remapping will have to be done
_after_ the connect worker has set up the new device.
But there's an ordering problem:
call_allocate, which invokes xprt_rdma_allocate which calls
rpcrdma_alloc_regbuf to allocate Send buffers, happens _before_
the connect worker can run to set up the new device.
Instead, at transport creation, allocate each buffer, but leave it
unmapped. Once the RPC carries these buffers into ->send_request, by
which time a transport connection should have been established,
check to see that the RPC's buffers have been DMA mapped. If not,
map them there.
When device driver unplug support is added, it will simply unmap all
the transport's regbufs, but it doesn't have to deallocate the
underlying memory.
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: Anna Schumaker <Anna.Schumaker@Netapp.com>
2016-09-15 21:56:18 +07:00
|
|
|
bool __rpcrdma_dma_map_regbuf(struct rpcrdma_ia *, struct rpcrdma_regbuf *);
|
2016-09-15 21:56:26 +07:00
|
|
|
void rpcrdma_free_regbuf(struct rpcrdma_regbuf *);
|
2015-01-21 23:04:00 +07:00
|
|
|
|
xprtrdma: Delay DMA mapping Send and Receive buffers
Currently, each regbuf is allocated and DMA mapped at the same time.
This is done during transport creation.
When a device driver is unloaded, every DMA-mapped buffer in use by
a transport has to be unmapped, and then remapped to the new
device if the driver is loaded again. Remapping will have to be done
_after_ the connect worker has set up the new device.
But there's an ordering problem:
call_allocate, which invokes xprt_rdma_allocate which calls
rpcrdma_alloc_regbuf to allocate Send buffers, happens _before_
the connect worker can run to set up the new device.
Instead, at transport creation, allocate each buffer, but leave it
unmapped. Once the RPC carries these buffers into ->send_request, by
which time a transport connection should have been established,
check to see that the RPC's buffers have been DMA mapped. If not,
map them there.
When device driver unplug support is added, it will simply unmap all
the transport's regbufs, but it doesn't have to deallocate the
underlying memory.
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: Anna Schumaker <Anna.Schumaker@Netapp.com>
2016-09-15 21:56:18 +07:00
|
|
|
static inline bool
|
|
|
|
rpcrdma_regbuf_is_mapped(struct rpcrdma_regbuf *rb)
|
|
|
|
{
|
|
|
|
return rb->rg_device != NULL;
|
|
|
|
}
|
|
|
|
|
|
|
|
static inline bool
|
|
|
|
rpcrdma_dma_map_regbuf(struct rpcrdma_ia *ia, struct rpcrdma_regbuf *rb)
|
|
|
|
{
|
|
|
|
if (likely(rpcrdma_regbuf_is_mapped(rb)))
|
|
|
|
return true;
|
|
|
|
return __rpcrdma_dma_map_regbuf(ia, rb);
|
|
|
|
}
|
|
|
|
|
2015-10-25 04:27:10 +07:00
|
|
|
int rpcrdma_alloc_wq(void);
|
|
|
|
void rpcrdma_destroy_wq(void);
|
|
|
|
|
2015-03-31 01:35:44 +07:00
|
|
|
/*
|
|
|
|
* Wrappers for chunk registration, shared by read/write chunk code.
|
|
|
|
*/
|
|
|
|
|
|
|
|
static inline enum dma_data_direction
|
|
|
|
rpcrdma_data_dir(bool writing)
|
|
|
|
{
|
|
|
|
return writing ? DMA_FROM_DEVICE : DMA_TO_DEVICE;
|
|
|
|
}
|
|
|
|
|
2007-09-11 00:50:12 +07:00
|
|
|
/*
|
|
|
|
* RPC/RDMA protocol calls - xprtrdma/rpc_rdma.c
|
|
|
|
*/
|
xprtrdma: Use gathered Send for large inline messages
An RPC Call message that is sent inline but that has a data payload
(ie, one or more items in rq_snd_buf's page list) must be "pulled
up:"
- call_allocate has to reserve enough RPC Call buffer space to
accommodate the data payload
- call_transmit has to memcopy the rq_snd_buf's page list and tail
into its head iovec before it is sent
As the inline threshold is increased beyond its current 1KB default,
however, this means data payloads of more than a few KB are copied
by the host CPU. For example, if the inline threshold is increased
just to 4KB, then NFS WRITE requests up to 4KB would involve a
memcpy of the NFS WRITE's payload data into the RPC Call buffer.
This is an undesirable amount of participation by the host CPU.
The inline threshold may be much larger than 4KB in the future,
after negotiation with a peer server.
Instead of copying the components of rq_snd_buf into its head iovec,
construct a gather list of these components, and send them all in
place. The same approach is already used in the Linux server's
RPC-over-RDMA reply path.
This mechanism also eliminates the need for rpcrdma_tail_pullup,
which is used to manage the XDR pad and trailing inline content when
a Read list is present.
This requires that the pages in rq_snd_buf's page list be DMA-mapped
during marshaling, and unmapped when a data-bearing RPC is
completed. This is slightly less efficient for very small I/O
payloads, but significantly more efficient as data payload size and
inline threshold increase past a kilobyte.
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: Anna Schumaker <Anna.Schumaker@Netapp.com>
2016-09-15 21:57:24 +07:00
|
|
|
|
|
|
|
enum rpcrdma_chunktype {
|
|
|
|
rpcrdma_noch = 0,
|
|
|
|
rpcrdma_readch,
|
|
|
|
rpcrdma_areadch,
|
|
|
|
rpcrdma_writech,
|
|
|
|
rpcrdma_replych
|
|
|
|
};
|
|
|
|
|
2017-10-20 21:47:55 +07:00
|
|
|
int rpcrdma_prepare_send_sges(struct rpcrdma_xprt *r_xprt,
|
|
|
|
struct rpcrdma_req *req, u32 hdrlen,
|
|
|
|
struct xdr_buf *xdr,
|
|
|
|
enum rpcrdma_chunktype rtype);
|
xprtrdma: Add data structure to manage RDMA Send arguments
Problem statement:
Recently Sagi Grimberg <sagi@grimberg.me> observed that kernel RDMA-
enabled storage initiators don't handle delayed Send completion
correctly. If Send completion is delayed beyond the end of a ULP
transaction, the ULP may release resources that are still being used
by the HCA to complete a long-running Send operation.
This is a common design trait amongst our initiators. Most Send
operations are faster than the ULP transaction they are part of.
Waiting for a completion for these is typically unnecessary.
Infrequently, a network partition or some other problem crops up
where an ordering problem can occur. In NFS parlance, the RPC Reply
arrives and completes the RPC, but the HCA is still retrying the
Send WR that conveyed the RPC Call. In this case, the HCA can try
to use memory that has been invalidated or DMA unmapped, and the
connection is lost. If that memory has been re-used for something
else (possibly not related to NFS), and the Send retransmission
exposes that data on the wire.
Thus we cannot assume that it is safe to release Send-related
resources just because a ULP reply has arrived.
After some analysis, we have determined that the completion
housekeeping will not be difficult for xprtrdma:
- Inline Send buffers are registered via the local DMA key, and
are already left DMA mapped for the lifetime of a transport
connection, thus no additional handling is necessary for those
- Gathered Sends involving page cache pages _will_ need to
DMA unmap those pages after the Send completes. But like
inline send buffers, they are registered via the local DMA key,
and thus will not need to be invalidated
In addition, RPC completion will need to wait for Send completion
in the latter case. However, nearly always, the Send that conveys
the RPC Call will have completed long before the RPC Reply
arrives, and thus no additional latency will be accrued.
Design notes:
In this patch, the rpcrdma_sendctx object is introduced, and a
lock-free circular queue is added to manage a set of them per
transport.
The RPC client's send path already prevents sending more than one
RPC Call at the same time. This allows us to treat the consumer
side of the queue (rpcrdma_sendctx_get_locked) as if there is a
single consumer thread.
The producer side of the queue (rpcrdma_sendctx_put_locked) is
invoked only from the Send completion handler, which is a single
thread of execution (soft IRQ).
The only care that needs to be taken is with the tail index, which
is shared between the producer and consumer. Only the producer
updates the tail index. The consumer compares the head with the
tail to ensure that the a sendctx that is in use is never handed
out again (or, expressed more conventionally, the queue is empty).
When the sendctx queue empties completely, there are enough Sends
outstanding that posting more Send operations can result in a Send
Queue overflow. In this case, the ULP is told to wait and try again.
This introduces strong Send Queue accounting to xprtrdma.
As a final touch, Jason Gunthorpe <jgunthorpe@obsidianresearch.com>
suggested a mechanism that does not require signaling every Send.
We signal once every N Sends, and perform SGE unmapping of N Send
operations during that one completion.
Reported-by: Sagi Grimberg <sagi@grimberg.me>
Suggested-by: Jason Gunthorpe <jgunthorpe@obsidianresearch.com>
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: Anna Schumaker <Anna.Schumaker@Netapp.com>
2017-10-20 21:48:12 +07:00
|
|
|
void rpcrdma_unmap_sendctx(struct rpcrdma_sendctx *sc);
|
2017-08-10 23:47:12 +07:00
|
|
|
int rpcrdma_marshal_req(struct rpcrdma_xprt *r_xprt, struct rpc_rqst *rqst);
|
2016-09-15 21:57:07 +07:00
|
|
|
void rpcrdma_set_max_header_sizes(struct rpcrdma_xprt *);
|
2017-10-17 02:01:22 +07:00
|
|
|
void rpcrdma_complete_rqst(struct rpcrdma_rep *rep);
|
2017-10-17 02:01:30 +07:00
|
|
|
void rpcrdma_reply_handler(struct rpcrdma_rep *rep);
|
2017-10-20 21:48:28 +07:00
|
|
|
void rpcrdma_release_rqst(struct rpcrdma_xprt *r_xprt,
|
|
|
|
struct rpcrdma_req *req);
|
2017-10-17 02:01:30 +07:00
|
|
|
void rpcrdma_deferred_completion(struct work_struct *work);
|
2007-09-11 00:50:12 +07:00
|
|
|
|
2017-08-04 01:30:03 +07:00
|
|
|
static inline void rpcrdma_set_xdrlen(struct xdr_buf *xdr, size_t len)
|
|
|
|
{
|
|
|
|
xdr->head[0].iov_len = len;
|
|
|
|
xdr->len = len;
|
|
|
|
}
|
|
|
|
|
2015-06-04 22:21:42 +07:00
|
|
|
/* RPC/RDMA module init - xprtrdma/transport.c
|
|
|
|
*/
|
2016-01-08 02:50:10 +07:00
|
|
|
extern unsigned int xprt_rdma_max_inline_read;
|
|
|
|
void xprt_rdma_format_addresses(struct rpc_xprt *xprt, struct sockaddr *sap);
|
|
|
|
void xprt_rdma_free_addresses(struct rpc_xprt *xprt);
|
2016-11-29 22:53:37 +07:00
|
|
|
void rpcrdma_connect_worker(struct work_struct *work);
|
2016-01-08 02:50:10 +07:00
|
|
|
void xprt_rdma_print_stats(struct rpc_xprt *xprt, struct seq_file *seq);
|
2015-06-04 22:21:42 +07:00
|
|
|
int xprt_rdma_init(void);
|
|
|
|
void xprt_rdma_cleanup(void);
|
|
|
|
|
2015-10-25 04:27:43 +07:00
|
|
|
/* Backchannel calls - xprtrdma/backchannel.c
|
|
|
|
*/
|
|
|
|
#if defined(CONFIG_SUNRPC_BACKCHANNEL)
|
|
|
|
int xprt_rdma_bc_setup(struct rpc_xprt *, unsigned int);
|
2015-10-25 04:28:32 +07:00
|
|
|
int xprt_rdma_bc_up(struct svc_serv *, struct net *);
|
2016-05-03 01:40:40 +07:00
|
|
|
size_t xprt_rdma_bc_maxpayload(struct rpc_xprt *);
|
2015-10-25 04:27:43 +07:00
|
|
|
int rpcrdma_bc_post_recv(struct rpcrdma_xprt *, unsigned int);
|
2015-10-25 04:28:08 +07:00
|
|
|
void rpcrdma_bc_receive_call(struct rpcrdma_xprt *, struct rpcrdma_rep *);
|
2017-12-15 08:57:31 +07:00
|
|
|
int xprt_rdma_bc_send_reply(struct rpc_rqst *rqst);
|
2015-10-25 04:27:43 +07:00
|
|
|
void xprt_rdma_bc_free_rqst(struct rpc_rqst *);
|
|
|
|
void xprt_rdma_bc_destroy(struct rpc_xprt *, unsigned int);
|
|
|
|
#endif /* CONFIG_SUNRPC_BACKCHANNEL */
|
|
|
|
|
2016-01-08 02:50:10 +07:00
|
|
|
extern struct xprt_class xprt_rdma_bc;
|
2012-02-16 00:30:00 +07:00
|
|
|
|
2007-09-11 00:50:12 +07:00
|
|
|
#endif /* _LINUX_SUNRPC_XPRT_RDMA_H */
|