linux/drivers/infiniband/hw/qib/qib_keys.c

239 lines
5.9 KiB
C
Raw Normal View History

/*
* Copyright (c) 2006, 2007, 2009 QLogic Corporation. All rights reserved.
* Copyright (c) 2005, 2006 PathScale, Inc. All rights reserved.
*
* This software is available to you under a choice of one of two
* licenses. You may choose to be licensed under the terms of the GNU
* General Public License (GPL) Version 2, available from the file
* COPYING in the main directory of this source tree, or the
* OpenIB.org BSD license below:
*
* Redistribution and use in source and binary forms, with or
* without modification, are permitted provided that the following
* conditions are met:
*
* - Redistributions of source code must retain the above
* copyright notice, this list of conditions and the following
* disclaimer.
*
* - Redistributions in binary form must reproduce the above
* copyright notice, this list of conditions and the following
* disclaimer in the documentation and/or other materials
* provided with the distribution.
*
* THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
* EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
* MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
* NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
* BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
* ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
* CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
* SOFTWARE.
*/
#include "qib.h"
/**
* qib_alloc_lkey - allocate an lkey
* @mr: memory region that this lkey protects
IB/qib: Avoid returning EBUSY from MR deregister A timing issue can occur where qib_mr_dereg can return -EBUSY if the MR use count is not zero. This can occur if the MR is de-registered while RDMA read response packets are being progressed from the SDMA ring. The suspicion is that the peer sent an RDMA read request, which has already been copied across to the peer. The peer sees the completion of his request and then communicates to the responder that the MR is not needed any longer. The responder tries to de-register the MR, catching some responses remaining in the SDMA ring holding the MR use count. The code now uses a get/put paradigm to track MR use counts and coordinates with the MR de-registration process using a completion when the count has reached zero. A timeout on the delay is in place to catch other EBUSY issues. The reference count protocol is as follows: - The return to the user counts as 1 - A reference from the lk_table or the qib_ibdev counts as 1. - Transient I/O operations increase/decrease as necessary A lot of code duplication has been folded into the new routines init_qib_mregion() and deinit_qib_mregion(). Additionally, explicit initialization of fields to zero is now handled by kzalloc(). Also, duplicated code 'while.*num_sge' that decrements reference counts have been consolidated in qib_put_ss(). Reviewed-by: Ramkrishna Vepa <ramkrishna.vepa@intel.com> Signed-off-by: Mike Marciniszyn <mike.marciniszyn@intel.com> Signed-off-by: Roland Dreier <roland@purestorage.com>
2012-06-28 06:33:12 +08:00
* @dma_region: 0->normal key, 1->restricted DMA key
*
* Returns 0 if successful, otherwise returns -errno.
*
* Increments mr reference count as required.
IB/qib: Avoid returning EBUSY from MR deregister A timing issue can occur where qib_mr_dereg can return -EBUSY if the MR use count is not zero. This can occur if the MR is de-registered while RDMA read response packets are being progressed from the SDMA ring. The suspicion is that the peer sent an RDMA read request, which has already been copied across to the peer. The peer sees the completion of his request and then communicates to the responder that the MR is not needed any longer. The responder tries to de-register the MR, catching some responses remaining in the SDMA ring holding the MR use count. The code now uses a get/put paradigm to track MR use counts and coordinates with the MR de-registration process using a completion when the count has reached zero. A timeout on the delay is in place to catch other EBUSY issues. The reference count protocol is as follows: - The return to the user counts as 1 - A reference from the lk_table or the qib_ibdev counts as 1. - Transient I/O operations increase/decrease as necessary A lot of code duplication has been folded into the new routines init_qib_mregion() and deinit_qib_mregion(). Additionally, explicit initialization of fields to zero is now handled by kzalloc(). Also, duplicated code 'while.*num_sge' that decrements reference counts have been consolidated in qib_put_ss(). Reviewed-by: Ramkrishna Vepa <ramkrishna.vepa@intel.com> Signed-off-by: Mike Marciniszyn <mike.marciniszyn@intel.com> Signed-off-by: Roland Dreier <roland@purestorage.com>
2012-06-28 06:33:12 +08:00
*
* Sets the lkey field mr for non-dma regions.
*
*/
int qib_alloc_lkey(struct rvt_mregion *mr, int dma_region)
{
unsigned long flags;
u32 r;
u32 n;
IB/qib: Avoid returning EBUSY from MR deregister A timing issue can occur where qib_mr_dereg can return -EBUSY if the MR use count is not zero. This can occur if the MR is de-registered while RDMA read response packets are being progressed from the SDMA ring. The suspicion is that the peer sent an RDMA read request, which has already been copied across to the peer. The peer sees the completion of his request and then communicates to the responder that the MR is not needed any longer. The responder tries to de-register the MR, catching some responses remaining in the SDMA ring holding the MR use count. The code now uses a get/put paradigm to track MR use counts and coordinates with the MR de-registration process using a completion when the count has reached zero. A timeout on the delay is in place to catch other EBUSY issues. The reference count protocol is as follows: - The return to the user counts as 1 - A reference from the lk_table or the qib_ibdev counts as 1. - Transient I/O operations increase/decrease as necessary A lot of code duplication has been folded into the new routines init_qib_mregion() and deinit_qib_mregion(). Additionally, explicit initialization of fields to zero is now handled by kzalloc(). Also, duplicated code 'while.*num_sge' that decrements reference counts have been consolidated in qib_put_ss(). Reviewed-by: Ramkrishna Vepa <ramkrishna.vepa@intel.com> Signed-off-by: Mike Marciniszyn <mike.marciniszyn@intel.com> Signed-off-by: Roland Dreier <roland@purestorage.com>
2012-06-28 06:33:12 +08:00
int ret = 0;
struct qib_ibdev *dev = to_idev(mr->pd->device);
struct rvt_lkey_table *rkt = &dev->lk_table;
spin_lock_irqsave(&rkt->lock, flags);
IB/qib: Avoid returning EBUSY from MR deregister A timing issue can occur where qib_mr_dereg can return -EBUSY if the MR use count is not zero. This can occur if the MR is de-registered while RDMA read response packets are being progressed from the SDMA ring. The suspicion is that the peer sent an RDMA read request, which has already been copied across to the peer. The peer sees the completion of his request and then communicates to the responder that the MR is not needed any longer. The responder tries to de-register the MR, catching some responses remaining in the SDMA ring holding the MR use count. The code now uses a get/put paradigm to track MR use counts and coordinates with the MR de-registration process using a completion when the count has reached zero. A timeout on the delay is in place to catch other EBUSY issues. The reference count protocol is as follows: - The return to the user counts as 1 - A reference from the lk_table or the qib_ibdev counts as 1. - Transient I/O operations increase/decrease as necessary A lot of code duplication has been folded into the new routines init_qib_mregion() and deinit_qib_mregion(). Additionally, explicit initialization of fields to zero is now handled by kzalloc(). Also, duplicated code 'while.*num_sge' that decrements reference counts have been consolidated in qib_put_ss(). Reviewed-by: Ramkrishna Vepa <ramkrishna.vepa@intel.com> Signed-off-by: Mike Marciniszyn <mike.marciniszyn@intel.com> Signed-off-by: Roland Dreier <roland@purestorage.com>
2012-06-28 06:33:12 +08:00
/* special case for dma_mr lkey == 0 */
if (dma_region) {
struct rvt_mregion *tmr;
IB/qib: Fix lockdep splat in qib_alloc_lkey() The following backtrace is reported with CONFIG_PROVE_RCU: drivers/infiniband/hw/qib/qib_keys.c:64 suspicious rcu_dereference_check() usage! other info that might help us debug this: rcu_scheduler_active = 1, debug_locks = 1 4 locks held by kworker/0:1/56: #0: (events){.+.+.+}, at: [<ffffffff8107a4f5>] process_one_work+0x165/0x4a0 #1: ((&wfc.work)){+.+.+.}, at: [<ffffffff8107a4f5>] process_one_work+0x165/0x4a0 #2: (device_mutex){+.+.+.}, at: [<ffffffffa0148dd8>] ib_register_device+0x38/0x220 [ib_core] #3: (&(&dev->lk_table.lock)->rlock){......}, at: [<ffffffffa017e81c>] qib_alloc_lkey+0x3c/0x1b0 [ib_qib] stack backtrace: Pid: 56, comm: kworker/0:1 Not tainted 3.10.0-rc1+ #6 Call Trace: [<ffffffff810c0b85>] lockdep_rcu_suspicious+0xe5/0x130 [<ffffffffa017e8e1>] qib_alloc_lkey+0x101/0x1b0 [ib_qib] [<ffffffffa0184886>] qib_get_dma_mr+0xa6/0xd0 [ib_qib] [<ffffffffa01461aa>] ib_get_dma_mr+0x1a/0x50 [ib_core] [<ffffffffa01678dc>] ib_mad_port_open+0x12c/0x390 [ib_mad] [<ffffffff810c2c55>] ? trace_hardirqs_on_caller+0x105/0x190 [<ffffffffa0167b92>] ib_mad_init_device+0x52/0x110 [ib_mad] [<ffffffffa01917c0>] ? sl2vl_attr_show+0x30/0x30 [ib_qib] [<ffffffffa0148f49>] ib_register_device+0x1a9/0x220 [ib_core] [<ffffffffa01b1685>] qib_register_ib_device+0x735/0xa40 [ib_qib] [<ffffffff8106ba98>] ? mod_timer+0x118/0x220 [<ffffffffa017d425>] qib_init_one+0x1e5/0x400 [ib_qib] [<ffffffff812ce86e>] local_pci_probe+0x4e/0x90 [<ffffffff81078118>] work_for_cpu_fn+0x18/0x30 [<ffffffff8107a566>] process_one_work+0x1d6/0x4a0 [<ffffffff8107a4f5>] ? process_one_work+0x165/0x4a0 [<ffffffff8107c9c9>] worker_thread+0x119/0x370 [<ffffffff8107c8b0>] ? manage_workers+0x180/0x180 [<ffffffff8108294e>] kthread+0xee/0x100 [<ffffffff81082860>] ? __init_kthread_worker+0x70/0x70 [<ffffffff815c04ac>] ret_from_fork+0x7c/0xb0 [<ffffffff81082860>] ? __init_kthread_worker+0x70/0x70 Per Documentation/RCU/lockdep-splat.txt, the code now uses rcu_access_pointer() vs. rcu_dereference(). Reported-by: Jay Fenlason <fenlason@redhat.com> Reviewed-by: Dean Luick <dean.luick@intel.com> Signed-off-by: Mike Marciniszyn <mike.marciniszyn@intel.com> Signed-off-by: Roland Dreier <roland@purestorage.com>
2013-05-17 20:40:32 +08:00
tmr = rcu_access_pointer(dev->dma_mr);
if (!tmr) {
IB/qib: Avoid returning EBUSY from MR deregister A timing issue can occur where qib_mr_dereg can return -EBUSY if the MR use count is not zero. This can occur if the MR is de-registered while RDMA read response packets are being progressed from the SDMA ring. The suspicion is that the peer sent an RDMA read request, which has already been copied across to the peer. The peer sees the completion of his request and then communicates to the responder that the MR is not needed any longer. The responder tries to de-register the MR, catching some responses remaining in the SDMA ring holding the MR use count. The code now uses a get/put paradigm to track MR use counts and coordinates with the MR de-registration process using a completion when the count has reached zero. A timeout on the delay is in place to catch other EBUSY issues. The reference count protocol is as follows: - The return to the user counts as 1 - A reference from the lk_table or the qib_ibdev counts as 1. - Transient I/O operations increase/decrease as necessary A lot of code duplication has been folded into the new routines init_qib_mregion() and deinit_qib_mregion(). Additionally, explicit initialization of fields to zero is now handled by kzalloc(). Also, duplicated code 'while.*num_sge' that decrements reference counts have been consolidated in qib_put_ss(). Reviewed-by: Ramkrishna Vepa <ramkrishna.vepa@intel.com> Signed-off-by: Mike Marciniszyn <mike.marciniszyn@intel.com> Signed-off-by: Roland Dreier <roland@purestorage.com>
2012-06-28 06:33:12 +08:00
qib_get_mr(mr);
rcu_assign_pointer(dev->dma_mr, mr);
IB/qib: Avoid returning EBUSY from MR deregister A timing issue can occur where qib_mr_dereg can return -EBUSY if the MR use count is not zero. This can occur if the MR is de-registered while RDMA read response packets are being progressed from the SDMA ring. The suspicion is that the peer sent an RDMA read request, which has already been copied across to the peer. The peer sees the completion of his request and then communicates to the responder that the MR is not needed any longer. The responder tries to de-register the MR, catching some responses remaining in the SDMA ring holding the MR use count. The code now uses a get/put paradigm to track MR use counts and coordinates with the MR de-registration process using a completion when the count has reached zero. A timeout on the delay is in place to catch other EBUSY issues. The reference count protocol is as follows: - The return to the user counts as 1 - A reference from the lk_table or the qib_ibdev counts as 1. - Transient I/O operations increase/decrease as necessary A lot of code duplication has been folded into the new routines init_qib_mregion() and deinit_qib_mregion(). Additionally, explicit initialization of fields to zero is now handled by kzalloc(). Also, duplicated code 'while.*num_sge' that decrements reference counts have been consolidated in qib_put_ss(). Reviewed-by: Ramkrishna Vepa <ramkrishna.vepa@intel.com> Signed-off-by: Mike Marciniszyn <mike.marciniszyn@intel.com> Signed-off-by: Roland Dreier <roland@purestorage.com>
2012-06-28 06:33:12 +08:00
mr->lkey_published = 1;
}
goto success;
}
/* Find the next available LKEY */
r = rkt->next;
n = r;
for (;;) {
if (rkt->table[r] == NULL)
break;
r = (r + 1) & (rkt->max - 1);
IB/qib: Avoid returning EBUSY from MR deregister A timing issue can occur where qib_mr_dereg can return -EBUSY if the MR use count is not zero. This can occur if the MR is de-registered while RDMA read response packets are being progressed from the SDMA ring. The suspicion is that the peer sent an RDMA read request, which has already been copied across to the peer. The peer sees the completion of his request and then communicates to the responder that the MR is not needed any longer. The responder tries to de-register the MR, catching some responses remaining in the SDMA ring holding the MR use count. The code now uses a get/put paradigm to track MR use counts and coordinates with the MR de-registration process using a completion when the count has reached zero. A timeout on the delay is in place to catch other EBUSY issues. The reference count protocol is as follows: - The return to the user counts as 1 - A reference from the lk_table or the qib_ibdev counts as 1. - Transient I/O operations increase/decrease as necessary A lot of code duplication has been folded into the new routines init_qib_mregion() and deinit_qib_mregion(). Additionally, explicit initialization of fields to zero is now handled by kzalloc(). Also, duplicated code 'while.*num_sge' that decrements reference counts have been consolidated in qib_put_ss(). Reviewed-by: Ramkrishna Vepa <ramkrishna.vepa@intel.com> Signed-off-by: Mike Marciniszyn <mike.marciniszyn@intel.com> Signed-off-by: Roland Dreier <roland@purestorage.com>
2012-06-28 06:33:12 +08:00
if (r == n)
goto bail;
}
rkt->next = (r + 1) & (rkt->max - 1);
/*
* Make sure lkey is never zero which is reserved to indicate an
* unrestricted LKEY.
*/
rkt->gen++;
/*
* bits are capped in qib_verbs.c to insure enough bits
* for generation number
*/
mr->lkey = (r << (32 - ib_rvt_lkey_table_size)) |
((((1 << (24 - ib_rvt_lkey_table_size)) - 1) & rkt->gen)
<< 8);
if (mr->lkey == 0) {
mr->lkey |= 1 << 8;
rkt->gen++;
}
IB/qib: Avoid returning EBUSY from MR deregister A timing issue can occur where qib_mr_dereg can return -EBUSY if the MR use count is not zero. This can occur if the MR is de-registered while RDMA read response packets are being progressed from the SDMA ring. The suspicion is that the peer sent an RDMA read request, which has already been copied across to the peer. The peer sees the completion of his request and then communicates to the responder that the MR is not needed any longer. The responder tries to de-register the MR, catching some responses remaining in the SDMA ring holding the MR use count. The code now uses a get/put paradigm to track MR use counts and coordinates with the MR de-registration process using a completion when the count has reached zero. A timeout on the delay is in place to catch other EBUSY issues. The reference count protocol is as follows: - The return to the user counts as 1 - A reference from the lk_table or the qib_ibdev counts as 1. - Transient I/O operations increase/decrease as necessary A lot of code duplication has been folded into the new routines init_qib_mregion() and deinit_qib_mregion(). Additionally, explicit initialization of fields to zero is now handled by kzalloc(). Also, duplicated code 'while.*num_sge' that decrements reference counts have been consolidated in qib_put_ss(). Reviewed-by: Ramkrishna Vepa <ramkrishna.vepa@intel.com> Signed-off-by: Mike Marciniszyn <mike.marciniszyn@intel.com> Signed-off-by: Roland Dreier <roland@purestorage.com>
2012-06-28 06:33:12 +08:00
qib_get_mr(mr);
rcu_assign_pointer(rkt->table[r], mr);
IB/qib: Avoid returning EBUSY from MR deregister A timing issue can occur where qib_mr_dereg can return -EBUSY if the MR use count is not zero. This can occur if the MR is de-registered while RDMA read response packets are being progressed from the SDMA ring. The suspicion is that the peer sent an RDMA read request, which has already been copied across to the peer. The peer sees the completion of his request and then communicates to the responder that the MR is not needed any longer. The responder tries to de-register the MR, catching some responses remaining in the SDMA ring holding the MR use count. The code now uses a get/put paradigm to track MR use counts and coordinates with the MR de-registration process using a completion when the count has reached zero. A timeout on the delay is in place to catch other EBUSY issues. The reference count protocol is as follows: - The return to the user counts as 1 - A reference from the lk_table or the qib_ibdev counts as 1. - Transient I/O operations increase/decrease as necessary A lot of code duplication has been folded into the new routines init_qib_mregion() and deinit_qib_mregion(). Additionally, explicit initialization of fields to zero is now handled by kzalloc(). Also, duplicated code 'while.*num_sge' that decrements reference counts have been consolidated in qib_put_ss(). Reviewed-by: Ramkrishna Vepa <ramkrishna.vepa@intel.com> Signed-off-by: Mike Marciniszyn <mike.marciniszyn@intel.com> Signed-off-by: Roland Dreier <roland@purestorage.com>
2012-06-28 06:33:12 +08:00
mr->lkey_published = 1;
success:
spin_unlock_irqrestore(&rkt->lock, flags);
IB/qib: Avoid returning EBUSY from MR deregister A timing issue can occur where qib_mr_dereg can return -EBUSY if the MR use count is not zero. This can occur if the MR is de-registered while RDMA read response packets are being progressed from the SDMA ring. The suspicion is that the peer sent an RDMA read request, which has already been copied across to the peer. The peer sees the completion of his request and then communicates to the responder that the MR is not needed any longer. The responder tries to de-register the MR, catching some responses remaining in the SDMA ring holding the MR use count. The code now uses a get/put paradigm to track MR use counts and coordinates with the MR de-registration process using a completion when the count has reached zero. A timeout on the delay is in place to catch other EBUSY issues. The reference count protocol is as follows: - The return to the user counts as 1 - A reference from the lk_table or the qib_ibdev counts as 1. - Transient I/O operations increase/decrease as necessary A lot of code duplication has been folded into the new routines init_qib_mregion() and deinit_qib_mregion(). Additionally, explicit initialization of fields to zero is now handled by kzalloc(). Also, duplicated code 'while.*num_sge' that decrements reference counts have been consolidated in qib_put_ss(). Reviewed-by: Ramkrishna Vepa <ramkrishna.vepa@intel.com> Signed-off-by: Mike Marciniszyn <mike.marciniszyn@intel.com> Signed-off-by: Roland Dreier <roland@purestorage.com>
2012-06-28 06:33:12 +08:00
out:
return ret;
IB/qib: Avoid returning EBUSY from MR deregister A timing issue can occur where qib_mr_dereg can return -EBUSY if the MR use count is not zero. This can occur if the MR is de-registered while RDMA read response packets are being progressed from the SDMA ring. The suspicion is that the peer sent an RDMA read request, which has already been copied across to the peer. The peer sees the completion of his request and then communicates to the responder that the MR is not needed any longer. The responder tries to de-register the MR, catching some responses remaining in the SDMA ring holding the MR use count. The code now uses a get/put paradigm to track MR use counts and coordinates with the MR de-registration process using a completion when the count has reached zero. A timeout on the delay is in place to catch other EBUSY issues. The reference count protocol is as follows: - The return to the user counts as 1 - A reference from the lk_table or the qib_ibdev counts as 1. - Transient I/O operations increase/decrease as necessary A lot of code duplication has been folded into the new routines init_qib_mregion() and deinit_qib_mregion(). Additionally, explicit initialization of fields to zero is now handled by kzalloc(). Also, duplicated code 'while.*num_sge' that decrements reference counts have been consolidated in qib_put_ss(). Reviewed-by: Ramkrishna Vepa <ramkrishna.vepa@intel.com> Signed-off-by: Mike Marciniszyn <mike.marciniszyn@intel.com> Signed-off-by: Roland Dreier <roland@purestorage.com>
2012-06-28 06:33:12 +08:00
bail:
spin_unlock_irqrestore(&rkt->lock, flags);
ret = -ENOMEM;
goto out;
}
/**
* qib_free_lkey - free an lkey
IB/qib: Avoid returning EBUSY from MR deregister A timing issue can occur where qib_mr_dereg can return -EBUSY if the MR use count is not zero. This can occur if the MR is de-registered while RDMA read response packets are being progressed from the SDMA ring. The suspicion is that the peer sent an RDMA read request, which has already been copied across to the peer. The peer sees the completion of his request and then communicates to the responder that the MR is not needed any longer. The responder tries to de-register the MR, catching some responses remaining in the SDMA ring holding the MR use count. The code now uses a get/put paradigm to track MR use counts and coordinates with the MR de-registration process using a completion when the count has reached zero. A timeout on the delay is in place to catch other EBUSY issues. The reference count protocol is as follows: - The return to the user counts as 1 - A reference from the lk_table or the qib_ibdev counts as 1. - Transient I/O operations increase/decrease as necessary A lot of code duplication has been folded into the new routines init_qib_mregion() and deinit_qib_mregion(). Additionally, explicit initialization of fields to zero is now handled by kzalloc(). Also, duplicated code 'while.*num_sge' that decrements reference counts have been consolidated in qib_put_ss(). Reviewed-by: Ramkrishna Vepa <ramkrishna.vepa@intel.com> Signed-off-by: Mike Marciniszyn <mike.marciniszyn@intel.com> Signed-off-by: Roland Dreier <roland@purestorage.com>
2012-06-28 06:33:12 +08:00
* @mr: mr to free from tables
*/
void qib_free_lkey(struct rvt_mregion *mr)
{
unsigned long flags;
u32 lkey = mr->lkey;
u32 r;
IB/qib: Avoid returning EBUSY from MR deregister A timing issue can occur where qib_mr_dereg can return -EBUSY if the MR use count is not zero. This can occur if the MR is de-registered while RDMA read response packets are being progressed from the SDMA ring. The suspicion is that the peer sent an RDMA read request, which has already been copied across to the peer. The peer sees the completion of his request and then communicates to the responder that the MR is not needed any longer. The responder tries to de-register the MR, catching some responses remaining in the SDMA ring holding the MR use count. The code now uses a get/put paradigm to track MR use counts and coordinates with the MR de-registration process using a completion when the count has reached zero. A timeout on the delay is in place to catch other EBUSY issues. The reference count protocol is as follows: - The return to the user counts as 1 - A reference from the lk_table or the qib_ibdev counts as 1. - Transient I/O operations increase/decrease as necessary A lot of code duplication has been folded into the new routines init_qib_mregion() and deinit_qib_mregion(). Additionally, explicit initialization of fields to zero is now handled by kzalloc(). Also, duplicated code 'while.*num_sge' that decrements reference counts have been consolidated in qib_put_ss(). Reviewed-by: Ramkrishna Vepa <ramkrishna.vepa@intel.com> Signed-off-by: Mike Marciniszyn <mike.marciniszyn@intel.com> Signed-off-by: Roland Dreier <roland@purestorage.com>
2012-06-28 06:33:12 +08:00
struct qib_ibdev *dev = to_idev(mr->pd->device);
struct rvt_lkey_table *rkt = &dev->lk_table;
IB/qib: Avoid returning EBUSY from MR deregister A timing issue can occur where qib_mr_dereg can return -EBUSY if the MR use count is not zero. This can occur if the MR is de-registered while RDMA read response packets are being progressed from the SDMA ring. The suspicion is that the peer sent an RDMA read request, which has already been copied across to the peer. The peer sees the completion of his request and then communicates to the responder that the MR is not needed any longer. The responder tries to de-register the MR, catching some responses remaining in the SDMA ring holding the MR use count. The code now uses a get/put paradigm to track MR use counts and coordinates with the MR de-registration process using a completion when the count has reached zero. A timeout on the delay is in place to catch other EBUSY issues. The reference count protocol is as follows: - The return to the user counts as 1 - A reference from the lk_table or the qib_ibdev counts as 1. - Transient I/O operations increase/decrease as necessary A lot of code duplication has been folded into the new routines init_qib_mregion() and deinit_qib_mregion(). Additionally, explicit initialization of fields to zero is now handled by kzalloc(). Also, duplicated code 'while.*num_sge' that decrements reference counts have been consolidated in qib_put_ss(). Reviewed-by: Ramkrishna Vepa <ramkrishna.vepa@intel.com> Signed-off-by: Mike Marciniszyn <mike.marciniszyn@intel.com> Signed-off-by: Roland Dreier <roland@purestorage.com>
2012-06-28 06:33:12 +08:00
spin_lock_irqsave(&rkt->lock, flags);
if (!mr->lkey_published)
goto out;
if (lkey == 0)
RCU_INIT_POINTER(dev->dma_mr, NULL);
else {
r = lkey >> (32 - ib_rvt_lkey_table_size);
RCU_INIT_POINTER(rkt->table[r], NULL);
}
qib_put_mr(mr);
mr->lkey_published = 0;
IB/qib: Avoid returning EBUSY from MR deregister A timing issue can occur where qib_mr_dereg can return -EBUSY if the MR use count is not zero. This can occur if the MR is de-registered while RDMA read response packets are being progressed from the SDMA ring. The suspicion is that the peer sent an RDMA read request, which has already been copied across to the peer. The peer sees the completion of his request and then communicates to the responder that the MR is not needed any longer. The responder tries to de-register the MR, catching some responses remaining in the SDMA ring holding the MR use count. The code now uses a get/put paradigm to track MR use counts and coordinates with the MR de-registration process using a completion when the count has reached zero. A timeout on the delay is in place to catch other EBUSY issues. The reference count protocol is as follows: - The return to the user counts as 1 - A reference from the lk_table or the qib_ibdev counts as 1. - Transient I/O operations increase/decrease as necessary A lot of code duplication has been folded into the new routines init_qib_mregion() and deinit_qib_mregion(). Additionally, explicit initialization of fields to zero is now handled by kzalloc(). Also, duplicated code 'while.*num_sge' that decrements reference counts have been consolidated in qib_put_ss(). Reviewed-by: Ramkrishna Vepa <ramkrishna.vepa@intel.com> Signed-off-by: Mike Marciniszyn <mike.marciniszyn@intel.com> Signed-off-by: Roland Dreier <roland@purestorage.com>
2012-06-28 06:33:12 +08:00
out:
spin_unlock_irqrestore(&rkt->lock, flags);
}
/**
* qib_rkey_ok - check the IB virtual address, length, and RKEY
* @qp: qp for validation
* @sge: SGE state
* @len: length of data
* @vaddr: virtual address to place data
* @rkey: rkey to check
* @acc: access flags
*
* Return 1 if successful, otherwise 0.
*
* increments the reference count upon success
*/
int qib_rkey_ok(struct rvt_qp *qp, struct rvt_sge *sge,
u32 len, u64 vaddr, u32 rkey, int acc)
{
struct rvt_lkey_table *rkt = &to_idev(qp->ibqp.device)->lk_table;
struct rvt_mregion *mr;
unsigned n, m;
size_t off;
/*
* We use RKEY == zero for kernel virtual addresses
* (see qib_get_dma_mr and qib_dma.c).
*/
rcu_read_lock();
if (rkey == 0) {
struct rvt_pd *pd = ibpd_to_rvtpd(qp->ibqp.pd);
struct qib_ibdev *dev = to_idev(pd->ibpd.device);
if (pd->user)
goto bail;
mr = rcu_dereference(dev->dma_mr);
if (!mr)
goto bail;
if (unlikely(!atomic_inc_not_zero(&mr->refcount)))
goto bail;
rcu_read_unlock();
sge->mr = mr;
sge->vaddr = (void *) vaddr;
sge->length = len;
sge->sge_length = len;
sge->m = 0;
sge->n = 0;
goto ok;
}
mr = rcu_dereference(
rkt->table[(rkey >> (32 - ib_rvt_lkey_table_size))]);
if (unlikely(!mr || mr->lkey != rkey || qp->ibqp.pd != mr->pd))
goto bail;
off = vaddr - mr->iova;
if (unlikely(vaddr < mr->iova || off + len > mr->length ||
(mr->access_flags & acc) == 0))
goto bail;
if (unlikely(!atomic_inc_not_zero(&mr->refcount)))
goto bail;
rcu_read_unlock();
off += mr->offset;
if (mr->page_shift) {
/*
page sizes are uniform power of 2 so no loop is necessary
entries_spanned_by_off is the number of times the loop below
would have executed.
*/
size_t entries_spanned_by_off;
entries_spanned_by_off = off >> mr->page_shift;
off -= (entries_spanned_by_off << mr->page_shift);
m = entries_spanned_by_off / RVT_SEGSZ;
n = entries_spanned_by_off % RVT_SEGSZ;
} else {
m = 0;
n = 0;
while (off >= mr->map[m]->segs[n].length) {
off -= mr->map[m]->segs[n].length;
n++;
if (n >= RVT_SEGSZ) {
m++;
n = 0;
}
}
}
sge->mr = mr;
sge->vaddr = mr->map[m]->segs[n].vaddr + off;
sge->length = mr->map[m]->segs[n].length - off;
sge->sge_length = len;
sge->m = m;
sge->n = n;
ok:
return 1;
bail:
rcu_read_unlock();
return 0;
}