diff options
author | Mike Marciniszyn <mike.marciniszyn@intel.com> | 2017-08-28 20:24:10 +0200 |
---|---|---|
committer | Doug Ledford <dledford@redhat.com> | 2017-08-29 01:12:31 +0200 |
commit | 0208da90def5776cef940f9de4ffe6ecef346207 (patch) | |
tree | 05833a7dffa3de7e1eea51a004871ae9780e40d5 /drivers/infiniband | |
parent | IB/qib: Convert qp_stats debugfs interface to use new iterator API (diff) | |
download | linux-0208da90def5776cef940f9de4ffe6ecef346207.tar.xz linux-0208da90def5776cef940f9de4ffe6ecef346207.zip |
IB/rdmavt: Handle dereg of inuse MRs properly
A destroy of an MR prior to destroying the QP can cause the following
diagnostic if the QP is referencing the MR being de-registered:
hfi1 0000:05:00.0: hfi1_0: rvt_dereg_mr timeout mr ffff8808562108
00 pd ffff880859b20b00
The solution is to when the a non-zero refcount is encountered when
the MR is destroyed the QPs needs to be iterated looking for QPs in
the same PD as the MR. If rvt_qp_mr_clean() detects any such QP
references the rkey/lkey, the QP needs to be put into an error state
via a call to rvt_qp_error() which will trigger the clean up of any
stuck references.
This solution is as specified in IBTA 1.3 Volume 1 11.2.10.5.
[This is reproduced with the 0.4.9 version of qperf and the rc_bw test]
Reviewed-by: Dennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: Mike Marciniszyn <mike.marciniszyn@intel.com>
Signed-off-by: Dennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>
Diffstat (limited to 'drivers/infiniband')
-rw-r--r-- | drivers/infiniband/sw/rdmavt/mr.c | 121 | ||||
-rw-r--r-- | drivers/infiniband/sw/rdmavt/qp.c | 112 |
2 files changed, 212 insertions, 21 deletions
diff --git a/drivers/infiniband/sw/rdmavt/mr.c b/drivers/infiniband/sw/rdmavt/mr.c index 1b3801f78e78..42713511b53b 100644 --- a/drivers/infiniband/sw/rdmavt/mr.c +++ b/drivers/infiniband/sw/rdmavt/mr.c @@ -441,6 +441,105 @@ bail_umem: } /** + * rvt_dereg_clean_qp_cb - callback from iterator + * @qp - the qp + * @v - the mregion (as u64) + * + * This routine fields the callback for all QPs and + * for QPs in the same PD as the MR will call the + * rvt_qp_mr_clean() to potentially cleanup references. + */ +static void rvt_dereg_clean_qp_cb(struct rvt_qp *qp, u64 v) +{ + struct rvt_mregion *mr = (struct rvt_mregion *)v; + + /* skip PDs that are not ours */ + if (mr->pd != qp->ibqp.pd) + return; + rvt_qp_mr_clean(qp, mr->lkey); +} + +/** + * rvt_dereg_clean_qps - find QPs for reference cleanup + * @mr - the MR that is being deregistered + * + * This routine iterates RC QPs looking for references + * to the lkey noted in mr. + */ +static void rvt_dereg_clean_qps(struct rvt_mregion *mr) +{ + struct rvt_dev_info *rdi = ib_to_rvt(mr->pd->device); + + rvt_qp_iter(rdi, (u64)mr, rvt_dereg_clean_qp_cb); +} + +/** + * rvt_check_refs - check references + * @mr - the megion + * @t - the caller identification + * + * This routine checks MRs holding a reference during + * when being de-registered. + * + * If the count is non-zero, the code calls a clean routine then + * waits for the timeout for the count to zero. + */ +static int rvt_check_refs(struct rvt_mregion *mr, const char *t) +{ + unsigned long timeout; + struct rvt_dev_info *rdi = ib_to_rvt(mr->pd->device); + + if (percpu_ref_is_zero(&mr->refcount)) + return 0; + /* avoid dma mr */ + if (mr->lkey) + rvt_dereg_clean_qps(mr); + timeout = wait_for_completion_timeout(&mr->comp, 5 * HZ); + if (!timeout) { + rvt_pr_err(rdi, + "%s timeout mr %p pd %p lkey %x refcount %ld\n", + t, mr, mr->pd, mr->lkey, + atomic_long_read(&mr->refcount.count)); + rvt_get_mr(mr); + return -EBUSY; + } + return 0; +} + +/** + * rvt_mr_has_lkey - is MR + * @mr - the mregion + * @lkey - the lkey + */ +bool rvt_mr_has_lkey(struct rvt_mregion *mr, u32 lkey) +{ + return mr && lkey == mr->lkey; +} + +/** + * rvt_ss_has_lkey - is mr in sge tests + * @ss - the sge state + * @lkey + * + * This code tests for an MR in the indicated + * sge state. + */ +bool rvt_ss_has_lkey(struct rvt_sge_state *ss, u32 lkey) +{ + int i; + bool rval = false; + + if (!ss->num_sge) + return rval; + /* first one */ + rval = rvt_mr_has_lkey(ss->sge.mr, lkey); + /* any others */ + for (i = 0; !rval && i < ss->num_sge - 1; i++) + rval = rvt_mr_has_lkey(ss->sg_list[i].mr, lkey); + return rval; +} + +/** * rvt_dereg_mr - unregister and free a memory region * @ibmr: the memory region to free * @@ -453,22 +552,14 @@ bail_umem: int rvt_dereg_mr(struct ib_mr *ibmr) { struct rvt_mr *mr = to_imr(ibmr); - struct rvt_dev_info *rdi = ib_to_rvt(ibmr->pd->device); - int ret = 0; - unsigned long timeout; + int ret; rvt_free_lkey(&mr->mr); rvt_put_mr(&mr->mr); /* will set completion if last */ - timeout = wait_for_completion_timeout(&mr->mr.comp, 5 * HZ); - if (!timeout) { - rvt_pr_err(rdi, - "rvt_dereg_mr timeout mr %p pd %p\n", - mr, mr->mr.pd); - rvt_get_mr(&mr->mr); - ret = -EBUSY; + ret = rvt_check_refs(&mr->mr, __func__); + if (ret) goto out; - } rvt_deinit_mregion(&mr->mr); if (mr->umem) ib_umem_release(mr->umem); @@ -761,16 +852,12 @@ int rvt_dealloc_fmr(struct ib_fmr *ibfmr) { struct rvt_fmr *fmr = to_ifmr(ibfmr); int ret = 0; - unsigned long timeout; rvt_free_lkey(&fmr->mr); rvt_put_mr(&fmr->mr); /* will set completion if last */ - timeout = wait_for_completion_timeout(&fmr->mr.comp, 5 * HZ); - if (!timeout) { - rvt_get_mr(&fmr->mr); - ret = -EBUSY; + ret = rvt_check_refs(&fmr->mr, __func__); + if (ret) goto out; - } rvt_deinit_mregion(&fmr->mr); kfree(fmr); out: diff --git a/drivers/infiniband/sw/rdmavt/qp.c b/drivers/infiniband/sw/rdmavt/qp.c index 9f70fd8665ab..22df09ae809e 100644 --- a/drivers/infiniband/sw/rdmavt/qp.c +++ b/drivers/infiniband/sw/rdmavt/qp.c @@ -458,10 +458,7 @@ static void rvt_clear_mr_refs(struct rvt_qp *qp, int clr_sends) } } - if (qp->ibqp.qp_type != IB_QPT_RC) - return; - - for (n = 0; n < rvt_max_atomic(rdi); n++) { + for (n = 0; qp->s_ack_queue && n < rvt_max_atomic(rdi); n++) { struct rvt_ack_entry *e = &qp->s_ack_queue[n]; if (e->rdma_sge.mr) { @@ -472,6 +469,113 @@ static void rvt_clear_mr_refs(struct rvt_qp *qp, int clr_sends) } /** + * rvt_swqe_has_lkey - return true if lkey is used by swqe + * @wqe - the send wqe + * @lkey - the lkey + * + * Test the swqe for using lkey + */ +static bool rvt_swqe_has_lkey(struct rvt_swqe *wqe, u32 lkey) +{ + int i; + + for (i = 0; i < wqe->wr.num_sge; i++) { + struct rvt_sge *sge = &wqe->sg_list[i]; + + if (rvt_mr_has_lkey(sge->mr, lkey)) + return true; + } + return false; +} + +/** + * rvt_qp_sends_has_lkey - return true is qp sends use lkey + * @qp - the rvt_qp + * @lkey - the lkey + */ +static bool rvt_qp_sends_has_lkey(struct rvt_qp *qp, u32 lkey) +{ + u32 s_last = qp->s_last; + + while (s_last != qp->s_head) { + struct rvt_swqe *wqe = rvt_get_swqe_ptr(qp, s_last); + + if (rvt_swqe_has_lkey(wqe, lkey)) + return true; + + if (++s_last >= qp->s_size) + s_last = 0; + } + if (qp->s_rdma_mr) + if (rvt_mr_has_lkey(qp->s_rdma_mr, lkey)) + return true; + return false; +} + +/** + * rvt_qp_acks_has_lkey - return true if acks have lkey + * @qp - the qp + * @lkey - the lkey + */ +static bool rvt_qp_acks_has_lkey(struct rvt_qp *qp, u32 lkey) +{ + int i; + struct rvt_dev_info *rdi = ib_to_rvt(qp->ibqp.device); + + for (i = 0; qp->s_ack_queue && i < rvt_max_atomic(rdi); i++) { + struct rvt_ack_entry *e = &qp->s_ack_queue[i]; + + if (rvt_mr_has_lkey(e->rdma_sge.mr, lkey)) + return true; + } + return false; +} + +/* + * rvt_qp_mr_clean - clean up remote ops for lkey + * @qp - the qp + * @lkey - the lkey that is being de-registered + * + * This routine checks if the lkey is being used by + * the qp. + * + * If so, the qp is put into an error state to elminate + * any references from the qp. + */ +void rvt_qp_mr_clean(struct rvt_qp *qp, u32 lkey) +{ + bool lastwqe = false; + + if (qp->ibqp.qp_type == IB_QPT_SMI || + qp->ibqp.qp_type == IB_QPT_GSI) + /* avoid special QPs */ + return; + spin_lock_irq(&qp->r_lock); + spin_lock(&qp->s_hlock); + spin_lock(&qp->s_lock); + + if (qp->state == IB_QPS_ERR || qp->state == IB_QPS_RESET) + goto check_lwqe; + + if (rvt_ss_has_lkey(&qp->r_sge, lkey) || + rvt_qp_sends_has_lkey(qp, lkey) || + rvt_qp_acks_has_lkey(qp, lkey)) + lastwqe = rvt_error_qp(qp, IB_WC_LOC_PROT_ERR); +check_lwqe: + spin_unlock(&qp->s_lock); + spin_unlock(&qp->s_hlock); + spin_unlock_irq(&qp->r_lock); + if (lastwqe) { + struct ib_event ev; + + ev.device = qp->ibqp.device; + ev.element.qp = &qp->ibqp; + ev.event = IB_EVENT_QP_LAST_WQE_REACHED; + qp->ibqp.event_handler(&ev, qp->ibqp.qp_context); + } +} + +/** * rvt_remove_qp - remove qp form table * @rdi: rvt dev struct * @qp: qp to remove |