IB/qib: RDMA lkey/rkey validation is inefficient for large MRs
The current code loops during rkey/lkey validiation to isolate the MR
for the RDMA, which is expensive when the current operation is inside
a very large memory region.
This fix optimizes rkey/lkey validation routines for user memory
regions and fast memory regions. The MR entry can be isolated by
shifts/mods instead of looping. The existing loop is preserved for
phys memory regions for now.
Signed-off-by: Mike Marciniszyn <mike.marciniszyn@qlogic.com>
Signed-off-by: Roland Dreier <rolandd@cisco.com>
diff --git a/drivers/infiniband/hw/qib/qib_keys.c b/drivers/infiniband/hw/qib/qib_keys.c
index 4b80eb1..756d160 100644
--- a/drivers/infiniband/hw/qib/qib_keys.c
+++ b/drivers/infiniband/hw/qib/qib_keys.c
@@ -158,31 +158,47 @@
isge->sge_length = sge->length;
isge->m = 0;
isge->n = 0;
+ spin_unlock_irqrestore(&rkt->lock, flags);
goto ok;
}
mr = rkt->table[(sge->lkey >> (32 - ib_qib_lkey_table_size))];
if (unlikely(mr == NULL || mr->lkey != sge->lkey ||
mr->pd != &pd->ibpd))
goto bail;
+ atomic_inc(&mr->refcount);
+ spin_unlock_irqrestore(&rkt->lock, flags);
off = sge->addr - mr->user_base;
if (unlikely(sge->addr < mr->user_base ||
off + sge->length > mr->length ||
(mr->access_flags & acc) != acc))
- goto bail;
+ return ret;
off += mr->offset;
- m = 0;
- n = 0;
- while (off >= mr->map[m]->segs[n].length) {
- off -= mr->map[m]->segs[n].length;
- n++;
- if (n >= QIB_SEGSZ) {
- m++;
- n = 0;
+ 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/QIB_SEGSZ;
+ n = entries_spanned_by_off%QIB_SEGSZ;
+ } else {
+ m = 0;
+ n = 0;
+ while (off >= mr->map[m]->segs[n].length) {
+ off -= mr->map[m]->segs[n].length;
+ n++;
+ if (n >= QIB_SEGSZ) {
+ m++;
+ n = 0;
+ }
}
}
- atomic_inc(&mr->refcount);
isge->mr = mr;
isge->vaddr = mr->map[m]->segs[n].vaddr + off;
isge->length = mr->map[m]->segs[n].length - off;
@@ -191,6 +207,7 @@
isge->n = n;
ok:
ret = 1;
+ return ret;
bail:
spin_unlock_irqrestore(&rkt->lock, flags);
return ret;
@@ -237,30 +254,46 @@
sge->sge_length = len;
sge->m = 0;
sge->n = 0;
+ spin_unlock_irqrestore(&rkt->lock, flags);
goto ok;
}
mr = rkt->table[(rkey >> (32 - ib_qib_lkey_table_size))];
if (unlikely(mr == NULL || mr->lkey != rkey || qp->ibqp.pd != mr->pd))
goto bail;
+ atomic_inc(&mr->refcount);
+ spin_unlock_irqrestore(&rkt->lock, flags);
off = vaddr - mr->iova;
if (unlikely(vaddr < mr->iova || off + len > mr->length ||
(mr->access_flags & acc) == 0))
- goto bail;
+ return ret;
off += mr->offset;
- m = 0;
- n = 0;
- while (off >= mr->map[m]->segs[n].length) {
- off -= mr->map[m]->segs[n].length;
- n++;
- if (n >= QIB_SEGSZ) {
- m++;
- n = 0;
+ 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/QIB_SEGSZ;
+ n = entries_spanned_by_off%QIB_SEGSZ;
+ } else {
+ m = 0;
+ n = 0;
+ while (off >= mr->map[m]->segs[n].length) {
+ off -= mr->map[m]->segs[n].length;
+ n++;
+ if (n >= QIB_SEGSZ) {
+ m++;
+ n = 0;
+ }
}
}
- atomic_inc(&mr->refcount);
sge->mr = mr;
sge->vaddr = mr->map[m]->segs[n].vaddr + off;
sge->length = mr->map[m]->segs[n].length - off;
@@ -269,6 +302,7 @@
sge->n = n;
ok:
ret = 1;
+ return ret;
bail:
spin_unlock_irqrestore(&rkt->lock, flags);
return ret;