IB/qib: Support creating qps with GFP_NOIO flag
authorVinit Agnihotri <vinit.abhay.agnihotri@intel.com>
Mon, 11 Jan 2016 17:57:25 +0000 (12:57 -0500)
committerDoug Ledford <dledford@redhat.com>
Tue, 19 Jan 2016 20:17:40 +0000 (15:17 -0500)
The current code is problematic when the QP creation and ipoib is used to
support NFS and NFS desires to do IO for paging purposes. In that case, the
GFP_KERNEL allocation in qib_qp.c causes a deadlock in tight memory
situations.

This fix adds support to create queue pair with GFP_NOIO flag for connected
mode only to cleanly fail the create queue pair in those situations.

Cc: <stable@vger.kernel.org> # 3.16+
Reviewed-by: Mike Marciniszyn <mike.marciniszyn@intel.com>
Signed-off-by: Vinit Agnihotri <vinit.abhay.agnihotri@intel.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>
drivers/infiniband/hw/qib/qib_qp.c

index 40f85bb..3eff35c 100644 (file)
@@ -100,9 +100,10 @@ static u32 credit_table[31] = {
        32768                   /* 1E */
 };
 
-static void get_map_page(struct qib_qpn_table *qpt, struct qpn_map *map)
+static void get_map_page(struct qib_qpn_table *qpt, struct qpn_map *map,
+                        gfp_t gfp)
 {
-       unsigned long page = get_zeroed_page(GFP_KERNEL);
+       unsigned long page = get_zeroed_page(gfp);
 
        /*
         * Free the page if someone raced with us installing it.
@@ -121,7 +122,7 @@ static void get_map_page(struct qib_qpn_table *qpt, struct qpn_map *map)
  * zero/one for QP type IB_QPT_SMI/IB_QPT_GSI.
  */
 static int alloc_qpn(struct qib_devdata *dd, struct qib_qpn_table *qpt,
-                    enum ib_qp_type type, u8 port)
+                    enum ib_qp_type type, u8 port, gfp_t gfp)
 {
        u32 i, offset, max_scan, qpn;
        struct qpn_map *map;
@@ -151,7 +152,7 @@ static int alloc_qpn(struct qib_devdata *dd, struct qib_qpn_table *qpt,
        max_scan = qpt->nmaps - !offset;
        for (i = 0;;) {
                if (unlikely(!map->page)) {
-                       get_map_page(qpt, map);
+                       get_map_page(qpt, map, gfp);
                        if (unlikely(!map->page))
                                break;
                }
@@ -983,13 +984,21 @@ struct ib_qp *qib_create_qp(struct ib_pd *ibpd,
        size_t sz;
        size_t sg_list_sz;
        struct ib_qp *ret;
+       gfp_t gfp;
+
 
        if (init_attr->cap.max_send_sge > ib_qib_max_sges ||
            init_attr->cap.max_send_wr > ib_qib_max_qp_wrs ||
-           init_attr->create_flags) {
-               ret = ERR_PTR(-EINVAL);
-               goto bail;
-       }
+           init_attr->create_flags & ~(IB_QP_CREATE_USE_GFP_NOIO))
+               return ERR_PTR(-EINVAL);
+
+       /* GFP_NOIO is applicable in RC QPs only */
+       if (init_attr->create_flags & IB_QP_CREATE_USE_GFP_NOIO &&
+           init_attr->qp_type != IB_QPT_RC)
+               return ERR_PTR(-EINVAL);
+
+       gfp = init_attr->create_flags & IB_QP_CREATE_USE_GFP_NOIO ?
+                       GFP_NOIO : GFP_KERNEL;
 
        /* Check receive queue parameters if no SRQ is specified. */
        if (!init_attr->srq) {
@@ -1021,7 +1030,8 @@ struct ib_qp *qib_create_qp(struct ib_pd *ibpd,
                sz = sizeof(struct qib_sge) *
                        init_attr->cap.max_send_sge +
                        sizeof(struct qib_swqe);
-               swq = vmalloc((init_attr->cap.max_send_wr + 1) * sz);
+               swq = __vmalloc((init_attr->cap.max_send_wr + 1) * sz,
+                               gfp, PAGE_KERNEL);
                if (swq == NULL) {
                        ret = ERR_PTR(-ENOMEM);
                        goto bail;
@@ -1037,13 +1047,13 @@ struct ib_qp *qib_create_qp(struct ib_pd *ibpd,
                } else if (init_attr->cap.max_recv_sge > 1)
                        sg_list_sz = sizeof(*qp->r_sg_list) *
                                (init_attr->cap.max_recv_sge - 1);
-               qp = kzalloc(sz + sg_list_sz, GFP_KERNEL);
+               qp = kzalloc(sz + sg_list_sz, gfp);
                if (!qp) {
                        ret = ERR_PTR(-ENOMEM);
                        goto bail_swq;
                }
                RCU_INIT_POINTER(qp->next, NULL);
-               qp->s_hdr = kzalloc(sizeof(*qp->s_hdr), GFP_KERNEL);
+               qp->s_hdr = kzalloc(sizeof(*qp->s_hdr), gfp);
                if (!qp->s_hdr) {
                        ret = ERR_PTR(-ENOMEM);
                        goto bail_qp;
@@ -1058,8 +1068,16 @@ struct ib_qp *qib_create_qp(struct ib_pd *ibpd,
                        qp->r_rq.max_sge = init_attr->cap.max_recv_sge;
                        sz = (sizeof(struct ib_sge) * qp->r_rq.max_sge) +
                                sizeof(struct qib_rwqe);
-                       qp->r_rq.wq = vmalloc_user(sizeof(struct qib_rwq) +
-                                                  qp->r_rq.size * sz);
+                       if (gfp != GFP_NOIO)
+                               qp->r_rq.wq = vmalloc_user(
+                                               sizeof(struct qib_rwq) +
+                                               qp->r_rq.size * sz);
+                       else
+                               qp->r_rq.wq = __vmalloc(
+                                               sizeof(struct qib_rwq) +
+                                               qp->r_rq.size * sz,
+                                               gfp, PAGE_KERNEL);
+
                        if (!qp->r_rq.wq) {
                                ret = ERR_PTR(-ENOMEM);
                                goto bail_qp;
@@ -1090,7 +1108,7 @@ struct ib_qp *qib_create_qp(struct ib_pd *ibpd,
                dev = to_idev(ibpd->device);
                dd = dd_from_dev(dev);
                err = alloc_qpn(dd, &dev->qpn_table, init_attr->qp_type,
-                               init_attr->port_num);
+                               init_attr->port_num, gfp);
                if (err < 0) {
                        ret = ERR_PTR(err);
                        vfree(qp->r_rq.wq);