RDMA/uverbs: Export XRC domains to user space
authorSean Hefty <sean.hefty@intel.com>
Tue, 24 May 2011 15:33:46 +0000 (08:33 -0700)
committerRoland Dreier <roland@purestorage.com>
Thu, 13 Oct 2011 16:21:24 +0000 (09:21 -0700)
Allow user space to create XRC domains.  Because XRCDs are expected to
be shared among multiple processes, we use inodes to identify an XRCD.

Based on patches by Jack Morgenstein <jackm@dev.mellanox.co.il>

Signed-off-by: Sean Hefty <sean.hefty@intel.com>
Signed-off-by: Roland Dreier <roland@purestorage.com>
drivers/infiniband/core/uverbs.h
drivers/infiniband/core/uverbs_cmd.c
drivers/infiniband/core/uverbs_main.c
drivers/infiniband/core/verbs.c
include/rdma/ib_user_verbs.h
include/rdma/ib_verbs.h

index a078e5624d22f700c0a526538e6e2926f5d2775f..461e2f357c1e09d5ff6292abd24c0f1e1ddddd18 100644 (file)
@@ -76,6 +76,8 @@ struct ib_uverbs_device {
        struct ib_device                       *ib_dev;
        int                                     devnum;
        struct cdev                             cdev;
+       struct rb_root                          xrcd_tree;
+       struct mutex                            xrcd_tree_mutex;
 };
 
 struct ib_uverbs_event_file {
@@ -120,6 +122,11 @@ struct ib_uevent_object {
        u32                     events_reported;
 };
 
+struct ib_uxrcd_object {
+       struct ib_uobject       uobject;
+       atomic_t                refcnt;
+};
+
 struct ib_uqp_object {
        struct ib_uevent_object uevent;
        struct list_head        mcast_list;
@@ -142,6 +149,7 @@ extern struct idr ib_uverbs_ah_idr;
 extern struct idr ib_uverbs_cq_idr;
 extern struct idr ib_uverbs_qp_idr;
 extern struct idr ib_uverbs_srq_idr;
+extern struct idr ib_uverbs_xrcd_idr;
 
 void idr_remove_uobj(struct idr *idp, struct ib_uobject *uobj);
 
@@ -161,6 +169,7 @@ void ib_uverbs_qp_event_handler(struct ib_event *event, void *context_ptr);
 void ib_uverbs_srq_event_handler(struct ib_event *event, void *context_ptr);
 void ib_uverbs_event_handler(struct ib_event_handler *handler,
                             struct ib_event *event);
+void ib_uverbs_dealloc_xrcd(struct ib_uverbs_device *dev, struct ib_xrcd *xrcd);
 
 #define IB_UVERBS_DECLARE_CMD(name)                                    \
        ssize_t ib_uverbs_##name(struct ib_uverbs_file *file,           \
@@ -195,5 +204,7 @@ IB_UVERBS_DECLARE_CMD(create_srq);
 IB_UVERBS_DECLARE_CMD(modify_srq);
 IB_UVERBS_DECLARE_CMD(query_srq);
 IB_UVERBS_DECLARE_CMD(destroy_srq);
+IB_UVERBS_DECLARE_CMD(open_xrcd);
+IB_UVERBS_DECLARE_CMD(close_xrcd);
 
 #endif /* UVERBS_H */
index 48d21d09afe7dc746034c4696c4687e7370d1bd8..c8b2a843fa007d1da31f768b0dc283a66a58432a 100644 (file)
@@ -47,6 +47,7 @@ static struct lock_class_key cq_lock_key;
 static struct lock_class_key qp_lock_key;
 static struct lock_class_key ah_lock_key;
 static struct lock_class_key srq_lock_key;
+static struct lock_class_key xrcd_lock_key;
 
 #define INIT_UDATA(udata, ibuf, obuf, ilen, olen)                      \
        do {                                                            \
@@ -255,6 +256,18 @@ static void put_srq_read(struct ib_srq *srq)
        put_uobj_read(srq->uobject);
 }
 
+static struct ib_xrcd *idr_read_xrcd(int xrcd_handle, struct ib_ucontext *context,
+                                    struct ib_uobject **uobj)
+{
+       *uobj = idr_read_uobj(&ib_uverbs_xrcd_idr, xrcd_handle, context, 0);
+       return *uobj ? (*uobj)->object : NULL;
+}
+
+static void put_xrcd_read(struct ib_uobject *uobj)
+{
+       put_uobj_read(uobj);
+}
+
 ssize_t ib_uverbs_get_context(struct ib_uverbs_file *file,
                              const char __user *buf,
                              int in_len, int out_len)
@@ -298,6 +311,7 @@ ssize_t ib_uverbs_get_context(struct ib_uverbs_file *file,
        INIT_LIST_HEAD(&ucontext->qp_list);
        INIT_LIST_HEAD(&ucontext->srq_list);
        INIT_LIST_HEAD(&ucontext->ah_list);
+       INIT_LIST_HEAD(&ucontext->xrcd_list);
        ucontext->closing = 0;
 
        resp.num_comp_vectors = file->device->num_comp_vectors;
@@ -579,6 +593,310 @@ ssize_t ib_uverbs_dealloc_pd(struct ib_uverbs_file *file,
        return in_len;
 }
 
+struct xrcd_table_entry {
+       struct rb_node  node;
+       struct ib_xrcd *xrcd;
+       struct inode   *inode;
+};
+
+static int xrcd_table_insert(struct ib_uverbs_device *dev,
+                           struct inode *inode,
+                           struct ib_xrcd *xrcd)
+{
+       struct xrcd_table_entry *entry, *scan;
+       struct rb_node **p = &dev->xrcd_tree.rb_node;
+       struct rb_node *parent = NULL;
+
+       entry = kmalloc(sizeof *entry, GFP_KERNEL);
+       if (!entry)
+               return -ENOMEM;
+
+       entry->xrcd  = xrcd;
+       entry->inode = inode;
+
+       while (*p) {
+               parent = *p;
+               scan = rb_entry(parent, struct xrcd_table_entry, node);
+
+               if (inode < scan->inode) {
+                       p = &(*p)->rb_left;
+               } else if (inode > scan->inode) {
+                       p = &(*p)->rb_right;
+               } else {
+                       kfree(entry);
+                       return -EEXIST;
+               }
+       }
+
+       rb_link_node(&entry->node, parent, p);
+       rb_insert_color(&entry->node, &dev->xrcd_tree);
+       igrab(inode);
+       return 0;
+}
+
+static struct xrcd_table_entry *xrcd_table_search(struct ib_uverbs_device *dev,
+                                                 struct inode *inode)
+{
+       struct xrcd_table_entry *entry;
+       struct rb_node *p = dev->xrcd_tree.rb_node;
+
+       while (p) {
+               entry = rb_entry(p, struct xrcd_table_entry, node);
+
+               if (inode < entry->inode)
+                       p = p->rb_left;
+               else if (inode > entry->inode)
+                       p = p->rb_right;
+               else
+                       return entry;
+       }
+
+       return NULL;
+}
+
+static struct ib_xrcd *find_xrcd(struct ib_uverbs_device *dev, struct inode *inode)
+{
+       struct xrcd_table_entry *entry;
+
+       entry = xrcd_table_search(dev, inode);
+       if (!entry)
+               return NULL;
+
+       return entry->xrcd;
+}
+
+static void xrcd_table_delete(struct ib_uverbs_device *dev,
+                             struct inode *inode)
+{
+       struct xrcd_table_entry *entry;
+
+       entry = xrcd_table_search(dev, inode);
+       if (entry) {
+               iput(inode);
+               rb_erase(&entry->node, &dev->xrcd_tree);
+               kfree(entry);
+       }
+}
+
+ssize_t ib_uverbs_open_xrcd(struct ib_uverbs_file *file,
+                           const char __user *buf, int in_len,
+                           int out_len)
+{
+       struct ib_uverbs_open_xrcd      cmd;
+       struct ib_uverbs_open_xrcd_resp resp;
+       struct ib_udata                 udata;
+       struct ib_uxrcd_object         *obj;
+       struct ib_xrcd                 *xrcd = NULL;
+       struct file                    *f = NULL;
+       struct inode                   *inode = NULL;
+       int                             ret = 0;
+       int                             new_xrcd = 0;
+
+       if (out_len < sizeof resp)
+               return -ENOSPC;
+
+       if (copy_from_user(&cmd, buf, sizeof cmd))
+               return -EFAULT;
+
+       INIT_UDATA(&udata, buf + sizeof cmd,
+                  (unsigned long) cmd.response + sizeof resp,
+                  in_len - sizeof cmd, out_len - sizeof  resp);
+
+       mutex_lock(&file->device->xrcd_tree_mutex);
+
+       if (cmd.fd != -1) {
+               /* search for file descriptor */
+               f = fget(cmd.fd);
+               if (!f) {
+                       ret = -EBADF;
+                       goto err_tree_mutex_unlock;
+               }
+
+               inode = f->f_dentry->d_inode;
+               if (!inode) {
+                       ret = -EBADF;
+                       goto err_tree_mutex_unlock;
+               }
+
+               xrcd = find_xrcd(file->device, inode);
+               if (!xrcd && !(cmd.oflags & O_CREAT)) {
+                       /* no file descriptor. Need CREATE flag */
+                       ret = -EAGAIN;
+                       goto err_tree_mutex_unlock;
+               }
+
+               if (xrcd && cmd.oflags & O_EXCL) {
+                       ret = -EINVAL;
+                       goto err_tree_mutex_unlock;
+               }
+       }
+
+       obj = kmalloc(sizeof *obj, GFP_KERNEL);
+       if (!obj) {
+               ret = -ENOMEM;
+               goto err_tree_mutex_unlock;
+       }
+
+       init_uobj(&obj->uobject, 0, file->ucontext, &xrcd_lock_key);
+
+       down_write(&obj->uobject.mutex);
+
+       if (!xrcd) {
+               xrcd = file->device->ib_dev->alloc_xrcd(file->device->ib_dev,
+                                                       file->ucontext, &udata);
+               if (IS_ERR(xrcd)) {
+                       ret = PTR_ERR(xrcd);
+                       goto err;
+               }
+
+               xrcd->inode   = inode;
+               xrcd->device  = file->device->ib_dev;
+               atomic_set(&xrcd->usecnt, 0);
+               mutex_init(&xrcd->tgt_qp_mutex);
+               INIT_LIST_HEAD(&xrcd->tgt_qp_list);
+               new_xrcd = 1;
+       }
+
+       atomic_set(&obj->refcnt, 0);
+       obj->uobject.object = xrcd;
+       ret = idr_add_uobj(&ib_uverbs_xrcd_idr, &obj->uobject);
+       if (ret)
+               goto err_idr;
+
+       memset(&resp, 0, sizeof resp);
+       resp.xrcd_handle = obj->uobject.id;
+
+       if (inode) {
+               if (new_xrcd) {
+                       /* create new inode/xrcd table entry */
+                       ret = xrcd_table_insert(file->device, inode, xrcd);
+                       if (ret)
+                               goto err_insert_xrcd;
+               }
+               atomic_inc(&xrcd->usecnt);
+       }
+
+       if (copy_to_user((void __user *) (unsigned long) cmd.response,
+                        &resp, sizeof resp)) {
+               ret = -EFAULT;
+               goto err_copy;
+       }
+
+       if (f)
+               fput(f);
+
+       mutex_lock(&file->mutex);
+       list_add_tail(&obj->uobject.list, &file->ucontext->xrcd_list);
+       mutex_unlock(&file->mutex);
+
+       obj->uobject.live = 1;
+       up_write(&obj->uobject.mutex);
+
+       mutex_unlock(&file->device->xrcd_tree_mutex);
+       return in_len;
+
+err_copy:
+       if (inode) {
+               if (new_xrcd)
+                       xrcd_table_delete(file->device, inode);
+               atomic_dec(&xrcd->usecnt);
+       }
+
+err_insert_xrcd:
+       idr_remove_uobj(&ib_uverbs_xrcd_idr, &obj->uobject);
+
+err_idr:
+       ib_dealloc_xrcd(xrcd);
+
+err:
+       put_uobj_write(&obj->uobject);
+
+err_tree_mutex_unlock:
+       if (f)
+               fput(f);
+
+       mutex_unlock(&file->device->xrcd_tree_mutex);
+
+       return ret;
+}
+
+ssize_t ib_uverbs_close_xrcd(struct ib_uverbs_file *file,
+                            const char __user *buf, int in_len,
+                            int out_len)
+{
+       struct ib_uverbs_close_xrcd cmd;
+       struct ib_uobject           *uobj;
+       struct ib_xrcd              *xrcd = NULL;
+       struct inode                *inode = NULL;
+       struct ib_uxrcd_object      *obj;
+       int                         live;
+       int                         ret = 0;
+
+       if (copy_from_user(&cmd, buf, sizeof cmd))
+               return -EFAULT;
+
+       mutex_lock(&file->device->xrcd_tree_mutex);
+       uobj = idr_write_uobj(&ib_uverbs_xrcd_idr, cmd.xrcd_handle, file->ucontext);
+       if (!uobj) {
+               ret = -EINVAL;
+               goto out;
+       }
+
+       xrcd  = uobj->object;
+       inode = xrcd->inode;
+       obj   = container_of(uobj, struct ib_uxrcd_object, uobject);
+       if (atomic_read(&obj->refcnt)) {
+               put_uobj_write(uobj);
+               ret = -EBUSY;
+               goto out;
+       }
+
+       if (!inode || atomic_dec_and_test(&xrcd->usecnt)) {
+               ret = ib_dealloc_xrcd(uobj->object);
+               if (!ret)
+                       uobj->live = 0;
+       }
+
+       live = uobj->live;
+       if (inode && ret)
+               atomic_inc(&xrcd->usecnt);
+
+       put_uobj_write(uobj);
+
+       if (ret)
+               goto out;
+
+       if (inode && !live)
+               xrcd_table_delete(file->device, inode);
+
+       idr_remove_uobj(&ib_uverbs_xrcd_idr, uobj);
+       mutex_lock(&file->mutex);
+       list_del(&uobj->list);
+       mutex_unlock(&file->mutex);
+
+       put_uobj(uobj);
+       ret = in_len;
+
+out:
+       mutex_unlock(&file->device->xrcd_tree_mutex);
+       return ret;
+}
+
+void ib_uverbs_dealloc_xrcd(struct ib_uverbs_device *dev,
+                           struct ib_xrcd *xrcd)
+{
+       struct inode *inode;
+
+       inode = xrcd->inode;
+       if (inode && !atomic_dec_and_test(&xrcd->usecnt))
+               return;
+
+       ib_dealloc_xrcd(xrcd);
+
+       if (inode)
+               xrcd_table_delete(dev, inode);
+}
+
 ssize_t ib_uverbs_reg_mr(struct ib_uverbs_file *file,
                         const char __user *buf, int in_len,
                         int out_len)
index 56898b6578a49e4e4b3d006763b6f6a861a78d93..bb9dcea8feded3803f1b026ab585acf0c87f6db9 100644 (file)
@@ -72,6 +72,7 @@ DEFINE_IDR(ib_uverbs_ah_idr);
 DEFINE_IDR(ib_uverbs_cq_idr);
 DEFINE_IDR(ib_uverbs_qp_idr);
 DEFINE_IDR(ib_uverbs_srq_idr);
+DEFINE_IDR(ib_uverbs_xrcd_idr);
 
 static DEFINE_SPINLOCK(map_lock);
 static DECLARE_BITMAP(dev_map, IB_UVERBS_MAX_DEVICES);
@@ -107,6 +108,8 @@ static ssize_t (*uverbs_cmd_table[])(struct ib_uverbs_file *file,
        [IB_USER_VERBS_CMD_MODIFY_SRQ]          = ib_uverbs_modify_srq,
        [IB_USER_VERBS_CMD_QUERY_SRQ]           = ib_uverbs_query_srq,
        [IB_USER_VERBS_CMD_DESTROY_SRQ]         = ib_uverbs_destroy_srq,
+       [IB_USER_VERBS_CMD_OPEN_XRCD]           = ib_uverbs_open_xrcd,
+       [IB_USER_VERBS_CMD_CLOSE_XRCD]          = ib_uverbs_close_xrcd,
 };
 
 static void ib_uverbs_add_one(struct ib_device *device);
@@ -241,6 +244,18 @@ static int ib_uverbs_cleanup_ucontext(struct ib_uverbs_file *file,
                kfree(uobj);
        }
 
+       mutex_lock(&file->device->xrcd_tree_mutex);
+       list_for_each_entry_safe(uobj, tmp, &context->xrcd_list, list) {
+               struct ib_xrcd *xrcd = uobj->object;
+               struct ib_uxrcd_object *uxrcd =
+                       container_of(uobj, struct ib_uxrcd_object, uobject);
+
+               idr_remove_uobj(&ib_uverbs_xrcd_idr, uobj);
+               ib_uverbs_dealloc_xrcd(file->device, xrcd);
+               kfree(uxrcd);
+       }
+       mutex_unlock(&file->device->xrcd_tree_mutex);
+
        list_for_each_entry_safe(uobj, tmp, &context->pd_list, list) {
                struct ib_pd *pd = uobj->object;
 
@@ -741,6 +756,8 @@ static void ib_uverbs_add_one(struct ib_device *device)
 
        kref_init(&uverbs_dev->ref);
        init_completion(&uverbs_dev->comp);
+       uverbs_dev->xrcd_tree = RB_ROOT;
+       mutex_init(&uverbs_dev->xrcd_tree_mutex);
 
        spin_lock(&map_lock);
        devnum = find_first_zero_bit(dev_map, IB_UVERBS_MAX_DEVICES);
index 8c6da5bda4c68e9070d57ed404aa9e343da79934..a6d95e635699030c9518796d023212db7b1f5faf 100644 (file)
@@ -1096,6 +1096,7 @@ struct ib_xrcd *ib_alloc_xrcd(struct ib_device *device)
        xrcd = device->alloc_xrcd(device, NULL, NULL);
        if (!IS_ERR(xrcd)) {
                xrcd->device = device;
+               xrcd->inode = NULL;
                atomic_set(&xrcd->usecnt, 0);
                mutex_init(&xrcd->tgt_qp_mutex);
                INIT_LIST_HEAD(&xrcd->tgt_qp_list);
index fe5b05177a2cb4fe9195cb2b95e4f1aca0242759..b5cb63f41f0f9f239765d2ae4cdeeb8dbdde0afe 100644 (file)
@@ -81,7 +81,9 @@ enum {
        IB_USER_VERBS_CMD_MODIFY_SRQ,
        IB_USER_VERBS_CMD_QUERY_SRQ,
        IB_USER_VERBS_CMD_DESTROY_SRQ,
-       IB_USER_VERBS_CMD_POST_SRQ_RECV
+       IB_USER_VERBS_CMD_POST_SRQ_RECV,
+       IB_USER_VERBS_CMD_OPEN_XRCD,
+       IB_USER_VERBS_CMD_CLOSE_XRCD
 };
 
 /*
@@ -222,6 +224,21 @@ struct ib_uverbs_dealloc_pd {
        __u32 pd_handle;
 };
 
+struct ib_uverbs_open_xrcd {
+       __u64 response;
+       __u32 fd;
+       __u32 oflags;
+       __u64 driver_data[0];
+};
+
+struct ib_uverbs_open_xrcd_resp {
+       __u32 xrcd_handle;
+};
+
+struct ib_uverbs_close_xrcd {
+       __u32 xrcd_handle;
+};
+
 struct ib_uverbs_reg_mr {
        __u64 response;
        __u64 start;
index ac46dcf043587889912cbae8d28cc44fa195acac..dfd9b87b7ffd0eb05fe74edd9fa39160e21e1d71 100644 (file)
@@ -851,6 +851,7 @@ struct ib_ucontext {
        struct list_head        qp_list;
        struct list_head        srq_list;
        struct list_head        ah_list;
+       struct list_head        xrcd_list;
        int                     closing;
 };
 
@@ -881,6 +882,7 @@ struct ib_pd {
 struct ib_xrcd {
        struct ib_device       *device;
        atomic_t                usecnt; /* count all exposed resources */
+       struct inode           *inode;
 
        struct mutex            tgt_qp_mutex;
        struct list_head        tgt_qp_list;