On 6/22/2020 3:52 PM, Jason Gunthorpe wrote:
On Sun, Jun 21, 2020 at 10:01:24AM +0300, Yishai Hadas wrote:
On 6/19/2020 3:29 PM, Jason Gunthorpe wrote:
On Wed, Jun 17, 2020 at 10:45:46AM +0300, Yishai Hadas wrote:
+static struct ibv_context *verbs_import_device(int cmd_fd)
+{
+ struct verbs_device *verbs_device = NULL;
+ struct verbs_context *context_ex;
+ struct ibv_device **dev_list;
+ struct ibv_context *ctx = NULL;
+ struct stat st;
+ int i;
+
+ if (fstat(cmd_fd, &st) || !S_ISCHR(st.st_mode)) {
+ errno = EINVAL;
+ return NULL;
+ }
+
+ dev_list = ibv_get_device_list(NULL);
+ if (!dev_list) {
+ errno = ENODEV;
+ return NULL;
+ }
+
+ for (i = 0; dev_list[i]; ++i) {
+ if (verbs_get_device(dev_list[i])->sysfs->sysfs_cdev ==
+ st.st_rdev) {
+ verbs_device = verbs_get_device(dev_list[i]);
+ break;
+ }
+ }
Unfortunately it looks like there is a small race here, the struct
ib_uverbs_file can exist beyond the lifetime of the cdev number - the
uverbs_ida is freed in ib_uverbs_remove_one() and files can still be
open past that point.
Are you referring to the option that we might end up with importing a device
that was already dissociated ? the below call to ops->import_context() will
just fail with -EIO upon calling on this FD to the query_context method, so
I believe that we should be fine here.
Okay, lets have a comment then, it is tricky
Sure, I have updated the PR accordingly.
It includes also fixes to other notes that you already pointed on, thanks.
Yishai
