[bug report] RDMA: Convert put_page() to put_user_page*()

[Dan Carpenter <dan.carpenter@oracle.com>]

Hi John,

I'm not really sure who to send this bug report to so you got picked
a bit at random...

The patch ea996974589e: "RDMA: Convert put_page() to
put_user_page*()" from May 24, 2019, leads to the following Smatch
static checker warning:

	./include/linux/pagemap.h:897 folio_lock()
	warn: sleeping in atomic context

./include/linux/pagemap.h
    895 static inline void folio_lock(struct folio *folio)
    896 {
--> 897         might_sleep();
    898         if (!folio_trylock(folio))
    899                 __folio_lock(folio);
    900 }

The problem is that unpin_user_pages_dirty_lock() calls folio_lock()
which can sleep.

Here is the raw Smatch preempt output.  As you can see there are
several places which seem to call unpin_user_pages_dirty_lock() with
preempt disabled.

__usnic_uiom_reg_release() <- disables preempt
usnic_uiom_reg_get() <- disables preempt
-> usnic_uiom_put_pages()
rds_tcp_write_space() <- disables preempt
-> rds_send_path_drop_acked()
   -> rds_send_remove_from_sock()
      -> rds_message_put()
         -> rds_message_purge()
            -> rds_rdma_free_op()
rds_message_purge() <duplicate>
-> rds_atomic_free_op()
               -> unpin_user_pages_dirty_lock()
                  -> folio_lock()

Let's pull out the first example:

drivers/infiniband/hw/usnic/usnic_uiom.c
   228                spin_lock(&pd->lock);
   229                usnic_uiom_remove_interval(&pd->root, vpn_start,
   230                                                vpn_last, &rm_intervals);
   231                usnic_uiom_unmap_sorted_intervals(&rm_intervals, pd);
   232
   233                list_for_each_entry_safe(interval, tmp, &rm_intervals, link) {
   234                        if (interval->flags & IOMMU_WRITE)
   235                                writable = 1;
   236                        list_del(&interval->link);
   237                        kfree(interval);
   238                }
   239
   240                usnic_uiom_put_pages(&uiomr->chunk_list, dirty & writable);
                      ^^^^^^^^^^^^^^^^^^^^
We're holding a spin lock, but _put_pages() calls unpin_user_pages_dirty_lock().

   241                spin_unlock(&pd->lock);
   242        }

regards,
dan carpenter