From a discussion with @lokokung about making mapping thread-safe in Dawn (without taking device-wide locks).
GetMappedRange()/GetConstMappedRange() return raw pointers which means that calling code always needs to be thread-safe w.r.t. Unmap(). For example this can never be safe:
- Thread 1:
p = GetMappedRange(); mutex A { use p }
- Thread 2:
mutex A { Unmap() }
because the unmap could happen between GetMappedRange and a lock being taken on A. The only way to be safe is to include GetMappedRange in the lock:
- Thread 1:
mutex A { p = GetMappedRange(); use p }
- Thread 2:
mutex A { Unmap() }
Thus, I think there is no use in having GetMappedRange be thread-safe w.r.t. Unmap.
GetMappedRange should still be safe w.r.t. MapAsync, as you could do something like:
- Thread 1:
MapAsync()
- Thread 2:
p = GetMappedRange(); if (p) { do something if the map happens to be complete; Unmap() }
(even though there are alternatives like using GetMapState or polling the MapAsync future).
Aside: technically it's not very useful for MapAsync and Unmap to be thread-safe w.r.t. one another, but I think they still should be, as they're not as hot.
From a discussion with @lokokung about making mapping thread-safe in Dawn (without taking device-wide locks).
GetMappedRange()/GetConstMappedRange()return raw pointers which means that calling code always needs to be thread-safe w.r.t.Unmap(). For example this can never be safe:p = GetMappedRange(); mutex A { use p }mutex A { Unmap() }because the unmap could happen between GetMappedRange and a lock being taken on A. The only way to be safe is to include GetMappedRange in the lock:
mutex A { p = GetMappedRange(); use p }mutex A { Unmap() }Thus, I think there is no use in having GetMappedRange be thread-safe w.r.t. Unmap.
GetMappedRange should still be safe w.r.t. MapAsync, as you could do something like:
MapAsync()p = GetMappedRange(); if (p) { do something if the map happens to be complete; Unmap() }(even though there are alternatives like using GetMapState or polling the MapAsync future).
Aside: technically it's not very useful for MapAsync and Unmap to be thread-safe w.r.t. one another, but I think they still should be, as they're not as hot.