Łukasz Plewa

> [@lplewa](https://github.com/lplewa) in fact in [#1063](https://github.com/oneapi-src/unified-memory-framework/pull/1063) I implemented a mix of 2a + b - lets call it just "2": Ok - my bad - i did not see macros...

> You can map version to size. Again, I don't care either way. We need size - user might get osprovider ops, duplicate them, and then i.e change one function...

There is multiple options: ``` c const umf_memory_provider_ops_t * ops = umfOsMemoryProviderOps(void); umf_memory_provider_ops_t *myops = malloc(ops.size); memcpy(myops, ops, ops.size); myops.free = &my_custom_free_function; // this code is wrong //umf_memory_provider_ops_t *myops =...

> First, I would prefer the `umfOsMemoryProviderOps` function to return a const pointer to `ops` so that the user cannot modify in-place. This is separate issue - please see: #1080...

> Furthermore, now I am thinking that even copying the ops structure and modifying a particular callback is error-prone because it breaks encapsulation. For example, to substitute `free` callback you...

Let me summarize discussion, - We stay with Version in ops structure (eventually we replace it with size) - If we stay with version we must replace UMF_VERSION_CURRENT for provider...

> Do you have a reproducer? Only as #899 - you have to remove sleep which i added to workaround this issue.

> What is the purpose of this PR, could you please describe the scenario and how these changes help? We have a race, between insert and delete of entry in...

Pool nesting detection. We are checking if there is region in the tracker that contains region that we are adding. So we are looking for region with address lower then...

@breiters Thanks for the issue. I created a PR with a fix for this problem #1309 - can you check if it's resolves the issue on your side?