Problem

Multi-threaded workloads with many syscalls stress the VMA subsystem a lot, because almost all syscalls verify their buffers for read/write access using the following functions:

• is_user_memory_readable()
• is_user_memory_writable()
• is_user_string_readable()
• is_user_memory_writable_no_skip()

All these functions call test_user_memory() helper:

This helper in turn calls the is_in_adjacent_user_vmas() func:

The important part is spinlock_lock(&vma_tree_lock) and spinlock_unlock(&vma_tree_lock). On a multi-threaded app, this lock contention becomes the bottleneck.

Gramine introduced a workaround to sidestep this bottleneck, via the libos.check_invalid_pointers manifest option; it translates to the g_check_invalid_ptrs variable. However, this cannot be used in all cases:

• Some runtimes like Java rely on being able to check invalid pointers. Thus they cannot set libos.check_invalid_pointers = false; this would lead to Java apps failing.
• The is_user_memory_writable_no_skip() function does not honor the libos.check_invalid_pointers manifest option; this is because in certain situations Gramine really must decide whether the VMA is writable or read-only, see e.g. the ppoll() case which emulates how Linux works.

Solution

Use the RW lock that was previously introduced in the Gramine codebase: https://github.com/gramineproject/gramine/blob/master/libos/include/libos_rwlock.h

Example usage of this RW lock: f071450

Benchmark results

TODO