summaryrefslogtreecommitdiff
path: root/Kernel/Syscalls/futex.cpp
diff options
context:
space:
mode:
authorAndreas Kling <kling@serenityos.org>2022-08-23 17:58:05 +0200
committerAndreas Kling <kling@serenityos.org>2022-08-24 14:57:51 +0200
commitcf16b2c8e64709d570c5f54a981017d217e95ed0 (patch)
tree16c9efdaaa579ae51682a51a58949ce02c3d2092 /Kernel/Syscalls/futex.cpp
parentd6ef18f587d4a7e4f58487c84e0b9eb260f3ec5a (diff)
downloadserenity-cf16b2c8e64709d570c5f54a981017d217e95ed0.zip
Kernel: Wrap process address spaces in SpinlockProtected
This forces anyone who wants to look into and/or manipulate an address space to lock it. And this replaces the previous, more flimsy, manual spinlock use. Note that pointers *into* the address space are not safe to use after you unlock the space. We've got many issues like this, and we'll have to track those down as wlel.
Diffstat (limited to 'Kernel/Syscalls/futex.cpp')
-rw-r--r--Kernel/Syscalls/futex.cpp66
1 files changed, 34 insertions, 32 deletions
diff --git a/Kernel/Syscalls/futex.cpp b/Kernel/Syscalls/futex.cpp
index 8289385065..6b12bd584a 100644
--- a/Kernel/Syscalls/futex.cpp
+++ b/Kernel/Syscalls/futex.cpp
@@ -18,7 +18,7 @@ static Singleton<SpinlockProtected<HashMap<GlobalFutexKey, NonnullLockRefPtr<Fut
void Process::clear_futex_queues_on_exec()
{
s_global_futex_queues->with([this](auto& queues) {
- auto const* address_space = &this->address_space();
+ auto const* address_space = this->address_space().with([](auto& space) { return space.ptr(); });
queues.remove_all_matching([address_space](auto& futex_key, auto& futex_queue) {
if ((futex_key.raw.offset & futex_key_private_flag) == 0)
return false;
@@ -45,45 +45,47 @@ ErrorOr<GlobalFutexKey> Process::get_futex_key(FlatPtr user_address, bool shared
if (!shared) { // If this is thread-shared, we can skip searching the matching region
return GlobalFutexKey {
.private_ = {
- .address_space = &address_space(),
+ .address_space = this->address_space().with([](auto& space) { return space.ptr(); }),
.user_address = user_address | futex_key_private_flag,
}
};
}
- auto* matching_region = address_space().find_region_containing(range);
- if (!matching_region)
- return EFAULT;
+ return address_space().with([&](auto& space) -> ErrorOr<GlobalFutexKey> {
+ auto* matching_region = space->find_region_containing(range);
+ if (!matching_region)
+ return EFAULT;
+
+ // The user wants to share this futex, but if the address doesn't point to a shared resource, there's not
+ // much sharing to be done, so let's mark this as private
+ if (!matching_region->is_shared()) {
+ return GlobalFutexKey {
+ .private_ = {
+ .address_space = space.ptr(),
+ .user_address = user_address | futex_key_private_flag,
+ }
+ };
+ }
+
+ // This address is backed by a shared VMObject, if it's an AnonymousVMObject, it can be shared between processes
+ // via forking, and shared regions that are cloned during a fork retain their original AnonymousVMObject.
+ // On the other hand, if it's a SharedInodeVMObject, it can be shared by two processes mapping the same file as
+ // MAP_SHARED, but since they are deduplicated based on the inode, in all cases the VMObject pointer should be
+ // a unique global identifier.
+ // NOTE: This assumes that a program will not unmap the only region keeping the vmobject alive while waiting on it,
+ // if it does, it will get stuck waiting forever until interrupted by a signal, but since that use case is defined as
+ // a programmer error, we are fine with it.
+
+ auto const& vmobject = matching_region->vmobject();
+ if (vmobject.is_inode())
+ VERIFY(vmobject.is_shared_inode());
- // The user wants to share this futex, but if the address doesn't point to a shared resource, there's not
- // much sharing to be done, so let's mark this as private
- if (!matching_region->is_shared()) {
return GlobalFutexKey {
- .private_ = {
- .address_space = &address_space(),
- .user_address = user_address | futex_key_private_flag,
- }
+ .shared = {
+ .vmobject = &vmobject,
+ .offset = matching_region->offset_in_vmobject_from_vaddr(range.base()) }
};
- }
-
- // This address is backed by a shared VMObject, if it's an AnonymousVMObject, it can be shared between processes
- // via forking, and shared regions that are cloned during a fork retain their original AnonymousVMObject.
- // On the other hand, if it's a SharedInodeVMObject, it can be shared by two processes mapping the same file as
- // MAP_SHARED, but since they are deduplicated based on the inode, in all cases the VMObject pointer should be
- // a unique global identifier.
- // NOTE: This assumes that a program will not unmap the only region keeping the vmobject alive while waiting on it,
- // if it does, it will get stuck waiting forever until interrupted by a signal, but since that use case is defined as
- // a programmer error, we are fine with it.
-
- auto const& vmobject = matching_region->vmobject();
- if (vmobject.is_inode())
- VERIFY(vmobject.is_shared_inode());
-
- return GlobalFutexKey {
- .shared = {
- .vmobject = &vmobject,
- .offset = matching_region->offset_in_vmobject_from_vaddr(range.base()) }
- };
+ });
}
ErrorOr<FlatPtr> Process::sys$futex(Userspace<Syscall::SC_futex_params const*> user_params)