From 033b2876351f77d7f15c23e9dd4e232c2d30b25f Mon Sep 17 00:00:00 2001 From: Brian Gianforcaro Date: Mon, 19 Apr 2021 23:55:44 -0700 Subject: [PATCH] Kernel: Make AsyncDeviceRequest sub-req management alloc free The previous implementation could allocate on insertion into the completed / pending sub request vectors. There's no reason these can't be intrusive lists instead. This is a very minor step towards improving the ability to handle OOM, as tracked by #6369 It might also help improve performance on the IO path in certain situations. I'll benchmark that later. --- Kernel/Devices/AsyncDeviceRequest.cpp | 30 +++++++++++++-------------- Kernel/Devices/AsyncDeviceRequest.h | 11 +++++++--- 2 files changed, 22 insertions(+), 19 deletions(-) diff --git a/Kernel/Devices/AsyncDeviceRequest.cpp b/Kernel/Devices/AsyncDeviceRequest.cpp index 25977d1ee57..ad421389ebd 100644 --- a/Kernel/Devices/AsyncDeviceRequest.cpp +++ b/Kernel/Devices/AsyncDeviceRequest.cpp @@ -49,10 +49,13 @@ AsyncDeviceRequest::~AsyncDeviceRequest() // sub-requests should be completed (either succeeded, failed, or cancelled). // Which means there should be no more pending sub-requests and the // entire AsyncDeviceRequest hierarchy should be immutable. - for (auto& sub_request : m_sub_requests_complete) { - VERIFY(is_completed_result(sub_request.m_result)); // Shouldn't need any locking anymore - VERIFY(sub_request.m_parent_request == this); - sub_request.m_parent_request = nullptr; + while (!m_sub_requests_complete.is_empty()) { + // Note: sub_request is ref-counted, and we use this specific pattern + // to allow make sure the refcount is dropped properly. + auto sub_request = m_sub_requests_complete.take_first(); + VERIFY(is_completed_result(sub_request->m_result)); // Shouldn't need any locking anymore + VERIFY(sub_request->m_parent_request == this); + sub_request->m_parent_request = nullptr; } } @@ -104,24 +107,19 @@ void AsyncDeviceRequest::sub_request_finished(AsyncDeviceRequest& sub_request) { ScopedSpinLock lock(m_lock); VERIFY(m_result == Started); - size_t index; - for (index = 0; index < m_sub_requests_pending.size(); index++) { - if (&m_sub_requests_pending[index] == &sub_request) { - NonnullRefPtr request(m_sub_requests_pending[index]); - m_sub_requests_pending.remove(index); - m_sub_requests_complete.append(move(request)); - break; - } + + if (m_sub_requests_pending.contains(sub_request)) { + // Note: append handles removing from any previous intrusive list internally. + m_sub_requests_complete.append(sub_request); } - VERIFY(index < m_sub_requests_pending.size()); + all_completed = m_sub_requests_pending.is_empty(); if (all_completed) { // Aggregate any errors bool any_failures = false; bool any_memory_faults = false; - for (index = 0; index < m_sub_requests_complete.size(); index++) { - auto& sub_request = m_sub_requests_complete[index]; - auto sub_result = sub_request.get_request_result(); + for (auto& com_sub_request : m_sub_requests_complete) { + auto sub_result = com_sub_request.get_request_result(); VERIFY(is_completed_result(sub_result)); switch (sub_result) { case Failure: diff --git a/Kernel/Devices/AsyncDeviceRequest.h b/Kernel/Devices/AsyncDeviceRequest.h index 8084f5b971e..9c49ecfad12 100644 --- a/Kernel/Devices/AsyncDeviceRequest.h +++ b/Kernel/Devices/AsyncDeviceRequest.h @@ -26,7 +26,8 @@ #pragma once -#include +#include +#include #include #include #include @@ -160,8 +161,12 @@ private: AsyncDeviceRequest* m_parent_request { nullptr }; RequestResult m_result { Pending }; - NonnullRefPtrVector m_sub_requests_pending; - NonnullRefPtrVector m_sub_requests_complete; + IntrusiveListNode> m_list_node; + + typedef IntrusiveList, &AsyncDeviceRequest::m_list_node> AsyncDeviceSubRequestList; + + AsyncDeviceSubRequestList m_sub_requests_pending; + AsyncDeviceSubRequestList m_sub_requests_complete; WaitQueue m_queue; NonnullRefPtr m_process; void* m_private { nullptr };