From b6a1aa2e24c0e06099146134974caad159c0b4a0 Mon Sep 17 00:00:00 2001
From: Gleb Natapov <gleb@cloudius-systems.com>
Date: Mon, 3 Feb 2014 15:15:44 +0200
Subject: [PATCH] mmu: write back only dirty pages during file sync
Walk page table and write out only dirty pages during file sync instead
of writing back entire mapping.
Reviewed-by: Glauber Costa <glommer@cloudius-systems.com>
Signed-off-by: Gleb Natapov <gleb@cloudius-systems.com>
Signed-off-by: Pekka Enberg <penberg@cloudius-systems.com>
---
core/mmu.cc | 77 +++++++++++++++++++++++++++++++++++++++++++++++++----
1 file changed, 72 insertions(+), 5 deletions(-)
diff --git a/core/mmu.cc b/core/mmu.cc
index bea973340..3a4b7d0c0 100644
--- a/core/mmu.cc
+++ b/core/mmu.cc
@@ -570,6 +570,40 @@ public:
}
};
+template <typename T, account_opt Account = account_opt::no>
+class dirty_cleaner : public vma_operation<allocate_intermediate_opt::no, skip_empty_opt::yes, Account> {
+private:
+ bool do_flush;
+ T handler;
+public:
+ dirty_cleaner(T handler) : do_flush(false), handler(handler) {}
+ void small_page(hw_ptep ptep, uintptr_t offset) {
+ pt_element pte = ptep.read();
+ if (!pte.dirty()) {
+ return;
+ }
+ do_flush |= true;
+ pte.set_dirty(false);
+ ptep.write(pte);
+ handler(ptep.read().addr(false), offset);
+ }
+ bool huge_page(hw_ptep ptep, uintptr_t offset) {
+ pt_element pte = ptep.read();
+ if (!pte.dirty()) {
+ return true;
+ }
+ do_flush |= true;
+ pte.set_dirty(false);
+ ptep.write(pte);
+ handler(ptep.read().addr(true), offset, huge_page_size);
+ return true;
+ }
+ bool tlb_flush_needed(void) {return do_flush;}
+ void finalize() {
+ handler.finalize();
+ }
+};
+
class virt_to_phys_map :
public page_table_operation<allocate_intermediate_opt::no, skip_empty_opt::yes,
descend_opt::yes, once_opt::yes, split_opt::no> {
@@ -1228,18 +1262,51 @@ void file_vma::split(uintptr_t edge)
vma_list.insert(*n);
}
+class dirty_page_sync {
+ friend dirty_cleaner<dirty_page_sync, account_opt::yes>;
+ friend file_vma;
+private:
+ file *_file;
+ f_offset _offset;
+ uint64_t _size;
+ struct elm {
+ iovec iov;
+ off_t offset;
+ };
+ std::stack<elm> queue;
+ dirty_page_sync(file *file, f_offset offset, uint64_t size) : _file(file), _offset(offset), _size(size) {}
+ void operator()(phys addr, uintptr_t offset, size_t size) {
+ off_t off = _offset + offset;
+ size_t len = std::min(size, _size - off);
+ queue.push(elm{{phys_to_virt(addr), len}, off});
+ }
+ void operator()(phys addr, uintptr_t offset) {
+ (*this)(addr, offset, page_size);
+ }
+ void finalize() {
+ while(!queue.empty()) {
+ elm w = queue.top();
+ uio data{&w.iov, 1, w.offset, ssize_t(w.iov.iov_len), UIO_WRITE};
+ _file->write(&data, FOF_OFFSET);
+ queue.pop();
+ }
+ }
+};
+
error file_vma::sync(uintptr_t start, uintptr_t end)
{
if (!_shared)
return make_error(ENOMEM);
start = std::max(start, _range.start());
end = std::min(end, _range.end());
- auto fsize = ::size(_file);
uintptr_t size = end - start;
- auto off = offset(start);
- write(_file, addr(), off, std::min(size, fsize - off));
- auto err = sys_fsync(_file.get());
- return make_error(err);
+
+ dirty_page_sync sync(_file.get(), _offset, ::size(_file));
+ error err = no_error();
+ if (operate_range(dirty_cleaner<dirty_page_sync, account_opt::yes>(sync), (void*)start, size) != 0) {
+ err = make_error(sys_fsync(_file.get()));
+ }
+ return err;
}
int file_vma::validate_perm(unsigned perm)
--
GitLab