bool ProcMemInfo::ReadVmaStats(int pagemap_fd, Vma& vma, bool get_wss, bool use_pageidle,
bool swap_only) {
PageAcct& pinfo = PageAcct::Instance();
if (get_wss && use_pageidle && !pinfo.InitPageAcct(true)) {
LOG(ERROR) << "Failed to init idle page accounting";
return false;
}
uint64_t pagesz_kb = getpagesize() / 1024;
size_t num_pages = (vma.end - vma.start) / (pagesz_kb * 1024);
size_t first_page = vma.start / (pagesz_kb * 1024);
std::vector<uint64_t> page_cache;
size_t cur_page_cache_index = 0;
size_t num_in_page_cache = 0;
size_t num_leftover_pages = num_pages;
for (size_t cur_page = first_page; cur_page < first_page + num_pages; ++cur_page) {
// Cache page map data.
if (cur_page_cache_index == num_in_page_cache) {
static constexpr size_t kMaxPages = 2048;
num_leftover_pages -= num_in_page_cache;
if (num_leftover_pages > kMaxPages) {
num_in_page_cache = kMaxPages;
} else {
num_in_page_cache = num_leftover_pages;
}
page_cache.resize(num_in_page_cache);
size_t total_bytes = page_cache.size() * sizeof(uint64_t);
ssize_t bytes = pread64(pagemap_fd, page_cache.data(), total_bytes,
cur_page * sizeof(uint64_t));
if (bytes != total_bytes) {
if (bytes == -1) {
PLOG(ERROR) << "Failed to read page data at offset 0x" << std::hex
<< cur_page * sizeof(uint64_t);
} else {
LOG(ERROR) << "Failed to read page data at offset 0x" << std::hex
<< cur_page * sizeof(uint64_t) << std::dec << " read bytes " << bytes
<< " expected bytes " << total_bytes;
}
return false;
}
cur_page_cache_index = 0;
}
uint64_t page_info = page_cache[cur_page_cache_index++];
if (!PAGE_PRESENT(page_info) && !PAGE_SWAPPED(page_info)) continue;
if (PAGE_SWAPPED(page_info)) {
vma.usage.swap += pagesz_kb;
swap_offsets_.emplace_back(PAGE_SWAP_OFFSET(page_info));
continue;
}
if (swap_only)
continue;
uint64_t page_frame = PAGE_PFN(page_info);
uint64_t cur_page_flags;
if (!pinfo.PageFlags(page_frame, &cur_page_flags)) {
LOG(ERROR) << "Failed to get page flags for " << page_frame << " in process " << pid_;
swap_offsets_.clear();
return false;
}
if (KPAGEFLAG_THP(cur_page_flags)) {
vma.usage.thp += pagesz_kb;
}
// skip unwanted pages from the count
if ((cur_page_flags & pgflags_mask_) != pgflags_) continue;
uint64_t cur_page_counts;
if (!pinfo.PageMapCount(page_frame, &cur_page_counts)) {
LOG(ERROR) << "Failed to get page count for " << page_frame << " in process " << pid_;
swap_offsets_.clear();
return false;
}
// Page was unmapped between the presence check at the beginning of the loop and here.
if (cur_page_counts == 0) {
continue;
}
bool is_dirty = !!(cur_page_flags & (1 << KPF_DIRTY));
bool is_private = (cur_page_counts == 1);
// Working set
if (get_wss) {
bool is_referenced = use_pageidle ? (pinfo.IsPageIdle(page_frame) == 1)
: !!(cur_page_flags & (1 << KPF_REFERENCED));
if (!is_referenced) {
continue;
}
// This effectively makes vss = rss for the working set is requested.
// The libpagemap implementation returns vss > rss for
// working set, which doesn't make sense.
vma.usage.vss += pagesz_kb;
}
vma.usage.rss += pagesz_kb;
vma.usage.uss += is_private ? pagesz_kb : 0;
vma.usage.pss += pagesz_kb / cur_page_counts;
if (is_private) {
vma.usage.private_dirty += is_dirty ? pagesz_kb : 0;
vma.usage.private_clean += is_dirty ? 0 : pagesz_kb;
} else {
vma.usage.shared_dirty += is_dirty ? pagesz_kb : 0;
vma.usage.shared_clean += is_dirty ? 0 : pagesz_kb;
}
}
if (!get_wss) {
vma.usage.vss += pagesz_kb * num_pages;
}
return true;
}
为什么要使用条件cur_page_cache_index == num_in_page_cache
这个条件的含义是:只有当cur_page_cache_index
等于num_in_page_cache
时,才会执行缓存页映射数据的逻辑。这样的设计有几个考虑因素:
Efficient Cache Usage:
Minimize File Reads:
Optimize I/O Operations:
pread64
函数用于从文件中读取数据,而文件I/O是相对耗时的操作。只在必要时读取数据,可以降低系统的开销。Control Cache Size:
kMaxPages
),可以防止过度使用内存。这是有意义的,因为在某些情况下,VMA范围可能很大,但您不想一次性将所有数据都加载到内存中。这个条件的目的是确保在需要时才从文件中读取新的页映射数据,以最大程度地减少文件I/O操作,提高代码的效率。
参考