On Sun, Jan 19, 2025 at 11:32:05AM +0100, Mateusz Guzik wrote:
> Dumping processes with large allocated and mostly not-faulted areas is
> very slow.
>
> Borrowing a test case from Tavian Barnes:
>
> int main(void) {
> char *mem = mmap(NULL, 1ULL << 40, PROT_READ | PROT_WRITE,
> MAP_ANONYMOUS | MAP_NORESERVE | MAP_PRIVATE, -1, 0);
> printf("%p %m\n", mem);
> if (mem != MAP_FAILED) {
> mem[0] = 1;
> }
> abort();
> }
>
> That's 1TB of almost completely not-populated area.
>
> On my test box it takes 13-14 seconds to dump.
>
> The profile shows:
> - 99.89% 0.00% a.out
> entry_SYSCALL_64_after_hwframe
> do_syscall_64
> syscall_exit_to_user_mode
> arch_do_signal_or_restart
> - get_signal
> - 99.89% do_coredump
> - 99.88% elf_core_dump
> - dump_user_range
> - 98.12% get_dump_page
> - 64.19% __get_user_pages
> - 40.92% gup_vma_lookup
> - find_vma
> - mt_find
> 4.21% __rcu_read_lock
> 1.33% __rcu_read_unlock
> - 3.14% check_vma_flags
> 0.68% vma_is_secretmem
> 0.61% __cond_resched
> 0.60% vma_pgtable_walk_end
> 0.59% vma_pgtable_walk_begin
> 0.58% no_page_table
> - 15.13% down_read_killable
> 0.69% __cond_resched
> 13.84% up_read
> 0.58% __cond_resched
>
> Almost 29% of the time is spent relocking the mmap semaphore between
> calls to get_dump_page() which find nothing.
>
> Whacking that results in times of 10 seconds (down from 13-14).
>
> While here make the thing killable.
>
> The real problem is the page-sized iteration and the real fix would
> patch it up instead. It is left as an exercise for the mm-familiar
> reader.
>
> Signed-off-by: Mateusz Guzik <mjguzik@xxxxxxxxx>
> ---
Seems like a good improvement to me.
Let's get it tested.
