The percpu counters can only deal with long ints but max_files is declared as an unsigned long. We have already started to reject anything exceeding LONG_INT in the file-max sysctl as well [1]. Should we ever have the need and precision for unsigned long percpu counters we can simply switch to that api for the sysctls and then lift the LONG_MAX cap from files_maxfiles_init() as well.
Note, that in a quick discussion with Tejun investigating percpu unsigned long counters it became obvious that technically, both LONG_MAX and ULONG_MAX should be unreachable. Given that an fd is usually larger than 2 bytes one would likely run out of address space before even reaching the LONG_MAX limit. [1]: https://lkml.org/lkml/2018/10/17/74 Cc: Al Viro <[email protected]> Cc: Kees Cook <[email protected]> Signed-off-by: Christian Brauner <[email protected]> --- fs/file_table.c | 8 ++++++++ 1 file changed, 8 insertions(+) diff --git a/fs/file_table.c b/fs/file_table.c index e49af4caf15d..2e3cddce20d6 100644 --- a/fs/file_table.c +++ b/fs/file_table.c @@ -10,6 +10,7 @@ #include <linux/file.h> #include <linux/fdtable.h> #include <linux/init.h> +#include <linux/kernel.h> #include <linux/module.h> #include <linux/fs.h> #include <linux/security.h> @@ -386,4 +387,11 @@ void __init files_maxfiles_init(void) n = ((totalram_pages - memreserve) * (PAGE_SIZE / 1024)) / 10; files_stat.max_files = max_t(unsigned long, n, NR_FILE); + + /* + * The percpu counters only handle long ints so cap maximum number of + * files at LONG_MAX. + */ + if (files_stat.max_files > LONG_MAX) + files_stat.max_files = LONG_MAX; } -- 2.19.1
