On Wed, Feb 17, 2021 at 06:06:48PM +0100, Andrew Lunn wrote: > > I'm wondering whether we need to add __acquires() and __releases() > > annotations to some of these functions so that sparse can catch > > these cases. Thoughts? > > Hi Russell > > The more tools we have for catching locking problems the better. > Jakubs patchwork bot should then catch them when a patch is submitted, > if the developer did not run sparse themselves. Here is how I wrote the check for Smatch. The code in the kernel looks like: oldpage = phy_select_page(phydev, 0x0007); ... phy_restore_page(phydev, oldpage, 0); So what I said is that if phy_select_page() returns an error code then set "phydev" to &selected state. Then if we call phy_restore_page() set it to &undefined. When we hit a return, check if we have any "phydev" variables can possibly be in &selected state and print a warning. The code is below. regards, dan carpenter #include "smatch.h" #include "smatch_slist.h" static int my_id; STATE(selected); static sval_t err_min = { .type = &int_ctype, .value = -4095 }; static sval_t err_max = { .type = &int_ctype, .value = -1 }; static void match_phy_select_page(struct expression *expr, const char *name, struct symbol *sym, void *data) { set_state(my_id, name, sym, &selected); } static void match_phy_restore_page(struct expression *expr, const char *name, struct symbol *sym, void *data) { set_state(my_id, name, sym, &undefined); } static void match_return(struct expression *expr) { struct sm_state *sm; FOR_EACH_MY_SM(my_id, __get_cur_stree(), sm) { if (slist_has_state(sm->possible, &selected)) { sm_warning("phy_select_page() requires restore on error"); return; } } END_FOR_EACH_SM(sm); } void check_phy_select_page_fail(int id) { if (option_project != PROJ_KERNEL) return; my_id = id; return_implies_param_key("phy_select_page", err_min, err_max, &match_phy_select_page, 0, "$", NULL); add_function_param_key_hook("phy_restore_page", &match_phy_restore_page, 0, "$", NULL); add_hook(&match_return, RETURN_HOOK); }