On Wed, Aug 31, 2022 at 04:58:58AM +0300, Jarkko Sakkinen wrote: > On Wed, Aug 31, 2022 at 04:55:24AM +0300, Jarkko Sakkinen wrote: > > On Tue, Aug 30, 2022 at 03:54:27PM -0700, Reinette Chatre wrote: > > > Hi Jarkko, > > > > > > On 8/29/2022 8:12 PM, Jarkko Sakkinen wrote: > > > > In sgx_init(), if misc_register() for the provision device fails, and > > > > neither sgx_drv_init() nor sgx_vepc_init() succeeds, then ksgxd will be > > > > prematurely stopped. > > > > > > I do not think misc_register() is required to fail for the scenario to > > > be triggered (rather use "or" than "and"?). Perhaps just > > > "In sgx_init(), if a failure is encountered after ksgxd is started > > > (via sgx_page_reclaimer_init()) ...". > > > > This would be the fixed version of the sentence: > > > > " > > In sgx_init(), if misc_register() fails or misc_register() succeeds but > > neither sgx_drv_init() nor sgx_vepc_init() succeeds, then ksgxd will be > > prematurely stopped. This may leave some unsanitized pages, which does > > not matter, because SGX will be disabled for the whole power cycle. > > " > > > > I want to keep the end states listed and not make it more abstract. > > > > The second sentence addresses the remark below. > > > > > To help the reader understand the subject of this patch it may help > > > to explain that prematurely stopping ksgxd may leave some > > > unsanitized pages, but that is not a problem since SGX cannot > > > be used on the platform anyway. > > > > > > > This triggers WARN_ON() because sgx_dirty_page_list ends up being > > > > non-empty, and dumps the call stack: > > > > > > > > > > Traces like below can be frowned upon. I recommend that you follow the > > > guidance in "Backtraces in commit mesages"(sic) in > > > Documentation/process/submitting-patches.rst. > > > > > > > [ 0.268592] WARNING: CPU: 6 PID: 83 at > > > > arch/x86/kernel/cpu/sgx/main.c:401 ksgxd+0x1b7/0x1d0 > > > > Is this good enough? I had not actually spotted this section before but > > nice that it exists. Apparently has been added in 5.12. > > > > >> > > > > > Ultimately this can crash the kernel, if the following is set: > > > > > > > > /proc/sys/kernel/panic_on_warn > > > > > > > > Print a simple warning instead, and improve the output by printing the > > > > number of unsanitized pages, in order to provide debug informnation for > > > > future needs. > > > > > > informnation -> information > > > > +1 > > > > > > > > > > > ... > > > > > > > Link: https://lore.kernel.org/linux-sgx/20220825051827.246698-1-jarkko@xxxxxxxxxx/T/#u > > > > Reported-by: Paul Menzel <pmenzel@xxxxxxxxxxxxx> > > > > Tested-by: Paul Menzel <pmenzel@xxxxxxxxxxxxx> > > > > Fixes: 51ab30eb2ad4 ("x86/sgx: Replace section->init_laundry_list with sgx_dirty_page_list") > > > > Signed-off-by: Jarkko Sakkinen <jarkko@xxxxxxxxxx> > > > > > > Should this go to stable? > > > > I guess it should. The hard reason for this that it can panic > > the kernel. > > > > > > > > > > > > > diff --git a/arch/x86/kernel/cpu/sgx/main.c b/arch/x86/kernel/cpu/sgx/main.c > > > > index 515e2a5f25bb..903100fcfce3 100644 > > > > --- a/arch/x86/kernel/cpu/sgx/main.c > > > > +++ b/arch/x86/kernel/cpu/sgx/main.c > > > > @@ -49,17 +49,20 @@ static LIST_HEAD(sgx_dirty_page_list); > > > > * Reset post-kexec EPC pages to the uninitialized state. The pages are removed > > > > * from the input list, and made available for the page allocator. SECS pages > > > > * prepending their children in the input list are left intact. > > > > + * > > > > + * Contents of the @dirty_page_list must be thread-local, i.e. > > > > + * not shared by multiple threads. > > > > > > Did you intend to mention something about the needed locking here? It looks > > > like some information is lost during the move to the function description. > > > > Nothing about the locking that concerns the parameter, as the > > sentence defines clear constraints for the caller. > > > > > > > > > */ > > > > -static void __sgx_sanitize_pages(struct list_head *dirty_page_list) > > > > +static int __sgx_sanitize_pages(struct list_head *dirty_page_list) > > > > { > > > > struct sgx_epc_page *page; > > > > + int left_dirty = 0; > > > > > > I do not know how many pages this code should be ready for but at least > > > this could handle more by being an unsigned int considering that it is > > > always positive ... maybe even unsigned long? > > > > I would go for 'long'. More information below. > > > > > > > > > LIST_HEAD(dirty); > > > > int ret; > > > > > > > > - /* dirty_page_list is thread-local, no need for a lock: */ > > > > while (!list_empty(dirty_page_list)) { > > > > if (kthread_should_stop()) > > > > - return; > > > > + break; > > > > > > > > page = list_first_entry(dirty_page_list, struct sgx_epc_page, list); > > > > > > > > @@ -92,12 +95,14 @@ static void __sgx_sanitize_pages(struct list_head *dirty_page_list) > > > > } else { > > > > /* The page is not yet clean - move to the dirty list. */ > > > > list_move_tail(&page->list, &dirty); > > > > + left_dirty++; > > > > } > > > > > > > > cond_resched(); > > > > } > > > > > > > > list_splice(&dirty, dirty_page_list); > > > > + return left_dirty; > > > > } > > > > > > > > static bool sgx_reclaimer_age(struct sgx_epc_page *epc_page) > > > > @@ -388,6 +393,8 @@ void sgx_reclaim_direct(void) > > > > > > > > static int ksgxd(void *p) > > > > { > > > > + int left_dirty; > > > > + > > > > set_freezable(); > > > > > > > > /* > > > > @@ -395,10 +402,10 @@ static int ksgxd(void *p) > > > > * required for SECS pages, whose child pages blocked EREMOVE. > > > > */ > > > > __sgx_sanitize_pages(&sgx_dirty_page_list); > > > > - __sgx_sanitize_pages(&sgx_dirty_page_list); > > > > > > > > - /* sanity check: */ > > > > - WARN_ON(!list_empty(&sgx_dirty_page_list)); > > > > + left_dirty = __sgx_sanitize_pages(&sgx_dirty_page_list); > > > > + if (left_dirty) > > > > + pr_warn("%d unsanitized pages\n", left_dirty); > > > > > > > > while (!kthread_should_stop()) { > > > > if (try_to_freeze()) > > > > > > > > > Reinette > > > > We need to return -ECANCELED on premature stop, and number of > > pages otherwise. > > > > In premature stop, nothing should be printed, as the number > > is by practical means a random number. Otherwise, it is an > > indicator of a bug in the driver, and therefore a non-zero > > number should be printed pr_err(), if that happens after the > > second call. > > I.e. even though we print less we get more *information* what > is going inside the kernel. Warning is not correct for either > path IMHO. Oh, sorry, I forgot one thing. The devices should be actually deinitialized in the error case. We do not want to leave a broken driver running. BR, Jarkko
