Content-Type: multipart/mixed; boundary="===============6243888562832111493==" MIME-Version: 1.0 From: Masami Hiramatsu To: lkp@lists.01.org Subject: Re: [uaccess] 780464aed0: WARNING:at_arch/x86/include/asm/uaccess.h:#strnlen_user/0x Date: Tue, 05 Mar 2019 11:36:35 +0900 Message-ID: <20190305113635.18f80ea3b1f4fca54b9d21e4@kernel.org> In-Reply-To: List-Id: --===============6243888562832111493== Content-Type: text/plain; charset="utf-8" MIME-Version: 1.0 Content-Transfer-Encoding: quoted-printable On Mon, 4 Mar 2019 10:59:22 -0800 Linus Torvalds wrote: > On Mon, Mar 4, 2019 at 1:06 AM Masami Hiramatsu w= rote: > > > > On Sun, 3 Mar 2019 18:37:59 -0800 > > Linus Torvalds wrote: > > > We've had this before. We've gotten rid of the actual "use system > > > calls", but we still have some of the init sequence in particular just > > > calling the wrappers instead. > > > > Are those safe if we are in init sequence? > = > Yes, they are, it runs with set_fs(KERNEL_DS). > = > But the patches made that now complain about copying from non-user > space, even though it's fine. > = > Basically, "strncpy_from_user()" shouldn't use "user_access_ok()", > since it actually can take a kernel address (with set_fs()). OK, so strncpy_from_user() or any other copy_from_user() should be available for copying kernel memory if set_fs(KERNEL_DS). > Your "unsafe" version for tracing that actually sets "set_fs(USER_DS)" > is thje only thing that should use that helper. I see, it ensures it is accessing user-memory. > = > > > And yes, ksys_mount() takes __user pointers. > > > > > > It would be a lot better to use "do_mount()", which is the interface > > > that takes actual "char *" pointers. > > > > Unfortunately, it still takes a __user pointer. > = > Ahh, yes, the name remains in user space. > = > Besides, I'm sure you'd just hit other cases instead where people use > set_fs() and copy strings. Yeah, under init/ I saw such cases. > = > > So what we need is > > > > long do_mount(const char *dev_name, struct path *dir_path, > > const char *type_page, unsigned long flags, void *data_= page) > > > > or introduce kern_do_mount()? > = > It's actually fairly painful. Particularly because of that "void *data_pa= ge". Yeah, that is what I've hit while testing :-( > = > Your second email with "Would this work?" helper function _wopuldn't_ > work correctly, exactly because you passed in a regular string to the > data page. > = > Also, I don't want to see code that replaces the unconditional "copy > path from user space" with a conditional "do we have path in kernel > space". Yes, that's just a hack :) > = > So together with the whole "uyou'll hit other peoblems anyway", I > don't think this is a good approach. > = > I think you simply need to have a separate "unsafe_strncpy()" > function, and not change the existing "strncpy_from_user()". Would you mean implementing yet another "strncpy_from_user without pagefault"? = What we changed here is just use user_access_ok() instead access_ok() = in user_access_begin() because access_ok() may cause false-positive warning if we use it in IRQ. I think the better way to do this is allowing strncpy_from_user() if some conditions are match, like - strncpy_from_user() will be able to copy user memory with set_fs(USER_DS) - strncpy_from_user() can copy kernel memory with set_fs(KERNEL_DS) - strncpy_from_user() can access unsafe memory in IRQ context if pagefault is disabled. This is almost done, except for CONFIG_DEBUG_ATOMIC_SLEEP=3Dy on x86. So, what about adding a condition to WARN_ON_IN_IRQ() like below instead of introducing user_access_ok() ? diff --git a/arch/x86/include/asm/uaccess.h b/arch/x86/include/asm/uaccess.h index 780f2b42c8ef..ec0f0b74c9ab 100644 --- a/arch/x86/include/asm/uaccess.h +++ b/arch/x86/include/asm/uaccess.h @@ -70,7 +70,7 @@ static inline bool __chk_range_not_ok(unsigned long addr,= unsigned long size, un }) = #ifdef CONFIG_DEBUG_ATOMIC_SLEEP -# define WARN_ON_IN_IRQ() WARN_ON_ONCE(!in_task()) +# define WARN_ON_IN_IRQ() WARN_ON_ONCE(pagefault_disabled() && !in_ta= sk()) #else # define WARN_ON_IN_IRQ() #endif Of course we have to move pagefault_disabled() macro in somewhere better pl= ace. Thank you, -- = Masami Hiramatsu --===============6243888562832111493==--