Arnd Bergmann <arnd@xxxxxxxx> writes: > On Fri, Nov 30, 2018 at 7:56 AM Christian Brauner <christian@xxxxxxxxxx> wrote: >> On Thu, Nov 29, 2018 at 11:13:57PM -0600, Eric W. Biederman wrote: >> > Arnd Bergmann <arnd@xxxxxxxx> writes: >> > > On Thu, Nov 29, 2018 at 9:14 PM Andy Lutomirski <luto@xxxxxxxxxxxxxx> wrote: >> > > >> > > It looks like we already have a 'struct signalfd_siginfo' that is defined in a >> > > sane architecture-independent way, so I'd suggest we use that. >> > >> > Unfortunately it isn't maintained very well. What you can >> > express with signalfd_siginfo is a subset what you can express with >> > siginfo. Many of the linux extensions to siginfo for exception >> > information add pointers and have integers right after those pointers. >> > Not all of those linux specific extentions for exceptions are handled >> > by signalfd_siginfo (it needs new fields). > > Would those fit in the 30 remaining padding bytes? Probably. I expect at some point the technique signalfd has been using won't scale. Most of what are missing are synchronous exceptions but the crazy seccomp extensions might be missing as well. I say crazy because seccomp places an integer immediately after a pointer which makes 32bit 64bit compatibility impossible. >> > As originally defined siginfo had the sigval_t union at the end so it >> > was perfectly fine on both 32bit and 64bit as it only had a single >> > pointer in the structure and the other fields were 32bits in size. > > The problem with sigval_t of course is that it is incompatible between > 32-bit and 64-bit. We can add the same information, but at least on > the syscall level that would have to be a __u64. But sigval_t with nothing after it is not incompatible between 32bit and 64bit. With nothing after it sigval_t in struct siginfo already has the extra 32bits you would need. It is sort of like transparently adding a __u64 member to the union. That is where we really are with sigval_t. Except for some crazy corner cases. >> > Although I do feel the pain as x86_64 has to deal with 3 versions >> > of siginfo. A 64bit one. A 32bit one for ia32. A 32bit one for x32 >> > with a 64bit si_clock_t. > > At least you and Al have managed to get it down to a single source-level > definition across all architectures, but there is also the (lesser) problem > that the structure has a slightly different binary layout on each of the > classic architectures. > > If we go with Andy's suggestion of having only a single binary layout > on x86 for the new call, I'd argue that we should also make it have > the exact same layout on all other architectures. Yes. >> > > We may then also want to make sure that any system call that takes a >> > > siginfo has a replacement that takes a signalfd_siginfo, and that this >> > > replacement can be used to implement the old version purely in >> > > user space. >> > >> > If you are not implementing CRIU and reinserting exceptions to yourself. >> > At most user space wants the ability to implement sigqueue: >> > >> > AKA: >> > sigqueue(pid_t pid, int sig, const union sigval value); >> > >> > Well sigqueue with it's own si_codes so the following would cover all >> > the use cases I know of: >> > int sendnewsig(pid_t pid, int sig, int si_code, const union sigval value); >> > >> > The si_code could even be set to SI_USER to request that the normal >> > trusted SI_USER values be filled in. si_code values of < 0 if not >> > recognized could reasonably safely be treated as the _rt member of >> > the siginfo union. Or perhaps better we error out in such a case. >> > >> > If we want to be flexible and not have N kinds of system calls that >> > is the direction I would go. That is simple, and you don't need any of >> > the rest. > > I'm not sure I understand what you are suggesting here. Would the > low-level implementation of this be based on procfd, or do you > mean that would be done for pid_t at the kernel level, plus another > syscall for procfd? I was thinking in general and for something that would be enough to back sigqueue, and normal kill semantics. For Christians proposed system call I can think of two ways we could go: long procfd_sigsend(int fd, int sig, int si_code, __u64 sigval); Where sigval is an appropriately written version of union sigval that works on both 32bit and 64bit. Or we could go with: long procfd_sigqueueinfo(int fd, struct siginfo_subset *info) Both would support the same set of system calls today and both would have some similar challenges. Both would take an si_code value of SI_USER as a request that the kernel fill in the struct siginfo as kill would. As otherwise it is invalid for userspace to send a signal with SI_USER. I would not worry about the signal injection case for CRIU where we explicitly allow any siginfo to be sent if we are sending it to ourselves. We already have a system call for that. What I would do is carefully define siginfo_subset architecturally as a proper subset of siginfo on each architecture. That in most if not all cases will work on 32bit and 64bit. Because of the extra padding after union sigval. That siginfo_subset userspace could just consider a siginfo_t. We would just leave out the trouble some bits. So we could do a simple untranslated copy_from_user on it, and then validate that si_code is in a range we support. That definition of siginfo_subset doesn't preclude extensions in the future. And so it could very much be a pass through interface on 32bit on 64bit and into everyone's existing stack frames that embed siginfo. So if we want to fix the situation that is what I would do, as userspace is not allowed to send any of the problematic siginfo definitions today. There might be some finicky detail of union sigval or some architure might have funny alignment between 32bit and 64bit, but I don't think there is. If we want to be more flexible than my procfd_sigsend above I think we should definitely look at that. >> > Alternatively we abandon the mistake of sigqueueinfo and not allow >> > a signal number in the arguments that differs from the si_signo in the >> > siginfo and allow passing the entire thing unchanged from sender to >> > receiver. That is maximum flexibility. > > This would be regardless of which flavor of siginfo (today's arch > specific one, signalfd_siginfo, or a new one) we pass, right? No. signalfd_siginfo offers much less future extensibility because it must be translated. Once you can not just pass the structure through you fail. As you can't be transparently forward compatible. In a similar way Linux filesystems written assuming all the world was ascii wound up forward compatible with utf8 as they are both byte encodings of strings that can use a NULL terminator. >> > signalfd_siginfo just sucks in practice. It is larger that siginfo 104 >> > bytes versus 48. We must deliver it to userspace as a siginfo so it >> > must be translated. Because of the translation signalfd_siginfo adds >> > no flexiblity in practice, because it can not just be passed through. >> > Finallay signalfd_siginfo does not have encodings for all of the >> > siginfo union members, so it fails to be fully general. >> > >> > Personally if I was to define signalfd_siginfo today I would make it: >> > struct siginfo_subset { >> > __u32 sis_signo; >> > __s32 sis_errno; >> > __s32 sis_code; >> > __u32 sis_pad; >> > __u32 sis_pid; >> > __u32 sis_uid; >> > __u64 sis_data (A pointer or integer data field); >> > }; >> > >> > That is just 32bytes, and is all that is needed for everything >> > except for synchronous exceptions. Oh and that happens to be a proper >> > subset of a any sane siginfo layout, on both 32bit and 64bit. > > And that would work for signalfd and waitid_time64, but not for > procfd_signal/kill/sendsiginfo? No. I was bringing this up because it will work for procfd_signal. See above. It supports every signal userspace can send today. We should be able to specify it in a way that is compatible with the existing siginfo definitions on all architectures. For signalfd or let me say a future signalfd2 the practical problem is that we have Linux specific signal extensions that are not synchronous signals like seccomp and sigchld that have fields in a bad location. If we successfully define a siginfo_subset that is the same on 32bit and 64bit I would include the defintion in the kernels definition of siginfo_t so new extensions are automatically picked up and require any extensions to siginfo_t to be added to siginfo_subset_t, in the appropriate 32bit/64bit way. >> > This is one of those rare times where POSIX is sane and what linux >> > has implemented is not. >> >> Thanks for the in-depth explanation. So your point is that we are better >> off if we stick with siginfo_t instead of struct signalfd_siginfo in >> procfd_signal()? My point is that we are better sticking with siginfo_t or just cherry picking the relevant fields and making them system call parameters. The only advantage of using siginfo_t is forwards compatibility. If we give up forward compatibility we can just as easily have our system call prototype include "...(..., int sig, int code, union sigval value)" for the definition of the signal bits. Simple easy and enough for all posix signals. Eric