Re: [RFC 2/4] lib/strncpy_from_user: Remove redundant user space pointer range check
On 1/15/20 6:42 AM, Andrey Konovalov wrote: >> - max_addr = user_addr_max(); >> - src_addr = (unsigned long)untagged_addr(src); > > If you end up changing this code, you need to keep the untagged_addr() > logic, otherwise this breaks arm64 tagged address ABI [1]. It is moot point now, but fwiw untagged_addr() would not have been needed anymore as it was only needed to compute the pointer difference which my patch got rid of. > > [1] https://www.kernel.org/doc/html/latest/arm64/tagged-address-abi.html > >> - if (likely(src_addr < max_addr)) { >> - unsigned long max = max_addr - src_addr; >> + kasan_check_write(dst, count); >> + check_object_size(dst, count, false); >> + if (user_access_begin(src, count)) { ___ linux-snps-arc mailing list linux-snps-arc@lists.infradead.org http://lists.infradead.org/mailman/listinfo/linux-snps-arc
Re: [RFC 2/4] lib/strncpy_from_user: Remove redundant user space pointer range check
On Tue, Jan 14, 2020 at 01:22:07PM -0800, Linus Torvalds wrote: > The fact is, copying a string from user space is *very* different from > copying a fixed number of bytes, and that whole dance with > > max_addr = user_addr_max(); > > is absolutely required and necessary. > > You completely broke string copying. BTW, a quick grep through the callers has found something odd - static ssize_t kmemleak_write(struct file *file, const char __user *user_buf, size_t size, loff_t *ppos) { char buf[64]; int buf_size; int ret; buf_size = min(size, (sizeof(buf) - 1)); if (strncpy_from_user(buf, user_buf, buf_size) < 0) return -EFAULT; buf[buf_size] = 0; What the hell? If somebody is calling write(fd, buf, n) they'd better be ready to see any byte from buf[0] up to buf[n - 1] fetched, and if something is unmapped - deal with -EFAULT. Is something really doing that and if so, why does kmemleak try to accomodate that idiocy? The same goes for several more ->write() instances - mtrr_write(), armada_debugfs_crtc_reg_write() and cio_ignore_write(); IMO that's seriously misguided (and cio one ought use vmemdup_user() instead of what it's doing)... ___ linux-snps-arc mailing list linux-snps-arc@lists.infradead.org http://lists.infradead.org/mailman/listinfo/linux-snps-arc
Re: [RFC 2/4] lib/strncpy_from_user: Remove redundant user space pointer range check
On 1/14/20 1:22 PM, Linus Torvalds wrote: > On Tue, Jan 14, 2020 at 12:09 PM Vineet Gupta > wrote: >> >> This came up when switching ARC to word-at-a-time interface and using >> generic/optimized strncpy_from_user >> >> It seems the existing code checks for user buffer/string range multiple >> times and one of tem cn be avoided. > > NO! > > DO NOT DO THIS. > > This is seriously buggy. > >> long strncpy_from_user(char *dst, const char __user *src, long count) >> { >> - unsigned long max_addr, src_addr; >> - >> if (unlikely(count <= 0)) >> return 0; >> >> - max_addr = user_addr_max(); >> - src_addr = (unsigned long)untagged_addr(src); >> - if (likely(src_addr < max_addr)) { >> - unsigned long max = max_addr - src_addr; >> + kasan_check_write(dst, count); >> + check_object_size(dst, count, false); >> + if (user_access_begin(src, count)) { > > You can't do that "user_access_begin(src, count)", because "count" is > the maximum _possible_ length, but it is *NOT* necessarily the actual > length of the string we really get from user space! > > Think of this situation: > > - user has a 5-byte string at the end of the address space > > - kernel does a > > n = strncpy_from_user(uaddr, page, PAGE_SIZE) > > now your "user_access_begin(src, count)" will _fail_, because "uaddr" > is close to the end of the user address space, and there's not room > for PAGE_SIZE bytes any more. Oops indeed that was the case I didn't comprehend. In my initial tests with debugger, every single hit on strncpy_from_user() had user addresses well into the address space such that @max was ridiculously large (0x_ - ptr) compared to @count. > But "count" isn't actually how many bytes we will access from user > space, it's only the maximum limit on the *target*. IOW, it's about a > kernel buffer size, not about the user access size. Right I understood all that, but missed the case when user buffer is towards end of address space and access_ok() will erroneously flag it. > Because we'll only access that 5-byte string, which fits just fine in > the user space, and doing that "user_access_begin(src, count)" gives > the wrong answer. > > The fact is, copying a string from user space is *very* different from > copying a fixed number of bytes, and that whole dance with > > max_addr = user_addr_max(); > > is absolutely required and necessary. > > You completely broke string copying. I'm sorry and I wasn't sure to begin with hence the disclaimer in 0/4 > It is very possible that string copying was horribly broken on ARC > before too - almost nobody ever gets this right, but the generic > routine does. No it is not. It is just dog slow since it does byte copy and uses the Zero delay loops which I'm trying to get rid of. That's when I recalled the word-at-a-time API which I'd meaning to go back to for last 7 years :-) -Vineet ___ linux-snps-arc mailing list linux-snps-arc@lists.infradead.org http://lists.infradead.org/mailman/listinfo/linux-snps-arc
Re: [RFC 2/4] lib/strncpy_from_user: Remove redundant user space pointer range check
On Tue, Jan 14, 2020 at 12:09 PM Vineet Gupta wrote: > > This came up when switching ARC to word-at-a-time interface and using > generic/optimized strncpy_from_user > > It seems the existing code checks for user buffer/string range multiple > times and one of tem cn be avoided. NO! DO NOT DO THIS. This is seriously buggy. > long strncpy_from_user(char *dst, const char __user *src, long count) > { > - unsigned long max_addr, src_addr; > - > if (unlikely(count <= 0)) > return 0; > > - max_addr = user_addr_max(); > - src_addr = (unsigned long)untagged_addr(src); > - if (likely(src_addr < max_addr)) { > - unsigned long max = max_addr - src_addr; > + kasan_check_write(dst, count); > + check_object_size(dst, count, false); > + if (user_access_begin(src, count)) { You can't do that "user_access_begin(src, count)", because "count" is the maximum _possible_ length, but it is *NOT* necessarily the actual length of the string we really get from user space! Think of this situation: - user has a 5-byte string at the end of the address space - kernel does a n = strncpy_from_user(uaddr, page, PAGE_SIZE) now your "user_access_begin(src, count)" will _fail_, because "uaddr" is close to the end of the user address space, and there's not room for PAGE_SIZE bytes any more. But "count" isn't actually how many bytes we will access from user space, it's only the maximum limit on the *target*. IOW, it's about a kernel buffer size, not about the user access size. Because we'll only access that 5-byte string, which fits just fine in the user space, and doing that "user_access_begin(src, count)" gives the wrong answer. The fact is, copying a string from user space is *very* different from copying a fixed number of bytes, and that whole dance with max_addr = user_addr_max(); is absolutely required and necessary. You completely broke string copying. It is very possible that string copying was horribly broken on ARC before too - almost nobody ever gets this right, but the generic routine does. So the generic routine is not only faster, it is *correct*, and your change broke it. Don't touch generic code. If you want to use the generic code, please do so. But DO NOT TOUCH IT. It is correct, your patch is wrong. The exact same issue is true in strnlen_user(). Don't break it. Linus ___ linux-snps-arc mailing list linux-snps-arc@lists.infradead.org http://lists.infradead.org/mailman/listinfo/linux-snps-arc