On Wed, Aug 7, 2024 at 5:57 AM Pedro Falcato <pedro.falcato@xxxxxxxxx> wrote: > > On Wed, Aug 7, 2024 at 2:40 AM Jeff Xu <jeffxu@xxxxxxxxxx> wrote: > > > > On Tue, Aug 6, 2024 at 5:49 PM Pedro Falcato <pedro.falcato@xxxxxxxxx> wrote: > > > > > > On Tue, Aug 6, 2024 at 11:25 PM Jeff Xu <jeffxu@xxxxxxxxxx> wrote: > > > > > > > > On Tue, Aug 6, 2024 at 2:28 PM Pedro Falcato <pedro.falcato@xxxxxxxxx> wrote: > > > > > > > > > > Optimize mseal checks by removing the separate can_modify_mm() step, and > > > > > just doing checks on the individual vmas, when various operations are > > > > > themselves iterating through the tree. This provides a nice speedup. > > > > > > > > > > While I was at it, I found that is_madv_discard() was completely bogus. > > > > > > > > > Thanks for catching this! > > > > Is it possible to separate this fix out from this series and send it > > > > separately and merge first ? > > > > > > Sure. This series is definitely too risky to catch this release, so > > > sending it out as a fix (tomorrow, it's late here) sounds ok. > > > > > Do you mind if I send out a fix ? (I will also include a test case to > > cover that ) > > No need, I'll handle it before the end of the day. > > > > > > > > > > > > Note that my series ignores arch_unmap(), which seems to generally be what we're trending towards[2]. It should > > > > > be applied on top of any powerpc vdso ->close patch to avoid regressions on the PPC architecture. No other > > > > > architecture seems to use arch_unmap. > > > > > > > > > > Note2: This series does not pass all mseal_tests on my end (test_seal_mremap_move_dontunmap_anyaddr fails twice). But the > > > > > top of Linus's tree does not pass these for me either (neither does my Arch Linux 6.10.2 kernel), > > > > > for some reason (mremap regression?). > > > > > > > > > I just sync to Linus's main and I was able to run the test (except two > > > > pkeys related test are skipped because I m on VM) > > > > > > Okay. Fun bug. > > > > > > I was really confused as to why no one could repro this except me :) > > > > > > It looks like recently[1] glibc started consuming the new_address > > > variadic argument when MREMAP_DONTUNMAP. As to the why, > > > MREMAP_DONTUNMAP also seems to take new_address as a hint (this is not > > > documented in the man page, and strace also doesn't know this). > > > However, this trips up some checks that were always fine before > > > (because glibc always passed NULL, and musl still does): > > > > > > if (offset_in_page(new_addr)) > > > if (new_len > TASK_SIZE || new_addr > TASK_SIZE - new_len) > > > if (addr + old_len > new_addr && new_addr + new_len > addr) > > > > > > ^^ These all look at the address without looking at MREMAP_FIXED, and > > > return -EINVAL if they fail. > > > > > > So, test_seal_mremap_move_dontunmap_anyaddr passes 0xdeadbeef For Some > > > Reason (why are you testing mremap in mseal_test.c??), it trips up > > > offset_in_page(new_addr) in mremap_to, and we crash and burn. > > > > > > As to why no one else could repro this: I guess you're not running a > > > glibc new enough ;) > > > > > That makes sense, mystery resolved. > > > > I added sys_ functions for mmap/munmap/mprotect, etc, so that the test > > does not depend on libc, but I didn't do that for mremap, I think the > > fix will be to add sys_mremap as well. > > I disagree, I don't understand why you're doing this test. The test is for testing the return error code should be EPERM in case of sealing. > And even if > you are rightfully doing the test, the test is wrong (and > mremap_dontunmap.c tests agree, and always pass 0 as new_address). the new_address don't need to be 0. E.g. mremap(ptr, size, size, MREMAP_MAYMOVE | MREMAP_DONTUNMAP, 0xdead0000); Will success and relocate memory to 0xdead0000 > The manpage needs to be updated to reflect this, and this test either > needs the 0xdeadbeef removed, or the whole thing. > > Adding a sys_mremap wrapper is inconsequential here, because you'll > need to decide whether to pick up new_address from the flags argument > and, if you do, it'll fail with the same error, but for everyone. > I will send a patch and you can try on your sys (since I don't have the new glibc installed) > -- > Pedro
