On Wed, Sep 03, 2025 at 10:47:17AM +0100, Peter Maydell wrote: > On Wed, 3 Sept 2025 at 06:03, chen CJ <cjc...@igel.co.jp> wrote: > > > > I would like to withdraw this patch series. > > > > Sorry for the inconvenience, and thank you for your understanding. > > That's unfortunate; I think it's an issue we really do need to fix, > but I entirely understand if you don't have the time to work > on it further. > > I might pick it up if I have the time to do so.
I worked on this problem a bit more in the past few days while almost everyone will be at the forum. It's almost because I saw similar issues that I have commented before on old versions, but they still existed in the core patch 5. Then I found more issues. Keep commenting on that might be awkward because there will be quite a few dependency changes. One example is, I kept thinking we should not worry about MMIO out-of-bound over mr->size when reaching as deep as access_with_adjusted_size(). There are still quite a few places in patch 5 of this series that does the calculation and it's not obvious what happens if mr->size violated. Peter, if you want to pick it up, please consider reading the replies I left in this series, alone with this version below as comparison reading material. The hope is the reworked patchset below _might_ be easier to read (at least I did add rich comments, because the unaligned changes are tricky and not easy to follow): https://gitlab.com/peterx/qemu/-/commits/mem-unaligned-fix-v0.1?ref_type=tags Especially this patch: https://gitlab.com/peterx/qemu/-/commit/8a8f0f5728a7adc6ecb2cf4358366d2d663a5ed9 However that won't pass the test cases. I still doubt the test case is wrong but I didn't go further modifying the test cases yet (or any better way to test this as you suggested in the other reply). I think that can be the 1st thing we figure out, not the best way to test, but the correctness of the current test case, because IIUC it shouldn't be relevant to impl of unaligned access. To me, if we can reach a consensus on what is the correct (test) behavior on all kinds of unaligned access emulations, fixing the impl should be relatively easy. Thanks, -- Peter Xu
