On Sun, Feb 4, 2024 at 2:27 AM Mike Rapoport <rppt@xxxxxxxxxx> wrote: > > On Tue, Jan 30, 2024 at 06:24:24PM -0800, Lokesh Gidra wrote: > > On Tue, Jan 30, 2024 at 9:28 AM Liam R. Howlett <Liam.Howlett@xxxxxxxxxx> wrote: > > > > > > * Mike Rapoport <rppt@xxxxxxxxxx> [240130 03:55]: > > > > On Mon, Jan 29, 2024 at 10:46:27PM -0500, Liam R. Howlett wrote: > > > > > * Lokesh Gidra <lokeshgidra@xxxxxxxxxx> [240129 17:35]: > > > > > > > > > > > > > diff --git a/fs/userfaultfd.c b/fs/userfaultfd.c > > > > > > > > index 58331b83d648..c00a021bcce4 100644 > > > > > > > > --- a/fs/userfaultfd.c > > > > > > > > +++ b/fs/userfaultfd.c > > > > > > > > @@ -685,12 +685,15 @@ int dup_userfaultfd(struct vm_area_struct *vma, struct list_head *fcs) > > > > > > > > ctx->flags = octx->flags; > > > > > > > > ctx->features = octx->features; > > > > > > > > ctx->released = false; > > > > > > > > + init_rwsem(&ctx->map_changing_lock); > > > > > > > > atomic_set(&ctx->mmap_changing, 0); > > > > > > > > ctx->mm = vma->vm_mm; > > > > > > > > mmgrab(ctx->mm); > > > > > > > > > > > > > > > > userfaultfd_ctx_get(octx); > > > > > > > > + down_write(&octx->map_changing_lock); > > > > > > > > atomic_inc(&octx->mmap_changing); > > > > > > > > + up_write(&octx->map_changing_lock); > > > > > > > > > > On init, I don't think taking the lock is strictly necessary - unless > > > > > there is a way to access it before this increment? Not that it would > > > > > cost much. > > > > > > > > It's fork, the lock is for the context of the parent process and there > > > > could be uffdio ops running in parallel on its VM. > > > > > > Is this necessary then? We are getting the octx from another mm but the > > > mm is locked for forking. Why does it matter if there are readers of > > > the octx? > > > > > > I assume, currently, there is no way the userfaultfd ctx can > > > be altered under mmap_lock held for writing. I would think it matters if > > > there are writers (which, I presume are blocked by the mmap_lock for > > > now?) Shouldn't we hold the write lock for the entire dup process, I > > > mean, if we remove the userfaultfd from the mmap_lock, we cannot let the > > > structure being duplicated change half way through the dup process? > > > > > > I must be missing something with where this is headed? > > > > > AFAIU, the purpose of mmap_changing is to serialize uffdio operations > > with non-cooperative events if and when such events are being > > monitored by userspace (in case you missed, in all the cases of writes > > to mmap_changing, we only do it if that non-cooperative event has been > > requested by the user). As you pointed out there are no correctness > > concerns as far as userfaultfd operations are concerned. But these > > events are essential for the uffd monitor's functioning. > > > > For example: say the uffd monitor wants to be notified for REMAP > > operations while doing uffdio_copy operations. When COPY ioctls start > > failing with -EAGAIN and uffdio_copy.copy == 0, then it knows it must > > be due to mremap(), in which case it waits for the REMAP event > > notification before attempting COPY again. > > > > But there are few things that I didn't get after going through the > > history of non-cooperative events. Hopefully Mike (or someone else > > familiar) can clarify: > > > > IIUC, the idea behind non-cooperative events was to block uffdio > > operations from happening *before* the page tables are manipulated by > > the event (like mremap), and that the uffdio ops are resumed after the > > event notification is received by the monitor. > > The idea was to give userspace some way to serialize processing of > non-cooperative event notifications and uffdio operations running in > parallel. It's not necessary to block uffdio operations from happening > before changes to the memory map, but with the mmap_lock synchronization > that already was there adding mmap_chaning that will prevent uffdio > operations when mmap_lock is taken for write was the simplest thing to do. > > When CRIU does post-copy restore of a process, its uffd monitor reacts to > page fault and non-cooperative notifications and also performs a background > copy of the memory contents from the saved state to the address space of > the process being restored. > > Since non-cooperative events may happen completely independent from the > uffd monitor, there are cases when the uffd monitor couldn't identify the > order of events, like e.g. what won the race on mmap_lock, the process > thread doing fork or the uffd monitor's uffdio_copy. > > In the fork vs uffdio_copy example, without mmap_changing, if the > uffdio_copy takes the mmap_lock first, the new page will be present in the > parent by the time copy_page_range() is called and the page will appear in > the child's memory mappings by the time uffd monitor gets notification > about the fork event. However, if the fork() is the first to take the > mmap_lock, the new page will appear in the parent address space after > copy_page_range() and it won't be mapped in the child's address space. > > With mmap_changing and current locking with mmap_lock, we have a guarantee > that uffdio_copy will bail out if fork already took mmap_lock and the > monitor can act appropriately. > Thanks for the explanation. Really helpful! > > 1) Why in the case of REMAP prep() is done after page-tables are > > moved? Shouldn't it be done before? All other non-cooperative > > operations do the prep() before. > > mremap_userfaultfd_prep() is done after page tables are moved because it > initializes uffd context on the new_vma and if the actual remap fails, > there's no point of doing it. > Since mrpemap holds mmap_lock for write it does not matter if mmap_changed > is updated before or after page tables are moved. In the time between > mmap_lock is released and the UFFD_EVENT_REMAP is delivered to the uffd > monitor, mmap_chaging will remain >0 and uffdio operations will bail out. > Yes this makes sense. Even with per-vma locks, I see that the new_vma is write-locked (vma_start_write()) in vma_link() guaranteeing the same. > > 2) UFFD_FEATURE_EVENT_REMOVE only notifies user space. It is not > > consistently blocking uffdio operations (as both sides are acquiring > > mmap_lock in read-mode) when remove operation is taking place. I can > > understand this was intentionally left as is in the interest of not > > acquiring mmap_lock in write-mode during madvise. But is only getting > > the notification any useful? Can we say this patch fixes it? And in > > that case shouldn't I split userfaultfd_remove() into two functions > > (like other non-cooperative operations)? > > The notifications are useful because uffd monitor knows what memory should > not be filled with uffdio_copy. Indeed there was no interest in taking > mmap_lock for write in madvise, so there could be race between madvise and > uffdio operations. This race essentially prevents uffd monitor from running > the background copy in a separate thread, and with your change this should > be possible. > Makes sense. Thanks! > > 3) Based on [1] I see how mmap_changing helps in eliminating duplicate > > work (background copy) by uffd monitor, but didn't get if there is a > > correctness aspect too that I'm missing? I concur with Amit's point in > > [1] that getting -EEXIST when setting up the pte will avoid memory > > corruption, no? > > In the fork case without mmap_changing the child process may be get data or > zeroes depending on the race for mmap_lock between the fork and > uffdio_copy and -EEXIST is not enough for monitor to detect what was the > ordering between fork and uffdio_copy. This is extremely helpful. IIUC, there is a window after mmap_lock (write-mode) is released and before the uffd monitor thread is notified of fork. In that window, the monitor doesn't know that fork has already happened. So, without mmap_changing it would have done background copy only in the parent, thereby causing data inconsistency between parent and child processes. It seems to me that the correctness argument for mmap_changing is there in case of FORK event and REMAP when mremap is called with MREMAP_DONTUNMAP. In all other cases its only benefit is by avoiding unnecessary background copies, right? > > > > > > > > > @@ -783,7 +788,9 @@ bool userfaultfd_remove(struct vm_area_struct *vma, > > > > > > > > return true; > > > > > > > > > > > > > > > > userfaultfd_ctx_get(ctx); > > > > > > > > + down_write(&ctx->map_changing_lock); > > > > > > > > atomic_inc(&ctx->mmap_changing); > > > > > > > > + up_write(&ctx->map_changing_lock); > > > > > > > > mmap_read_unlock(mm); > > > > > > > > > > > > > > > > msg_init(&ewq.msg); > > > > > > > > > > If this happens in read mode, then why are you waiting for the readers > > > > > to leave? Can't you just increment the atomic? It's fine happening in > > > > > read mode today, so it should be fine with this new rwsem. > > > > > > > > It's been a while and the details are blurred now, but if I remember > > > > correctly, having this in read mode forced non-cooperative uffd monitor to > > > > be single threaded. If a monitor runs, say uffdio_copy, and in parallel a > > > > thread in the monitored process does MADV_DONTNEED, the latter will wait > > > > for userfaultfd_remove notification to be processed in the monitor and drop > > > > the VMA contents only afterwards. If a non-cooperative monitor would > > > > process notification in parallel with uffdio ops, MADV_DONTNEED could > > > > continue and race with uffdio_copy, so read mode wouldn't be enough. > > > > > > > > > > Right now this function won't stop to wait for readers to exit the > > > critical section, but with this change there will be a pause (since the > > > down_write() will need to wait for the readers with the read lock). So > > > this is adding a delay in this call path that isn't necessary (?) nor > > > existed before. If you have non-cooperative uffd monitors, then you > > > will have to wait for them to finish to mark the uffd as being removed, > > > where as before it was a fire & forget, this is now a wait to tell. > > > > > I think a lot will be clearer once we get a response to my questions > > above. IMHO not only this write-lock is needed here, we need to fix > > userfaultfd_remove() by splitting it into userfaultfd_remove_prep() > > and userfaultfd_remove_complete() (like all other non-cooperative > > operations) as well. This patch enables us to do that as we remove > > mmap_changing's dependency on mmap_lock for synchronization. > > The write-lock is not a requirement here for correctness and I don't see > why we would need userfaultfd_remove_prep(). > > As I've said earlier, having a write-lock here will let CRIU to run > background copy in parallel with processing of uffd events, but I don't > feel strongly about doing it. > Got it. Anyways, such a change needn't be part of this patch, so I'm going to keep it unchanged. > > > > There was no much sense to make MADV_DONTNEED take mmap_lock in write mode > > > > just for this, but now taking the rwsem in write mode here sounds > > > > reasonable. > > > > > > > > > > I see why there was no need for a mmap_lock in write mode, but I think > > > taking the new rwsem in write mode is unnecessary. > > > > > > Basically, I see this as a signal to new readers to abort, but we don't > > > need to wait for current readers to finish before this one increments > > > the atomic. > > > > > > Unless I missed something, I don't think you want to take the write lock > > > here. > > What I understood from the history of mmap_changing is that the > > intention was to enable informing the uffd monitor about the correct > > state of which pages are filled and which aren't. Going through this > > thread was very helpful [2] > > > > [2] https://lore.kernel.org/lkml/1527061324-19949-1-git-send-email-rppt@xxxxxxxxxxxxxxxxxx/ > > -- > Sincerely yours, > Mike.
