- 20 Aug, 2019 13 commits
-
-
Christoph Hellwig authored
Now that we can rely errors in the normal control flow there is no need for this flag, remove it. Link: https://lore.kernel.org/r/20190814075928.23766-9-hch@lst.deSigned-off-by:
Christoph Hellwig <hch@lst.de> Reviewed-by:
Ralph Campbell <rcampbell@nvidia.com> Tested-by:
Jason Gunthorpe <jgg@mellanox.com>
-
Christoph Hellwig authored
Factor the main copy page to vram routine out into a helper that acts on a single page and which doesn't require the nouveau_dmem_migrate structure for argument passing. As an added benefit the new version only allocates the dma address array once and reuses it for each subsequent chunk of work. Link: https://lore.kernel.org/r/20190814075928.23766-8-hch@lst.deSigned-off-by:
-
Christoph Hellwig authored
Factor the main copy page to ram routine out into a helper that acts on a single page and which doesn't require the nouveau_dmem_fault structure for argument passing. Also remove the loop over multiple pages as we only handle one at the moment, although the structure of the main worker function makes it relatively easy to add multi page support back if needed in the future. But at least for now this avoid the needed to dynamically allocate memory for the dma addresses in what is essentially the page fault path. Link: https://lore.kernel.org/r/20190814075928.23766-7-hch@lst.deSigned-off-by:
-
Christoph Hellwig authored
nouveau_dmem_migrate_vma and nouveau_dmem_convert_pfn are only called when CONFIG_DRM_NOUVEAU_SVM is enabled, so there is no need to provide !CONFIG_DRM_NOUVEAU_SVM stubs for them. Link: https://lore.kernel.org/r/20190814075928.23766-6-hch@lst.deSigned-off-by:
-
Christoph Hellwig authored
Factor out the end of fencing logic from the two migration routines. Link: https://lore.kernel.org/r/20190814075928.23766-5-hch@lst.deSigned-off-by:
-
Christoph Hellwig authored
Factor out the repeated device memory address calculation into a helper. Link: https://lore.kernel.org/r/20190814075928.23766-4-hch@lst.deSigned-off-by:
-
Christoph Hellwig authored
When we start a new batch of dma_map operations we need to reset dma_nr, as we start filling a newly allocated array. Link: https://lore.kernel.org/r/20190814075928.23766-3-hch@lst.deSigned-off-by:
-
Christoph Hellwig authored
There isn't any good reason to pass callbacks to migrate_vma. Instead we can just export the three steps done by this function to drivers and let them sequence the operation without callbacks. This removes a lot of boilerplate code as-is, and will allow the drivers to drastically improve code flow and error handling further on. Link: https://lore.kernel.org/r/20190814075928.23766-2-hch@lst.deSigned-off-by:
-
Jason Gunthorpe authored
Jason Gunthorpe says: ==================== Add mmu_notifier_get/put for managing mmu notifier registrations This series introduces a new registration flow for mmu_notifiers based on the idea that the user would like to get a single refcounted piece of memory for a mm, keyed to its use. For instance many users of mmu_notifiers use an interval tree or similar to dispatch notifications to some object. There are many objects but only one notifier subscription per mm holding the tree. Of the 12 places that call mmu_notifier_register: - 7 are maintaining some kind of obvious mapping of mm_struct to mmu_notifier registration, ie in some linked list or hash table. Of the 7 this series converts 4 (gru, hmm, RDMA, radeon) - 3 (hfi1, gntdev, vhost) are registering multiple notifiers, but each one immediately does some VA range filtering, ie with an interval tree. These would be better with a global subsystem-wide range filter and could convert to this API. - 2 (kvm, amd_iommu) are deliberately using a single mm at a time, and really can't use this API. One of the intel-svm's modes is also in this list The 3/7 unconverted drivers are: - intel-svm This driver tracks mm's in a global linked list 'global_svm_list' and would benefit from this API. Its flow is a bit complex, since it also wants a set of non-shared notifiers. - i915_gem_usrptr This driver tracks mm's in a per-device hash table (dev_priv->mm_structs), but only has an optional use of mmu_notifiers. Since it still seems to need the hash table it is difficult to convert. - amdkfd/kfd_process This driver is using a global SRCU hash table to track mm's The control flow here is very complicated and the driver is relying on this hash table to be fast on the ioctl syscall path. It would definitely benefit, but only if the ioctl path didn't need to do the search so often. ==================== Signed-off-by: -
Jason Gunthorpe authored
The sequence of mmu_notifier_unregister_no_release(), mmu_notifier_call_srcu() is identical to mmu_notifier_put() with the free_notifier callback. As this is the last user of those APIs, converting it means we can drop them. Link: https://lore.kernel.org/r/20190806231548.25242-11-jgg@ziepe.caReviewed-by:
Felix Kuehling <Felix.Kuehling@amd.com> Signed-off-by:
-
Jason Gunthorpe authored
When using mmu_notifer_unregister_no_release() the caller must ensure there is a SRCU synchronize before the mn memory is freed, otherwise use after free races are possible, for instance: CPU0 CPU1 invalidate_range_start hlist_for_each_entry_rcu(..) mmu_notifier_unregister_no_release(&p->mn) kfree(mn) if (mn->ops->invalidate_range_end) The error unwind in amdkfd misses the SRCU synchronization. amdkfd keeps the kfd_process around until the mm is released, so split the flow to fully initialize the kfd_process and register it for find_process, and with the notifier. Past this point the kfd_process does not need to be cleaned up as it is fully ready. The final failable step does a vm_mmap() and does not seem to impact the kfd_process global state. Since it also cannot be undone (and already has problems with undo if it internally fails), it has to be last. This way we don't have to try to unwind the mmu_notifier_register() and avoid the problem with the SRCU. Along the way this also fixes various other error unwind bugs in the flow. Fixes: 45102048 ("amdkfd: Add process queue manager module") Link: https://lore.kernel.org/r/20190806231548.25242-10-jgg@ziepe.caReviewed-by: -
Jason Gunthorpe authored
radeon is using a device global hash table to track what mmu_notifiers have been registered on struct mm. This is better served with the new get/put scheme instead. radeon has a bug where it was not blocking notifier release() until all the BO's had been invalidated. This could result in a use after free of pages the BOs. This is tied into a second bug where radeon left the notifiers running endlessly even once the interval tree became empty. This could result in a use after free with module unload. Both are fixed by changing the lifetime model, the BOs exist in the interval tree with their natural lifetimes independent of the mm_struct lifetime using the get/put scheme. The release runs synchronously and just does invalidate_start across the entire interval tree to create the required DMA fence. Additions to the interval tree after release are already impossible as only current->mm is used during the add. Link: https://lore.kernel.org/r/20190806231548.25242-9-jgg@ziepe.caAcked-by:
Christian König <christian.koenig@amd.com> Signed-off-by:
-
Jason Gunthorpe authored
This is a significant simplification, it eliminates all the remaining 'hmm' stuff in mm_struct, eliminates krefing along the critical notifier paths, and takes away all the ugly locking and abuse of page_table_lock. mmu_notifier_get() provides the single struct hmm per struct mm which eliminates mm->hmm. It also directly guarantees that no mmu_notifier op callback is callable while concurrent free is possible, this eliminates all the krefs inside the mmu_notifier callbacks. The remaining krefs in the range code were overly cautious, drivers are already not permitted to free the mirror while a range exists. Link: https://lore.kernel.org/r/20190806231548.25242-6-jgg@ziepe.caReviewed-by:
-
- 16 Aug, 2019 4 commits
-
-
Jason Gunthorpe authored
GRU is already using almost the same algorithm as get/put, it even helpfully has a 10 year old comment to make this algorithm common, which is finally happening. There are a few differences and fixes from this conversion: - GRU used rcu not srcu to read the hlist - Unclear how the locking worked to prevent gru_register_mmu_notifier() from running concurrently with gru_drop_mmu_notifier() - this version is safe - GRU had a release function which only set a variable without any locking that skiped the synchronize_srcu during unregister which looks racey, but this makes it reliable via the integrated call_srcu(). - It is unclear if the mmap_sem is actually held when __mmu_notifier_register() was called, lockdep will now warn if this is wrong Link: https://lore.kernel.org/r/20190806231548.25242-5-jgg@ziepe.caReviewed-by:
Dimitri Sivanich <sivanich@hpe.com> Signed-off-by:
-
Jason Gunthorpe authored
Many places in the kernel have a flow where userspace will create some object and that object will need to connect to the subsystem's mmu_notifier subscription for the duration of its lifetime. In this case the subsystem is usually tracking multiple mm_structs and it is difficult to keep track of what struct mmu_notifier's have been allocated for what mm's. Since this has been open coded in a variety of exciting ways, provide core functionality to do this safely. This approach uses the struct mmu_notifier_ops * as a key to determine if the subsystem has a notifier registered on the mm or not. If there is a registration then the existing notifier struct is returned, otherwise the ops->alloc_notifiers() is used to create a new per-subsystem notifier for the mm. The destroy side incorporates an async call_srcu based destruction which will avoid bugs in the callers such as commit 6d7c3cde ("mm/hmm: fix use after free with struct hmm in the mmu notifiers"). Since we are inside the mmu notifier core locking is fairly simple, the allocation uses the same approach as for mmu_notifier_mm, the write side of the mmap_sem makes everything deterministic and we only need to do hlist_add_head_rcu() under the mm_take_all_locks(). The new users count and the discoverability in the hlist is fully serialized by the mmu_notifier_mm->lock. Link: https://lore.kernel.org/r/20190806231548.25242-4-jgg@ziepe.caCo-developed-by:
Christoph Hellwig <hch@infradead.org> Signed-off-by:
-
Jason Gunthorpe authored
A prior commit e0f3c3f7 ("mm/mmu_notifier: init notifier if necessary") made an attempt at doing this, but had to be reverted as calling the GFP_KERNEL allocator under the i_mmap_mutex causes deadlock, see commit 35cfa2b0 ("mm/mmu_notifier: allocate mmu_notifier in advance"). However, we can avoid that problem by doing the allocation only under the mmap_sem, which is already happening. Since all writers to mm->mmu_notifier_mm hold the write side of the mmap_sem reading it under that sem is deterministic and we can use that to decide if the allocation path is required, without speculation. The actual update to mmu_notifier_mm must still be done under the mm_take_all_locks() to ensure read-side coherency. Link: https://lore.kernel.org/r/20190806231548.25242-3-jgg@ziepe.caReviewed-by:
-
Jason Gunthorpe authored
This simplifies the code to not have so many one line functions and extra logic. __mmu_notifier_register() simply becomes the entry point to register the notifier, and the other one calls it under lock. Also add a lockdep_assert to check that the callers are holding the lock as expected. Link: https://lore.kernel.org/r/20190806231548.25242-2-jgg@ziepe.caSuggested-by:
-
- 07 Aug, 2019 13 commits
-
-
Christoph Hellwig authored
Make HMM_MIRROR an option that is selected by drivers wanting to use it instead of a user visible option as it is just a low-level implementation detail. Link: https://lore.kernel.org/r/20190806160554.14046-15-hch@lst.deSigned-off-by:
-
Christoph Hellwig authored
There isn't really any architecture specific code in this page table walk implementation, so drop the dependencies. Link: https://lore.kernel.org/r/20190806160554.14046-14-hch@lst.deSigned-off-by:
-
Christoph Hellwig authored
Stub out the whole function and assign NULL to the .hugetlb_entry method if CONFIG_HUGETLB_PAGE is not set, as the method won't ever be called in that case. Link: https://lore.kernel.org/r/20190806160554.14046-13-hch@lst.deSigned-off-by:
-
Christoph Hellwig authored
Stub out the whole function when CONFIG_TRANSPARENT_HUGEPAGE is not set to make the function easier to read. Link: https://lore.kernel.org/r/20190806160554.14046-12-hch@lst.deSigned-off-by:
-
Christoph Hellwig authored
We only need the special pud_entry walker if PUD-sized hugepages and pte mappings are supported, else the common pagewalk code will take care of the iteration. Not implementing this callback reduced the amount of code compiled for non-x86 platforms, and also fixes compile failures with other architectures when helpers like pud_pfn are not implemented. Link: https://lore.kernel.org/r/20190806160554.14046-11-hch@lst.deSigned-off-by:
-
Christoph Hellwig authored
pte_index is an internal arch helper in various architectures, without consistent semantics. Open code that calculation of a PMD index based on the virtual address instead. Link: https://lore.kernel.org/r/20190806160554.14046-10-hch@lst.deSigned-off-by:
-
Christoph Hellwig authored
The pagewalk code already passes the value as the hmask parameter. Link: https://lore.kernel.org/r/20190806160554.14046-9-hch@lst.deSigned-off-by:
-
Christoph Hellwig authored
All users pass PAGE_SIZE here, and if we wanted to support single entries for huge pages we should really just add a HMM_FAULT_HUGEPAGE flag instead that uses the huge page size instead of having the caller calculate that size once, just for the hmm code to verify it. Link: https://lore.kernel.org/r/20190806160554.14046-8-hch@lst.deSigned-off-by:
-
Christoph Hellwig authored
The start, end and page_shift values are all saved in the range structure, so we might as well use that for argument passing. Link: https://lore.kernel.org/r/20190806160554.14046-7-hch@lst.deSigned-off-by:
-
Christoph Hellwig authored
Link: https://lore.kernel.org/r/20190806160554.14046-6-hch@lst.deSigned-off-by:
-
Christoph Hellwig authored
We'll need the nouveau_svmm structure to improve the function soon. For now this allows using the svmm->mm reference to unlock the mmap_sem, and thus the same dereference chain that the caller uses to lock and unlock it. Link: https://lore.kernel.org/r/20190806160554.14046-4-hch@lst.deSigned-off-by:
-
Christoph Hellwig authored
The list is used to add the range to another list as an entry in the core hmm code, and intended as a private member not exposed to drivers. There is no need to initialize it in a driver. Link: https://lore.kernel.org/r/20190806160554.14046-3-hch@lst.deSigned-off-by:
-
Christoph Hellwig authored
hmm_range_fault can only return -EAGAIN if called with the HMM_FAULT_ALLOW_RETRY flag, which amdgpu never does. Remove the handling for the -EAGAIN case with its non-standard locking scheme. Link: https://lore.kernel.org/r/20190806160554.14046-2-hch@lst.deSigned-off-by:
-
- 26 Jul, 2019 6 commits
-
-
Ralph Campbell authored
Since hmm_range_fault() doesn't use the struct hmm_range vma field, remove it. Link: https://lore.kernel.org/r/20190726005650.2566-8-rcampbell@nvidia.comSuggested-by:
-
Ralph Campbell authored
walk_page_range() will only call hmm_vma_walk_hugetlb_entry() for hugetlbfs pages and doesn't call hmm_vma_walk_pmd() in this case. Therefore, it is safe to remove the check for vma->vm_flags & VM_HUGETLB in hmm_vma_walk_pmd(). Link: https://lore.kernel.org/r/20190726005650.2566-7-rcampbell@nvidia.comSigned-off-by:
-
Christoph Hellwig authored
Add a HMM_FAULT_SNAPSHOT flag so that hmm_range_snapshot can be merged into the almost identical hmm_range_fault function. Link: https://lore.kernel.org/r/20190726005650.2566-5-rcampbell@nvidia.comSigned-off-by:
-
Christoph Hellwig authored
This allows easier expansion to other flags, and also makes the callers a little easier to read. Link: https://lore.kernel.org/r/20190726005650.2566-4-rcampbell@nvidia.comSigned-off-by:
-
Ralph Campbell authored
A few more comments and minor programming style clean ups. There should be no functional changes. Link: https://lore.kernel.org/r/20190726005650.2566-3-rcampbell@nvidia.comSigned-off-by:
-
Ralph Campbell authored
The hmm_mirror_ops callback function sync_cpu_device_pagetables() passes a struct hmm_update which is a simplified version of struct mmu_notifier_range. This is unnecessary so replace hmm_update with mmu_notifier_range directly. Link: https://lore.kernel.org/r/20190726005650.2566-2-rcampbell@nvidia.comSigned-off-by:
-
- 25 Jul, 2019 4 commits
-
-
Jason Gunthorpe authored
The magic dropping of mmap_sem when handle_mm_fault returns VM_FAULT_RETRY is rather subtile. Add a comment explaining it. Link: https://lore.kernel.org/r/20190724065258.16603-8-hch@lst.deTested-by:
-
Christoph Hellwig authored
Switch the one remaining user in nouveau over to its replacement, and remove all the wrappers. Link: https://lore.kernel.org/r/20190724065258.16603-7-hch@lst.deTested-by:
-
Christoph Hellwig authored
-EAGAIN has a magic meaning for non-blocking faults, so don't overload it. Given that the caller doesn't check for specific error codes this change is purely cosmetic. Link: https://lore.kernel.org/r/20190724065258.16603-6-hch@lst.deTested-by:
-
Christoph Hellwig authored
Currently nouveau_svm_fault expects nouveau_range_fault to never unlock mmap_sem, but the latter unlocks it for a random selection of error codes. Fix this up by always unlocking mmap_sem for non-zero return values in nouveau_range_fault, and only unlocking it in the caller for successful returns. Link: https://lore.kernel.org/r/20190724065258.16603-5-hch@lst.deTested-by:
-
