mm/hmm: merge hmm_range_snapshot into hmm_range_fault
authorChristoph Hellwig <hch@lst.de>
Fri, 26 Jul 2019 00:56:47 +0000 (17:56 -0700)
committerJason Gunthorpe <jgg@mellanox.com>
Fri, 26 Jul 2019 14:10:53 +0000 (11:10 -0300)
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.com
Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Ralph Campbell <rcampbell@nvidia.com>
Reviewed-by: Jason Gunthorpe <jgg@mellanox.com>
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
Documentation/vm/hmm.rst
include/linux/hmm.h
mm/hmm.c

index 710ce1c..ddcb5ca 100644 (file)
@@ -192,15 +192,14 @@ read only, or fully unmap, etc.). The device must complete the update before
 the driver callback returns.
 
 When the device driver wants to populate a range of virtual addresses, it can
-use either::
+use::
 
-  long hmm_range_snapshot(struct hmm_range *range);
-  long hmm_range_fault(struct hmm_range *range, bool block);
+  long hmm_range_fault(struct hmm_range *range, unsigned int flags);
 
-The first one (hmm_range_snapshot()) will only fetch present CPU page table
+With the HMM_RANGE_SNAPSHOT flag, it will only fetch present CPU page table
 entries and will not trigger a page fault on missing or non-present entries.
-The second one does trigger a page fault on missing or read-only entries if
-write access is requested (see below). Page faults use the generic mm page
+Without that flag, it does trigger a page fault on missing or read-only entries
+if write access is requested (see below). Page faults use the generic mm page
 fault code path just like a CPU page fault.
 
 Both functions copy CPU page table entries into their pfns array argument. Each
@@ -227,20 +226,20 @@ The usage pattern is::
 
       /*
        * Just wait for range to be valid, safe to ignore return value as we
-       * will use the return value of hmm_range_snapshot() below under the
+       * will use the return value of hmm_range_fault() below under the
        * mmap_sem to ascertain the validity of the range.
        */
       hmm_range_wait_until_valid(&range, TIMEOUT_IN_MSEC);
 
  again:
       down_read(&mm->mmap_sem);
-      ret = hmm_range_snapshot(&range);
+      ret = hmm_range_fault(&range, HMM_RANGE_SNAPSHOT);
       if (ret) {
           up_read(&mm->mmap_sem);
           if (ret == -EBUSY) {
             /*
              * No need to check hmm_range_wait_until_valid() return value
-             * on retry we will get proper error with hmm_range_snapshot()
+             * on retry we will get proper error with hmm_range_fault()
              */
             hmm_range_wait_until_valid(&range, TIMEOUT_IN_MSEC);
             goto again;
index 89a1410..90dc594 100644 (file)
@@ -413,7 +413,9 @@ void hmm_range_unregister(struct hmm_range *range);
  */
 #define HMM_FAULT_ALLOW_RETRY          (1 << 0)
 
-long hmm_range_snapshot(struct hmm_range *range);
+/* Don't fault in missing PTEs, just snapshot the current state. */
+#define HMM_FAULT_SNAPSHOT             (1 << 1)
+
 long hmm_range_fault(struct hmm_range *range, unsigned int flags);
 
 long hmm_range_dma_map(struct hmm_range *range,
index 84f2791..1bc014c 100644 (file)
--- a/mm/hmm.c
+++ b/mm/hmm.c
@@ -280,7 +280,6 @@ struct hmm_vma_walk {
        struct hmm_range        *range;
        struct dev_pagemap      *pgmap;
        unsigned long           last;
-       bool                    fault;
        unsigned int            flags;
 };
 
@@ -373,7 +372,7 @@ static inline void hmm_pte_need_fault(const struct hmm_vma_walk *hmm_vma_walk,
 {
        struct hmm_range *range = hmm_vma_walk->range;
 
-       if (!hmm_vma_walk->fault)
+       if (hmm_vma_walk->flags & HMM_FAULT_SNAPSHOT)
                return;
 
        /*
@@ -418,7 +417,7 @@ static void hmm_range_need_fault(const struct hmm_vma_walk *hmm_vma_walk,
 {
        unsigned long i;
 
-       if (!hmm_vma_walk->fault) {
+       if (hmm_vma_walk->flags & HMM_FAULT_SNAPSHOT) {
                *fault = *write_fault = false;
                return;
        }
@@ -936,85 +935,6 @@ void hmm_range_unregister(struct hmm_range *range)
 }
 EXPORT_SYMBOL(hmm_range_unregister);
 
-/*
- * hmm_range_snapshot() - snapshot CPU page table for a range
- * @range: range
- * Return: -EINVAL if invalid argument, -ENOMEM out of memory, -EPERM invalid
- *          permission (for instance asking for write and range is read only),
- *          -EBUSY if you need to retry, -EFAULT invalid (ie either no valid
- *          vma or it is illegal to access that range), number of valid pages
- *          in range->pfns[] (from range start address).
- *
- * This snapshots the CPU page table for a range of virtual addresses. Snapshot
- * validity is tracked by range struct. See in include/linux/hmm.h for example
- * on how to use.
- */
-long hmm_range_snapshot(struct hmm_range *range)
-{
-       const unsigned long device_vma = VM_IO | VM_PFNMAP | VM_MIXEDMAP;
-       unsigned long start = range->start, end;
-       struct hmm_vma_walk hmm_vma_walk;
-       struct hmm *hmm = range->hmm;
-       struct vm_area_struct *vma;
-       struct mm_walk mm_walk;
-
-       lockdep_assert_held(&hmm->mm->mmap_sem);
-       do {
-               /* If range is no longer valid force retry. */
-               if (!range->valid)
-                       return -EBUSY;
-
-               vma = find_vma(hmm->mm, start);
-               if (vma == NULL || (vma->vm_flags & device_vma))
-                       return -EFAULT;
-
-               if (is_vm_hugetlb_page(vma)) {
-                       if (huge_page_shift(hstate_vma(vma)) !=
-                                   range->page_shift &&
-                           range->page_shift != PAGE_SHIFT)
-                               return -EINVAL;
-               } else {
-                       if (range->page_shift != PAGE_SHIFT)
-                               return -EINVAL;
-               }
-
-               if (!(vma->vm_flags & VM_READ)) {
-                       /*
-                        * If vma do not allow read access, then assume that it
-                        * does not allow write access, either. HMM does not
-                        * support architecture that allow write without read.
-                        */
-                       hmm_pfns_clear(range, range->pfns,
-                               range->start, range->end);
-                       return -EPERM;
-               }
-
-               range->vma = vma;
-               hmm_vma_walk.pgmap = NULL;
-               hmm_vma_walk.last = start;
-               hmm_vma_walk.fault = false;
-               hmm_vma_walk.range = range;
-               mm_walk.private = &hmm_vma_walk;
-               end = min(range->end, vma->vm_end);
-
-               mm_walk.vma = vma;
-               mm_walk.mm = vma->vm_mm;
-               mm_walk.pte_entry = NULL;
-               mm_walk.test_walk = NULL;
-               mm_walk.hugetlb_entry = NULL;
-               mm_walk.pud_entry = hmm_vma_walk_pud;
-               mm_walk.pmd_entry = hmm_vma_walk_pmd;
-               mm_walk.pte_hole = hmm_vma_walk_hole;
-               mm_walk.hugetlb_entry = hmm_vma_walk_hugetlb_entry;
-
-               walk_page_range(start, end, &mm_walk);
-               start = end;
-       } while (start < range->end);
-
-       return (hmm_vma_walk.last - range->start) >> PAGE_SHIFT;
-}
-EXPORT_SYMBOL(hmm_range_snapshot);
-
 /**
  * hmm_range_fault - try to fault some address in a virtual address range
  * @range:     range being faulted
@@ -1088,7 +1008,6 @@ long hmm_range_fault(struct hmm_range *range, unsigned int flags)
                range->vma = vma;
                hmm_vma_walk.pgmap = NULL;
                hmm_vma_walk.last = start;
-               hmm_vma_walk.fault = true;
                hmm_vma_walk.flags = flags;
                hmm_vma_walk.range = range;
                mm_walk.private = &hmm_vma_walk;