dma debug: account for cachelines and read-only mappings in overlap tracking
authorDan Williams <dan.j.williams@intel.com>
Mon, 3 Mar 2014 23:38:21 +0000 (15:38 -0800)
committerLinus Torvalds <torvalds@linux-foundation.org>
Tue, 4 Mar 2014 15:55:47 +0000 (07:55 -0800)
While debug_dma_assert_idle() checks if a given *page* is actively
undergoing dma the valid granularity of a dma mapping is a *cacheline*.
Sander's testing shows that the warning message "DMA-API: exceeded 7
overlapping mappings of pfn..." is falsely triggering.  The test is
simply mapping multiple cachelines in a given page.

Ultimately we want overlap tracking to be valid as it is a real api
violation, so we need to track active mappings by cachelines.  Update
the active dma tracking to use the page-frame-relative cacheline of the
mapping as the key, and update debug_dma_assert_idle() to check for all
possible mapped cachelines for a given page.

However, the need to track active mappings is only relevant when the
dma-mapping is writable by the device.  In fact it is fairly standard
for read-only mappings to have hundreds or thousands of overlapping
mappings at once.  Limiting the overlap tracking to writable
(!DMA_TO_DEVICE) eliminates this class of false-positive overlap
reports.

Note, the radix gang lookup is sub-optimal.  It would be best if it
stopped fetching entries once the search passed a page boundary.
Nevertheless, this implementation does not perturb the original net_dma
failing case.  That is to say the extra overhead does not show up in
terms of making the failing case pass due to a timing change.

References:
  http://marc.info/?l=linux-netdev&m=139232263419315&w=2
  http://marc.info/?l=linux-netdev&m=139217088107122&w=2

Signed-off-by: Dan Williams <dan.j.williams@intel.com>
Reported-by: Sander Eikelenboom <linux@eikelenboom.it>
Reported-by: Dave Jones <davej@redhat.com>
Tested-by: Dave Jones <davej@redhat.com>
Tested-by: Sander Eikelenboom <linux@eikelenboom.it>
Cc: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com>
Cc: Francois Romieu <romieu@fr.zoreil.com>
Cc: Eric Dumazet <eric.dumazet@gmail.com>
Cc: Wei Liu <wei.liu2@citrix.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
lib/dma-debug.c

index 2defd1308b045c46389a6232391999914573deb6..98f2d7e91a9114e5e124e42bef00d7935708e521 100644 (file)
@@ -424,111 +424,134 @@ void debug_dma_dump_mappings(struct device *dev)
 EXPORT_SYMBOL(debug_dma_dump_mappings);
 
 /*
- * For each page mapped (initial page in the case of
- * dma_alloc_coherent/dma_map_{single|page}, or each page in a
- * scatterlist) insert into this tree using the pfn as the key. At
+ * For each mapping (initial cacheline in the case of
+ * dma_alloc_coherent/dma_map_page, initial cacheline in each page of a
+ * scatterlist, or the cacheline specified in dma_map_single) insert
+ * into this tree using the cacheline as the key. At
  * dma_unmap_{single|sg|page} or dma_free_coherent delete the entry.  If
- * the pfn already exists at insertion time add a tag as a reference
+ * the entry already exists at insertion time add a tag as a reference
  * count for the overlapping mappings.  For now, the overlap tracking
- * just ensures that 'unmaps' balance 'maps' before marking the pfn
- * idle, but we should also be flagging overlaps as an API violation.
+ * just ensures that 'unmaps' balance 'maps' before marking the
+ * cacheline idle, but we should also be flagging overlaps as an API
+ * violation.
  *
  * Memory usage is mostly constrained by the maximum number of available
  * dma-debug entries in that we need a free dma_debug_entry before
- * inserting into the tree.  In the case of dma_map_{single|page} and
- * dma_alloc_coherent there is only one dma_debug_entry and one pfn to
- * track per event.  dma_map_sg(), on the other hand,
- * consumes a single dma_debug_entry, but inserts 'nents' entries into
- * the tree.
+ * inserting into the tree.  In the case of dma_map_page and
+ * dma_alloc_coherent there is only one dma_debug_entry and one
+ * dma_active_cacheline entry to track per event.  dma_map_sg(), on the
+ * other hand, consumes a single dma_debug_entry, but inserts 'nents'
+ * entries into the tree.
  *
  * At any time debug_dma_assert_idle() can be called to trigger a
- * warning if the given page is in the active set.
+ * warning if any cachelines in the given page are in the active set.
  */
-static RADIX_TREE(dma_active_pfn, GFP_NOWAIT);
+static RADIX_TREE(dma_active_cacheline, GFP_NOWAIT);
 static DEFINE_SPINLOCK(radix_lock);
-#define ACTIVE_PFN_MAX_OVERLAP ((1 << RADIX_TREE_MAX_TAGS) - 1)
+#define ACTIVE_CACHELINE_MAX_OVERLAP ((1 << RADIX_TREE_MAX_TAGS) - 1)
+#define CACHELINE_PER_PAGE_SHIFT (PAGE_SHIFT - L1_CACHE_SHIFT)
+#define CACHELINES_PER_PAGE (1 << CACHELINE_PER_PAGE_SHIFT)
 
-static int active_pfn_read_overlap(unsigned long pfn)
+static phys_addr_t to_cacheline_number(struct dma_debug_entry *entry)
+{
+       return (entry->pfn << CACHELINE_PER_PAGE_SHIFT) +
+               (entry->offset >> L1_CACHE_SHIFT);
+}
+
+static int active_cacheline_read_overlap(phys_addr_t cln)
 {
        int overlap = 0, i;
 
        for (i = RADIX_TREE_MAX_TAGS - 1; i >= 0; i--)
-               if (radix_tree_tag_get(&dma_active_pfn, pfn, i))
+               if (radix_tree_tag_get(&dma_active_cacheline, cln, i))
                        overlap |= 1 << i;
        return overlap;
 }
 
-static int active_pfn_set_overlap(unsigned long pfn, int overlap)
+static int active_cacheline_set_overlap(phys_addr_t cln, int overlap)
 {
        int i;
 
-       if (overlap > ACTIVE_PFN_MAX_OVERLAP || overlap < 0)
+       if (overlap > ACTIVE_CACHELINE_MAX_OVERLAP || overlap < 0)
                return overlap;
 
        for (i = RADIX_TREE_MAX_TAGS - 1; i >= 0; i--)
                if (overlap & 1 << i)
-                       radix_tree_tag_set(&dma_active_pfn, pfn, i);
+                       radix_tree_tag_set(&dma_active_cacheline, cln, i);
                else
-                       radix_tree_tag_clear(&dma_active_pfn, pfn, i);
+                       radix_tree_tag_clear(&dma_active_cacheline, cln, i);
 
        return overlap;
 }
 
-static void active_pfn_inc_overlap(unsigned long pfn)
+static void active_cacheline_inc_overlap(phys_addr_t cln)
 {
-       int overlap = active_pfn_read_overlap(pfn);
+       int overlap = active_cacheline_read_overlap(cln);
 
-       overlap = active_pfn_set_overlap(pfn, ++overlap);
+       overlap = active_cacheline_set_overlap(cln, ++overlap);
 
        /* If we overflowed the overlap counter then we're potentially
         * leaking dma-mappings.  Otherwise, if maps and unmaps are
         * balanced then this overflow may cause false negatives in
-        * debug_dma_assert_idle() as the pfn may be marked idle
+        * debug_dma_assert_idle() as the cacheline may be marked idle
         * prematurely.
         */
-       WARN_ONCE(overlap > ACTIVE_PFN_MAX_OVERLAP,
-                 "DMA-API: exceeded %d overlapping mappings of pfn %lx\n",
-                 ACTIVE_PFN_MAX_OVERLAP, pfn);
+       WARN_ONCE(overlap > ACTIVE_CACHELINE_MAX_OVERLAP,
+                 "DMA-API: exceeded %d overlapping mappings of cacheline %pa\n",
+                 ACTIVE_CACHELINE_MAX_OVERLAP, &cln);
 }
 
-static int active_pfn_dec_overlap(unsigned long pfn)
+static int active_cacheline_dec_overlap(phys_addr_t cln)
 {
-       int overlap = active_pfn_read_overlap(pfn);
+       int overlap = active_cacheline_read_overlap(cln);
 
-       return active_pfn_set_overlap(pfn, --overlap);
+       return active_cacheline_set_overlap(cln, --overlap);
 }
 
-static int active_pfn_insert(struct dma_debug_entry *entry)
+static int active_cacheline_insert(struct dma_debug_entry *entry)
 {
+       phys_addr_t cln = to_cacheline_number(entry);
        unsigned long flags;
        int rc;
 
+       /* If the device is not writing memory then we don't have any
+        * concerns about the cpu consuming stale data.  This mitigates
+        * legitimate usages of overlapping mappings.
+        */
+       if (entry->direction == DMA_TO_DEVICE)
+               return 0;
+
        spin_lock_irqsave(&radix_lock, flags);
-       rc = radix_tree_insert(&dma_active_pfn, entry->pfn, entry);
+       rc = radix_tree_insert(&dma_active_cacheline, cln, entry);
        if (rc == -EEXIST)
-               active_pfn_inc_overlap(entry->pfn);
+               active_cacheline_inc_overlap(cln);
        spin_unlock_irqrestore(&radix_lock, flags);
 
        return rc;
 }
 
-static void active_pfn_remove(struct dma_debug_entry *entry)
+static void active_cacheline_remove(struct dma_debug_entry *entry)
 {
+       phys_addr_t cln = to_cacheline_number(entry);
        unsigned long flags;
 
+       /* ...mirror the insert case */
+       if (entry->direction == DMA_TO_DEVICE)
+               return;
+
        spin_lock_irqsave(&radix_lock, flags);
        /* since we are counting overlaps the final put of the
-        * entry->pfn will occur when the overlap count is 0.
-        * active_pfn_dec_overlap() returns -1 in that case
+        * cacheline will occur when the overlap count is 0.
+        * active_cacheline_dec_overlap() returns -1 in that case
         */
-       if (active_pfn_dec_overlap(entry->pfn) < 0)
-               radix_tree_delete(&dma_active_pfn, entry->pfn);
+       if (active_cacheline_dec_overlap(cln) < 0)
+               radix_tree_delete(&dma_active_cacheline, cln);
        spin_unlock_irqrestore(&radix_lock, flags);
 }
 
 /**
  * debug_dma_assert_idle() - assert that a page is not undergoing dma
- * @page: page to lookup in the dma_active_pfn tree
+ * @page: page to lookup in the dma_active_cacheline tree
  *
  * Place a call to this routine in cases where the cpu touching the page
  * before the dma completes (page is dma_unmapped) will lead to data
@@ -536,22 +559,38 @@ static void active_pfn_remove(struct dma_debug_entry *entry)
  */
 void debug_dma_assert_idle(struct page *page)
 {
+       static struct dma_debug_entry *ents[CACHELINES_PER_PAGE];
+       struct dma_debug_entry *entry = NULL;
+       void **results = (void **) &ents;
+       unsigned int nents, i;
        unsigned long flags;
-       struct dma_debug_entry *entry;
+       phys_addr_t cln;
 
        if (!page)
                return;
 
+       cln = (phys_addr_t) page_to_pfn(page) << CACHELINE_PER_PAGE_SHIFT;
        spin_lock_irqsave(&radix_lock, flags);
-       entry = radix_tree_lookup(&dma_active_pfn, page_to_pfn(page));
+       nents = radix_tree_gang_lookup(&dma_active_cacheline, results, cln,
+                                      CACHELINES_PER_PAGE);
+       for (i = 0; i < nents; i++) {
+               phys_addr_t ent_cln = to_cacheline_number(ents[i]);
+
+               if (ent_cln == cln) {
+                       entry = ents[i];
+                       break;
+               } else if (ent_cln >= cln + CACHELINES_PER_PAGE)
+                       break;
+       }
        spin_unlock_irqrestore(&radix_lock, flags);
 
        if (!entry)
                return;
 
+       cln = to_cacheline_number(entry);
        err_printk(entry->dev, entry,
-                  "DMA-API: cpu touching an active dma mapped page "
-                  "[pfn=0x%lx]\n", entry->pfn);
+                  "DMA-API: cpu touching an active dma mapped cacheline [cln=%pa]\n",
+                  &cln);
 }
 
 /*
@@ -568,9 +607,9 @@ static void add_dma_entry(struct dma_debug_entry *entry)
        hash_bucket_add(bucket, entry);
        put_hash_bucket(bucket, &flags);
 
-       rc = active_pfn_insert(entry);
+       rc = active_cacheline_insert(entry);
        if (rc == -ENOMEM) {
-               pr_err("DMA-API: pfn tracking ENOMEM, dma-debug disabled\n");
+               pr_err("DMA-API: cacheline tracking ENOMEM, dma-debug disabled\n");
                global_disable = true;
        }
 
@@ -631,7 +670,7 @@ static void dma_entry_free(struct dma_debug_entry *entry)
 {
        unsigned long flags;
 
-       active_pfn_remove(entry);
+       active_cacheline_remove(entry);
 
        /*
         * add to beginning of the list - this way the entries are