devm_memremap_pages: fix vmem_altmap lifetime + alignment handling

to_vmem_altmap() needs to return valid results until
arch_remove_memory() completes.  It also needs to be valid for any pfn
in a section regardless of whether that pfn maps to data.  This escape
was a result of a bug in the unit test.

The signature of this bug is that free_pagetable() fails to retrieve a
vmem_altmap and goes off into the weeds:

 BUG: unable to handle kernel NULL pointer dereference at           (null)
 IP: [<ffffffff811d2629>] get_pfnblock_flags_mask+0x49/0x60
 [..]
 Call Trace:
  [<ffffffff811d3477>] free_hot_cold_page+0x97/0x1d0
  [<ffffffff811d367a>] __free_pages+0x2a/0x40
  [<ffffffff8191e669>] free_pagetable+0x8c/0xd4
  [<ffffffff8191ef4e>] remove_pagetable+0x37a/0x808
  [<ffffffff8191b210>] vmemmap_free+0x10/0x20

Fixes: 4b94ffdc4163 ("x86, mm: introduce vmem_altmap to augment vmemmap_populate()")
Cc: Andrew Morton <akpm@linux-foundation.org>
Reported-by: Jeff Moyer <jmoyer@redhat.com>
Signed-off-by: Dan Williams <dan.j.williams@intel.com>

diff --git a/kernel/memremap.c b/kernel/memremap.c
index e517a16..cbc3e97 100644 (file)
--- a/kernel/memremap.c
+++ b/kernel/memremap.c
@@ -183,7 +183,11 @@ EXPORT_SYMBOL(put_zone_device_page);
 
 static void pgmap_radix_release(struct resource *res)
 {
-       resource_size_t key;
+       resource_size_t key, align_start, align_size, align_end;
+
+       align_start = res->start & ~(SECTION_SIZE - 1);
+       align_size = ALIGN(resource_size(res), SECTION_SIZE);
+       align_end = align_start + align_size - 1;
 
        mutex_lock(&pgmap_lock);
        for (key = res->start; key <= res->end; key += SECTION_SIZE)
@@ -226,12 +230,11 @@ static void devm_memremap_pages_release(struct device *dev, void *data)
                percpu_ref_put(pgmap->ref);
        }
 
-       pgmap_radix_release(res);
-
        /* pages are dead and unused, undo the arch mapping */
        align_start = res->start & ~(SECTION_SIZE - 1);
        align_size = ALIGN(resource_size(res), SECTION_SIZE);
        arch_remove_memory(align_start, align_size);
+       pgmap_radix_release(res);
        dev_WARN_ONCE(dev, pgmap->altmap && pgmap->altmap->alloc,
                        "%s: failed to free all reserved pages\n", __func__);
 }
@@ -267,7 +270,7 @@ void *devm_memremap_pages(struct device *dev, struct resource *res,
 {
        int is_ram = region_intersects(res->start, resource_size(res),
                        "System RAM");
-       resource_size_t key, align_start, align_size;
+       resource_size_t key, align_start, align_size, align_end;
        struct dev_pagemap *pgmap;
        struct page_map *page_map;
        unsigned long pfn;
@@ -309,7 +312,10 @@ void *devm_memremap_pages(struct device *dev, struct resource *res,
 
        mutex_lock(&pgmap_lock);
        error = 0;
-       for (key = res->start; key <= res->end; key += SECTION_SIZE) {
+       align_start = res->start & ~(SECTION_SIZE - 1);
+       align_size = ALIGN(resource_size(res), SECTION_SIZE);
+       align_end = align_start + align_size - 1;
+       for (key = align_start; key <= align_end; key += SECTION_SIZE) {
                struct dev_pagemap *dup;
 
                rcu_read_lock();
@@ -336,8 +342,6 @@ void *devm_memremap_pages(struct device *dev, struct resource *res,
        if (nid < 0)
                nid = numa_mem_id();
 
-       align_start = res->start & ~(SECTION_SIZE - 1);
-       align_size = ALIGN(resource_size(res), SECTION_SIZE);
        error = arch_add_memory(nid, align_start, align_size, true);
        if (error)
                goto err_add_memory;