libnvdimm/altmap: Track namespace boundaries in altmap
authorAneesh Kumar K.V <aneesh.kumar@linux.ibm.com>
Tue, 10 Sep 2019 06:28:25 +0000 (11:58 +0530)
committerDan Williams <dan.j.williams@intel.com>
Tue, 24 Sep 2019 17:24:12 +0000 (10:24 -0700)
With PFN_MODE_PMEM namespace, the memmap area is allocated from the device
area. Some architectures map the memmap area with large page size. On
architectures like ppc64, 16MB page for memap mapping can map 262144 pfns.
This maps a namespace size of 16G.

When populating memmap region with 16MB page from the device area,
make sure the allocated space is not used to map resources outside this
namespace. Such usage of device area will prevent a namespace destroy.

Add resource end pnf in altmap and use that to check if the memmap area
allocation can map pfn outside the namespace. On ppc64 in such case we fallback
to allocation from memory.

This fix kernel crash reported below:

[  132.034989] WARNING: CPU: 13 PID: 13719 at mm/memremap.c:133 devm_memremap_pages_release+0x2d8/0x2e0
[  133.464754] BUG: Unable to handle kernel data access at 0xc00c00010b204000
[  133.464760] Faulting instruction address: 0xc00000000007580c
[  133.464766] Oops: Kernel access of bad area, sig: 11 [#1]
[  133.464771] LE PAGE_SIZE=64K MMU=Hash SMP NR_CPUS=2048 NUMA pSeries
.....
[  133.464901] NIP [c00000000007580c] vmemmap_free+0x2ac/0x3d0
[  133.464906] LR [c0000000000757f8] vmemmap_free+0x298/0x3d0
[  133.464910] Call Trace:
[  133.464914] [c000007cbfd0f7b0] [c0000000000757f8] vmemmap_free+0x298/0x3d0 (unreliable)
[  133.464921] [c000007cbfd0f8d0] [c000000000370a44] section_deactivate+0x1a4/0x240
[  133.464928] [c000007cbfd0f980] [c000000000386270] __remove_pages+0x3a0/0x590
[  133.464935] [c000007cbfd0fa50] [c000000000074158] arch_remove_memory+0x88/0x160
[  133.464942] [c000007cbfd0fae0] [c0000000003be8c0] devm_memremap_pages_release+0x150/0x2e0
[  133.464949] [c000007cbfd0fb70] [c000000000738ea0] devm_action_release+0x30/0x50
[  133.464955] [c000007cbfd0fb90] [c00000000073a5a4] release_nodes+0x344/0x400
[  133.464961] [c000007cbfd0fc40] [c00000000073378c] device_release_driver_internal+0x15c/0x250
[  133.464968] [c000007cbfd0fc80] [c00000000072fd14] unbind_store+0x104/0x110
[  133.464973] [c000007cbfd0fcd0] [c00000000072ee24] drv_attr_store+0x44/0x70
[  133.464981] [c000007cbfd0fcf0] [c0000000004a32bc] sysfs_kf_write+0x6c/0xa0
[  133.464987] [c000007cbfd0fd10] [c0000000004a1dfc] kernfs_fop_write+0x17c/0x250
[  133.464993] [c000007cbfd0fd60] [c0000000003c348c] __vfs_write+0x3c/0x70
[  133.464999] [c000007cbfd0fd80] [c0000000003c75d0] vfs_write+0xd0/0x250

djbw: Aneesh notes that this crash can likely be triggered in any kernel that
supports 'papr_scm', so flagging that commit for -stable consideration.

Fixes: b5beae5e224f ("powerpc/pseries: Add driver for PAPR SCM regions")
Cc: <stable@vger.kernel.org>
Reported-by: Sachin Sant <sachinp@linux.vnet.ibm.com>
Signed-off-by: Aneesh Kumar K.V <aneesh.kumar@linux.ibm.com>
Reviewed-by: Pankaj Gupta <pagupta@redhat.com>
Tested-by: Santosh Sivaraj <santosh@fossix.org>
Reviewed-by: Johannes Thumshirn <jthumshirn@suse.de>
Link: https://lore.kernel.org/r/20190910062826.10041-1-aneesh.kumar@linux.ibm.com
Signed-off-by: Dan Williams <dan.j.williams@intel.com>
arch/powerpc/mm/init_64.c
drivers/nvdimm/pfn_devs.c
include/linux/memremap.h

index a44f6281ca3af729e3ecb512b73ce669729df389..4e08246acd79ad54d81a2acf0c2909475a704e63 100644 (file)
@@ -172,6 +172,21 @@ static __meminit void vmemmap_list_populate(unsigned long phys,
        vmemmap_list = vmem_back;
 }
 
+static bool altmap_cross_boundary(struct vmem_altmap *altmap, unsigned long start,
+                               unsigned long page_size)
+{
+       unsigned long nr_pfn = page_size / sizeof(struct page);
+       unsigned long start_pfn = page_to_pfn((struct page *)start);
+
+       if ((start_pfn + nr_pfn) > altmap->end_pfn)
+               return true;
+
+       if (start_pfn < altmap->base_pfn)
+               return true;
+
+       return false;
+}
+
 int __meminit vmemmap_populate(unsigned long start, unsigned long end, int node,
                struct vmem_altmap *altmap)
 {
@@ -194,7 +209,7 @@ int __meminit vmemmap_populate(unsigned long start, unsigned long end, int node,
                 * fail due to alignment issues when using 16MB hugepages, so
                 * fall back to system memory if the altmap allocation fail.
                 */
-               if (altmap) {
+               if (altmap && !altmap_cross_boundary(altmap, start, page_size)) {
                        p = altmap_alloc_block_buf(page_size, altmap);
                        if (!p)
                                pr_debug("altmap block allocation failed, falling back to system memory");
index 934cdcaaae97cd83ac0521f0d76e6c70e01d23f3..80c7992bc53899ce9aa70fcb34dc2e630eabdcc0 100644 (file)
@@ -672,9 +672,11 @@ static int __nvdimm_setup_pfn(struct nd_pfn *nd_pfn, struct dev_pagemap *pgmap)
        struct nd_namespace_common *ndns = nd_pfn->ndns;
        struct nd_namespace_io *nsio = to_nd_namespace_io(&ndns->dev);
        resource_size_t base = nsio->res.start + start_pad;
+       resource_size_t end = nsio->res.end - end_trunc;
        struct vmem_altmap __altmap = {
                .base_pfn = init_altmap_base(base),
                .reserve = init_altmap_reserve(base),
+               .end_pfn = PHYS_PFN(end),
        };
 
        memcpy(res, &nsio->res, sizeof(*res));
index f8a5b2a199459c71bdfa343f48a9665f34aff080..c70996fe48c85cac55f20f2fed766405644e2eec 100644 (file)
@@ -17,6 +17,7 @@ struct device;
  */
 struct vmem_altmap {
        const unsigned long base_pfn;
+       const unsigned long end_pfn;
        const unsigned long reserve;
        unsigned long free;
        unsigned long align;