libnvdimm, pmem: adjust for section collisions with 'System RAM'
authorDan Williams <dan.j.williams@intel.com>
Thu, 3 Mar 2016 17:38:00 +0000 (09:38 -0800)
committerDan Williams <dan.j.williams@intel.com>
Sat, 5 Mar 2016 20:25:45 +0000 (12:25 -0800)
On a platform where 'Persistent Memory' and 'System RAM' are mixed
within a given sparsemem section, trim the namespace and notify about the
sub-optimal alignment.

Cc: Toshi Kani <toshi.kani@hpe.com>
Cc: Ross Zwisler <ross.zwisler@linux.intel.com>
Signed-off-by: Dan Williams <dan.j.williams@intel.com>
drivers/nvdimm/namespace_devs.c
drivers/nvdimm/pfn.h
drivers/nvdimm/pfn_devs.c
drivers/nvdimm/pmem.c

index 9edf7eb7d17cc39713e67e4dd62d4e16c0d93e07..f5cb886013591bfa384fc5f19d95eb823ee1e1c3 100644 (file)
@@ -133,6 +133,7 @@ bool nd_is_uuid_unique(struct device *dev, u8 *uuid)
 bool pmem_should_map_pages(struct device *dev)
 {
        struct nd_region *nd_region = to_nd_region(dev->parent);
+       struct nd_namespace_io *nsio;
 
        if (!IS_ENABLED(CONFIG_ZONE_DEVICE))
                return false;
@@ -143,6 +144,12 @@ bool pmem_should_map_pages(struct device *dev)
        if (is_nd_pfn(dev) || is_nd_btt(dev))
                return false;
 
+       nsio = to_nd_namespace_io(dev);
+       if (region_intersects(nsio->res.start, resource_size(&nsio->res),
+                               IORESOURCE_SYSTEM_RAM,
+                               IORES_DESC_NONE) == REGION_MIXED)
+               return false;
+
 #ifdef ARCH_MEMREMAP_PMEM
        return ARCH_MEMREMAP_PMEM == MEMREMAP_WB;
 #else
index 6ee707e5b279785170a3def843aaae5e43d8a28e..8e343a3ca873012a7124c46564f9bcc4833c7a67 100644 (file)
@@ -27,10 +27,13 @@ struct nd_pfn_sb {
        __le32 flags;
        __le16 version_major;
        __le16 version_minor;
-       __le64 dataoff;
+       __le64 dataoff; /* relative to namespace_base + start_pad */
        __le64 npfns;
        __le32 mode;
-       u8 padding[4012];
+       /* minor-version-1 additions for section alignment */
+       __le32 start_pad;
+       __le32 end_trunc;
+       u8 padding[4004];
        __le64 checksum;
 };
 
@@ -45,4 +48,7 @@ struct nd_pfn_sb {
 #define PFN_SECTION_ALIGN_DOWN(x) (x)
 #define PFN_SECTION_ALIGN_UP(x) (x)
 #endif
+
+#define PHYS_SECTION_ALIGN_DOWN(x) PFN_PHYS(PFN_SECTION_ALIGN_DOWN(PHYS_PFN(x)))
+#define PHYS_SECTION_ALIGN_UP(x) PFN_PHYS(PFN_SECTION_ALIGN_UP(PHYS_PFN(x)))
 #endif /* __NVDIMM_PFN_H */
index ae81a2f1da50b1e9f9cf5a14bd180df19b3ccab1..75a31a7359fb4c57eedad346d2e8623ad23e8da7 100644 (file)
@@ -299,6 +299,11 @@ int nd_pfn_validate(struct nd_pfn *nd_pfn)
        if (memcmp(pfn_sb->parent_uuid, parent_uuid, 16) != 0)
                return -ENODEV;
 
+       if (__le16_to_cpu(pfn_sb->version_minor) < 1) {
+               pfn_sb->start_pad = 0;
+               pfn_sb->end_trunc = 0;
+       }
+
        switch (le32_to_cpu(pfn_sb->mode)) {
        case PFN_MODE_RAM:
        case PFN_MODE_PMEM:
index 59d568ab75562ef4cd32e2f08a855bdfe0019b92..0cb450e1b400b5535562fd1c9bf795afa4bc0983 100644 (file)
@@ -43,7 +43,10 @@ struct pmem_device {
        phys_addr_t             data_offset;
        u64                     pfn_flags;
        void __pmem             *virt_addr;
+       /* immutable base size of the namespace */
        size_t                  size;
+       /* trim size when namespace capacity has been section aligned */
+       u32                     pfn_pad;
        struct badblocks        bb;
 };
 
@@ -145,7 +148,7 @@ static long pmem_direct_access(struct block_device *bdev, sector_t sector,
        *kaddr = pmem->virt_addr + offset;
        *pfn = phys_to_pfn_t(pmem->phys_addr + offset, pmem->pfn_flags);
 
-       return pmem->size - offset;
+       return pmem->size - pmem->pfn_pad - offset;
 }
 
 static const struct block_device_operations pmem_fops = {
@@ -236,7 +239,8 @@ static int pmem_attach_disk(struct device *dev,
        disk->flags             = GENHD_FL_EXT_DEVT;
        nvdimm_namespace_disk_name(ndns, disk->disk_name);
        disk->driverfs_dev = dev;
-       set_capacity(disk, (pmem->size - pmem->data_offset) / 512);
+       set_capacity(disk, (pmem->size - pmem->pfn_pad - pmem->data_offset)
+                       / 512);
        pmem->pmem_disk = disk;
        devm_exit_badblocks(dev, &pmem->bb);
        if (devm_init_badblocks(dev, &pmem->bb))
@@ -279,6 +283,9 @@ static int nd_pfn_init(struct nd_pfn *nd_pfn)
        struct nd_pfn_sb *pfn_sb = kzalloc(sizeof(*pfn_sb), GFP_KERNEL);
        struct pmem_device *pmem = dev_get_drvdata(&nd_pfn->dev);
        struct nd_namespace_common *ndns = nd_pfn->ndns;
+       u32 start_pad = 0, end_trunc = 0;
+       resource_size_t start, size;
+       struct nd_namespace_io *nsio;
        struct nd_region *nd_region;
        unsigned long npfns;
        phys_addr_t offset;
@@ -304,21 +311,56 @@ static int nd_pfn_init(struct nd_pfn *nd_pfn)
        }
 
        memset(pfn_sb, 0, sizeof(*pfn_sb));
-       npfns = (pmem->size - SZ_8K) / SZ_4K;
+
+       /*
+        * Check if pmem collides with 'System RAM' when section aligned and
+        * trim it accordingly
+        */
+       nsio = to_nd_namespace_io(&ndns->dev);
+       start = PHYS_SECTION_ALIGN_DOWN(nsio->res.start);
+       size = resource_size(&nsio->res);
+       if (region_intersects(start, size, IORESOURCE_SYSTEM_RAM,
+                               IORES_DESC_NONE) == REGION_MIXED) {
+
+               start = nsio->res.start;
+               start_pad = PHYS_SECTION_ALIGN_UP(start) - start;
+       }
+
+       start = nsio->res.start;
+       size = PHYS_SECTION_ALIGN_UP(start + size) - start;
+       if (region_intersects(start, size, IORESOURCE_SYSTEM_RAM,
+                               IORES_DESC_NONE) == REGION_MIXED) {
+               size = resource_size(&nsio->res);
+               end_trunc = start + size - PHYS_SECTION_ALIGN_DOWN(start + size);
+       }
+
+       if (start_pad + end_trunc)
+               dev_info(&nd_pfn->dev, "%s section collision, truncate %d bytes\n",
+                               dev_name(&ndns->dev), start_pad + end_trunc);
+
        /*
         * Note, we use 64 here for the standard size of struct page,
         * debugging options may cause it to be larger in which case the
         * implementation will limit the pfns advertised through
         * ->direct_access() to those that are included in the memmap.
         */
+       start += start_pad;
+       npfns = (pmem->size - start_pad - end_trunc - SZ_8K) / SZ_4K;
        if (nd_pfn->mode == PFN_MODE_PMEM)
-               offset = ALIGN(SZ_8K + 64 * npfns, nd_pfn->align);
+               offset = ALIGN(start + SZ_8K + 64 * npfns, nd_pfn->align)
+                       - start;
        else if (nd_pfn->mode == PFN_MODE_RAM)
-               offset = ALIGN(SZ_8K, nd_pfn->align);
+               offset = ALIGN(start + SZ_8K, nd_pfn->align) - start;
        else
                goto err;
 
-       npfns = (pmem->size - offset) / SZ_4K;
+       if (offset + start_pad + end_trunc >= pmem->size) {
+               dev_err(&nd_pfn->dev, "%s unable to satisfy requested alignment\n",
+                               dev_name(&ndns->dev));
+               goto err;
+       }
+
+       npfns = (pmem->size - offset - start_pad - end_trunc) / SZ_4K;
        pfn_sb->mode = cpu_to_le32(nd_pfn->mode);
        pfn_sb->dataoff = cpu_to_le64(offset);
        pfn_sb->npfns = cpu_to_le64(npfns);
@@ -326,6 +368,9 @@ static int nd_pfn_init(struct nd_pfn *nd_pfn)
        memcpy(pfn_sb->uuid, nd_pfn->uuid, 16);
        memcpy(pfn_sb->parent_uuid, nd_dev_to_uuid(&ndns->dev), 16);
        pfn_sb->version_major = cpu_to_le16(1);
+       pfn_sb->version_minor = cpu_to_le16(1);
+       pfn_sb->start_pad = cpu_to_le32(start_pad);
+       pfn_sb->end_trunc = cpu_to_le32(end_trunc);
        checksum = nd_sb_checksum((struct nd_gen_sb *) pfn_sb);
        pfn_sb->checksum = cpu_to_le64(checksum);
 
@@ -376,41 +421,36 @@ static unsigned long init_altmap_reserve(resource_size_t base)
        return reserve;
 }
 
-static int nvdimm_namespace_attach_pfn(struct nd_namespace_common *ndns)
+static int __nvdimm_namespace_attach_pfn(struct nd_pfn *nd_pfn)
 {
-       struct nd_namespace_io *nsio = to_nd_namespace_io(&ndns->dev);
-       struct nd_pfn *nd_pfn = to_nd_pfn(ndns->claim);
-       struct device *dev = &nd_pfn->dev;
-       struct nd_region *nd_region;
-       struct vmem_altmap *altmap;
-       struct nd_pfn_sb *pfn_sb;
-       struct pmem_device *pmem;
-       struct request_queue *q;
-       phys_addr_t offset;
        int rc;
+       struct resource res;
+       struct request_queue *q;
+       struct pmem_device *pmem;
+       struct vmem_altmap *altmap;
+       struct device *dev = &nd_pfn->dev;
+       struct nd_pfn_sb *pfn_sb = nd_pfn->pfn_sb;
+       struct nd_namespace_common *ndns = nd_pfn->ndns;
+       u32 start_pad = __le32_to_cpu(pfn_sb->start_pad);
+       u32 end_trunc = __le32_to_cpu(pfn_sb->end_trunc);
+       struct nd_namespace_io *nsio = to_nd_namespace_io(&ndns->dev);
+       resource_size_t base = nsio->res.start + start_pad;
        struct vmem_altmap __altmap = {
-               .base_pfn = init_altmap_base(nsio->res.start),
-               .reserve = init_altmap_reserve(nsio->res.start),
+               .base_pfn = init_altmap_base(base),
+               .reserve = init_altmap_reserve(base),
        };
 
-       if (!nd_pfn->uuid || !nd_pfn->ndns)
-               return -ENODEV;
-
-       nd_region = to_nd_region(dev->parent);
-       rc = nd_pfn_init(nd_pfn);
-       if (rc)
-               return rc;
-
-       pfn_sb = nd_pfn->pfn_sb;
-       offset = le64_to_cpu(pfn_sb->dataoff);
+       pmem = dev_get_drvdata(dev);
+       pmem->data_offset = le64_to_cpu(pfn_sb->dataoff);
+       pmem->pfn_pad = start_pad + end_trunc;
        nd_pfn->mode = le32_to_cpu(nd_pfn->pfn_sb->mode);
        if (nd_pfn->mode == PFN_MODE_RAM) {
-               if (offset < SZ_8K)
+               if (pmem->data_offset < SZ_8K)
                        return -EINVAL;
                nd_pfn->npfns = le64_to_cpu(pfn_sb->npfns);
                altmap = NULL;
        } else if (nd_pfn->mode == PFN_MODE_PMEM) {
-               nd_pfn->npfns = (resource_size(&nsio->res) - offset)
+               nd_pfn->npfns = (pmem->size - pmem->pfn_pad - pmem->data_offset)
                        / PAGE_SIZE;
                if (le64_to_cpu(nd_pfn->pfn_sb->npfns) > nd_pfn->npfns)
                        dev_info(&nd_pfn->dev,
@@ -418,7 +458,7 @@ static int nvdimm_namespace_attach_pfn(struct nd_namespace_common *ndns)
                                        le64_to_cpu(nd_pfn->pfn_sb->npfns),
                                        nd_pfn->npfns);
                altmap = & __altmap;
-               altmap->free = __phys_to_pfn(offset - SZ_8K);
+               altmap->free = __phys_to_pfn(pmem->data_offset - SZ_8K);
                altmap->alloc = 0;
        } else {
                rc = -ENXIO;
@@ -426,10 +466,12 @@ static int nvdimm_namespace_attach_pfn(struct nd_namespace_common *ndns)
        }
 
        /* establish pfn range for lookup, and switch to direct map */
-       pmem = dev_get_drvdata(dev);
        q = pmem->pmem_queue;
+       memcpy(&res, &nsio->res, sizeof(res));
+       res.start += start_pad;
+       res.end -= end_trunc;
        devm_memunmap(dev, (void __force *) pmem->virt_addr);
-       pmem->virt_addr = (void __pmem *) devm_memremap_pages(dev, &nsio->res,
+       pmem->virt_addr = (void __pmem *) devm_memremap_pages(dev, &res,
                        &q->q_usage_counter, altmap);
        pmem->pfn_flags |= PFN_MAP;
        if (IS_ERR(pmem->virt_addr)) {
@@ -438,7 +480,6 @@ static int nvdimm_namespace_attach_pfn(struct nd_namespace_common *ndns)
        }
 
        /* attach pmem disk in "pfn-mode" */
-       pmem->data_offset = offset;
        rc = pmem_attach_disk(dev, ndns, pmem);
        if (rc)
                goto err;
@@ -447,6 +488,22 @@ static int nvdimm_namespace_attach_pfn(struct nd_namespace_common *ndns)
  err:
        nvdimm_namespace_detach_pfn(ndns);
        return rc;
+
+}
+
+static int nvdimm_namespace_attach_pfn(struct nd_namespace_common *ndns)
+{
+       struct nd_pfn *nd_pfn = to_nd_pfn(ndns->claim);
+       int rc;
+
+       if (!nd_pfn->uuid || !nd_pfn->ndns)
+               return -ENODEV;
+
+       rc = nd_pfn_init(nd_pfn);
+       if (rc)
+               return rc;
+       /* we need a valid pfn_sb before we can init a vmem_altmap */
+       return __nvdimm_namespace_attach_pfn(nd_pfn);
 }
 
 static int nd_pmem_probe(struct device *dev)