libnvdimm: fix clear poison locking with spinlock and GFP_NOWAIT allocation
authorDave Jiang <dave.jiang@intel.com>
Thu, 13 Apr 2017 21:25:17 +0000 (14:25 -0700)
committerDan Williams <dan.j.williams@intel.com>
Thu, 13 Apr 2017 21:23:51 +0000 (14:23 -0700)
The following warning results from holding a lane spinlock,
preempt_disable(), or the btt map spinlock and then trying to take the
reconfig_mutex to walk the poison list and potentially add new entries.

BUG: sleeping function called from invalid context at kernel/locking/mutex.
c:747
in_atomic(): 1, irqs_disabled(): 0, pid: 17159, name: dd
[..]
Call Trace:
dump_stack+0x85/0xc8
___might_sleep+0x184/0x250
__might_sleep+0x4a/0x90
__mutex_lock+0x58/0x9b0
? nvdimm_bus_lock+0x21/0x30 [libnvdimm]
? __nvdimm_bus_badblocks_clear+0x2f/0x60 [libnvdimm]
? acpi_nfit_forget_poison+0x79/0x80 [nfit]
? _raw_spin_unlock+0x27/0x40
mutex_lock_nested+0x1b/0x20
nvdimm_bus_lock+0x21/0x30 [libnvdimm]
nvdimm_forget_poison+0x25/0x50 [libnvdimm]
nvdimm_clear_poison+0x106/0x140 [libnvdimm]
nsio_rw_bytes+0x164/0x270 [libnvdimm]
btt_write_pg+0x1de/0x3e0 [nd_btt]
? blk_queue_enter+0x30/0x290
btt_make_request+0x11a/0x310 [nd_btt]
? blk_queue_enter+0xb7/0x290
? blk_queue_enter+0x30/0x290
generic_make_request+0x118/0x3b0

A spinlock is introduced to protect the poison list. This allows us to not
having to acquire the reconfig_mutex for touching the poison list. The
add_poison() function has been broken out into two helper functions. One to
allocate the poison entry and the other to apppend the entry. This allows us
to unlock the poison_lock in non-I/O path and continue to be able to allocate
the poison entry with GFP_KERNEL. We will use GFP_NOWAIT in the I/O path in
order to satisfy being in atomic context.

Reviewed-by: Vishal Verma <vishal.l.verma@intel.com>
Signed-off-by: Dave Jiang <dave.jiang@intel.com>
Signed-off-by: Dan Williams <dan.j.williams@intel.com>
drivers/nvdimm/bus.c
drivers/nvdimm/core.c
drivers/nvdimm/nd-core.h
include/linux/libnvdimm.h

index 5ad2e59..d214ac4 100644 (file)
@@ -296,6 +296,7 @@ struct nvdimm_bus *nvdimm_bus_register(struct device *parent,
        init_waitqueue_head(&nvdimm_bus->probe_wait);
        nvdimm_bus->id = ida_simple_get(&nd_ida, 0, 0, GFP_KERNEL);
        mutex_init(&nvdimm_bus->reconfig_mutex);
+       spin_lock_init(&nvdimm_bus->poison_lock);
        if (nvdimm_bus->id < 0) {
                kfree(nvdimm_bus);
                return NULL;
@@ -364,9 +365,9 @@ static int nd_bus_remove(struct device *dev)
        nd_synchronize();
        device_for_each_child(&nvdimm_bus->dev, NULL, child_unregister);
 
-       nvdimm_bus_lock(&nvdimm_bus->dev);
+       spin_lock(&nvdimm_bus->poison_lock);
        free_poison_list(&nvdimm_bus->poison_list);
-       nvdimm_bus_unlock(&nvdimm_bus->dev);
+       spin_unlock(&nvdimm_bus->poison_lock);
 
        nvdimm_bus_destroy_ndctl(nvdimm_bus);
 
@@ -990,7 +991,7 @@ static int __nd_ioctl(struct nvdimm_bus *nvdimm_bus, struct nvdimm *nvdimm,
 
                if (clear_err->cleared) {
                        /* clearing the poison list we keep track of */
-                       __nvdimm_forget_poison(nvdimm_bus, clear_err->address,
+                       nvdimm_forget_poison(nvdimm_bus, clear_err->address,
                                        clear_err->cleared);
 
                        /* now sync the badblocks lists */
index 40a3da0..2dee908 100644 (file)
@@ -518,6 +518,15 @@ void nvdimm_badblocks_populate(struct nd_region *nd_region,
 }
 EXPORT_SYMBOL_GPL(nvdimm_badblocks_populate);
 
+static void append_poison_entry(struct nvdimm_bus *nvdimm_bus,
+               struct nd_poison *pl, u64 addr, u64 length)
+{
+       lockdep_assert_held(&nvdimm_bus->poison_lock);
+       pl->start = addr;
+       pl->length = length;
+       list_add_tail(&pl->list, &nvdimm_bus->poison_list);
+}
+
 static int add_poison(struct nvdimm_bus *nvdimm_bus, u64 addr, u64 length,
                        gfp_t flags)
 {
@@ -527,19 +536,24 @@ static int add_poison(struct nvdimm_bus *nvdimm_bus, u64 addr, u64 length,
        if (!pl)
                return -ENOMEM;
 
-       pl->start = addr;
-       pl->length = length;
-       list_add_tail(&pl->list, &nvdimm_bus->poison_list);
-
+       append_poison_entry(nvdimm_bus, pl, addr, length);
        return 0;
 }
 
 static int bus_add_poison(struct nvdimm_bus *nvdimm_bus, u64 addr, u64 length)
 {
-       struct nd_poison *pl;
+       struct nd_poison *pl, *pl_new;
 
-       if (list_empty(&nvdimm_bus->poison_list))
-               return add_poison(nvdimm_bus, addr, length, GFP_KERNEL);
+       spin_unlock(&nvdimm_bus->poison_lock);
+       pl_new = kzalloc(sizeof(*pl_new), GFP_KERNEL);
+       spin_lock(&nvdimm_bus->poison_lock);
+
+       if (list_empty(&nvdimm_bus->poison_list)) {
+               if (!pl_new)
+                       return -ENOMEM;
+               append_poison_entry(nvdimm_bus, pl_new, addr, length);
+               return 0;
+       }
 
        /*
         * There is a chance this is a duplicate, check for those first.
@@ -551,6 +565,7 @@ static int bus_add_poison(struct nvdimm_bus *nvdimm_bus, u64 addr, u64 length)
                        /* If length has changed, update this list entry */
                        if (pl->length != length)
                                pl->length = length;
+                       kfree(pl_new);
                        return 0;
                }
 
@@ -559,30 +574,33 @@ static int bus_add_poison(struct nvdimm_bus *nvdimm_bus, u64 addr, u64 length)
         * as any overlapping ranges will get resolved when the list is consumed
         * and converted to badblocks
         */
-       return add_poison(nvdimm_bus, addr, length, GFP_KERNEL);
+       if (!pl_new)
+               return -ENOMEM;
+       append_poison_entry(nvdimm_bus, pl_new, addr, length);
+
+       return 0;
 }
 
 int nvdimm_bus_add_poison(struct nvdimm_bus *nvdimm_bus, u64 addr, u64 length)
 {
        int rc;
 
-       nvdimm_bus_lock(&nvdimm_bus->dev);
+       spin_lock(&nvdimm_bus->poison_lock);
        rc = bus_add_poison(nvdimm_bus, addr, length);
-       nvdimm_bus_unlock(&nvdimm_bus->dev);
+       spin_unlock(&nvdimm_bus->poison_lock);
 
        return rc;
 }
 EXPORT_SYMBOL_GPL(nvdimm_bus_add_poison);
 
-void __nvdimm_forget_poison(struct nvdimm_bus *nvdimm_bus, phys_addr_t start,
+void nvdimm_forget_poison(struct nvdimm_bus *nvdimm_bus, phys_addr_t start,
                unsigned int len)
 {
        struct list_head *poison_list = &nvdimm_bus->poison_list;
        u64 clr_end = start + len - 1;
        struct nd_poison *pl, *next;
 
-       lockdep_assert_held(&nvdimm_bus->reconfig_mutex);
-
+       spin_lock(&nvdimm_bus->poison_lock);
        WARN_ON_ONCE(list_empty(poison_list));
 
        /*
@@ -629,21 +647,13 @@ void __nvdimm_forget_poison(struct nvdimm_bus *nvdimm_bus, phys_addr_t start,
                        u64 new_len = pl_end - new_start + 1;
 
                        /* Add new entry covering the right half */
-                       add_poison(nvdimm_bus, new_start, new_len, GFP_NOIO);
+                       add_poison(nvdimm_bus, new_start, new_len, GFP_NOWAIT);
                        /* Adjust this entry to cover the left half */
                        pl->length = start - pl->start;
                        continue;
                }
        }
-}
-EXPORT_SYMBOL_GPL(__nvdimm_forget_poison);
-
-void nvdimm_forget_poison(struct nvdimm_bus *nvdimm_bus,
-               phys_addr_t start, unsigned int len)
-{
-       nvdimm_bus_lock(&nvdimm_bus->dev);
-       __nvdimm_forget_poison(nvdimm_bus, start, len);
-       nvdimm_bus_unlock(&nvdimm_bus->dev);
+       spin_unlock(&nvdimm_bus->poison_lock);
 }
 EXPORT_SYMBOL_GPL(nvdimm_forget_poison);
 
index 8623e57..4c4bd20 100644 (file)
@@ -32,6 +32,7 @@ struct nvdimm_bus {
        struct list_head poison_list;
        struct list_head mapping_list;
        struct mutex reconfig_mutex;
+       spinlock_t poison_lock;
 };
 
 struct nvdimm {
index 1c609e8..98b2076 100644 (file)
@@ -122,8 +122,6 @@ static inline struct nd_blk_region_desc *to_blk_region_desc(
 int nvdimm_bus_add_poison(struct nvdimm_bus *nvdimm_bus, u64 addr, u64 length);
 void nvdimm_forget_poison(struct nvdimm_bus *nvdimm_bus,
                phys_addr_t start, unsigned int len);
-void __nvdimm_forget_poison(struct nvdimm_bus *nvdimm_bus,
-               phys_addr_t start, unsigned int len);
 struct nvdimm_bus *nvdimm_bus_register(struct device *parent,
                struct nvdimm_bus_descriptor *nfit_desc);
 void nvdimm_bus_unregister(struct nvdimm_bus *nvdimm_bus);