x86, pmem: fix broken __copy_user_nocache cache-bypass assumptions
authorDan Williams <dan.j.williams@intel.com>
Thu, 6 Apr 2017 16:04:31 +0000 (09:04 -0700)
committerDan Williams <dan.j.williams@intel.com>
Wed, 12 Apr 2017 20:45:18 +0000 (13:45 -0700)
Before we rework the "pmem api" to stop abusing __copy_user_nocache()
for memcpy_to_pmem() we need to fix cases where we may strand dirty data
in the cpu cache. The problem occurs when copy_from_iter_pmem() is used
for arbitrary data transfers from userspace. There is no guarantee that
these transfers, performed by dax_iomap_actor(), will have aligned
destinations or aligned transfer lengths. Backstop the usage
__copy_user_nocache() with explicit cache management in these unaligned
cases.

Yes, copy_from_iter_pmem() is now too big for an inline, but addressing
that is saved for a later patch that moves the entirety of the "pmem
api" into the pmem driver directly.

Fixes: 5de490daec8b ("pmem: add copy_from_iter_pmem() and clear_pmem()")
Cc: <stable@vger.kernel.org>
Cc: <x86@kernel.org>
Cc: Jan Kara <jack@suse.cz>
Cc: Jeff Moyer <jmoyer@redhat.com>
Cc: Ingo Molnar <mingo@redhat.com>
Cc: Christoph Hellwig <hch@lst.de>
Cc: "H. Peter Anvin" <hpa@zytor.com>
Cc: Al Viro <viro@zeniv.linux.org.uk>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Matthew Wilcox <mawilcox@microsoft.com>
Reviewed-by: Ross Zwisler <ross.zwisler@linux.intel.com>
Signed-off-by: Toshi Kani <toshi.kani@hpe.com>
Signed-off-by: Dan Williams <dan.j.williams@intel.com>
arch/x86/include/asm/pmem.h

index 2c1ebeb4d7376db6350b7266048a1d37d800ac86..529bb4a6487a9e42d07d0c4ceb226e41b5808f0f 100644 (file)
@@ -55,7 +55,8 @@ static inline int arch_memcpy_from_pmem(void *dst, const void *src, size_t n)
  * @size:      number of bytes to write back
  *
  * Write back a cache range using the CLWB (cache line write back)
- * instruction.
+ * instruction. Note that @size is internally rounded up to be cache
+ * line size aligned.
  */
 static inline void arch_wb_cache_pmem(void *addr, size_t size)
 {
@@ -69,15 +70,6 @@ static inline void arch_wb_cache_pmem(void *addr, size_t size)
                clwb(p);
 }
 
-/*
- * copy_from_iter_nocache() on x86 only uses non-temporal stores for iovec
- * iterators, so for other types (bvec & kvec) we must do a cache write-back.
- */
-static inline bool __iter_needs_pmem_wb(struct iov_iter *i)
-{
-       return iter_is_iovec(i) == false;
-}
-
 /**
  * arch_copy_from_iter_pmem - copy data from an iterator to PMEM
  * @addr:      PMEM destination address
@@ -94,7 +86,35 @@ static inline size_t arch_copy_from_iter_pmem(void *addr, size_t bytes,
        /* TODO: skip the write-back by always using non-temporal stores */
        len = copy_from_iter_nocache(addr, bytes, i);
 
-       if (__iter_needs_pmem_wb(i))
+       /*
+        * In the iovec case on x86_64 copy_from_iter_nocache() uses
+        * non-temporal stores for the bulk of the transfer, but we need
+        * to manually flush if the transfer is unaligned. A cached
+        * memory copy is used when destination or size is not naturally
+        * aligned. That is:
+        *   - Require 8-byte alignment when size is 8 bytes or larger.
+        *   - Require 4-byte alignment when size is 4 bytes.
+        *
+        * In the non-iovec case the entire destination needs to be
+        * flushed.
+        */
+       if (iter_is_iovec(i)) {
+               unsigned long flushed, dest = (unsigned long) addr;
+
+               if (bytes < 8) {
+                       if (!IS_ALIGNED(dest, 4) || (bytes != 4))
+                               arch_wb_cache_pmem(addr, 1);
+               } else {
+                       if (!IS_ALIGNED(dest, 8)) {
+                               dest = ALIGN(dest, boot_cpu_data.x86_clflush_size);
+                               arch_wb_cache_pmem(addr, 1);
+                       }
+
+                       flushed = dest - (unsigned long) addr;
+                       if (bytes > flushed && !IS_ALIGNED(bytes - flushed, 8))
+                               arch_wb_cache_pmem(addr + bytes - 1, 1);
+               }
+       } else
                arch_wb_cache_pmem(addr, bytes);
 
        return len;