summaryrefslogtreecommitdiff
path: root/doc
diff options
context:
space:
mode:
Diffstat (limited to 'doc')
-rw-r--r--doc/README.arm-unaligned-accesses122
-rw-r--r--doc/README.unaligned-memory-access.txt240
2 files changed, 240 insertions, 122 deletions
diff --git a/doc/README.arm-unaligned-accesses b/doc/README.arm-unaligned-accesses
deleted file mode 100644
index c37d135..0000000
--- a/doc/README.arm-unaligned-accesses
+++ /dev/null
@@ -1,122 +0,0 @@
-If you are reading this because of a data abort: the following MIGHT
-be relevant to your abort, if it was caused by an alignment violation.
-In order to determine this, use the PC from the abort dump along with
-an objdump -s -S of the u-boot ELF binary to locate the function where
-the abort happened; then compare this function with the examples below.
-If they match, then you've been hit with a compiler generated unaligned
-access, and you should rewrite your code or add -mno-unaligned-access
-to the command line of the offending file.
-
-Note that the PC shown in the abort message is relocated. In order to
-be able to match it to an address in the ELF binary dump, you will need
-to know the relocation offset. If your target defines CONFIG_CMD_BDI
-and if you can get to the prompt and enter commands before the abort
-happens, then command "bdinfo" will give you the offset. Otherwise you
-will need to try a build with DEBUG set, which will display the offset,
-or use a debugger and set a breakpoint at relocate_code() to see the
-offset (passed as an argument).
-
-*
-
-Since U-Boot runs on a variety of hardware, some only able to perform
-unaligned accesses with a strong penalty, some unable to perform them
-at all, the policy regarding unaligned accesses is to not perform any,
-unless absolutely necessary because of hardware or standards.
-
-Also, on hardware which permits it, the core is configured to throw
-data abort exceptions on unaligned accesses in order to catch these
-unallowed accesses as early as possible.
-
-Until version 4.7, the gcc default for performing unaligned accesses
-(-mno-unaligned-access) is to emulate unaligned accesses using aligned
-loads and stores plus shifts and masks. Emulated unaligned accesses
-will not be caught by hardware. These accesses may be costly and may
-be actually unnecessary. In order to catch these accesses and remove
-or optimize them, option -munaligned-access is explicitly set for all
-versions of gcc which support it.
-
-From gcc 4.7 onward starting at armv7 architectures, the default for
-performing unaligned accesses is to use unaligned native loads and
-stores (-munaligned-access), because the cost of unaligned accesses
-has dropped on armv7 and beyond. This should not affect U-Boot's
-policy of controlling unaligned accesses, however the compiler may
-generate uncontrolled unaligned accesses on its own in at least one
-known case: when declaring a local initialized char array, e.g.
-
-function foo()
-{
- char buffer[] = "initial value";
-/* or */
- char buffer[] = { 'i', 'n', 'i', 't', 0 };
- ...
-}
-
-Under -munaligned-accesses with optimizations on, this declaration
-causes the compiler to generate native loads from the literal string
-and native stores to the buffer, and the literal string alignment
-cannot be controlled. If it is misaligned, then the core will throw
-a data abort exception.
-
-Quite probably the same might happen for 16-bit array initializations
-where the constant is aligned on a boundary which is a multiple of 2
-but not of 4:
-
-function foo()
-{
- u16 buffer[] = { 1, 2, 3 };
- ...
-}
-
-The long term solution to this issue is to add an option to gcc to
-allow controlling the general alignment of data, including constant
-initialization values.
-
-However this will only apply to the version of gcc which will have such
-an option. For other versions, there are four workarounds:
-
-a) Enforce as a rule that array initializations as described above
- are forbidden. This is generally not acceptable as they are valid,
- and usual, C constructs. The only case where they could be rejected
- is when they actually equate to a const char* declaration, i.e. the
- array is initialized and never modified in the function's scope.
-
-b) Drop the requirement on unaligned accesses at least for ARMv7,
- i.e. do not throw a data abort exception upon unaligned accesses.
- But that will allow adding badly aligned code to U-Boot, only for
- it to fail when re-used with a stricter target, possibly once the
- bad code is already in mainline.
-
-c) Relax the -munaligned-access rule globally. This will prevent native
- unaligned accesses of course, but that will also hide any bug caused
- by a bad unaligned access, making it much harder to diagnose it. It
- is actually what already happens when building ARM targets with a
- pre-4.7 gcc, and it may actually already hide some bugs yet unseen
- until the target gets compiled with -munaligned-access.
-
-d) Relax the -munaligned-access rule only for for files susceptible to
- the local initialized array issue and for armv7 architectures and
- beyond. This minimizes the quantity of code which can hide unwanted
- misaligned accesses.
-
-The option retained is d).
-
-Considering that actual occurrences of the issue are rare (as of this
-writing, 5 files out of 7840 in U-Boot, or .3%, contain an initialized
-local char array which cannot actually be replaced with a const char*),
-contributors should not be required to systematically try and detect
-the issue in their patches.
-
-Detecting files susceptible to the issue can be automated through a
-filter installed as a hook in .git which recognizes local char array
-initializations. Automation should err on the false positive side, for
-instance flagging non-local arrays as if they were local if they cannot
-be told apart.
-
-In any case, detection shall not prevent committing the patch, but
-shall pre-populate the commit message with a note to the effect that
-this patch contains an initialized local char or 16-bit array and thus
-should be protected from the gcc 4.7 issue.
-
-Upon a positive detection, either $(PLATFORM_NO_UNALIGNED) should be
-added to CFLAGS for the affected file(s), or if the array is a pseudo
-const char*, it should be replaced by an actual one.
diff --git a/doc/README.unaligned-memory-access.txt b/doc/README.unaligned-memory-access.txt
new file mode 100644
index 0000000..00529f5
--- /dev/null
+++ b/doc/README.unaligned-memory-access.txt
@@ -0,0 +1,240 @@
+Editors note: This document is _heavily_ cribbed from the Linux Kernel, with
+really only the section about "Alignment vs. Networking" removed.
+
+UNALIGNED MEMORY ACCESSES
+=========================
+
+Linux runs on a wide variety of architectures which have varying behaviour
+when it comes to memory access. This document presents some details about
+unaligned accesses, why you need to write code that doesn't cause them,
+and how to write such code!
+
+
+The definition of an unaligned access
+=====================================
+
+Unaligned memory accesses occur when you try to read N bytes of data starting
+from an address that is not evenly divisible by N (i.e. addr % N != 0).
+For example, reading 4 bytes of data from address 0x10004 is fine, but
+reading 4 bytes of data from address 0x10005 would be an unaligned memory
+access.
+
+The above may seem a little vague, as memory access can happen in different
+ways. The context here is at the machine code level: certain instructions read
+or write a number of bytes to or from memory (e.g. movb, movw, movl in x86
+assembly). As will become clear, it is relatively easy to spot C statements
+which will compile to multiple-byte memory access instructions, namely when
+dealing with types such as u16, u32 and u64.
+
+
+Natural alignment
+=================
+
+The rule mentioned above forms what we refer to as natural alignment:
+When accessing N bytes of memory, the base memory address must be evenly
+divisible by N, i.e. addr % N == 0.
+
+When writing code, assume the target architecture has natural alignment
+requirements.
+
+In reality, only a few architectures require natural alignment on all sizes
+of memory access. However, we must consider ALL supported architectures;
+writing code that satisfies natural alignment requirements is the easiest way
+to achieve full portability.
+
+
+Why unaligned access is bad
+===========================
+
+The effects of performing an unaligned memory access vary from architecture
+to architecture. It would be easy to write a whole document on the differences
+here; a summary of the common scenarios is presented below:
+
+ - Some architectures are able to perform unaligned memory accesses
+ transparently, but there is usually a significant performance cost.
+ - Some architectures raise processor exceptions when unaligned accesses
+ happen. The exception handler is able to correct the unaligned access,
+ at significant cost to performance.
+ - Some architectures raise processor exceptions when unaligned accesses
+ happen, but the exceptions do not contain enough information for the
+ unaligned access to be corrected.
+ - Some architectures are not capable of unaligned memory access, but will
+ silently perform a different memory access to the one that was requested,
+ resulting in a subtle code bug that is hard to detect!
+
+It should be obvious from the above that if your code causes unaligned
+memory accesses to happen, your code will not work correctly on certain
+platforms and will cause performance problems on others.
+
+
+Code that does not cause unaligned access
+=========================================
+
+At first, the concepts above may seem a little hard to relate to actual
+coding practice. After all, you don't have a great deal of control over
+memory addresses of certain variables, etc.
+
+Fortunately things are not too complex, as in most cases, the compiler
+ensures that things will work for you. For example, take the following
+structure:
+
+ struct foo {
+ u16 field1;
+ u32 field2;
+ u8 field3;
+ };
+
+Let us assume that an instance of the above structure resides in memory
+starting at address 0x10000. With a basic level of understanding, it would
+not be unreasonable to expect that accessing field2 would cause an unaligned
+access. You'd be expecting field2 to be located at offset 2 bytes into the
+structure, i.e. address 0x10002, but that address is not evenly divisible
+by 4 (remember, we're reading a 4 byte value here).
+
+Fortunately, the compiler understands the alignment constraints, so in the
+above case it would insert 2 bytes of padding in between field1 and field2.
+Therefore, for standard structure types you can always rely on the compiler
+to pad structures so that accesses to fields are suitably aligned (assuming
+you do not cast the field to a type of different length).
+
+Similarly, you can also rely on the compiler to align variables and function
+parameters to a naturally aligned scheme, based on the size of the type of
+the variable.
+
+At this point, it should be clear that accessing a single byte (u8 or char)
+will never cause an unaligned access, because all memory addresses are evenly
+divisible by one.
+
+On a related topic, with the above considerations in mind you may observe
+that you could reorder the fields in the structure in order to place fields
+where padding would otherwise be inserted, and hence reduce the overall
+resident memory size of structure instances. The optimal layout of the
+above example is:
+
+ struct foo {
+ u32 field2;
+ u16 field1;
+ u8 field3;
+ };
+
+For a natural alignment scheme, the compiler would only have to add a single
+byte of padding at the end of the structure. This padding is added in order
+to satisfy alignment constraints for arrays of these structures.
+
+Another point worth mentioning is the use of __attribute__((packed)) on a
+structure type. This GCC-specific attribute tells the compiler never to
+insert any padding within structures, useful when you want to use a C struct
+to represent some data that comes in a fixed arrangement 'off the wire'.
+
+You might be inclined to believe that usage of this attribute can easily
+lead to unaligned accesses when accessing fields that do not satisfy
+architectural alignment requirements. However, again, the compiler is aware
+of the alignment constraints and will generate extra instructions to perform
+the memory access in a way that does not cause unaligned access. Of course,
+the extra instructions obviously cause a loss in performance compared to the
+non-packed case, so the packed attribute should only be used when avoiding
+structure padding is of importance.
+
+
+Code that causes unaligned access
+=================================
+
+With the above in mind, let's move onto a real life example of a function
+that can cause an unaligned memory access. The following function taken
+from the Linux Kernel's include/linux/etherdevice.h is an optimized routine
+to compare two ethernet MAC addresses for equality.
+
+bool ether_addr_equal(const u8 *addr1, const u8 *addr2)
+{
+#ifdef CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS
+ u32 fold = ((*(const u32 *)addr1) ^ (*(const u32 *)addr2)) |
+ ((*(const u16 *)(addr1 + 4)) ^ (*(const u16 *)(addr2 + 4)));
+
+ return fold == 0;
+#else
+ const u16 *a = (const u16 *)addr1;
+ const u16 *b = (const u16 *)addr2;
+ return ((a[0] ^ b[0]) | (a[1] ^ b[1]) | (a[2] ^ b[2])) != 0;
+#endif
+}
+
+In the above function, when the hardware has efficient unaligned access
+capability, there is no issue with this code. But when the hardware isn't
+able to access memory on arbitrary boundaries, the reference to a[0] causes
+2 bytes (16 bits) to be read from memory starting at address addr1.
+
+Think about what would happen if addr1 was an odd address such as 0x10003.
+(Hint: it'd be an unaligned access.)
+
+Despite the potential unaligned access problems with the above function, it
+is included in the kernel anyway but is understood to only work normally on
+16-bit-aligned addresses. It is up to the caller to ensure this alignment or
+not use this function at all. This alignment-unsafe function is still useful
+as it is a decent optimization for the cases when you can ensure alignment,
+which is true almost all of the time in ethernet networking context.
+
+
+Here is another example of some code that could cause unaligned accesses:
+ void myfunc(u8 *data, u32 value)
+ {
+ [...]
+ *((u32 *) data) = cpu_to_le32(value);
+ [...]
+ }
+
+This code will cause unaligned accesses every time the data parameter points
+to an address that is not evenly divisible by 4.
+
+In summary, the 2 main scenarios where you may run into unaligned access
+problems involve:
+ 1. Casting variables to types of different lengths
+ 2. Pointer arithmetic followed by access to at least 2 bytes of data
+
+
+Avoiding unaligned accesses
+===========================
+
+The easiest way to avoid unaligned access is to use the get_unaligned() and
+put_unaligned() macros provided by the <asm/unaligned.h> header file.
+
+Going back to an earlier example of code that potentially causes unaligned
+access:
+
+ void myfunc(u8 *data, u32 value)
+ {
+ [...]
+ *((u32 *) data) = cpu_to_le32(value);
+ [...]
+ }
+
+To avoid the unaligned memory access, you would rewrite it as follows:
+
+ void myfunc(u8 *data, u32 value)
+ {
+ [...]
+ value = cpu_to_le32(value);
+ put_unaligned(value, (u32 *) data);
+ [...]
+ }
+
+The get_unaligned() macro works similarly. Assuming 'data' is a pointer to
+memory and you wish to avoid unaligned access, its usage is as follows:
+
+ u32 value = get_unaligned((u32 *) data);
+
+These macros work for memory accesses of any length (not just 32 bits as
+in the examples above). Be aware that when compared to standard access of
+aligned memory, using these macros to access unaligned memory can be costly in
+terms of performance.
+
+If use of such macros is not convenient, another option is to use memcpy(),
+where the source or destination (or both) are of type u8* or unsigned char*.
+Due to the byte-wise nature of this operation, unaligned accesses are avoided.
+
+--
+In the Linux Kernel,
+Authors: Daniel Drake <dsd@gentoo.org>,
+ Johannes Berg <johannes@sipsolutions.net>
+With help from: Alan Cox, Avuton Olrich, Heikki Orsila, Jan Engelhardt,
+Kyle McMartin, Kyle Moffett, Randy Dunlap, Robert Hancock, Uli Kunitz,
+Vadim Lobanov