io: block: fix block_read/write may read/write overlap buffer

The block operations were trying to optimize the number of memory
copies, and it tried to use directly the buffer supplied by the user
to them. This was a mistake because it created too many corner cases:

	1- It was possible to generate unaligned
	   operations to unaligned buffers. Drivers that were using
	   DMA transfer failed in that case.

	2- It was possible to generate read operations
	   with sizes that weren't a multiple of the block size. Some
	   low level drivers assumed that condition and they calculated
	   the number of blocks dividing the number of bytes by the
	   size of the block, without considering the remaining bytes.

	3- The block_* operations didn't control the
	   number of bytes actually copied to memory, because the
	   low level drivers were writing directly to the user buffer.

This patch rewrite block_read and block_write to use always the device
buffer, which the platform ensures that has the correct aligment and
the correct size.

Change-Id: I5e479bb7bc137e6ec205a8573eb250acd5f40420
Signed-off-by: Qixiang Xu <qixiang.xu@arm.com>
Signed-off-by: Roberto Vargas <roberto.vargas@arm.com>

1 files changed, 260 insertions, 140 deletions

diff --git a/drivers/io/io_block.c b/drivers/io/io_block.c
index 128246fe..8226554d 100644
--- a/drivers/io/io_block.c
+++ b/drivers/io/io_block.c
@@ -167,15 +167,98 @@ static int block_seek(io_entity_t *entity, int mode, ssize_t offset)
 	return 0;
 }
 
+/*
+ * This function allows the caller to read any number of bytes
+ * from any position. It hides from the caller that the low level
+ * driver only can read aligned blocks of data. For this reason
+ * we need to handle the use case where the first byte to be read is not
+ * aligned to start of the block, the last byte to be read is also not
+ * aligned to the end of a block, and there are zero or more blocks-worth
+ * of data in between.
+ *
+ * In such a case we need to read more bytes than requested (i.e. full
+ * blocks) and strip-out the leading bytes (aka skip) and the trailing
+ * bytes (aka padding). See diagram below
+ *
+ * cur->file_pos ------------
+ *                          |
+ * cur->base                |
+ *  |                       |
+ *  v                       v<----  length   ---->
+ *  --------------------------------------------------------------
+ * |           |         block#1    |        |   block#n          |
+ * |  block#0  |            +       |   ...  |     +              |
+ * |           | <- skip -> +       |        |     + <- padding ->|
+ *  ------------------------+----------------------+--------------
+ *             ^                                                  ^
+ *             |                                                  |
+ *             v    iteration#1                iteration#n        v
+ *              --------------------------------------------------
+ *             |                    |        |                    |
+ *             |<----  request ---->|  ...   |<----- request ---->|
+ *             |                    |        |                    |
+ *              --------------------------------------------------
+ *            /                   /          |                    |
+ *           /                   /           |                    |
+ *          /                   /            |                    |
+ *         /                   /             |                    |
+ *        /                   /              |                    |
+ *       /                   /               |                    |
+ *      /                   /                |                    |
+ *     /                   /                 |                    |
+ *    /                   /                  |                    |
+ *   /                   /                   |                    |
+ *  <---- request ------>                    <------ request  ----->
+ *  ---------------------                    -----------------------
+ *  |        |          |                    |          |           |
+ *  |<-skip->|<-nbytes->|           -------->|<-nbytes->|<-padding->|
+ *  |        |          |           |        |          |           |
+ *  ---------------------           |        -----------------------
+ *  ^        \           \          |        |          |
+ *  |         \           \         |        |          |
+ *  |          \           \        |        |          |
+ *  buf->offset \           \   buf->offset  |          |
+ *               \           \               |          |
+ *                \           \              |          |
+ *                 \           \             |          |
+ *                  \           \            |          |
+ *                   \           \           |          |
+ *                    \           \          |          |
+ *                     \           \         |          |
+ *                      --------------------------------
+ *                      |           |        |         |
+ * buffer-------------->|           | ...    |         |
+ *                      |           |        |         |
+ *                      --------------------------------
+ *                      <-count#1->|                   |
+ *                      <----------  count#n   -------->
+ *                      <----------  length  ---------->
+ *
+ * Additionally, the IO driver has an underlying buffer that is at least
+ * one block-size and may be big enough to allow.
+ */
 static int block_read(io_entity_t *entity, uintptr_t buffer, size_t length,
 		      size_t *length_read)
 {
 	block_dev_state_t *cur;
 	io_block_spec_t *buf;
 	io_block_ops_t *ops;
-	size_t aligned_length, skip, count, left, padding, block_size;
 	int lba;
-	int buffer_not_aligned;
+	size_t block_size, left;
+	size_t nbytes;  /* number of bytes read in one iteration */
+	size_t request; /* number of requested bytes in one iteration */
+	size_t count;   /* number of bytes already read */
+	/*
+	 * number of leading bytes from start of the block
+	 * to the first byte to be read
+	 */
+	size_t skip;
+
+	/*
+	 * number of trailing bytes between the last byte
+	 * to be read and the end of the block
+	 */
+	size_t padding;
 
 	assert(entity->info != (uintptr_t)NULL);
 	cur = (block_dev_state_t *)entity->info;
@@ -186,102 +269,107 @@ static int block_read(io_entity_t *entity, uintptr_t buffer, size_t length,
 	       (length > 0) &&
 	       (ops->read != 0));
 
-	if ((buffer & (block_size - 1)) != 0) {
+	/*
+	 * We don't know the number of bytes that we are going
+	 * to read in every iteration, because it will depend
+	 * on the low level driver.
+	 */
+	count = 0;
+	for (left = length; left > 0; left -= nbytes) {
 		/*
-		 * buffer isn't aligned with block size.
-		 * Block device always relies on DMA operation.
-		 * It's better to make the buffer as block size aligned.
+		 * We must only request operations aligned to the block
+		 * size. Therefore if file_pos is not block-aligned,
+		 * we have to request the operation to start at the
+		 * previous block boundary and skip the leading bytes. And
+		 * similarly, the number of bytes requested must be a
+		 * block size multiple
 		 */
-		buffer_not_aligned = 1;
-	} else {
-		buffer_not_aligned = 0;
-	}
+		skip = cur->file_pos & (block_size - 1);
 
-	skip = cur->file_pos % block_size;
-	aligned_length = ((skip + length) + (block_size - 1)) &
-			 ~(block_size - 1);
-	padding = aligned_length - (skip + length);
-	left = aligned_length;
-	do {
+		/*
+		 * Calculate the block number containing file_pos
+		 * - e.g. block 3.
+		 */
 		lba = (cur->file_pos + cur->base) / block_size;
-		if (left >= buf->length) {
+
+		if (skip + left > buf->length) {
 			/*
-			 * Since left is larger, it's impossible to padding.
-			 *
-			 * If buffer isn't aligned, we need to use aligned
-			 * buffer instead.
+			 * The underlying read buffer is too small to
+			 * read all the required data - limit to just
+			 * fill the buffer, and then read again.
 			 */
-			if (skip || buffer_not_aligned) {
-				/*
-				 * The beginning address (file_pos) isn't
-				 * aligned with block size, we need to use
-				 * block buffer to read block. Since block
-				 * device is always relied on DMA operation.
-				 */
-				count = ops->read(lba, buf->offset,
-						  buf->length);
-			} else {
-				count = ops->read(lba, buffer, buf->length);
-			}
-			assert(count == buf->length);
-			cur->file_pos += count - skip;
-			if (skip || buffer_not_aligned) {
-				/*
-				 * Since there's not aligned block size caused
-				 * by skip or not aligned buffer, block buffer
-				 * is used to store data.
-				 */
-				memcpy((void *)buffer,
-				       (void *)(buf->offset + skip),
-				       count - skip);
-			}
-			left = left - (count - skip);
+			request = buf->length;
 		} else {
-			if (skip || padding || buffer_not_aligned) {
-				/*
-				 * The beginning address (file_pos) isn't
-				 * aligned with block size, we have to read
-				 * full block by block buffer instead.
-				 * The size isn't aligned with block size.
-				 * Use block buffer to avoid overflow.
-				 *
-				 * If buffer isn't aligned, use block buffer
-				 * to avoid DMA error.
-				 */
-				count = ops->read(lba, buf->offset, left);
-			} else
-				count = ops->read(lba, buffer, left);
-			assert(count == left);
-			left = left - (skip + padding);
-			cur->file_pos += left;
-			if (skip || padding || buffer_not_aligned) {
-				/*
-				 * Since there's not aligned block size or
-				 * buffer, block buffer is used to store data.
-				 */
-				memcpy((void *)buffer,
-				       (void *)(buf->offset + skip),
-				       left);
-			}
-			/* It's already the last block operation */
-			left = 0;
+			/*
+			 * The underlying read buffer is big enough to
+			 * read all the required data. Calculate the
+			 * number of bytes to read to align with the
+			 * block size.
+			 */
+			request = skip + left;
+			request = (request + (block_size - 1)) & ~(block_size - 1);
 		}
-		skip = cur->file_pos % block_size;
-	} while (left > 0);
-	*length_read = length;
+		request = ops->read(lba, buf->offset, request);
+
+		if (request <= skip) {
+			/*
+			 * We couldn't read enough bytes to jump over
+			 * the skip bytes, so we should have to read
+			 * again the same block, thus generating
+			 * the same error.
+			 */
+			return -EIO;
+		}
+
+		/*
+		 * Need to remove skip and padding bytes,if any, from
+		 * the read data when copying to the user buffer.
+		 */
+		nbytes = request - skip;
+		padding = (nbytes > left) ? nbytes - left : 0;
+		nbytes -= padding;
+
+		memcpy((void *)(buffer + count),
+		       (void *)(buf->offset + skip),
+		       nbytes);
+
+		cur->file_pos += nbytes;
+		count += nbytes;
+	}
+	assert(count == length);
+	*length_read = count;
 
 	return 0;
 }
 
+/*
+ * This function allows the caller to write any number of bytes
+ * from any position. It hides from the caller that the low level
+ * driver only can write aligned blocks of data.
+ * See comments for block_read for more details.
+ */
 static int block_write(io_entity_t *entity, const uintptr_t buffer,
 		       size_t length, size_t *length_written)
 {
 	block_dev_state_t *cur;
 	io_block_spec_t *buf;
 	io_block_ops_t *ops;
-	size_t aligned_length, skip, count, left, padding, block_size;
 	int lba;
-	int buffer_not_aligned;
+	size_t block_size, left;
+	size_t nbytes;  /* number of bytes read in one iteration */
+	size_t request; /* number of requested bytes in one iteration */
+	size_t count;   /* number of bytes already read */
+	/*
+	 * number of leading bytes from start of the block
+	 * to the first byte to be read
+	 */
+	size_t skip;
+
+	/*
+	 * number of trailing bytes between the last byte
+	 * to be read and the end of the block
+	 */
+	size_t padding;
 
 	assert(entity->info != (uintptr_t)NULL);
 	cur = (block_dev_state_t *)entity->info;
@@ -293,75 +381,107 @@ static int block_write(io_entity_t *entity, const uintptr_t buffer,
 	       (ops->read != 0) &&
 	       (ops->write != 0));
 
-	if ((buffer & (block_size - 1)) != 0) {
+	/*
+	 * We don't know the number of bytes that we are going
+	 * to write in every iteration, because it will depend
+	 * on the low level driver.
+	 */
+	count = 0;
+	for (left = length; left > 0; left -= nbytes) {
 		/*
-		 * buffer isn't aligned with block size.
-		 * Block device always relies on DMA operation.
-		 * It's better to make the buffer as block size aligned.
+		 * We must only request operations aligned to the block
+		 * size. Therefore if file_pos is not block-aligned,
+		 * we have to request the operation to start at the
+		 * previous block boundary and skip the leading bytes. And
+		 * similarly, the number of bytes requested must be a
+		 * block size multiple
 		 */
-		buffer_not_aligned = 1;
-	} else {
-		buffer_not_aligned = 0;
-	}
+		skip = cur->file_pos & (block_size - 1);
 
-	skip = cur->file_pos % block_size;
-	aligned_length = ((skip + length) + (block_size - 1)) &
-			 ~(block_size - 1);
-	padding = aligned_length - (skip + length);
-	left = aligned_length;
-	do {
+		/*
+		 * Calculate the block number containing file_pos
+		 * - e.g. block 3.
+		 */
 		lba = (cur->file_pos + cur->base) / block_size;
-		if (left >= buf->length) {
-			/* Since left is larger, it's impossible to padding. */
-			if (skip || buffer_not_aligned) {
-				/*
-				 * The beginning address (file_pos) isn't
-				 * aligned with block size or buffer isn't
-				 * aligned, we need to use block buffer to
-				 * write block.
-				 */
-				count = ops->read(lba, buf->offset,
-						  buf->length);
-				assert(count == buf->length);
-				memcpy((void *)(buf->offset + skip),
-				       (void *)buffer,
-				       count - skip);
-				count = ops->write(lba, buf->offset,
-						   buf->length);
-			} else
-				count = ops->write(lba, buffer, buf->length);
-			assert(count == buf->length);
-			cur->file_pos += count - skip;
-			left = left - (count - skip);
+
+		if (skip + left > buf->length) {
+			/*
+			 * The underlying read buffer is too small to
+			 * read all the required data - limit to just
+			 * fill the buffer, and then read again.
+			 */
+			request = buf->length;
 		} else {
-			if (skip || padding || buffer_not_aligned) {
+			/*
+			 * The underlying read buffer is big enough to
+			 * read all the required data. Calculate the
+			 * number of bytes to read to align with the
+			 * block size.
+			 */
+			request = skip + left;
+			request = (request + (block_size - 1)) & ~(block_size - 1);
+		}
+
+		/*
+		 * The number of bytes that we are going to write
+		 * from the user buffer will depend of the size
+		 * of the current request.
+		 */
+		nbytes = request - skip;
+		padding = (nbytes > left) ? nbytes - left : 0;
+		nbytes -= padding;
+
+		/*
+		 * If we have skip or padding bytes then we have to preserve
+		 * some content and it means that we have to read before
+		 * writing
+		 */
+		if (skip > 0 || padding > 0) {
+			request = ops->read(lba, buf->offset, request);
+			/*
+			 * The read may return size less than
+			 * requested. Round down to the nearest block
+			 * boundary
+			 */
+			request &= ~(block_size-1);
+			if (request <= skip) {
 				/*
-				 * The beginning address (file_pos) isn't
-				 * aligned with block size, we need to avoid
-				 * poluate data in the beginning. Reading and
-				 * skipping the beginning is the only way.
-				 * The size isn't aligned with block size.
-				 * Use block buffer to avoid overflow.
-				 *
-				 * If buffer isn't aligned, use block buffer
-				 * to avoid DMA error.
+				 * We couldn't read enough bytes to jump over
+				 * the skip bytes, so we should have to read
+				 * again the same block, thus generating
+				 * the same error.
 				 */
-				count = ops->read(lba, buf->offset, left);
-				assert(count == left);
-				memcpy((void *)(buf->offset + skip),
-				       (void *)buffer,
-				       left - skip - padding);
-				count = ops->write(lba, buf->offset, left);
-			} else
-				count = ops->write(lba, buffer, left);
-			assert(count == left);
-			cur->file_pos += left - (skip + padding);
-			/* It's already the last block operation */
-			left = 0;
+				return -EIO;
+			}
+			nbytes = request - skip;
+			padding = (nbytes > left) ? nbytes - left : 0;
+			nbytes -= padding;
 		}
-		skip = cur->file_pos % block_size;
-	} while (left > 0);
-	*length_written = length;
+
+		memcpy((void *)(buf->offset + skip),
+		       (void *)(buffer + count),
+		       nbytes);
+
+		request = ops->write(lba, buf->offset, request);
+		if (request <= skip)
+			return -EIO;
+
+		/*
+		 * And the previous write operation may modify the size
+		 * of the request, so again, we have to calculate the
+		 * number of bytes that we consumed from the user
+		 * buffer
+		 */
+		nbytes = request - skip;
+		padding = (nbytes > left) ? nbytes - left : 0;
+		nbytes -= padding;
+
+		cur->file_pos += nbytes;
+		count += nbytes;
+	}
+	assert(count == length);
+	*length_written = count;
+
 	return 0;
 }