[PATCH] libata: fix ata_pio_sector for CONFIG_HIGHMEM

[PATCH] libata: fix ata_pio_sector for CONFIG_HIGHMEM

[Christoph Hellwig]

Data transfers are not required to be block aligned in memory, so they
span two pages.  Fix this by splitting the call to >sff_data_xfer into
two for that case.

This has been broken since the initial libata import before the damn
of git, but was uncovered by the legacy ide driver removal.

Reported-by: kernel test robot <oliver.sang@intel.com>
Signed-off-by: Christoph Hellwig <hch@lst.de>
---
 drivers/ata/libata-sff.c | 35 +++++++++++++++++++++++++++--------
 1 file changed, 27 insertions(+), 8 deletions(-)

diff --git a/drivers/ata/libata-sff.c b/drivers/ata/libata-sff.c
index ae7189d1a568..b71ea4a680b0 100644
--- a/drivers/ata/libata-sff.c
+++ b/drivers/ata/libata-sff.c
@@ -637,6 +637,20 @@ unsigned int ata_sff_data_xfer32(struct ata_queued_cmd *qc, unsigned char *buf,
 }
 EXPORT_SYMBOL_GPL(ata_sff_data_xfer32);
 
+static void ata_pio_xfer(struct ata_queued_cmd *qc, struct page *page,
+		unsigned int offset, size_t xfer_size)
+{
+	bool do_write = (qc->tf.flags & ATA_TFLAG_WRITE);
+	unsigned char *buf;
+
+	buf = kmap_atomic(page);
+	qc->ap->ops->sff_data_xfer(qc, buf + offset, xfer_size, do_write);
+	kunmap_atomic(buf);
+
+	if (!do_write && !PageSlab(page))
+		flush_dcache_page(page);
+}
+
 /**
  *	ata_pio_sector - Transfer a sector of data.
  *	@qc: Command on going
@@ -648,11 +662,9 @@ EXPORT_SYMBOL_GPL(ata_sff_data_xfer32);
  */
 static void ata_pio_sector(struct ata_queued_cmd *qc)
 {
-	int do_write = (qc->tf.flags & ATA_TFLAG_WRITE);
 	struct ata_port *ap = qc->ap;
 	struct page *page;
 	unsigned int offset;
-	unsigned char *buf;
 
 	if (!qc->cursg) {
 		qc->curbytes = qc->nbytes;
@@ -670,13 +682,20 @@ static void ata_pio_sector(struct ata_queued_cmd *qc)
 
 	DPRINTK("data %s\n", qc->tf.flags & ATA_TFLAG_WRITE ? "write" : "read");
 
-	/* do the actual data transfer */
-	buf = kmap_atomic(page);
-	ap->ops->sff_data_xfer(qc, buf + offset, qc->sect_size, do_write);
-	kunmap_atomic(buf);
+	/*
+	 * Split the transfer when it splits a page boundary.  Note that the
+	 * split still has to be dword aligned like all ATA data transfers.
+	 */
+	WARN_ON_ONCE(offset % 4);
+	if (offset + qc->sect_size > PAGE_SIZE) {
+		unsigned int split_len = PAGE_SIZE - offset;
 
-	if (!do_write && !PageSlab(page))
-		flush_dcache_page(page);
+		ata_pio_xfer(qc, page, offset, split_len);
+		ata_pio_xfer(qc, nth_page(page, 1), 0,
+			     qc->sect_size - split_len);
+	} else {
+		ata_pio_xfer(qc, page, offset, qc->sect_size);
+	}
 
 	qc->curbytes += qc->sect_size;
 	qc->cursg_ofs += qc->sect_size;
-- 
2.30.2

Re: [PATCH] libata: fix ata_pio_sector for CONFIG_HIGHMEM

[Jens Axboe]

On 7/9/21 7:02 AM, Christoph Hellwig wrote:
> Data transfers are not required to be block aligned in memory, so they
> span two pages.  Fix this by splitting the call to >sff_data_xfer into
> two for that case.
> 
> This has been broken since the initial libata import before the damn
> of git, but was uncovered by the legacy ide driver removal.

Applied, thanks.

-- 
Jens Axboe

Re: [PATCH] libata: fix ata_pio_sector for CONFIG_HIGHMEM

[Sergei Shtylyov]

Hello!

On 7/9/21 4:02 PM, Christoph Hellwig wrote:

> Data transfers are not required to be block aligned in memory, so they

> span two pages.  Fix this by splitting the call to >sff_data_xfer into
> two for that case.
> 
> This has been broken since the initial libata import before the damn
> of git, but was uncovered by the legacy ide driver removal.

   Damn of git, hehe... :-)

> Reported-by: kernel test robot <oliver.sang@intel.com>
> Signed-off-by: Christoph Hellwig <hch@lst.de>
[...]

MBR, Sergei