On 2023/08/18 12:08, Michael Schmitz wrote:
Hi Damien,
thanks for the review ...
Am 18.08.2023 um 12:51 schrieb Damien Le Moal:
On 2023/08/18 7:12, Michael Schmitz wrote:
Some users of pata_falcon on Q40 have IDE disks in default
IDE little endian byte order, whereas legacy disks use
host-native big-endian byte order as on the Atari Falcon.
Add module parameter 'data_swab' to allow connecting drives
with non-native data byte order. Drives selected by the
data_swap bit mask will have their user data byte-swapped to
host byte order, i.e. 'pata_falcon.data_swab=2' will byte-swap
all user data on drive B, leaving data on drive A in native
byte order. On Q40, drives on a second IDE interface may be
added to the bit mask as bits 2 and 3.
Default setting is no byte swapping, i.e. compatibility with
the native Falcon or Q40 operating system disk format.
Cc: William R Sowerbutts <will@xxxxxxxxxxxxxx>
Cc: Finn Thain <fthain@xxxxxxxxxxxxxx>
Cc: Geert Uytterhoeven <geert@xxxxxxxxxxxxxx>
Signed-off-by: Michael Schmitz <schmitzmic@xxxxxxxxx>
---
Changes since RFC v4:
Geert Uytterhoeven:
- don't shift static module parameter for drive 3/4 bitmask
- simplify bit mask calculation to always use pdev->id
Finn Thain:
- correct bit numbers for drive 3/4
Changes since RFC v3:
- split off this byte swap handling into separate patch
- add hint regarding third and fourth drive on Q40
Finn Thain:
- rename module parameter to 'data_swab' to better reflect its use
William Sowerbutts:
- correct IDE drive number used in data swap conditional
---
drivers/ata/pata_falcon.c | 26 +++++++++++++++++++++++++-
1 file changed, 25 insertions(+), 1 deletion(-)
diff --git a/drivers/ata/pata_falcon.c b/drivers/ata/pata_falcon.c
index 346259e3bbc8..90488f565d6f 100644
--- a/drivers/ata/pata_falcon.c
+++ b/drivers/ata/pata_falcon.c
@@ -33,6 +33,16 @@
#define DRV_NAME "pata_falcon"
#define DRV_VERSION "0.1.0"
+static int pata_falcon_swap_mask;
+
+module_param_named(data_swab, pata_falcon_swap_mask, int, 0444);
+MODULE_PARM_DESC(data_swab, "Data byte swap enable/disable bitmap (0x1==drive1, 0x2==drive2, 0x4==drive3, 0x8==drive4, default==0)");
+
+struct pata_falcon_priv {
+ unsigned int swap_mask;
+ bool swap_data;
+};
+
static const struct scsi_host_template pata_falcon_sht = {
ATA_PIO_SHT(DRV_NAME),
};
@@ -44,13 +54,15 @@ static unsigned int pata_falcon_data_xfer(struct ata_queued_cmd *qc,
struct ata_device *dev = qc->dev;
struct ata_port *ap = dev->link->ap;
void __iomem *data_addr = ap->ioaddr.data_addr;
+ struct pata_falcon_priv *priv = ap->private_data;
unsigned int words = buflen >> 1;
struct scsi_cmnd *cmd = qc->scsicmd;
+ int dev_id = dev->devno;
bool swap = 1;
if (dev->class == ATA_DEV_ATA && cmd &&
!blk_rq_is_passthrough(scsi_cmd_to_rq(cmd)))
- swap = 0;
+ swap = priv->swap_data && (priv->swap_mask & BIT(dev_id));
/* Transfer multiple of 2 bytes */
if (rw == READ) {
@@ -123,6 +135,7 @@ static int __init pata_falcon_init_one(struct platform_device *pdev)
struct resource *base_res, *ctl_res, *irq_res;
struct ata_host *host;
struct ata_port *ap;
+ struct pata_falcon_priv *priv;
void __iomem *base, *ctl_base;
int irq = 0, io_offset = 1, reg_scale = 4;
@@ -165,6 +178,13 @@ static int __init pata_falcon_init_one(struct platform_device *pdev)
ap->pio_mask = ATA_PIO4;
ap->flags |= ATA_FLAG_SLAVE_POSS | ATA_FLAG_NO_IORDY;
+ priv = devm_kzalloc(&pdev->dev,
+ sizeof(struct pata_falcon_priv), GFP_KERNEL);
+ if (!priv)
+ return -ENOMEM;
+
+ ap->private_data = priv;
+
/* N.B. this assumes data_addr will be used for word-sized I/O only */
ap->ioaddr.data_addr = (void __iomem *)base_mem_res->start;
@@ -199,6 +219,10 @@ static int __init pata_falcon_init_one(struct platform_device *pdev)
ap->ioaddr.altstatus_addr = ctl_base + io_offset;
ap->ioaddr.ctl_addr = ctl_base + io_offset;
+ priv->swap_mask = pata_falcon_swap_mask >> (2 * pdev->id);
I do not understand the shift here. It seems that it will lead to
priv->swap_mask always being 0...
On Q40, it is possible to have two ISA IDE adapters, and two platform
devices are defined (in arch/m68k/q40/config.c) One will have
pdev->id==0, the other will have pdev->id==1.
In the pdev->id==0 case, there's no shift of the bit mask passed in as
module parameter, so the data transfer function will examine bits 0 and
1. That case we've verified in testing.
In the other case, we shift down by two bits, so the data transfer
function will now examine bits 2 and 3 from the module parameter.
+ if (priv->swap_mask)
+ priv->swap_data = 1;
I do not understand why priv->swap_data is needed, given that it is set to 1 if
priv->swap_mask != 0, the above:
swap = priv->swap_data && (priv->swap_mask & BIT(dev_id));
should be equivalent to the simpler:
swap = priv->swap_mask & BIT(dev_id);
No ? Am I missing something ?
I had hoped to avoid the pointer dereference and bit shift in the
default (and by far most common) case where none of the bits are set.
Compared to the simpler version, I actually just save the bit shift, so
it's probably a pointless optimization.
Finn had suggested to simplify this even further, and use ap->private as
bit mask directly (saving the kzalloc()). If you're OK with that, I'll
change the code accordingly.
Sounds good. And given that we are talking about old IDE devices, which are
*really* slow, I would not worry too much about optimizing for pointer
dereference :)
Cheers,
Michael
+
irq_res = platform_get_resource(pdev, IORESOURCE_IRQ, 0);
if (irq_res && irq_res->start > 0) {
irq = irq_res->start;
--
Damien Le Moal
Western Digital Research
