kernel_optimize_test/drivers/ata/pata_ali.c

653 lines
18 KiB
C
Raw Normal View History

/*
* pata_ali.c - ALI 15x3 PATA for new ATA layer
* (C) 2005 Red Hat Inc
*
* based in part upon
* linux/drivers/ide/pci/alim15x3.c Version 0.17 2003/01/02
*
* Copyright (C) 1998-2000 Michel Aubry, Maintainer
* Copyright (C) 1998-2000 Andrzej Krzysztofowicz, Maintainer
* Copyright (C) 1999-2000 CJ, cjtsai@ali.com.tw, Maintainer
*
* Copyright (C) 1998-2000 Andre Hedrick (andre@linux-ide.org)
* May be copied or modified under the terms of the GNU General Public License
* Copyright (C) 2002 Alan Cox <alan@redhat.com>
* ALi (now ULi M5228) support by Clear Zhang <Clear.Zhang@ali.com.tw>
*
* Documentation
* Chipset documentation available under NDA only
*
* TODO/CHECK
* Cannot have ATAPI on both master & slave for rev < c2 (???) but
* otherwise should do atapi DMA (For now for old we do PIO only for
* ATAPI)
* Review Sunblade workaround.
*/
#include <linux/kernel.h>
#include <linux/module.h>
#include <linux/pci.h>
#include <linux/init.h>
#include <linux/blkdev.h>
#include <linux/delay.h>
#include <scsi/scsi_host.h>
#include <linux/libata.h>
#include <linux/dmi.h>
#define DRV_NAME "pata_ali"
#define DRV_VERSION "0.7.8"
static int ali_atapi_dma = 0;
module_param_named(atapi_dma, ali_atapi_dma, int, 0644);
MODULE_PARM_DESC(atapi_dma, "Enable ATAPI DMA (0=disable, 1=enable)");
static struct pci_dev *ali_isa_bridge;
/*
* Cable special cases
*/
static const struct dmi_system_id cable_dmi_table[] = {
{
.ident = "HP Pavilion N5430",
.matches = {
DMI_MATCH(DMI_BOARD_VENDOR, "Hewlett-Packard"),
DMI_MATCH(DMI_BOARD_VERSION, "OmniBook N32N-736"),
},
},
{
.ident = "Toshiba Satellite S1800-814",
.matches = {
DMI_MATCH(DMI_SYS_VENDOR, "TOSHIBA"),
DMI_MATCH(DMI_PRODUCT_NAME, "S1800-814"),
},
},
{ }
};
static int ali_cable_override(struct pci_dev *pdev)
{
/* Fujitsu P2000 */
if (pdev->subsystem_vendor == 0x10CF && pdev->subsystem_device == 0x10AF)
return 1;
/* Mitac 8317 (Winbook-A) and relatives */
if (pdev->subsystem_vendor == 0x1071 && pdev->subsystem_device == 0x8317)
return 1;
/* Systems by DMI */
if (dmi_check_system(cable_dmi_table))
return 1;
return 0;
}
/**
* ali_c2_cable_detect - cable detection
* @ap: ATA port
*
* Perform cable detection for C2 and later revisions
*/
static int ali_c2_cable_detect(struct ata_port *ap)
{
struct pci_dev *pdev = to_pci_dev(ap->host->dev);
u8 ata66;
/* Certain laptops use short but suitable cables and don't
implement the detect logic */
if (ali_cable_override(pdev))
return ATA_CBL_PATA40_SHORT;
/* Host view cable detect 0x4A bit 0 primary bit 1 secondary
Bit set for 40 pin */
pci_read_config_byte(pdev, 0x4A, &ata66);
if (ata66 & (1 << ap->port_no))
return ATA_CBL_PATA40;
else
return ATA_CBL_PATA80;
}
/**
* ali_20_filter - filter for earlier ALI DMA
* @ap: ALi ATA port
* @adev: attached device
*
* Ensure that we do not do DMA on CD devices. We may be able to
* fix that later on. Also ensure we do not do UDMA on WDC drives
*/
static unsigned long ali_20_filter(struct ata_device *adev, unsigned long mask)
{
char model_num[ATA_ID_PROD_LEN + 1];
/* No DMA on anything but a disk for now */
if (adev->class != ATA_DEV_ATA)
mask &= ~(ATA_MASK_MWDMA | ATA_MASK_UDMA);
ata_id_c_string(adev->id, model_num, ATA_ID_PROD, sizeof(model_num));
if (strstr(model_num, "WDC"))
return mask &= ~ATA_MASK_UDMA;
libata-sff: clean up BMDMA initialization When BMDMA initialization failed or BMDMA was not available for whatever reason, bmdma_addr was left at zero and used as an indication that BMDMA shouldn't be used. This leads to the following problems. p1. For BMDMA drivers which don't use traditional BMDMA register, ata_bmdma_mode_filter() incorrectly inhibits DMA modes. Those drivers either have to inherit from ata_sff_port_ops or clear ->mode_filter explicitly. p2. non-BMDMA drivers call into BMDMA PRD table allocation. It doesn't actually allocate PRD table if bmdma_addr is not initialized but is still confusing. p3. For BMDMA drivers which don't use traditional BMDMA register, some methods might not be invoked as expected (e.g. bmdma_stop from ata_sff_post_internal_cmd()). p4. SFF drivers w/ custom DMA interface implement noop BMDMA ops worrying libata core might call into one of them. These problems are caused by the muddy line between SFF and BMDMA and the assumption that all BMDMA controllers initialize bmdma_addr. This patch fixes p1 and p2 by removing the bmdma_addr assumption and moving prd allocation to BMDMA port start. Later patches will fix the remaining issues. This patch improves BMDMA initialization such that * When BMDMA register initialization fails, falls back to PIO instead of failing. ata_pci_bmdma_init() never fails now. * When ata_pci_bmdma_init() falls back to PIO, it clears ap->mwdma_mask and udma_mask instead of depending on ata_bmdma_mode_filter(). This makes ata_bmdma_mode_filter() unnecessary thus resolving p1. * ata_port_start() which actually is BMDMA specific is moved to ata_bmdma_port_start(). ata_port_start() and ata_sff_port_start() are killed. * ata_sff_port_start32() is moved and renamed to ata_bmdma_port_start32(). Drivers which no longer call into PRD table allocation are... pdc_adma, sata_inic162x, sata_qstor, sata_sx4, pata_cmd640 and all drivers which inherit from ata_sff_port_ops. pata_icside sets ->port_start to ATA_OP_NULL as it doesn't need PRD but is a BMDMA controller and doesn't have custom port_start like other such controllers. Note that with the previous patch which makes all and only BMDMA drivers inherit from ata_bmdma_port_ops, this change doesn't break drivers which need PRD table. Signed-off-by: Tejun Heo <tj@kernel.org> Signed-off-by: Jeff Garzik <jgarzik@redhat.com>
2010-05-11 03:41:34 +08:00
return mask;
}
/**
* ali_fifo_control - FIFO manager
* @ap: ALi channel to control
* @adev: device for FIFO control
* @on: 0 for off 1 for on
*
* Enable or disable the FIFO on a given device. Because of the way the
* ALi FIFO works it provides a boost on ATA disk but can be confused by
* ATAPI and we must therefore manage it.
*/
static void ali_fifo_control(struct ata_port *ap, struct ata_device *adev, int on)
{
struct pci_dev *pdev = to_pci_dev(ap->host->dev);
int pio_fifo = 0x54 + ap->port_no;
u8 fifo;
int shift = 4 * adev->devno;
/* ATA - FIFO on set nibble to 0x05, ATAPI - FIFO off, set nibble to
0x00. Not all the docs agree but the behaviour we now use is the
one stated in the BIOS Programming Guide */
pci_read_config_byte(pdev, pio_fifo, &fifo);
fifo &= ~(0x0F << shift);
fifo |= (on << shift);
pci_write_config_byte(pdev, pio_fifo, fifo);
}
/**
* ali_program_modes - load mode registers
* @ap: ALi channel to load
* @adev: Device the timing is for
* @t: timing data
* @ultra: UDMA timing or zero for off
*
* Loads the timing registers for cmd/data and disable UDMA if
* ultra is zero. If ultra is set then load and enable the UDMA
* timing but do not touch the command/data timing.
*/
static void ali_program_modes(struct ata_port *ap, struct ata_device *adev, struct ata_timing *t, u8 ultra)
{
struct pci_dev *pdev = to_pci_dev(ap->host->dev);
int cas = 0x58 + 4 * ap->port_no; /* Command timing */
int cbt = 0x59 + 4 * ap->port_no; /* Command timing */
int drwt = 0x5A + 4 * ap->port_no + adev->devno; /* R/W timing */
int udmat = 0x56 + ap->port_no; /* UDMA timing */
int shift = 4 * adev->devno;
u8 udma;
if (t != NULL) {
t->setup = clamp_val(t->setup, 1, 8) & 7;
t->act8b = clamp_val(t->act8b, 1, 8) & 7;
t->rec8b = clamp_val(t->rec8b, 1, 16) & 15;
t->active = clamp_val(t->active, 1, 8) & 7;
t->recover = clamp_val(t->recover, 1, 16) & 15;
pci_write_config_byte(pdev, cas, t->setup);
pci_write_config_byte(pdev, cbt, (t->act8b << 4) | t->rec8b);
pci_write_config_byte(pdev, drwt, (t->active << 4) | t->recover);
}
/* Set up the UDMA enable */
pci_read_config_byte(pdev, udmat, &udma);
udma &= ~(0x0F << shift);
udma |= ultra << shift;
pci_write_config_byte(pdev, udmat, udma);
}
/**
* ali_set_piomode - set initial PIO mode data
* @ap: ATA interface
* @adev: ATA device
*
* Program the ALi registers for PIO mode.
*/
static void ali_set_piomode(struct ata_port *ap, struct ata_device *adev)
{
struct ata_device *pair = ata_dev_pair(adev);
struct ata_timing t;
unsigned long T = 1000000000 / 33333; /* PCI clock based */
ata_timing_compute(adev, adev->pio_mode, &t, T, 1);
if (pair) {
struct ata_timing p;
ata_timing_compute(pair, pair->pio_mode, &p, T, 1);
ata_timing_merge(&p, &t, &t, ATA_TIMING_SETUP|ATA_TIMING_8BIT);
if (pair->dma_mode) {
ata_timing_compute(pair, pair->dma_mode, &p, T, 1);
ata_timing_merge(&p, &t, &t, ATA_TIMING_SETUP|ATA_TIMING_8BIT);
}
}
/* PIO FIFO is only permitted on ATA disk */
if (adev->class != ATA_DEV_ATA)
ali_fifo_control(ap, adev, 0x00);
ali_program_modes(ap, adev, &t, 0);
if (adev->class == ATA_DEV_ATA)
ali_fifo_control(ap, adev, 0x05);
}
/**
* ali_set_dmamode - set initial DMA mode data
* @ap: ATA interface
* @adev: ATA device
*
* Program the ALi registers for DMA mode.
*/
static void ali_set_dmamode(struct ata_port *ap, struct ata_device *adev)
{
static u8 udma_timing[7] = { 0xC, 0xB, 0xA, 0x9, 0x8, 0xF, 0xD };
struct ata_device *pair = ata_dev_pair(adev);
struct ata_timing t;
unsigned long T = 1000000000 / 33333; /* PCI clock based */
struct pci_dev *pdev = to_pci_dev(ap->host->dev);
if (adev->class == ATA_DEV_ATA)
ali_fifo_control(ap, adev, 0x08);
if (adev->dma_mode >= XFER_UDMA_0) {
ali_program_modes(ap, adev, NULL, udma_timing[adev->dma_mode - XFER_UDMA_0]);
if (adev->dma_mode >= XFER_UDMA_3) {
u8 reg4b;
pci_read_config_byte(pdev, 0x4B, &reg4b);
reg4b |= 1;
pci_write_config_byte(pdev, 0x4B, reg4b);
}
} else {
ata_timing_compute(adev, adev->dma_mode, &t, T, 1);
if (pair) {
struct ata_timing p;
ata_timing_compute(pair, pair->pio_mode, &p, T, 1);
ata_timing_merge(&p, &t, &t, ATA_TIMING_SETUP|ATA_TIMING_8BIT);
if (pair->dma_mode) {
ata_timing_compute(pair, pair->dma_mode, &p, T, 1);
ata_timing_merge(&p, &t, &t, ATA_TIMING_SETUP|ATA_TIMING_8BIT);
}
}
ali_program_modes(ap, adev, &t, 0);
}
}
/**
* ali_warn_atapi_dma - Warn about ATAPI DMA disablement
* @adev: Device
*
* Whine about ATAPI DMA disablement if @adev is an ATAPI device.
* Can be used as ->dev_config.
*/
static void ali_warn_atapi_dma(struct ata_device *adev)
{
struct ata_eh_context *ehc = &adev->link->eh_context;
int print_info = ehc->i.flags & ATA_EHI_PRINTINFO;
if (print_info && adev->class == ATA_DEV_ATAPI && !ali_atapi_dma) {
ata_dev_warn(adev,
"WARNING: ATAPI DMA disabled for reliability issues. It can be enabled\n");
ata_dev_warn(adev,
"WARNING: via pata_ali.atapi_dma modparam or corresponding sysfs node.\n");
}
}
/**
* ali_lock_sectors - Keep older devices to 255 sector mode
* @adev: Device
*
* Called during the bus probe for each device that is found. We use
* this call to lock the sector count of the device to 255 or less on
* older ALi controllers. If we didn't do this then large I/O's would
* require LBA48 commands which the older ALi requires are issued by
* slower PIO methods
*/
static void ali_lock_sectors(struct ata_device *adev)
{
adev->max_sectors = 255;
ali_warn_atapi_dma(adev);
}
/**
* ali_check_atapi_dma - DMA check for most ALi controllers
* @adev: Device
*
* Called to decide whether commands should be sent by DMA or PIO
*/
static int ali_check_atapi_dma(struct ata_queued_cmd *qc)
{
if (!ali_atapi_dma) {
/* FIXME: pata_ali can't do ATAPI DMA reliably but the
* IDE alim15x3 driver can. I tried lots of things
* but couldn't find what the actual difference was.
* If you got an idea, please write it to
* linux-ide@vger.kernel.org and cc htejun@gmail.com.
*
* Disable ATAPI DMA for now.
*/
return -EOPNOTSUPP;
}
/* If its not a media command, its not worth it */
if (atapi_cmd_type(qc->cdb[0]) == ATAPI_MISC)
return -EOPNOTSUPP;
return 0;
}
static void ali_c2_c3_postreset(struct ata_link *link, unsigned int *classes)
{
u8 r;
int port_bit = 4 << link->ap->port_no;
/* If our bridge is an ALI 1533 then do the extra work */
if (ali_isa_bridge) {
/* Tristate and re-enable the bus signals */
pci_read_config_byte(ali_isa_bridge, 0x58, &r);
r &= ~port_bit;
pci_write_config_byte(ali_isa_bridge, 0x58, r);
r |= port_bit;
pci_write_config_byte(ali_isa_bridge, 0x58, r);
}
ata_sff_postreset(link, classes);
}
static struct scsi_host_template ali_sht = {
ATA_BMDMA_SHT(DRV_NAME),
};
/*
* Port operations for PIO only ALi
*/
static struct ata_port_operations ali_early_port_ops = {
libata: implement and use ops inheritance libata lets low level drivers build ata_port_operations table and register it with libata core layer. This allows low level drivers high level of flexibility but also burdens them with lots of boilerplate entries. This becomes worse for drivers which support related similar controllers which differ slightly. They share most of the operations except for a few. However, the driver still needs to list all operations for each variant. This results in large number of duplicate entries, which is not only inefficient but also error-prone as it becomes very difficult to tell what the actual differences are. This duplicate boilerplates all over the low level drivers also make updating the core layer exteremely difficult and error-prone. When compounded with multi-branched development model, it ends up accumulating inconsistencies over time. Some of those inconsistencies cause immediate problems and fixed. Others just remain there dormant making maintenance increasingly difficult. To rectify the problem, this patch implements ata_port_operations inheritance. To allow LLDs to easily re-use their own ops tables overriding only specific methods, this patch implements poor man's class inheritance. An ops table has ->inherits field which can be set to any ops table as long as it doesn't create a loop. When the host is started, the inheritance chain is followed and any operation which isn't specified is taken from the nearest ancestor which has it specified. This operation is called finalization and done only once per an ops table and the LLD doesn't have to do anything special about it other than making the ops table non-const such that libata can update it. libata provides four base ops tables lower drivers can inherit from - base, sata, pmp, sff and bmdma. To avoid overriding these ops accidentaly, these ops are declared const and LLDs should always inherit these instead of using them directly. After finalization, all the ops table are identical before and after the patch except for setting .irq_handler to ata_interrupt in drivers which didn't use to. The .irq_handler doesn't have any actual effect and the field will soon be removed by later patch. * sata_sx4 is still using old style EH and currently doesn't take advantage of ops inheritance. Signed-off-by: Tejun Heo <htejun@gmail.com>
2008-03-25 11:22:49 +08:00
.inherits = &ata_sff_port_ops,
.cable_detect = ata_cable_40wire,
libata: implement and use ops inheritance libata lets low level drivers build ata_port_operations table and register it with libata core layer. This allows low level drivers high level of flexibility but also burdens them with lots of boilerplate entries. This becomes worse for drivers which support related similar controllers which differ slightly. They share most of the operations except for a few. However, the driver still needs to list all operations for each variant. This results in large number of duplicate entries, which is not only inefficient but also error-prone as it becomes very difficult to tell what the actual differences are. This duplicate boilerplates all over the low level drivers also make updating the core layer exteremely difficult and error-prone. When compounded with multi-branched development model, it ends up accumulating inconsistencies over time. Some of those inconsistencies cause immediate problems and fixed. Others just remain there dormant making maintenance increasingly difficult. To rectify the problem, this patch implements ata_port_operations inheritance. To allow LLDs to easily re-use their own ops tables overriding only specific methods, this patch implements poor man's class inheritance. An ops table has ->inherits field which can be set to any ops table as long as it doesn't create a loop. When the host is started, the inheritance chain is followed and any operation which isn't specified is taken from the nearest ancestor which has it specified. This operation is called finalization and done only once per an ops table and the LLD doesn't have to do anything special about it other than making the ops table non-const such that libata can update it. libata provides four base ops tables lower drivers can inherit from - base, sata, pmp, sff and bmdma. To avoid overriding these ops accidentaly, these ops are declared const and LLDs should always inherit these instead of using them directly. After finalization, all the ops table are identical before and after the patch except for setting .irq_handler to ata_interrupt in drivers which didn't use to. The .irq_handler doesn't have any actual effect and the field will soon be removed by later patch. * sata_sx4 is still using old style EH and currently doesn't take advantage of ops inheritance. Signed-off-by: Tejun Heo <htejun@gmail.com>
2008-03-25 11:22:49 +08:00
.set_piomode = ali_set_piomode,
.sff_data_xfer = ata_sff_data_xfer32,
libata: implement and use ops inheritance libata lets low level drivers build ata_port_operations table and register it with libata core layer. This allows low level drivers high level of flexibility but also burdens them with lots of boilerplate entries. This becomes worse for drivers which support related similar controllers which differ slightly. They share most of the operations except for a few. However, the driver still needs to list all operations for each variant. This results in large number of duplicate entries, which is not only inefficient but also error-prone as it becomes very difficult to tell what the actual differences are. This duplicate boilerplates all over the low level drivers also make updating the core layer exteremely difficult and error-prone. When compounded with multi-branched development model, it ends up accumulating inconsistencies over time. Some of those inconsistencies cause immediate problems and fixed. Others just remain there dormant making maintenance increasingly difficult. To rectify the problem, this patch implements ata_port_operations inheritance. To allow LLDs to easily re-use their own ops tables overriding only specific methods, this patch implements poor man's class inheritance. An ops table has ->inherits field which can be set to any ops table as long as it doesn't create a loop. When the host is started, the inheritance chain is followed and any operation which isn't specified is taken from the nearest ancestor which has it specified. This operation is called finalization and done only once per an ops table and the LLD doesn't have to do anything special about it other than making the ops table non-const such that libata can update it. libata provides four base ops tables lower drivers can inherit from - base, sata, pmp, sff and bmdma. To avoid overriding these ops accidentaly, these ops are declared const and LLDs should always inherit these instead of using them directly. After finalization, all the ops table are identical before and after the patch except for setting .irq_handler to ata_interrupt in drivers which didn't use to. The .irq_handler doesn't have any actual effect and the field will soon be removed by later patch. * sata_sx4 is still using old style EH and currently doesn't take advantage of ops inheritance. Signed-off-by: Tejun Heo <htejun@gmail.com>
2008-03-25 11:22:49 +08:00
};
libata: implement and use ops inheritance libata lets low level drivers build ata_port_operations table and register it with libata core layer. This allows low level drivers high level of flexibility but also burdens them with lots of boilerplate entries. This becomes worse for drivers which support related similar controllers which differ slightly. They share most of the operations except for a few. However, the driver still needs to list all operations for each variant. This results in large number of duplicate entries, which is not only inefficient but also error-prone as it becomes very difficult to tell what the actual differences are. This duplicate boilerplates all over the low level drivers also make updating the core layer exteremely difficult and error-prone. When compounded with multi-branched development model, it ends up accumulating inconsistencies over time. Some of those inconsistencies cause immediate problems and fixed. Others just remain there dormant making maintenance increasingly difficult. To rectify the problem, this patch implements ata_port_operations inheritance. To allow LLDs to easily re-use their own ops tables overriding only specific methods, this patch implements poor man's class inheritance. An ops table has ->inherits field which can be set to any ops table as long as it doesn't create a loop. When the host is started, the inheritance chain is followed and any operation which isn't specified is taken from the nearest ancestor which has it specified. This operation is called finalization and done only once per an ops table and the LLD doesn't have to do anything special about it other than making the ops table non-const such that libata can update it. libata provides four base ops tables lower drivers can inherit from - base, sata, pmp, sff and bmdma. To avoid overriding these ops accidentaly, these ops are declared const and LLDs should always inherit these instead of using them directly. After finalization, all the ops table are identical before and after the patch except for setting .irq_handler to ata_interrupt in drivers which didn't use to. The .irq_handler doesn't have any actual effect and the field will soon be removed by later patch. * sata_sx4 is still using old style EH and currently doesn't take advantage of ops inheritance. Signed-off-by: Tejun Heo <htejun@gmail.com>
2008-03-25 11:22:49 +08:00
static const struct ata_port_operations ali_dma_base_ops = {
.inherits = &ata_bmdma32_port_ops,
libata: implement and use ops inheritance libata lets low level drivers build ata_port_operations table and register it with libata core layer. This allows low level drivers high level of flexibility but also burdens them with lots of boilerplate entries. This becomes worse for drivers which support related similar controllers which differ slightly. They share most of the operations except for a few. However, the driver still needs to list all operations for each variant. This results in large number of duplicate entries, which is not only inefficient but also error-prone as it becomes very difficult to tell what the actual differences are. This duplicate boilerplates all over the low level drivers also make updating the core layer exteremely difficult and error-prone. When compounded with multi-branched development model, it ends up accumulating inconsistencies over time. Some of those inconsistencies cause immediate problems and fixed. Others just remain there dormant making maintenance increasingly difficult. To rectify the problem, this patch implements ata_port_operations inheritance. To allow LLDs to easily re-use their own ops tables overriding only specific methods, this patch implements poor man's class inheritance. An ops table has ->inherits field which can be set to any ops table as long as it doesn't create a loop. When the host is started, the inheritance chain is followed and any operation which isn't specified is taken from the nearest ancestor which has it specified. This operation is called finalization and done only once per an ops table and the LLD doesn't have to do anything special about it other than making the ops table non-const such that libata can update it. libata provides four base ops tables lower drivers can inherit from - base, sata, pmp, sff and bmdma. To avoid overriding these ops accidentaly, these ops are declared const and LLDs should always inherit these instead of using them directly. After finalization, all the ops table are identical before and after the patch except for setting .irq_handler to ata_interrupt in drivers which didn't use to. The .irq_handler doesn't have any actual effect and the field will soon be removed by later patch. * sata_sx4 is still using old style EH and currently doesn't take advantage of ops inheritance. Signed-off-by: Tejun Heo <htejun@gmail.com>
2008-03-25 11:22:49 +08:00
.set_piomode = ali_set_piomode,
.set_dmamode = ali_set_dmamode,
};
/*
* Port operations for DMA capable ALi without cable
* detect
*/
static struct ata_port_operations ali_20_port_ops = {
libata: implement and use ops inheritance libata lets low level drivers build ata_port_operations table and register it with libata core layer. This allows low level drivers high level of flexibility but also burdens them with lots of boilerplate entries. This becomes worse for drivers which support related similar controllers which differ slightly. They share most of the operations except for a few. However, the driver still needs to list all operations for each variant. This results in large number of duplicate entries, which is not only inefficient but also error-prone as it becomes very difficult to tell what the actual differences are. This duplicate boilerplates all over the low level drivers also make updating the core layer exteremely difficult and error-prone. When compounded with multi-branched development model, it ends up accumulating inconsistencies over time. Some of those inconsistencies cause immediate problems and fixed. Others just remain there dormant making maintenance increasingly difficult. To rectify the problem, this patch implements ata_port_operations inheritance. To allow LLDs to easily re-use their own ops tables overriding only specific methods, this patch implements poor man's class inheritance. An ops table has ->inherits field which can be set to any ops table as long as it doesn't create a loop. When the host is started, the inheritance chain is followed and any operation which isn't specified is taken from the nearest ancestor which has it specified. This operation is called finalization and done only once per an ops table and the LLD doesn't have to do anything special about it other than making the ops table non-const such that libata can update it. libata provides four base ops tables lower drivers can inherit from - base, sata, pmp, sff and bmdma. To avoid overriding these ops accidentaly, these ops are declared const and LLDs should always inherit these instead of using them directly. After finalization, all the ops table are identical before and after the patch except for setting .irq_handler to ata_interrupt in drivers which didn't use to. The .irq_handler doesn't have any actual effect and the field will soon be removed by later patch. * sata_sx4 is still using old style EH and currently doesn't take advantage of ops inheritance. Signed-off-by: Tejun Heo <htejun@gmail.com>
2008-03-25 11:22:49 +08:00
.inherits = &ali_dma_base_ops,
.cable_detect = ata_cable_40wire,
.mode_filter = ali_20_filter,
.check_atapi_dma = ali_check_atapi_dma,
.dev_config = ali_lock_sectors,
};
/*
* Port operations for DMA capable ALi with cable detect
*/
static struct ata_port_operations ali_c2_port_ops = {
libata: implement and use ops inheritance libata lets low level drivers build ata_port_operations table and register it with libata core layer. This allows low level drivers high level of flexibility but also burdens them with lots of boilerplate entries. This becomes worse for drivers which support related similar controllers which differ slightly. They share most of the operations except for a few. However, the driver still needs to list all operations for each variant. This results in large number of duplicate entries, which is not only inefficient but also error-prone as it becomes very difficult to tell what the actual differences are. This duplicate boilerplates all over the low level drivers also make updating the core layer exteremely difficult and error-prone. When compounded with multi-branched development model, it ends up accumulating inconsistencies over time. Some of those inconsistencies cause immediate problems and fixed. Others just remain there dormant making maintenance increasingly difficult. To rectify the problem, this patch implements ata_port_operations inheritance. To allow LLDs to easily re-use their own ops tables overriding only specific methods, this patch implements poor man's class inheritance. An ops table has ->inherits field which can be set to any ops table as long as it doesn't create a loop. When the host is started, the inheritance chain is followed and any operation which isn't specified is taken from the nearest ancestor which has it specified. This operation is called finalization and done only once per an ops table and the LLD doesn't have to do anything special about it other than making the ops table non-const such that libata can update it. libata provides four base ops tables lower drivers can inherit from - base, sata, pmp, sff and bmdma. To avoid overriding these ops accidentaly, these ops are declared const and LLDs should always inherit these instead of using them directly. After finalization, all the ops table are identical before and after the patch except for setting .irq_handler to ata_interrupt in drivers which didn't use to. The .irq_handler doesn't have any actual effect and the field will soon be removed by later patch. * sata_sx4 is still using old style EH and currently doesn't take advantage of ops inheritance. Signed-off-by: Tejun Heo <htejun@gmail.com>
2008-03-25 11:22:49 +08:00
.inherits = &ali_dma_base_ops,
.check_atapi_dma = ali_check_atapi_dma,
.cable_detect = ali_c2_cable_detect,
libata: implement and use ops inheritance libata lets low level drivers build ata_port_operations table and register it with libata core layer. This allows low level drivers high level of flexibility but also burdens them with lots of boilerplate entries. This becomes worse for drivers which support related similar controllers which differ slightly. They share most of the operations except for a few. However, the driver still needs to list all operations for each variant. This results in large number of duplicate entries, which is not only inefficient but also error-prone as it becomes very difficult to tell what the actual differences are. This duplicate boilerplates all over the low level drivers also make updating the core layer exteremely difficult and error-prone. When compounded with multi-branched development model, it ends up accumulating inconsistencies over time. Some of those inconsistencies cause immediate problems and fixed. Others just remain there dormant making maintenance increasingly difficult. To rectify the problem, this patch implements ata_port_operations inheritance. To allow LLDs to easily re-use their own ops tables overriding only specific methods, this patch implements poor man's class inheritance. An ops table has ->inherits field which can be set to any ops table as long as it doesn't create a loop. When the host is started, the inheritance chain is followed and any operation which isn't specified is taken from the nearest ancestor which has it specified. This operation is called finalization and done only once per an ops table and the LLD doesn't have to do anything special about it other than making the ops table non-const such that libata can update it. libata provides four base ops tables lower drivers can inherit from - base, sata, pmp, sff and bmdma. To avoid overriding these ops accidentaly, these ops are declared const and LLDs should always inherit these instead of using them directly. After finalization, all the ops table are identical before and after the patch except for setting .irq_handler to ata_interrupt in drivers which didn't use to. The .irq_handler doesn't have any actual effect and the field will soon be removed by later patch. * sata_sx4 is still using old style EH and currently doesn't take advantage of ops inheritance. Signed-off-by: Tejun Heo <htejun@gmail.com>
2008-03-25 11:22:49 +08:00
.dev_config = ali_lock_sectors,
.postreset = ali_c2_c3_postreset,
};
/*
* Port operations for DMA capable ALi with cable detect
*/
static struct ata_port_operations ali_c4_port_ops = {
.inherits = &ali_dma_base_ops,
.check_atapi_dma = ali_check_atapi_dma,
.cable_detect = ali_c2_cable_detect,
.dev_config = ali_lock_sectors,
};
/*
* Port operations for DMA capable ALi with cable detect and LBA48
*/
static struct ata_port_operations ali_c5_port_ops = {
libata: implement and use ops inheritance libata lets low level drivers build ata_port_operations table and register it with libata core layer. This allows low level drivers high level of flexibility but also burdens them with lots of boilerplate entries. This becomes worse for drivers which support related similar controllers which differ slightly. They share most of the operations except for a few. However, the driver still needs to list all operations for each variant. This results in large number of duplicate entries, which is not only inefficient but also error-prone as it becomes very difficult to tell what the actual differences are. This duplicate boilerplates all over the low level drivers also make updating the core layer exteremely difficult and error-prone. When compounded with multi-branched development model, it ends up accumulating inconsistencies over time. Some of those inconsistencies cause immediate problems and fixed. Others just remain there dormant making maintenance increasingly difficult. To rectify the problem, this patch implements ata_port_operations inheritance. To allow LLDs to easily re-use their own ops tables overriding only specific methods, this patch implements poor man's class inheritance. An ops table has ->inherits field which can be set to any ops table as long as it doesn't create a loop. When the host is started, the inheritance chain is followed and any operation which isn't specified is taken from the nearest ancestor which has it specified. This operation is called finalization and done only once per an ops table and the LLD doesn't have to do anything special about it other than making the ops table non-const such that libata can update it. libata provides four base ops tables lower drivers can inherit from - base, sata, pmp, sff and bmdma. To avoid overriding these ops accidentaly, these ops are declared const and LLDs should always inherit these instead of using them directly. After finalization, all the ops table are identical before and after the patch except for setting .irq_handler to ata_interrupt in drivers which didn't use to. The .irq_handler doesn't have any actual effect and the field will soon be removed by later patch. * sata_sx4 is still using old style EH and currently doesn't take advantage of ops inheritance. Signed-off-by: Tejun Heo <htejun@gmail.com>
2008-03-25 11:22:49 +08:00
.inherits = &ali_dma_base_ops,
.check_atapi_dma = ali_check_atapi_dma,
.dev_config = ali_warn_atapi_dma,
.cable_detect = ali_c2_cable_detect,
};
/**
* ali_init_chipset - chip setup function
* @pdev: PCI device of ATA controller
*
* Perform the setup on the device that must be done both at boot
* and at resume time.
*/
static void ali_init_chipset(struct pci_dev *pdev)
{
u8 tmp;
struct pci_dev *north;
/*
* The chipset revision selects the driver operations and
* mode data.
*/
if (pdev->revision <= 0x20) {
pci_read_config_byte(pdev, 0x53, &tmp);
tmp |= 0x03;
pci_write_config_byte(pdev, 0x53, tmp);
} else {
pci_read_config_byte(pdev, 0x4a, &tmp);
pci_write_config_byte(pdev, 0x4a, tmp | 0x20);
pci_read_config_byte(pdev, 0x4B, &tmp);
if (pdev->revision < 0xC2)
/* 1543-E/F, 1543C-C, 1543C-D, 1543C-E */
/* Clear CD-ROM DMA write bit */
tmp &= 0x7F;
/* Cable and UDMA */
if (pdev->revision >= 0xc2)
tmp |= 0x01;
pci_write_config_byte(pdev, 0x4B, tmp | 0x08);
/*
* CD_ROM DMA on (0x53 bit 0). Enable this even if we want
* to use PIO. 0x53 bit 1 (rev 20 only) - enable FIFO control
* via 0x54/55.
*/
pci_read_config_byte(pdev, 0x53, &tmp);
if (pdev->revision >= 0xc7)
tmp |= 0x03;
else
tmp |= 0x01; /* CD_ROM enable for DMA */
pci_write_config_byte(pdev, 0x53, tmp);
}
north = pci_get_domain_bus_and_slot(pci_domain_nr(pdev->bus), 0,
PCI_DEVFN(0, 0));
if (north && north->vendor == PCI_VENDOR_ID_AL && ali_isa_bridge) {
/* Configure the ALi bridge logic. For non ALi rely on BIOS.
Set the south bridge enable bit */
pci_read_config_byte(ali_isa_bridge, 0x79, &tmp);
if (pdev->revision == 0xC2)
pci_write_config_byte(ali_isa_bridge, 0x79, tmp | 0x04);
else if (pdev->revision > 0xC2 && pdev->revision < 0xC5)
pci_write_config_byte(ali_isa_bridge, 0x79, tmp | 0x02);
}
pci_dev_put(north);
ata_pci_bmdma_clear_simplex(pdev);
}
/**
* ali_init_one - discovery callback
* @pdev: PCI device ID
* @id: PCI table info
*
* An ALi IDE interface has been discovered. Figure out what revision
* and perform configuration work before handing it to the ATA layer
*/
static int ali_init_one(struct pci_dev *pdev, const struct pci_device_id *id)
{
libata: clean up SFF init mess The intention of using port_mask in SFF init helpers was to eventually support exoctic configurations such as combination of legacy and native port on the same controller. This never became actually necessary and the related code always has been subtly broken one way or the other. Now that new init model is in place, there is no reason to make common helpers capable of handling all corner cases. Exotic cases can simply dealt within LLDs as necessary. This patch removes port_mask handling in SFF init helpers. SFF init helpers don't take n_ports argument and interpret it into port_mask anymore. All information is carried via port_info. n_ports argument is dropped and always two ports are allocated. LLD can tell SFF to skip certain port by marking it dummy. Note that SFF code has been treating unuvailable ports this way for a long time until recent breakage fix from Linus and is consistent with how other drivers handle with unavailable ports. This fixes 1-port legacy host handling still broken after the recent native mode fix and simplifies SFF init logic. The following changes are made... * ata_pci_init_native_host() and ata_init_legacy_host() both now try to initialized whatever they can and mark failed ports dummy. They return 0 if any port is successfully initialized. * ata_pci_prepare_native_host() and ata_pci_init_one() now doesn't take n_ports argument. All info should be specified via port_info array. Always two ports are allocated. * ata_pci_init_bmdma() exported to be used by LLDs in exotic cases. * port_info handling in all LLDs are standardized - all port_info arrays are const stack variable named ppi. Unless the second port is different from the first, its port_info is specified as NULL (tells libata that it's identical to the last non-NULL port_info). * pata_hpt37x/hpt3x2n: don't modify static variable directly. Make an on-stack copy instead as ata_piix does. * pata_uli: It has 4 ports instead of 2. Don't use ata_pci_prepare_native_host(). Allocate the host explicitly and use init helpers. It's simple enough. Signed-off-by: Tejun Heo <htejun@gmail.com> Signed-off-by: Jeff Garzik <jeff@garzik.org>
2007-05-04 18:43:58 +08:00
static const struct ata_port_info info_early = {
.flags = ATA_FLAG_SLAVE_POSS,
.pio_mask = ATA_PIO4,
.port_ops = &ali_early_port_ops
};
/* Revision 0x20 added DMA */
libata: clean up SFF init mess The intention of using port_mask in SFF init helpers was to eventually support exoctic configurations such as combination of legacy and native port on the same controller. This never became actually necessary and the related code always has been subtly broken one way or the other. Now that new init model is in place, there is no reason to make common helpers capable of handling all corner cases. Exotic cases can simply dealt within LLDs as necessary. This patch removes port_mask handling in SFF init helpers. SFF init helpers don't take n_ports argument and interpret it into port_mask anymore. All information is carried via port_info. n_ports argument is dropped and always two ports are allocated. LLD can tell SFF to skip certain port by marking it dummy. Note that SFF code has been treating unuvailable ports this way for a long time until recent breakage fix from Linus and is consistent with how other drivers handle with unavailable ports. This fixes 1-port legacy host handling still broken after the recent native mode fix and simplifies SFF init logic. The following changes are made... * ata_pci_init_native_host() and ata_init_legacy_host() both now try to initialized whatever they can and mark failed ports dummy. They return 0 if any port is successfully initialized. * ata_pci_prepare_native_host() and ata_pci_init_one() now doesn't take n_ports argument. All info should be specified via port_info array. Always two ports are allocated. * ata_pci_init_bmdma() exported to be used by LLDs in exotic cases. * port_info handling in all LLDs are standardized - all port_info arrays are const stack variable named ppi. Unless the second port is different from the first, its port_info is specified as NULL (tells libata that it's identical to the last non-NULL port_info). * pata_hpt37x/hpt3x2n: don't modify static variable directly. Make an on-stack copy instead as ata_piix does. * pata_uli: It has 4 ports instead of 2. Don't use ata_pci_prepare_native_host(). Allocate the host explicitly and use init helpers. It's simple enough. Signed-off-by: Tejun Heo <htejun@gmail.com> Signed-off-by: Jeff Garzik <jeff@garzik.org>
2007-05-04 18:43:58 +08:00
static const struct ata_port_info info_20 = {
.flags = ATA_FLAG_SLAVE_POSS | ATA_FLAG_PIO_LBA48 |
ATA_FLAG_IGN_SIMPLEX,
.pio_mask = ATA_PIO4,
.mwdma_mask = ATA_MWDMA2,
.port_ops = &ali_20_port_ops
};
/* Revision 0x20 with support logic added UDMA */
libata: clean up SFF init mess The intention of using port_mask in SFF init helpers was to eventually support exoctic configurations such as combination of legacy and native port on the same controller. This never became actually necessary and the related code always has been subtly broken one way or the other. Now that new init model is in place, there is no reason to make common helpers capable of handling all corner cases. Exotic cases can simply dealt within LLDs as necessary. This patch removes port_mask handling in SFF init helpers. SFF init helpers don't take n_ports argument and interpret it into port_mask anymore. All information is carried via port_info. n_ports argument is dropped and always two ports are allocated. LLD can tell SFF to skip certain port by marking it dummy. Note that SFF code has been treating unuvailable ports this way for a long time until recent breakage fix from Linus and is consistent with how other drivers handle with unavailable ports. This fixes 1-port legacy host handling still broken after the recent native mode fix and simplifies SFF init logic. The following changes are made... * ata_pci_init_native_host() and ata_init_legacy_host() both now try to initialized whatever they can and mark failed ports dummy. They return 0 if any port is successfully initialized. * ata_pci_prepare_native_host() and ata_pci_init_one() now doesn't take n_ports argument. All info should be specified via port_info array. Always two ports are allocated. * ata_pci_init_bmdma() exported to be used by LLDs in exotic cases. * port_info handling in all LLDs are standardized - all port_info arrays are const stack variable named ppi. Unless the second port is different from the first, its port_info is specified as NULL (tells libata that it's identical to the last non-NULL port_info). * pata_hpt37x/hpt3x2n: don't modify static variable directly. Make an on-stack copy instead as ata_piix does. * pata_uli: It has 4 ports instead of 2. Don't use ata_pci_prepare_native_host(). Allocate the host explicitly and use init helpers. It's simple enough. Signed-off-by: Tejun Heo <htejun@gmail.com> Signed-off-by: Jeff Garzik <jeff@garzik.org>
2007-05-04 18:43:58 +08:00
static const struct ata_port_info info_20_udma = {
.flags = ATA_FLAG_SLAVE_POSS | ATA_FLAG_PIO_LBA48 |
ATA_FLAG_IGN_SIMPLEX,
.pio_mask = ATA_PIO4,
.mwdma_mask = ATA_MWDMA2,
.udma_mask = ATA_UDMA2,
.port_ops = &ali_20_port_ops
};
/* Revision 0xC2 adds UDMA66 */
libata: clean up SFF init mess The intention of using port_mask in SFF init helpers was to eventually support exoctic configurations such as combination of legacy and native port on the same controller. This never became actually necessary and the related code always has been subtly broken one way or the other. Now that new init model is in place, there is no reason to make common helpers capable of handling all corner cases. Exotic cases can simply dealt within LLDs as necessary. This patch removes port_mask handling in SFF init helpers. SFF init helpers don't take n_ports argument and interpret it into port_mask anymore. All information is carried via port_info. n_ports argument is dropped and always two ports are allocated. LLD can tell SFF to skip certain port by marking it dummy. Note that SFF code has been treating unuvailable ports this way for a long time until recent breakage fix from Linus and is consistent with how other drivers handle with unavailable ports. This fixes 1-port legacy host handling still broken after the recent native mode fix and simplifies SFF init logic. The following changes are made... * ata_pci_init_native_host() and ata_init_legacy_host() both now try to initialized whatever they can and mark failed ports dummy. They return 0 if any port is successfully initialized. * ata_pci_prepare_native_host() and ata_pci_init_one() now doesn't take n_ports argument. All info should be specified via port_info array. Always two ports are allocated. * ata_pci_init_bmdma() exported to be used by LLDs in exotic cases. * port_info handling in all LLDs are standardized - all port_info arrays are const stack variable named ppi. Unless the second port is different from the first, its port_info is specified as NULL (tells libata that it's identical to the last non-NULL port_info). * pata_hpt37x/hpt3x2n: don't modify static variable directly. Make an on-stack copy instead as ata_piix does. * pata_uli: It has 4 ports instead of 2. Don't use ata_pci_prepare_native_host(). Allocate the host explicitly and use init helpers. It's simple enough. Signed-off-by: Tejun Heo <htejun@gmail.com> Signed-off-by: Jeff Garzik <jeff@garzik.org>
2007-05-04 18:43:58 +08:00
static const struct ata_port_info info_c2 = {
.flags = ATA_FLAG_SLAVE_POSS | ATA_FLAG_PIO_LBA48 |
ATA_FLAG_IGN_SIMPLEX,
.pio_mask = ATA_PIO4,
.mwdma_mask = ATA_MWDMA2,
.udma_mask = ATA_UDMA4,
.port_ops = &ali_c2_port_ops
};
/* Revision 0xC3 is UDMA66 for now */
libata: clean up SFF init mess The intention of using port_mask in SFF init helpers was to eventually support exoctic configurations such as combination of legacy and native port on the same controller. This never became actually necessary and the related code always has been subtly broken one way or the other. Now that new init model is in place, there is no reason to make common helpers capable of handling all corner cases. Exotic cases can simply dealt within LLDs as necessary. This patch removes port_mask handling in SFF init helpers. SFF init helpers don't take n_ports argument and interpret it into port_mask anymore. All information is carried via port_info. n_ports argument is dropped and always two ports are allocated. LLD can tell SFF to skip certain port by marking it dummy. Note that SFF code has been treating unuvailable ports this way for a long time until recent breakage fix from Linus and is consistent with how other drivers handle with unavailable ports. This fixes 1-port legacy host handling still broken after the recent native mode fix and simplifies SFF init logic. The following changes are made... * ata_pci_init_native_host() and ata_init_legacy_host() both now try to initialized whatever they can and mark failed ports dummy. They return 0 if any port is successfully initialized. * ata_pci_prepare_native_host() and ata_pci_init_one() now doesn't take n_ports argument. All info should be specified via port_info array. Always two ports are allocated. * ata_pci_init_bmdma() exported to be used by LLDs in exotic cases. * port_info handling in all LLDs are standardized - all port_info arrays are const stack variable named ppi. Unless the second port is different from the first, its port_info is specified as NULL (tells libata that it's identical to the last non-NULL port_info). * pata_hpt37x/hpt3x2n: don't modify static variable directly. Make an on-stack copy instead as ata_piix does. * pata_uli: It has 4 ports instead of 2. Don't use ata_pci_prepare_native_host(). Allocate the host explicitly and use init helpers. It's simple enough. Signed-off-by: Tejun Heo <htejun@gmail.com> Signed-off-by: Jeff Garzik <jeff@garzik.org>
2007-05-04 18:43:58 +08:00
static const struct ata_port_info info_c3 = {
.flags = ATA_FLAG_SLAVE_POSS | ATA_FLAG_PIO_LBA48 |
ATA_FLAG_IGN_SIMPLEX,
.pio_mask = ATA_PIO4,
.mwdma_mask = ATA_MWDMA2,
.udma_mask = ATA_UDMA4,
.port_ops = &ali_c2_port_ops
};
/* Revision 0xC4 is UDMA100 */
libata: clean up SFF init mess The intention of using port_mask in SFF init helpers was to eventually support exoctic configurations such as combination of legacy and native port on the same controller. This never became actually necessary and the related code always has been subtly broken one way or the other. Now that new init model is in place, there is no reason to make common helpers capable of handling all corner cases. Exotic cases can simply dealt within LLDs as necessary. This patch removes port_mask handling in SFF init helpers. SFF init helpers don't take n_ports argument and interpret it into port_mask anymore. All information is carried via port_info. n_ports argument is dropped and always two ports are allocated. LLD can tell SFF to skip certain port by marking it dummy. Note that SFF code has been treating unuvailable ports this way for a long time until recent breakage fix from Linus and is consistent with how other drivers handle with unavailable ports. This fixes 1-port legacy host handling still broken after the recent native mode fix and simplifies SFF init logic. The following changes are made... * ata_pci_init_native_host() and ata_init_legacy_host() both now try to initialized whatever they can and mark failed ports dummy. They return 0 if any port is successfully initialized. * ata_pci_prepare_native_host() and ata_pci_init_one() now doesn't take n_ports argument. All info should be specified via port_info array. Always two ports are allocated. * ata_pci_init_bmdma() exported to be used by LLDs in exotic cases. * port_info handling in all LLDs are standardized - all port_info arrays are const stack variable named ppi. Unless the second port is different from the first, its port_info is specified as NULL (tells libata that it's identical to the last non-NULL port_info). * pata_hpt37x/hpt3x2n: don't modify static variable directly. Make an on-stack copy instead as ata_piix does. * pata_uli: It has 4 ports instead of 2. Don't use ata_pci_prepare_native_host(). Allocate the host explicitly and use init helpers. It's simple enough. Signed-off-by: Tejun Heo <htejun@gmail.com> Signed-off-by: Jeff Garzik <jeff@garzik.org>
2007-05-04 18:43:58 +08:00
static const struct ata_port_info info_c4 = {
.flags = ATA_FLAG_SLAVE_POSS | ATA_FLAG_PIO_LBA48 |
ATA_FLAG_IGN_SIMPLEX,
.pio_mask = ATA_PIO4,
.mwdma_mask = ATA_MWDMA2,
.udma_mask = ATA_UDMA5,
.port_ops = &ali_c4_port_ops
};
/* Revision 0xC5 is UDMA133 with LBA48 DMA */
libata: clean up SFF init mess The intention of using port_mask in SFF init helpers was to eventually support exoctic configurations such as combination of legacy and native port on the same controller. This never became actually necessary and the related code always has been subtly broken one way or the other. Now that new init model is in place, there is no reason to make common helpers capable of handling all corner cases. Exotic cases can simply dealt within LLDs as necessary. This patch removes port_mask handling in SFF init helpers. SFF init helpers don't take n_ports argument and interpret it into port_mask anymore. All information is carried via port_info. n_ports argument is dropped and always two ports are allocated. LLD can tell SFF to skip certain port by marking it dummy. Note that SFF code has been treating unuvailable ports this way for a long time until recent breakage fix from Linus and is consistent with how other drivers handle with unavailable ports. This fixes 1-port legacy host handling still broken after the recent native mode fix and simplifies SFF init logic. The following changes are made... * ata_pci_init_native_host() and ata_init_legacy_host() both now try to initialized whatever they can and mark failed ports dummy. They return 0 if any port is successfully initialized. * ata_pci_prepare_native_host() and ata_pci_init_one() now doesn't take n_ports argument. All info should be specified via port_info array. Always two ports are allocated. * ata_pci_init_bmdma() exported to be used by LLDs in exotic cases. * port_info handling in all LLDs are standardized - all port_info arrays are const stack variable named ppi. Unless the second port is different from the first, its port_info is specified as NULL (tells libata that it's identical to the last non-NULL port_info). * pata_hpt37x/hpt3x2n: don't modify static variable directly. Make an on-stack copy instead as ata_piix does. * pata_uli: It has 4 ports instead of 2. Don't use ata_pci_prepare_native_host(). Allocate the host explicitly and use init helpers. It's simple enough. Signed-off-by: Tejun Heo <htejun@gmail.com> Signed-off-by: Jeff Garzik <jeff@garzik.org>
2007-05-04 18:43:58 +08:00
static const struct ata_port_info info_c5 = {
.flags = ATA_FLAG_SLAVE_POSS | ATA_FLAG_IGN_SIMPLEX,
.pio_mask = ATA_PIO4,
.mwdma_mask = ATA_MWDMA2,
.udma_mask = ATA_UDMA6,
.port_ops = &ali_c5_port_ops
};
libata: clean up SFF init mess The intention of using port_mask in SFF init helpers was to eventually support exoctic configurations such as combination of legacy and native port on the same controller. This never became actually necessary and the related code always has been subtly broken one way or the other. Now that new init model is in place, there is no reason to make common helpers capable of handling all corner cases. Exotic cases can simply dealt within LLDs as necessary. This patch removes port_mask handling in SFF init helpers. SFF init helpers don't take n_ports argument and interpret it into port_mask anymore. All information is carried via port_info. n_ports argument is dropped and always two ports are allocated. LLD can tell SFF to skip certain port by marking it dummy. Note that SFF code has been treating unuvailable ports this way for a long time until recent breakage fix from Linus and is consistent with how other drivers handle with unavailable ports. This fixes 1-port legacy host handling still broken after the recent native mode fix and simplifies SFF init logic. The following changes are made... * ata_pci_init_native_host() and ata_init_legacy_host() both now try to initialized whatever they can and mark failed ports dummy. They return 0 if any port is successfully initialized. * ata_pci_prepare_native_host() and ata_pci_init_one() now doesn't take n_ports argument. All info should be specified via port_info array. Always two ports are allocated. * ata_pci_init_bmdma() exported to be used by LLDs in exotic cases. * port_info handling in all LLDs are standardized - all port_info arrays are const stack variable named ppi. Unless the second port is different from the first, its port_info is specified as NULL (tells libata that it's identical to the last non-NULL port_info). * pata_hpt37x/hpt3x2n: don't modify static variable directly. Make an on-stack copy instead as ata_piix does. * pata_uli: It has 4 ports instead of 2. Don't use ata_pci_prepare_native_host(). Allocate the host explicitly and use init helpers. It's simple enough. Signed-off-by: Tejun Heo <htejun@gmail.com> Signed-off-by: Jeff Garzik <jeff@garzik.org>
2007-05-04 18:43:58 +08:00
const struct ata_port_info *ppi[] = { NULL, NULL };
u8 tmp;
int rc;
rc = pcim_enable_device(pdev);
if (rc)
return rc;
/*
* The chipset revision selects the driver operations and
* mode data.
*/
if (pdev->revision < 0x20) {
libata: clean up SFF init mess The intention of using port_mask in SFF init helpers was to eventually support exoctic configurations such as combination of legacy and native port on the same controller. This never became actually necessary and the related code always has been subtly broken one way or the other. Now that new init model is in place, there is no reason to make common helpers capable of handling all corner cases. Exotic cases can simply dealt within LLDs as necessary. This patch removes port_mask handling in SFF init helpers. SFF init helpers don't take n_ports argument and interpret it into port_mask anymore. All information is carried via port_info. n_ports argument is dropped and always two ports are allocated. LLD can tell SFF to skip certain port by marking it dummy. Note that SFF code has been treating unuvailable ports this way for a long time until recent breakage fix from Linus and is consistent with how other drivers handle with unavailable ports. This fixes 1-port legacy host handling still broken after the recent native mode fix and simplifies SFF init logic. The following changes are made... * ata_pci_init_native_host() and ata_init_legacy_host() both now try to initialized whatever they can and mark failed ports dummy. They return 0 if any port is successfully initialized. * ata_pci_prepare_native_host() and ata_pci_init_one() now doesn't take n_ports argument. All info should be specified via port_info array. Always two ports are allocated. * ata_pci_init_bmdma() exported to be used by LLDs in exotic cases. * port_info handling in all LLDs are standardized - all port_info arrays are const stack variable named ppi. Unless the second port is different from the first, its port_info is specified as NULL (tells libata that it's identical to the last non-NULL port_info). * pata_hpt37x/hpt3x2n: don't modify static variable directly. Make an on-stack copy instead as ata_piix does. * pata_uli: It has 4 ports instead of 2. Don't use ata_pci_prepare_native_host(). Allocate the host explicitly and use init helpers. It's simple enough. Signed-off-by: Tejun Heo <htejun@gmail.com> Signed-off-by: Jeff Garzik <jeff@garzik.org>
2007-05-04 18:43:58 +08:00
ppi[0] = &info_early;
} else if (pdev->revision < 0xC2) {
libata: clean up SFF init mess The intention of using port_mask in SFF init helpers was to eventually support exoctic configurations such as combination of legacy and native port on the same controller. This never became actually necessary and the related code always has been subtly broken one way or the other. Now that new init model is in place, there is no reason to make common helpers capable of handling all corner cases. Exotic cases can simply dealt within LLDs as necessary. This patch removes port_mask handling in SFF init helpers. SFF init helpers don't take n_ports argument and interpret it into port_mask anymore. All information is carried via port_info. n_ports argument is dropped and always two ports are allocated. LLD can tell SFF to skip certain port by marking it dummy. Note that SFF code has been treating unuvailable ports this way for a long time until recent breakage fix from Linus and is consistent with how other drivers handle with unavailable ports. This fixes 1-port legacy host handling still broken after the recent native mode fix and simplifies SFF init logic. The following changes are made... * ata_pci_init_native_host() and ata_init_legacy_host() both now try to initialized whatever they can and mark failed ports dummy. They return 0 if any port is successfully initialized. * ata_pci_prepare_native_host() and ata_pci_init_one() now doesn't take n_ports argument. All info should be specified via port_info array. Always two ports are allocated. * ata_pci_init_bmdma() exported to be used by LLDs in exotic cases. * port_info handling in all LLDs are standardized - all port_info arrays are const stack variable named ppi. Unless the second port is different from the first, its port_info is specified as NULL (tells libata that it's identical to the last non-NULL port_info). * pata_hpt37x/hpt3x2n: don't modify static variable directly. Make an on-stack copy instead as ata_piix does. * pata_uli: It has 4 ports instead of 2. Don't use ata_pci_prepare_native_host(). Allocate the host explicitly and use init helpers. It's simple enough. Signed-off-by: Tejun Heo <htejun@gmail.com> Signed-off-by: Jeff Garzik <jeff@garzik.org>
2007-05-04 18:43:58 +08:00
ppi[0] = &info_20;
} else if (pdev->revision == 0xC2) {
libata: clean up SFF init mess The intention of using port_mask in SFF init helpers was to eventually support exoctic configurations such as combination of legacy and native port on the same controller. This never became actually necessary and the related code always has been subtly broken one way or the other. Now that new init model is in place, there is no reason to make common helpers capable of handling all corner cases. Exotic cases can simply dealt within LLDs as necessary. This patch removes port_mask handling in SFF init helpers. SFF init helpers don't take n_ports argument and interpret it into port_mask anymore. All information is carried via port_info. n_ports argument is dropped and always two ports are allocated. LLD can tell SFF to skip certain port by marking it dummy. Note that SFF code has been treating unuvailable ports this way for a long time until recent breakage fix from Linus and is consistent with how other drivers handle with unavailable ports. This fixes 1-port legacy host handling still broken after the recent native mode fix and simplifies SFF init logic. The following changes are made... * ata_pci_init_native_host() and ata_init_legacy_host() both now try to initialized whatever they can and mark failed ports dummy. They return 0 if any port is successfully initialized. * ata_pci_prepare_native_host() and ata_pci_init_one() now doesn't take n_ports argument. All info should be specified via port_info array. Always two ports are allocated. * ata_pci_init_bmdma() exported to be used by LLDs in exotic cases. * port_info handling in all LLDs are standardized - all port_info arrays are const stack variable named ppi. Unless the second port is different from the first, its port_info is specified as NULL (tells libata that it's identical to the last non-NULL port_info). * pata_hpt37x/hpt3x2n: don't modify static variable directly. Make an on-stack copy instead as ata_piix does. * pata_uli: It has 4 ports instead of 2. Don't use ata_pci_prepare_native_host(). Allocate the host explicitly and use init helpers. It's simple enough. Signed-off-by: Tejun Heo <htejun@gmail.com> Signed-off-by: Jeff Garzik <jeff@garzik.org>
2007-05-04 18:43:58 +08:00
ppi[0] = &info_c2;
} else if (pdev->revision == 0xC3) {
libata: clean up SFF init mess The intention of using port_mask in SFF init helpers was to eventually support exoctic configurations such as combination of legacy and native port on the same controller. This never became actually necessary and the related code always has been subtly broken one way or the other. Now that new init model is in place, there is no reason to make common helpers capable of handling all corner cases. Exotic cases can simply dealt within LLDs as necessary. This patch removes port_mask handling in SFF init helpers. SFF init helpers don't take n_ports argument and interpret it into port_mask anymore. All information is carried via port_info. n_ports argument is dropped and always two ports are allocated. LLD can tell SFF to skip certain port by marking it dummy. Note that SFF code has been treating unuvailable ports this way for a long time until recent breakage fix from Linus and is consistent with how other drivers handle with unavailable ports. This fixes 1-port legacy host handling still broken after the recent native mode fix and simplifies SFF init logic. The following changes are made... * ata_pci_init_native_host() and ata_init_legacy_host() both now try to initialized whatever they can and mark failed ports dummy. They return 0 if any port is successfully initialized. * ata_pci_prepare_native_host() and ata_pci_init_one() now doesn't take n_ports argument. All info should be specified via port_info array. Always two ports are allocated. * ata_pci_init_bmdma() exported to be used by LLDs in exotic cases. * port_info handling in all LLDs are standardized - all port_info arrays are const stack variable named ppi. Unless the second port is different from the first, its port_info is specified as NULL (tells libata that it's identical to the last non-NULL port_info). * pata_hpt37x/hpt3x2n: don't modify static variable directly. Make an on-stack copy instead as ata_piix does. * pata_uli: It has 4 ports instead of 2. Don't use ata_pci_prepare_native_host(). Allocate the host explicitly and use init helpers. It's simple enough. Signed-off-by: Tejun Heo <htejun@gmail.com> Signed-off-by: Jeff Garzik <jeff@garzik.org>
2007-05-04 18:43:58 +08:00
ppi[0] = &info_c3;
} else if (pdev->revision == 0xC4) {
libata: clean up SFF init mess The intention of using port_mask in SFF init helpers was to eventually support exoctic configurations such as combination of legacy and native port on the same controller. This never became actually necessary and the related code always has been subtly broken one way or the other. Now that new init model is in place, there is no reason to make common helpers capable of handling all corner cases. Exotic cases can simply dealt within LLDs as necessary. This patch removes port_mask handling in SFF init helpers. SFF init helpers don't take n_ports argument and interpret it into port_mask anymore. All information is carried via port_info. n_ports argument is dropped and always two ports are allocated. LLD can tell SFF to skip certain port by marking it dummy. Note that SFF code has been treating unuvailable ports this way for a long time until recent breakage fix from Linus and is consistent with how other drivers handle with unavailable ports. This fixes 1-port legacy host handling still broken after the recent native mode fix and simplifies SFF init logic. The following changes are made... * ata_pci_init_native_host() and ata_init_legacy_host() both now try to initialized whatever they can and mark failed ports dummy. They return 0 if any port is successfully initialized. * ata_pci_prepare_native_host() and ata_pci_init_one() now doesn't take n_ports argument. All info should be specified via port_info array. Always two ports are allocated. * ata_pci_init_bmdma() exported to be used by LLDs in exotic cases. * port_info handling in all LLDs are standardized - all port_info arrays are const stack variable named ppi. Unless the second port is different from the first, its port_info is specified as NULL (tells libata that it's identical to the last non-NULL port_info). * pata_hpt37x/hpt3x2n: don't modify static variable directly. Make an on-stack copy instead as ata_piix does. * pata_uli: It has 4 ports instead of 2. Don't use ata_pci_prepare_native_host(). Allocate the host explicitly and use init helpers. It's simple enough. Signed-off-by: Tejun Heo <htejun@gmail.com> Signed-off-by: Jeff Garzik <jeff@garzik.org>
2007-05-04 18:43:58 +08:00
ppi[0] = &info_c4;
} else
libata: clean up SFF init mess The intention of using port_mask in SFF init helpers was to eventually support exoctic configurations such as combination of legacy and native port on the same controller. This never became actually necessary and the related code always has been subtly broken one way or the other. Now that new init model is in place, there is no reason to make common helpers capable of handling all corner cases. Exotic cases can simply dealt within LLDs as necessary. This patch removes port_mask handling in SFF init helpers. SFF init helpers don't take n_ports argument and interpret it into port_mask anymore. All information is carried via port_info. n_ports argument is dropped and always two ports are allocated. LLD can tell SFF to skip certain port by marking it dummy. Note that SFF code has been treating unuvailable ports this way for a long time until recent breakage fix from Linus and is consistent with how other drivers handle with unavailable ports. This fixes 1-port legacy host handling still broken after the recent native mode fix and simplifies SFF init logic. The following changes are made... * ata_pci_init_native_host() and ata_init_legacy_host() both now try to initialized whatever they can and mark failed ports dummy. They return 0 if any port is successfully initialized. * ata_pci_prepare_native_host() and ata_pci_init_one() now doesn't take n_ports argument. All info should be specified via port_info array. Always two ports are allocated. * ata_pci_init_bmdma() exported to be used by LLDs in exotic cases. * port_info handling in all LLDs are standardized - all port_info arrays are const stack variable named ppi. Unless the second port is different from the first, its port_info is specified as NULL (tells libata that it's identical to the last non-NULL port_info). * pata_hpt37x/hpt3x2n: don't modify static variable directly. Make an on-stack copy instead as ata_piix does. * pata_uli: It has 4 ports instead of 2. Don't use ata_pci_prepare_native_host(). Allocate the host explicitly and use init helpers. It's simple enough. Signed-off-by: Tejun Heo <htejun@gmail.com> Signed-off-by: Jeff Garzik <jeff@garzik.org>
2007-05-04 18:43:58 +08:00
ppi[0] = &info_c5;
ali_init_chipset(pdev);
if (ali_isa_bridge && pdev->revision >= 0x20 && pdev->revision < 0xC2) {
/* Are we paired with a UDMA capable chip */
pci_read_config_byte(ali_isa_bridge, 0x5E, &tmp);
if ((tmp & 0x1E) == 0x12)
libata: clean up SFF init mess The intention of using port_mask in SFF init helpers was to eventually support exoctic configurations such as combination of legacy and native port on the same controller. This never became actually necessary and the related code always has been subtly broken one way or the other. Now that new init model is in place, there is no reason to make common helpers capable of handling all corner cases. Exotic cases can simply dealt within LLDs as necessary. This patch removes port_mask handling in SFF init helpers. SFF init helpers don't take n_ports argument and interpret it into port_mask anymore. All information is carried via port_info. n_ports argument is dropped and always two ports are allocated. LLD can tell SFF to skip certain port by marking it dummy. Note that SFF code has been treating unuvailable ports this way for a long time until recent breakage fix from Linus and is consistent with how other drivers handle with unavailable ports. This fixes 1-port legacy host handling still broken after the recent native mode fix and simplifies SFF init logic. The following changes are made... * ata_pci_init_native_host() and ata_init_legacy_host() both now try to initialized whatever they can and mark failed ports dummy. They return 0 if any port is successfully initialized. * ata_pci_prepare_native_host() and ata_pci_init_one() now doesn't take n_ports argument. All info should be specified via port_info array. Always two ports are allocated. * ata_pci_init_bmdma() exported to be used by LLDs in exotic cases. * port_info handling in all LLDs are standardized - all port_info arrays are const stack variable named ppi. Unless the second port is different from the first, its port_info is specified as NULL (tells libata that it's identical to the last non-NULL port_info). * pata_hpt37x/hpt3x2n: don't modify static variable directly. Make an on-stack copy instead as ata_piix does. * pata_uli: It has 4 ports instead of 2. Don't use ata_pci_prepare_native_host(). Allocate the host explicitly and use init helpers. It's simple enough. Signed-off-by: Tejun Heo <htejun@gmail.com> Signed-off-by: Jeff Garzik <jeff@garzik.org>
2007-05-04 18:43:58 +08:00
ppi[0] = &info_20_udma;
}
if (!ppi[0]->mwdma_mask && !ppi[0]->udma_mask)
return ata_pci_sff_init_one(pdev, ppi, &ali_sht, NULL, 0);
else
return ata_pci_bmdma_init_one(pdev, ppi, &ali_sht, NULL, 0);
}
#ifdef CONFIG_PM_SLEEP
static int ali_reinit_one(struct pci_dev *pdev)
{
struct ata_host *host = pci_get_drvdata(pdev);
int rc;
rc = ata_pci_device_do_resume(pdev);
if (rc)
return rc;
ali_init_chipset(pdev);
ata_host_resume(host);
return 0;
}
#endif
static const struct pci_device_id ali[] = {
{ PCI_VDEVICE(AL, PCI_DEVICE_ID_AL_M5228), },
{ PCI_VDEVICE(AL, PCI_DEVICE_ID_AL_M5229), },
{ },
};
static struct pci_driver ali_pci_driver = {
.name = DRV_NAME,
.id_table = ali,
.probe = ali_init_one,
.remove = ata_pci_remove_one,
#ifdef CONFIG_PM_SLEEP
.suspend = ata_pci_device_suspend,
.resume = ali_reinit_one,
#endif
};
static int __init ali_init(void)
{
int ret;
ali_isa_bridge = pci_get_device(PCI_VENDOR_ID_AL, PCI_DEVICE_ID_AL_M1533, NULL);
ret = pci_register_driver(&ali_pci_driver);
if (ret < 0)
pci_dev_put(ali_isa_bridge);
return ret;
}
static void __exit ali_exit(void)
{
pci_unregister_driver(&ali_pci_driver);
pci_dev_put(ali_isa_bridge);
}
MODULE_AUTHOR("Alan Cox");
MODULE_DESCRIPTION("low-level driver for ALi PATA");
MODULE_LICENSE("GPL");
MODULE_DEVICE_TABLE(pci, ali);
MODULE_VERSION(DRV_VERSION);
module_init(ali_init);
module_exit(ali_exit);