kernel-ark/drivers/macintosh/via-cuda.c

634 lines
16 KiB
C
Raw Normal View History

/*
* Device driver for the via-cuda on Apple Powermacs.
*
* The VIA (versatile interface adapter) interfaces to the CUDA,
* a 6805 microprocessor core which controls the ADB (Apple Desktop
* Bus) which connects to the keyboard and mouse. The CUDA also
* controls system power and the RTC (real time clock) chip.
*
* Copyright (C) 1996 Paul Mackerras.
*/
#include <stdarg.h>
#include <linux/types.h>
#include <linux/errno.h>
#include <linux/kernel.h>
#include <linux/delay.h>
#include <linux/adb.h>
#include <linux/cuda.h>
#include <linux/spinlock.h>
#include <linux/interrupt.h>
#ifdef CONFIG_PPC
#include <asm/prom.h>
#include <asm/machdep.h>
#else
#include <asm/macintosh.h>
#include <asm/macints.h>
#include <asm/machw.h>
#include <asm/mac_via.h>
#endif
#include <asm/io.h>
#include <asm/system.h>
#include <linux/init.h>
static volatile unsigned char __iomem *via;
static DEFINE_SPINLOCK(cuda_lock);
/* VIA registers - spaced 0x200 bytes apart */
#define RS 0x200 /* skip between registers */
#define B 0 /* B-side data */
#define A RS /* A-side data */
#define DIRB (2*RS) /* B-side direction (1=output) */
#define DIRA (3*RS) /* A-side direction (1=output) */
#define T1CL (4*RS) /* Timer 1 ctr/latch (low 8 bits) */
#define T1CH (5*RS) /* Timer 1 counter (high 8 bits) */
#define T1LL (6*RS) /* Timer 1 latch (low 8 bits) */
#define T1LH (7*RS) /* Timer 1 latch (high 8 bits) */
#define T2CL (8*RS) /* Timer 2 ctr/latch (low 8 bits) */
#define T2CH (9*RS) /* Timer 2 counter (high 8 bits) */
#define SR (10*RS) /* Shift register */
#define ACR (11*RS) /* Auxiliary control register */
#define PCR (12*RS) /* Peripheral control register */
#define IFR (13*RS) /* Interrupt flag register */
#define IER (14*RS) /* Interrupt enable register */
#define ANH (15*RS) /* A-side data, no handshake */
/* Bits in B data register: all active low */
#define TREQ 0x08 /* Transfer request (input) */
#define TACK 0x10 /* Transfer acknowledge (output) */
#define TIP 0x20 /* Transfer in progress (output) */
/* Bits in ACR */
#define SR_CTRL 0x1c /* Shift register control bits */
#define SR_EXT 0x0c /* Shift on external clock */
#define SR_OUT 0x10 /* Shift out if 1 */
/* Bits in IFR and IER */
#define IER_SET 0x80 /* set bits in IER */
#define IER_CLR 0 /* clear bits in IER */
#define SR_INT 0x04 /* Shift register full/empty */
static enum cuda_state {
idle,
sent_first_byte,
sending,
reading,
read_done,
awaiting_reply
} cuda_state;
static struct adb_request *current_req;
static struct adb_request *last_req;
static unsigned char cuda_rbuf[16];
static unsigned char *reply_ptr;
static int reading_reply;
static int data_index;
static int cuda_irq;
#ifdef CONFIG_PPC
static struct device_node *vias;
#endif
static int cuda_fully_inited;
#ifdef CONFIG_ADB
static int cuda_probe(void);
static int cuda_init(void);
static int cuda_send_request(struct adb_request *req, int sync);
static int cuda_adb_autopoll(int devs);
static int cuda_reset_adb_bus(void);
#endif /* CONFIG_ADB */
static int cuda_init_via(void);
static void cuda_start(void);
IRQ: Maintain regs pointer globally rather than passing to IRQ handlers Maintain a per-CPU global "struct pt_regs *" variable which can be used instead of passing regs around manually through all ~1800 interrupt handlers in the Linux kernel. The regs pointer is used in few places, but it potentially costs both stack space and code to pass it around. On the FRV arch, removing the regs parameter from all the genirq function results in a 20% speed up of the IRQ exit path (ie: from leaving timer_interrupt() to leaving do_IRQ()). Where appropriate, an arch may override the generic storage facility and do something different with the variable. On FRV, for instance, the address is maintained in GR28 at all times inside the kernel as part of general exception handling. Having looked over the code, it appears that the parameter may be handed down through up to twenty or so layers of functions. Consider a USB character device attached to a USB hub, attached to a USB controller that posts its interrupts through a cascaded auxiliary interrupt controller. A character device driver may want to pass regs to the sysrq handler through the input layer which adds another few layers of parameter passing. I've build this code with allyesconfig for x86_64 and i386. I've runtested the main part of the code on FRV and i386, though I can't test most of the drivers. I've also done partial conversion for powerpc and MIPS - these at least compile with minimal configurations. This will affect all archs. Mostly the changes should be relatively easy. Take do_IRQ(), store the regs pointer at the beginning, saving the old one: struct pt_regs *old_regs = set_irq_regs(regs); And put the old one back at the end: set_irq_regs(old_regs); Don't pass regs through to generic_handle_irq() or __do_IRQ(). In timer_interrupt(), this sort of change will be necessary: - update_process_times(user_mode(regs)); - profile_tick(CPU_PROFILING, regs); + update_process_times(user_mode(get_irq_regs())); + profile_tick(CPU_PROFILING); I'd like to move update_process_times()'s use of get_irq_regs() into itself, except that i386, alone of the archs, uses something other than user_mode(). Some notes on the interrupt handling in the drivers: (*) input_dev() is now gone entirely. The regs pointer is no longer stored in the input_dev struct. (*) finish_unlinks() in drivers/usb/host/ohci-q.c needs checking. It does something different depending on whether it's been supplied with a regs pointer or not. (*) Various IRQ handler function pointers have been moved to type irq_handler_t. Signed-Off-By: David Howells <dhowells@redhat.com> (cherry picked from 1b16e7ac850969f38b375e511e3fa2f474a33867 commit)
2006-10-05 13:55:46 +00:00
static irqreturn_t cuda_interrupt(int irq, void *arg);
static void cuda_input(unsigned char *buf, int nb);
void cuda_poll(void);
static int cuda_write(struct adb_request *req);
int cuda_request(struct adb_request *req,
void (*done)(struct adb_request *), int nbytes, ...);
#ifdef CONFIG_ADB
struct adb_driver via_cuda_driver = {
"CUDA",
cuda_probe,
cuda_init,
cuda_send_request,
cuda_adb_autopoll,
cuda_poll,
cuda_reset_adb_bus
};
#endif /* CONFIG_ADB */
#ifdef CONFIG_PPC
int __init find_via_cuda(void)
{
struct adb_request req;
phys_addr_t taddr;
const u32 *reg;
int err;
if (vias != 0)
return 1;
vias = of_find_node_by_name(NULL, "via-cuda");
if (vias == 0)
return 0;
reg = of_get_property(vias, "reg", NULL);
if (reg == NULL) {
printk(KERN_ERR "via-cuda: No \"reg\" property !\n");
goto fail;
}
taddr = of_translate_address(vias, reg);
if (taddr == 0) {
printk(KERN_ERR "via-cuda: Can't translate address !\n");
goto fail;
}
via = ioremap(taddr, 0x2000);
if (via == NULL) {
printk(KERN_ERR "via-cuda: Can't map address !\n");
goto fail;
}
cuda_state = idle;
sys_ctrler = SYS_CTRLER_CUDA;
err = cuda_init_via();
if (err) {
printk(KERN_ERR "cuda_init_via() failed\n");
via = NULL;
return 0;
}
/* Clear and enable interrupts, but only on PPC. On 68K it's done */
/* for us by the main VIA driver in arch/m68k/mac/via.c */
out_8(&via[IFR], 0x7f); /* clear interrupts by writing 1s */
out_8(&via[IER], IER_SET|SR_INT); /* enable interrupt from SR */
/* enable autopoll */
cuda_request(&req, NULL, 3, CUDA_PACKET, CUDA_AUTOPOLL, 1);
while (!req.complete)
cuda_poll();
return 1;
fail:
of_node_put(vias);
vias = NULL;
return 0;
}
#endif /* CONFIG_PPC */
static int __init via_cuda_start(void)
{
if (via == NULL)
return -ENODEV;
2006-07-03 11:36:01 +00:00
#ifdef CONFIG_MAC
cuda_irq = IRQ_MAC_ADB;
2006-07-03 11:36:01 +00:00
#else /* CONFIG_MAC */
cuda_irq = irq_of_parse_and_map(vias, 0);
if (cuda_irq == NO_IRQ) {
2006-07-03 11:36:01 +00:00
printk(KERN_ERR "via-cuda: can't map interrupts for %s\n",
vias->full_name);
return -ENODEV;
}
#endif /* CONFIG_MAC */
2006-07-03 11:36:01 +00:00
if (request_irq(cuda_irq, cuda_interrupt, 0, "ADB", cuda_interrupt)) {
printk(KERN_ERR "via-cuda: can't request irq %d\n", cuda_irq);
return -EAGAIN;
}
printk("Macintosh CUDA driver v0.5 for Unified ADB.\n");
cuda_fully_inited = 1;
return 0;
}
device_initcall(via_cuda_start);
#ifdef CONFIG_ADB
static int
cuda_probe(void)
{
#ifdef CONFIG_PPC
if (sys_ctrler != SYS_CTRLER_CUDA)
return -ENODEV;
#else
if (macintosh_config->adb_type != MAC_ADB_CUDA)
return -ENODEV;
via = via1;
#endif
return 0;
}
static int __init
cuda_init(void)
{
#ifdef CONFIG_PPC
if (via == NULL)
return -ENODEV;
return 0;
#else
int err = cuda_init_via();
if (err) {
printk(KERN_ERR "cuda_init_via() failed\n");
return -ENODEV;
}
out_8(&via[IER], IER_SET|SR_INT); /* enable interrupt from SR */
return via_cuda_start();
#endif
}
#endif /* CONFIG_ADB */
#define WAIT_FOR(cond, what) \
do { \
int x; \
for (x = 1000; !(cond); --x) { \
if (x == 0) { \
printk("Timeout waiting for " what "\n"); \
return -ENXIO; \
} \
udelay(100); \
} \
} while (0)
static int
cuda_init_via(void)
{
out_8(&via[DIRB], (in_8(&via[DIRB]) | TACK | TIP) & ~TREQ); /* TACK & TIP out */
out_8(&via[B], in_8(&via[B]) | TACK | TIP); /* negate them */
out_8(&via[ACR] ,(in_8(&via[ACR]) & ~SR_CTRL) | SR_EXT); /* SR data in */
(void)in_8(&via[SR]); /* clear any left-over data */
#ifdef CONFIG_PPC
out_8(&via[IER], 0x7f); /* disable interrupts from VIA */
(void)in_8(&via[IER]);
#else
out_8(&via[IER], SR_INT); /* disable SR interrupt from VIA */
#endif
/* delay 4ms and then clear any pending interrupt */
mdelay(4);
(void)in_8(&via[SR]);
out_8(&via[IFR], SR_INT);
/* sync with the CUDA - assert TACK without TIP */
out_8(&via[B], in_8(&via[B]) & ~TACK);
/* wait for the CUDA to assert TREQ in response */
WAIT_FOR((in_8(&via[B]) & TREQ) == 0, "CUDA response to sync");
/* wait for the interrupt and then clear it */
WAIT_FOR(in_8(&via[IFR]) & SR_INT, "CUDA response to sync (2)");
(void)in_8(&via[SR]);
out_8(&via[IFR], SR_INT);
/* finish the sync by negating TACK */
out_8(&via[B], in_8(&via[B]) | TACK);
/* wait for the CUDA to negate TREQ and the corresponding interrupt */
WAIT_FOR(in_8(&via[B]) & TREQ, "CUDA response to sync (3)");
WAIT_FOR(in_8(&via[IFR]) & SR_INT, "CUDA response to sync (4)");
(void)in_8(&via[SR]);
out_8(&via[IFR], SR_INT);
out_8(&via[B], in_8(&via[B]) | TIP); /* should be unnecessary */
return 0;
}
#ifdef CONFIG_ADB
/* Send an ADB command */
static int
cuda_send_request(struct adb_request *req, int sync)
{
int i;
if ((via == NULL) || !cuda_fully_inited) {
req->complete = 1;
return -ENXIO;
}
req->reply_expected = 1;
i = cuda_write(req);
if (i)
return i;
if (sync) {
while (!req->complete)
cuda_poll();
}
return 0;
}
/* Enable/disable autopolling */
static int
cuda_adb_autopoll(int devs)
{
struct adb_request req;
if ((via == NULL) || !cuda_fully_inited)
return -ENXIO;
cuda_request(&req, NULL, 3, CUDA_PACKET, CUDA_AUTOPOLL, (devs? 1: 0));
while (!req.complete)
cuda_poll();
return 0;
}
/* Reset adb bus - how do we do this?? */
static int
cuda_reset_adb_bus(void)
{
struct adb_request req;
if ((via == NULL) || !cuda_fully_inited)
return -ENXIO;
cuda_request(&req, NULL, 2, ADB_PACKET, 0); /* maybe? */
while (!req.complete)
cuda_poll();
return 0;
}
#endif /* CONFIG_ADB */
/* Construct and send a cuda request */
int
cuda_request(struct adb_request *req, void (*done)(struct adb_request *),
int nbytes, ...)
{
va_list list;
int i;
if (via == NULL) {
req->complete = 1;
return -ENXIO;
}
req->nbytes = nbytes;
req->done = done;
va_start(list, nbytes);
for (i = 0; i < nbytes; ++i)
req->data[i] = va_arg(list, int);
va_end(list);
req->reply_expected = 1;
return cuda_write(req);
}
static int
cuda_write(struct adb_request *req)
{
unsigned long flags;
if (req->nbytes < 2 || req->data[0] > CUDA_PACKET) {
req->complete = 1;
return -EINVAL;
}
req->next = NULL;
req->sent = 0;
req->complete = 0;
req->reply_len = 0;
spin_lock_irqsave(&cuda_lock, flags);
if (current_req != 0) {
last_req->next = req;
last_req = req;
} else {
current_req = req;
last_req = req;
if (cuda_state == idle)
cuda_start();
}
spin_unlock_irqrestore(&cuda_lock, flags);
return 0;
}
static void
cuda_start(void)
{
struct adb_request *req;
/* assert cuda_state == idle */
/* get the packet to send */
req = current_req;
if (req == 0)
return;
if ((in_8(&via[B]) & TREQ) == 0)
return; /* a byte is coming in from the CUDA */
/* set the shift register to shift out and send a byte */
out_8(&via[ACR], in_8(&via[ACR]) | SR_OUT);
out_8(&via[SR], req->data[0]);
out_8(&via[B], in_8(&via[B]) & ~TIP);
cuda_state = sent_first_byte;
}
void
cuda_poll(void)
{
/* cuda_interrupt only takes a normal lock, we disable
* interrupts here to avoid re-entering and thus deadlocking.
*/
disable_irq(cuda_irq);
cuda_interrupt(0, NULL);
enable_irq(cuda_irq);
}
static irqreturn_t
IRQ: Maintain regs pointer globally rather than passing to IRQ handlers Maintain a per-CPU global "struct pt_regs *" variable which can be used instead of passing regs around manually through all ~1800 interrupt handlers in the Linux kernel. The regs pointer is used in few places, but it potentially costs both stack space and code to pass it around. On the FRV arch, removing the regs parameter from all the genirq function results in a 20% speed up of the IRQ exit path (ie: from leaving timer_interrupt() to leaving do_IRQ()). Where appropriate, an arch may override the generic storage facility and do something different with the variable. On FRV, for instance, the address is maintained in GR28 at all times inside the kernel as part of general exception handling. Having looked over the code, it appears that the parameter may be handed down through up to twenty or so layers of functions. Consider a USB character device attached to a USB hub, attached to a USB controller that posts its interrupts through a cascaded auxiliary interrupt controller. A character device driver may want to pass regs to the sysrq handler through the input layer which adds another few layers of parameter passing. I've build this code with allyesconfig for x86_64 and i386. I've runtested the main part of the code on FRV and i386, though I can't test most of the drivers. I've also done partial conversion for powerpc and MIPS - these at least compile with minimal configurations. This will affect all archs. Mostly the changes should be relatively easy. Take do_IRQ(), store the regs pointer at the beginning, saving the old one: struct pt_regs *old_regs = set_irq_regs(regs); And put the old one back at the end: set_irq_regs(old_regs); Don't pass regs through to generic_handle_irq() or __do_IRQ(). In timer_interrupt(), this sort of change will be necessary: - update_process_times(user_mode(regs)); - profile_tick(CPU_PROFILING, regs); + update_process_times(user_mode(get_irq_regs())); + profile_tick(CPU_PROFILING); I'd like to move update_process_times()'s use of get_irq_regs() into itself, except that i386, alone of the archs, uses something other than user_mode(). Some notes on the interrupt handling in the drivers: (*) input_dev() is now gone entirely. The regs pointer is no longer stored in the input_dev struct. (*) finish_unlinks() in drivers/usb/host/ohci-q.c needs checking. It does something different depending on whether it's been supplied with a regs pointer or not. (*) Various IRQ handler function pointers have been moved to type irq_handler_t. Signed-Off-By: David Howells <dhowells@redhat.com> (cherry picked from 1b16e7ac850969f38b375e511e3fa2f474a33867 commit)
2006-10-05 13:55:46 +00:00
cuda_interrupt(int irq, void *arg)
{
int status;
struct adb_request *req = NULL;
unsigned char ibuf[16];
int ibuf_len = 0;
int complete = 0;
spin_lock(&cuda_lock);
/* On powermacs, this handler is registered for the VIA IRQ. But it uses
* just the shift register IRQ -- other VIA interrupt sources are disabled.
* On m68k macs, the VIA IRQ sources are dispatched individually. Unless
* we are polling, the shift register IRQ flag has already been cleared.
*/
#ifdef CONFIG_MAC
if (!arg)
#endif
{
if ((in_8(&via[IFR]) & SR_INT) == 0) {
spin_unlock(&cuda_lock);
return IRQ_NONE;
} else {
out_8(&via[IFR], SR_INT);
}
}
status = (~in_8(&via[B]) & (TIP|TREQ)) | (in_8(&via[ACR]) & SR_OUT);
/* printk("cuda_interrupt: state=%d status=%x\n", cuda_state, status); */
switch (cuda_state) {
case idle:
/* CUDA has sent us the first byte of data - unsolicited */
if (status != TREQ)
printk("cuda: state=idle, status=%x\n", status);
(void)in_8(&via[SR]);
out_8(&via[B], in_8(&via[B]) & ~TIP);
cuda_state = reading;
reply_ptr = cuda_rbuf;
reading_reply = 0;
break;
case awaiting_reply:
/* CUDA has sent us the first byte of data of a reply */
if (status != TREQ)
printk("cuda: state=awaiting_reply, status=%x\n", status);
(void)in_8(&via[SR]);
out_8(&via[B], in_8(&via[B]) & ~TIP);
cuda_state = reading;
reply_ptr = current_req->reply;
reading_reply = 1;
break;
case sent_first_byte:
if (status == TREQ + TIP + SR_OUT) {
/* collision */
out_8(&via[ACR], in_8(&via[ACR]) & ~SR_OUT);
(void)in_8(&via[SR]);
out_8(&via[B], in_8(&via[B]) | TIP | TACK);
cuda_state = idle;
} else {
/* assert status == TIP + SR_OUT */
if (status != TIP + SR_OUT)
printk("cuda: state=sent_first_byte status=%x\n", status);
out_8(&via[SR], current_req->data[1]);
out_8(&via[B], in_8(&via[B]) ^ TACK);
data_index = 2;
cuda_state = sending;
}
break;
case sending:
req = current_req;
if (data_index >= req->nbytes) {
out_8(&via[ACR], in_8(&via[ACR]) & ~SR_OUT);
(void)in_8(&via[SR]);
out_8(&via[B], in_8(&via[B]) | TACK | TIP);
req->sent = 1;
if (req->reply_expected) {
cuda_state = awaiting_reply;
} else {
current_req = req->next;
complete = 1;
/* not sure about this */
cuda_state = idle;
cuda_start();
}
} else {
out_8(&via[SR], req->data[data_index++]);
out_8(&via[B], in_8(&via[B]) ^ TACK);
}
break;
case reading:
*reply_ptr++ = in_8(&via[SR]);
if (status == TIP) {
/* that's all folks */
out_8(&via[B], in_8(&via[B]) | TACK | TIP);
cuda_state = read_done;
} else {
/* assert status == TIP | TREQ */
if (status != TIP + TREQ)
printk("cuda: state=reading status=%x\n", status);
out_8(&via[B], in_8(&via[B]) ^ TACK);
}
break;
case read_done:
(void)in_8(&via[SR]);
if (reading_reply) {
req = current_req;
req->reply_len = reply_ptr - req->reply;
if (req->data[0] == ADB_PACKET) {
/* Have to adjust the reply from ADB commands */
if (req->reply_len <= 2 || (req->reply[1] & 2) != 0) {
/* the 0x2 bit indicates no response */
req->reply_len = 0;
} else {
/* leave just the command and result bytes in the reply */
req->reply_len -= 2;
memmove(req->reply, req->reply + 2, req->reply_len);
}
}
current_req = req->next;
complete = 1;
} else {
/* This is tricky. We must break the spinlock to call
* cuda_input. However, doing so means we might get
* re-entered from another CPU getting an interrupt
* or calling cuda_poll(). I ended up using the stack
* (it's only for 16 bytes) and moving the actual
* call to cuda_input to outside of the lock.
*/
ibuf_len = reply_ptr - cuda_rbuf;
memcpy(ibuf, cuda_rbuf, ibuf_len);
}
if (status == TREQ) {
out_8(&via[B], in_8(&via[B]) & ~TIP);
cuda_state = reading;
reply_ptr = cuda_rbuf;
reading_reply = 0;
} else {
cuda_state = idle;
cuda_start();
}
break;
default:
printk("cuda_interrupt: unknown cuda_state %d?\n", cuda_state);
}
spin_unlock(&cuda_lock);
if (complete && req) {
void (*done)(struct adb_request *) = req->done;
mb();
req->complete = 1;
/* Here, we assume that if the request has a done member, the
* struct request will survive to setting req->complete to 1
*/
if (done)
(*done)(req);
}
if (ibuf_len)
IRQ: Maintain regs pointer globally rather than passing to IRQ handlers Maintain a per-CPU global "struct pt_regs *" variable which can be used instead of passing regs around manually through all ~1800 interrupt handlers in the Linux kernel. The regs pointer is used in few places, but it potentially costs both stack space and code to pass it around. On the FRV arch, removing the regs parameter from all the genirq function results in a 20% speed up of the IRQ exit path (ie: from leaving timer_interrupt() to leaving do_IRQ()). Where appropriate, an arch may override the generic storage facility and do something different with the variable. On FRV, for instance, the address is maintained in GR28 at all times inside the kernel as part of general exception handling. Having looked over the code, it appears that the parameter may be handed down through up to twenty or so layers of functions. Consider a USB character device attached to a USB hub, attached to a USB controller that posts its interrupts through a cascaded auxiliary interrupt controller. A character device driver may want to pass regs to the sysrq handler through the input layer which adds another few layers of parameter passing. I've build this code with allyesconfig for x86_64 and i386. I've runtested the main part of the code on FRV and i386, though I can't test most of the drivers. I've also done partial conversion for powerpc and MIPS - these at least compile with minimal configurations. This will affect all archs. Mostly the changes should be relatively easy. Take do_IRQ(), store the regs pointer at the beginning, saving the old one: struct pt_regs *old_regs = set_irq_regs(regs); And put the old one back at the end: set_irq_regs(old_regs); Don't pass regs through to generic_handle_irq() or __do_IRQ(). In timer_interrupt(), this sort of change will be necessary: - update_process_times(user_mode(regs)); - profile_tick(CPU_PROFILING, regs); + update_process_times(user_mode(get_irq_regs())); + profile_tick(CPU_PROFILING); I'd like to move update_process_times()'s use of get_irq_regs() into itself, except that i386, alone of the archs, uses something other than user_mode(). Some notes on the interrupt handling in the drivers: (*) input_dev() is now gone entirely. The regs pointer is no longer stored in the input_dev struct. (*) finish_unlinks() in drivers/usb/host/ohci-q.c needs checking. It does something different depending on whether it's been supplied with a regs pointer or not. (*) Various IRQ handler function pointers have been moved to type irq_handler_t. Signed-Off-By: David Howells <dhowells@redhat.com> (cherry picked from 1b16e7ac850969f38b375e511e3fa2f474a33867 commit)
2006-10-05 13:55:46 +00:00
cuda_input(ibuf, ibuf_len);
return IRQ_HANDLED;
}
static void
IRQ: Maintain regs pointer globally rather than passing to IRQ handlers Maintain a per-CPU global "struct pt_regs *" variable which can be used instead of passing regs around manually through all ~1800 interrupt handlers in the Linux kernel. The regs pointer is used in few places, but it potentially costs both stack space and code to pass it around. On the FRV arch, removing the regs parameter from all the genirq function results in a 20% speed up of the IRQ exit path (ie: from leaving timer_interrupt() to leaving do_IRQ()). Where appropriate, an arch may override the generic storage facility and do something different with the variable. On FRV, for instance, the address is maintained in GR28 at all times inside the kernel as part of general exception handling. Having looked over the code, it appears that the parameter may be handed down through up to twenty or so layers of functions. Consider a USB character device attached to a USB hub, attached to a USB controller that posts its interrupts through a cascaded auxiliary interrupt controller. A character device driver may want to pass regs to the sysrq handler through the input layer which adds another few layers of parameter passing. I've build this code with allyesconfig for x86_64 and i386. I've runtested the main part of the code on FRV and i386, though I can't test most of the drivers. I've also done partial conversion for powerpc and MIPS - these at least compile with minimal configurations. This will affect all archs. Mostly the changes should be relatively easy. Take do_IRQ(), store the regs pointer at the beginning, saving the old one: struct pt_regs *old_regs = set_irq_regs(regs); And put the old one back at the end: set_irq_regs(old_regs); Don't pass regs through to generic_handle_irq() or __do_IRQ(). In timer_interrupt(), this sort of change will be necessary: - update_process_times(user_mode(regs)); - profile_tick(CPU_PROFILING, regs); + update_process_times(user_mode(get_irq_regs())); + profile_tick(CPU_PROFILING); I'd like to move update_process_times()'s use of get_irq_regs() into itself, except that i386, alone of the archs, uses something other than user_mode(). Some notes on the interrupt handling in the drivers: (*) input_dev() is now gone entirely. The regs pointer is no longer stored in the input_dev struct. (*) finish_unlinks() in drivers/usb/host/ohci-q.c needs checking. It does something different depending on whether it's been supplied with a regs pointer or not. (*) Various IRQ handler function pointers have been moved to type irq_handler_t. Signed-Off-By: David Howells <dhowells@redhat.com> (cherry picked from 1b16e7ac850969f38b375e511e3fa2f474a33867 commit)
2006-10-05 13:55:46 +00:00
cuda_input(unsigned char *buf, int nb)
{
int i;
switch (buf[0]) {
case ADB_PACKET:
#ifdef CONFIG_XMON
if (nb == 5 && buf[2] == 0x2c) {
extern int xmon_wants_key, xmon_adb_keycode;
if (xmon_wants_key) {
xmon_adb_keycode = buf[3];
return;
}
}
#endif /* CONFIG_XMON */
#ifdef CONFIG_ADB
IRQ: Maintain regs pointer globally rather than passing to IRQ handlers Maintain a per-CPU global "struct pt_regs *" variable which can be used instead of passing regs around manually through all ~1800 interrupt handlers in the Linux kernel. The regs pointer is used in few places, but it potentially costs both stack space and code to pass it around. On the FRV arch, removing the regs parameter from all the genirq function results in a 20% speed up of the IRQ exit path (ie: from leaving timer_interrupt() to leaving do_IRQ()). Where appropriate, an arch may override the generic storage facility and do something different with the variable. On FRV, for instance, the address is maintained in GR28 at all times inside the kernel as part of general exception handling. Having looked over the code, it appears that the parameter may be handed down through up to twenty or so layers of functions. Consider a USB character device attached to a USB hub, attached to a USB controller that posts its interrupts through a cascaded auxiliary interrupt controller. A character device driver may want to pass regs to the sysrq handler through the input layer which adds another few layers of parameter passing. I've build this code with allyesconfig for x86_64 and i386. I've runtested the main part of the code on FRV and i386, though I can't test most of the drivers. I've also done partial conversion for powerpc and MIPS - these at least compile with minimal configurations. This will affect all archs. Mostly the changes should be relatively easy. Take do_IRQ(), store the regs pointer at the beginning, saving the old one: struct pt_regs *old_regs = set_irq_regs(regs); And put the old one back at the end: set_irq_regs(old_regs); Don't pass regs through to generic_handle_irq() or __do_IRQ(). In timer_interrupt(), this sort of change will be necessary: - update_process_times(user_mode(regs)); - profile_tick(CPU_PROFILING, regs); + update_process_times(user_mode(get_irq_regs())); + profile_tick(CPU_PROFILING); I'd like to move update_process_times()'s use of get_irq_regs() into itself, except that i386, alone of the archs, uses something other than user_mode(). Some notes on the interrupt handling in the drivers: (*) input_dev() is now gone entirely. The regs pointer is no longer stored in the input_dev struct. (*) finish_unlinks() in drivers/usb/host/ohci-q.c needs checking. It does something different depending on whether it's been supplied with a regs pointer or not. (*) Various IRQ handler function pointers have been moved to type irq_handler_t. Signed-Off-By: David Howells <dhowells@redhat.com> (cherry picked from 1b16e7ac850969f38b375e511e3fa2f474a33867 commit)
2006-10-05 13:55:46 +00:00
adb_input(buf+2, nb-2, buf[1] & 0x40);
#endif /* CONFIG_ADB */
break;
default:
printk("data from cuda (%d bytes):", nb);
for (i = 0; i < nb; ++i)
printk(" %.2x", buf[i]);
printk("\n");
}
}