uti: futex: cache remote va to remote pa result

Change-Id: Idbbb3f2981b76a0235615fceaa6281d2c7134ca2
This commit is contained in:
Masamichi Takagi
2021-02-04 09:25:32 +09:00
parent 5d16ce9dcc
commit adc5b7102f
9 changed files with 138 additions and 23 deletions

View File

@ -1301,6 +1301,7 @@ void mcctrl_put_per_proc_data(struct mcctrl_per_proc_data *ppd)
ihk_ikc_spinlock_unlock(&ppd->wq_list_lock, flags);
pager_remove_process(ppd);
futex_remove_process(ppd);
kfree(ppd);
}
@ -1901,6 +1902,7 @@ int mcexec_create_per_process_data(ihk_os_t os,
spin_lock_init(&ppd->wq_list_lock);
memset(&ppd->cpu_set, 0, sizeof(cpumask_t));
ppd->ikc_target_cpu = 0;
ppd->rva_to_rpa_cache = RB_ROOT;
/* Final ref will be dropped in release_handler() through
* mcexec_destroy_per_process_data() */
atomic_set(&ppd->refcount, 1);
@ -3014,6 +3016,8 @@ void mcctrl_futex_wake(struct ikc_scd_packet *pisp)
}
resp->done = 1;
dprintk("%s: cpu: %d\n", __func__, ihk_ikc_get_processor_id());
wake_up_interruptible(&resp->wq);
}

View File

@ -10,6 +10,7 @@
#include <linux/semaphore.h>
#include <linux/interrupt.h>
#include <linux/cpumask.h>
#include <linux/rbtree.h>
#include <asm/uaccess.h>
#include <asm/delay.h>
#include <asm/io.h>
@ -213,6 +214,76 @@ retry_alloc:
return error;
}
struct rva_to_rpa_cache_node {
struct rb_node node;
unsigned long rva;
unsigned long rpa;
};
void futex_remove_process(struct mcctrl_per_proc_data *ppd)
{
struct rb_node *node;
while ((node = rb_first(&ppd->rva_to_rpa_cache))) {
struct rva_to_rpa_cache_node *cache_node;
cache_node = container_of(node, struct rva_to_rpa_cache_node,
node);
rb_erase(node, &ppd->rva_to_rpa_cache);
kfree(cache_node);
}
}
struct rva_to_rpa_cache_node *rva_to_rpa_cache_search(struct rb_root *root,
unsigned long rva)
{
struct rb_node **iter = &root->rb_node, *parent = NULL;
while (*iter) {
struct rva_to_rpa_cache_node *inode =
container_of(*iter, struct rva_to_rpa_cache_node, node);
parent = *iter;
if (rva == inode->rva) {
return inode;
}
if (rva < inode->rva)
iter = &((*iter)->rb_left);
else
iter = &((*iter)->rb_right);
}
return NULL;
}
int rva_to_rpa_cache_insert(struct rb_root *root,
struct rva_to_rpa_cache_node *cache_node)
{
struct rb_node **iter = &root->rb_node, *parent = NULL;
while (*iter) {
struct rva_to_rpa_cache_node *inode =
container_of(*iter, struct rva_to_rpa_cache_node, node);
parent = *iter;
if (cache_node->rva == inode->rva)
return -EINVAL;
if (cache_node->rva < inode->rva)
iter = &((*iter)->rb_left);
else
iter = &((*iter)->rb_right);
}
rb_link_node(&cache_node->node, parent, iter);
rb_insert_color(&cache_node->node, root);
return 0;
}
/**
* get_futex_key() - Get parameters which are the keys for a futex
* @uaddr: virtual address of the futex
@ -238,6 +309,7 @@ get_futex_key(uint32_t *uaddr, int fshared, union futex_key *key,
struct mcctrl_usrdata *usrdata;
struct mcctrl_per_proc_data *ppd;
int ret = 0, error = 0;
struct rva_to_rpa_cache_node *cache_node;
/*
* The futex address must be "naturally" aligned.
@ -281,6 +353,17 @@ get_futex_key(uint32_t *uaddr, int fshared, union futex_key *key,
goto out;
}
/* cache because translate_rva_to_rpa calls smp_ihk_arch_dcache_flush
* via ihk_device_unmap_virtual
*/
cache_node = rva_to_rpa_cache_search(&ppd->rva_to_rpa_cache,
(unsigned long)uaddr);
if (cache_node) {
phys = cache_node->rpa;
dprintk("%s: cache hit, rva: %lx, rpa: %lx\n",
__func__, (unsigned long)uaddr, phys);
goto found;
}
retry_v2p:
error = translate_rva_to_rpa((ihk_os_t)uti_info->os, ppd->rpgtable,
(unsigned long)uaddr, &phys, &pgsize);
@ -299,6 +382,23 @@ retry_v2p:
goto retry_v2p;
}
cache_node = kmalloc(sizeof(struct rva_to_rpa_cache_node), GFP_KERNEL);
if (!cache_node) {
ret = -ENOMEM;
goto put_out;
}
cache_node->rva = (unsigned long)uaddr;
cache_node->rpa = phys;
dprintk("%s: cache insert, rva: %lx, rpa: %lx\n",
__func__, (unsigned long)uaddr, phys);
ret = rva_to_rpa_cache_insert(&ppd->rva_to_rpa_cache, cache_node);
if (ret) {
pr_err("%s: error: cache entry found, rva: %lx, rpa: %lx\n",
__func__, (unsigned long)uaddr, phys);
goto put_out;
}
found:
key->shared.phys = (void *)phys;
key->shared.pgoff = 0;
@ -444,8 +544,7 @@ static int uti_sched_wakeup_thread(struct futex_q *q, int valid_states,
irqstate = ihk_mc_spinlock_lock(
(_ihk_spinlock_t *)q->th_spin_sleep_lock);
if (*(int *)q->th_spin_sleep == 1) {
dprintk("%s: spin wakeup: cpu_id: %d\n",
__func__, uti_info->cpu);
dprintk("%s: spin wakeup: cpu_id: %d\n", __func__, uti_info->cpu);
status = 0;
}
*(int *)q->th_spin_sleep = 0;
@ -477,11 +576,11 @@ static int uti_sched_wakeup_thread(struct futex_q *q, int valid_states,
ihk_mc_spinlock_unlock((_ihk_spinlock_t *)q->runq_lock, irqstate);
if (!status) {
dprintk("%s: issuing IPI, thread->cpu_id=%d\n",
__func__, uti_info->cpu);
dprintk("%s: issuing IPI, thread->cpu_id=%d, intr_id: %d\n",
__func__, uti_info->cpu, q->intr_id);
ihk_os_issue_interrupt(uti_info->os,
q->intr_id, q->intr_vector);
ihk_os_issue_interrupt(uti_info->os, q->intr_id,
q->intr_vector);
}
return status;
@ -582,6 +681,7 @@ static int64_t futex_wait_queue_me(struct futex_hash_bucket *hb, struct futex_q
{
int64_t time_remain = 0;
unsigned long irqstate;
/*
* The task state is guaranteed to be set before another task can
* wake it.
@ -606,8 +706,8 @@ static int64_t futex_wait_queue_me(struct futex_hash_bucket *hb, struct futex_q
queue_me(q, hb, uti_info);
if (!mc_plist_node_empty(&q->list)) {
dprintk("%s: tid: %d is trying to sleep\n", __func__,
uti_info->tid);
dprintk("%s: tid: %d is trying to sleep, cpu: %d\n",
__func__, uti_info->tid, ihk_ikc_get_processor_id());
/* Note that the unit of timeout is nsec */
time_remain = uti_wait_event(q->uti_futex_resp, timeout);
@ -619,7 +719,8 @@ static int64_t futex_wait_queue_me(struct futex_hash_bucket *hb, struct futex_q
pr_err("%s: ERROR: wait_event returned %lld\n", __func__, time_remain);
}
}
dprintk("%s: tid: %d woken up\n", __func__, uti_info->tid);
dprintk("%s: tid: %d woken up, cpu: %d\n",
__func__, uti_info->tid, ihk_ikc_get_processor_id());
}
/* This does not need to be serialized */
@ -1052,7 +1153,7 @@ long do_futex(int n, unsigned long arg0, unsigned long arg1,
}
op = (op & FUTEX_CMD_MASK);
dprintk("futex op=[%x, %s],uaddr=%lx, val=%x, utime=%p, uaddr2=%p, val3=%x, []=%x, shared: %d\n",
dprintk("futex op=[%x, %s],uaddr=%lx, val=%x, utime=%p, uaddr2=%p, val3=%x, shared: %d\n",
flags,
(op == FUTEX_WAIT) ? "FUTEX_WAIT" :
(op == FUTEX_WAIT_BITSET) ? "FUTEX_WAIT_BITSET" :
@ -1061,7 +1162,7 @@ long do_futex(int n, unsigned long arg0, unsigned long arg1,
(op == FUTEX_WAKE_BITSET) ? "FUTEX_WAKE_BITSET" :
(op == FUTEX_CMP_REQUEUE) ? "FUTEX_CMP_REQUEUE" :
(op == FUTEX_REQUEUE) ? "FUTEX_REQUEUE (NOT IMPL!)" : "unknown",
(unsigned long)uaddr, val, utime, uaddr2, val3, *uaddr, fshared);
(unsigned long)uaddr, val, utime, uaddr2, val3, fshared);
if (utime && (op == FUTEX_WAIT_BITSET || op == FUTEX_WAIT)) {
if (copy_from_user(&ts, utime, sizeof(ts)) != 0) {
@ -1100,7 +1201,7 @@ long do_futex(int n, unsigned long arg0, unsigned long arg1,
ret = futex(uaddr, op, val, timeout, uaddr2,
val2, val3, fshared, uti_info);
dprintk("futex op=[%x, %s],uaddr=%lx, val=%x, utime=%p, uaddr2=%p, val3=%x, []=%x, shared: %d, ret: %d\n",
dprintk("futex op=[%x, %s],uaddr=%lx, val=%x, utime=%p, uaddr2=%p, val3=%x, shared: %d, ret: %d\n",
op,
(op == FUTEX_WAIT) ? "FUTEX_WAIT" :
(op == FUTEX_WAIT_BITSET) ? "FUTEX_WAIT_BITSET" :
@ -1109,7 +1210,7 @@ long do_futex(int n, unsigned long arg0, unsigned long arg1,
(op == FUTEX_WAKE_BITSET) ? "FUTEX_WAKE_BITSET" :
(op == FUTEX_CMP_REQUEUE) ? "FUTEX_CMP_REQUEUE" :
(op == FUTEX_REQUEUE) ? "FUTEX_REQUEUE (NOT IMPL!)" : "unknown",
(unsigned long)uaddr, val, utime, uaddr2, val3, *uaddr, fshared, ret);
(unsigned long)uaddr, val, utime, uaddr2, val3, fshared, ret);
return ret;
}

View File

@ -166,4 +166,6 @@ long do_futex(int n, unsigned long arg0, unsigned long arg1,
struct uti_info *uti_info,
void *uti_futex_resp);
void futex_remove_process(struct mcctrl_per_proc_data *ppd);
#endif

View File

@ -264,6 +264,8 @@ struct mcctrl_per_proc_data {
struct list_head devobj_pager_list;
struct semaphore devobj_pager_lock;
int enable_tofu;
struct rb_root rva_to_rpa_cache;
};
struct sysfsm_req {

View File

@ -239,7 +239,7 @@ static void wake_futex(struct futex_q *q)
if (q->uti_futex_resp) {
int rc;
dkprintf("%s: waking up migrated-to-Linux thread (tid %d),uti_futex_resp=%p\n",
__func__, p->tid, q->uti_futex_resp);
__func__, p->tid, q->uti_futex_resp);
struct ikc_scd_packet pckt;
struct ihk_ikc_channel_desc *resp_channel = cpu_local_var(ikc2linux);

View File

@ -3649,6 +3649,8 @@ void spin_sleep_or_schedule(void)
}
if (woken) {
dkprintf("%s: woken while spinning, cpu: %d, do_schedule: %d\n",
__func__, ihk_ikc_get_processor_id(), do_schedule);
if (do_schedule) {
irqstate = ihk_mc_spinlock_lock(&v->runq_lock);
v->flags |= CPU_FLAG_NEED_RESCHED;
@ -3667,6 +3669,8 @@ void spin_sleep_or_schedule(void)
out_schedule:
schedule();
dkprintf("%s: woken while sleeping, cpu: %d\n",
__func__, ihk_ikc_get_processor_id());
}
void schedule(void)
@ -3680,7 +3684,7 @@ void schedule(void)
if (cpu_local_var(no_preempt)) {
kprintf("%s: WARNING can't schedule() while no preemption, cnt: %d\n",
__FUNCTION__, cpu_local_var(no_preempt));
__func__, cpu_local_var(no_preempt));
irqstate = cpu_disable_interrupt_save();
ihk_mc_spinlock_lock_noirq(

View File

@ -2,6 +2,7 @@
USELTP=1
USEOSTEST=0
MCREBOOT=0
. ../../common.sh
@ -23,6 +24,7 @@ pushd ${UTI_TEST_DIR}
make
popd
mcreboot
for tno in `seq 12 20`
do
tname=`printf "C${issue}T%02d" ${tid}`
@ -46,6 +48,7 @@ for tno in `seq 31 34`
do
sudo ${UTI_TEST_DIR}/CT${tno} -l &> ./lnx_CT${tno}.txt
done
echo "*** Boot mckernel"
mcreboot
echo ""
@ -84,4 +87,3 @@ do
let tid++
echo ""
done

View File

@ -40,8 +40,8 @@ void *util_thread(void *arg)
FUTEX_WAKE, 1, NULL, NULL, 0);
if (rc != 1) {
fprintf(stderr,
"CT13101 FUTEX_WAKE NG (%d,%s)\n",
rc, strerror(errno));
"CT13101 FUTEX_WAKE NG (%d,%d)\n",
rc, errno);
}
else {
fprintf(stderr, "CT13101 FUTEX_WAKE OK\n");

View File

@ -20,11 +20,11 @@
static inline void FIXED_SIZE_WORK(unsigned long *ptr)
{
asm volatile("mov %x0, x20\n"
"add x20, x20, #1\n"
"mov x20, %x0\n"
: "+rm" (*ptr)
:
: "x20", "cc", "memory");
"add x20, x20, #1\n"
"mov x20, %x0\n"
: "+rm" (*ptr)
:
: "x20", "cc", "memory");
}
static inline void BULK_FSW(unsigned long n,