/**
* test_fork - test if forks are handled properly
* Description:
* Called by xpmem_master, but do nothing. xpmem_proc1 does the fork.
* Return Values:
* Success: 0
* Failure: -2
*/
int test_fork(test_args *xpmem_args)
{
xpmem_segid_t segid;
xpmem_apid_t apid;
struct xpmem_addr addr;
int i, ret=0, *data;
segid = strtol(xpmem_args->share, NULL, 16);
apid = xpmem_get(segid, XPMEM_RDWR, XPMEM_PERMIT_MODE, NULL);
addr.apid = apid;
addr.offset = PAGE_SIZE;
data = (int *)xpmem_attach(addr, PAGE_SIZE, NULL);
if (data == (void *)-1) {
perror("xpmem_attach");
return -2;
}
printf("xpmem_proc2: mypid = %d\n", getpid());
printf("xpmem_proc2: segid = %llx\n", segid);
printf("xpmem_proc2: attached at %p\n", data);
printf("xpmem_proc2: reading to pin pages\n");
for (i = 0; i < PAGE_INT_SIZE; i++) {
if (*(data + i) != PAGE_INT_SIZE + i) {
printf("xpmem_proc2: ***mismatch at %d: expected %lu "
"got %d\n", i, PAGE_INT_SIZE + i, *(data + i));
ret = -2;
}
}
/* Now wait for xpmem_proc1 to invoke COW */
printf("xpmem_proc2: waiting for COW...\n\n");
while (xpmem_args->share[COW_LOCK_INDEX] == 0) {
xpmem_args->share[COW_LOCK_INDEX] = 1;
}
sleep(1);
printf("xpmem_proc2: adding 1 to all elems\n\n");
for (i = 0; i < PAGE_INT_SIZE; i++)
*(data + i) += 1;
xpmem_detach(data);
xpmem_release(apid);
return ret;
}
static ucs_status_t uct_xpmem_attach(uct_mm_id_t mmid, size_t length,
void *remote_address, void **local_address,
uint64_t *cookie, const char *path)
{
struct xpmem_addr addr;
ucs_status_t status;
ptrdiff_t offset;
void *address;
addr.offset = 0;
addr.apid = xpmem_get(mmid, XPMEM_RDWR, XPMEM_PERMIT_MODE, NULL);
VALGRIND_MAKE_MEM_DEFINED(&addr.apid, sizeof(addr.apid));
if (addr.apid < 0) {
ucs_error("Failed to acquire xpmem segment 0x%"PRIx64": %m", mmid);
status = UCS_ERR_IO_ERROR;
goto err_xget;
}
ucs_trace("xpmem acquired segment 0x%"PRIx64" apid 0x%llx remote_address %p",
mmid, addr.apid, remote_address);
offset = ((uintptr_t)remote_address) % ucs_get_page_size();
address = xpmem_attach(addr, length + offset, NULL);
VALGRIND_MAKE_MEM_DEFINED(&address, sizeof(address));
if (address == MAP_FAILED) {
ucs_error("Failed to attach xpmem segment 0x%"PRIx64" apid 0x%llx "
"with length %zu: %m", mmid, addr.apid, length);
status = UCS_ERR_IO_ERROR;
goto err_xattach;
}
VALGRIND_MAKE_MEM_DEFINED(address + offset, length);
*local_address = address + offset;
*cookie = addr.apid;
ucs_trace("xpmem attached segment 0x%"PRIx64" apid 0x%llx %p..%p at %p (+%zd)",
mmid, addr.apid, remote_address, remote_address + length, address, offset);
return UCS_OK;
err_xattach:
xpmem_release(addr.apid);
err_xget:
return status;
}
/* look up the remote pointer in the peer rcache and attach if
* necessary */
mca_mpool_base_registration_t *vader_get_registation (struct mca_btl_base_endpoint_t *ep, void *rem_ptr,
size_t size, int flags, void **local_ptr)
{
struct mca_rcache_base_module_t *rcache = ep->rcache;
mca_mpool_base_registration_t *regs[10], *reg = NULL;
xpmem_addr_t xpmem_addr;
uintptr_t base, bound;
int rc, i;
/* protect rcache access */
OPAL_THREAD_LOCK(&ep->lock);
/* use btl/self for self communication */
assert (ep->peer_smp_rank != MCA_BTL_VADER_LOCAL_RANK);
base = (uintptr_t) down_align_addr(rem_ptr, mca_btl_vader_component.log_attach_align);
bound = (uintptr_t) up_align_addr((void *)((uintptr_t) rem_ptr + size - 1),
mca_btl_vader_component.log_attach_align) + 1;
if (OPAL_UNLIKELY(bound > VADER_MAX_ADDRESS)) {
bound = VADER_MAX_ADDRESS;
}
/* several segments may match the base pointer */
rc = rcache->rcache_find_all (rcache, (void *) base, bound - base, regs, 10);
for (i = 0 ; i < rc ; ++i) {
if (bound <= (uintptr_t)regs[i]->bound && base >= (uintptr_t)regs[i]->base) {
opal_atomic_add (®s[i]->ref_count, 1);
reg = regs[i];
goto reg_found;
}
if (regs[i]->flags & MCA_MPOOL_FLAGS_PERSIST) {
continue;
}
/* remove this pointer from the rcache and decrement its reference count
(so it is detached later) */
rc = rcache->rcache_delete (rcache, regs[i]);
if (OPAL_UNLIKELY(0 != rc)) {
/* someone beat us to it? */
break;
}
/* start the new segment from the lower of the two bases */
base = (uintptr_t) regs[i]->base < base ? (uintptr_t) regs[i]->base : base;
opal_atomic_add (®s[i]->ref_count, -1);
if (OPAL_LIKELY(0 == regs[i]->ref_count)) {
/* this pointer is not in use */
(void) xpmem_detach (regs[i]->alloc_base);
OBJ_RELEASE(regs[i]);
}
break;
}
reg = OBJ_NEW(mca_mpool_base_registration_t);
if (OPAL_LIKELY(NULL != reg)) {
/* stick around for awhile */
reg->ref_count = 2;
reg->base = (unsigned char *) base;
reg->bound = (unsigned char *) bound;
reg->flags = flags;
#if defined(HAVE_SN_XPMEM_H)
xpmem_addr.id = ep->apid;
#else
xpmem_addr.apid = ep->apid;
#endif
xpmem_addr.offset = base;
reg->alloc_base = xpmem_attach (xpmem_addr, bound - base, NULL);
if (OPAL_UNLIKELY((void *)-1 == reg->alloc_base)) {
OPAL_THREAD_UNLOCK(&ep->lock);
OBJ_RELEASE(reg);
return NULL;
}
opal_memchecker_base_mem_defined (reg->alloc_base, bound - base);
rcache->rcache_insert (rcache, reg, 0);
}
reg_found:
opal_atomic_wmb ();
*local_ptr = (void *) ((uintptr_t) reg->alloc_base +
(ptrdiff_t)((uintptr_t) rem_ptr - (uintptr_t) reg->base));
OPAL_THREAD_UNLOCK(&ep->lock);
return reg;
}
int
shmem_transport_xpmem_startup(void)
{
int ret, i, peer_num, num_on_node = 0;
struct share_info_t info;
struct xpmem_addr addr;
for (i = 0 ; i < shmem_internal_num_pes; ++i) {
if (-1 != SHMEM_GET_RANK_SAME_NODE(i)) {
num_on_node++;
}
}
/* allocate space for local peers */
shmem_transport_xpmem_peers = calloc(num_on_node,
sizeof(struct shmem_transport_xpmem_peer_info_t));
if (NULL == shmem_transport_xpmem_peers) return 1;
/* get local peer info and map into our address space ... */
for (i = 0 ; i < shmem_internal_num_pes; ++i) {
peer_num = SHMEM_GET_RANK_SAME_NODE(i);
if (-1 == peer_num) continue;
if (shmem_internal_my_pe == i) {
shmem_transport_xpmem_peers[peer_num].data_ptr =
shmem_internal_data_base;
shmem_transport_xpmem_peers[peer_num].heap_ptr =
shmem_internal_heap_base;
} else {
ret = shmem_runtime_get(i, "xpmem-segids", &info, sizeof(struct share_info_t));
if (0 != ret) {
fprintf(stderr, "[%03d] ERROR: runtime_get failed: %d\n",
shmem_internal_my_pe, ret);
return 1;
}
shmem_transport_xpmem_peers[peer_num].data_apid =
xpmem_get(info.data_seg, XPMEM_RDWR, XPMEM_PERMIT_MODE, (void*)0666);
if (shmem_transport_xpmem_peers[peer_num].data_apid < 0) {
fprintf(stderr, "[%03d] ERROR: could not get data apid: %s\n",
shmem_internal_my_pe, strerror(errno));
return 1;
}
addr.apid = shmem_transport_xpmem_peers[peer_num].data_apid;
addr.offset = 0;
shmem_transport_xpmem_peers[peer_num].data_attach_ptr =
xpmem_attach(addr, info.data_len, NULL);
if ((size_t) shmem_transport_xpmem_peers[peer_num].data_ptr == XPMEM_MAXADDR_SIZE) {
fprintf(stderr, "[%03d] ERROR: could not get data segment: %s\n",
shmem_internal_my_pe, strerror(errno));
return 1;
}
shmem_transport_xpmem_peers[peer_num].data_ptr =
(char*) shmem_transport_xpmem_peers[peer_num].data_attach_ptr + info.data_off;
shmem_transport_xpmem_peers[peer_num].heap_apid =
xpmem_get(info.heap_seg, XPMEM_RDWR, XPMEM_PERMIT_MODE, (void*)0666);
if (shmem_transport_xpmem_peers[peer_num].heap_apid < 0) {
fprintf(stderr, "[%03d] ERROR: could not get heap apid: %s\n",
shmem_internal_my_pe, strerror(errno));
return 1;
}
addr.apid = shmem_transport_xpmem_peers[peer_num].heap_apid;
addr.offset = 0;
shmem_transport_xpmem_peers[peer_num].heap_attach_ptr =
xpmem_attach(addr, info.heap_len, NULL);
if ((size_t) shmem_transport_xpmem_peers[peer_num].heap_ptr == XPMEM_MAXADDR_SIZE) {
fprintf(stderr, "[%03d] ERROR: could not get data segment: %s\n",
shmem_internal_my_pe, strerror(errno));
return 1;
}
shmem_transport_xpmem_peers[peer_num].heap_ptr =
(char*) shmem_transport_xpmem_peers[peer_num].heap_attach_ptr + info.heap_off;
}
}
return 0;
}
/* look up the remote pointer in the peer rcache and attach if
* necessary */
mca_rcache_base_registration_t *vader_get_registation (struct mca_btl_base_endpoint_t *ep, void *rem_ptr,
size_t size, int flags, void **local_ptr)
{
mca_rcache_base_vma_module_t *vma_module = mca_btl_vader_component.vma_module;
uint64_t attach_align = 1 << mca_btl_vader_component.log_attach_align;
mca_rcache_base_registration_t *reg = NULL;
vader_check_reg_ctx_t check_ctx = {.ep = ep, .reg = ®, .vma_module = vma_module};
xpmem_addr_t xpmem_addr;
uintptr_t base, bound;
int rc;
base = OPAL_DOWN_ALIGN((uintptr_t) rem_ptr, attach_align, uintptr_t);
bound = OPAL_ALIGN((uintptr_t) rem_ptr + size - 1, attach_align, uintptr_t) + 1;
if (OPAL_UNLIKELY(bound > VADER_MAX_ADDRESS)) {
bound = VADER_MAX_ADDRESS;
}
check_ctx.base = base;
check_ctx.bound = bound;
/* several segments may match the base pointer */
rc = mca_rcache_base_vma_iterate (vma_module, (void *) base, bound - base, vader_check_reg, &check_ctx);
if (2 == rc) {
/* start the new segment from the lower of the two bases */
base = (uintptr_t) reg->base < base ? (uintptr_t) reg->base : base;
if (OPAL_LIKELY(0 == opal_atomic_add_32 (®->ref_count, -1))) {
/* this pointer is not in use */
(void) xpmem_detach (reg->rcache_context);
OBJ_RELEASE(reg);
}
reg = NULL;
}
if (NULL == reg) {
reg = OBJ_NEW(mca_rcache_base_registration_t);
if (OPAL_LIKELY(NULL != reg)) {
/* stick around for awhile */
reg->ref_count = 2;
reg->base = (unsigned char *) base;
reg->bound = (unsigned char *) bound;
reg->flags = flags;
reg->alloc_base = (void *) (intptr_t) ep->peer_smp_rank;
#if defined(HAVE_SN_XPMEM_H)
xpmem_addr.id = ep->segment_data.xpmem.apid;
#else
xpmem_addr.apid = ep->segment_data.xpmem.apid;
#endif
xpmem_addr.offset = base;
reg->rcache_context = xpmem_attach (xpmem_addr, bound - base, NULL);
if (OPAL_UNLIKELY((void *)-1 == reg->rcache_context)) {
OBJ_RELEASE(reg);
return NULL;
}
opal_memchecker_base_mem_defined (reg->rcache_context, bound - base);
mca_rcache_base_vma_insert (vma_module, reg, 0);
}
}
opal_atomic_wmb ();
*local_ptr = (void *) ((uintptr_t) reg->rcache_context +
(ptrdiff_t)((uintptr_t) rem_ptr - (uintptr_t) reg->base));
return reg;
}
/*
* User ioctl to the XPMEM driver. Only 64-bit user applications are
* supported.
*/
static long
xpmem_ioctl(struct file *file, unsigned int cmd, unsigned long arg)
{
long ret;
switch (cmd) {
case XPMEM_CMD_VERSION: {
return XPMEM_CURRENT_VERSION;
}
case XPMEM_CMD_MAKE: {
struct xpmem_cmd_make make_info;
xpmem_segid_t segid;
if (copy_from_user(&make_info, (void __user *)arg,
sizeof(struct xpmem_cmd_make)))
return -EFAULT;
ret = xpmem_make(make_info.vaddr, make_info.size,
make_info.permit_type,
(void *)make_info.permit_value, &segid);
if (ret != 0)
return ret;
if (put_user(segid,
&((struct xpmem_cmd_make __user *)arg)->segid)) {
(void)xpmem_remove(segid);
return -EFAULT;
}
return 0;
}
case XPMEM_CMD_REMOVE: {
struct xpmem_cmd_remove remove_info;
if (copy_from_user(&remove_info, (void __user *)arg,
sizeof(struct xpmem_cmd_remove)))
return -EFAULT;
return xpmem_remove(remove_info.segid);
}
case XPMEM_CMD_GET: {
struct xpmem_cmd_get get_info;
xpmem_apid_t apid;
if (copy_from_user(&get_info, (void __user *)arg,
sizeof(struct xpmem_cmd_get)))
return -EFAULT;
ret = xpmem_get(get_info.segid, get_info.flags,
get_info.permit_type,
(void *)get_info.permit_value, &apid);
if (ret != 0)
return ret;
if (put_user(apid,
&((struct xpmem_cmd_get __user *)arg)->apid)) {
(void)xpmem_release(apid);
return -EFAULT;
}
return 0;
}
case XPMEM_CMD_RELEASE: {
struct xpmem_cmd_release release_info;
if (copy_from_user(&release_info, (void __user *)arg,
sizeof(struct xpmem_cmd_release)))
return -EFAULT;
return xpmem_release(release_info.apid);
}
case XPMEM_CMD_ATTACH: {
struct xpmem_cmd_attach attach_info;
u64 at_vaddr;
if (copy_from_user(&attach_info, (void __user *)arg,
sizeof(struct xpmem_cmd_attach)))
return -EFAULT;
ret = xpmem_attach(file, attach_info.apid, attach_info.offset,
attach_info.size, attach_info.vaddr,
attach_info.fd, attach_info.flags,
&at_vaddr);
if (ret != 0)
return ret;
if (put_user(at_vaddr,
&((struct xpmem_cmd_attach __user *)arg)->vaddr)) {
(void)xpmem_detach(at_vaddr);
return -EFAULT;
}
return 0;
}
case XPMEM_CMD_DETACH: {
struct xpmem_cmd_detach detach_info;
if (copy_from_user(&detach_info, (void __user *)arg,
sizeof(struct xpmem_cmd_detach)))
return -EFAULT;
return xpmem_detach(detach_info.vaddr);
}
case XPMEM_CMD_FORK_BEGIN: {
return xpmem_fork_begin();
}
case XPMEM_CMD_FORK_END: {
return xpmem_fork_end();
}
default:
break;
}
return -ENOIOCTLCMD;
}
