Hello Roland,
Here is the new version of the patch. It tries to address most of your comments.
I've looked at possibility to use autoconf for detecting MADV_* defines,
but I haven't found AC_CHECK_DEFUN or something like this to check for
available defines. Besides what should we do in case the define is
available? Define HAVE_MADV_DOFORK and check this instead of MADV_DOFORK?
It seem redundant to me.
Index: libibverbs/include/infiniband/verbs.h
===================================================================
--- libibverbs/include/infiniband/verbs.h (revision 7141)
+++ libibverbs/include/infiniband/verbs.h (working copy)
@@ -289,6 +289,8 @@ struct ibv_mr {
uint32_t handle;
uint32_t lkey;
uint32_t rkey;
+ void *addr;
+ size_t length;
};
struct ibv_global_route {
Index: libibverbs/src/verbs.c
===================================================================
--- libibverbs/src/verbs.c (revision 7141)
+++ libibverbs/src/verbs.c (working copy)
@@ -154,10 +154,13 @@ struct ibv_mr *ibv_reg_mr(struct ibv_pd
{
struct ibv_mr *mr;
+ ibv_dontfork_range(addr, length);
mr = pd->context->ops.reg_mr(pd, addr, length, access);
if (mr) {
mr->context = pd->context;
mr->pd = pd;
+ } else {
+ ibv_dofork_range(addr, length);
}
return mr;
@@ -165,7 +168,12 @@ struct ibv_mr *ibv_reg_mr(struct ibv_pd
int ibv_dereg_mr(struct ibv_mr *mr)
{
- return mr->context->ops.dereg_mr(mr);
+ int rc = mr->context->ops.dereg_mr(mr);
+
+ if (!rc)
+ ibv_dofork_range(mr->addr, mr->length);
+
+ return rc;
}
static struct ibv_comp_channel *ibv_create_comp_channel_v2(struct ibv_context
*context)
Index: libibverbs/src/ibverbs.h
===================================================================
--- libibverbs/src/ibverbs.h (revision 7141)
+++ libibverbs/src/ibverbs.h (working copy)
@@ -61,8 +61,8 @@ extern HIDDEN int abi_ver;
extern HIDDEN int ibverbs_init(struct ibv_device ***list);
extern HIDDEN int ibv_init_mem_map(void);
-extern HIDDEN int ibv_lock_range(void *base, size_t size);
-extern HIDDEN int ibv_unlock_range(void *base, size_t size);
+extern HIDDEN int ibv_dontfork_range(void *base, size_t size);
+extern HIDDEN int ibv_dofork_range(void *base, size_t size);
#define IBV_INIT_CMD(cmd, size, opcode)
\
do { \
Index: libibverbs/src/cmd.c
===================================================================
--- libibverbs/src/cmd.c (revision 7141)
+++ libibverbs/src/cmd.c (working copy)
@@ -238,6 +238,8 @@ int ibv_cmd_reg_mr(struct ibv_pd *pd, vo
mr->handle = resp.mr_handle;
mr->lkey = resp.lkey;
mr->rkey = resp.rkey;
+ mr->addr = addr;
+ mr->length = length;
return 0;
}
Index: libibverbs/src/memory.c
===================================================================
--- libibverbs/src/memory.c (revision 7141)
+++ libibverbs/src/memory.c (working copy)
@@ -43,6 +43,13 @@
#include "ibverbs.h"
+#ifndef MADV_DONTFORK
+#define MADV_DONTFORK 10
+#endif
+#ifndef MADV_DOFORK
+#define MADV_DOFORK 11
+#endif
+
/*
* We keep a linked list of page ranges that have been locked along with a
* reference count to manage overlapping registrations, etc.
@@ -136,7 +143,7 @@ static void __mm_remove(struct ibv_mem_n
node->next->prev = node->prev;
}
-int ibv_lock_range(void *base, size_t size)
+int ibv_dontfork_range(void *base, size_t size)
{
uintptr_t start, end;
struct ibv_mem_node *node, *tmp;
@@ -187,8 +194,8 @@ int ibv_lock_range(void *base, size_t si
if (node->refcnt++ == 0) {
- ret = mlock((void *) node->start,
- node->end - node->start + 1);
+ ret = madvise((void *) node->start,
+ node->end - node->start + 1, MADV_DONTFORK);
if (ret)
goto out;
}
@@ -202,7 +209,7 @@ out:
return ret;
}
-int ibv_unlock_range(void *base, size_t size)
+int ibv_dofork_range(void *base, size_t size)
{
uintptr_t start, end;
struct ibv_mem_node *node, *tmp;
@@ -226,8 +233,8 @@ int ibv_unlock_range(void *base, size_t
while (node && node->end <= end) {
if (--node->refcnt == 0) {
- ret = munlock((void *) node->start,
- node->end - node->start + 1);
+ ret = madvise((void *) node->start,
+ node->end - node->start + 1, MADV_DOFORK);
}
if (__mm_prev(node) && node->refcnt == __mm_prev(node)->refcnt)
{
Index: libmthca/src/qp.c
===================================================================
--- libmthca/src/qp.c (revision 7141)
+++ libmthca/src/qp.c (working copy)
@@ -819,8 +819,10 @@ int mthca_alloc_qp_buf(struct ibv_pd *pd
qp->buf_size = qp->send_wqe_offset + (qp->sq.max << qp->sq.wqe_shift);
- if (posix_memalign(&qp->buf, to_mdev(pd->context->device)->page_size,
- align(qp->buf_size,
to_mdev(pd->context->device)->page_size))) {
+ if (mthca_memalign_dontfork(&qp->buf,
+ to_mdev(pd->context->device)->page_size,
+ align(qp->buf_size,
+ to_mdev(pd->context->device)->page_size))) {
free(qp->wrid);
return -1;
}
Index: libmthca/src/mthca.h
===================================================================
--- libmthca/src/mthca.h (revision 7141)
+++ libmthca/src/mthca.h (working copy)
@@ -36,6 +36,8 @@
#ifndef MTHCA_H
#define MTHCA_H
+#include <sys/mman.h>
+
#include <infiniband/driver.h>
#include <infiniband/arch.h>
@@ -341,4 +343,32 @@ void mthca_free_av(struct mthca_ah *ah);
int mthca_attach_mcast(struct ibv_qp *qp, union ibv_gid *gid, uint16_t lid);
int mthca_detach_mcast(struct ibv_qp *qp, union ibv_gid *gid, uint16_t lid);
+#ifndef MADV_DONTFORK
+#define MADV_DONTFORK 10
+#endif
+#ifndef MADV_DOFORK
+#define MADV_DOFORK 11
+#endif
+
+static inline int mthca_memalign_dontfork(void **memptr, size_t alignment,
+ size_t size)
+{
+ int ret;
+
+ ret = posix_memalign(memptr, alignment, size);
+
+ if (ret)
+ return ret;
+
+ madvise(*memptr, size, MADV_DONTFORK);
+
+ return 0;
+}
+
+static inline void mthca_free_dofork(void *ptr, size_t size)
+{
+ madvise(ptr, size, MADV_DOFORK);
+ free(ptr);
+}
+
#endif /* MTHCA_H */
Index: libmthca/src/verbs.c
===================================================================
--- libmthca/src/verbs.c (revision 7141)
+++ libmthca/src/verbs.c (working copy)
@@ -247,7 +247,7 @@ err_unreg:
mthca_dereg_mr(cq->mr);
err_buf:
- free(cq->buf);
+ mthca_free_dofork(cq->buf, cqe * MTHCA_CQ_ENTRY_SIZE);
err:
free(cq);
@@ -263,6 +263,7 @@ int mthca_resize_cq(struct ibv_cq *ibcq,
void *buf;
int old_cqe;
int ret;
+ size_t length;
pthread_spin_lock(&cq->lock);
@@ -282,7 +283,7 @@ int mthca_resize_cq(struct ibv_cq *ibcq,
cqe * MTHCA_CQ_ENTRY_SIZE,
0, IBV_ACCESS_LOCAL_WRITE);
if (!mr) {
- free(buf);
+ mthca_free_dofork(buf, cqe * MTHCA_CQ_ENTRY_SIZE);
ret = ENOMEM;
goto out;
}
@@ -295,14 +296,15 @@ int mthca_resize_cq(struct ibv_cq *ibcq,
ret = ibv_cmd_resize_cq(ibcq, cqe - 1, &cmd.ibv_cmd, sizeof cmd);
if (ret) {
mthca_dereg_mr(mr);
- free(buf);
+ mthca_free_dofork(buf, cqe * MTHCA_CQ_ENTRY_SIZE);
goto out;
}
mthca_cq_resize_copy_cqes(cq, buf, old_cqe);
+ length = cq->mr->length;
mthca_dereg_mr(cq->mr);
- free(cq->buf);
+ mthca_free_dofork(cq->buf, length);
cq->buf = buf;
cq->mr = mr;
@@ -315,6 +317,7 @@ out:
int mthca_destroy_cq(struct ibv_cq *cq)
{
int ret;
+ size_t length;
ret = ibv_cmd_destroy_cq(cq);
if (ret)
@@ -327,9 +330,10 @@ int mthca_destroy_cq(struct ibv_cq *cq)
to_mcq(cq)->arm_db_index);
}
+ length = to_mcq(cq)->mr->length;
mthca_dereg_mr(to_mcq(cq)->mr);
- free(to_mcq(cq)->buf);
+ mthca_free_dofork(to_mcq(cq)->buf, length);
free(to_mcq(cq));
return 0;
@@ -422,7 +426,7 @@ err_unreg:
err_free:
free(srq->wrid);
- free(srq->buf);
+ mthca_free_dofork(srq->buf, srq->buf_size);
err:
free(srq);
@@ -461,7 +465,7 @@ int mthca_destroy_srq(struct ibv_srq *sr
mthca_dereg_mr(to_msrq(srq)->mr);
- free(to_msrq(srq)->buf);
+ mthca_free_dofork(to_msrq(srq)->buf, to_msrq(srq)->buf_size);
free(to_msrq(srq)->wrid);
free(to_msrq(srq));
@@ -566,7 +570,7 @@ err_unreg:
err_free:
free(qp->wrid);
- free(qp->buf);
+ mthca_free_dofork(qp->buf, qp->buf_size);
err:
free(qp);
@@ -648,7 +652,7 @@ int mthca_destroy_qp(struct ibv_qp *qp)
mthca_dereg_mr(to_mqp(qp)->mr);
- free(to_mqp(qp)->buf);
+ mthca_free_dofork(to_mqp(qp)->buf, to_mqp(qp)->buf_size);
free(to_mqp(qp)->wrid);
free(to_mqp(qp));
Index: libmthca/src/cq.c
===================================================================
--- libmthca/src/cq.c (revision 7141)
+++ libmthca/src/cq.c (working copy)
@@ -606,8 +606,9 @@ void *mthca_alloc_cq_buf(struct mthca_de
void *buf;
int i;
- if (posix_memalign(&buf, dev->page_size,
- align(nent * MTHCA_CQ_ENTRY_SIZE, dev->page_size)))
+ if (mthca_memalign_dontfork(&buf, dev->page_size,
+ align(nent * MTHCA_CQ_ENTRY_SIZE,
+ dev->page_size)))
return NULL;
for (i = 0; i < nent; ++i)
Index: libmthca/src/srq.c
===================================================================
--- libmthca/src/srq.c (revision 7141)
+++ libmthca/src/srq.c (working copy)
@@ -291,8 +291,10 @@ int mthca_alloc_srq_buf(struct ibv_pd *p
srq->buf_size = srq->max << srq->wqe_shift;
- if (posix_memalign(&srq->buf, to_mdev(pd->context->device)->page_size,
- align(srq->buf_size,
to_mdev(pd->context->device)->page_size))) {
+ if (mthca_memalign_dontfork(&srq->buf,
+ to_mdev(pd->context->device)->page_size,
+ align(srq->buf_size,
+ to_mdev(pd->context->device)->page_size))) {
free(srq->wrid);
return -1;
}
Index: libmthca/src/ah.c
===================================================================
--- libmthca/src/ah.c (revision 7141)
+++ libmthca/src/ah.c (working copy)
@@ -59,14 +59,14 @@ static struct mthca_ah_page *__add_page(
if (!page)
return NULL;
- if (posix_memalign(&page->buf, page_size, page_size)) {
+ if (mthca_memalign_dontfork(&page->buf, page_size, page_size)) {
free(page);
return NULL;
}
page->mr = mthca_reg_mr(&pd->ibv_pd, page->buf, page_size, 0);
if (!page->mr) {
- free(page->buf);
+ mthca_free_dofork(page->buf, page_size);
free(page);
return NULL;
}
@@ -175,6 +175,8 @@ void mthca_free_av(struct mthca_ah *ah)
page->free[i / (8 * sizeof (int))] |= 1 << (i % (8 * sizeof
(int)));
if (!--page->use_cnt) {
+ size_t length;
+
if (page->prev)
page->prev->next = page->next;
else
@@ -182,8 +184,9 @@ void mthca_free_av(struct mthca_ah *ah)
if (page->next)
page->next->prev = page->prev;
+ length = page->mr->length;
mthca_dereg_mr(page->mr);
- free(page->buf);
+ mthca_free_dofork(page->buf, length);
free(page);
}
--
Gleb.
_______________________________________________
openib-general mailing list
[email protected]
http://openib.org/mailman/listinfo/openib-general
To unsubscribe, please visit http://openib.org/mailman/listinfo/openib-general
