On Thu, Jul 08, 2010 at 11:40:05AM +0300, Andriy Gapon wrote: > on 08/07/2010 11:29 Kostik Belousov said the following: > > Right, the patch maps the page in sf buffer read-only (on i386 only). > > But note the parallel posting with m_cat() change. It is still not enough, > > and I am not set up for the real network testing ATM. > > Could you also try to experiment with mb_dupcl? > Namely transfer M_RDONLY from source mbuf.
Right, it is it.
Below is my current patch including debugging facilities that seems to work.
Real changes that needed are in m_cat and mb_dupcl.
diff --git a/sys/i386/i386/vm_machdep.c b/sys/i386/i386/vm_machdep.c
index 6f86895..530f704 100644
--- a/sys/i386/i386/vm_machdep.c
+++ b/sys/i386/i386/vm_machdep.c
@@ -810,6 +810,12 @@ sf_buf_alloc(struct vm_page *m, int flags)
nsfbufsused++;
nsfbufspeak = imax(nsfbufspeak, nsfbufsused);
}
+ if ((flags & SFB_RONLY) == 0) {
+ ptep = vtopte(sf->kva);
+ opte = *ptep;
+ if ((opte & PG_RW) == 0)
+ *ptep |= PG_RW | PG_A;
+ }
#ifdef SMP
goto shootdown;
#else
@@ -854,8 +860,9 @@ sf_buf_alloc(struct vm_page *m, int flags)
PT_SET_MA(sf->kva, xpmap_ptom(VM_PAGE_TO_PHYS(m)) | pgeflag
| PG_RW | PG_V | pmap_cache_bits(m->md.pat_mode, 0));
#else
- *ptep = VM_PAGE_TO_PHYS(m) | pgeflag | PG_RW | PG_V |
- pmap_cache_bits(m->md.pat_mode, 0);
+ *ptep = VM_PAGE_TO_PHYS(m) | pgeflag |
+ ((flags & SFB_RONLY) ? 0 : PG_RW) | PG_V |
+ pmap_cache_bits(m->md.pat_mode, 0);
#endif
/*
diff --git a/sys/kern/uipc_mbuf.c b/sys/kern/uipc_mbuf.c
index f41eb03..1701ef2 100644
--- a/sys/kern/uipc_mbuf.c
+++ b/sys/kern/uipc_mbuf.c
@@ -301,7 +301,7 @@ mb_dupcl(struct mbuf *n, struct mbuf *m)
n->m_ext.ext_size = m->m_ext.ext_size;
n->m_ext.ref_cnt = m->m_ext.ref_cnt;
n->m_ext.ext_type = m->m_ext.ext_type;
- n->m_flags |= M_EXT;
+ n->m_flags |= M_EXT | (M_RDONLY & m->m_flags);
}
/*
@@ -911,7 +911,8 @@ m_cat(struct mbuf *m, struct mbuf *n)
m = m->m_next;
while (n) {
if (m->m_flags & M_EXT ||
- m->m_data + m->m_len + n->m_len >= &m->m_dat[MLEN]) {
+ m->m_data + m->m_len + n->m_len >= &m->m_dat[MLEN] ||
+ !M_WRITABLE(m)) {
/* just join the two chains */
m->m_next = n;
return;
diff --git a/sys/kern/uipc_sockbuf.c b/sys/kern/uipc_sockbuf.c
index 2060a2e..95fae7e 100644
--- a/sys/kern/uipc_sockbuf.c
+++ b/sys/kern/uipc_sockbuf.c
@@ -776,6 +776,8 @@ sbcompress(struct sockbuf *sb, struct mbuf *m, struct mbuf
*n)
m->m_len <= MCLBYTES / 4 && /* XXX: Don't copy too much */
m->m_len <= M_TRAILINGSPACE(n) &&
n->m_type == m->m_type) {
+ KASSERT(n->m_ext.ext_type != EXT_SFBUF,
+ ("destroying file page %p", n));
bcopy(mtod(m, caddr_t), mtod(n, caddr_t) + n->m_len,
(unsigned)m->m_len);
n->m_len += m->m_len;
diff --git a/sys/kern/uipc_syscalls.c b/sys/kern/uipc_syscalls.c
index 3165dab..df9cd69 100644
--- a/sys/kern/uipc_syscalls.c
+++ b/sys/kern/uipc_syscalls.c
@@ -2131,7 +2131,8 @@ retry_space:
* as necessary, but this wait can be interrupted.
*/
if ((sf = sf_buf_alloc(pg,
- (mnw ? SFB_NOWAIT : SFB_CATCH))) == NULL) {
+ (mnw ? SFB_NOWAIT : SFB_CATCH) | SFB_RONLY))
+ == NULL) {
mbstat.sf_allocfail++;
vm_page_lock(pg);
vm_page_unwire(pg, 0);
@@ -2162,6 +2163,8 @@ retry_space:
m_cat(m, m0);
else
m = m0;
+ KASSERT((m0->m_flags & M_RDONLY) != 0,
+ ("lost M_RDONLY"));
/* Keep track of bits processed. */
loopbytes += xfsize;
diff --git a/sys/sys/sf_buf.h b/sys/sys/sf_buf.h
index af42065..fcb31f8 100644
--- a/sys/sys/sf_buf.h
+++ b/sys/sys/sf_buf.h
@@ -41,6 +41,7 @@
#define SFB_CPUPRIVATE 2 /* Create a CPU private
mapping. */
#define SFB_DEFAULT 0
#define SFB_NOWAIT 4 /* Return NULL if all bufs are
used. */
+#define SFB_RONLY 8 /* Map page read-only, if
possible. */
struct vm_page;
pgpJitHDuL82I.pgp
Description: PGP signature
