/*
- * Copyright (c) 1998-2014 Apple Inc. All rights reserved.
+ * Copyright (c) 1998-2016 Apple Inc. All rights reserved.
*
* @APPLE_OSREFERENCE_LICENSE_HEADER_START@
*
#include <kern/simple_lock.h>
#include <kern/queue.h>
#include <kern/sched_prim.h>
+#include <kern/backtrace.h>
#include <kern/cpu_number.h>
#include <kern/zalloc.h>
static lck_grp_attr_t *mbuf_mlock_grp_attr;
/* Back-end (common) layer */
-static void *mbuf_worker_run; /* wait channel for worker thread */
+static boolean_t mbuf_worker_needs_wakeup; /* wait channel for mbuf worker */
static int mbuf_worker_ready; /* worker thread is runnable */
static int mbuf_expand_mcl; /* number of cluster creation requets */
static int mbuf_expand_big; /* number of big cluster creation requests */
static lck_grp_t *mleak_lock_grp;
static lck_grp_attr_t *mleak_lock_grp_attr;
+/* Lock to protect the completion callback table */
+static lck_grp_attr_t *mbuf_tx_compl_tbl_lck_grp_attr = NULL;
+static lck_attr_t *mbuf_tx_compl_tbl_lck_attr = NULL;
+static lck_grp_t *mbuf_tx_compl_tbl_lck_grp = NULL;
+decl_lck_rw_data(, mbuf_tx_compl_tbl_lck_rw_data);
+lck_rw_t *mbuf_tx_compl_tbl_lock = &mbuf_tx_compl_tbl_lck_rw_data;
+
extern u_int32_t high_sb_max;
/* The minimum number of objects that are allocated, to start. */
static unsigned int mb_watchdog = 0;
static unsigned int mb_drain_maxint = 0;
+uintptr_t mb_obscure_extfree __attribute__((visibility("hidden")));
+uintptr_t mb_obscure_extref __attribute__((visibility("hidden")));
+
/* Red zone */
static u_int32_t mb_redzone_cookie;
static void m_redzone_init(struct mbuf *);
static char *mbuf_dump(void);
static void mbuf_table_init(void);
static inline void m_incref(struct mbuf *);
-static inline u_int32_t m_decref(struct mbuf *);
+static inline u_int16_t m_decref(struct mbuf *);
static int m_clalloc(const u_int32_t, const int, const u_int32_t);
static void mbuf_worker_thread_init(void);
static mcache_obj_t *slab_alloc(mbuf_class_t, int);
* cleared.
*/
#define EXTF_READONLY 0x2
-#define EXTF_MASK (EXTF_COMPOSITE | EXTF_READONLY)
-
-#define MEXT_RFA(m) ((m)->m_ext.ext_refflags)
-#define MEXT_REF(m) (MEXT_RFA(m)->refcnt)
-#define MEXT_FLAGS(m) (MEXT_RFA(m)->flags)
-#define MBUF_IS_COMPOSITE(m) \
- (MEXT_REF(m) == 0 && (MEXT_FLAGS(m) & EXTF_MASK) == EXTF_COMPOSITE)
+/*
+ * This flag indicates that the external cluster is paired with the mbuf.
+ * Pairing implies an external free routine defined which will be invoked
+ * when the reference count drops to the minimum at m_free time. This
+ * flag is never cleared.
+ */
+#define EXTF_PAIRED 0x4
+
+#define EXTF_MASK \
+ (EXTF_COMPOSITE | EXTF_READONLY | EXTF_PAIRED)
+
+#define MEXT_MINREF(m) ((m_get_rfa(m))->minref)
+#define MEXT_REF(m) ((m_get_rfa(m))->refcnt)
+#define MEXT_PREF(m) ((m_get_rfa(m))->prefcnt)
+#define MEXT_FLAGS(m) ((m_get_rfa(m))->flags)
+#define MEXT_PRIV(m) ((m_get_rfa(m))->priv)
+#define MEXT_PMBUF(m) ((m_get_rfa(m))->paired)
+#define MEXT_TOKEN(m) ((m_get_rfa(m))->ext_token)
+#define MBUF_IS_COMPOSITE(m) \
+ (MEXT_REF(m) == MEXT_MINREF(m) && \
+ (MEXT_FLAGS(m) & EXTF_MASK) == EXTF_COMPOSITE)
+/*
+ * This macro can be used to test if the mbuf is paired to an external
+ * cluster. The test for MEXT_PMBUF being equal to the mbuf in subject
+ * is important, as EXTF_PAIRED alone is insufficient since it is immutable,
+ * and thus survives calls to m_free_paired.
+ */
+#define MBUF_IS_PAIRED(m) \
+ (((m)->m_flags & M_EXT) && \
+ (MEXT_FLAGS(m) & EXTF_MASK) == EXTF_PAIRED && \
+ MEXT_PMBUF(m) == (m))
/*
* Macros used to verify the integrity of the mbuf.
*/
#define _MCHECK(m) { \
- if ((m)->m_type != MT_FREE) { \
+ if ((m)->m_type != MT_FREE && !MBUF_IS_PAIRED(m)) { \
if (mclaudit == NULL) \
panic("MCHECK: m_type=%d m=%p", \
(u_int16_t)(m)->m_type, m); \
} \
}
-#define MEXT_INIT(m, buf, size, free, arg, rfa, ref, flag) { \
+#define MEXT_INIT(m, buf, size, free, arg, rfa, min, ref, pref, flag, \
+ priv, pm) { \
(m)->m_data = (m)->m_ext.ext_buf = (buf); \
(m)->m_flags |= M_EXT; \
+ m_set_ext((m), (rfa), (free), (arg)); \
(m)->m_ext.ext_size = (size); \
- (m)->m_ext.ext_free = (free); \
- (m)->m_ext.ext_arg = (arg); \
- (m)->m_ext.ext_refs.forward = (m)->m_ext.ext_refs.backward = \
- &(m)->m_ext.ext_refs; \
- MEXT_RFA(m) = (rfa); \
+ MEXT_MINREF(m) = (min); \
MEXT_REF(m) = (ref); \
+ MEXT_PREF(m) = (pref); \
MEXT_FLAGS(m) = (flag); \
+ MEXT_PRIV(m) = (priv); \
+ MEXT_PMBUF(m) = (pm); \
}
#define MBUF_CL_INIT(m, buf, rfa, ref, flag) \
- MEXT_INIT(m, buf, m_maxsize(MC_CL), NULL, NULL, rfa, ref, flag)
+ MEXT_INIT(m, buf, m_maxsize(MC_CL), NULL, NULL, rfa, 0, \
+ ref, 0, flag, 0, NULL)
#define MBUF_BIGCL_INIT(m, buf, rfa, ref, flag) \
- MEXT_INIT(m, buf, m_maxsize(MC_BIGCL), m_bigfree, NULL, rfa, ref, flag)
+ MEXT_INIT(m, buf, m_maxsize(MC_BIGCL), m_bigfree, NULL, rfa, 0, \
+ ref, 0, flag, 0, NULL)
#define MBUF_16KCL_INIT(m, buf, rfa, ref, flag) \
- MEXT_INIT(m, buf, m_maxsize(MC_16KCL), m_16kfree, NULL, rfa, ref, flag)
+ MEXT_INIT(m, buf, m_maxsize(MC_16KCL), m_16kfree, NULL, rfa, 0, \
+ ref, 0, flag, 0, NULL)
/*
* Macro to convert BSD malloc sleep flag to mcache's
static inline void
m_incref(struct mbuf *m)
{
- UInt32 old, new;
- volatile UInt32 *addr = (volatile UInt32 *)&MEXT_REF(m);
+ UInt16 old, new;
+ volatile UInt16 *addr = (volatile UInt16 *)&MEXT_REF(m);
do {
old = *addr;
new = old + 1;
ASSERT(new != 0);
- } while (!OSCompareAndSwap(old, new, addr));
+ } while (!OSCompareAndSwap16(old, new, addr));
/*
* If cluster is shared, mark it with (sticky) EXTF_READONLY;
- * we don't clear the flag when the refcount goes back to 1
- * to simplify code calling m_mclhasreference().
+ * we don't clear the flag when the refcount goes back to the
+ * minimum, to simplify code calling m_mclhasreference().
*/
- if (new > 1 && !(MEXT_FLAGS(m) & EXTF_READONLY))
- (void) OSBitOrAtomic(EXTF_READONLY, &MEXT_FLAGS(m));
+ if (new > (MEXT_MINREF(m) + 1) && !(MEXT_FLAGS(m) & EXTF_READONLY))
+ (void) OSBitOrAtomic16(EXTF_READONLY, &MEXT_FLAGS(m));
}
-static inline u_int32_t
+static inline u_int16_t
m_decref(struct mbuf *m)
{
- UInt32 old, new;
- volatile UInt32 *addr = (volatile UInt32 *)&MEXT_REF(m);
+ UInt16 old, new;
+ volatile UInt16 *addr = (volatile UInt16 *)&MEXT_REF(m);
do {
old = *addr;
new = old - 1;
ASSERT(old != 0);
- } while (!OSCompareAndSwap(old, new, addr));
+ } while (!OSCompareAndSwap16(old, new, addr));
return (new);
}
_CASSERT(sizeof (mb_redzone_cookie) ==
sizeof (((struct pkthdr *)0)->redzone));
read_random(&mb_redzone_cookie, sizeof (mb_redzone_cookie));
+ read_random(&mb_obscure_extref, sizeof (mb_obscure_extref));
+ read_random(&mb_obscure_extfree, sizeof (mb_obscure_extfree));
+ mb_obscure_extref |= 0x3;
+ mb_obscure_extfree |= 0x3;
/* Make sure we don't save more than we should */
_CASSERT(MCA_SAVED_MBUF_SIZE <= sizeof (struct mbuf));
(nmbclusters << MCLSHIFT) >> MBSHIFT,
(nclusters << MCLSHIFT) >> MBSHIFT,
(njcl << MCLSHIFT) >> MBSHIFT);
+
+ /* initialize lock form tx completion callback table */
+ mbuf_tx_compl_tbl_lck_grp_attr = lck_grp_attr_alloc_init();
+ if (mbuf_tx_compl_tbl_lck_grp_attr == NULL) {
+ panic("%s: lck_grp_attr_alloc_init failed", __func__);
+ /* NOTREACHED */
+ }
+ mbuf_tx_compl_tbl_lck_grp = lck_grp_alloc_init("mbuf_tx_compl_tbl",
+ mbuf_tx_compl_tbl_lck_grp_attr);
+ if (mbuf_tx_compl_tbl_lck_grp == NULL) {
+ panic("%s: lck_grp_alloc_init failed", __func__);
+ /* NOTREACHED */
+ }
+ mbuf_tx_compl_tbl_lck_attr = lck_attr_alloc_init();
+ if (mbuf_tx_compl_tbl_lck_attr == NULL) {
+ panic("%s: lck_attr_alloc_init failed", __func__);
+ /* NOTREACHED */
+ }
+ lck_rw_init(mbuf_tx_compl_tbl_lock, mbuf_tx_compl_tbl_lck_grp,
+ mbuf_tx_compl_tbl_lck_attr);
+
}
/*
* A 4K cluster slab can have NBCLPG references.
*/
VERIFY(sp->sl_refcnt >= 1 && sp->sl_chunks == NBCLPG &&
- sp->sl_len == PAGE_SIZE &&
+ sp->sl_len == PAGE_SIZE &&
(sp->sl_refcnt < NBCLPG || sp->sl_head == NULL));
} else if (class == MC_16KCL) {
mcl_slab_t *nsp;
cl = m->m_ext.ext_buf;
clsp = slab_get(cl);
VERIFY(m->m_flags == M_EXT && cl != NULL);
- VERIFY(MEXT_RFA(m) != NULL && MBUF_IS_COMPOSITE(m));
+ VERIFY(m_get_rfa(m) != NULL && MBUF_IS_COMPOSITE(m));
if (class == MC_MBUF_CL) {
VERIFY(clsp->sl_refcnt >= 1 &&
}
VERIFY(ms->m_type == MT_FREE);
VERIFY(ms->m_flags == M_EXT);
- VERIFY(MEXT_RFA(ms) != NULL && MBUF_IS_COMPOSITE(ms));
+ VERIFY(m_get_rfa(ms) != NULL && MBUF_IS_COMPOSITE(ms));
if (cl_class == MC_CL) {
VERIFY(clsp->sl_refcnt >= 1 &&
clsp->sl_refcnt <= NCLPG);
} else {
- VERIFY(clsp->sl_refcnt >= 1 &&
+ VERIFY(clsp->sl_refcnt >= 1 &&
clsp->sl_refcnt <= NBCLPG);
}
if (cl_class == MC_16KCL) {
if (mclaudit != NULL)
mcl_audit_restore_mbuf(m, mca, TRUE);
+ MEXT_MINREF(m) = 0;
MEXT_REF(m) = 0;
+ MEXT_PREF(m) = 0;
MEXT_FLAGS(m) = 0;
+ MEXT_PRIV(m) = 0;
+ MEXT_PMBUF(m) = NULL;
+ MEXT_TOKEN(m) = 0;
- rfa = (mcache_obj_t *)(void *)MEXT_RFA(m);
+ rfa = (mcache_obj_t *)(void *)m_get_rfa(m);
+ m_set_ext(m, NULL, NULL, NULL);
rfa->obj_next = ref_list;
ref_list = rfa;
- MEXT_RFA(m) = NULL;
m->m_type = MT_FREE;
m->m_flags = m->m_len = 0;
MBUF_CL_INIT(ms, cl, rfa, 0, EXTF_COMPOSITE);
}
VERIFY(ms->m_flags == M_EXT);
- VERIFY(MEXT_RFA(ms) != NULL && MBUF_IS_COMPOSITE(ms));
+ VERIFY(m_get_rfa(ms) != NULL && MBUF_IS_COMPOSITE(ms));
*list = (mcache_obj_t *)m;
(*list)->obj_next = NULL;
cl = ms->m_ext.ext_buf;
clsp = slab_get(cl);
VERIFY(ms->m_flags == M_EXT && cl != NULL);
- VERIFY(MEXT_RFA(ms) != NULL && MBUF_IS_COMPOSITE(ms));
+ VERIFY(m_get_rfa(ms) != NULL && MBUF_IS_COMPOSITE(ms));
if (class == MC_MBUF_CL)
VERIFY(clsp->sl_refcnt >= 1 &&
clsp->sl_refcnt <= NCLPG);
} else {
/*
* if multiple 4K pages are being used for a
- * 16K cluster
- */
+ * 16K cluster
+ */
needed = numpages / NSLABSP16KB;
}
union m16kcluster *m16kcl = (union m16kcluster *)page;
mcl_slab_t *nsp;
int k;
-
+
/* One for the entire 16KB */
sp = slab_get(m16kcl);
if (mclaudit != NULL)
* When non-blocking we kick a thread if we have to grow the
* pool or if the number of free clusters is less than requested.
*/
+ if (i > 0 && mbuf_worker_ready && mbuf_worker_needs_wakeup) {
+ wakeup((caddr_t)&mbuf_worker_needs_wakeup);
+ mbuf_worker_needs_wakeup = FALSE;
+ }
if (class == MC_BIGCL) {
if (i > 0) {
/*
i += m_total(MC_BIGCL);
if (i > mbuf_expand_big) {
mbuf_expand_big = i;
- if (mbuf_worker_ready)
- wakeup((caddr_t)&mbuf_worker_run);
}
}
-
if (m_infree(MC_BIGCL) >= num)
return (1);
} else {
i += m_total(MC_16KCL);
if (i > mbuf_expand_16k) {
mbuf_expand_16k = i;
- if (mbuf_worker_ready)
- wakeup((caddr_t)&mbuf_worker_run);
}
}
-
if (m_infree(MC_16KCL) >= num)
return (1);
}
* MC_CL, verify that the reference count will match that
* assumption
*/
- VERIFY(sp->sl_refcnt == 1 && slab_is_detached(sp));
+ VERIFY(sp->sl_refcnt == 1 && slab_is_detached(sp));
VERIFY((sp->sl_flags & (SLF_MAPPED | SLF_PARTIAL)) == SLF_MAPPED);
/*
* Make sure that the cluster is unmolested
return (m);
}
+static int
+m_free_paired(struct mbuf *m)
+{
+ VERIFY((m->m_flags & M_EXT) && (MEXT_FLAGS(m) & EXTF_PAIRED));
+
+ membar_sync();
+ if (MEXT_PMBUF(m) == m) {
+ volatile UInt16 *addr = (volatile UInt16 *)&MEXT_PREF(m);
+ int16_t oprefcnt, prefcnt;
+
+ /*
+ * Paired ref count might be negative in case we lose
+ * against another thread clearing MEXT_PMBUF, in the
+ * event it occurs after the above memory barrier sync.
+ * In that case just ignore as things have been unpaired.
+ */
+ do {
+ oprefcnt = *addr;
+ prefcnt = oprefcnt - 1;
+ } while (!OSCompareAndSwap16(oprefcnt, prefcnt, addr));
+
+ if (prefcnt > 1) {
+ return (1);
+ } else if (prefcnt == 1) {
+ (*(m_get_ext_free(m)))(m->m_ext.ext_buf,
+ m->m_ext.ext_size, m_get_ext_arg(m));
+ return (1);
+ } else if (prefcnt == 0) {
+ VERIFY(MBUF_IS_PAIRED(m));
+
+ /*
+ * Restore minref to its natural value, so that
+ * the caller will be able to free the cluster
+ * as appropriate.
+ */
+ MEXT_MINREF(m) = 0;
+
+ /*
+ * Clear MEXT_PMBUF, but leave EXTF_PAIRED intact
+ * as it is immutable. atomic_set_ptr also causes
+ * memory barrier sync.
+ */
+ atomic_set_ptr(&MEXT_PMBUF(m), NULL);
+
+ switch (m->m_ext.ext_size) {
+ case MCLBYTES:
+ m_set_ext(m, m_get_rfa(m), NULL, NULL);
+ break;
+
+ case MBIGCLBYTES:
+ m_set_ext(m, m_get_rfa(m), m_bigfree, NULL);
+ break;
+
+ case M16KCLBYTES:
+ m_set_ext(m, m_get_rfa(m), m_16kfree, NULL);
+ break;
+
+ default:
+ VERIFY(0);
+ /* NOTREACHED */
+ }
+ }
+ }
+
+ /*
+ * Tell caller the unpair has occurred, and that the reference
+ * count on the external cluster held for the paired mbuf should
+ * now be dropped.
+ */
+ return (0);
+}
+
struct mbuf *
m_free(struct mbuf *m)
{
m_redzone_verify(m);
/* Free the aux data and tags if there is any */
m_tag_delete_chain(m, NULL);
+
+ m_do_tx_compl_callback(m, NULL);
}
if (m->m_flags & M_EXT) {
- u_int32_t refcnt;
+ u_int16_t refcnt;
u_int32_t composite;
+ m_ext_free_func_t m_free_func;
+
+ if (MBUF_IS_PAIRED(m) && m_free_paired(m))
+ return (n);
refcnt = m_decref(m);
composite = (MEXT_FLAGS(m) & EXTF_COMPOSITE);
- if (refcnt == 0 && !composite) {
- if (m->m_ext.ext_free == NULL) {
+ m_free_func = m_get_ext_free(m);
+
+ if (refcnt == MEXT_MINREF(m) && !composite) {
+ if (m_free_func == NULL) {
mcache_free(m_cache(MC_CL), m->m_ext.ext_buf);
- } else if (m->m_ext.ext_free == m_bigfree) {
+ } else if (m_free_func == m_bigfree) {
mcache_free(m_cache(MC_BIGCL),
m->m_ext.ext_buf);
- } else if (m->m_ext.ext_free == m_16kfree) {
+ } else if (m_free_func == m_16kfree) {
mcache_free(m_cache(MC_16KCL),
m->m_ext.ext_buf);
} else {
- (*(m->m_ext.ext_free))(m->m_ext.ext_buf,
- m->m_ext.ext_size, m->m_ext.ext_arg);
+ (*m_free_func)(m->m_ext.ext_buf,
+ m->m_ext.ext_size, m_get_ext_arg(m));
}
- mcache_free(ref_cache, MEXT_RFA(m));
- MEXT_RFA(m) = NULL;
- } else if (refcnt == 0 && composite) {
+ mcache_free(ref_cache, m_get_rfa(m));
+ m_set_ext(m, NULL, NULL, NULL);
+ } else if (refcnt == MEXT_MINREF(m) && composite) {
+ VERIFY(!(MEXT_FLAGS(m) & EXTF_PAIRED));
VERIFY(m->m_type != MT_FREE);
mtype_stat_dec(m->m_type);
MEXT_FLAGS(m) &= ~EXTF_READONLY;
/* "Free" into the intermediate cache */
- if (m->m_ext.ext_free == NULL) {
+ if (m_free_func == NULL) {
mcache_free(m_cache(MC_MBUF_CL), m);
- } else if (m->m_ext.ext_free == m_bigfree) {
+ } else if (m_free_func == m_bigfree) {
mcache_free(m_cache(MC_MBUF_BIGCL), m);
} else {
- VERIFY(m->m_ext.ext_free == m_16kfree);
+ VERIFY(m_free_func == m_16kfree);
mcache_free(m_cache(MC_MBUF_16KCL), m);
}
return (n);
__private_extern__ struct mbuf *
m_clattach(struct mbuf *m, int type, caddr_t extbuf,
void (*extfree)(caddr_t, u_int, caddr_t), u_int extsize, caddr_t extarg,
- int wait)
+ int wait, int pair)
{
struct ext_ref *rfa = NULL;
- if (m == NULL && (m = _M_GETHDR(wait, type)) == NULL)
+ /*
+ * If pairing is requested and an existing mbuf is provided, reject
+ * it if it's already been paired to another cluster. Otherwise,
+ * allocate a new one or free any existing below.
+ */
+ if ((m != NULL && MBUF_IS_PAIRED(m)) ||
+ (m == NULL && (m = _M_GETHDR(wait, type)) == NULL))
return (NULL);
if (m->m_flags & M_EXT) {
- u_int32_t refcnt;
+ u_int16_t refcnt;
u_int32_t composite;
+ m_ext_free_func_t m_free_func;
refcnt = m_decref(m);
composite = (MEXT_FLAGS(m) & EXTF_COMPOSITE);
- if (refcnt == 0 && !composite) {
- if (m->m_ext.ext_free == NULL) {
+ VERIFY(!(MEXT_FLAGS(m) & EXTF_PAIRED) && MEXT_PMBUF(m) == NULL);
+ m_free_func = m_get_ext_free(m);
+ if (refcnt == MEXT_MINREF(m) && !composite) {
+ if (m_free_func == NULL) {
mcache_free(m_cache(MC_CL), m->m_ext.ext_buf);
- } else if (m->m_ext.ext_free == m_bigfree) {
+ } else if (m_free_func == m_bigfree) {
mcache_free(m_cache(MC_BIGCL),
m->m_ext.ext_buf);
- } else if (m->m_ext.ext_free == m_16kfree) {
+ } else if (m_free_func == m_16kfree) {
mcache_free(m_cache(MC_16KCL),
m->m_ext.ext_buf);
} else {
- (*(m->m_ext.ext_free))(m->m_ext.ext_buf,
- m->m_ext.ext_size, m->m_ext.ext_arg);
+ (*m_free_func)(m->m_ext.ext_buf,
+ m->m_ext.ext_size, m_get_ext_arg(m));
}
/* Re-use the reference structure */
- rfa = MEXT_RFA(m);
- } else if (refcnt == 0 && composite) {
+ rfa = m_get_rfa(m);
+ } else if (refcnt == MEXT_MINREF(m) && composite) {
VERIFY(m->m_type != MT_FREE);
mtype_stat_dec(m->m_type);
MEXT_FLAGS(m) &= ~EXTF_READONLY;
/* "Free" into the intermediate cache */
- if (m->m_ext.ext_free == NULL) {
+ if (m_free_func == NULL) {
mcache_free(m_cache(MC_MBUF_CL), m);
- } else if (m->m_ext.ext_free == m_bigfree) {
+ } else if (m_free_func == m_bigfree) {
mcache_free(m_cache(MC_MBUF_BIGCL), m);
} else {
- VERIFY(m->m_ext.ext_free == m_16kfree);
+ VERIFY(m_free_func == m_16kfree);
mcache_free(m_cache(MC_MBUF_16KCL), m);
}
/*
return (NULL);
}
- MEXT_INIT(m, extbuf, extsize, extfree, extarg, rfa, 1, 0);
+ if (!pair) {
+ MEXT_INIT(m, extbuf, extsize, extfree, extarg, rfa,
+ 0, 1, 0, 0, 0, NULL);
+ } else {
+ MEXT_INIT(m, extbuf, extsize, extfree, (caddr_t)m, rfa,
+ 1, 1, 1, EXTF_PAIRED, 0, m);
+ }
return (m);
}
m = mcache_alloc(m_cache(MC_MBUF_CL), mcflags);
if (m != NULL) {
- u_int32_t flag;
+ u_int16_t flag;
struct ext_ref *rfa;
void *cl;
VERIFY(m->m_type == MT_FREE && m->m_flags == M_EXT);
cl = m->m_ext.ext_buf;
- rfa = MEXT_RFA(m);
+ rfa = m_get_rfa(m);
ASSERT(cl != NULL && rfa != NULL);
- VERIFY(MBUF_IS_COMPOSITE(m) && m->m_ext.ext_free == NULL);
+ VERIFY(MBUF_IS_COMPOSITE(m) && m_get_ext_free(m) == NULL);
flag = MEXT_FLAGS(m);
if (!(m->m_flags & M_EXT))
return (0);
- ASSERT(MEXT_RFA(m) != NULL);
+ ASSERT(m_get_rfa(m) != NULL);
return ((MEXT_FLAGS(m) & EXTF_READONLY) ? 1 : 0);
}
void
m_copy_pftag(struct mbuf *to, struct mbuf *from)
{
- to->m_pkthdr.pf_mtag = from->m_pkthdr.pf_mtag;
+ memcpy(m_pftag(to), m_pftag(from), sizeof(struct pf_mtag));
#if PF_ECN
- to->m_pkthdr.pf_mtag.pftag_hdr = NULL;
- to->m_pkthdr.pf_mtag.pftag_flags &= ~(PF_TAG_HDR_INET|PF_TAG_HDR_INET6);
+ m_pftag(to)->pftag_hdr = NULL;
+ m_pftag(to)->pftag_flags &= ~(PF_TAG_HDR_INET|PF_TAG_HDR_INET6);
#endif /* PF_ECN */
}
#if MEASURE_BW
m->m_pkthdr.pkt_bwseq = 0;
#endif /* MEASURE_BW */
- m->m_pkthdr.pkt_enqueue_ts = 0;
+ m->m_pkthdr.pkt_timestamp = 0;
}
void
unsigned int pnum, needed = *num_needed;
mcache_obj_t *mp_list = NULL;
int mcflags = MSLEEPF(wait);
- u_int32_t flag;
+ u_int16_t flag;
struct ext_ref *rfa;
mcache_t *cp;
void *cl;
VERIFY(m->m_type == MT_FREE && m->m_flags == M_EXT);
cl = m->m_ext.ext_buf;
- rfa = MEXT_RFA(m);
+ rfa = m_get_rfa(m);
ASSERT(cl != NULL && rfa != NULL);
VERIFY(MBUF_IS_COMPOSITE(m));
printf("%s: File a radar related to <rdar://10146739>. \
needed = %u, pnum = %u, num_needed = %u \n",
__func__, needed, pnum, *num_needed);
- }
+ }
*num_needed = pnum;
return (top);
for (;;) {
struct mbuf *m;
- u_int32_t flag;
+ u_int16_t flag;
struct ext_ref *rfa;
void *cl;
int pkthdr;
+ m_ext_free_func_t m_free_func;
++num;
if (nsegs == 1 || (num % nsegs) != 0 || resid == 0) {
m = (struct mbuf *)rmp_list;
rmp_list = rmp_list->obj_next;
}
+ m_free_func = m_get_ext_free(m);
ASSERT(m != NULL);
VERIFY(m->m_type == MT_FREE && m->m_flags == M_EXT);
- VERIFY(m->m_ext.ext_free == NULL ||
- m->m_ext.ext_free == m_bigfree ||
- m->m_ext.ext_free == m_16kfree);
+ VERIFY(m_free_func == NULL || m_free_func == m_bigfree ||
+ m_free_func == m_16kfree);
cl = m->m_ext.ext_buf;
- rfa = MEXT_RFA(m);
+ rfa = m_get_rfa(m);
ASSERT(cl != NULL && rfa != NULL);
VERIFY(MBUF_IS_COMPOSITE(m));
if (pkthdr)
first = m;
MBUF_INIT(m, pkthdr, MT_DATA);
- if (m->m_ext.ext_free == m_16kfree) {
+ if (m_free_func == m_16kfree) {
MBUF_16KCL_INIT(m, cl, rfa, 1, flag);
- } else if (m->m_ext.ext_free == m_bigfree) {
+ } else if (m_free_func == m_bigfree) {
MBUF_BIGCL_INIT(m, cl, rfa, 1, flag);
} else {
MBUF_CL_INIT(m, cl, rfa, 1, flag);
while (m != NULL) {
struct mbuf *next = m->m_next;
mcache_obj_t *o, *rfa;
- u_int32_t refcnt, composite;
+ u_int32_t composite;
+ u_int16_t refcnt;
+ m_ext_free_func_t m_free_func;
if (m->m_type == MT_FREE)
panic("m_free: freeing an already freed mbuf");
- if (m->m_type != MT_FREE)
- mt_free++;
-
if (m->m_flags & M_PKTHDR) {
/* Check for scratch area overflow */
m_redzone_verify(m);
m_tag_delete_chain(m, NULL);
}
- if (!(m->m_flags & M_EXT))
+ if (!(m->m_flags & M_EXT)) {
+ mt_free++;
goto simple_free;
+ }
+
+ if (MBUF_IS_PAIRED(m) && m_free_paired(m)) {
+ m = next;
+ continue;
+ }
+
+ mt_free++;
o = (mcache_obj_t *)(void *)m->m_ext.ext_buf;
refcnt = m_decref(m);
composite = (MEXT_FLAGS(m) & EXTF_COMPOSITE);
- if (refcnt == 0 && !composite) {
- if (m->m_ext.ext_free == NULL) {
+ m_free_func = m_get_ext_free(m);
+ if (refcnt == MEXT_MINREF(m) && !composite) {
+ if (m_free_func == NULL) {
o->obj_next = mcl_list;
mcl_list = o;
- } else if (m->m_ext.ext_free == m_bigfree) {
+ } else if (m_free_func == m_bigfree) {
o->obj_next = mbc_list;
mbc_list = o;
- } else if (m->m_ext.ext_free == m_16kfree) {
+ } else if (m_free_func == m_16kfree) {
o->obj_next = m16k_list;
m16k_list = o;
} else {
- (*(m->m_ext.ext_free))((caddr_t)o,
+ (*(m_free_func))((caddr_t)o,
m->m_ext.ext_size,
- m->m_ext.ext_arg);
+ m_get_ext_arg(m));
}
- rfa = (mcache_obj_t *)(void *)MEXT_RFA(m);
+ rfa = (mcache_obj_t *)(void *)m_get_rfa(m);
rfa->obj_next = ref_list;
ref_list = rfa;
- MEXT_RFA(m) = NULL;
- } else if (refcnt == 0 && composite) {
+ m_set_ext(m, NULL, NULL, NULL);
+ } else if (refcnt == MEXT_MINREF(m) && composite) {
+ VERIFY(!(MEXT_FLAGS(m) & EXTF_PAIRED));
VERIFY(m->m_type != MT_FREE);
/*
* Amortize the costs of atomic operations
/* "Free" into the intermediate cache */
o = (mcache_obj_t *)m;
- if (m->m_ext.ext_free == NULL) {
+ if (m_free_func == NULL) {
o->obj_next = m_mcl_list;
m_mcl_list = o;
- } else if (m->m_ext.ext_free == m_bigfree) {
+ } else if (m_free_func == m_bigfree) {
o->obj_next = m_mbc_list;
m_mbc_list = o;
} else {
- VERIFY(m->m_ext.ext_free == m_16kfree);
+ VERIFY(m_free_func == m_16kfree);
o->obj_next = m_m16k_list;
m_m16k_list = o;
}
n->m_len = MIN(n->m_len, MLEN);
if (MTOD(n, char *) + n->m_len > ((char *)n) + MSIZE)
- panic("%s n %p copy overflow",
+ panic("%s n %p copy overflow",
__func__, n);
bcopy(MTOD(m, caddr_t)+off, MTOD(n, caddr_t),
if (off == 0 && (m->m_flags & M_PKTHDR))
copyhdr = 1;
-
+
if (m_lastm != NULL && *m_lastm != NULL) {
m = *m_lastm;
off = *m_off;
n->m_flags |= M_EXT;
} else {
if (MTOD(n, char *) + n->m_len > ((char *)n) + MSIZE)
- panic("%s n %p copy overflow",
+ panic("%s n %p copy overflow",
__func__, n);
bcopy(MTOD(m, caddr_t)+off, MTOD(n, caddr_t),
}
while (m != NULL) {
#if BLUE_DEBUG
- kprintf("<%x: %x, %x, %x\n", m, m->m_flags, m->m_len,
+ printf("<%x: %x, %x, %x\n", m, m->m_flags, m->m_len,
m->m_data);
#endif
if (copyhdr)
m = m->m_next;
np = &n->m_next;
#if BLUE_DEBUG
- kprintf(">%x: %x, %x, %x\n", n, n->m_flags, n->m_len,
+ printf(">%x: %x, %x, %x\n", n, n->m_flags, n->m_len,
n->m_data);
#endif
}
space = remainder;
bcopy(cp, mtod(m, caddr_t) + m->m_len, space);
m->m_len += space;
- cp += space, remainder -= space;
+ cp += space;
+ remainder -= space;
}
while (remainder > 0) {
/*
return (mcache_retry);
}
+__attribute__((noreturn))
static void
mbuf_worker_thread(void)
{
while (1) {
lck_mtx_lock(mbuf_mlock);
-
mbuf_expand = 0;
if (mbuf_expand_mcl) {
int n;
}
}
+ mbuf_worker_needs_wakeup = TRUE;
+ assert_wait((caddr_t)&mbuf_worker_needs_wakeup,
+ THREAD_UNINT);
lck_mtx_unlock(mbuf_mlock);
-
- assert_wait(&mbuf_worker_run, THREAD_UNINT);
(void) thread_block((thread_continue_t)mbuf_worker_thread);
}
}
+__attribute__((noreturn))
static void
mbuf_worker_thread_init(void)
{
if ((slg = slabstbl[ix]) == NULL) {
/*
- * In the current implementation, we never shrink the slabs
+ * In the current implementation, we never shrink the slabs
* table; if we attempt to reallocate a cluster group when
* it's already allocated, panic since this is a sign of a
* memory corruption (slabstbl[ix] got nullified).
ix = MTOPG(buf);
VERIFY(ix < maxclaudit);
-
+
if (mclaudit[ix].cl_audit[0] != NULL) {
mca_list = mclaudit[ix].cl_audit[0];
for (i = 0; i < num; i++) {
if (composite) {
struct mbuf *next = m->m_next;
- VERIFY(ms->m_flags == M_EXT && MEXT_RFA(ms) != NULL &&
+ VERIFY(ms->m_flags == M_EXT && m_get_rfa(ms) != NULL &&
MBUF_IS_COMPOSITE(ms));
VERIFY(mca->mca_contents_size == AUDIT_CONTENTS_SIZE);
/*
if ((temp % mleak_table.mleak_sample_factor) == 0 && addr != NULL) {
uintptr_t bt[MLEAK_STACK_DEPTH];
- int logged = fastbacktrace(bt, MLEAK_STACK_DEPTH);
+ int logged = backtrace(bt, MLEAK_STACK_DEPTH);
mleak_log(bt, addr, logged, num);
}
}
return (ret);
}
+int
+m_ext_set_prop(struct mbuf *m, uint32_t o, uint32_t n)
+{
+ ASSERT(m->m_flags & M_EXT);
+ return (atomic_test_set_32(&MEXT_PRIV(m), o, n));
+}
+
+uint32_t
+m_ext_get_prop(struct mbuf *m)
+{
+ ASSERT(m->m_flags & M_EXT);
+ return (MEXT_PRIV(m));
+}
+
+int
+m_ext_paired_is_active(struct mbuf *m)
+{
+ return (MBUF_IS_PAIRED(m) ? (MEXT_PREF(m) > MEXT_MINREF(m)) : 1);
+}
+
+void
+m_ext_paired_activate(struct mbuf *m)
+{
+ struct ext_ref *rfa;
+ int hdr, type;
+ caddr_t extbuf;
+ m_ext_free_func_t extfree;
+ u_int extsize;
+
+ VERIFY(MBUF_IS_PAIRED(m));
+ VERIFY(MEXT_REF(m) == MEXT_MINREF(m));
+ VERIFY(MEXT_PREF(m) == MEXT_MINREF(m));
+
+ hdr = (m->m_flags & M_PKTHDR);
+ type = m->m_type;
+ extbuf = m->m_ext.ext_buf;
+ extfree = m_get_ext_free(m);
+ extsize = m->m_ext.ext_size;
+ rfa = m_get_rfa(m);
+
+ VERIFY(extbuf != NULL && rfa != NULL);
+
+ /*
+ * Safe to reinitialize packet header tags, since it's
+ * already taken care of at m_free() time. Similar to
+ * what's done in m_clattach() for the cluster. Bump
+ * up MEXT_PREF to indicate activation.
+ */
+ MBUF_INIT(m, hdr, type);
+ MEXT_INIT(m, extbuf, extsize, extfree, (caddr_t)m, rfa,
+ 1, 1, 2, EXTF_PAIRED, MEXT_PRIV(m), m);
+}
+
void
m_scratch_init(struct mbuf *m)
{
}
}
+__private_extern__ inline void
+m_set_ext(struct mbuf *m, struct ext_ref *rfa, m_ext_free_func_t ext_free,
+ caddr_t ext_arg)
+{
+ VERIFY(m->m_flags & M_EXT);
+ if (rfa != NULL) {
+ m->m_ext.ext_refflags =
+ (struct ext_ref *)(((uintptr_t)rfa) ^ mb_obscure_extref);
+ if (ext_free != NULL) {
+ rfa->ext_token = ((uintptr_t)&rfa->ext_token) ^
+ mb_obscure_extfree;
+ m->m_ext.ext_free = (m_ext_free_func_t)
+ (((uintptr_t)ext_free) ^ rfa->ext_token);
+ if (ext_arg != NULL) {
+ m->m_ext.ext_arg = (((uintptr_t)ext_arg) ^
+ rfa->ext_token);
+ } else {
+ m->m_ext.ext_arg = NULL;
+ }
+ } else {
+ rfa->ext_token = 0;
+ m->m_ext.ext_free = NULL;
+ m->m_ext.ext_arg = NULL;
+ }
+ } else {
+ /*
+ * If we are going to loose the cookie in ext_token by
+ * resetting the rfa, we should use the global cookie
+ * to obscure the ext_free and ext_arg pointers.
+ */
+ if (ext_free != NULL) {
+ m->m_ext.ext_free = ((uintptr_t)ext_free ^
+ mb_obscure_extfree);
+ if (ext_arg != NULL) {
+ m->m_ext.ext_arg = ((uintptr_t)ext_arg ^
+ mb_obscure_extfree);
+ } else {
+ m->m_ext.ext_arg = NULL;
+ }
+ } else {
+ m->m_ext.ext_free = NULL;
+ m->m_ext.ext_arg = NULL;
+ }
+ m->m_ext.ext_refflags = NULL;
+ }
+}
+
+__private_extern__ inline struct ext_ref *
+m_get_rfa(struct mbuf *m)
+{
+ if (m->m_ext.ext_refflags == NULL)
+ return (NULL);
+ else
+ return ((struct ext_ref *)(((uintptr_t)m->m_ext.ext_refflags) ^ mb_obscure_extref));
+}
+
+__private_extern__ inline m_ext_free_func_t
+m_get_ext_free(struct mbuf *m)
+{
+ struct ext_ref *rfa;
+ if (m->m_ext.ext_free == NULL)
+ return (NULL);
+
+ rfa = m_get_rfa(m);
+ if (rfa == NULL)
+ return ((uintptr_t)m->m_ext.ext_free ^ mb_obscure_extfree);
+ else
+ return ((m_ext_free_func_t)(((uintptr_t)m->m_ext.ext_free)
+ ^ rfa->ext_token));
+}
+
+__private_extern__ inline caddr_t
+m_get_ext_arg(struct mbuf *m)
+{
+ struct ext_ref *rfa;
+ if (m->m_ext.ext_arg == NULL)
+ return (NULL);
+
+ rfa = m_get_rfa(m);
+ if (rfa == NULL) {
+ return ((uintptr_t)m->m_ext.ext_arg ^ mb_obscure_extfree);
+ } else {
+ return ((caddr_t)(((uintptr_t)m->m_ext.ext_arg) ^
+ rfa->ext_token));
+ }
+}
+
/*
* Send a report of mbuf usage if the usage is at least 6% of max limit
* or if there has been at least 3% increase since the last report.
*
* The values 6% and 3% are chosen so that we can do simple arithmetic
* with shift operations.
- */
+ */
static boolean_t
mbuf_report_usage(mbuf_class_t cl)
{
__private_extern__ void
mbuf_report_peak_usage(void)
{
- int i = 0;
+ int i = 0;
u_int64_t uptime;
struct nstat_sysinfo_data ns_data;
uint32_t memreleased = 0;
lck_mtx_lock(mbuf_mlock);
/* Generate an initial report after 1 week of uptime */
- if (!mb_peak_firstreport &&
+ if (!mb_peak_firstreport &&
uptime > MBUF_PEAK_FIRST_REPORT_THRESHOLD) {
mb_peak_newreport = TRUE;
mb_peak_firstreport = TRUE;
}
/*
- * Since a report is being generated before 1 week,
+ * Since a report is being generated before 1 week,
* we do not need to force another one later
*/
if (uptime < MBUF_PEAK_FIRST_REPORT_THRESHOLD)
ns_data.u.mb_stats.sb_atmbuflimit = sbmb_limreached;
ns_data.u.mb_stats.draincnt = mbstat.m_drain;
ns_data.u.mb_stats.memreleased = memreleased;
+ ns_data.u.mb_stats.sbmb_floor = total_sbmb_cnt_floor;
nstat_sysinfo_send_data(&ns_data);
+
+ /*
+ * Reset the floor whenever we report a new
+ * peak to track the trend (increase peek usage
+ * is not a leak if mbufs get released
+ * between reports and the floor stays low)
+ */
+ total_sbmb_cnt_floor = total_sbmb_cnt_peak;
}
/*
mbuf_class_t mc;
mcl_slab_t *sp, *sp_tmp, *nsp;
unsigned int num, k, interval, released = 0;
- unsigned int total_mem = 0, use_mem = 0;
+ unsigned long total_mem = 0, use_mem = 0;
boolean_t ret, purge_caches = FALSE;
ppnum_t offset;
mcache_obj_t *obj;
- float per;
+ unsigned long per;
static uint64_t last_drain = 0;
static unsigned char scratch[32];
static ppnum_t scratch_pa = 0;
lck_mtx_unlock(mbuf_mlock);
return;
}
- interval = net_uptime() - last_drain;
+ interval = net_uptime() - last_drain;
if (interval <= mb_drain_maxint) {
lck_mtx_unlock(mbuf_mlock);
return;
- }
+ }
if (interval <= mb_drain_maxint * 5)
purge_caches = TRUE;
last_drain = net_uptime();
total_mem += m_total(mc) * m_maxsize(mc);
use_mem += m_active(mc) * m_maxsize(mc);
}
- per = (float)use_mem / (float)total_mem;
- if (per >= 0.6) {
+ per = (use_mem * 100) / total_mem;
+ if (per >= 60) {
lck_mtx_unlock(mbuf_mlock);
return;
}
* total in the freelist.
*/
for (mc = 0; mc < NELEM(mbuf_table); mc++) {
- while (m_cobjlist(mc) &&
+ while (m_cobjlist(mc) &&
m_total(mc) < m_avgtotal(mc) &&
m_infree(mc) > 0.1 * m_avgtotal(mc) + m_minlimit(mc)) {
obj = m_cobjlist(mc);
m_total(mc)--;
for (nsp = sp, k = 1; k < NSLABSP16KB; k++) {
nsp = nsp->sl_next;
- VERIFY(nsp->sl_refcnt == 0 &&
+ VERIFY(nsp->sl_refcnt == 0 &&
nsp->sl_base != NULL &&
nsp->sl_len == 0);
slab_init(nsp, 0, 0, NULL, NULL, 0, 0,
*/
IOMapperInsertPage(mcl_paddr_base, offset, scratch_pa);
mcl_paddr[offset] = 0;
- kmem_free(mb_map, (vm_offset_t)sp->sl_base,
+ kmem_free(mb_map, (vm_offset_t)sp->sl_base,
sp->sl_len);
slab_init(sp, 0, 0, NULL, NULL, 0, 0, 0);
sp->sl_flags = 0;
{
#pragma unused(arg1, arg2)
int val = 0, err;
-
+
err = sysctl_handle_int(oidp, &val, 0, req);
if (err != 0 || req->newptr == USER_ADDR_NULL)
return (err);
SYSCTL_INT(_kern_ipc, OID_AUTO, mb_watchdog,
CTLFLAG_RW | CTLFLAG_LOCKED, &mb_watchdog, 0, "");
SYSCTL_PROC(_kern_ipc, OID_AUTO, mb_drain_force,
- CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_LOCKED, NULL, 0,
+ CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_LOCKED, NULL, 0,
m_drain_force_sysctl, "I",
"Forces the mbuf garbage collection to run");
SYSCTL_INT(_kern_ipc, OID_AUTO, mb_drain_maxint,
projects / apple / xnu.git / blobdiff