static int c4iw_mmap(struct ib_ucontext *context, struct vm_area_struct *vma)
{
int len = vma->vm_end - vma->vm_start;
u32 key = vma->vm_pgoff << PAGE_SHIFT;
struct c4iw_rdev *rdev;
int ret = 0;
struct c4iw_mm_entry *mm;
struct c4iw_ucontext *ucontext;
u64 addr, paddr;
u64 va_regs_res = 0, va_udbs_res = 0;
u64 len_regs_res = 0, len_udbs_res = 0;
CTR3(KTR_IW_CXGBE, "%s:1 ctx %p vma %p", __func__, context, vma);
CTR4(KTR_IW_CXGBE, "%s:1a pgoff 0x%lx key 0x%x len %d", __func__,
vma->vm_pgoff, key, len);
if (vma->vm_start & (PAGE_SIZE-1)) {
CTR3(KTR_IW_CXGBE, "%s:2 unaligned vm_start %u vma %p",
__func__, vma->vm_start, vma);
return -EINVAL;
}
rdev = &(to_c4iw_dev(context->device)->rdev);
ucontext = to_c4iw_ucontext(context);
mm = remove_mmap(ucontext, key, len);
if (!mm) {
CTR4(KTR_IW_CXGBE, "%s:3 ucontext %p key %u len %u", __func__,
ucontext, key, len);
return -EINVAL;
}
addr = mm->addr;
kfree(mm);
va_regs_res = (u64)rman_get_virtual(rdev->adap->regs_res);
len_regs_res = (u64)rman_get_size(rdev->adap->regs_res);
va_udbs_res = (u64)rman_get_virtual(rdev->adap->udbs_res);
len_udbs_res = (u64)rman_get_size(rdev->adap->udbs_res);
CTR6(KTR_IW_CXGBE,
"%s:4 addr %p, masync region %p:%p, udb region %p:%p", __func__,
addr, va_regs_res, va_regs_res+len_regs_res, va_udbs_res,
va_udbs_res+len_udbs_res);
if (addr >= va_regs_res && addr < va_regs_res + len_regs_res) {
CTR4(KTR_IW_CXGBE, "%s:5 MA_SYNC addr %p region %p, reglen %u",
__func__, addr, va_regs_res, len_regs_res);
/*
* MA_SYNC register...
*/
paddr = vtophys(addr);
vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot);
ret = io_remap_pfn_range(vma, vma->vm_start,
paddr >> PAGE_SHIFT,
len, vma->vm_page_prot);
} else {
void
pmap_kenter_attr(vm_offset_t va, vm_offset_t pa, vm_memattr_t ma)
{
CTR4(KTR_PMAP, "%s(%#x, %#x, %#x)", __func__, va, pa, ma);
MMU_KENTER_ATTR(mmu_obj, va, pa, ma);
}
void
pmap_change_wiring(pmap_t pmap, vm_offset_t va, boolean_t wired)
{
CTR4(KTR_PMAP, "%s(%p, %#x, %u)", __func__, pmap, va, wired);
MMU_CHANGE_WIRING(mmu_obj, pmap, va, wired);
}
vm_offset_t
pmap_dumpsys_map(struct pmap_md *md, vm_size_t ofs, vm_size_t *sz)
{
CTR4(KTR_PMAP, "%s(%p, %#x, %#x)", __func__, md, ofs, *sz);
return (MMU_DUMPSYS_MAP(mmu_obj, md, ofs, sz));
}
void
pmap_dumpsys_unmap(struct pmap_md *md, vm_size_t ofs, vm_offset_t va)
{
CTR4(KTR_PMAP, "%s(%p, %#x, %#x)", __func__, md, ofs, va);
return (MMU_DUMPSYS_UNMAP(mmu_obj, md, ofs, va));
}
static void
release_offload_resources(struct toepcb *toep)
{
struct tom_data *td = toep->td;
struct adapter *sc = td_adapter(td);
int tid = toep->tid;
KASSERT(toepcb_flag(toep, TPF_CPL_PENDING) == 0,
("%s: %p has CPL pending.", __func__, toep));
KASSERT(toepcb_flag(toep, TPF_ATTACHED) == 0,
("%s: %p is still attached.", __func__, toep));
CTR4(KTR_CXGBE, "%s: toep %p (tid %d, l2te %p)",
__func__, toep, tid, toep->l2te);
if (toep->l2te)
t4_l2t_release(toep->l2te);
if (tid >= 0) {
remove_tid(sc, tid);
release_tid(sc, tid, toep->ctrlq);
}
mtx_lock(&td->toep_list_lock);
TAILQ_REMOVE(&td->toep_list, toep, link);
mtx_unlock(&td->toep_list_lock);
free_toepcb(toep);
}
void
pmap_unwire(pmap_t pmap, vm_offset_t start, vm_offset_t end)
{
CTR4(KTR_PMAP, "%s(%p, %#x, %#x)", __func__, pmap, start, end);
MMU_UNWIRE(mmu_obj, pmap, start, end);
}
void
pmap_sync_icache(pmap_t pm, vm_offset_t va, vm_size_t sz)
{
CTR4(KTR_PMAP, "%s(%p, %#x, %#x)", __func__, pm, va, sz);
return (MMU_SYNC_ICACHE(mmu_obj, pm, va, sz));
}
static int
do_rx_iscsi_hdr(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m)
{
struct adapter *sc = iq->adapter;
struct cpl_iscsi_hdr *cpl = mtod(m, struct cpl_iscsi_hdr *);
u_int tid = GET_TID(cpl);
struct toepcb *toep = lookup_tid(sc, tid);
struct icl_pdu *ip;
struct icl_cxgbei_pdu *icp;
M_ASSERTPKTHDR(m);
ip = icl_cxgbei_new_pdu(M_NOWAIT);
if (ip == NULL)
CXGBE_UNIMPLEMENTED("PDU allocation failure");
icp = ip_to_icp(ip);
bcopy(mtod(m, caddr_t) + sizeof(*cpl), icp->ip.ip_bhs, sizeof(struct
iscsi_bhs));
icp->pdu_seq = ntohl(cpl->seq);
icp->pdu_flags = SBUF_ULP_FLAG_HDR_RCVD;
/* This is the start of a new PDU. There should be no old state. */
MPASS(toep->ulpcb2 == NULL);
toep->ulpcb2 = icp;
#if 0
CTR4(KTR_CXGBE, "%s: tid %u, cpl->len hlen %u, m->m_len hlen %u",
__func__, tid, ntohs(cpl->len), m->m_len);
#endif
m_freem(m);
return (0);
}
static int
do_rx_iscsi_data(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m)
{
struct adapter *sc = iq->adapter;
struct cpl_iscsi_data *cpl = mtod(m, struct cpl_iscsi_data *);
u_int tid = GET_TID(cpl);
struct toepcb *toep = lookup_tid(sc, tid);
struct icl_cxgbei_pdu *icp = toep->ulpcb2;
M_ASSERTPKTHDR(m);
/* Must already have received the header (but not the data). */
MPASS(icp != NULL);
MPASS(icp->pdu_flags == SBUF_ULP_FLAG_HDR_RCVD);
MPASS(icp->ip.ip_data_mbuf == NULL);
MPASS(icp->ip.ip_data_len == 0);
m_adj(m, sizeof(*cpl));
icp->pdu_flags |= SBUF_ULP_FLAG_DATA_RCVD;
icp->ip.ip_data_mbuf = m;
icp->ip.ip_data_len = m->m_pkthdr.len;
#if 0
CTR4(KTR_CXGBE, "%s: tid %u, cpl->len dlen %u, m->m_len dlen %u",
__func__, tid, ntohs(cpl->len), m->m_len);
#endif
return (0);
}
void
dumpsys_unmap_chunk(vm_paddr_t pa, size_t sz, void *va)
{
CTR4(KTR_PMAP, "%s(%#jx, %#zx, %p)", __func__, (uintmax_t)pa, sz, va);
return (MMU_DUMPSYS_UNMAP(mmu_obj, pa, sz, va));
}
void
pmap_zero_page_area(vm_page_t m, int off, int size)
{
CTR4(KTR_PMAP, "%s(%p, %d, %d)", __func__, m, off, size);
MMU_ZERO_PAGE_AREA(mmu_obj, m, off, size);
}
void *
pmap_mapdev_attr(vm_paddr_t pa, vm_size_t size, vm_memattr_t attr)
{
CTR4(KTR_PMAP, "%s(%#x, %#x, %#x)", __func__, pa, size, attr);
return (MMU_MAPDEV_ATTR(mmu_obj, pa, size, attr));
}
void
pmap_remove(pmap_t pmap, vm_offset_t start, vm_offset_t end)
{
CTR4(KTR_PMAP, "%s(%p, %#x, %#x)", __func__, pmap, start, end);
MMU_REMOVE(mmu_obj, pmap, start, end);
}
void
pmap_qenter(vm_offset_t start, vm_page_t *m, int count)
{
CTR4(KTR_PMAP, "%s(%#x, %p, %d)", __func__, start, m, count);
MMU_QENTER(mmu_obj, start, m, count);
}
vm_page_t
pmap_extract_and_hold(pmap_t pmap, vm_offset_t va, vm_prot_t prot)
{
CTR4(KTR_PMAP, "%s(%p, %#x, %#x)", __func__, pmap, va, prot);
return (MMU_EXTRACT_AND_HOLD(mmu_obj, pmap, va, prot));
}
/*
* Intercept the return address from a freshly forked process that has NOT
* been scheduled yet.
*
* This is needed to make kernel threads stay in kernel mode.
*/
void
cpu_set_fork_handler(struct thread *td, void (*func)(void *), void *arg)
{
struct callframe *cf;
CTR4(KTR_PROC, "%s called with td=%p func=%p arg=%p",
__func__, td, func, arg);
cf = (struct callframe *)td->td_pcb->pcb_sp;
cf->cf_func = (register_t)func;
cf->cf_arg0 = (register_t)arg;
}
/*
* Handle the return of a child process from fork1(). This function
* is called from the MD fork_trampoline() entry point.
*/
void
fork_exit(void (*callout)(void *, struct trapframe *), void *arg,
struct trapframe *frame)
{
struct proc *p;
struct thread *td;
struct thread *dtd;
td = curthread;
p = td->td_proc;
KASSERT(p->p_state == PRS_NORMAL, ("executing process is still new"));
CTR4(KTR_PROC, "fork_exit: new thread %p (td_sched %p, pid %d, %s)",
td, td_get_sched(td), p->p_pid, td->td_name);
sched_fork_exit(td);
/*
* Processes normally resume in mi_switch() after being
* cpu_switch()'ed to, but when children start up they arrive here
* instead, so we must do much the same things as mi_switch() would.
*/
if ((dtd = PCPU_GET(deadthread))) {
PCPU_SET(deadthread, NULL);
thread_stash(dtd);
}
thread_unlock(td);
/*
* cpu_fork_kthread_handler intercepts this function call to
* have this call a non-return function to stay in kernel mode.
* initproc has its own fork handler, but it does return.
*/
KASSERT(callout != NULL, ("NULL callout in fork_exit"));
callout(arg, frame);
/*
* Check if a kernel thread misbehaved and returned from its main
* function.
*/
if (p->p_flag & P_KPROC) {
printf("Kernel thread \"%s\" (pid %d) exited prematurely.\n",
td->td_name, p->p_pid);
kthread_exit();
}
mtx_assert(&Giant, MA_NOTOWNED);
if (p->p_sysent->sv_schedtail != NULL)
(p->p_sysent->sv_schedtail)(td);
td->td_pflags &= ~TDP_FORKING;
}
void
sema_init(struct sema *sema, int value, const char *description)
{
KASSERT((value >= 0), ("%s(): negative value\n", __func__));
bzero(sema, sizeof(*sema));
mtx_init(&sema->sema_mtx, description, "sema backing lock",
MTX_DEF | MTX_NOWITNESS | MTX_QUIET);
cv_init(&sema->sema_cv, description);
sema->sema_value = value;
CTR4(KTR_LOCK, "%s(%p, %d, \"%s\")", __func__, sema, value, description);
}
static int alloc_host_sq(struct c4iw_rdev *rdev, struct t4_sq *sq)
{
sq->queue = contigmalloc(sq->memsize, M_DEVBUF, M_NOWAIT, 0ul, ~0ul,
4096, 0);
if (sq->queue)
sq->dma_addr = vtophys(sq->queue);
else
return -ENOMEM;
sq->phys_addr = vtophys(sq->queue);
pci_unmap_addr_set(sq, mapping, sq->dma_addr);
CTR4(KTR_IW_CXGBE, "%s sq %p dma_addr %p phys_addr %p", __func__,
sq->queue, sq->dma_addr, sq->phys_addr);
return 0;
}
static __inline void
call_trapsignal(struct thread *td, int sig, int code, vm_offset_t addr)
{
ksiginfo_t ksi;
CTR4(KTR_TRAP, "%s: addr: %#x, sig: %d, code: %d",
__func__, addr, sig, code);
/*
* TODO: some info would be nice to know
* if we are serving data or prefetch abort.
*/
ksiginfo_init_trap(&ksi);
ksi.ksi_signo = sig;
ksi.ksi_code = code;
ksi.ksi_addr = (void *)addr;
trapsignal(td, &ksi);
}
static void post_qp_event(struct c4iw_dev *dev, struct c4iw_cq *chp,
struct c4iw_qp *qhp,
struct t4_cqe *err_cqe,
enum ib_event_type ib_event)
{
struct ib_event event;
struct c4iw_qp_attributes attrs;
unsigned long flag;
if ((qhp->attr.state == C4IW_QP_STATE_ERROR) ||
(qhp->attr.state == C4IW_QP_STATE_TERMINATE)) {
CTR4(KTR_IW_CXGBE, "%s AE received after RTS - "
"qp state %d qpid 0x%x status 0x%x", __func__,
qhp->attr.state, qhp->wq.sq.qid, CQE_STATUS(err_cqe));
return;
}
printf("AE qpid 0x%x opcode %d status 0x%x "
"type %d wrid.hi 0x%x wrid.lo 0x%x\n",
CQE_QPID(err_cqe), CQE_OPCODE(err_cqe),
CQE_STATUS(err_cqe), CQE_TYPE(err_cqe),
CQE_WRID_HI(err_cqe), CQE_WRID_LOW(err_cqe));
if (qhp->attr.state == C4IW_QP_STATE_RTS) {
attrs.next_state = C4IW_QP_STATE_TERMINATE;
c4iw_modify_qp(qhp->rhp, qhp, C4IW_QP_ATTR_NEXT_STATE,
&attrs, 0);
}
event.event = ib_event;
event.device = chp->ibcq.device;
if (ib_event == IB_EVENT_CQ_ERR)
event.element.cq = &chp->ibcq;
else
event.element.qp = &qhp->ibqp;
if (qhp->ibqp.event_handler)
(*qhp->ibqp.event_handler)(&event, qhp->ibqp.qp_context);
spin_lock_irqsave(&chp->comp_handler_lock, flag);
(*chp->ibcq.comp_handler)(&chp->ibcq, chp->ibcq.cq_context);
spin_unlock_irqrestore(&chp->comp_handler_lock, flag);
}
/*
* Called when a connection is established to translate the TCP options
* reported by HW to FreeBSD's native format.
*/
static void
assign_rxopt(struct tcpcb *tp, unsigned int opt)
{
struct toepcb *toep = tp->t_toe;
struct inpcb *inp = tp->t_inpcb;
struct adapter *sc = td_adapter(toep->td);
int n;
INP_LOCK_ASSERT(inp);
if (inp->inp_inc.inc_flags & INC_ISIPV6)
n = sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
else
n = sizeof(struct ip) + sizeof(struct tcphdr);
tp->t_maxseg = sc->params.mtus[G_TCPOPT_MSS(opt)] - n;
CTR4(KTR_CXGBE, "%s: tid %d, mtu_idx %u (%u)", __func__, toep->tid,
G_TCPOPT_MSS(opt), sc->params.mtus[G_TCPOPT_MSS(opt)]);
if (G_TCPOPT_TSTAMP(opt)) {
tp->t_flags |= TF_RCVD_TSTMP; /* timestamps ok */
tp->ts_recent = 0; /* hmmm */
tp->ts_recent_age = tcp_ts_getsbintime();
tp->t_maxseg -= TCPOLEN_TSTAMP_APPA;
}
if (G_TCPOPT_SACK(opt))
tp->t_flags |= TF_SACK_PERMIT; /* should already be set */
else
tp->t_flags &= ~TF_SACK_PERMIT; /* sack disallowed by peer */
if (G_TCPOPT_WSCALE_OK(opt))
tp->t_flags |= TF_RCVD_SCALE;
/* Doing window scaling? */
if ((tp->t_flags & (TF_RCVD_SCALE | TF_REQ_SCALE)) ==
(TF_RCVD_SCALE | TF_REQ_SCALE)) {
tp->rcv_scale = tp->request_r_scale;
tp->snd_scale = G_TCPOPT_SND_WSCALE(opt);
}
}
int
t3_ddp_copy(const struct mbuf *m, int offset, struct uio *uio, int len)
{
int resid_init, err;
struct ddp_gather_list *gl = (struct ddp_gather_list *)m->m_ddp_gl;
resid_init = uio->uio_resid;
if (!gl->dgl_pages)
panic("pages not set\n");
CTR4(KTR_TOM, "t3_ddp_copy: offset=%d dgl_offset=%d cur_offset=%d len=%d",
offset, gl->dgl_offset, m->m_cur_offset, len);
offset += gl->dgl_offset + m->m_cur_offset;
KASSERT(len <= gl->dgl_length,
("len=%d > dgl_length=%d in ddp_copy\n", len, gl->dgl_length));
err = uiomove_fromphys(gl->dgl_pages, offset, len, uio);
return (err);
}
void cxio_dump_pbl(struct cxio_rdev *rdev, uint32_t pbl_addr, uint32_t len, u8 shift)
{
struct ch_mem_range *m;
u64 *data;
int rc;
int size, npages;
shift += 12;
npages = (len + (1ULL << shift) - 1) >> shift;
size = npages * sizeof(u64);
m = kmalloc(sizeof(*m) + size, M_NOWAIT);
if (!m) {
CTR1(KTR_IW_CXGB, "%s couldn't allocate memory.", __FUNCTION__);
return;
}
m->mem_id = MEM_PMRX;
m->addr = pbl_addr;
m->len = size;
CTR4(KTR_IW_CXGB, "%s PBL addr 0x%x len %d depth %d",
__FUNCTION__, m->addr, m->len, npages);
rc = rdev->t3cdev_p->ctl(rdev->t3cdev_p, RDMA_GET_MEM, m);
if (rc) {
CTR2(KTR_IW_CXGB, "%s toectl returned error %d", __FUNCTION__, rc);
free(m, M_DEVBUF);
return;
}
data = (u64 *)m->buf;
while (size > 0) {
CTR2(KTR_IW_CXGB, "PBL %08x: %016llx", m->addr, (unsigned long long) *data);
size -= 8;
data++;
m->addr += 8;
}
free(m, M_DEVBUF);
}
/*
* _mtx_lock_sleep: the tougher part of acquiring an MTX_DEF lock.
*
* We call this if the lock is either contested (i.e. we need to go to
* sleep waiting for it), or if we need to recurse on it.
*/
void
_mtx_lock_sleep(struct mtx *m, uintptr_t tid, int opts, const char *file,
int line)
{
struct turnstile *ts;
uintptr_t v;
#ifdef ADAPTIVE_MUTEXES
volatile struct thread *owner;
#endif
#ifdef KTR
int cont_logged = 0;
#endif
#ifdef LOCK_PROFILING
int contested = 0;
uint64_t waittime = 0;
#endif
#ifdef KDTRACE_HOOKS
uint64_t spin_cnt = 0;
uint64_t sleep_cnt = 0;
int64_t sleep_time = 0;
#endif
if (mtx_owned(m)) {
KASSERT((m->lock_object.lo_flags & LO_RECURSABLE) != 0,
("_mtx_lock_sleep: recursed on non-recursive mutex %s @ %s:%d\n",
m->lock_object.lo_name, file, line));
m->mtx_recurse++;
atomic_set_ptr(&m->mtx_lock, MTX_RECURSED);
if (LOCK_LOG_TEST(&m->lock_object, opts))
CTR1(KTR_LOCK, "_mtx_lock_sleep: %p recursing", m);
return;
}
lock_profile_obtain_lock_failed(&m->lock_object,
&contested, &waittime);
if (LOCK_LOG_TEST(&m->lock_object, opts))
CTR4(KTR_LOCK,
"_mtx_lock_sleep: %s contested (lock=%p) at %s:%d",
m->lock_object.lo_name, (void *)m->mtx_lock, file, line);
while (!_obtain_lock(m, tid)) {
#ifdef KDTRACE_HOOKS
spin_cnt++;
#endif
#ifdef ADAPTIVE_MUTEXES
/*
* If the owner is running on another CPU, spin until the
* owner stops running or the state of the lock changes.
*/
v = m->mtx_lock;
if (v != MTX_UNOWNED) {
owner = (struct thread *)(v & ~MTX_FLAGMASK);
if (TD_IS_RUNNING(owner)) {
if (LOCK_LOG_TEST(&m->lock_object, 0))
CTR3(KTR_LOCK,
"%s: spinning on %p held by %p",
__func__, m, owner);
while (mtx_owner(m) == owner &&
TD_IS_RUNNING(owner)) {
cpu_spinwait();
#ifdef KDTRACE_HOOKS
spin_cnt++;
#endif
}
continue;
}
}
#endif
ts = turnstile_trywait(&m->lock_object);
v = m->mtx_lock;
/*
* Check if the lock has been released while spinning for
* the turnstile chain lock.
*/
if (v == MTX_UNOWNED) {
turnstile_cancel(ts);
continue;
}
#ifdef ADAPTIVE_MUTEXES
/*
* The current lock owner might have started executing
* on another CPU (or the lock could have changed
* owners) while we were waiting on the turnstile
* chain lock. If so, drop the turnstile lock and try
* again.
*/
owner = (struct thread *)(v & ~MTX_FLAGMASK);
if (TD_IS_RUNNING(owner)) {
turnstile_cancel(ts);
continue;
}
#endif
/*
* If the mutex isn't already contested and a failure occurs
* setting the contested bit, the mutex was either released
* or the state of the MTX_RECURSED bit changed.
*/
if ((v & MTX_CONTESTED) == 0 &&
!atomic_cmpset_ptr(&m->mtx_lock, v, v | MTX_CONTESTED)) {
turnstile_cancel(ts);
continue;
}
/*
* We definitely must sleep for this lock.
*/
mtx_assert(m, MA_NOTOWNED);
#ifdef KTR
if (!cont_logged) {
CTR6(KTR_CONTENTION,
"contention: %p at %s:%d wants %s, taken by %s:%d",
(void *)tid, file, line, m->lock_object.lo_name,
WITNESS_FILE(&m->lock_object),
WITNESS_LINE(&m->lock_object));
cont_logged = 1;
}
#endif
/*
* Block on the turnstile.
*/
#ifdef KDTRACE_HOOKS
sleep_time -= lockstat_nsecs();
#endif
turnstile_wait(ts, mtx_owner(m), TS_EXCLUSIVE_QUEUE);
#ifdef KDTRACE_HOOKS
sleep_time += lockstat_nsecs();
sleep_cnt++;
#endif
}
#ifdef KTR
if (cont_logged) {
CTR4(KTR_CONTENTION,
"contention end: %s acquired by %p at %s:%d",
m->lock_object.lo_name, (void *)tid, file, line);
}
#endif
LOCKSTAT_PROFILE_OBTAIN_LOCK_SUCCESS(LS_MTX_LOCK_ACQUIRE, m, contested,
waittime, file, line);
#ifdef KDTRACE_HOOKS
if (sleep_time)
LOCKSTAT_RECORD1(LS_MTX_LOCK_BLOCK, m, sleep_time);
/*
* Only record the loops spinning and not sleeping.
*/
if (spin_cnt > sleep_cnt)
LOCKSTAT_RECORD1(LS_MTX_LOCK_SPIN, m, (spin_cnt - sleep_cnt));
#endif
}
void
_rw_runlock_cookie(volatile uintptr_t *c, const char *file, int line)
{
struct rwlock *rw;
struct turnstile *ts;
uintptr_t x, v, queue;
if (SCHEDULER_STOPPED())
return;
rw = rwlock2rw(c);
KASSERT(rw->rw_lock != RW_DESTROYED,
("rw_runlock() of destroyed rwlock @ %s:%d", file, line));
__rw_assert(c, RA_RLOCKED, file, line);
WITNESS_UNLOCK(&rw->lock_object, 0, file, line);
LOCK_LOG_LOCK("RUNLOCK", &rw->lock_object, 0, 0, file, line);
/* TODO: drop "owner of record" here. */
for (;;) {
/*
* See if there is more than one read lock held. If so,
* just drop one and return.
*/
x = rw->rw_lock;
if (RW_READERS(x) > 1) {
if (atomic_cmpset_rel_ptr(&rw->rw_lock, x,
x - RW_ONE_READER)) {
if (LOCK_LOG_TEST(&rw->lock_object, 0))
CTR4(KTR_LOCK,
"%s: %p succeeded %p -> %p",
__func__, rw, (void *)x,
(void *)(x - RW_ONE_READER));
break;
}
continue;
}
/*
* If there aren't any waiters for a write lock, then try
* to drop it quickly.
*/
if (!(x & RW_LOCK_WAITERS)) {
MPASS((x & ~RW_LOCK_WRITE_SPINNER) ==
RW_READERS_LOCK(1));
if (atomic_cmpset_rel_ptr(&rw->rw_lock, x,
RW_UNLOCKED)) {
if (LOCK_LOG_TEST(&rw->lock_object, 0))
CTR2(KTR_LOCK, "%s: %p last succeeded",
__func__, rw);
break;
}
continue;
}
/*
* Ok, we know we have waiters and we think we are the
* last reader, so grab the turnstile lock.
*/
turnstile_chain_lock(&rw->lock_object);
v = rw->rw_lock & (RW_LOCK_WAITERS | RW_LOCK_WRITE_SPINNER);
MPASS(v & RW_LOCK_WAITERS);
/*
* Try to drop our lock leaving the lock in a unlocked
* state.
*
* If you wanted to do explicit lock handoff you'd have to
* do it here. You'd also want to use turnstile_signal()
* and you'd have to handle the race where a higher
* priority thread blocks on the write lock before the
* thread you wakeup actually runs and have the new thread
* "steal" the lock. For now it's a lot simpler to just
* wakeup all of the waiters.
*
* As above, if we fail, then another thread might have
* acquired a read lock, so drop the turnstile lock and
* restart.
*/
x = RW_UNLOCKED;
if (v & RW_LOCK_WRITE_WAITERS) {
queue = TS_EXCLUSIVE_QUEUE;
x |= (v & RW_LOCK_READ_WAITERS);
} else
queue = TS_SHARED_QUEUE;
if (!atomic_cmpset_rel_ptr(&rw->rw_lock, RW_READERS_LOCK(1) | v,
x)) {
turnstile_chain_unlock(&rw->lock_object);
continue;
}
if (LOCK_LOG_TEST(&rw->lock_object, 0))
CTR2(KTR_LOCK, "%s: %p last succeeded with waiters",
__func__, rw);
/*
* Ok. The lock is released and all that's left is to
* wake up the waiters. Note that the lock might not be
* free anymore, but in that case the writers will just
* block again if they run before the new lock holder(s)
* release the lock.
*/
ts = turnstile_lookup(&rw->lock_object);
MPASS(ts != NULL);
turnstile_broadcast(ts, queue);
turnstile_unpend(ts, TS_SHARED_LOCK);
turnstile_chain_unlock(&rw->lock_object);
break;
}
LOCKSTAT_PROFILE_RELEASE_RWLOCK(rw__release, rw, LOCKSTAT_READER);
curthread->td_locks--;
curthread->td_rw_rlocks--;
}
void
__rw_rlock(volatile uintptr_t *c, const char *file, int line)
{
struct rwlock *rw;
struct turnstile *ts;
#ifdef ADAPTIVE_RWLOCKS
volatile struct thread *owner;
int spintries = 0;
int i;
#endif
#ifdef LOCK_PROFILING
uint64_t waittime = 0;
int contested = 0;
#endif
uintptr_t v;
#ifdef KDTRACE_HOOKS
uintptr_t state;
uint64_t spin_cnt = 0;
uint64_t sleep_cnt = 0;
int64_t sleep_time = 0;
int64_t all_time = 0;
#endif
if (SCHEDULER_STOPPED())
return;
rw = rwlock2rw(c);
KASSERT(kdb_active != 0 || !TD_IS_IDLETHREAD(curthread),
("rw_rlock() by idle thread %p on rwlock %s @ %s:%d",
curthread, rw->lock_object.lo_name, file, line));
KASSERT(rw->rw_lock != RW_DESTROYED,
("rw_rlock() of destroyed rwlock @ %s:%d", file, line));
KASSERT(rw_wowner(rw) != curthread,
("rw_rlock: wlock already held for %s @ %s:%d",
rw->lock_object.lo_name, file, line));
WITNESS_CHECKORDER(&rw->lock_object, LOP_NEWORDER, file, line, NULL);
#ifdef KDTRACE_HOOKS
all_time -= lockstat_nsecs(&rw->lock_object);
state = rw->rw_lock;
#endif
for (;;) {
/*
* Handle the easy case. If no other thread has a write
* lock, then try to bump up the count of read locks. Note
* that we have to preserve the current state of the
* RW_LOCK_WRITE_WAITERS flag. If we fail to acquire a
* read lock, then rw_lock must have changed, so restart
* the loop. Note that this handles the case of a
* completely unlocked rwlock since such a lock is encoded
* as a read lock with no waiters.
*/
v = rw->rw_lock;
if (RW_CAN_READ(v)) {
/*
* The RW_LOCK_READ_WAITERS flag should only be set
* if the lock has been unlocked and write waiters
* were present.
*/
if (atomic_cmpset_acq_ptr(&rw->rw_lock, v,
v + RW_ONE_READER)) {
if (LOCK_LOG_TEST(&rw->lock_object, 0))
CTR4(KTR_LOCK,
"%s: %p succeed %p -> %p", __func__,
rw, (void *)v,
(void *)(v + RW_ONE_READER));
break;
}
continue;
}
#ifdef KDTRACE_HOOKS
spin_cnt++;
#endif
#ifdef HWPMC_HOOKS
PMC_SOFT_CALL( , , lock, failed);
#endif
lock_profile_obtain_lock_failed(&rw->lock_object,
&contested, &waittime);
#ifdef ADAPTIVE_RWLOCKS
/*
* If the owner is running on another CPU, spin until
* the owner stops running or the state of the lock
* changes.
*/
if ((v & RW_LOCK_READ) == 0) {
owner = (struct thread *)RW_OWNER(v);
if (TD_IS_RUNNING(owner)) {
if (LOCK_LOG_TEST(&rw->lock_object, 0))
CTR3(KTR_LOCK,
"%s: spinning on %p held by %p",
__func__, rw, owner);
KTR_STATE1(KTR_SCHED, "thread",
sched_tdname(curthread), "spinning",
"lockname:\"%s\"", rw->lock_object.lo_name);
while ((struct thread*)RW_OWNER(rw->rw_lock) ==
owner && TD_IS_RUNNING(owner)) {
cpu_spinwait();
#ifdef KDTRACE_HOOKS
spin_cnt++;
#endif
}
KTR_STATE0(KTR_SCHED, "thread",
sched_tdname(curthread), "running");
continue;
}
} else if (spintries < rowner_retries) {
spintries++;
KTR_STATE1(KTR_SCHED, "thread", sched_tdname(curthread),
"spinning", "lockname:\"%s\"",
rw->lock_object.lo_name);
for (i = 0; i < rowner_loops; i++) {
v = rw->rw_lock;
if ((v & RW_LOCK_READ) == 0 || RW_CAN_READ(v))
break;
cpu_spinwait();
}
#ifdef KDTRACE_HOOKS
spin_cnt += rowner_loops - i;
#endif
KTR_STATE0(KTR_SCHED, "thread", sched_tdname(curthread),
"running");
if (i != rowner_loops)
continue;
}
#endif
/*
* Okay, now it's the hard case. Some other thread already
* has a write lock or there are write waiters present,
* acquire the turnstile lock so we can begin the process
* of blocking.
*/
ts = turnstile_trywait(&rw->lock_object);
/*
* The lock might have been released while we spun, so
* recheck its state and restart the loop if needed.
*/
v = rw->rw_lock;
if (RW_CAN_READ(v)) {
turnstile_cancel(ts);
continue;
}
#ifdef ADAPTIVE_RWLOCKS
/*
* The current lock owner might have started executing
* on another CPU (or the lock could have changed
* owners) while we were waiting on the turnstile
* chain lock. If so, drop the turnstile lock and try
* again.
*/
if ((v & RW_LOCK_READ) == 0) {
owner = (struct thread *)RW_OWNER(v);
if (TD_IS_RUNNING(owner)) {
turnstile_cancel(ts);
continue;
}
}
#endif
/*
* The lock is held in write mode or it already has waiters.
*/
MPASS(!RW_CAN_READ(v));
/*
* If the RW_LOCK_READ_WAITERS flag is already set, then
* we can go ahead and block. If it is not set then try
* to set it. If we fail to set it drop the turnstile
* lock and restart the loop.
*/
if (!(v & RW_LOCK_READ_WAITERS)) {
if (!atomic_cmpset_ptr(&rw->rw_lock, v,
v | RW_LOCK_READ_WAITERS)) {
turnstile_cancel(ts);
continue;
}
if (LOCK_LOG_TEST(&rw->lock_object, 0))
CTR2(KTR_LOCK, "%s: %p set read waiters flag",
__func__, rw);
}
/*
* We were unable to acquire the lock and the read waiters
* flag is set, so we must block on the turnstile.
*/
if (LOCK_LOG_TEST(&rw->lock_object, 0))
CTR2(KTR_LOCK, "%s: %p blocking on turnstile", __func__,
rw);
#ifdef KDTRACE_HOOKS
sleep_time -= lockstat_nsecs(&rw->lock_object);
#endif
turnstile_wait(ts, rw_owner(rw), TS_SHARED_QUEUE);
#ifdef KDTRACE_HOOKS
sleep_time += lockstat_nsecs(&rw->lock_object);
sleep_cnt++;
#endif
if (LOCK_LOG_TEST(&rw->lock_object, 0))
CTR2(KTR_LOCK, "%s: %p resuming from turnstile",
__func__, rw);
}
#ifdef KDTRACE_HOOKS
all_time += lockstat_nsecs(&rw->lock_object);
if (sleep_time)
LOCKSTAT_RECORD4(rw__block, rw, sleep_time,
LOCKSTAT_READER, (state & RW_LOCK_READ) == 0,
(state & RW_LOCK_READ) == 0 ? 0 : RW_READERS(state));
/* Record only the loops spinning and not sleeping. */
if (spin_cnt > sleep_cnt)
LOCKSTAT_RECORD4(rw__spin, rw, all_time - sleep_time,
LOCKSTAT_READER, (state & RW_LOCK_READ) == 0,
(state & RW_LOCK_READ) == 0 ? 0 : RW_READERS(state));
#endif
/*
* TODO: acquire "owner of record" here. Here be turnstile dragons
* however. turnstiles don't like owners changing between calls to
* turnstile_wait() currently.
*/
LOCKSTAT_PROFILE_OBTAIN_RWLOCK_SUCCESS(rw__acquire, rw, contested,
waittime, file, line, LOCKSTAT_READER);
LOCK_LOG_LOCK("RLOCK", &rw->lock_object, 0, 0, file, line);
WITNESS_LOCK(&rw->lock_object, 0, file, line);
curthread->td_locks++;
curthread->td_rw_rlocks++;
}
static int
do_rx_iscsi_ddp(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m)
{
struct adapter *sc = iq->adapter;
struct cxgbei_data *ci = sc->iscsi_ulp_softc;
const struct cpl_rx_data_ddp *cpl = (const void *)(rss + 1);
u_int tid = GET_TID(cpl);
struct toepcb *toep = lookup_tid(sc, tid);
struct inpcb *inp = toep->inp;
struct socket *so;
struct sockbuf *sb;
struct tcpcb *tp;
struct icl_cxgbei_conn *icc;
struct icl_conn *ic;
struct icl_cxgbei_pdu *icp = toep->ulpcb2;
struct icl_pdu *ip;
u_int pdu_len, val;
MPASS(m == NULL);
/* Must already be assembling a PDU. */
MPASS(icp != NULL);
MPASS(icp->icp_flags & ICPF_RX_HDR); /* Data is optional. */
MPASS((icp->icp_flags & ICPF_RX_STATUS) == 0);
pdu_len = be16toh(cpl->len); /* includes everything. */
val = be32toh(cpl->ddpvld);
#if 0
CTR5(KTR_CXGBE,
"%s: tid %u, cpl->len %u, ddpvld 0x%08x, icp_flags 0x%08x",
__func__, tid, pdu_len, val, icp->icp_flags);
#endif
icp->icp_flags |= ICPF_RX_STATUS;
ip = &icp->ip;
if (val & F_DDP_PADDING_ERR)
icp->icp_flags |= ICPF_PAD_ERR;
if (val & F_DDP_HDRCRC_ERR)
icp->icp_flags |= ICPF_HCRC_ERR;
if (val & F_DDP_DATACRC_ERR)
icp->icp_flags |= ICPF_DCRC_ERR;
if (val & F_DDP_PDU && ip->ip_data_mbuf == NULL) {
MPASS((icp->icp_flags & ICPF_RX_FLBUF) == 0);
MPASS(ip->ip_data_len > 0);
icp->icp_flags |= ICPF_RX_DDP;
counter_u64_add(ci->ddp_pdus, 1);
counter_u64_add(ci->ddp_bytes, ip->ip_data_len);
}
INP_WLOCK(inp);
if (__predict_false(inp->inp_flags & (INP_DROPPED | INP_TIMEWAIT))) {
CTR4(KTR_CXGBE, "%s: tid %u, rx (%d bytes), inp_flags 0x%x",
__func__, tid, pdu_len, inp->inp_flags);
INP_WUNLOCK(inp);
icl_cxgbei_conn_pdu_free(NULL, ip);
#ifdef INVARIANTS
toep->ulpcb2 = NULL;
#endif
return (0);
}
tp = intotcpcb(inp);
MPASS(icp->icp_seq == tp->rcv_nxt);
MPASS(tp->rcv_wnd >= pdu_len);
tp->rcv_nxt += pdu_len;
tp->rcv_wnd -= pdu_len;
tp->t_rcvtime = ticks;
/* update rx credits */
toep->rx_credits += pdu_len;
t4_rcvd(&toep->td->tod, tp); /* XXX: sc->tom_softc.tod */
so = inp->inp_socket;
sb = &so->so_rcv;
SOCKBUF_LOCK(sb);
icc = toep->ulpcb;
if (__predict_false(icc == NULL || sb->sb_state & SBS_CANTRCVMORE)) {
CTR5(KTR_CXGBE,
"%s: tid %u, excess rx (%d bytes), icc %p, sb_state 0x%x",
__func__, tid, pdu_len, icc, sb->sb_state);
SOCKBUF_UNLOCK(sb);
INP_WUNLOCK(inp);
INP_INFO_RLOCK(&V_tcbinfo);
INP_WLOCK(inp);
tp = tcp_drop(tp, ECONNRESET);
if (tp)
INP_WUNLOCK(inp);
INP_INFO_RUNLOCK(&V_tcbinfo);
icl_cxgbei_conn_pdu_free(NULL, ip);
#ifdef INVARIANTS
toep->ulpcb2 = NULL;
#endif
return (0);
}
MPASS(icc->icc_signature == CXGBEI_CONN_SIGNATURE);
ic = &icc->ic;
icl_cxgbei_new_pdu_set_conn(ip, ic);
MPASS(m == NULL); /* was unused, we'll use it now. */
m = sbcut_locked(sb, sbused(sb)); /* XXXNP: toep->sb_cc accounting? */
if (__predict_false(m != NULL)) {
int len = m_length(m, NULL);
/*
* PDUs were received before the tid transitioned to ULP mode.
* Convert them to icl_cxgbei_pdus and send them to ICL before
* the PDU in icp/ip.
*/
CTR3(KTR_CXGBE, "%s: tid %u, %u bytes in so_rcv", __func__, tid,
len);
/* XXXNP: needs to be rewritten. */
if (len == sizeof(struct iscsi_bhs) || len == 4 + sizeof(struct
iscsi_bhs)) {
struct icl_cxgbei_pdu *icp0;
struct icl_pdu *ip0;
ip0 = icl_cxgbei_new_pdu(M_NOWAIT);
icl_cxgbei_new_pdu_set_conn(ip0, ic);
if (ip0 == NULL)
CXGBE_UNIMPLEMENTED("PDU allocation failure");
icp0 = ip_to_icp(ip0);
icp0->icp_seq = 0; /* XXX */
icp0->icp_flags = ICPF_RX_HDR | ICPF_RX_STATUS;
m_copydata(m, 0, sizeof(struct iscsi_bhs), (void *)ip0->ip_bhs);
STAILQ_INSERT_TAIL(&icc->rcvd_pdus, ip0, ip_next);
}
m_freem(m);
}
STAILQ_INSERT_TAIL(&icc->rcvd_pdus, ip, ip_next);
if ((icc->rx_flags & RXF_ACTIVE) == 0) {
struct cxgbei_worker_thread_softc *cwt = &cwt_softc[icc->cwt];
mtx_lock(&cwt->cwt_lock);
icc->rx_flags |= RXF_ACTIVE;
TAILQ_INSERT_TAIL(&cwt->rx_head, icc, rx_link);
if (cwt->cwt_state == CWT_SLEEPING) {
cwt->cwt_state = CWT_RUNNING;
cv_signal(&cwt->cwt_cv);
}
mtx_unlock(&cwt->cwt_lock);
}
SOCKBUF_UNLOCK(sb);
INP_WUNLOCK(inp);
#ifdef INVARIANTS
toep->ulpcb2 = NULL;
#endif
return (0);
}
void
sendsig(sig_t catcher, ksiginfo_t *ksi, sigset_t *mask)
{
struct thread *td;
struct proc *p;
struct trapframe *tf;
struct sigframe *fp, frame;
struct sigacts *psp;
int code, onstack, sig;
td = curthread;
p = td->td_proc;
PROC_LOCK_ASSERT(p, MA_OWNED);
sig = ksi->ksi_signo;
code = ksi->ksi_code;
psp = p->p_sigacts;
mtx_assert(&psp->ps_mtx, MA_OWNED);
tf = td->td_frame;
onstack = sigonstack(tf->tf_sp);
CTR4(KTR_SIG, "sendsig: td=%p (%s) catcher=%p sig=%d", td, p->p_comm,
catcher, sig);
/* Allocate and validate space for the signal handler context. */
if ((td->td_pflags & TDP_ALTSTACK) != 0 && !onstack &&
SIGISMEMBER(psp->ps_sigonstack, sig)) {
fp = (struct sigframe *)(td->td_sigstk.ss_sp +
td->td_sigstk.ss_size);
#if defined(COMPAT_43)
td->td_sigstk.ss_flags |= SS_ONSTACK;
#endif
} else {
fp = (struct sigframe *)td->td_frame->tf_sp;
}
/* Make room, keeping the stack aligned */
fp--;
fp = (struct sigframe *)STACKALIGN(fp);
/* Fill in the frame to copy out */
get_mcontext(td, &frame.sf_uc.uc_mcontext, 0);
get_fpcontext(td, &frame.sf_uc.uc_mcontext);
frame.sf_si = ksi->ksi_info;
frame.sf_uc.uc_sigmask = *mask;
frame.sf_uc.uc_stack.ss_flags = (td->td_pflags & TDP_ALTSTACK) ?
((onstack) ? SS_ONSTACK : 0) : SS_DISABLE;
frame.sf_uc.uc_stack = td->td_sigstk;
mtx_unlock(&psp->ps_mtx);
PROC_UNLOCK(td->td_proc);
/* Copy the sigframe out to the user's stack. */
if (copyout(&frame, fp, sizeof(*fp)) != 0) {
/* Process has trashed its stack. Kill it. */
CTR2(KTR_SIG, "sendsig: sigexit td=%p fp=%p", td, fp);
PROC_LOCK(p);
sigexit(td, SIGILL);
}
tf->tf_x[0]= sig;
tf->tf_x[1] = (register_t)&fp->sf_si;
tf->tf_x[2] = (register_t)&fp->sf_uc;
tf->tf_elr = (register_t)catcher;
tf->tf_sp = (register_t)fp;
tf->tf_lr = (register_t)(PS_STRINGS - *(p->p_sysent->sv_szsigcode));
CTR3(KTR_SIG, "sendsig: return td=%p pc=%#x sp=%#x", td, tf->tf_elr,
tf->tf_sp);
PROC_LOCK(p);
mtx_lock(&psp->ps_mtx);
}
