void
dmar_qi_invalidate_locked(struct dmar_domain *domain, dmar_gaddr_t base,
dmar_gaddr_t size, struct dmar_qi_genseq *pseq)
{
struct dmar_unit *unit;
dmar_gaddr_t isize;
int am;
unit = domain->dmar;
DMAR_ASSERT_LOCKED(unit);
for (; size > 0; base += isize, size -= isize) {
am = calc_am(unit, base, size, &isize);
dmar_qi_ensure(unit, 1);
dmar_qi_emit(unit, DMAR_IQ_DESCR_IOTLB_INV |
DMAR_IQ_DESCR_IOTLB_PAGE | DMAR_IQ_DESCR_IOTLB_DW |
DMAR_IQ_DESCR_IOTLB_DR |
DMAR_IQ_DESCR_IOTLB_DID(domain->domain),
base | am);
}
if (pseq != NULL) {
dmar_qi_ensure(unit, 1);
dmar_qi_emit_wait_seq(unit, pseq);
}
dmar_qi_advance_tail(unit);
}
void
dmar_qi_invalidate_iec_glob(struct dmar_unit *unit)
{
struct dmar_qi_genseq gseq;
DMAR_ASSERT_LOCKED(unit);
dmar_qi_ensure(unit, 2);
dmar_qi_emit(unit, DMAR_IQ_DESCR_IEC_INV, 0);
dmar_qi_emit_wait_seq(unit, &gseq, true);
dmar_qi_advance_tail(unit);
dmar_qi_wait_for_seq(unit, &gseq, false);
}
void
dmar_qi_invalidate_ctx_glob_locked(struct dmar_unit *unit)
{
struct dmar_qi_genseq gseq;
DMAR_ASSERT_LOCKED(unit);
dmar_qi_ensure(unit, 2);
dmar_qi_emit(unit, DMAR_IQ_DESCR_CTX_INV | DMAR_IQ_DESCR_CTX_GLOB, 0);
dmar_qi_emit_wait_seq(unit, &gseq, true);
dmar_qi_advance_tail(unit);
dmar_qi_wait_for_seq(unit, &gseq, false);
}
static void
dmar_qi_emit_wait_descr(struct dmar_unit *unit, uint32_t seq, bool intr,
bool memw, bool fence)
{
DMAR_ASSERT_LOCKED(unit);
dmar_qi_emit(unit, DMAR_IQ_DESCR_WAIT_ID |
(intr ? DMAR_IQ_DESCR_WAIT_IF : 0) |
(memw ? DMAR_IQ_DESCR_WAIT_SW : 0) |
(fence ? DMAR_IQ_DESCR_WAIT_FN : 0) |
(memw ? DMAR_IQ_DESCR_WAIT_SD(seq) : 0),
memw ? unit->inv_waitd_seq_hw_phys : 0);
}
void
dmar_qi_invalidate_iotlb_glob_locked(struct dmar_unit *unit)
{
struct dmar_qi_genseq gseq;
DMAR_ASSERT_LOCKED(unit);
dmar_qi_ensure(unit, 2);
dmar_qi_emit(unit, DMAR_IQ_DESCR_IOTLB_INV | DMAR_IQ_DESCR_IOTLB_GLOB |
DMAR_IQ_DESCR_IOTLB_DW | DMAR_IQ_DESCR_IOTLB_DR, 0);
dmar_qi_emit_wait_seq(unit, &gseq);
dmar_qi_advance_tail(unit);
dmar_qi_wait_for_seq(unit, &gseq, false);
}
void
dmar_qi_invalidate_iec(struct dmar_unit *unit, u_int start, u_int cnt)
{
struct dmar_qi_genseq gseq;
u_int c, l;
DMAR_ASSERT_LOCKED(unit);
KASSERT(start < unit->irte_cnt && start < start + cnt &&
start + cnt <= unit->irte_cnt,
("inv iec overflow %d %d %d", unit->irte_cnt, start, cnt));
for (; cnt > 0; cnt -= c, start += c) {
l = ffs(start | cnt) - 1;
c = 1 << l;
dmar_qi_ensure(unit, 1);
dmar_qi_emit(unit, DMAR_IQ_DESCR_IEC_INV |
DMAR_IQ_DESCR_IEC_IDX | DMAR_IQ_DESCR_IEC_IIDX(start) |
DMAR_IQ_DESCR_IEC_IM(l), 0);
}
dmar_qi_ensure(unit, 1);
dmar_qi_emit_wait_seq(unit, &gseq, true);
dmar_qi_advance_tail(unit);
/*
* The caller of the function, in particular,
* dmar_ir_program_irte(), may be called from the context
* where the sleeping is forbidden (in fact, the
* intr_table_lock mutex may be held, locked from
* intr_shuffle_irqs()). Wait for the invalidation completion
* using the busy wait.
*
* The impact on the interrupt input setup code is small, the
* expected overhead is comparable with the chipset register
* read. It is more harmful for the parallel DMA operations,
* since we own the dmar unit lock until whole invalidation
* queue is processed, which includes requests possibly issued
* before our request.
*/
dmar_qi_wait_for_seq(unit, &gseq, true);
}