Hi Eric,
On Tue, Apr 02, 2024 at 07:15:20PM +0200, Eric Auger wrote:
> Hi Mostafa,
>
> On 3/25/24 11:13, Mostafa Saleh wrote:
> > TLBs for nesting will be extended to be combined, a new index is added
> > "stage", with 2 valid values:
> > - SMMU_STAGE_1: Meaning this translates VA to PADDR, this entry can
> > be cached from fully nested configuration or from stage-1 only.
> > We don't support separate cached entries (VA to IPA)
> >
> > - SMMU_STAGE_2: Meaning this translates IPA to PADDR, cached from
> > stage-2 only configuration.
> >
> > For TLB invalidation:
> > - by VA: Invalidate TLBs tagged with SMMU_STAGE_1
> > - by IPA: Invalidate TLBs tagged with SMMU_STAGE_2
> > - All: Will invalidate both, this is communicated to the TLB as
> > SMMU_NESTED which is (SMMU_STAGE_1 | SMMU_STAGE_2) which uses
> > it as a mask.
>
> I don't really get why you need this extra stage field in the key. Why
> aren't the asid and vmid tags enough?
>
Looking again, I think we can do it with ASID and VMID only, but that
requires some rework in the invalidation path.
With nested SMMUs, we can cache entries from:
- Stage-1 (or nested): Tagged with VMID and ASID
- Stage-2: Tagged with VMID only (ASID = -1)
That should be enough for caching/lookup, but for invalidation, we
should be able to invalidate IPAs which are cached from stage-2.
At the moment, we represent ASIDs with < 0 as a wildcard for
invalidation or stage-2 and they were mutually exclusive.
An example is:
- CMD_TLBI_NH_VAA: Invalidate stage-1 for a VMID, all ASIDs (we use ASID = -1)
- CMD_TLBI_NH_VA: Invalidate stage-1 for a VMID, an ASID ( > 0)
- CMD_TLBI_S2_IPA: Invalidate stage-2 for a VMID (we use ASID = -1)
We need to distinguish between case 1) and 3) otherwise we over invalidate.
Similarly, CMD_TLBI_NH_ALL(invalidate all stage-1 by VMID) and
CMD_TLBI_S12_VMALL(invalidate both stages by VMID).
I guess we can add variants of these functions that operate on ASIDs
(>= 0) or (< 0) which is basically stage-1 or stage-2.
Another case I can think of which is not implemented in QEMU is
global entries, where we would like to look up entries for all ASIDs
(-1), but that’s not a problem for now.
I don’t have a strong opinion, I can try to do it this way.
Thanks,
Mostafa
> Eric
> >
> > This briefly described in the user manual (ARM IHI 0070 F.b) in
> > "16.2.1 Caching combined structures".
> >
> > Signed-off-by: Mostafa Saleh <smost...@google.com>
> > ---
> > hw/arm/smmu-common.c | 27 +++++++++++++++++----------
> > hw/arm/smmu-internal.h | 2 ++
> > hw/arm/smmuv3.c | 5 +++--
> > hw/arm/trace-events | 3 ++-
> > include/hw/arm/smmu-common.h | 8 ++++++--
> > 5 files changed, 30 insertions(+), 15 deletions(-)
> >
> > diff --git a/hw/arm/smmu-common.c b/hw/arm/smmu-common.c
> > index 20630eb670..677dcf9a13 100644
> > --- a/hw/arm/smmu-common.c
> > +++ b/hw/arm/smmu-common.c
> > @@ -38,7 +38,7 @@ static guint smmu_iotlb_key_hash(gconstpointer v)
> >
> > /* Jenkins hash */
> > a = b = c = JHASH_INITVAL + sizeof(*key);
> > - a += key->asid + key->vmid + key->level + key->tg;
> > + a += key->asid + key->vmid + key->level + key->tg + key->stage;
> > b += extract64(key->iova, 0, 32);
> > c += extract64(key->iova, 32, 32);
> >
> > @@ -54,14 +54,14 @@ static gboolean smmu_iotlb_key_equal(gconstpointer v1,
> > gconstpointer v2)
> >
> > return (k1->asid == k2->asid) && (k1->iova == k2->iova) &&
> > (k1->level == k2->level) && (k1->tg == k2->tg) &&
> > - (k1->vmid == k2->vmid);
> > + (k1->vmid == k2->vmid) && (k1->stage == k2->stage);
> > }
> >
> > SMMUIOTLBKey smmu_get_iotlb_key(uint16_t asid, uint16_t vmid, uint64_t
> > iova,
> > - uint8_t tg, uint8_t level)
> > + uint8_t tg, uint8_t level, SMMUStage stage)
> > {
> > SMMUIOTLBKey key = {.asid = asid, .vmid = vmid, .iova = iova,
> > - .tg = tg, .level = level};
> > + .tg = tg, .level = level, .stage = stage};
> >
> > return key;
> > }
> > @@ -81,7 +81,8 @@ SMMUTLBEntry *smmu_iotlb_lookup(SMMUState *bs,
> > SMMUTransCfg *cfg,
> > SMMUIOTLBKey key;
> >
> > key = smmu_get_iotlb_key(cfg->asid, cfg->s2cfg.vmid,
> > - iova & ~mask, tg, level);
> > + iova & ~mask, tg, level,
> > + SMMU_STAGE_TO_TLB_TAG(cfg->stage));
> > entry = g_hash_table_lookup(bs->iotlb, &key);
> > if (entry) {
> > break;
> > @@ -109,15 +110,16 @@ void smmu_iotlb_insert(SMMUState *bs, SMMUTransCfg
> > *cfg, SMMUTLBEntry *new)
> > {
> > SMMUIOTLBKey *key = g_new0(SMMUIOTLBKey, 1);
> > uint8_t tg = (new->granule - 10) / 2;
> > + SMMUStage stage_tag = SMMU_STAGE_TO_TLB_TAG(cfg->stage);
> >
> > if (g_hash_table_size(bs->iotlb) >= SMMU_IOTLB_MAX_SIZE) {
> > smmu_iotlb_inv_all(bs);
> > }
> >
> > *key = smmu_get_iotlb_key(cfg->asid, cfg->s2cfg.vmid, new->entry.iova,
> > - tg, new->level);
> > + tg, new->level, stage_tag);
> > trace_smmu_iotlb_insert(cfg->asid, cfg->s2cfg.vmid, new->entry.iova,
> > - tg, new->level);
> > + tg, new->level, stage_tag);
> > g_hash_table_insert(bs->iotlb, key, new);
> > }
> >
> > @@ -159,18 +161,22 @@ static gboolean
> > smmu_hash_remove_by_asid_vmid_iova(gpointer key, gpointer value,
> > if (info->vmid >= 0 && info->vmid != SMMU_IOTLB_VMID(iotlb_key)) {
> > return false;
> > }
> > + if (!(info->stage & SMMU_IOTLB_STAGE(iotlb_key))) {
> > + return false;
> > + }
> > return ((info->iova & ~entry->addr_mask) == entry->iova) ||
> > ((entry->iova & ~info->mask) == info->iova);
> > }
> >
> > void smmu_iotlb_inv_iova(SMMUState *s, int asid, int vmid, dma_addr_t iova,
> > - uint8_t tg, uint64_t num_pages, uint8_t ttl)
> > + uint8_t tg, uint64_t num_pages, uint8_t ttl,
> > + SMMUStage stage)
> > {
> > /* if tg is not set we use 4KB range invalidation */
> > uint8_t granule = tg ? tg * 2 + 10 : 12;
> >
> > if (ttl && (num_pages == 1) && (asid >= 0)) {
> > - SMMUIOTLBKey key = smmu_get_iotlb_key(asid, vmid, iova, tg, ttl);
> > + SMMUIOTLBKey key = smmu_get_iotlb_key(asid, vmid, iova, tg, ttl,
> > stage);
> >
> > if (g_hash_table_remove(s->iotlb, &key)) {
> > return;
> > @@ -184,6 +190,7 @@ void smmu_iotlb_inv_iova(SMMUState *s, int asid, int
> > vmid, dma_addr_t iova,
> > SMMUIOTLBPageInvInfo info = {
> > .asid = asid, .iova = iova,
> > .vmid = vmid,
> > + .stage = stage,
> > .mask = (num_pages * 1 << granule) - 1};
> >
> > g_hash_table_foreach_remove(s->iotlb,
> > @@ -597,7 +604,7 @@ SMMUTLBEntry *smmu_translate(SMMUState *bs,
> > SMMUTransCfg *cfg, dma_addr_t addr,
> > if (cached_entry) {
> > if ((flag & IOMMU_WO) && !(cached_entry->entry.perm & IOMMU_WO)) {
> > info->type = SMMU_PTW_ERR_PERMISSION;
> > - info->stage = cfg->stage;
> > + info->stage = SMMU_STAGE_TO_TLB_TAG(cfg->stage);
> > return NULL;
> > }
> > return cached_entry;
> > diff --git a/hw/arm/smmu-internal.h b/hw/arm/smmu-internal.h
> > index 843bebb185..6caa0ddf21 100644
> > --- a/hw/arm/smmu-internal.h
> > +++ b/hw/arm/smmu-internal.h
> > @@ -133,12 +133,14 @@ static inline int pgd_concat_idx(int start_level, int
> > granule_sz,
> >
> > #define SMMU_IOTLB_ASID(key) ((key).asid)
> > #define SMMU_IOTLB_VMID(key) ((key).vmid)
> > +#define SMMU_IOTLB_STAGE(key) ((key).stage)
> >
> > typedef struct SMMUIOTLBPageInvInfo {
> > int asid;
> > int vmid;
> > uint64_t iova;
> > uint64_t mask;
> > + SMMUStage stage;
> > } SMMUIOTLBPageInvInfo;
> >
> > typedef struct SMMUSIDRange {
> > diff --git a/hw/arm/smmuv3.c b/hw/arm/smmuv3.c
> > index f081ff0cc4..b27bf297e1 100644
> > --- a/hw/arm/smmuv3.c
> > +++ b/hw/arm/smmuv3.c
> > @@ -1087,7 +1087,7 @@ static void smmuv3_range_inval(SMMUState *s, Cmd *cmd)
> > if (!tg) {
> > trace_smmuv3_range_inval(vmid, asid, addr, tg, 1, ttl, leaf);
> > smmuv3_inv_notifiers_iova(s, asid, vmid, addr, tg, 1);
> > - smmu_iotlb_inv_iova(s, asid, vmid, addr, tg, 1, ttl);
> > + smmu_iotlb_inv_iova(s, asid, vmid, addr, tg, 1, ttl, SMMU_NESTED);
> > return;
> > }
> >
> > @@ -1105,7 +1105,8 @@ static void smmuv3_range_inval(SMMUState *s, Cmd *cmd)
> > num_pages = (mask + 1) >> granule;
> > trace_smmuv3_range_inval(vmid, asid, addr, tg, num_pages, ttl,
> > leaf);
> > smmuv3_inv_notifiers_iova(s, asid, vmid, addr, tg, num_pages);
> > - smmu_iotlb_inv_iova(s, asid, vmid, addr, tg, num_pages, ttl);
> > + smmu_iotlb_inv_iova(s, asid, vmid, addr, tg,
> > + num_pages, ttl, SMMU_NESTED);
> > addr += mask + 1;
> > }
> > }
> > diff --git a/hw/arm/trace-events b/hw/arm/trace-events
> > index cc12924a84..3000c3bf14 100644
> > --- a/hw/arm/trace-events
> > +++ b/hw/arm/trace-events
> > @@ -14,10 +14,11 @@ smmu_iotlb_inv_all(void) "IOTLB invalidate all"
> > smmu_iotlb_inv_asid(uint16_t asid) "IOTLB invalidate asid=%d"
> > smmu_iotlb_inv_vmid(uint16_t vmid) "IOTLB invalidate vmid=%d"
> > smmu_iotlb_inv_iova(uint16_t asid, uint64_t addr) "IOTLB invalidate
> > asid=%d addr=0x%"PRIx64
> > +smmu_iotlb_inv_stage(int stage) "Stage invalidate stage=%d"
> > smmu_inv_notifiers_mr(const char *name) "iommu mr=%s"
> > smmu_iotlb_lookup_hit(uint16_t asid, uint16_t vmid, uint64_t addr,
> > uint32_t hit, uint32_t miss, uint32_t p) "IOTLB cache HIT asid=%d vmid=%d
> > addr=0x%"PRIx64" hit=%d miss=%d hit rate=%d"
> > smmu_iotlb_lookup_miss(uint16_t asid, uint16_t vmid, uint64_t addr,
> > uint32_t hit, uint32_t miss, uint32_t p) "IOTLB cache MISS asid=%d vmid=%d
> > addr=0x%"PRIx64" hit=%d miss=%d hit rate=%d"
> > -smmu_iotlb_insert(uint16_t asid, uint16_t vmid, uint64_t addr, uint8_t tg,
> > uint8_t level) "IOTLB ++ asid=%d vmid=%d addr=0x%"PRIx64" tg=%d level=%d"
> > +smmu_iotlb_insert(uint16_t asid, uint16_t vmid, uint64_t addr, uint8_t tg,
> > uint8_t level, int stage) "IOTLB ++ asid=%d vmid=%d addr=0x%"PRIx64" tg=%d
> > level=%d stage=%d"
> >
> > # smmuv3.c
> > smmuv3_read_mmio(uint64_t addr, uint64_t val, unsigned size, uint32_t r)
> > "addr: 0x%"PRIx64" val:0x%"PRIx64" size: 0x%x(%d)"
> > diff --git a/include/hw/arm/smmu-common.h b/include/hw/arm/smmu-common.h
> > index 876e78975c..695d6d10ad 100644
> > --- a/include/hw/arm/smmu-common.h
> > +++ b/include/hw/arm/smmu-common.h
> > @@ -37,6 +37,8 @@
> > #define VMSA_IDXMSK(isz, strd, lvl) ((1ULL << \
> > VMSA_BIT_LVL(isz, strd, lvl))
> > - 1)
> >
> > +#define SMMU_STAGE_TO_TLB_TAG(stage) (((stage) == SMMU_NESTED) ? \
> > + SMMU_STAGE_1 : (stage))
> > /*
> > * Page table walk error types
> > */
> > @@ -136,6 +138,7 @@ typedef struct SMMUIOTLBKey {
> > uint16_t vmid;
> > uint8_t tg;
> > uint8_t level;
> > + SMMUStage stage;
> > } SMMUIOTLBKey;
> >
> > struct SMMUState {
> > @@ -203,12 +206,13 @@ SMMUTLBEntry *smmu_iotlb_lookup(SMMUState *bs,
> > SMMUTransCfg *cfg,
> > SMMUTransTableInfo *tt, hwaddr iova);
> > void smmu_iotlb_insert(SMMUState *bs, SMMUTransCfg *cfg, SMMUTLBEntry
> > *entry);
> > SMMUIOTLBKey smmu_get_iotlb_key(uint16_t asid, uint16_t vmid, uint64_t
> > iova,
> > - uint8_t tg, uint8_t level);
> > + uint8_t tg, uint8_t level, SMMUStage
> > stage);
> > void smmu_iotlb_inv_all(SMMUState *s);
> > void smmu_iotlb_inv_asid(SMMUState *s, uint16_t asid);
> > void smmu_iotlb_inv_vmid(SMMUState *s, uint16_t vmid);
> > void smmu_iotlb_inv_iova(SMMUState *s, int asid, int vmid, dma_addr_t iova,
> > - uint8_t tg, uint64_t num_pages, uint8_t ttl);
> > + uint8_t tg, uint64_t num_pages, uint8_t ttl,
> > + SMMUStage stage);
> >
> > /* Unmap the range of all the notifiers registered to any IOMMU mr */
> > void smmu_inv_notifiers_all(SMMUState *s);
>
