| Message ID | 1505999838-52530-4-git-send-email-guohanjun@huawei.com |
|---|---|
| State | New |
| Headers | show |
| Series | IORT SMMUv3 MSI support | expand |
On Thu, Sep 21, 2017 at 09:17:17PM +0800, Hanjun Guo wrote: > From: Hanjun Guo <hanjun.guo@linaro.org> > > IORT revision C introduced SMMUv3 MSI support which adding a > device ID mapping index in SMMUv3 sub table, to get the SMMUv3 > device ID mapping for the output ID (dev ID for ITS) and the > link to which ITS. > > So if a platform supports SMMUv3 MSI for control interrupt, > there will be a additional single map entry under SMMU, this > will not introduce any difference for devices just use one > step map to get its output ID and parent (ITS or SMMU), such > as PCI/NC/PMCG ---> ITS or PCI/NC ---> SMMU, but we need to > do the special handling for two steps map case such as > PCI/NC--->SMMUv3--->ITS. > > Take a PCI hostbridge for example, > > |----------------------| > | Root Complex Node | > |----------------------| > | map entry[x] | > |----------------------| > | id value | > | output_reference | > |---|------------------| > | > | |----------------------| > |-->| SMMUv3 | > |----------------------| > | SMMU dev ID | > | mapping index 0 | > |----------------------| > | map entry[0] | > |----------------------| > | id value | > | output_reference-----------> ITS 1 (SMMU MSI domain) > |----------------------| > | map entry[1] | > |----------------------| > | id value | > | output_reference-----------> ITS 2 (PCI MSI domain) > |----------------------| > > When the SMMU dev ID mapping index is 0, there is entry[0] > to map to a ITS, we need to skip that map entry for PCI > or NC (named component), or we may get the wrong ITS parent. We do skip it because it is a single mapping that it is currently not allowed for SMMUv3 components, right ? Ok, we barf with a printk log message if we encounter such mapping but the mapping won't resolve to the SMMUv3 MSI in the current kernel. Anyway, I think patch 3 and patch 4 should be partially squashed, more about that in patch 4 comments. Lorenzo > For now we have two APIs for ID mapping, iort_node_map_id() > and iort_node_map_platform_id(), and iort_node_map_id() is > used for optional two steps mapping, so we just need to > skip the map entry in iort_node_map_id() for non-SMMUv3 > devices. > > Signed-off-by: Hanjun Guo <hanjun.guo@linaro.org> > --- > drivers/acpi/arm64/iort.c | 43 ++++++++++++++++++++++++++++++++++++++++++- > 1 file changed, 42 insertions(+), 1 deletion(-) > > diff --git a/drivers/acpi/arm64/iort.c b/drivers/acpi/arm64/iort.c > index db71d7f..269959e 100644 > --- a/drivers/acpi/arm64/iort.c > +++ b/drivers/acpi/arm64/iort.c > @@ -366,6 +366,34 @@ struct acpi_iort_node *iort_node_get_id(struct acpi_iort_node *node, > return NULL; > } > > +static int iort_get_smmu_v3_id_mapping_index(struct acpi_iort_node *node, > + u32 *index) > +{ > + struct acpi_iort_smmu_v3 *smmu; > + > + /* > + * SMMUv3 dev ID mapping index was introdueced in revision 1 > + * table, not avaible in revision 0 > + */ > + if (node->revision < 1) > + return -EINVAL; > + > + smmu = (struct acpi_iort_smmu_v3 *)node->node_data; > + /* if any of the gsi for control interrupts is not 0, ignore the MSI */ > + if (smmu->event_gsiv || smmu->pri_gsiv || smmu->gerr_gsiv > + || smmu->sync_gsiv) > + return -EINVAL; > + > + if (smmu->id_mapping_index >= node->mapping_count) { > + pr_err(FW_BUG "[node %p type %d] ID mapping index overflows valid mappings\n", > + node, node->type); > + return -EINVAL; > + } > + > + *index = smmu->id_mapping_index; > + return 0; > +} > + > static struct acpi_iort_node *iort_node_map_id(struct acpi_iort_node *node, > u32 id_in, u32 *id_out, > u8 type_mask) > @@ -375,7 +403,9 @@ static struct acpi_iort_node *iort_node_map_id(struct acpi_iort_node *node, > /* Parse the ID mapping tree to find specified node type */ > while (node) { > struct acpi_iort_id_mapping *map; > - int i; > + int i, ret = -EINVAL; > + /* big enough for an invalid id index in practical */ > + u32 index = U32_MAX; > > if (IORT_TYPE_MASK(node->type) & type_mask) { > if (id_out) > @@ -396,8 +426,19 @@ static struct acpi_iort_node *iort_node_map_id(struct acpi_iort_node *node, > goto fail_map; > } > > + /* > + * we need to get SMMUv3 dev ID mapping index and skip its > + * associated ID map for single mapping cases. > + */ > + if (node->type == ACPI_IORT_NODE_SMMU_V3) > + ret = iort_get_smmu_v3_id_mapping_index(node, &index); > + > /* Do the ID translation */ > for (i = 0; i < node->mapping_count; i++, map++) { > + /* if it's a SMMUv3 device id mapping index, skip it */ > + if (!ret && i == index) > + continue; > + > if (!iort_id_map(map, node->type, id, &id)) > break; > } > -- > 1.7.12.4 > > -- > To unsubscribe from this list: send the line "unsubscribe linux-acpi" in > the body of a message to majordomo@vger.kernel.org > More majordomo info at http://vger.kernel.org/majordomo-info.html -- To unsubscribe from this list: send the line "unsubscribe linux-acpi" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
On 21/09/17 14:17, Hanjun Guo wrote: > From: Hanjun Guo <hanjun.guo@linaro.org> > > IORT revision C introduced SMMUv3 MSI support which adding a > device ID mapping index in SMMUv3 sub table, to get the SMMUv3 > device ID mapping for the output ID (dev ID for ITS) and the > link to which ITS. > > So if a platform supports SMMUv3 MSI for control interrupt, > there will be a additional single map entry under SMMU, this > will not introduce any difference for devices just use one > step map to get its output ID and parent (ITS or SMMU), such > as PCI/NC/PMCG ---> ITS or PCI/NC ---> SMMU, but we need to > do the special handling for two steps map case such as > PCI/NC--->SMMUv3--->ITS. > > Take a PCI hostbridge for example, > > |----------------------| > | Root Complex Node | > |----------------------| > | map entry[x] | > |----------------------| > | id value | > | output_reference | > |---|------------------| > | > | |----------------------| > |-->| SMMUv3 | > |----------------------| > | SMMU dev ID | > | mapping index 0 | > |----------------------| > | map entry[0] | > |----------------------| > | id value | > | output_reference-----------> ITS 1 (SMMU MSI domain) > |----------------------| > | map entry[1] | > |----------------------| > | id value | > | output_reference-----------> ITS 2 (PCI MSI domain) > |----------------------| > > When the SMMU dev ID mapping index is 0, there is entry[0] > to map to a ITS, we need to skip that map entry for PCI > or NC (named component), or we may get the wrong ITS parent. > > For now we have two APIs for ID mapping, iort_node_map_id() > and iort_node_map_platform_id(), and iort_node_map_id() is > used for optional two steps mapping, so we just need to > skip the map entry in iort_node_map_id() for non-SMMUv3 > devices. > > Signed-off-by: Hanjun Guo <hanjun.guo@linaro.org> > --- > drivers/acpi/arm64/iort.c | 43 ++++++++++++++++++++++++++++++++++++++++++- > 1 file changed, 42 insertions(+), 1 deletion(-) > > diff --git a/drivers/acpi/arm64/iort.c b/drivers/acpi/arm64/iort.c > index db71d7f..269959e 100644 > --- a/drivers/acpi/arm64/iort.c > +++ b/drivers/acpi/arm64/iort.c > @@ -366,6 +366,34 @@ struct acpi_iort_node *iort_node_get_id(struct acpi_iort_node *node, > return NULL; > } > > +static int iort_get_smmu_v3_id_mapping_index(struct acpi_iort_node *node, > + u32 *index) > +{ > + struct acpi_iort_smmu_v3 *smmu; > + > + /* > + * SMMUv3 dev ID mapping index was introdueced in revision 1 > + * table, not avaible in revision 0 > + */ > + if (node->revision < 1) > + return -EINVAL; > + > + smmu = (struct acpi_iort_smmu_v3 *)node->node_data; > + /* if any of the gsi for control interrupts is not 0, ignore the MSI */ > + if (smmu->event_gsiv || smmu->pri_gsiv || smmu->gerr_gsiv > + || smmu->sync_gsiv) To repeat my previous comments, the ID mapping index is only ignored if *all* interrupts are GSIV-based. It would be quite reasonable for gerr to be wired while everything else uses MSIs, especially since that's effectively the only reliable way to get MSI aborts reported if things are completely hosed. > + return -EINVAL; > + > + if (smmu->id_mapping_index >= node->mapping_count) { > + pr_err(FW_BUG "[node %p type %d] ID mapping index overflows valid mappings\n", > + node, node->type); > + return -EINVAL; > + } > + > + *index = smmu->id_mapping_index; > + return 0; > +} > + > static struct acpi_iort_node *iort_node_map_id(struct acpi_iort_node *node, > u32 id_in, u32 *id_out, > u8 type_mask) > @@ -375,7 +403,9 @@ static struct acpi_iort_node *iort_node_map_id(struct acpi_iort_node *node, > /* Parse the ID mapping tree to find specified node type */ > while (node) { > struct acpi_iort_id_mapping *map; > - int i; > + int i, ret = -EINVAL; > + /* big enough for an invalid id index in practical */ > + u32 index = U32_MAX; And again, more of a style nit, but indices must be sufficiently small to fit into the positive half of an int, so *_id_mapping_index() really doesn't need this complication of passing a separate output argument. Robin. > if (IORT_TYPE_MASK(node->type) & type_mask) { > if (id_out) > @@ -396,8 +426,19 @@ static struct acpi_iort_node *iort_node_map_id(struct acpi_iort_node *node, > goto fail_map; > } > > + /* > + * we need to get SMMUv3 dev ID mapping index and skip its > + * associated ID map for single mapping cases. > + */ > + if (node->type == ACPI_IORT_NODE_SMMU_V3) > + ret = iort_get_smmu_v3_id_mapping_index(node, &index); > + > /* Do the ID translation */ > for (i = 0; i < node->mapping_count; i++, map++) { > + /* if it's a SMMUv3 device id mapping index, skip it */ > + if (!ret && i == index) > + continue; > + > if (!iort_id_map(map, node->type, id, &id)) > break; > } > -- To unsubscribe from this list: send the line "unsubscribe linux-acpi" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
Hi Robin, Sorry for the late reply, in Linaro Connect now.. On 09/22/2017 09:21 PM, Robin Murphy wrote: > On 21/09/17 14:17, Hanjun Guo wrote: >> From: Hanjun Guo <hanjun.guo@linaro.org> >> >> IORT revision C introduced SMMUv3 MSI support which adding a >> device ID mapping index in SMMUv3 sub table, to get the SMMUv3 >> device ID mapping for the output ID (dev ID for ITS) and the >> link to which ITS. >> >> So if a platform supports SMMUv3 MSI for control interrupt, >> there will be a additional single map entry under SMMU, this >> will not introduce any difference for devices just use one >> step map to get its output ID and parent (ITS or SMMU), such >> as PCI/NC/PMCG ---> ITS or PCI/NC ---> SMMU, but we need to >> do the special handling for two steps map case such as >> PCI/NC--->SMMUv3--->ITS. >> >> Take a PCI hostbridge for example, >> >> |----------------------| >> | Root Complex Node | >> |----------------------| >> | map entry[x] | >> |----------------------| >> | id value | >> | output_reference | >> |---|------------------| >> | >> | |----------------------| >> |-->| SMMUv3 | >> |----------------------| >> | SMMU dev ID | >> | mapping index 0 | >> |----------------------| >> | map entry[0] | >> |----------------------| >> | id value | >> | output_reference-----------> ITS 1 (SMMU MSI domain) >> |----------------------| >> | map entry[1] | >> |----------------------| >> | id value | >> | output_reference-----------> ITS 2 (PCI MSI domain) >> |----------------------| >> >> When the SMMU dev ID mapping index is 0, there is entry[0] >> to map to a ITS, we need to skip that map entry for PCI >> or NC (named component), or we may get the wrong ITS parent. >> >> For now we have two APIs for ID mapping, iort_node_map_id() >> and iort_node_map_platform_id(), and iort_node_map_id() is >> used for optional two steps mapping, so we just need to >> skip the map entry in iort_node_map_id() for non-SMMUv3 >> devices. >> >> Signed-off-by: Hanjun Guo <hanjun.guo@linaro.org> >> --- >> drivers/acpi/arm64/iort.c | 43 ++++++++++++++++++++++++++++++++++++++++++- >> 1 file changed, 42 insertions(+), 1 deletion(-) >> >> diff --git a/drivers/acpi/arm64/iort.c b/drivers/acpi/arm64/iort.c >> index db71d7f..269959e 100644 >> --- a/drivers/acpi/arm64/iort.c >> +++ b/drivers/acpi/arm64/iort.c >> @@ -366,6 +366,34 @@ struct acpi_iort_node *iort_node_get_id(struct acpi_iort_node *node, >> return NULL; >> } >> >> +static int iort_get_smmu_v3_id_mapping_index(struct acpi_iort_node *node, >> + u32 *index) >> +{ >> + struct acpi_iort_smmu_v3 *smmu; >> + >> + /* >> + * SMMUv3 dev ID mapping index was introdueced in revision 1 >> + * table, not avaible in revision 0 >> + */ >> + if (node->revision < 1) >> + return -EINVAL; >> + >> + smmu = (struct acpi_iort_smmu_v3 *)node->node_data; >> + /* if any of the gsi for control interrupts is not 0, ignore the MSI */ >> + if (smmu->event_gsiv || smmu->pri_gsiv || smmu->gerr_gsiv >> + || smmu->sync_gsiv) > > To repeat my previous comments, the ID mapping index is only ignored if > *all* interrupts are GSIV-based. > > It would be quite reasonable for gerr to be wired while everything else > uses MSIs, especially since that's effectively the only reliable way to > get MSI aborts reported if things are completely hosed. OK, I missed this point, so I will modify the code as below if (smmu->event_gsiv && smmu->pri_gsiv && smmu->gerr_gsiv && smmu->sync_gsiv) return -EINVAL; Am I doing this right with above code? > >> + return -EINVAL; >> + >> + if (smmu->id_mapping_index >= node->mapping_count) { >> + pr_err(FW_BUG "[node %p type %d] ID mapping index overflows valid mappings\n", >> + node, node->type); >> + return -EINVAL; >> + } >> + >> + *index = smmu->id_mapping_index; >> + return 0; >> +} >> + >> static struct acpi_iort_node *iort_node_map_id(struct acpi_iort_node *node, >> u32 id_in, u32 *id_out, >> u8 type_mask) >> @@ -375,7 +403,9 @@ static struct acpi_iort_node *iort_node_map_id(struct acpi_iort_node *node, >> /* Parse the ID mapping tree to find specified node type */ >> while (node) { >> struct acpi_iort_id_mapping *map; >> - int i; >> + int i, ret = -EINVAL; >> + /* big enough for an invalid id index in practical */ >> + u32 index = U32_MAX; > > And again, more of a style nit, but indices must be sufficiently small > to fit into the positive half of an int, so *_id_mapping_index() really > doesn't need this complication of passing a separate output argument. Sorry, I reused the code which Lorenzo sent me, will update the code in next version. Thanks Hanjun -- To unsubscribe from this list: send the line "unsubscribe linux-acpi" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
On 09/22/2017 08:53 PM, Lorenzo Pieralisi wrote: > On Thu, Sep 21, 2017 at 09:17:17PM +0800, Hanjun Guo wrote: >> From: Hanjun Guo <hanjun.guo@linaro.org> >> >> IORT revision C introduced SMMUv3 MSI support which adding a >> device ID mapping index in SMMUv3 sub table, to get the SMMUv3 >> device ID mapping for the output ID (dev ID for ITS) and the >> link to which ITS. >> >> So if a platform supports SMMUv3 MSI for control interrupt, >> there will be a additional single map entry under SMMU, this >> will not introduce any difference for devices just use one >> step map to get its output ID and parent (ITS or SMMU), such >> as PCI/NC/PMCG ---> ITS or PCI/NC ---> SMMU, but we need to >> do the special handling for two steps map case such as >> PCI/NC--->SMMUv3--->ITS. >> >> Take a PCI hostbridge for example, >> >> |----------------------| >> | Root Complex Node | >> |----------------------| >> | map entry[x] | >> |----------------------| >> | id value | >> | output_reference | >> |---|------------------| >> | >> | |----------------------| >> |-->| SMMUv3 | >> |----------------------| >> | SMMU dev ID | >> | mapping index 0 | >> |----------------------| >> | map entry[0] | >> |----------------------| >> | id value | >> | output_reference-----------> ITS 1 (SMMU MSI domain) >> |----------------------| >> | map entry[1] | >> |----------------------| >> | id value | >> | output_reference-----------> ITS 2 (PCI MSI domain) >> |----------------------| >> >> When the SMMU dev ID mapping index is 0, there is entry[0] >> to map to a ITS, we need to skip that map entry for PCI >> or NC (named component), or we may get the wrong ITS parent. > > We do skip it because it is a single mapping that it is currently > not allowed for SMMUv3 components, right ? > > Ok, we barf with a printk log message if we encounter such mapping > but the mapping won't resolve to the SMMUv3 MSI in the current > kernel. This is not clear in the spec, maybe we also need to update the IORT spec about it. Thanks Hanjun -- To unsubscribe from this list: send the line "unsubscribe linux-acpi" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
diff --git a/drivers/acpi/arm64/iort.c b/drivers/acpi/arm64/iort.c index db71d7f..269959e 100644 --- a/drivers/acpi/arm64/iort.c +++ b/drivers/acpi/arm64/iort.c @@ -366,6 +366,34 @@ struct acpi_iort_node *iort_node_get_id(struct acpi_iort_node *node, return NULL; } +static int iort_get_smmu_v3_id_mapping_index(struct acpi_iort_node *node, + u32 *index) +{ + struct acpi_iort_smmu_v3 *smmu; + + /* + * SMMUv3 dev ID mapping index was introdueced in revision 1 + * table, not avaible in revision 0 + */ + if (node->revision < 1) + return -EINVAL; + + smmu = (struct acpi_iort_smmu_v3 *)node->node_data; + /* if any of the gsi for control interrupts is not 0, ignore the MSI */ + if (smmu->event_gsiv || smmu->pri_gsiv || smmu->gerr_gsiv + || smmu->sync_gsiv) + return -EINVAL; + + if (smmu->id_mapping_index >= node->mapping_count) { + pr_err(FW_BUG "[node %p type %d] ID mapping index overflows valid mappings\n", + node, node->type); + return -EINVAL; + } + + *index = smmu->id_mapping_index; + return 0; +} + static struct acpi_iort_node *iort_node_map_id(struct acpi_iort_node *node, u32 id_in, u32 *id_out, u8 type_mask) @@ -375,7 +403,9 @@ static struct acpi_iort_node *iort_node_map_id(struct acpi_iort_node *node, /* Parse the ID mapping tree to find specified node type */ while (node) { struct acpi_iort_id_mapping *map; - int i; + int i, ret = -EINVAL; + /* big enough for an invalid id index in practical */ + u32 index = U32_MAX; if (IORT_TYPE_MASK(node->type) & type_mask) { if (id_out) @@ -396,8 +426,19 @@ static struct acpi_iort_node *iort_node_map_id(struct acpi_iort_node *node, goto fail_map; } + /* + * we need to get SMMUv3 dev ID mapping index and skip its + * associated ID map for single mapping cases. + */ + if (node->type == ACPI_IORT_NODE_SMMU_V3) + ret = iort_get_smmu_v3_id_mapping_index(node, &index); + /* Do the ID translation */ for (i = 0; i < node->mapping_count; i++, map++) { + /* if it's a SMMUv3 device id mapping index, skip it */ + if (!ret && i == index) + continue; + if (!iort_id_map(map, node->type, id, &id)) break; }