hw/cxl: Passthrough HDM decoder emulation
cover hw/cxl: Passthrough HDM decoder emulation
Until now, testing using CXL has relied up always using two root ports below a host bridge, to work around a current assumption in the Linux kernel support that, in the single root port case, the implementation will use the allowed passthrough decoder implementation choice. If that choice is made all accesses are routed from the host bridge to the single root port that is present. Effectively we have a pass through decoder (it is called that in the kernel driver).
This patch series implements that functionality and makes it the default See patch 2 for a discussion of why I think we can make this change without backwards compatibility issues (basically if it didn't work before who are we breaking by making it work?)
Whilst this limitation has been known since the initial QEMU patch postings / kernel CXL region support, Fan Ni Ran into it recently reminding me that we should solve it.
https://lore.kernel.org/linux-cxl/20230113171044.GA24788@bgt-140510-bm03/
Tree with a large set of patches before this at:
https://gitlab.com/jic23/qemu/-/tree/cxl-2023-01-20
I've done some basic testing, though I did hit what appears to be a kernel race on region bring up of existing region / namespace in a 1HB 2RP 2EP test case. That is proving hard to replicate consistently but doesn't seem to have anything to do with the emulation other than perhaps we are opening up a race by responding slowly to something.
Jonathan Cameron (2):
hw/pci: Add pcie_count_ds_port() and pcie_find_port_first() helpers hw/pxb-cxl: Support passthrough HDM Decoders unless overridden
hw/cxl/cxl-host.c | 31 +++++++++++++-------- hw/pci-bridge/pci_expander_bridge.c | 43 +++++++++++++++++++++++++---- hw/pci/pcie_port.c | 38 +++++++++++++++++++++++++ include/hw/cxl/cxl.h | 1 + include/hw/cxl/cxl_component.h | 1 + include/hw/pci/pci_bridge.h | 1 + include/hw/pci/pcie_port.h | 2 ++ 7 files changed, 100 insertions(+), 17 deletions(-)
Signature · 1 line
2.37.2
RFC PATCH 1/2 hw/pci: Add pcie_count_ds_port() and pcie_find_port_first() helpers
These two helpers enable host bridges to operate differently depending on the number of downstream ports, in particular if there is only a single port.
Useful for CXL where HDM address decoders are allowed to be implicit in the host bridge if there is only a single root port.
--- hw/pci/pcie_port.c | 38 ++++++++++++++++++++++++++++++++++++++ include/hw/pci/pcie_port.h | 2 ++ 2 files changed, 40 insertions(+)
hw/pci/pcie_port.c +38 -0
@@ -161,6 +161,44 @@ PCIDevice *pcie_find_port_by_pn(PCIBus *bus, uint8_t pn)return NULL ; } +/* Find first port in devfn number */ +PCIDevice *pcie_find_port_first (PCIBus *bus) +{ +int devfn; + +for (devfn =0 ; devfn <ARRAY_SIZE (bus->devices); devfn++) { +PCIDevice *d = bus->devices[devfn]; + +if (!d || !pci_is_express (d) || !d->exp.exp_cap) { +continue ; + } + +if (object_dynamic_cast (OBJECT (d),TYPE_PCIE_PORT )) { +return d; + } + } + +return NULL ; +} + +int pcie_count_ds_ports (PCIBus *bus) +{ +int dsp_count =0 ; +int devfn; + +for (devfn =0 ; devfn <ARRAY_SIZE (bus->devices); devfn++) { +PCIDevice *d = bus->devices[devfn]; + +if (!d || !pci_is_express (d) || !d->exp.exp_cap) { +continue ; + } +if (object_dynamic_cast (OBJECT (d),TYPE_PCIE_PORT )) { + dsp_count++; + } + } +return dsp_count; +} +static const TypeInfo pcie_port_type_info = { .name =TYPE_PCIE_PORT , .parent =TYPE_PCI_BRIDGE ,
include/hw/pci/pcie_port.h +2 -0
@@ -41,6 +41,8 @@ struct PCIEPort {void pcie_port_init_reg (PCIDevice *d);PCIDevice *pcie_find_port_by_pn (PCIBus *bus,uint8_t pn); +PCIDevice *pcie_find_port_first (PCIBus *bus); +int pcie_count_ds_ports (PCIBus *bus);#define TYPE_PCIE_SLOT "pcie-slot"OBJECT_DECLARE_SIMPLE_TYPE (PCIESlot,PCIE_SLOT )
Signature · 1 line
2.37.2
RFC PATCH 2/2 hw/pxb-cxl: Support passthrough HDM Decoders unless overridden
In reply to Jonathan CameronThe CXL r3.0 specification allows for there to be no HDM decoders on CXL Host Bridges if they have only a single root port. Instead, all accesses directed to the host bridge (as specified in CXL Fixed Memory Windows) are assumed to be routed to the single root port.
Linux currently assumes this implementation choice. So to simplify testing, make QEMU emulation also default to no HDM decoders under these particular circumstances, but provide a hdm_for_passthrough boolean option to have HDM decoders as previously.
Technically this is breaking backwards compatibility, but given the only known software stack used with the QEMU emulation is the Linux kernel and this configuration did not work before this change, there are unlikely to be any complaints that it now works. The option is retained to allow testing of software that does allow for these HDM decoders to exist, once someone writes it.
--- hw/cxl/cxl-host.c | 31 +++++++++++++-------- hw/pci-bridge/pci_expander_bridge.c | 43 +++++++++++++++++++++++++---- include/hw/cxl/cxl.h | 1 + include/hw/cxl/cxl_component.h | 1 + include/hw/pci/pci_bridge.h | 1 + 5 files changed, 60 insertions(+), 17 deletions(-)
hw/cxl/cxl-host.c +19 -12
@@ -146,21 +146,28 @@ static PCIDevice *cxl_cfmws_find_device(CXLFixedWindow *fw, hwaddr addr)return NULL ; } - hb_cstate =cxl_get_hb_cstate (hb); -if (!hb_cstate) { -return NULL ; - } +if (cxl_get_hb_passthrough (hb)) { + rp =pcie_find_port_first (hb->bus); +if (!rp) { +return NULL ; + } + }else { + hb_cstate =cxl_get_hb_cstate (hb); +if (!hb_cstate) { +return NULL ; + } - cache_mem = hb_cstate->crb.cache_mem_registers; + cache_mem = hb_cstate->crb.cache_mem_registers; - target_found =cxl_hdm_find_target (cache_mem, addr, &target); -if (!target_found) { -return NULL ; - } + target_found =cxl_hdm_find_target (cache_mem, addr, &target); +if (!target_found) { +return NULL ; + } - rp =pcie_find_port_by_pn (hb->bus, target); -if (!rp) { -return NULL ; + rp =pcie_find_port_by_pn (hb->bus, target); +if (!rp) { +return NULL ; + } } d =pci_bridge_get_sec_bus (PCI_BRIDGE (rp))->devices[0 ];
hw/pci-bridge/pci_expander_bridge.c +38 -5
86 lines not on this page · raw message
include/hw/cxl/cxl.h +1 -0
@@ -50,6 +50,7 @@ struct CXLHost {PCIHostState parent_obj;CXLComponentState cxl_cstate; +bool passthrough; };#define TYPE_PXB_CXL_HOST "pxb-cxl-host"
include/hw/cxl/cxl_component.h +1 -0
@@ -251,6 +251,7 @@ static inline hwaddr cxl_decode_ig(int ig) }CXLComponentState *cxl_get_hb_cstate (PCIHostState *hb); +bool cxl_get_hb_passthrough (PCIHostState *hb);void cxl_doe_cdat_init (CXLComponentState *cxl_cstate,Error **errp);void cxl_doe_cdat_release (CXLComponentState *cxl_cstate);
include/hw/pci/pci_bridge.h +1 -0
@@ -92,6 +92,7 @@ struct PXBDev {uint8_t bus_nr;uint16_t numa_node;bool bypass_iommu; +bool hdm_for_passthrough;struct cxl_dev {CXLHost *cxl_host_bridge;/* Pointer to a CXLHost */ } cxl;
Signature · 1 line
2.37.2
Quoted 32 lines · Jonathan Cameron
Until now, testing using CXL has relied up always using two root ports below a host bridge, to work around a current assumption in the Linux kernel support that, in the single root port case, the implementation will use the allowed passthrough decoder implementation choice. If that choice is made all accesses are routed from the host bridge to the single root port that is present. Effectively we have a pass through decoder (it is called that in the kernel driver).
This patch series implements that functionality and makes it the default See patch 2 for a discussion of why I think we can make this change without backwards compatibility issues (basically if it didn't work before who are we breaking by making it work?)
Whilst this limitation has been known since the initial QEMU patch postings / kernel CXL region support, Fan Ni Ran into it recently reminding me that we should solve it.
https://lore.kernel.org/linux-cxl/20230113171044.GA24788@bgt-140510-bm03/
Tree with a large set of patches before this at:
https://gitlab.com/jic23/qemu/-/tree/cxl-2023-01-20I've done some basic testing, though I did hit what appears to be a kernel race on region bring up of existing region / namespace in a 1HB 2RP 2EP test case. That is proving hard to replicate consistently but doesn't seem to have anything to do with the emulation other than perhaps we are opening up a race by responding slowly to something.
Jonathan Cameron (2):
hw/pci: Addpcie_count_ds_port()andpcie_find_port_first()helpers hw/pxb-cxl: Support passthrough HDM Decoders unless overridden
Quoted 13 lines
hw/cxl/cxl-host.c | 31 +++++++++++++-------- hw/pci-bridge/pci_expander_bridge.c | 43 +++++++++++++++++++++++++---- hw/pci/pcie_port.c | 38 +++++++++++++++++++++++++ include/hw/cxl/cxl.h | 1 + include/hw/cxl/cxl_component.h | 1 + include/hw/pci/pci_bridge.h | 1 + include/hw/pci/pcie_port.h | 2 ++ 7 files changed, 100 insertions(+), 17 deletions(-)Signature · 1 line
2.37.2
Tried three different cxl topology setups (1HB1RP, 1HB2RP2Memdev, with switch), the patch works fine for me.
Btw, there seem some format issues with the patch, got warnings with checkpatch tool.
Quoted 45 lines · On Mon, 23 Jan 2023 17:53:24 +0000 Fan Ni <[email protected]> wrote:
Quoted 41 lines · Jonathan Cameron
Until now, testing using CXL has relied up always using two root ports below a host bridge, to work around a current assumption in the Linux kernel support that, in the single root port case, the implementation will use the allowed passthrough decoder implementation choice. If that choice is made all accesses are routed from the host bridge to the single root port that is present. Effectively we have a pass through decoder (it is called that in the kernel driver).
This patch series implements that functionality and makes it the default See patch 2 for a discussion of why I think we can make this change without backwards compatibility issues (basically if it didn't work before who are we breaking by making it work?)
Whilst this limitation has been known since the initial QEMU patch postings / kernel CXL region support, Fan Ni Ran into it recently reminding me that we should solve it.
https://lore.kernel.org/linux-cxl/20230113171044.GA24788@bgt-140510-bm03/
Tree with a large set of patches before this at:
https://gitlab.com/jic23/qemu/-/tree/cxl-2023-01-20I've done some basic testing, though I did hit what appears to be a kernel race on region bring up of existing region / namespace in a 1HB 2RP 2EP test case. That is proving hard to replicate consistently but doesn't seem to have anything to do with the emulation other than perhaps we are opening up a race by responding slowly to something.
Jonathan Cameron (2):
hw/pci: Addpcie_count_ds_port()andpcie_find_port_first()helpers hw/pxb-cxl: Support passthrough HDM Decoders unless overriddenhw/cxl/cxl-host.c | 31 +++++++++++++-------- hw/pci-bridge/pci_expander_bridge.c | 43 +++++++++++++++++++++++++---- hw/pci/pcie_port.c | 38 +++++++++++++++++++++++++ include/hw/cxl/cxl.h | 1 + include/hw/cxl/cxl_component.h | 1 + include/hw/pci/pci_bridge.h | 1 + include/hw/pci/pcie_port.h | 2 ++ 7 files changed, 100 insertions(+), 17 deletions(-)
Quoted 8 lines
Quoted 4 lines
Signature · 1 line
2.37.2Tried three different cxl topology setups (1HB1RP, 1HB2RP2Memdev, with switch), the patch works fine for me.
Btw, there seem some format issues with the patch, got warnings with checkpatch tool.
Thanks! I'll clean those up. Was being lazy on it as it's an RFC for now :) Given this is small and useful I'll probably pull it nearer the head of the queue.
When I repost, if you could give a Tested-by tag that would be great!
Thanks,
Jonathan
Quoted 60 lines · Jonathan Cameron
Quoted 45 lines · On Mon, 23 Jan 2023 17:53:24 +0000 Fan Ni <[email protected]> wrote:
Quoted 41 lines · Jonathan Cameron
Until now, testing using CXL has relied up always using two root ports below a host bridge, to work around a current assumption in the Linux kernel support that, in the single root port case, the implementation will use the allowed passthrough decoder implementation choice. If that choice is made all accesses are routed from the host bridge to the single root port that is present. Effectively we have a pass through decoder (it is called that in the kernel driver).
This patch series implements that functionality and makes it the default See patch 2 for a discussion of why I think we can make this change without backwards compatibility issues (basically if it didn't work before who are we breaking by making it work?)
Whilst this limitation has been known since the initial QEMU patch postings / kernel CXL region support, Fan Ni Ran into it recently reminding me that we should solve it.
Tree with a large set of patches before this at:
https://urldefense.com/v3/__https://gitlab.com/jic23/qemu/-/tree/cxl-2023-01-20__;!!EwVzqGoTKBqv-0DWAJBm!WsD6FV-KV_YnhhHWLll72cHLqLQ_8kpps3MlpAa6Bonsdz6aifuWT40-QnlRyFqWyeyaVb-RiC03_qajbeCGsPjbv12T$I've done some basic testing, though I did hit what appears to be a kernel race on region bring up of existing region / namespace in a 1HB 2RP 2EP test case. That is proving hard to replicate consistently but doesn't seem to have anything to do with the emulation other than perhaps we are opening up a race by responding slowly to something.
Jonathan Cameron (2):
hw/pci: Addpcie_count_ds_port()andpcie_find_port_first()helpers hw/pxb-cxl: Support passthrough HDM Decoders unless overriddenhw/cxl/cxl-host.c | 31 +++++++++++++-------- hw/pci-bridge/pci_expander_bridge.c | 43 +++++++++++++++++++++++++---- hw/pci/pcie_port.c | 38 +++++++++++++++++++++++++ include/hw/cxl/cxl.h | 1 + include/hw/cxl/cxl_component.h | 1 + include/hw/pci/pci_bridge.h | 1 + include/hw/pci/pcie_port.h | 2 ++ 7 files changed, 100 insertions(+), 17 deletions(-)Quoted 8 lines
Quoted 4 lines
Signature · 1 line
2.37.2Tried three different cxl topology setups (1HB1RP, 1HB2RP2Memdev, with switch), the patch works fine for me.
Btw, there seem some format issues with the patch, got warnings with checkpatch tool.Thanks! I'll clean those up. Was being lazy on it as it's an RFC for now :) Given this is small and useful I'll probably pull it nearer the head of the queue.
Quoted 6 lines
When I repost, if you could give a Tested-by tag that would be great!
Thanks,
Jonathan
Will do. Thanks.