ice: Fix enabling SR-IOV with Xen
Subsystems (3)
PATCH v2 1/1 ice: Fix enabling SR-IOV with Xen
When the PCI functions are created, Xen is informed about them and caches the number of MSI-X entries each function has. However, the number of MSI-X entries is not set until after the hardware has been configured and the VFs have been started. This prevents PCI-passthrough from working because Xen rejects mapping MSI-X interrupts to domains because it thinks the MSI-X interrupts don't exist.
Fix this by moving the call to pci_enable_sriov() later so that the number of MSI-X entries is set correctly in hardware by the time Xen reads it.
---
In v2:
* Fix cleanup on if pci_enable_sriov() fails.
drivers/net/ethernet/intel/ice/ice_sriov.c | 23 +++++++++++++--------- 1 file changed, 14 insertions(+), 9 deletions(-)
drivers/net/ethernet/intel/ice/ice_sriov.c +14 -9
@@ -864,6 +864,8 @@ static int ice_ena_vfs(struct ice_pf *pf, u16 num_vfs)int total_vectors = pf->hw.func_caps.common_cap.num_msix_vectors;struct device *dev =ice_pf_to_dev (pf);struct ice_hw *hw = &pf->hw; +struct ice_vf *vf; +unsigned int bkt;int ret; pf->sriov_irq_bm =bitmap_zalloc (total_vectors,GFP_KERNEL );
@@ -877,24 +879,20 @@ static int ice_ena_vfs(struct ice_pf *pf, u16 num_vfs)set_bit (ICE_OICR_INTR_DIS , pf->state);ice_flush (hw); - ret =pci_enable_sriov (pf->pdev, num_vfs); -if (ret) - goto err_unroll_intr; -mutex_lock (&pf->vfs.table_lock); ret =ice_set_per_vf_res (pf, num_vfs);if (ret) {dev_err (dev,"Not enough resources for %d VFs, err %d. Try with fewer number of VFs\n" , num_vfs, ret); - goto err_unroll_sriov; + goto err_unroll_intr; } ret =ice_create_vf_entries (pf, num_vfs);if (ret) {dev_err (dev,"Failed to allocate VF entries for %d VFs\n" , num_vfs); - goto err_unroll_sriov; + goto err_unroll_intr; }ice_eswitch_reserve_cp_queues (pf, num_vfs);
@@ -905,6 +903,10 @@ static int ice_ena_vfs(struct ice_pf *pf, u16 num_vfs) goto err_unroll_vf_entries; } + ret =pci_enable_sriov (pf->pdev, num_vfs); +if (ret) + goto err_unroll_start_vfs; +clear_bit (ICE_VF_DIS , pf->state);/* rearm global interrupts */
@@ -915,12 +917,15 @@ static int ice_ena_vfs(struct ice_pf *pf, u16 num_vfs)return 0 ; +err_unroll_start_vfs: +ice_for_each_vf (pf, bkt, vf) { +ice_dis_vf_mappings (vf); +ice_vf_vsi_release (vf); + } err_unroll_vf_entries:ice_free_vf_entries (pf); -err_unroll_sriov: -mutex_unlock (&pf->vfs.table_lock); -pci_disable_sriov (pf->pdev); err_unroll_intr: +mutex_unlock (&pf->vfs.table_lock);/* rearm interrupts here */ ice_irq_dynamic_ena (hw,NULL ,NULL );clear_bit (ICE_OICR_INTR_DIS , pf->state);
Signature · 1 line
2.43.0
Dear Ross,
Thank you for your patch.
Am 29.04.24 um 14:49 schrieb Ross Lagerwall:
Quoted 7 lines
When the PCI functions are created, Xen is informed about them and caches the number of MSI-X entries each function has. However, the number of MSI-X entries is not set until after the hardware has been configured and the VFs have been started. This prevents PCI-passthrough from working because Xen rejects mapping MSI-X interrupts to domains because it thinks the MSI-X interrupts don't exist.
Thank you for this great problem description. Is there any log message shown, you could paste, so people can find this commit when searching for the log message?
Do you have a minimal test case, so the maintainers can reproduce this to test the fix?
Quoted 3 lines
Fix this by moving the call to
pci_enable_sriov()later so that the number of MSI-X entries is set correctly in hardware by the time Xen reads it.
It’d be great if you could be more specific on “later”, and why this is the correct place.
Quoted 62 lines
Signed-off-by Ross Lagerwall <[email protected]>Signed-off-by Javi Merino <[email protected]>---
In v2:
* Fix cleanup on ifpci_enable_sriov()fails.drivers/net/ethernet/intel/ice/ice_sriov.c | 23 +++++++++++++--------- 1 file changed, 14 insertions(+), 9 deletions(-)drivers/net/ethernet/intel/ice/ice_sriov.c +8
-6@@ -864,6 +864,8 @@ static int ice_ena_vfs(struct ice_pf *pf, u16 num_vfs)int total_vectors = pf->hw.func_caps.common_cap.num_msix_vectors;struct device *dev =ice_pf_to_dev (pf);struct ice_hw *hw = &pf->hw; +struct ice_vf *vf; +unsigned int bkt;int ret; pf->sriov_irq_bm =bitmap_zalloc (total_vectors,GFP_KERNEL );@@ -877,24 +879,20 @@ static int ice_ena_vfs(struct ice_pf *pf, u16 num_vfs)set_bit (ICE_OICR_INTR_DIS , pf->state);ice_flush (hw); - ret =pci_enable_sriov (pf->pdev, num_vfs); -if (ret) - goto err_unroll_intr; -mutex_lock (&pf->vfs.table_lock); ret =ice_set_per_vf_res (pf, num_vfs);if (ret) {dev_err (dev,"Not enough resources for %d VFs, err %d. Try with fewer number of VFs\n" , num_vfs, ret); - goto err_unroll_sriov; + goto err_unroll_intr; } ret =ice_create_vf_entries (pf, num_vfs);if (ret) {dev_err (dev,"Failed to allocate VF entries for %d VFs\n" , num_vfs); - goto err_unroll_sriov; + goto err_unroll_intr; }ice_eswitch_reserve_cp_queues (pf, num_vfs);@@ -905,6 +903,10 @@ static int ice_ena_vfs(struct ice_pf *pf, u16 num_vfs) goto err_unroll_vf_entries; } + ret =pci_enable_sriov (pf->pdev, num_vfs); +if (ret) + goto err_unroll_start_vfs; +clear_bit (ICE_VF_DIS , pf->state);/* rearm global interrupts */
drivers/net/ethernet/intel/ice/ice_sriov.c static int ice_ena_vfs(struct ice_pf *pf, u16 num_vfs) · line 924
Quoted 8 lines
drivers/net/ethernet/intel/ice/ice_sriov.c +5
-0@@ -915,12 +917,15 @@ static int ice_ena_vfs(struct ice_pf *pf, u16 num_vfs)return 0 ; +err_unroll_start_vfs: +ice_for_each_vf (pf, bkt, vf) { +ice_dis_vf_mappings (vf); +ice_vf_vsi_release (vf); + }
Why wasn’t this needed with pci_enable_sriov() done earlier?
drivers/net/ethernet/intel/ice/ice_sriov.c static int ice_ena_vfs(struct ice_pf *pf, u16 num_vfs) · line 931
Quoted 10 lines
drivers/net/ethernet/intel/ice/ice_sriov.c +1
-3err_unroll_vf_entries:ice_free_vf_entries (pf); -err_unroll_sriov: -mutex_unlock (&pf->vfs.table_lock); -pci_disable_sriov (pf->pdev); err_unroll_intr: +mutex_unlock (&pf->vfs.table_lock);/* rearm interrupts here */ ice_irq_dynamic_ena (hw,NULL ,NULL );clear_bit (ICE_OICR_INTR_DIS , pf->state);
Kind regards,
Paul
Quoted 15 lines · Paul Menzel
Dear Ross,
Thank you for your patch.
Am 29.04.24 um 14:49 schrieb Ross Lagerwall:
Quoted 7 lines
When the PCI functions are created, Xen is informed about them and caches the number of MSI-X entries each function has. However, the number of MSI-X entries is not set until after the hardware has been configured and the VFs have been started. This prevents PCI-passthrough from working because Xen rejects mapping MSI-X interrupts to domains because it thinks the MSI-X interrupts don't exist.
Quoted 3 lines
Thank you for this great problem description. Is there any log message shown, you could paste, so people can find this commit when searching for the log message?
When this issue occurs, QEMU repeatedly reports:
msi_msix_setup: Error: Mapping of MSI-X (err: 22, vec: 0, entry 0x1)
Quoted 3 lines
Do you have a minimal test case, so the maintainers can reproduce this to test the fix?
Testing this requires setting up Xen which I wouldn't expect anyone to do unless they already have an environment set up.
In any case, a "minimal" test would be something like:
1. Set up Xen with dom0 and another VM running Linux.
2. Pass through a VF to the VM. See that QEMU reports the above message and the VF is not usable within the VM.
3. Rebuild the dom0 kernel with the attached patch.
4. Pass through a VF to the VM. See that the VF is usable within the VM.
Quoted 7 lines
Quoted 4 lines
Fix this by moving the call to
pci_enable_sriov()later so that the number of MSI-X entries is set correctly in hardware by the time Xen reads it.It’d be great if you could be more specific on “later”, and why this is the correct place.
"later" in this case means after ice_start_vfs() since it is at that point that the hardware sets the number of MSI-X entries. I expect that a maintainer or someone with more knowledge of the hardware could explain why the hardware only sets the number of MSI-X entries at this point.
Quoted 66 lines
Quoted 66 lines
Signed-off-by Ross Lagerwall <[email protected]>Signed-off-by Javi Merino <[email protected]>---
In v2:
* Fix cleanup on ifpci_enable_sriov()fails.drivers/net/ethernet/intel/ice/ice_sriov.c | 23 +++++++++++++--------- 1 file changed, 14 insertions(+), 9 deletions(-)drivers/net/ethernet/intel/ice/ice_sriov.c +8
-6@@ -864,6 +864,8 @@ static int ice_ena_vfs(struct ice_pf *pf, u16 num_vfs)int total_vectors = pf->hw.func_caps.common_cap.num_msix_vectors;struct device *dev =ice_pf_to_dev (pf);struct ice_hw *hw = &pf->hw; +struct ice_vf *vf; +unsigned int bkt;int ret; pf->sriov_irq_bm =bitmap_zalloc (total_vectors,GFP_KERNEL );@@ -877,24 +879,20 @@ static int ice_ena_vfs(struct ice_pf *pf, u16 num_vfs)set_bit (ICE_OICR_INTR_DIS , pf->state);ice_flush (hw); - ret =pci_enable_sriov (pf->pdev, num_vfs); -if (ret) - goto err_unroll_intr; -mutex_lock (&pf->vfs.table_lock); ret =ice_set_per_vf_res (pf, num_vfs);if (ret) {dev_err (dev,"Not enough resources for %d VFs, err %d. Try with fewer number of VFs\n" , num_vfs, ret); - goto err_unroll_sriov; + goto err_unroll_intr; } ret =ice_create_vf_entries (pf, num_vfs);if (ret) {dev_err (dev,"Failed to allocate VF entries for %d VFs\n" , num_vfs); - goto err_unroll_sriov; + goto err_unroll_intr; }ice_eswitch_reserve_cp_queues (pf, num_vfs);@@ -905,6 +903,10 @@ static int ice_ena_vfs(struct ice_pf *pf, u16 num_vfs) goto err_unroll_vf_entries; } + ret =pci_enable_sriov (pf->pdev, num_vfs); +if (ret) + goto err_unroll_start_vfs; +clear_bit (ICE_VF_DIS , pf->state);/* rearm global interrupts */ @@ -915,12 +917,15 @@ static int ice_ena_vfs(struct ice_pf *pf, u16 num_vfs)return 0 ;
Quoted 7 lines
drivers/net/ethernet/intel/ice/ice_sriov.c static int ice_ena_vfs(struct ice_pf *pf, u16 num_vfs) · line 924
Quoted 5 lines
drivers/net/ethernet/intel/ice/ice_sriov.c +5
-0+err_unroll_start_vfs: +ice_for_each_vf (pf, bkt, vf) { +ice_dis_vf_mappings (vf); +ice_vf_vsi_release (vf); + }Why wasn’t this needed with
pci_enable_sriov()done earlier?
Previously ice_start_vifs() was the last function call that may fail in this function. That is no longer the case so when pci_enable_sriov() fails, it needs to undo what was done in ice_start_vifs().
Thanks,
Ross
This patch makes sense since VFs need to be assigned resources (especially MSI-X interrupt count) before making these VFs visible, so that the kernel PCI enumeration code reads correct MSI-X interrupt count for the VFs.
Regards,
Sergey
Tony Nguyen
maintainerQuoted 8 lines · Ross Lagerwall
When the PCI functions are created, Xen is informed about them and caches the number of MSI-X entries each function has. However, the number of MSI-X entries is not set until after the hardware has been configured and the VFs have been started. This prevents PCI-passthrough from working because Xen rejects mapping MSI-X interrupts to domains because it thinks the MSI-X interrupts don't exist.
Quoted 4 lines
Fix this by moving the call to
pci_enable_sriov()later so that the number of MSI-X entries is set correctly in hardware by the time Xen reads it.
Sorry, I missed this on initial review, but bug fixes should have a Fixes: tag
I assume you are targeting this for net, if so, can you mark it as 'PATCH iwl-net'.
Quoted 2 lines
Signed-off-by Ross Lagerwall <[email protected]>Signed-off-by Javi Merino <[email protected]>
Also, sender should be the last sign-off.
Thanks,
Tony