i915 vgpu PV to improve vgpu performance
v4 cover i915 vgpu PV to improve vgpu performance
To improve vgpu performance, it could implement some PV optimization such as to reduce the mmio access trap numbers or eliminate certain piece of HW emulation within guest driver to reduce vm exit/vm enter cost.
the solutions in this patch set are implemented two PV optimizations based on the shared memory region between guest and GVTg for data communication. The shared memory region is allocated by guest driver and this region's memory guest physical address will be passed to GVTg through PVINFO register and later GVTg can access this region directly without trap cost to achieve data exchange purpose between guest and GVTg.
in this patch set, 2 kind of PV optimization implemented controlled by pv_caps PVINO register with different pv bit.
1. workload PV submission (context submission): reduce 4 traps to 1 trap and eliminated execlists HW behaviour emulation.
2. ppgtt PV update: eliminate the cost of ppgtt write protection.
based on the experiment, for small workloads, specifally, glxgears with vblank_mode off, the average performance gain on single vgpu is 30~50%. for large workload such as media and 3D, the average performance gain is about 4%.
based on the PV mechanism, it could achive more vgpu feature optimization such as globle GTT update, display plane and water mark update.
v0: RFC patch set
v1: addressed RFC review comments
v2: addressed v1 review comments, added pv callbacks for pv operations v3:
1. addressed v2 review comments, removed pv callbacks code duplication in v2 and unified pv calls under g2v notification register. different g2v pv notifications defined.
2. dropped pv master irq feature due to hard conflict with recnet i915 change and take time to rework.
v4:
1. addressed v3 review comments.
2. extended workload PV submission by skip execlists HW behaviour emulation and context switch interrupt injection.
Xiaolin Zhang (8):
drm/i915: introduced vgpu pv capability
drm/i915: vgpu shared memory setup for pv optimization drm/i915: vgpu ppgtt update pv optimization drm/i915: vgpu context submission pv optimization drm/i915/gvt: GVTg handle pv_caps PVINFO register drm/i915/gvt: GVTg handle shared_page setup drm/i915/gvt: GVTg support ppgtt pv optimization drm/i915/gvt: GVTg support context submission pv optimization
drivers/gpu/drm/i915/gvt/execlist.c | 6 + drivers/gpu/drm/i915/gvt/gtt.c | 317 +++++++++++++++++++++++++++++++++++ drivers/gpu/drm/i915/gvt/gtt.h | 9 + drivers/gpu/drm/i915/gvt/gvt.h | 10 +- drivers/gpu/drm/i915/gvt/handlers.c | 63 ++++++- drivers/gpu/drm/i915/gvt/vgpu.c | 31 ++++ drivers/gpu/drm/i915/i915_drv.h | 6 +- drivers/gpu/drm/i915/i915_gem.c | 3 +- drivers/gpu/drm/i915/i915_gem_gtt.c | 9 +- drivers/gpu/drm/i915/i915_gem_gtt.h | 8 + drivers/gpu/drm/i915/i915_irq.c | 2 + drivers/gpu/drm/i915/i915_pvinfo.h | 12 +- drivers/gpu/drm/i915/i915_vgpu.c | 323 +++++++++++++++++++++++++++++++++++- drivers/gpu/drm/i915/i915_vgpu.h | 52 ++++++ drivers/gpu/drm/i915/intel_lrc.c | 3 + 15 files changed, 845 insertions(+), 9 deletions(-)
Signature · 6 lines
2.7.4 _______________________________________________ Intel-gfx mailing list [email protected] https://lists.freedesktop.org/mailman/listinfo/intel-gfx
Patchwork
CI✗ Fi.CI.BAT: failure for i915 vgpu PV to improve vgpu performance
== Series Details ==
== Series Details ==
Series: i915 vgpu PV to improve vgpu performance
URL : https://patchwork.freedesktop.org/series/58707/
State : failure
== Summary ==
CALL scripts/checksyscalls.sh
CALL scripts/atomic/check-atomics.sh
DESCEND objtool
CHK include/generated/compile.h
CC [M] drivers/gpu/drm/i915/i915_vgpu.odrivers/gpu/drm/i915/i915_vgpu.c: In function ‘pv_submit’:drivers/gpu/drm/i915/i915_vgpu.c:337:2: error: implicit declaration of function ‘__raw_i915_write32’; did you mean ‘__raw_uncore_write32’? [-Werror=implicit-function-declaration]
__raw_i915_write32(uncore, vgtif_reg(g2v_notify),
^~~~~~~~~~~~~~~~~~
__raw_uncore_write32drivers/gpu/drm/i915/i915_vgpu.c: In function ‘intel_vgpu_setup_shared_page’:drivers/gpu/drm/i915/i915_vgpu.c:523:2: error: implicit declaration of function ‘__raw_i915_write64’; did you mean ‘__raw_uncore_write64’? [-Werror=implicit-function-declaration]
__raw_i915_write64(uncore, vgtif_reg(shared_page_gpa), gpa);
^~~~~~~~~~~~~~~~~~
__raw_uncore_write64drivers/gpu/drm/i915/i915_vgpu.c:524:13: error: implicit declaration of function ‘__raw_i915_read64’; did you mean ‘__raw_uncore_read64’? [-Werror=implicit-function-declaration]
if (gpa != __raw_i915_read64(uncore, vgtif_reg(shared_page_gpa))) {
^~~~~~~~~~~~~~~~~
__raw_uncore_read64
drivers/gpu/drm/i915/i915_vgpu.c: In function ‘intel_vgpu_check_pv_caps’:
drivers/gpu/drm/i915/i915_vgpu.c:582:15: error: implicit declaration of function ‘__raw_i915_read32’; did you mean ‘__raw_uncore_read32’? [-Werror=implicit-function-declaration]
gvt_pvcaps = __raw_i915_read32(uncore, vgtif_reg(pv_caps));
^~~~~~~~~~~~~~~~~
__raw_uncore_read32cc1: all warnings being treated as errorsscripts/Makefile.build:278: recipe for target 'drivers/gpu/drm/i915/i915_vgpu.o' failed make[5]: *** [drivers/gpu/drm/i915/i915_vgpu.o] Error 1 scripts/Makefile.build:489: recipe for target 'drivers/gpu/drm/i915' failed make[4]: *** [drivers/gpu/drm/i915] Error 2scripts/Makefile.build:489: recipe for target 'drivers/gpu/drm' failed make[3]: *** [drivers/gpu/drm] Error 2scripts/Makefile.build:489: recipe for target 'drivers/gpu' failed make[2]: *** [drivers/gpu] Error 2/home/cidrm/kernel/Makefile:1046: recipe for target 'drivers' failed make[1]: *** [drivers] Error 2
Makefile:170: recipe for target 'sub-make' failed
make: *** [sub-make] Error 2
_______________________________________________
Intel-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/intel-gfx
PATCH v4 1/8 drm/i915: introduced vgpu pv capability
In reply to Xiaolin Zhangpv capability for guest gpu was introduced by pv_caps in struct i915_virtual_gpu and a new pv_caps register for host GVT was defined in struct vgt_if for vgpu pv optimization.
both of them are used to control different feature pv optimization supported and implemented by both guest and host.
These fields are default zero, no any pv feature enabled.
it also adds VGT_CAPS_PV capability BIT for guest to check GVTg can support PV feature or not.
v0: RFC, introudced enable_pvmmio module parameter.
v1: addressed RFC comment to remove enable_pvmmio module parameter by pv capability check.
v2: rebase
v3: distinct pv caps from guest and host. renamed enable_pvmmio to pvmmio_caps which is used for host pv caps.
v4: consolidated all pv related functons into a single file i915_vgpu.c and renamed pvmmio to pv_caps.
--- drivers/gpu/drm/i915/i915_drv.h | 2 ++ drivers/gpu/drm/i915/i915_pvinfo.h | 5 ++++- drivers/gpu/drm/i915/i915_vgpu.c | 45 +++++++++++++++++++++++++++++++++++++- drivers/gpu/drm/i915/i915_vgpu.h | 8 +++++++ 4 files changed, 58 insertions(+), 2 deletions(-)
drivers/gpu/drm/i915/i915_drv.h +2 -0
@@ -86,6 +86,7 @@#include "i915_vma.h" #include "intel_gvt.h" +#include "i915_pvinfo.h" /* General customization: */
@@ -1246,6 +1247,7 @@ struct i915_frontbuffer_tracking {struct i915_virtual_gpu {bool active;u32 caps; +u32 pv_caps; };/* used in computing the new watermarks state */
drivers/gpu/drm/i915/i915_pvinfo.h +4 -1
@@ -55,6 +55,7 @@ enum vgt_g2v_type {#define VGT_CAPS_FULL_PPGTT BIT(2)#define VGT_CAPS_HWSP_EMULATION BIT(3)#define VGT_CAPS_HUGE_GTT BIT(4) +#define VGT_CAPS_PV BIT(5)struct vgt_if {u64 magic;/* VGT_MAGIC */
@@ -107,7 +108,9 @@ struct vgt_if {u32 execlist_context_descriptor_lo;u32 execlist_context_descriptor_hi; -u32 rsv7[0x200 -24 ];/* pad to one page */ +u32 pv_caps; + +u32 rsv7[0x200 -25 ];/* pad to one page */ } __packed;#define vgtif_reg (x) \
drivers/gpu/drm/i915/i915_vgpu.c +44 -1
54 lines not on this page · raw message
drivers/gpu/drm/i915/i915_vgpu.h +8 -0
@@ -42,7 +42,15 @@ intel_vgpu_has_huge_gtt(struct drm_i915_private *dev_priv) return dev_priv->vgpu.caps & VGT_CAPS_HUGE_GTT; } +static inline bool +intel_vgpu_has_pv_caps (struct drm_i915_private *dev_priv) +{ +return dev_priv->vgpu.caps &VGT_CAPS_PV ; +} +int intel_vgt_balloon (struct drm_i915_private *dev_priv);void intel_vgt_deballoon (struct drm_i915_private *dev_priv); +/* i915 vgpu pv related functions */ +bool intel_vgpu_check_pv_caps (struct drm_i915_private *dev_priv);#endif /* _I915_VGPU_H_ */
Signature · 6 lines
2.7.4 _______________________________________________ Intel-gfx mailing list [email protected] https://lists.freedesktop.org/mailman/listinfo/intel-gfx
PATCH v4 2/8 drm/i915: vgpu shared memory setup for pv optimization
In reply to Xiaolin ZhangTo enable vgpu pv features, we need to setup a shared memory page which will be used for data exchange directly accessed between both guest and backend i915 driver to avoid emulation trap cost.
guest i915 will allocate this page memory and then pass it's physical address to backend i915 driver through PVINFO register so that backend i915 driver can access this shared page meory without any trap cost with the help form hyperviser's read guest gpa functionality.
guest i915 will send VGT_G2V_SHARED_PAGE_SETUP notification to host GVT once shared memory setup finished.
the layout of the shared_page also defined as well in this patch which is used for pv features implementation.
v0: RFC
v1: addressed RFC comment to move both shared_page_lock and shared_page to i915_virtual_gpu structure
v2: packed i915_virtual_gpu structure
v3: added SHARED_PAGE_SETUP g2v notification for pv shared_page setup
v4: added intel_vgpu_setup_shared_page() in i915_vgpu_pv.c
--- drivers/gpu/drm/i915/i915_drv.h | 4 +++- drivers/gpu/drm/i915/i915_pvinfo.h | 5 ++++- drivers/gpu/drm/i915/i915_vgpu.c | 38 ++++++++++++++++++++++++++++++++++++++ drivers/gpu/drm/i915/i915_vgpu.h | 17 +++++++++++++++++ 4 files changed, 62 insertions(+), 2 deletions(-)
drivers/gpu/drm/i915/i915_drv.h +3 -1
@@ -1248,7 +1248,9 @@ struct i915_virtual_gpu {bool active;u32 caps;u32 pv_caps; -}; +spinlock_t shared_page_lock; +struct gvt_shared_page *shared_page; +} __packed;/* used in computing the new watermarks state */ struct intel_wm_config {
drivers/gpu/drm/i915/i915_pvinfo.h +4 -1
@@ -46,6 +46,7 @@ enum vgt_g2v_type {VGT_G2V_PPGTT_L4_PAGE_TABLE_DESTROY ,VGT_G2V_EXECLIST_CONTEXT_CREATE ,VGT_G2V_EXECLIST_CONTEXT_DESTROY , +VGT_G2V_SHARED_PAGE_SETUP , VGT_G2V_MAX, };
@@ -110,7 +111,9 @@ struct vgt_if {u32 pv_caps; -u32 rsv7[0x200 -25 ];/* pad to one page */ +u64 shared_page_gpa; + +u32 rsv7[0x200 -27 ];/* pad to one page */ } __packed;#define vgtif_reg (x) \
drivers/gpu/drm/i915/i915_vgpu.c +38 -0
@@ -135,6 +135,9 @@ void intel_vgt_deballoon(struct drm_i915_private *dev_priv)for (i =0 ; i <4 ; i++)vgt_deballoon_space (&dev_priv->ggtt, &bl_info.space[i]); + +if (dev_priv->vgpu.shared_page) +free_page ((unsigned long )dev_priv->vgpu.shared_page); } static int vgt_balloon_space(struct i915_ggtt *ggtt,
@@ -286,6 +289,36 @@ int intel_vgt_balloon(struct drm_i915_private *dev_priv)* i915 vgpu PV support for Linux */ +/* +* shared_page setup for VGPU PV features +*/ +static int intel_vgpu_setup_shared_page (struct drm_i915_private *dev_priv) +{ +struct intel_uncore *uncore = &dev_priv->uncore; +u64 gpa; + + dev_priv->vgpu.shared_page = (struct gvt_shared_page *) +get_zeroed_page (GFP_KERNEL ); +if (!dev_priv->vgpu.shared_page) { +DRM_ERROR ("out of memory for shared page memory\n" ); +return -ENOMEM ; + } + +/* pass guest memory pa address to GVT and then read back to verify */ + gpa =__pa (dev_priv->vgpu.shared_page); +__raw_i915_write64 (uncore,vgtif_reg (shared_page_gpa), gpa); +if (gpa !=__raw_i915_read64 (uncore,vgtif_reg (shared_page_gpa))) { +DRM_ERROR ("vgpu: passed shared_page_gpa failed\n" ); +free_page ((unsigned long )dev_priv->vgpu.shared_page); +return -EIO ; + } +__raw_i915_write32 (uncore,vgtif_reg (g2v_notify), +VGT_G2V_SHARED_PAGE_SETUP ); +spin_lock_init (&dev_priv->vgpu.shared_page_lock); + +return 0 ; +} +/** * intel_vgpu_check_pv_caps - detect virtual GPU PV capabilities * @dev_priv: i915 device private
@@ -313,6 +346,11 @@ bool intel_vgpu_check_pv_caps(struct drm_i915_private *dev_priv)if (!pvcaps)return false; +if (intel_vgpu_setup_shared_page (dev_priv)) { + dev_priv->vgpu.pv_caps =0 ; +return false; + } +__raw_i915_write32 (uncore,vgtif_reg (pv_caps), pvcaps);return true;
drivers/gpu/drm/i915/i915_vgpu.h +17 -0
@@ -26,6 +26,23 @@#include "i915_pvinfo.h" +/* +* A shared page(4KB) between gvt and VM, could be allocated by guest driver +* or a fixed location in PCI bar 0 region +*/ +struct pv_ppgtt_update { +u64 pdp; +u64 start; +u64 length; +u32 cache_level; +}; + +struct gvt_shared_page { +struct pv_ppgtt_update pv_ppgtt; +u32 ring_id; +u64 descs[EXECLIST_MAX_PORTS ]; +}; +void i915_check_vgpu (struct drm_i915_private *dev_priv);bool intel_vgpu_has_full_ppgtt (struct drm_i915_private *dev_priv);
Signature · 6 lines
2.7.4 _______________________________________________ Intel-gfx mailing list [email protected] https://lists.freedesktop.org/mailman/listinfo/intel-gfx
PATCH v4 3/8 drm/i915: vgpu ppgtt update pv optimization
In reply to Xiaolin ZhangThis patch extends vgpu ppgtt g2v notification to notify host GVT-g of ppgtt update from guest including alloc_4lvl, clear_4lv4 and insert_4lvl.
These updates use the shared memory page to pass struct pv_ppgtt_update from guest to GVT which is used for pv optimiation implemeation within host GVT side.
This patch also add one new pv_caps level to control ppgtt update.
Use PV_PPGTT_UPDATE to control this level of pv optimization.
v0: RFC
v1: rebased
v2: added pv callbacks for vm.{allocate_va_range, insert_entries, clear_range} within ppgtt.
v3: rebased, disable huge page ppgtt support when using PVMMIO ppgtt update due to complex and performance impact.
v4: moved alloc/insert/clear_4lvl pv callbacks into i915_vgpu_pv.c and added a single intel_vgpu_config_pv_caps() for vgpu pv callbacks setup.
--- drivers/gpu/drm/i915/i915_gem.c | 3 +- drivers/gpu/drm/i915/i915_gem_gtt.c | 9 ++-- drivers/gpu/drm/i915/i915_gem_gtt.h | 8 ++++ drivers/gpu/drm/i915/i915_pvinfo.h | 3 ++ drivers/gpu/drm/i915/i915_vgpu.c | 84 +++++++++++++++++++++++++++++++++++++ drivers/gpu/drm/i915/i915_vgpu.h | 17 ++++++++ 6 files changed, 120 insertions(+), 4 deletions(-)
drivers/gpu/drm/i915/i915_gem.c +2 -1
@@ -4826,7 +4826,8 @@ int i915_gem_init(struct drm_i915_private *dev_priv)int ret;/* We need to fallback to 4K pages if host doesn't support huge gtt. */ -if (intel_vgpu_active (dev_priv) && !intel_vgpu_has_huge_gtt (dev_priv)) +if ((intel_vgpu_active (dev_priv) && !intel_vgpu_has_huge_gtt (dev_priv)) + ||intel_vgpu_enabled_pv_caps (dev_priv,PV_PPGTT_UPDATE ))mkwrite_device_info (dev_priv)->page_sizes =I915_GTT_PAGE_SIZE_4K ;
drivers/gpu/drm/i915/i915_gem_gtt.c +6 -3
@@ -923,7 +923,7 @@ static void gen8_ppgtt_set_pml4e(struct i915_pml4 *pml4,* This is the top-level structure in 4-level page tables used on gen8+. * Empty entries are always scratch pml4e. */ -static void gen8_ppgtt_clear_4lvl (struct i915_address_space *vm, +void gen8_ppgtt_clear_4lvl (struct i915_address_space *vm,u64 start,u64 length) {struct i915_hw_ppgtt *ppgtt =i915_vm_to_ppgtt (vm);
@@ -1162,7 +1162,7 @@ static void gen8_ppgtt_insert_huge_entries(struct i915_vma *vma, }while (iter->sg); } -static void gen8_ppgtt_insert_4lvl (struct i915_address_space *vm, +void gen8_ppgtt_insert_4lvl (struct i915_address_space *vm,struct i915_vma *vma,enum i915_cache_level cache_level,u32 flags)
@@ -1444,7 +1444,7 @@ static int gen8_ppgtt_alloc_3lvl(struct i915_address_space *vm, &i915_vm_to_ppgtt(vm)->pdp, start, length); } -static int gen8_ppgtt_alloc_4lvl (struct i915_address_space *vm, +int gen8_ppgtt_alloc_4lvl (struct i915_address_space *vm,u64 start,u64 length) {struct i915_hw_ppgtt *ppgtt =i915_vm_to_ppgtt (vm);
@@ -1571,6 +1571,9 @@ static struct i915_hw_ppgtt *gen8_ppgtt_create(struct drm_i915_private *i915) ppgtt->vm.allocate_va_range = gen8_ppgtt_alloc_4lvl; ppgtt->vm.insert_entries = gen8_ppgtt_insert_4lvl; ppgtt->vm.clear_range = gen8_ppgtt_clear_4lvl; + +if (intel_vgpu_active (i915)) +intel_vgpu_config_pv_caps (i915,PV_PPGTT_UPDATE , ppgtt); }else { err =__pdp_init (&ppgtt->vm, &ppgtt->pdp);if (err)
drivers/gpu/drm/i915/i915_gem_gtt.h +8 -0
@@ -625,6 +625,14 @@ int gen6_ppgtt_pin(struct i915_hw_ppgtt *base);void gen6_ppgtt_unpin (struct i915_hw_ppgtt *base);void gen6_ppgtt_unpin_all (struct i915_hw_ppgtt *base); +void gen8_ppgtt_clear_4lvl (struct i915_address_space *vm, +u64 start,u64 length); +void gen8_ppgtt_insert_4lvl (struct i915_address_space *vm, +struct i915_vma *vma, +enum i915_cache_level cache_level,u32 flags); +int gen8_ppgtt_alloc_4lvl (struct i915_address_space *vm, +u64 start,u64 length); +void i915_check_and_clear_faults (struct drm_i915_private *dev_priv);void i915_gem_suspend_gtt_mappings (struct drm_i915_private *dev_priv);void i915_gem_restore_gtt_mappings (struct drm_i915_private *dev_priv);
drivers/gpu/drm/i915/i915_pvinfo.h +3 -0
@@ -47,6 +47,9 @@ enum vgt_g2v_type {VGT_G2V_EXECLIST_CONTEXT_CREATE ,VGT_G2V_EXECLIST_CONTEXT_DESTROY ,VGT_G2V_SHARED_PAGE_SETUP , +VGT_G2V_PPGTT_L4_ALLOC , +VGT_G2V_PPGTT_L4_CLEAR , +VGT_G2V_PPGTT_L4_INSERT , VGT_G2V_MAX, };
drivers/gpu/drm/i915/i915_vgpu.c +84 -0
102 lines not on this page · raw message
drivers/gpu/drm/i915/i915_vgpu.h +17 -0
@@ -27,6 +27,13 @@#include "i915_pvinfo.h" /* +* define different capabilities of PV optimization +*/ +enum pv_caps { +PV_PPGTT_UPDATE =0x1 , +}; + +/* * A shared page(4KB) between gvt and VM, could be allocated by guest driver * or a fixed location in PCI bar 0 region */
@@ -65,9 +72,19 @@ intel_vgpu_has_pv_caps(struct drm_i915_private *dev_priv) return dev_priv->vgpu.caps & VGT_CAPS_PV; } +static inline bool +intel_vgpu_enabled_pv_caps (struct drm_i915_private *dev_priv, +enum pv_caps cap) +{ +return intel_vgpu_active (dev_priv) &&intel_vgpu_has_pv_caps (dev_priv) + && (dev_priv->vgpu.pv_caps & cap); +} +int intel_vgt_balloon (struct drm_i915_private *dev_priv);void intel_vgt_deballoon (struct drm_i915_private *dev_priv);/* i915 vgpu pv related functions */ bool intel_vgpu_check_pv_caps (struct drm_i915_private *dev_priv); +void intel_vgpu_config_pv_caps (struct drm_i915_private *dev_priv, +enum pv_caps cap,void *data);#endif /* _I915_VGPU_H_ */
Signature · 6 lines
2.7.4 _______________________________________________ Intel-gfx mailing list [email protected] https://lists.freedesktop.org/mailman/listinfo/intel-gfx
PATCH v4 4/8 drm/i915: vgpu context submission pv optimization
In reply to Xiaolin ZhangIt is performance optimization to override the actual submisison backend in order to eliminate execlists csb process and reduce mmio trap numbers for workload submission without contextswith interrupt by talking with GVT via PV submisison notification mechanism between guest and GVT.
Use PV_SUBMISSION to control this level of pv optimization.
v0: RFC
v1: rebase
v2: added pv ops for pv context submission. to maximize code resuse, introduced 2 more ops (submit_ports & preempt_context) instead of 1 op (set_default_submission) in engine structure. pv version of submit_ports and preempt_context implemented.
v3:
1. to reduce more code duplication, code refactor and replaced 2 ops "submit_ports & preempt_contex" from v2 by 1 ops "write_desc" in engine structure. pv version of write_des implemented.
2. added VGT_G2V_ELSP_SUBMIT for g2v pv notification.
v4: implemented pv elsp submission tasklet as the backend workload submisison by talking to GVT with PV notificaiton mechanism and renamed VGT_G2V_ELSP_SUBMIT to VGT_G2V_PV_SUBMISIION.
--- drivers/gpu/drm/i915/i915_irq.c | 2 + drivers/gpu/drm/i915/i915_pvinfo.h | 1 + drivers/gpu/drm/i915/i915_vgpu.c | 158 ++++++++++++++++++++++++++++++++++++- drivers/gpu/drm/i915/i915_vgpu.h | 10 +++ drivers/gpu/drm/i915/intel_lrc.c | 3 + 5 files changed, 173 insertions(+), 1 deletion(-)
drivers/gpu/drm/i915/i915_irq.c +2 -0
@@ -37,6 +37,7 @@#include "i915_drv.h" #include "i915_trace.h" #include "intel_drv.h" +#include "i915_vgpu.h" /** * DOC: interrupt handling
@@ -1470,6 +1471,7 @@ gen8_cs_irq_handler(struct intel_engine_cs *engine, u32 iir)if (iir >_RENDER_USER_INTERRUPT ) {intel_engine_breadcrumbs_irq (engine); tasklet |=USES_GUC_SUBMISSION (engine->i915); + tasklet |=USES_PV_SUBMISSION (engine->i915); }if (tasklet)
drivers/gpu/drm/i915/i915_pvinfo.h +1 -0
@@ -50,6 +50,7 @@ enum vgt_g2v_type {VGT_G2V_PPGTT_L4_ALLOC ,VGT_G2V_PPGTT_L4_CLEAR ,VGT_G2V_PPGTT_L4_INSERT , +VGT_G2V_PV_SUBMISSION , VGT_G2V_MAX, };
drivers/gpu/drm/i915/i915_vgpu.c +157 -1
188 lines not on this page · raw message
drivers/gpu/drm/i915/i915_vgpu.h +10 -0
@@ -31,6 +31,7 @@*/ enum pv_caps {PV_PPGTT_UPDATE =0x1 , +PV_SUBMISSION =0x2 , };/*
@@ -80,6 +81,12 @@ intel_vgpu_enabled_pv_caps(struct drm_i915_private *dev_priv, && (dev_priv->vgpu.pv_caps & cap); } +static inline bool +intel_vgpu_is_using_pv_submission (struct drm_i915_private *dev_priv) +{ +return intel_vgpu_enabled_pv_caps (dev_priv,PV_SUBMISSION ); +} +int intel_vgt_balloon (struct drm_i915_private *dev_priv);void intel_vgt_deballoon (struct drm_i915_private *dev_priv);
@@ -87,4 +94,7 @@ void intel_vgt_deballoon(struct drm_i915_private *dev_priv);bool intel_vgpu_check_pv_caps (struct drm_i915_private *dev_priv);void intel_vgpu_config_pv_caps (struct drm_i915_private *dev_priv,enum pv_caps cap,void *data); +#define USES_PV_SUBMISSION (dev_priv) \ + intel_vgpu_is_using_pv_submission(dev_priv) +#endif /* _I915_VGPU_H_ */
drivers/gpu/drm/i915/intel_lrc.c +3 -0
@@ -2352,6 +2352,9 @@ logical_ring_default_vfuncs(struct intel_engine_cs *engine)*/ } engine->emit_bb_start = gen8_emit_bb_start; + +if (intel_vgpu_active (engine->i915)) +intel_vgpu_config_pv_caps (engine->i915,PV_SUBMISSION , engine); } static inline void
Signature · 6 lines
2.7.4 _______________________________________________ Intel-gfx mailing list [email protected] https://lists.freedesktop.org/mailman/listinfo/intel-gfx
Re: [PATCH v4 4/8] drm/i915: vgpu context submission pv optimization
Quoting Xiaolin Zhang (2019-03-29 13:32:40)
Quoted 44 lines
It is performance optimization to override the actual submisison backend in order to eliminate execlists csb process and reduce mmio trap numbers for workload submission without contextswith interrupt by talking with GVT via PV submisison notification mechanism between guest and GVT.
Use
PV_SUBMISSIONto control this level of pv optimization.v0: RFC
v1: rebase
v2: added pv ops for pv context submission. to maximize code resuse, introduced 2 more ops (submit_ports&preempt_context) instead of 1 op (set_default_submission) in engine structure. pv version ofsubmit_portsandpreempt_contextimplemented.
v3:
1. to reduce more code duplication, code refactor and replaced 2 ops "submit_ports&preempt_contex" from v2 by 1 ops "write_desc" in engine structure. pv version ofwrite_desimplemented.
2. addedVGT_G2V_ELSP_SUBMITfor g2v pv notification.
v4: implemented pv elsp submission tasklet as the backend workload submisison by talking to GVT with PV notificaiton mechanism and renamedVGT_G2V_ELSP_SUBMITtoVGT_G2V_PV_SUBMISIION.Signed-off-by Xiaolin Zhang <[email protected]>--- drivers/gpu/drm/i915/i915_irq.c | 2 + drivers/gpu/drm/i915/i915_pvinfo.h | 1 + drivers/gpu/drm/i915/i915_vgpu.c | 158 ++++++++++++++++++++++++++++++++++++- drivers/gpu/drm/i915/i915_vgpu.h | 10 +++ drivers/gpu/drm/i915/intel_lrc.c | 3 + 5 files changed, 173 insertions(+), 1 deletion(-)drivers/gpu/drm/i915/i915_irq.c +1
-0@@ -37,6 +37,7 @@#include "i915_drv.h" #include "i915_trace.h" #include "intel_drv.h" +#include "i915_vgpu.h" /** * DOC: interrupt handling
drivers/gpu/drm/i915/i915_irq.c gen8_cs_irq_handler(struct intel_engine_cs *engine, u32 iir) · line 1474
Quoted 4 lines
drivers/gpu/drm/i915/i915_irq.c +1
-0@@ -1470,6 +1471,7 @@ gen8_cs_irq_handler(struct intel_engine_cs *engine, u32 iir)if (iir >_RENDER_USER_INTERRUPT ) {intel_engine_breadcrumbs_irq (engine); tasklet |=USES_GUC_SUBMISSION (engine->i915); + tasklet |=USES_PV_SUBMISSION (engine->i915);
You call this an optimisation!
-Chris
_______________________________________________
Intel-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/intel-gfx
Re: [PATCH v4 4/8] drm/i915: vgpu context submission pv optimization
In reply to Xiaolin ZhangQuoting Xiaolin Zhang (2019-03-29 13:32:40)
Quoted 13 lines
drivers/gpu/drm/i915/i915_irq.c +1
-0@@ -37,6 +37,7 @@#include "i915_drv.h" #include "i915_trace.h" #include "intel_drv.h" +#include "i915_vgpu.h" /** * DOC: interrupt handling
drivers/gpu/drm/i915/i915_irq.c gen8_cs_irq_handler(struct intel_engine_cs *engine, u32 iir) · line 1474
Quoted 4 lines
drivers/gpu/drm/i915/i915_irq.c +1
-0@@ -1470,6 +1471,7 @@ gen8_cs_irq_handler(struct intel_engine_cs *engine, u32 iir)if (iir >_RENDER_USER_INTERRUPT ) {intel_engine_breadcrumbs_irq (engine); tasklet |=USES_GUC_SUBMISSION (engine->i915); + tasklet |=USES_PV_SUBMISSION (engine->i915);
We should move this to an engine->flag.
Quoted 9 lines
drivers/gpu/drm/i915/i915_irq.c +6
-0}if (tasklet) +static void vgpu_pv_set_default_submission(struct intel_engine_cs *engine) +{ +/* +* We inherit a bunch of functions from execlists that we'd like +* to keep using: +*
drivers/gpu/drm/i915/i915_irq.c gen8_cs_irq_handler(struct intel_engine_cs *engine, u32 iir)
Quoted 10 lines
drivers/gpu/drm/i915/i915_irq.c +10
-0+* engine->submit_request = execlists_submit_request; +* engine->cancel_requests = execlists_cancel_requests; +* engine->schedule = execlists_schedule; +* +* But we need to override the actual submission backend in order +* to talk to the GVT with PV notification message. +*/ +intel_execlists_set_default_submission (engine); + + engine->execlists.tasklet.func = vgpu_pv_submission_tasklet;
You need to pin the breadcrumbs irq, or it will not fire on every request.
I'd push for this to live in intel_pv_submission.c or intel_vgpu_submission.c
Quoted 17 lines
drivers/gpu/drm/i915/i915_irq.c +6
-0@@ -401,6 +551,12 @@ void intel_vgpu_config_pv_caps(struct drm_i915_private *dev_priv, ppgtt->vm.insert_entries = gen8_ppgtt_insert_4lvl_pv; ppgtt->vm.clear_range = gen8_ppgtt_clear_4lvl_pv; } + +if (cap ==PV_SUBMISSION ) { + engine = (struct intel_engine_cs *)data; + engine->set_default_submission = vgpu_pv_set_default_submission; + engine->set_default_submission (engine); + } }drivers/gpu/drm/i915/intel_lrc.c +0
-0
drivers/gpu/drm/i915/intel_lrc.c logical_ring_default_vfuncs(struct intel_engine_cs *engine) · line 2357
Quoted 6 lines
drivers/gpu/drm/i915/intel_lrc.c +3
-0@@ -2352,6 +2352,9 @@ logical_ring_default_vfuncs(struct intel_engine_cs *engine)*/ } engine->emit_bb_start = gen8_emit_bb_start; + +if (intel_vgpu_active (engine->i915)) +intel_vgpu_config_pv_caps (engine->i915,PV_SUBMISSION , engine);
That pair is ugly. Should clean up the engine initialisation so that it doesn't involve placing a chunk of code in a foreign class. -Chris
_______________________________________________
Intel-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/intel-gfx
Re: [PATCH v4 4/8] drm/i915: vgpu context submission pv optimization
Quoted 11 lines · Chris Wilson
Quoting Xiaolin Zhang (2019-03-29 13:32:40)
Quoted 10 lines
drivers/gpu/drm/i915/i915_irq.c +1
-0@@ -37,6 +37,7 @@#include "i915_drv.h" #include "i915_trace.h" #include "intel_drv.h" +#include "i915_vgpu.h"
Quoted 8 lines
drivers/gpu/drm/i915/i915_irq.c gen8_cs_irq_handler(struct intel_engine_cs *engine, u32 iir) · line 1474
Quoted 7 lines
drivers/gpu/drm/i915/i915_irq.c +1
-0/** * DOC: interrupt handling @@ -1470,6 +1471,7 @@ gen8_cs_irq_handler(struct intel_engine_cs *engine, u32 iir)if (iir >_RENDER_USER_INTERRUPT ) {intel_engine_breadcrumbs_irq (engine); tasklet |=USES_GUC_SUBMISSION (engine->i915); + tasklet |=USES_PV_SUBMISSION (engine->i915);We should move this to an
engine->flag.
sure, I will remove this next version.
Quoted 14 lines
Quoted 14 lines
drivers/gpu/drm/i915/i915_irq.c +10
-0}if (tasklet) +static void vgpu_pv_set_default_submission(struct intel_engine_cs *engine) +{ +/* +* We inherit a bunch of functions from execlists that we'd like +* to keep using: +* +* engine->submit_request = execlists_submit_request; +* engine->cancel_requests = execlists_cancel_requests; +* engine->schedule = execlists_schedule; +*
Quoted 8 lines
drivers/gpu/drm/i915/i915_irq.c
Quoted 6 lines
drivers/gpu/drm/i915/i915_irq.c +6
-0+* But we need to override the actual submission backend in order +* to talk to the GVT with PV notification message. +*/ +intel_execlists_set_default_submission (engine); + + engine->execlists.tasklet.func = vgpu_pv_submission_tasklet;You need to pin the breadcrumbs irq, or it will not fire on every request.
indeed, I should do. thanks your great point.
Quoted 2 lines
I'd push for this to live in
intel_pv_submission.corintel_vgpu_submission.c
this will create new file and make a change in Makefile. if this kind of change is OK,
I will create one to use intel_pv_submission.c name.
Quoted 17 lines
Quoted 17 lines
drivers/gpu/drm/i915/i915_irq.c +6
-0@@ -401,6 +551,12 @@ void intel_vgpu_config_pv_caps(struct drm_i915_private *dev_priv, ppgtt->vm.insert_entries = gen8_ppgtt_insert_4lvl_pv; ppgtt->vm.clear_range = gen8_ppgtt_clear_4lvl_pv; } + +if (cap ==PV_SUBMISSION ) { + engine = (struct intel_engine_cs *)data; + engine->set_default_submission = vgpu_pv_set_default_submission; + engine->set_default_submission (engine); + } }drivers/gpu/drm/i915/intel_lrc.c +0
-0@@ -2352,6 +2352,9 @@ logical_ring_default_vfuncs(struct intel_engine_cs *engine)*/
Quoted 9 lines
drivers/gpu/drm/i915/intel_lrc.c logical_ring_default_vfuncs(struct intel_engine_cs *engine) · line 2357
Quoted 5 lines
drivers/gpu/drm/i915/intel_lrc.c +3
-0} engine->emit_bb_start = gen8_emit_bb_start; + +if (intel_vgpu_active (engine->i915)) +intel_vgpu_config_pv_caps (engine->i915,PV_SUBMISSION , engine);That pair is ugly. Should clean up the engine initialisation so that it doesn't involve placing a chunk of code in a foreign class. -Chris
Yep, will move it to other location.
_______________________________________________
Intel-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/intel-gfx
Re: [PATCH v4 4/8] drm/i915: vgpu context submission pv optimization
In reply to Xiaolin ZhangQuoting Xiaolin Zhang (2019-03-29 13:32:40)
drivers/gpu/drm/i915/i915_vgpu.c int intel_vgt_balloon(struct drm_i915_private *dev_priv) · line 338
Quoted 7 lines
drivers/gpu/drm/i915/i915_vgpu.c +7
-0+spin_lock (&engine->i915->vgpu.shared_page_lock); + shared_page->ring_id = engine->id; +for (n =0 ; n <execlists_num_ports (execlists); n++) + shared_page->descs[n] = descs[n]; + +__raw_i915_write32 (uncore,vgtif_reg (g2v_notify), +VGT_G2V_PV_SUBMISSION );
There's no serialisation with the consumer here. The shared_page->descs[] will be clobbered by a second engine before they are read by pv. There should be a wait-for-ack here, or separate channels for each engine. In execlists, we avoid this by not dequeuing until after we get a response/ack from HW.
Preemption? I guess preemption was turned off, or else you have to worry about the same engine overwriting the desc[] (if no ack). -Chris
_______________________________________________
Intel-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/intel-gfx
Re: [PATCH v4 4/8] drm/i915: vgpu context submission pv optimization
Quoted 14 lines · Chris Wilson
Quoting Xiaolin Zhang (2019-03-29 13:32:40)
drivers/gpu/drm/i915/i915_vgpu.c int intel_vgt_balloon(struct drm_i915_private *dev_priv) · line 338
Quoted 7 lines
drivers/gpu/drm/i915/i915_vgpu.c +7
-0+spin_lock (&engine->i915->vgpu.shared_page_lock); + shared_page->ring_id = engine->id; +for (n =0 ; n <execlists_num_ports (execlists); n++) + shared_page->descs[n] = descs[n]; + +__raw_i915_write32 (uncore,vgtif_reg (g2v_notify), +VGT_G2V_PV_SUBMISSION );There's no serialisation with the consumer here. The shared_page->descs[] will be clobbered by a second engine before they are read by pv. There should be a wait-for-ack here, or separate channels for each engine. In execlists, we avoid this by not dequeuing until after we get a response/ack from HW.
Quoted 4 lines
Preemption? I guess preemption was turned off, or else you have to worry about the same engine overwriting the desc[] (if no ack). -Chris
Chris, thanks you to point it out my flaw here. will use the per-engine channel for notification. preemption is turned off and no support in current version.
_______________________________________________
Intel-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/intel-gfx
PATCH v4 5/8 drm/i915/gvt: GVTg handle pv_caps PVINFO register
In reply to Xiaolin Zhangimplement pv_caps PVINFO register handler in GVTg to control different level pv optimization within guest.
report VGT_CAPS_PV capability in pvinfo page for guest.
v0: RFC
v1: rebase
v2: rebase
v3: renamed enable_pvmmio to pvmmio_caps which is used for host pv caps.
v4: renamed pvmmio_caps to pv_caps
--- drivers/gpu/drm/i915/gvt/handlers.c | 4 ++++ drivers/gpu/drm/i915/gvt/vgpu.c | 3 +++ 2 files changed, 7 insertions(+)
drivers/gpu/drm/i915/gvt/handlers.c +4 -0
@@ -1167,6 +1167,7 @@ static int pvinfo_mmio_read(struct intel_vgpu *vgpu, unsigned int offset,break ;case 0x78010 :/* vgt_caps */ case 0x7881c : +case _vgtif_reg (pv_caps):break ;default : invalid_read = true;
@@ -1240,6 +1241,9 @@ static int pvinfo_mmio_write(struct intel_vgpu *vgpu, unsigned int offset,case _vgtif_reg (g2v_notify): ret =handle_g2v_notification (vgpu, data);break ; +case _vgtif_reg (pv_caps): +DRM_INFO ("vgpu id=%d pv caps =0x%x\n" , vgpu->id, data); +break ;/* add xhot and yhot to handled list to avoid error log */ case _vgtif_reg (cursor_x_hot):case _vgtif_reg (cursor_y_hot):
drivers/gpu/drm/i915/gvt/vgpu.c +3 -0
@@ -47,6 +47,7 @@ void populate_pvinfo_page(struct intel_vgpu *vgpu)vgpu_vreg_t (vgpu,vgtif_reg (vgt_caps)) =VGT_CAPS_FULL_PPGTT ;vgpu_vreg_t (vgpu,vgtif_reg (vgt_caps)) |=VGT_CAPS_HWSP_EMULATION ;vgpu_vreg_t (vgpu,vgtif_reg (vgt_caps)) |=VGT_CAPS_HUGE_GTT ; +vgpu_vreg_t (vgpu,vgtif_reg (vgt_caps)) |=VGT_CAPS_PV ;vgpu_vreg_t (vgpu,vgtif_reg (avail_rs.mappable_gmadr.base)) =vgpu_aperture_gmadr_base (vgpu);
@@ -531,6 +532,7 @@ void intel_gvt_reset_vgpu_locked(struct intel_vgpu *vgpu, bool dmlr,struct intel_gvt *gvt = vgpu->gvt;struct intel_gvt_workload_scheduler *scheduler = &gvt->scheduler;unsigned int resetting_eng = dmlr ?ALL_ENGINES : engine_mask; +int pv_caps =vgpu_vreg_t (vgpu,vgtif_reg (pv_caps));gvt_dbg_core ("------------------------------------------\n" ); gvt_dbg_core("resseting vgpu%d, dmlr %d, engine_mask %08x\n",
@@ -562,6 +564,7 @@ void intel_gvt_reset_vgpu_locked(struct intel_vgpu *vgpu, bool dmlr,intel_vgpu_reset_mmio (vgpu, dmlr);populate_pvinfo_page (vgpu); +vgpu_vreg_t (vgpu,vgtif_reg (pv_caps)) = pv_caps;intel_vgpu_reset_display (vgpu);if (dmlr) {
Signature · 6 lines
2.7.4 _______________________________________________ Intel-gfx mailing list [email protected] https://lists.freedesktop.org/mailman/listinfo/intel-gfx
PATCH v4 6/8 drm/i915/gvt: GVTg handle shared_page setup
In reply to Xiaolin ZhangGVTg implemented shared_page setup operation and read_shared_page functionality based on hypervisor_read_gpa().
the shared_page_gpa was passed from guest driver through PVINFO shared_page_gpa register.
v0: RFC
v1: rebase
v2: rebase
v3: added shared_page_gpa check and if read_gpa failure, return zero memory and handle VGT_G2V_SHARED_PAGE_SETUP g2v notification
v4: rebase
--- drivers/gpu/drm/i915/gvt/gvt.h | 6 +++++- drivers/gpu/drm/i915/gvt/handlers.c | 15 +++++++++++++++ drivers/gpu/drm/i915/gvt/vgpu.c | 24 ++++++++++++++++++++++++ 3 files changed, 44 insertions(+), 1 deletion(-)
drivers/gpu/drm/i915/gvt/gvt.h +5 -1
@@ -49,6 +49,7 @@#include "fb_decoder.h" #include "dmabuf.h" #include "page_track.h" +#include "i915_vgpu.h" #define GVT_MAX_VGPU 8
@@ -229,6 +230,8 @@ struct intel_vgpu {struct completion vblank_done;u32 scan_nonprivbb; +u64 shared_page_gpa; +bool shared_page_enabled; };/* validating GM healthy status*/
@@ -686,7 +689,8 @@ int intel_gvt_debugfs_add_vgpu(struct intel_vgpu *vgpu);void intel_gvt_debugfs_remove_vgpu (struct intel_vgpu *vgpu);int intel_gvt_debugfs_init (struct intel_gvt *gvt);void intel_gvt_debugfs_clean (struct intel_gvt *gvt); - +int intel_gvt_read_shared_page (struct intel_vgpu *vgpu, +unsigned int offset,void *buf,unsigned long len);#include "trace.h" #include "mpt.h"
drivers/gpu/drm/i915/gvt/handlers.c +15 -0
@@ -1168,6 +1168,8 @@ static int pvinfo_mmio_read(struct intel_vgpu *vgpu, unsigned int offset,case 0x78010 :/* vgt_caps */ case 0x7881c :case _vgtif_reg (pv_caps): +case _vgtif_reg (shared_page_gpa): +case _vgtif_reg (shared_page_gpa) +4 :break ;default : invalid_read = true;
@@ -1185,6 +1187,7 @@ static int handle_g2v_notification(struct intel_vgpu *vgpu, int notification)intel_gvt_gtt_type_t root_entry_type =GTT_TYPE_PPGTT_ROOT_L4_ENTRY ;struct intel_vgpu_mm *mm;u64 *pdps; +unsigned long gpa, gfn; pdps = (u64 *)&vgpu_vreg64_t (vgpu,vgtif_reg (pdp[0 ]));
@@ -1198,6 +1201,16 @@ static int handle_g2v_notification(struct intel_vgpu *vgpu, int notification)case VGT_G2V_PPGTT_L3_PAGE_TABLE_DESTROY :case VGT_G2V_PPGTT_L4_PAGE_TABLE_DESTROY :return intel_vgpu_put_ppgtt_mm (vgpu, pdps); +case VGT_G2V_SHARED_PAGE_SETUP : + gpa =vgpu_vreg64_t (vgpu,vgtif_reg (shared_page_gpa)); + gfn = gpa >>PAGE_SHIFT ; +if (!intel_gvt_hypervisor_is_valid_gfn (vgpu, gfn)) { +vgpu_vreg_t (vgpu,vgtif_reg (pv_caps)) =0 ; +return 0 ; + } + vgpu->shared_page_gpa = gpa; + vgpu->shared_page_enabled = true; +break ;case VGT_G2V_EXECLIST_CONTEXT_CREATE :case VGT_G2V_EXECLIST_CONTEXT_DESTROY :case 1 :/* Remove this in guest driver. */
@@ -1257,6 +1270,8 @@ static int pvinfo_mmio_write(struct intel_vgpu *vgpu, unsigned int offset,case _vgtif_reg (pdp[3 ].hi):case _vgtif_reg (execlist_context_descriptor_lo):case _vgtif_reg (execlist_context_descriptor_hi): +case _vgtif_reg (shared_page_gpa): +case _vgtif_reg (shared_page_gpa) +4 :break ; case _vgtif_reg(rsv5[0])..._vgtif_reg(rsv5[3]):enter_failsafe_mode (vgpu,GVT_FAILSAFE_INSUFFICIENT_RESOURCE );
drivers/gpu/drm/i915/gvt/vgpu.c +24 -0
@@ -63,6 +63,8 @@ void populate_pvinfo_page(struct intel_vgpu *vgpu)vgpu_vreg_t (vgpu,vgtif_reg (cursor_x_hot)) =UINT_MAX ;vgpu_vreg_t (vgpu,vgtif_reg (cursor_y_hot)) =UINT_MAX ; +vgpu_vreg64_t (vgpu,vgtif_reg (shared_page_gpa)) =0 ; +gvt_dbg_core ("Populate PVINFO PAGE for vGPU %d\n" , vgpu->id);gvt_dbg_core ("aperture base [GMADR] 0x%llx size 0x%llx\n" ,vgpu_aperture_gmadr_base (vgpu),vgpu_aperture_sz (vgpu));
@@ -593,3 +595,25 @@ void intel_gvt_reset_vgpu(struct intel_vgpu *vgpu)intel_gvt_reset_vgpu_locked (vgpu, true,0 );mutex_unlock (&vgpu->vgpu_lock); } + +/** +* intel_gvt_read_shared_page - read content from shared page +*/ +int intel_gvt_read_shared_page (struct intel_vgpu *vgpu, +unsigned int offset,void *buf,unsigned long len) +{ +int ret = -EINVAL ; +unsigned long gpa; + +if (offset >=sizeof (struct gvt_shared_page )) + goto err; + + gpa = vgpu->shared_page_gpa + offset; + ret =intel_gvt_hypervisor_read_gpa (vgpu, gpa, buf, len); +if (!ret) +return ret; +err: +gvt_vgpu_err ("read shared page (offset %x) failed" , offset); +memset (buf,0 , len); +return ret; +}
Signature · 6 lines
2.7.4 _______________________________________________ Intel-gfx mailing list [email protected] https://lists.freedesktop.org/mailman/listinfo/intel-gfx
PATCH v4 7/8 drm/i915/gvt: GVTg support ppgtt pv optimization
In reply to Xiaolin ZhangThis patch handles ppgtt update from g2v notification.
It read out ppgtt pte entries from guest pte tables page and convert them to host pfns.
It creates local ppgtt tables and insert the content pages into the local ppgtt tables directly, which does not track the usage of guest page table and removes the cost of write protection from the original shadow page mechansim.
v0: RFC
v1: rebase
v2: rebase
v3: report pv pggtt cap to guest
v4: renamed VGPU_PVMMIO with VGPU_PVCAP for name consistance, no PV support if gfx vtd enabled.
--- drivers/gpu/drm/i915/gvt/gtt.c | 317 ++++++++++++++++++++++++++++++++++++ drivers/gpu/drm/i915/gvt/gtt.h | 9 + drivers/gpu/drm/i915/gvt/gvt.h | 4 + drivers/gpu/drm/i915/gvt/handlers.c | 12 +- drivers/gpu/drm/i915/gvt/vgpu.c | 3 + 5 files changed, 344 insertions(+), 1 deletion(-)
drivers/gpu/drm/i915/gvt/gtt.c +317 -0
@@ -1744,6 +1744,25 @@ static int ppgtt_handle_guest_write_page_table_bytes(return 0 ; } +static void invalidate_mm_pv (struct intel_vgpu_mm *mm) +{ +struct intel_vgpu *vgpu = mm->vgpu; +struct intel_gvt *gvt = vgpu->gvt; +struct intel_gvt_gtt *gtt = &gvt->gtt; +struct intel_gvt_gtt_pte_ops *ops = gtt->pte_ops; +struct intel_gvt_gtt_entry se; + +i915_ppgtt_put (mm->ppgtt); + +ppgtt_get_shadow_root_entry (mm, &se,0 ); +if (!ops->test_present (&se)) +return ; + se.val64 =0 ; +ppgtt_set_shadow_root_entry (mm, &se,0 ); + + mm->ppgtt_mm.shadowed = false; +} +static void invalidate_ppgtt_mm (struct intel_vgpu_mm *mm) {struct intel_vgpu *vgpu = mm->vgpu;
@@ -1756,6 +1775,11 @@ static void invalidate_ppgtt_mm(struct intel_vgpu_mm *mm)if (!mm->ppgtt_mm.shadowed)return ; +if (VGPU_PVCAP (mm->vgpu,PV_PPGTT_UPDATE )) { +invalidate_mm_pv (mm); +return ; + } +for (index =0 ; index <ARRAY_SIZE (mm->ppgtt_mm.shadow_pdps); index++) {ppgtt_get_shadow_root_entry (mm, &se, index);
@@ -1773,6 +1797,26 @@ static void invalidate_ppgtt_mm(struct intel_vgpu_mm *mm) mm->ppgtt_mm.shadowed = false; } +static int shadow_mm_pv (struct intel_vgpu_mm *mm) +{ +struct intel_vgpu *vgpu = mm->vgpu; +struct intel_gvt *gvt = vgpu->gvt; +struct intel_gvt_gtt_entry se; + + mm->ppgtt =i915_ppgtt_create (gvt->dev_priv); +if (IS_ERR (mm->ppgtt)) { +gvt_vgpu_err ("fail to create ppgtt for pdp 0x%llx\n" , +px_dma (&mm->ppgtt->pml4)); +return PTR_ERR (mm->ppgtt); + } + + se.type =GTT_TYPE_PPGTT_ROOT_L4_ENTRY ; + se.val64 =px_dma (&mm->ppgtt->pml4); +ppgtt_set_shadow_root_entry (mm, &se,0 ); + mm->ppgtt_mm.shadowed = true; + +return 0 ; +}static int shadow_ppgtt_mm (struct intel_vgpu_mm *mm) {
@@ -1787,6 +1831,9 @@ static int shadow_ppgtt_mm(struct intel_vgpu_mm *mm)if (mm->ppgtt_mm.shadowed)return 0 ; +if (VGPU_PVCAP (mm->vgpu,PV_PPGTT_UPDATE )) +return shadow_mm_pv (mm); + mm->ppgtt_mm.shadowed = true;for (index =0 ; index <ARRAY_SIZE (mm->ppgtt_mm.guest_pdps); index++) {
@@ -2787,3 +2834,273 @@ void intel_vgpu_reset_gtt(struct intel_vgpu *vgpu)intel_vgpu_destroy_all_ppgtt_mm (vgpu);intel_vgpu_reset_ggtt (vgpu, true); } + +int intel_vgpu_g2v_pv_ppgtt_alloc_4lvl (struct intel_vgpu *vgpu, +u64 pdps[]) +{ +struct intel_vgpu_mm *mm; +int ret =0 ; +u32 offset; +struct pv_ppgtt_update pv_ppgtt; + + offset = offsetof(struct gvt_shared_page , pv_ppgtt); +intel_gvt_read_shared_page (vgpu, offset, &pv_ppgtt,sizeof (pv_ppgtt)); + + mm =intel_vgpu_find_ppgtt_mm (vgpu, &pv_ppgtt.pdp); +if (!mm) { +gvt_vgpu_err ("failed to find pdp 0x%llx\n" , pv_ppgtt.pdp); + ret = -EINVAL ; + }else { + ret = mm->ppgtt->vm.allocate_va_range (&mm->ppgtt->vm, + pv_ppgtt.start, pv_ppgtt.length); +if (ret) +gvt_vgpu_err ("failed to alloc %llx\n" , pv_ppgtt.pdp); + } + +return ret; +} + +int intel_vgpu_g2v_pv_ppgtt_clear_4lvl (struct intel_vgpu *vgpu, +u64 pdps[]) +{ +struct intel_vgpu_mm *mm; +int ret =0 ; +u32 offset; +struct pv_ppgtt_update pv_ppgtt; + + offset = offsetof(struct gvt_shared_page , pv_ppgtt); +intel_gvt_read_shared_page (vgpu, offset, &pv_ppgtt,sizeof (pv_ppgtt)); + mm =intel_vgpu_find_ppgtt_mm (vgpu, &pv_ppgtt.pdp); +if (!mm) { +gvt_vgpu_err ("failed to find pdp 0x%llx\n" , pv_ppgtt.pdp); + ret = -EINVAL ; + }else { + mm->ppgtt->vm.clear_range (&mm->ppgtt->vm, + pv_ppgtt.start, pv_ppgtt.length); + } + +return ret; +} + +#define GEN8_PML4E_SIZE (1UL << GEN8_PML4E_SHIFT) +#define GEN8_PML4E_SIZE_MASK (~(GEN8_PML4E_SIZE - 1)) +#define GEN8_PDPE_SIZE (1UL << GEN8_PDPE_SHIFT) +#define GEN8_PDPE_SIZE_MASK (~(GEN8_PDPE_SIZE - 1)) +#define GEN8_PDE_SIZE (1UL << GEN8_PDE_SHIFT) +#define GEN8_PDE_SIZE_MASK (~(GEN8_PDE_SIZE - 1)) + +#define pml4_addr_end (addr, end) \ +({ unsigned long __boundary = \ + ((addr) + GEN8_PML4E_SIZE) & GEN8_PML4E_SIZE_MASK; \ + (__boundary < (end)) ? __boundary : (end); \ +}) + +#define pdp_addr_end (addr, end) \ +({ unsigned long __boundary = \ + ((addr) + GEN8_PDPE_SIZE) & GEN8_PDPE_SIZE_MASK; \ + (__boundary < (end)) ? __boundary : (end); \ +}) + +#define pd_addr_end (addr, end) \ +({ unsigned long __boundary = \ + ((addr) + GEN8_PDE_SIZE) & GEN8_PDE_SIZE_MASK; \ + (__boundary < (end)) ? __boundary : (end); \ +}) + +struct ppgtt_walk { +unsigned long *mfns; +int mfn_index; +unsigned long *pt; +}; + +static int walk_pt_range (struct intel_vgpu *vgpu,u64 pt, +u64 start,u64 end,struct ppgtt_walk *walk) +{ +const struct intel_gvt_device_info *info = &vgpu->gvt->device_info; +struct intel_gvt_gtt_gma_ops *gma_ops = vgpu->gvt->gtt.gma_ops; +unsigned long start_index, end_index; +int ret; +int i; +unsigned long mfn, gfn; + + start_index = gma_ops->gma_to_pte_index (start); + end_index = ((end - start) >>PAGE_SHIFT ) + start_index; + + ret =intel_gvt_hypervisor_read_gpa (vgpu, + (pt &PAGE_MASK ) + (start_index << info->gtt_entry_size_shift), + walk->pt + start_index, + (end_index - start_index) << info->gtt_entry_size_shift); +if (ret) { +gvt_vgpu_err ("fail to read gpa %llx\n" , pt); +return ret; + } + +for (i = start_index; i < end_index; i++) { + gfn = walk->pt[i] >>PAGE_SHIFT ; + mfn =intel_gvt_hypervisor_gfn_to_mfn (vgpu, gfn); +if (mfn ==INTEL_GVT_INVALID_ADDR ) { +gvt_vgpu_err ("fail to translate gfn: 0x%lx\n" , gfn); +return -ENXIO ; + } + walk->mfns[walk->mfn_index++] = mfn <<PAGE_SHIFT ; + } + +return 0 ; +} + + +static int walk_pd_range (struct intel_vgpu *vgpu,u64 pd, +u64 start,u64 end,struct ppgtt_walk *walk) +{ +const struct intel_gvt_device_info *info = &vgpu->gvt->device_info; +struct intel_gvt_gtt_gma_ops *gma_ops = vgpu->gvt->gtt.gma_ops; +unsigned long index; +u64 pt, next; +int ret =0 ; + +do { + index = gma_ops->gma_to_pde_index (start); + + ret =intel_gvt_hypervisor_read_gpa (vgpu, + (pd &PAGE_MASK ) + (index << + info->gtt_entry_size_shift), &pt,8 ); +if (ret) +return ret; + next =pd_addr_end (start, end); +walk_pt_range (vgpu, pt, start, next, walk); + + start = next; + }while (start != end); + +return ret; +} + + +static int walk_pdp_range (struct intel_vgpu *vgpu,u64 pdp, +u64 start,u64 end,struct ppgtt_walk *walk) +{ +const struct intel_gvt_device_info *info = &vgpu->gvt->device_info; +struct intel_gvt_gtt_gma_ops *gma_ops = vgpu->gvt->gtt.gma_ops; +unsigned long index; +u64 pd, next; +int ret =0 ; + +do { + index = gma_ops->gma_to_l4_pdp_index (start); + + ret =intel_gvt_hypervisor_read_gpa (vgpu, + (pdp &PAGE_MASK ) + (index << + info->gtt_entry_size_shift), &pd,8 ); +if (ret) +return ret; + next =pdp_addr_end (start, end); +walk_pd_range (vgpu, pd, start, next, walk); + start = next; + }while (start != end); + +return ret; +} + + +static int walk_pml4_range (struct intel_vgpu *vgpu,u64 pml4, +u64 start,u64 end,struct ppgtt_walk *walk) +{ +const struct intel_gvt_device_info *info = &vgpu->gvt->device_info; +struct intel_gvt_gtt_gma_ops *gma_ops = vgpu->gvt->gtt.gma_ops; +unsigned long index; +u64 pdp, next; +int ret =0 ; + +do { + index = gma_ops->gma_to_pml4_index (start); + ret =intel_gvt_hypervisor_read_gpa (vgpu, + (pml4 &PAGE_MASK ) + (index << + info->gtt_entry_size_shift), &pdp,8 ); +if (ret) +return ret; + next =pml4_addr_end (start, end); +walk_pdp_range (vgpu, pdp, start, next, walk); + start = next; + }while (start != end); + +return ret; +} + +int intel_vgpu_g2v_pv_ppgtt_insert_4lvl (struct intel_vgpu *vgpu, +u64 pdps[]) +{ +struct intel_vgpu_mm *mm; +u64 pml4, start, length; +u32 cache_level; +int ret =0 ; +struct sg_table st; +struct scatterlist *sg =NULL ; +int num_pages; +struct i915_vma vma; +struct ppgtt_walk walk; +int i; +u32 offset; +struct pv_ppgtt_update pv_ppgtt; + + offset = offsetof(struct gvt_shared_page , pv_ppgtt); +intel_gvt_read_shared_page (vgpu, offset, &pv_ppgtt,sizeof (pv_ppgtt)); + pml4 = pv_ppgtt.pdp; + start = pv_ppgtt.start; + length = pv_ppgtt.length; + cache_level = pv_ppgtt.cache_level; + num_pages = length >>PAGE_SHIFT ; + + mm =intel_vgpu_find_ppgtt_mm (vgpu, &pml4); +if (!mm) { +gvt_vgpu_err ("fail to find mm for pml4 0x%llx\n" , pml4); +return -EINVAL ; + } + + walk.mfn_index =0 ; + walk.mfns =NULL ; + walk.pt =NULL ; + + walk.mfns =kmalloc_array (num_pages, +sizeof (unsigned long ),GFP_KERNEL ); +if (!walk.mfns) { + ret = -ENOMEM ; + goto fail; + } + + walk.pt = (unsigned long *)__get_free_pages (GFP_KERNEL ,0 ); +if (!walk.pt) { + ret = -ENOMEM ; + goto fail; + } + +if (sg_alloc_table (&st, num_pages,GFP_KERNEL )) { + ret = -ENOMEM ; + goto fail; + } + + ret =walk_pml4_range (vgpu, pml4, start, start + length, &walk); +if (ret) + goto fail_free_sg; + +WARN_ON (num_pages != walk.mfn_index); + +for_each_sg (st.sgl, sg, num_pages, i) { + sg->offset =0 ; + sg->length =PAGE_SIZE ; +sg_dma_address (sg) = walk.mfns[i]; +sg_dma_len (sg) =PAGE_SIZE ; + } + +memset (&vma,0 ,sizeof (vma)); + vma.node.start = start; + vma.pages = &st; + mm->ppgtt->vm.insert_entries (&mm->ppgtt->vm, &vma, cache_level,0 ); + +fail_free_sg: +sg_free_table (&st); +fail: +kfree (walk.mfns); +free_page ((unsigned long )walk.pt); + +return ret; +}
drivers/gpu/drm/i915/gvt/gtt.h +9 -0
@@ -141,6 +141,7 @@ struct intel_gvt_partial_pte {struct intel_vgpu_mm {enum intel_gvt_mm_type type; +struct i915_hw_ppgtt *ppgtt;struct intel_vgpu *vgpu;struct kref ref;
@@ -277,4 +278,12 @@ int intel_vgpu_emulate_ggtt_mmio_read(struct intel_vgpu *vgpu,int intel_vgpu_emulate_ggtt_mmio_write (struct intel_vgpu *vgpu,unsigned int off,void *p_data,unsigned int bytes); +int intel_vgpu_g2v_pv_ppgtt_alloc_4lvl (struct intel_vgpu *vgpu, +u64 pdps[]); + +int intel_vgpu_g2v_pv_ppgtt_clear_4lvl (struct intel_vgpu *vgpu, +u64 pdps[]); + +int intel_vgpu_g2v_pv_ppgtt_insert_4lvl (struct intel_vgpu *vgpu, +u64 pdps[]);#endif /* _GVT_GTT_H_ */
drivers/gpu/drm/i915/gvt/gvt.h +4 -0
@@ -53,6 +53,10 @@#define GVT_MAX_VGPU 8 +#define VGPU_PVCAP (vgpu, cap) \ + ((vgpu_vreg_t(vgpu, vgtif_reg(pv_caps)) & (cap)) \ + && vgpu->shared_page_enabled) +struct intel_gvt_host {struct device *dev;bool initialized;
drivers/gpu/drm/i915/gvt/handlers.c +11 -1
@@ -1188,6 +1188,7 @@ static int handle_g2v_notification(struct intel_vgpu *vgpu, int notification)struct intel_vgpu_mm *mm;u64 *pdps;unsigned long gpa, gfn; +int ret =0 ; pdps = (u64 *)&vgpu_vreg64_t (vgpu,vgtif_reg (pdp[0 ]));
@@ -1211,6 +1212,15 @@ static int handle_g2v_notification(struct intel_vgpu *vgpu, int notification) vgpu->shared_page_gpa = gpa; vgpu->shared_page_enabled = true;break ; +case VGT_G2V_PPGTT_L4_ALLOC : + ret =intel_vgpu_g2v_pv_ppgtt_alloc_4lvl (vgpu, pdps); +break ; +case VGT_G2V_PPGTT_L4_INSERT : + ret =intel_vgpu_g2v_pv_ppgtt_insert_4lvl (vgpu, pdps); +break ; +case VGT_G2V_PPGTT_L4_CLEAR : + ret =intel_vgpu_g2v_pv_ppgtt_clear_4lvl (vgpu, pdps); +break ;case VGT_G2V_EXECLIST_CONTEXT_CREATE :case VGT_G2V_EXECLIST_CONTEXT_DESTROY :case 1 :/* Remove this in guest driver. */
@@ -1218,7 +1228,7 @@ static int handle_g2v_notification(struct intel_vgpu *vgpu, int notification)default :gvt_vgpu_err ("Invalid PV notification %d\n" , notification); } -return 0 ; +return ret; }static int send_display_ready_uevent (struct intel_vgpu *vgpu,int ready)
drivers/gpu/drm/i915/gvt/vgpu.c +3 -0
@@ -49,6 +49,9 @@ void populate_pvinfo_page(struct intel_vgpu *vgpu)vgpu_vreg_t (vgpu,vgtif_reg (vgt_caps)) |=VGT_CAPS_HUGE_GTT ;vgpu_vreg_t (vgpu,vgtif_reg (vgt_caps)) |=VGT_CAPS_PV ; +if (!intel_vtd_active ()) +vgpu_vreg_t (vgpu,vgtif_reg (pv_caps)) =PV_PPGTT_UPDATE ; +vgpu_vreg_t (vgpu,vgtif_reg (avail_rs.mappable_gmadr.base)) =vgpu_aperture_gmadr_base (vgpu); vgpu_vreg_t(vgpu, vgtif_reg(avail_rs.mappable_gmadr.size)) =
Signature · 6 lines
2.7.4 _______________________________________________ Intel-gfx mailing list [email protected] https://lists.freedesktop.org/mailman/listinfo/intel-gfx
PATCH v4 8/8 drm/i915/gvt: GVTg support context submission pv optimization
In reply to Xiaolin Zhangimplemented context submission pv optimizaiton within GVTg.
GVTg to read context submission data (elsp_data) from the shared_page directly without trap cost and eliminate execlist HW behavior emulation without injecting context switch interrupt to guest under PV submisison mechanism.
v0: RFC
v1: rebase
v2: rebase
v3: report pv context submission cap and handle VGT_G2V_ELSP_SUBMIT g2v pv notification.
v4: eliminate execlist HW emulation and don't inject context switch interrupt to guest under PV submisison mechanism.
--- drivers/gpu/drm/i915/gvt/execlist.c | 6 ++++++ drivers/gpu/drm/i915/gvt/handlers.c | 32 ++++++++++++++++++++++++++++++++ drivers/gpu/drm/i915/gvt/vgpu.c | 1 + 3 files changed, 39 insertions(+)
drivers/gpu/drm/i915/gvt/execlist.c +6 -0
@@ -382,6 +382,9 @@ static int prepare_execlist_workload(struct intel_vgpu_workload *workload)int ring_id = workload->ring_id;int ret; +if (VGPU_PVCAP (vgpu,PV_SUBMISSION )) +return 0 ; +if (!workload->emulate_schedule_in)return 0 ;
@@ -429,6 +432,9 @@ static int complete_execlist_workload(struct intel_vgpu_workload *workload) goto out; } +if (VGPU_PVCAP (vgpu,PV_SUBMISSION )) + goto out; + ret =emulate_execlist_ctx_schedule_out (execlist, &workload->ctx_desc); out:intel_vgpu_unpin_mm (workload->shadow_mm);
drivers/gpu/drm/i915/gvt/handlers.c +32 -0
@@ -1182,6 +1182,35 @@ static int pvinfo_mmio_read(struct intel_vgpu *vgpu, unsigned int offset,return 0 ; } +static int intel_vgpu_g2v_pv_elsp_submit (struct intel_vgpu *vgpu) +{ +struct intel_vgpu_execlist *execlist; +u32 ring_id_off; +int ring_id; +u32 descs_off; + +int ret = -EINVAL ; + +if (!VGPU_PVCAP (vgpu,PV_SUBMISSION )) +return ret; + + ring_id_off = offsetof(struct gvt_shared_page , ring_id); +if (intel_gvt_read_shared_page (vgpu, ring_id_off, &ring_id,4 )) +return ret; + +if (WARN_ON (ring_id <0 || ring_id >=I915_NUM_ENGINES )) +return ret; + + execlist = &vgpu->submission.execlist[ring_id]; + + descs_off = offsetof(struct gvt_shared_page , descs); +if (intel_gvt_read_shared_page (vgpu, descs_off, + &execlist->elsp_dwords.data,8 *EXECLIST_MAX_PORTS )) +return ret; + +return intel_vgpu_submit_execlist (vgpu, ring_id); +} +static int handle_g2v_notification (struct intel_vgpu *vgpu,int notification) {intel_gvt_gtt_type_t root_entry_type =GTT_TYPE_PPGTT_ROOT_L4_ENTRY ;
@@ -1221,6 +1250,9 @@ static int handle_g2v_notification(struct intel_vgpu *vgpu, int notification)case VGT_G2V_PPGTT_L4_CLEAR : ret =intel_vgpu_g2v_pv_ppgtt_clear_4lvl (vgpu, pdps);break ; +case VGT_G2V_PV_SUBMISSION : + ret =intel_vgpu_g2v_pv_elsp_submit (vgpu); +break ;case VGT_G2V_EXECLIST_CONTEXT_CREATE :case VGT_G2V_EXECLIST_CONTEXT_DESTROY :case 1 :/* Remove this in guest driver. */
drivers/gpu/drm/i915/gvt/vgpu.c +1 -0
@@ -51,6 +51,7 @@ void populate_pvinfo_page(struct intel_vgpu *vgpu)if (!intel_vtd_active ())vgpu_vreg_t (vgpu,vgtif_reg (pv_caps)) =PV_PPGTT_UPDATE ; +vgpu_vreg_t (vgpu,vgtif_reg (pv_caps)) |=PV_SUBMISSION ;vgpu_vreg_t (vgpu,vgtif_reg (avail_rs.mappable_gmadr.base)) =vgpu_aperture_gmadr_base (vgpu);
Signature · 6 lines
2.7.4 _______________________________________________ Intel-gfx mailing list [email protected] https://lists.freedesktop.org/mailman/listinfo/intel-gfx
Quoting Xiaolin Zhang (2019-03-29 13:32:36)
Quoted 3 lines
To improve vgpu performance, it could implement some PV optimization such as to reduce the mmio access trap numbers or eliminate certain piece of HW emulation within guest driver to reduce vm exit/vm enter cost.
Where's the CI for this patchset? The lack of interrupts to drive submission should have shown up, and if not, we need some testcases to make sure it doesn't happen again. Everytime I see a gvt patch, I ask if we can get some coverage in intel-gfx-ci :)
-Chris
_______________________________________________
Intel-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/intel-gfx
Quoted 10 lines · Chris Wilson
Quoting Xiaolin Zhang (2019-03-29 13:32:36)
Quoted 3 lines
To improve vgpu performance, it could implement some PV optimization such as to reduce the mmio access trap numbers or eliminate certain piece of HW emulation within guest driver to reduce vm exit/vm enter cost.
Where's the CI for this patchset? The lack of interrupts to drive submission should have shown up, and if not, we need some testcases to make sure it doesn't happen again. Everytime I see a gvt patch, I ask if we can get some coverage in intel-gfx-ci :)
-Chris
The CI for this patchset was not generated due to build failure. will try next version to avoid build issue by rebasing to tip of branch.
_______________________________________________
Intel-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/intel-gfx