Skip to contentSkip to conversation

ocxl: Mmio invalidation support

PATCH Superseded Christophe Lombard 4 participants 13 messages Nov 20, 2020 – Nov 24, 2020

Subsystems (4)

v2 of 4 Superseded Latest is v4 →

lore
Share & tools
lore
#0 lore raw

v2 cover ocxl: Mmio invalidation support

OpenCAPI 4.0/5.0 with TLBI/SLBI Snooping, is not used due to performance problems caused by the PAU having to process all incoming TLBI/SLBI commands which will cause them to back up on the PowerBus.

When the Address Translation Mode requires TLB operations to be initiated using MMIO registers, a set of registers like the following is used:
• XTS MMIO ATSD0 LPARID register
• XTS MMIO ATSD0 AVA register
• XTS MMIO ATSD0 launch register, write access initiates a shoot down
• XTS MMIO ATSD0 status register

The MMIO based mechanism also blocks the NPU/PAU from snooping TLBIE commands from the PowerBus.

The Shootdown commands (ATSD) will be generated using MMIO registers in the NPU/PAU and sent to the device.

---
Changelog[v2]
- Rebase to latest upstream.
- Create a set of smaller patches
- Move the device tree parsing and ioremap() for the shootdown page in a platform-specific file (powernv)
- Release the shootdown page in release_xsl()
- Initialize atsd_lock
- Move the code to initiate the TLB Invalidate command in a platform-specific file (powernv)
- Use the notifier invalidate_range
---

Christophe Lombard (5):
ocxl: Assign a register set to a Logical Partition
ocxl: Initiate a TLB invalidate command
ocxl: Update the Process Element Entry
ocxl: Add mmu notifier
ocxl: Add new kernel traces

 arch/powerpc/include/asm/pnv-ocxl.h   |  53 +++++++++++++
 arch/powerpc/platforms/powernv/ocxl.c | 103 ++++++++++++++++++++++++++
 drivers/misc/ocxl/context.c           |   4 +-
 drivers/misc/ocxl/link.c              |  66 ++++++++++++++++-
 drivers/misc/ocxl/ocxl_internal.h     |   4 +-
 drivers/misc/ocxl/trace.h             |  64 ++++++++++++++++
 drivers/scsi/cxlflash/ocxl_hw.c       |   6 +-
 include/misc/ocxl.h                   |   2 +-
 8 files changed, 295 insertions(+), 7 deletions(-)
Signature · 1 line
2.28.0
#1 lore raw

PATCH v2 1/5 ocxl: Assign a register set to a Logical Partition

Platform specific function to assign a register set to a Logical Partition. The "ibm,mmio-atsd" property, provided by the firmware, contains the 16 base ATSD physical addresses (ATSD0 through ATSD15) of the set of MMIO registers (XTS MMIO ATSDx LPARID/AVA/launch/status register).

For the time being, the ATSD0 set of registers is used by default.

---
 arch/powerpc/include/asm/pnv-ocxl.h   |  3 ++
 arch/powerpc/platforms/powernv/ocxl.c | 48 +++++++++++++++++++++++++++
 2 files changed, 51 insertions(+)
arch/powerpc/include/asm/pnv-ocxl.h +3 -0
index d37ededca3ee..3f38aed7100c 100644
@@ -28,4 +28,7 @@ int pnv_ocxl_spa_setup(struct pci_dev *dev, void *spa_mem, int PE_mask, void **p
 void pnv_ocxl_spa_release(void *platform_data);
 int pnv_ocxl_spa_remove_pe_from_cache(void *platform_data, int pe_handle);
 
+int pnv_ocxl_map_lpar(struct pci_dev *dev, uint64_t lparid,
+		      uint64_t lpcr, void __iomem **arva);
+void pnv_ocxl_unmap_lpar(void __iomem **arva);
 #endif /* _ASM_PNV_OCXL_H */
arch/powerpc/platforms/powernv/ocxl.c +48 -0

51 lines not on this page · raw message

Signature · 1 line
2.28.0
#2 lore raw

Re: [PATCH V2 1/5] ocxl: Assign a register set to a Logical Partition

Quoted 47 lines · Christophe Lombard

Platform specific function to assign a register set to a Logical Partition. The "ibm,mmio-atsd" property, provided by the firmware, contains the 16 base ATSD physical addresses (ATSD0 through ATSD15) of the set of MMIO registers (XTS MMIO ATSDx LPARID/AVA/launch/status register).

For the time being, the ATSD0 set of registers is used by default.

---
  arch/powerpc/include/asm/pnv-ocxl.h   |  3 ++
  arch/powerpc/platforms/powernv/ocxl.c | 48 +++++++++++++++++++++++++++
  2 files changed, 51 insertions(+)
arch/powerpc/include/asm/pnv-ocxl.h +3 -0
index d37ededca3ee..3f38aed7100c 100644
@@ -28,4 +28,7 @@ int pnv_ocxl_spa_setup(struct pci_dev *dev, void *spa_mem, int PE_mask, void **p
  void pnv_ocxl_spa_release(void *platform_data);
  int pnv_ocxl_spa_remove_pe_from_cache(void *platform_data, int pe_handle);
 
+int pnv_ocxl_map_lpar(struct pci_dev *dev, uint64_t lparid,
+		      uint64_t lpcr, void __iomem **arva);
+void pnv_ocxl_unmap_lpar(void __iomem **arva);
  #endif /* _ASM_PNV_OCXL_H */
arch/powerpc/platforms/powernv/ocxl.c +14 -0
index ecdad219d704..bc20cf867900 100644
@@ -483,3 +483,51 @@ int pnv_ocxl_spa_remove_pe_from_cache(void *platform_data, int pe_handle)
  	return rc;
  }
  EXPORT_SYMBOL_GPL(pnv_ocxl_spa_remove_pe_from_cache);
+
+int pnv_ocxl_map_lpar(struct pci_dev *dev, uint64_t lparid,
+		      uint64_t lpcr, void __iomem **arva)
+{
+	struct pci_controller *hose = pci_bus_to_host(dev->bus);
+	struct pnv_phb *phb = hose->private_data;
+	u64 mmio_atsd;
+	int rc;
+
+	/* ATSD physical address.
+	 * ATSD LAUNCH register: write access initiates a shoot down to
+	 * initiate the TLB Invalidate command.
+	 */
+	rc = of_property_read_u64_index(hose->dn, "ibm,mmio-atsd",

arch/powerpc/platforms/powernv/ocxl.c int pnv_ocxl_spa_remove_pe_from_cache(void *platform_data, int pe_handle) · line 509

Quoted 10 lines
arch/powerpc/platforms/powernv/ocxl.c +10 -0
+					0, &mmio_atsd);
+	if (rc) {
+		dev_info(&dev->dev, "No available ATSD found\n");
+		return rc;
+	}
+
+	/* Assign a register set to a Logical Partition and MMIO ATSD
+	 * LPARID register to the required value.
+	 */
+	if (mmio_atsd)

If we don't have the "ibm,mmio-atsd", i.e on P9, then we've already exited above. So why not consider mmio_atsd as an error?

arch/powerpc/platforms/powernv/ocxl.c int pnv_ocxl_spa_remove_pe_from_cache(void *platform_data, int pe_handle) · line 517

Quoted 8 lines
arch/powerpc/platforms/powernv/ocxl.c +8 -0
+		rc = opal_npu_map_lpar(phb->opal_id, pci_dev_id(dev),
+				       lparid, lpcr);
+	if (rc) {
+		dev_err(&dev->dev, "Error mapping device to LPAR: %d\n", rc);
+		return rc;
+	}
+
+	if (mmio_atsd) {

Same here

arch/powerpc/platforms/powernv/ocxl.c int pnv_ocxl_spa_remove_pe_from_cache(void *platform_data, int pe_handle) · line 529

Quoted 12 lines
arch/powerpc/platforms/powernv/ocxl.c +12 -0
+		*arva = ioremap(mmio_atsd, 24);
+		if (!(*arva)) {
+			dev_warn(&dev->dev, "ioremap failed - mmio_atsd: %#llx\n", mmio_atsd);
+			rc = -ENOMEM;
+		}
+	}
+
+	return rc;
+}
+EXPORT_SYMBOL_GPL(pnv_ocxl_map_lpar);
+
+void pnv_ocxl_unmap_lpar(void __iomem **arva)

The arva argument doesn't need to be a double pointer. Void * is enough.

Fred

Quoted 5 lines
arch/powerpc/platforms/powernv/ocxl.c +4 -0
+{
+	iounmap(*arva);
+}
+EXPORT_SYMBOL_GPL(pnv_ocxl_unmap_lpar);
#3 lore raw

PATCH v2 2/5 ocxl: Initiate a TLB invalidate command

In reply to Christophe Lombard

When a TLB Invalidate is required for the Logical Partition, the following sequence has to be performed:

1. Load MMIO ATSD AVA register with the necessary value, if required.
2. Write the MMIO ATSD launch register to initiate the TLB Invalidate command.
3. Poll the MMIO ATSD status register to determine when the TLB Invalidate has been completed.

---
 arch/powerpc/include/asm/pnv-ocxl.h   | 50 ++++++++++++++++++++++++
 arch/powerpc/platforms/powernv/ocxl.c | 55 +++++++++++++++++++++++++++
 2 files changed, 105 insertions(+)
arch/powerpc/include/asm/pnv-ocxl.h +50 -0
index 3f38aed7100c..9c90e87e7659 100644
@@ -3,12 +3,59 @@
 #ifndef _ASM_PNV_OCXL_H
 #define _ASM_PNV_OCXL_H
 
+#include <linux/bitfield.h>
 #include <linux/pci.h>
 
 #define PNV_OCXL_TL_MAX_TEMPLATE        63
 #define PNV_OCXL_TL_BITS_PER_RATE       4
 #define PNV_OCXL_TL_RATE_BUF_SIZE       ((PNV_OCXL_TL_MAX_TEMPLATE+1) * PNV_OCXL_TL_BITS_PER_RATE / 8)
 
+#define PNV_OCXL_ATSD_TIMEOUT		1
+
+/* TLB Management Instructions */
+#define PNV_OCXL_ATSD_LNCH		0x00
+/* Radix Invalidate */
+#define   PNV_OCXL_ATSD_LNCH_R		PPC_BIT(0)
+/* Radix Invalidation Control
+ * 0b00 Just invalidate TLB.
+ * 0b01 Invalidate just Page Walk Cache.
+ * 0b10 Invalidate TLB, Page Walk Cache, and any
+ * caching of Partition and Process Table Entries.
+ */
+#define   PNV_OCXL_ATSD_LNCH_RIC	PPC_BITMASK(1, 2)
+/* Number and Page Size of translations to be invalidated */
+#define   PNV_OCXL_ATSD_LNCH_LP		PPC_BITMASK(3, 10)
+/* Invalidation Criteria
+ * 0b00 Invalidate just the target VA.
+ * 0b01 Invalidate matching PID.
+ */
+#define   PNV_OCXL_ATSD_LNCH_IS		PPC_BITMASK(11, 12)
+/* 0b1: Process Scope, 0b0: Partition Scope */
+#define   PNV_OCXL_ATSD_LNCH_PRS	PPC_BIT(13)
+/* Invalidation Flag */
+#define   PNV_OCXL_ATSD_LNCH_B		PPC_BIT(14)
+/* Actual Page Size to be invalidated
+ * 000 4KB
+ * 101 64KB
+ * 001 2MB
+ * 010 1GB
+ */
+#define   PNV_OCXL_ATSD_LNCH_AP		PPC_BITMASK(15, 17)
+/* Defines the large page select
+ * L=0b0 for 4KB pages
+ * L=0b1 for large pages)
+ */
+#define   PNV_OCXL_ATSD_LNCH_L		PPC_BIT(18)
+/* Process ID */
+#define   PNV_OCXL_ATSD_LNCH_PID	PPC_BITMASK(19, 38)
+/* NoFlush – Assumed to be 0b0 */
+#define   PNV_OCXL_ATSD_LNCH_F		PPC_BIT(39)
+#define   PNV_OCXL_ATSD_LNCH_OCAPI_SLBI	PPC_BIT(40)
+#define   PNV_OCXL_ATSD_LNCH_OCAPI_SINGLETON	PPC_BIT(41)
+#define PNV_OCXL_ATSD_AVA		0x08
+#define   PNV_OCXL_ATSD_AVA_AVA		PPC_BITMASK(0, 51)
+#define PNV_OCXL_ATSD_STAT		0x10
+
 int pnv_ocxl_get_actag(struct pci_dev *dev, u16 *base, u16 *enabled, u16 *supported);
 int pnv_ocxl_get_pasid_count(struct pci_dev *dev, int *count);
 
@@ -31,4 +78,7 @@ int pnv_ocxl_spa_remove_pe_from_cache(void *platform_data, int pe_handle);
 int pnv_ocxl_map_lpar(struct pci_dev *dev, uint64_t lparid,
 		      uint64_t lpcr, void __iomem **arva);
 void pnv_ocxl_unmap_lpar(void __iomem **arva);
+void pnv_ocxl_tlb_invalidate(void __iomem **arva,
+			     unsigned long pid,
+			     unsigned long addr);
 #endif /* _ASM_PNV_OCXL_H */
arch/powerpc/platforms/powernv/ocxl.c +55 -0

58 lines not on this page · raw message

Signature · 1 line
2.28.0
#4 lore raw

Re: [PATCH V2 2/5] ocxl: Initiate a TLB invalidate command

Quoted 92 lines · Christophe Lombard

When a TLB Invalidate is required for the Logical Partition, the following sequence has to be performed:

1. Load MMIO ATSD AVA register with the necessary value, if required.
2. Write the MMIO ATSD launch register to initiate the TLB Invalidate command.
3. Poll the MMIO ATSD status register to determine when the TLB Invalidate has been completed.

---
  arch/powerpc/include/asm/pnv-ocxl.h   | 50 ++++++++++++++++++++++++
  arch/powerpc/platforms/powernv/ocxl.c | 55 +++++++++++++++++++++++++++
  2 files changed, 105 insertions(+)
arch/powerpc/include/asm/pnv-ocxl.h +50 -0
index 3f38aed7100c..9c90e87e7659 100644
@@ -3,12 +3,59 @@
  #ifndef _ASM_PNV_OCXL_H
  #define _ASM_PNV_OCXL_H
 
+#include <linux/bitfield.h>
  #include <linux/pci.h>
 
  #define PNV_OCXL_TL_MAX_TEMPLATE        63
  #define PNV_OCXL_TL_BITS_PER_RATE       4
  #define PNV_OCXL_TL_RATE_BUF_SIZE       ((PNV_OCXL_TL_MAX_TEMPLATE+1) * PNV_OCXL_TL_BITS_PER_RATE / 8)
 
+#define PNV_OCXL_ATSD_TIMEOUT		1
+
+/* TLB Management Instructions */
+#define PNV_OCXL_ATSD_LNCH		0x00
+/* Radix Invalidate */
+#define   PNV_OCXL_ATSD_LNCH_R		PPC_BIT(0)
+/* Radix Invalidation Control
+ * 0b00 Just invalidate TLB.
+ * 0b01 Invalidate just Page Walk Cache.
+ * 0b10 Invalidate TLB, Page Walk Cache, and any
+ * caching of Partition and Process Table Entries.
+ */
+#define   PNV_OCXL_ATSD_LNCH_RIC	PPC_BITMASK(1, 2)
+/* Number and Page Size of translations to be invalidated */
+#define   PNV_OCXL_ATSD_LNCH_LP		PPC_BITMASK(3, 10)
+/* Invalidation Criteria
+ * 0b00 Invalidate just the target VA.
+ * 0b01 Invalidate matching PID.
+ */
+#define   PNV_OCXL_ATSD_LNCH_IS		PPC_BITMASK(11, 12)
+/* 0b1: Process Scope, 0b0: Partition Scope */
+#define   PNV_OCXL_ATSD_LNCH_PRS	PPC_BIT(13)
+/* Invalidation Flag */
+#define   PNV_OCXL_ATSD_LNCH_B		PPC_BIT(14)
+/* Actual Page Size to be invalidated
+ * 000 4KB
+ * 101 64KB
+ * 001 2MB
+ * 010 1GB
+ */
+#define   PNV_OCXL_ATSD_LNCH_AP		PPC_BITMASK(15, 17)
+/* Defines the large page select
+ * L=0b0 for 4KB pages
+ * L=0b1 for large pages)
+ */
+#define   PNV_OCXL_ATSD_LNCH_L		PPC_BIT(18)
+/* Process ID */
+#define   PNV_OCXL_ATSD_LNCH_PID	PPC_BITMASK(19, 38)
+/* NoFlush – Assumed to be 0b0 */
+#define   PNV_OCXL_ATSD_LNCH_F		PPC_BIT(39)
+#define   PNV_OCXL_ATSD_LNCH_OCAPI_SLBI	PPC_BIT(40)
+#define   PNV_OCXL_ATSD_LNCH_OCAPI_SINGLETON	PPC_BIT(41)
+#define PNV_OCXL_ATSD_AVA		0x08
+#define   PNV_OCXL_ATSD_AVA_AVA		PPC_BITMASK(0, 51)
+#define PNV_OCXL_ATSD_STAT		0x10
+
  int pnv_ocxl_get_actag(struct pci_dev *dev, u16 *base, u16 *enabled, u16 *supported);
  int pnv_ocxl_get_pasid_count(struct pci_dev *dev, int *count);
 
@@ -31,4 +78,7 @@ int pnv_ocxl_spa_remove_pe_from_cache(void *platform_data, int pe_handle);
  int pnv_ocxl_map_lpar(struct pci_dev *dev, uint64_t lparid,
  		      uint64_t lpcr, void __iomem **arva);
  void pnv_ocxl_unmap_lpar(void __iomem **arva);
+void pnv_ocxl_tlb_invalidate(void __iomem **arva,
+			     unsigned long pid,
+			     unsigned long addr);
  #endif /* _ASM_PNV_OCXL_H */
arch/powerpc/platforms/powernv/ocxl.c +0 -0
index bc20cf867900..07878496954b 100644

arch/powerpc/platforms/powernv/ocxl.c void pnv_ocxl_unmap_lpar(void __iomem **arva) · line 535

Quoted 5 lines
arch/powerpc/platforms/powernv/ocxl.c +2 -0
@@ -531,3 +531,58 @@ void pnv_ocxl_unmap_lpar(void __iomem **arva)
  	iounmap(*arva);
  }
  EXPORT_SYMBOL_GPL(pnv_ocxl_unmap_lpar);
+
+void pnv_ocxl_tlb_invalidate(void __iomem **arva,

Similarly to the previous patch, there's no reason why arva should be a double-pointer.

Quoted 22 lines
arch/powerpc/platforms/powernv/ocxl.c +22 -0
+			     unsigned long pid,
+			     unsigned long addr)
+{
+	unsigned long timeout = jiffies + (HZ * PNV_OCXL_ATSD_TIMEOUT);
+	uint64_t val = 0ull;
+	int pend;
+
+	if (!(*arva))
+		return;
+
+	if (addr) {
+		/* load Abbreviated Virtual Address register with
+		 * the necessary value
+		 */
+		val |= FIELD_PREP(PNV_OCXL_ATSD_AVA_AVA, addr >> (63-51));
+		out_be64(*arva + PNV_OCXL_ATSD_AVA, val);
+	}
+
+	/* Write access initiates a shoot down to initiate the
+	 * TLB Invalidate command
+	 */
+	val = PNV_OCXL_ATSD_LNCH_R;

arch/powerpc/platforms/powernv/ocxl.c void pnv_ocxl_unmap_lpar(void __iomem **arva) · line 567

Quoted 10 lines
arch/powerpc/platforms/powernv/ocxl.c +10 -0
+	if (addr) {
+		val |= FIELD_PREP(PNV_OCXL_ATSD_LNCH_RIC, 0b00);
+		val |= FIELD_PREP(PNV_OCXL_ATSD_LNCH_IS, 0b00);
+	} else {
+		val |= FIELD_PREP(PNV_OCXL_ATSD_LNCH_RIC, 0b10);
+		val |= FIELD_PREP(PNV_OCXL_ATSD_LNCH_IS, 0b01);
+		val |= PNV_OCXL_ATSD_LNCH_OCAPI_SINGLETON;
+	}
+	val |= PNV_OCXL_ATSD_LNCH_PRS;
+	val |= FIELD_PREP(PNV_OCXL_ATSD_LNCH_AP, 0b101);

So we hard code a page size of 64k. The mmu notifier loops over PAGE_SIZE. It would be cleaner to pass the page size as an argument and code AP based on it.

Fred

Quoted 22 lines
arch/powerpc/platforms/powernv/ocxl.c +21 -0
+	val |= FIELD_PREP(PNV_OCXL_ATSD_LNCH_PID, pid);
+	out_be64(*arva + PNV_OCXL_ATSD_LNCH, val);
+
+	/* Poll the ATSD status register to determine when the
+	 * TLB Invalidate has been completed.
+	 */
+	val = in_be64(*arva + PNV_OCXL_ATSD_STAT);
+	pend = val >> 63;
+
+	while (pend) {
+		if (time_after_eq(jiffies, timeout)) {
+			pr_err("%s - Timeout while reading XTS MMIO ATSD status register (val=%#llx, pidr=0x%lx)\n",
+			       __func__, val, pid);
+			return;
+		}
+		cpu_relax();
+		val = in_be64(*arva + PNV_OCXL_ATSD_STAT);
+		pend = val >> 63;
+	}
+}
+EXPORT_SYMBOL_GPL(pnv_ocxl_tlb_invalidate);
#5 lore raw

PATCH v2 3/5 ocxl: Update the Process Element Entry

In reply to Christophe Lombard

To complete the MMIO based mechanism, the fields: PASID, bus, device and function of the Process Element Entry have to be filled. (See OpenCAPI Power Platform Architecture document)

Hypervisor Process Element Entry
Word

    0 1 .... 7  8  ...... 12  13 ..15  16.... 19  20 ........... 31
0                  OSL Configuration State (0:31)
1                  OSL Configuration State (32:63)

2 PASID | Reserved
3 Bus | Device |Function | Reserved
4 Reserved
5 Reserved
6 ....

---
 drivers/misc/ocxl/context.c       | 4 +++-
 drivers/misc/ocxl/link.c          | 4 +++-
 drivers/misc/ocxl/ocxl_internal.h | 4 +++-
 drivers/scsi/cxlflash/ocxl_hw.c   | 6 ++++--
 include/misc/ocxl.h               | 2 +-
 5 files changed, 14 insertions(+), 6 deletions(-)
drivers/misc/ocxl/context.c +3 -1
index c21f65a5c762..9eb0d93b01c6 100644
@@ -70,6 +70,7 @@ int ocxl_context_attach(struct ocxl_context *ctx, u64 amr, struct mm_struct *mm)
 {
 	int rc;
 	unsigned long pidr = 0;
+	struct pci_dev *dev;
 
 	// Locks both status & tidr
 	mutex_lock(&ctx->status_mutex);
@@ -81,8 +82,9 @@ int ocxl_context_attach(struct ocxl_context *ctx, u64 amr, struct mm_struct *mm)
 	if (mm)
 		pidr = mm->context.id;
 
+	dev = to_pci_dev(ctx->afu->fn->dev.parent);
 	rc = ocxl_link_add_pe(ctx->afu->fn->link, ctx->pasid, pidr, ctx->tidr,
-			      amr, mm, xsl_fault_error, ctx);
+			      amr, pci_dev_id(dev), mm, xsl_fault_error, ctx);
 	if (rc)
 		goto out;
 
drivers/misc/ocxl/link.c +3 -1
index fd73d3bc0eb6..20444db8a2bb 100644
@@ -494,7 +494,7 @@ static u64 calculate_cfg_state(bool kernel)
 }
 
 int ocxl_link_add_pe(void *link_handle, int pasid, u32 pidr, u32 tidr,
-		u64 amr, struct mm_struct *mm,
+		u64 amr, u64 bdf, struct mm_struct *mm,
 		void (*xsl_err_cb)(void *data, u64 addr, u64 dsisr),
 		void *xsl_err_data)
 {
@@ -529,6 +529,8 @@ int ocxl_link_add_pe(void *link_handle, int pasid, u32 pidr, u32 tidr,
 
 	memset(pe, 0, sizeof(struct ocxl_process_element));
 	pe->config_state = cpu_to_be64(calculate_cfg_state(pidr == 0));
+	pe->pasid = cpu_to_be32(pasid << (31 - 19));
+	pe->bdf = cpu_to_be32(bdf << (31 - 15));
 	pe->lpid = cpu_to_be32(mfspr(SPRN_LPID));
 	pe->pid = cpu_to_be32(pidr);
 	pe->tid = cpu_to_be32(tidr);
drivers/misc/ocxl/ocxl_internal.h +3 -1
index 0bad0a123af6..c9ce2af21d6f 100644
@@ -84,7 +84,9 @@ struct ocxl_context {
 
 struct ocxl_process_element {
 	__be64 config_state;
-	__be32 reserved1[11];
+	__be32 pasid;
+	__be32 bdf;
+	__be32 reserved1[9];
 	__be32 lpid;
 	__be32 tid;
 	__be32 pid;
drivers/scsi/cxlflash/ocxl_hw.c +4 -2
index e4e0d767b98e..244fc27215dc 100644
@@ -329,6 +329,7 @@ static int start_context(struct ocxlflash_context *ctx)
 	struct ocxl_hw_afu *afu = ctx->hw_afu;
 	struct ocxl_afu_config *acfg = &afu->acfg;
 	void *link_token = afu->link_token;
+	struct pci_dev *pdev = afu->pdev;
 	struct device *dev = afu->dev;
 	bool master = ctx->master;
 	struct mm_struct *mm;
@@ -360,8 +361,9 @@ static int start_context(struct ocxlflash_context *ctx)
 		mm = current->mm;
 	}
 
-	rc = ocxl_link_add_pe(link_token, ctx->pe, pid, 0, 0, mm,
-			      ocxlflash_xsl_fault, ctx);
+	rc = ocxl_link_add_pe(link_token, ctx->pe, pid, 0, 0,
+			      pci_dev_id(pdev), mm, ocxlflash_xsl_fault,
+			      ctx);
 	if (unlikely(rc)) {
 		dev_err(dev, "%s: ocxl_link_add_pe failed rc=%d\n",
 			__func__, rc);
include/misc/ocxl.h +1 -1
index e013736e275d..d0f101f428dd 100644
@@ -447,7 +447,7 @@ void ocxl_link_release(struct pci_dev *dev, void *link_handle);
  * defined
  */
 int ocxl_link_add_pe(void *link_handle, int pasid, u32 pidr, u32 tidr,
-		u64 amr, struct mm_struct *mm,
+		u64 amr, u64 bdf, struct mm_struct *mm,
 		void (*xsl_err_cb)(void *data, u64 addr, u64 dsisr),
 		void *xsl_err_data);
 
Signature · 1 line
2.28.0
#6 lore raw

Re: [PATCH V2 3/5] ocxl: Update the Process Element Entry

Quoted 52 lines · Christophe Lombard

To complete the MMIO based mechanism, the fields: PASID, bus, device and function of the Process Element Entry have to be filled. (See OpenCAPI Power Platform Architecture document)

Hypervisor Process Element Entry
Word

     0 1 .... 7  8  ...... 12  13 ..15  16.... 19  20 ........... 31

0 OSL Configuration State (0:31)
1 OSL Configuration State (32:63)
2 PASID | Reserved
3 Bus | Device |Function | Reserved
4 Reserved
5 Reserved
6 ....

---
  drivers/misc/ocxl/context.c       | 4 +++-
  drivers/misc/ocxl/link.c          | 4 +++-
  drivers/misc/ocxl/ocxl_internal.h | 4 +++-
  drivers/scsi/cxlflash/ocxl_hw.c   | 6 ++++--
  include/misc/ocxl.h               | 2 +-
  5 files changed, 14 insertions(+), 6 deletions(-)
drivers/misc/ocxl/context.c +3 -1
index c21f65a5c762..9eb0d93b01c6 100644
@@ -70,6 +70,7 @@ int ocxl_context_attach(struct ocxl_context *ctx, u64 amr, struct mm_struct *mm)
  {
  	int rc;
  	unsigned long pidr = 0;
+	struct pci_dev *dev;
 
  	// Locks both status & tidr
  	mutex_lock(&ctx->status_mutex);
@@ -81,8 +82,9 @@ int ocxl_context_attach(struct ocxl_context *ctx, u64 amr, struct mm_struct *mm)
  	if (mm)
  		pidr = mm->context.id;
 
+	dev = to_pci_dev(ctx->afu->fn->dev.parent);
  	rc = ocxl_link_add_pe(ctx->afu->fn->link, ctx->pasid, pidr, ctx->tidr,
-			      amr, mm, xsl_fault_error, ctx);
+			      amr, pci_dev_id(dev), mm, xsl_fault_error, ctx);
  	if (rc)
  		goto out;
 
drivers/misc/ocxl/link.c +0 -0
index fd73d3bc0eb6..20444db8a2bb 100644

drivers/misc/ocxl/link.c static u64 calculate_cfg_state(bool kernel) · line 497

Quoted 5 lines
drivers/misc/ocxl/link.c +1 -1
@@ -494,7 +494,7 @@ static u64 calculate_cfg_state(bool kernel)
  }
 
  int ocxl_link_add_pe(void *link_handle, int pasid, u32 pidr, u32 tidr,
-		u64 amr, struct mm_struct *mm,
+		u64 amr, u64 bdf, struct mm_struct *mm,

bdf could/should be a u16, since that's per the PCI spec.

Fred

Quoted 64 lines
drivers/misc/ocxl/link.c +2 -0
  		void (*xsl_err_cb)(void *data, u64 addr, u64 dsisr),
  		void *xsl_err_data)
  {
@@ -529,6 +529,8 @@ int ocxl_link_add_pe(void *link_handle, int pasid, u32 pidr, u32 tidr,
 
  	memset(pe, 0, sizeof(struct ocxl_process_element));
  	pe->config_state = cpu_to_be64(calculate_cfg_state(pidr == 0));
+	pe->pasid = cpu_to_be32(pasid << (31 - 19));
+	pe->bdf = cpu_to_be32(bdf << (31 - 15));
  	pe->lpid = cpu_to_be32(mfspr(SPRN_LPID));
  	pe->pid = cpu_to_be32(pidr);
  	pe->tid = cpu_to_be32(tidr);
drivers/misc/ocxl/ocxl_internal.h +3 -1
index 0bad0a123af6..c9ce2af21d6f 100644
@@ -84,7 +84,9 @@ struct ocxl_context {
 
  struct ocxl_process_element {
  	__be64 config_state;
-	__be32 reserved1[11];
+	__be32 pasid;
+	__be32 bdf;
+	__be32 reserved1[9];
  	__be32 lpid;
  	__be32 tid;
  	__be32 pid;
drivers/scsi/cxlflash/ocxl_hw.c +4 -2
index e4e0d767b98e..244fc27215dc 100644
@@ -329,6 +329,7 @@ static int start_context(struct ocxlflash_context *ctx)
  	struct ocxl_hw_afu *afu = ctx->hw_afu;
  	struct ocxl_afu_config *acfg = &afu->acfg;
  	void *link_token = afu->link_token;
+	struct pci_dev *pdev = afu->pdev;
  	struct device *dev = afu->dev;
  	bool master = ctx->master;
  	struct mm_struct *mm;
@@ -360,8 +361,9 @@ static int start_context(struct ocxlflash_context *ctx)
  		mm = current->mm;
  	}
 
-	rc = ocxl_link_add_pe(link_token, ctx->pe, pid, 0, 0, mm,
-			      ocxlflash_xsl_fault, ctx);
+	rc = ocxl_link_add_pe(link_token, ctx->pe, pid, 0, 0,
+			      pci_dev_id(pdev), mm, ocxlflash_xsl_fault,
+			      ctx);
  	if (unlikely(rc)) {
  		dev_err(dev, "%s: ocxl_link_add_pe failed rc=%d\n",
  			__func__, rc);
include/misc/ocxl.h +1 -1
index e013736e275d..d0f101f428dd 100644
@@ -447,7 +447,7 @@ void ocxl_link_release(struct pci_dev *dev, void *link_handle);
   * defined
   */
  int ocxl_link_add_pe(void *link_handle, int pasid, u32 pidr, u32 tidr,
-		u64 amr, struct mm_struct *mm,
+		u64 amr, u64 bdf, struct mm_struct *mm,
  		void (*xsl_err_cb)(void *data, u64 addr, u64 dsisr),
  		void *xsl_err_data);
 
#7 lore raw

PATCH v2 4/5 ocxl: Add mmu notifier

In reply to Christophe Lombard

Add invalidate_range mmu notifier, when required (ATSD access of MMIO registers is available), to initiate TLB invalidation commands. For the time being, the ATSD0 set of registers is used by default.

The pasid and bdf values have to be configured in the Process Element Entry.
The PEE must be set up to match the BDF/PASID of the AFU.

---
 drivers/misc/ocxl/link.c | 58 +++++++++++++++++++++++++++++++++++++++-
 1 file changed, 57 insertions(+), 1 deletion(-)
drivers/misc/ocxl/link.c +57 -1
index 20444db8a2bb..100bdfe9ec37 100644
@@ -2,8 +2,10 @@
 // Copyright 2017 IBM Corp.
 #include <linux/sched/mm.h>
 #include <linux/mutex.h>
+#include <linux/mm.h>
 #include <linux/mm_types.h>
 #include <linux/mmu_context.h>
+#include <linux/mmu_notifier.h>
 #include <asm/copro.h>
 #include <asm/pnv-ocxl.h>
 #include <asm/xive.h>
@@ -33,6 +35,7 @@
 
 #define SPA_PE_VALID		0x80000000
 
+struct ocxl_link;
 
 struct pe_data {
 	struct mm_struct *mm;
@@ -41,6 +44,8 @@ struct pe_data {
 	/* opaque pointer to be passed to the above callback */
 	void *xsl_err_data;
 	struct rcu_head rcu;
+	struct ocxl_link *link;
+	struct mmu_notifier mmu_notifier;
 };
 
 struct spa {
@@ -83,6 +88,8 @@ struct ocxl_link {
 	int domain;
 	int bus;
 	int dev;
+	void __iomem *arva;     /* ATSD register virtual address */
+	spinlock_t atsd_lock;   /* to serialize shootdowns */
 	atomic_t irq_available;
 	struct spa *spa;
 	void *platform_data;
@@ -403,6 +410,11 @@ static int alloc_link(struct pci_dev *dev, int PE_mask, struct ocxl_link **out_l
 	if (rc)
 		goto err_xsl_irq;
 
+	rc = pnv_ocxl_map_lpar(dev, mfspr(SPRN_LPID), 0,
+					  &link->arva);
+	if (!rc)
+		spin_lock_init(&link->atsd_lock);
+
 	*out_link = link;
 	return 0;
 
@@ -454,6 +466,11 @@ static void release_xsl(struct kref *ref)
 {
 	struct ocxl_link *link = container_of(ref, struct ocxl_link, ref);
 
+	if (link->arva) {
+		pnv_ocxl_unmap_lpar(&link->arva);
+		link->arva = NULL;
+	}
+
 	list_del(&link->list);
 	/* call platform code before releasing data */
 	pnv_ocxl_spa_release(link->platform_data);
@@ -470,6 +487,26 @@ void ocxl_link_release(struct pci_dev *dev, void *link_handle)
 }
 EXPORT_SYMBOL_GPL(ocxl_link_release);
 
+static void invalidate_range(struct mmu_notifier *mn,
+			     struct mm_struct *mm,
+			     unsigned long start, unsigned long end)
+{
+	struct pe_data *pe_data = container_of(mn, struct pe_data, mmu_notifier);
+	struct ocxl_link *link = pe_data->link;
+	unsigned long addr, pid, page_size = PAGE_SIZE;
+
+	pid = mm->context.id;
+
+	spin_lock(&link->atsd_lock);
+	for (addr = start; addr < end; addr += page_size)
+		pnv_ocxl_tlb_invalidate(&link->arva, pid, addr);
+	spin_unlock(&link->atsd_lock);
+}
+
+static const struct mmu_notifier_ops ocxl_mmu_notifier_ops = {
+	.invalidate_range = invalidate_range,
+};
+
 static u64 calculate_cfg_state(bool kernel)
 {
 	u64 state;
@@ -526,6 +563,8 @@ int ocxl_link_add_pe(void *link_handle, int pasid, u32 pidr, u32 tidr,
 	pe_data->mm = mm;
 	pe_data->xsl_err_cb = xsl_err_cb;
 	pe_data->xsl_err_data = xsl_err_data;
+	pe_data->link = link;
+	pe_data->mmu_notifier.ops = &ocxl_mmu_notifier_ops;
 
 	memset(pe, 0, sizeof(struct ocxl_process_element));
 	pe->config_state = cpu_to_be64(calculate_cfg_state(pidr == 0));
@@ -542,8 +581,16 @@ int ocxl_link_add_pe(void *link_handle, int pasid, u32 pidr, u32 tidr,
 	 * by the nest MMU. If we have a kernel context, TLBIs are
 	 * already global.
 	 */
-	if (mm)
+	if (mm) {
 		mm_context_add_copro(mm);
+		if (link->arva) {
+			/* Use MMIO registers for the TLB Invalidate
+			 * operations.
+			 */
+			mmu_notifier_register(&pe_data->mmu_notifier, mm);
+		}
+	}
+
 	/*
 	 * Barrier is to make sure PE is visible in the SPA before it
 	 * is used by the device. It also helps with the global TLBI
@@ -674,6 +721,15 @@ int ocxl_link_remove_pe(void *link_handle, int pasid)
 		WARN(1, "Couldn't find pe data when removing PE\n");
 	} else {
 		if (pe_data->mm) {
+			if (link->arva) {
+				mmu_notifier_unregister(&pe_data->mmu_notifier,
+							pe_data->mm);
+				spin_lock(&link->atsd_lock);
+				pnv_ocxl_tlb_invalidate(&link->arva,
+							pe_data->mm->context.id,
+							0ull);
+				spin_unlock(&link->atsd_lock);
+			}
 			mm_context_remove_copro(pe_data->mm);
 			mmdrop(pe_data->mm);
 		}
Signature · 1 line
2.28.0
#8 lore raw

Re: [PATCH V2 4/5] ocxl: Add mmu notifier

Quoted 55 lines · Christophe Lombard

Add invalidate_range mmu notifier, when required (ATSD access of MMIO registers is available), to initiate TLB invalidation commands. For the time being, the ATSD0 set of registers is used by default.

The pasid and bdf values have to be configured in the Process Element Entry.
The PEE must be set up to match the BDF/PASID of the AFU.

---
  drivers/misc/ocxl/link.c | 58 +++++++++++++++++++++++++++++++++++++++-
  1 file changed, 57 insertions(+), 1 deletion(-)
drivers/misc/ocxl/link.c +7 -0
index 20444db8a2bb..100bdfe9ec37 100644
@@ -2,8 +2,10 @@
  // Copyright 2017 IBM Corp.
  #include <linux/sched/mm.h>
  #include <linux/mutex.h>
+#include <linux/mm.h>
  #include <linux/mm_types.h>
  #include <linux/mmu_context.h>
+#include <linux/mmu_notifier.h>
  #include <asm/copro.h>
  #include <asm/pnv-ocxl.h>
  #include <asm/xive.h>
@@ -33,6 +35,7 @@
 
  #define SPA_PE_VALID		0x80000000
 
+struct ocxl_link;
 
  struct pe_data {
  	struct mm_struct *mm;
@@ -41,6 +44,8 @@ struct pe_data {
  	/* opaque pointer to be passed to the above callback */
  	void *xsl_err_data;
  	struct rcu_head rcu;
+	struct ocxl_link *link;
+	struct mmu_notifier mmu_notifier;
  };
 
  struct spa {
@@ -83,6 +88,8 @@ struct ocxl_link {
  	int domain;
  	int bus;
  	int dev;
+	void __iomem *arva;     /* ATSD register virtual address */
+	spinlock_t atsd_lock;   /* to serialize shootdowns */
  	atomic_t irq_available;
  	struct spa *spa;
  	void *platform_data;

drivers/misc/ocxl/link.c static int alloc_link(struct pci_dev *dev, int PE_mask, struct ocxl_link **out_l · line 417

Quoted 8 lines
drivers/misc/ocxl/link.c +5 -0
@@ -403,6 +410,11 @@ static int alloc_link(struct pci_dev *dev, int PE_mask, struct ocxl_link **out_l
  	if (rc)
  		goto err_xsl_irq;
 
+	rc = pnv_ocxl_map_lpar(dev, mfspr(SPRN_LPID), 0,
+					  &link->arva);
+	if (!rc)
+		spin_lock_init(&link->atsd_lock);
+

We could use a comment to say that if arva = 0, then we don't need mmio shootdowns and we rely on hardware snooping.

Also, we could always initialize the spin lock, it doesn't hurt and make the code more readable.

Fred

Quoted 86 lines
drivers/misc/ocxl/link.c +45 -1
  	*out_link = link;
  	return 0;
 
@@ -454,6 +466,11 @@ static void release_xsl(struct kref *ref)
  {
  	struct ocxl_link *link = container_of(ref, struct ocxl_link, ref);
 
+	if (link->arva) {
+		pnv_ocxl_unmap_lpar(&link->arva);
+		link->arva = NULL;
+	}
+
  	list_del(&link->list);
  	/* call platform code before releasing data */
  	pnv_ocxl_spa_release(link->platform_data);
@@ -470,6 +487,26 @@ void ocxl_link_release(struct pci_dev *dev, void *link_handle)
  }
  EXPORT_SYMBOL_GPL(ocxl_link_release);
 
+static void invalidate_range(struct mmu_notifier *mn,
+			     struct mm_struct *mm,
+			     unsigned long start, unsigned long end)
+{
+	struct pe_data *pe_data = container_of(mn, struct pe_data, mmu_notifier);
+	struct ocxl_link *link = pe_data->link;
+	unsigned long addr, pid, page_size = PAGE_SIZE;
+
+	pid = mm->context.id;
+
+	spin_lock(&link->atsd_lock);
+	for (addr = start; addr < end; addr += page_size)
+		pnv_ocxl_tlb_invalidate(&link->arva, pid, addr);
+	spin_unlock(&link->atsd_lock);
+}
+
+static const struct mmu_notifier_ops ocxl_mmu_notifier_ops = {
+	.invalidate_range = invalidate_range,
+};
+
  static u64 calculate_cfg_state(bool kernel)
  {
  	u64 state;
@@ -526,6 +563,8 @@ int ocxl_link_add_pe(void *link_handle, int pasid, u32 pidr, u32 tidr,
  	pe_data->mm = mm;
  	pe_data->xsl_err_cb = xsl_err_cb;
  	pe_data->xsl_err_data = xsl_err_data;
+	pe_data->link = link;
+	pe_data->mmu_notifier.ops = &ocxl_mmu_notifier_ops;
 
  	memset(pe, 0, sizeof(struct ocxl_process_element));
  	pe->config_state = cpu_to_be64(calculate_cfg_state(pidr == 0));
@@ -542,8 +581,16 @@ int ocxl_link_add_pe(void *link_handle, int pasid, u32 pidr, u32 tidr,
  	 * by the nest MMU. If we have a kernel context, TLBIs are
  	 * already global.
  	 */
-	if (mm)
+	if (mm) {
  		mm_context_add_copro(mm);
+		if (link->arva) {
+			/* Use MMIO registers for the TLB Invalidate
+			 * operations.
+			 */
+			mmu_notifier_register(&pe_data->mmu_notifier, mm);
+		}
+	}
+
  	/*
  	 * Barrier is to make sure PE is visible in the SPA before it
  	 * is used by the device. It also helps with the global TLBI
@@ -674,6 +721,15 @@ int ocxl_link_remove_pe(void *link_handle, int pasid)
  		WARN(1, "Couldn't find pe data when removing PE\n");
  	} else {
  		if (pe_data->mm) {
+			if (link->arva) {
+				mmu_notifier_unregister(&pe_data->mmu_notifier,
+							pe_data->mm);
+				spin_lock(&link->atsd_lock);
+				pnv_ocxl_tlb_invalidate(&link->arva,
+							pe_data->mm->context.id,
+							0ull);
+				spin_unlock(&link->atsd_lock);
+			}
  			mm_context_remove_copro(pe_data->mm);
  			mmdrop(pe_data->mm);
  		}
#9 lore raw

Re: [PATCH V2 4/5] ocxl: Add mmu notifier

In reply to Christophe Lombard

You probably want to add Jason for an audit of new notifier uses.

Quoted 151 lines · Christophe Lombard

Add invalidate_range mmu notifier, when required (ATSD access of MMIO registers is available), to initiate TLB invalidation commands. For the time being, the ATSD0 set of registers is used by default.

The pasid and bdf values have to be configured in the Process Element Entry.
The PEE must be set up to match the BDF/PASID of the AFU.

---
 drivers/misc/ocxl/link.c | 58 +++++++++++++++++++++++++++++++++++++++-
 1 file changed, 57 insertions(+), 1 deletion(-)
drivers/misc/ocxl/link.c +57 -1
index 20444db8a2bb..100bdfe9ec37 100644
@@ -2,8 +2,10 @@
 // Copyright 2017 IBM Corp.
 #include <linux/sched/mm.h>
 #include <linux/mutex.h>
+#include <linux/mm.h>
 #include <linux/mm_types.h>
 #include <linux/mmu_context.h>
+#include <linux/mmu_notifier.h>
 #include <asm/copro.h>
 #include <asm/pnv-ocxl.h>
 #include <asm/xive.h>
@@ -33,6 +35,7 @@
 
 #define SPA_PE_VALID		0x80000000
 
+struct ocxl_link;
 
 struct pe_data {
 	struct mm_struct *mm;
@@ -41,6 +44,8 @@ struct pe_data {
 	/* opaque pointer to be passed to the above callback */
 	void *xsl_err_data;
 	struct rcu_head rcu;
+	struct ocxl_link *link;
+	struct mmu_notifier mmu_notifier;
 };
 
 struct spa {
@@ -83,6 +88,8 @@ struct ocxl_link {
 	int domain;
 	int bus;
 	int dev;
+	void __iomem *arva;     /* ATSD register virtual address */
+	spinlock_t atsd_lock;   /* to serialize shootdowns */
 	atomic_t irq_available;
 	struct spa *spa;
 	void *platform_data;
@@ -403,6 +410,11 @@ static int alloc_link(struct pci_dev *dev, int PE_mask, struct ocxl_link **out_l
 	if (rc)
 		goto err_xsl_irq;
 
+	rc = pnv_ocxl_map_lpar(dev, mfspr(SPRN_LPID), 0,
+					  &link->arva);
+	if (!rc)
+		spin_lock_init(&link->atsd_lock);
+
 	*out_link = link;
 	return 0;
 
@@ -454,6 +466,11 @@ static void release_xsl(struct kref *ref)
 {
 	struct ocxl_link *link = container_of(ref, struct ocxl_link, ref);
 
+	if (link->arva) {
+		pnv_ocxl_unmap_lpar(&link->arva);
+		link->arva = NULL;
+	}
+
 	list_del(&link->list);
 	/* call platform code before releasing data */
 	pnv_ocxl_spa_release(link->platform_data);
@@ -470,6 +487,26 @@ void ocxl_link_release(struct pci_dev *dev, void *link_handle)
 }
 EXPORT_SYMBOL_GPL(ocxl_link_release);
 
+static void invalidate_range(struct mmu_notifier *mn,
+			     struct mm_struct *mm,
+			     unsigned long start, unsigned long end)
+{
+	struct pe_data *pe_data = container_of(mn, struct pe_data, mmu_notifier);
+	struct ocxl_link *link = pe_data->link;
+	unsigned long addr, pid, page_size = PAGE_SIZE;
+
+	pid = mm->context.id;
+
+	spin_lock(&link->atsd_lock);
+	for (addr = start; addr < end; addr += page_size)
+		pnv_ocxl_tlb_invalidate(&link->arva, pid, addr);
+	spin_unlock(&link->atsd_lock);
+}
+
+static const struct mmu_notifier_ops ocxl_mmu_notifier_ops = {
+	.invalidate_range = invalidate_range,
+};
+
 static u64 calculate_cfg_state(bool kernel)
 {
 	u64 state;
@@ -526,6 +563,8 @@ int ocxl_link_add_pe(void *link_handle, int pasid, u32 pidr, u32 tidr,
 	pe_data->mm = mm;
 	pe_data->xsl_err_cb = xsl_err_cb;
 	pe_data->xsl_err_data = xsl_err_data;
+	pe_data->link = link;
+	pe_data->mmu_notifier.ops = &ocxl_mmu_notifier_ops;
 
 	memset(pe, 0, sizeof(struct ocxl_process_element));
 	pe->config_state = cpu_to_be64(calculate_cfg_state(pidr == 0));
@@ -542,8 +581,16 @@ int ocxl_link_add_pe(void *link_handle, int pasid, u32 pidr, u32 tidr,
 	 * by the nest MMU. If we have a kernel context, TLBIs are
 	 * already global.
 	 */
-	if (mm)
+	if (mm) {
 		mm_context_add_copro(mm);
+		if (link->arva) {
+			/* Use MMIO registers for the TLB Invalidate
+			 * operations.
+			 */
+			mmu_notifier_register(&pe_data->mmu_notifier, mm);
+		}
+	}
+
 	/*
 	 * Barrier is to make sure PE is visible in the SPA before it
 	 * is used by the device. It also helps with the global TLBI
@@ -674,6 +721,15 @@ int ocxl_link_remove_pe(void *link_handle, int pasid)
 		WARN(1, "Couldn't find pe data when removing PE\n");
 	} else {
 		if (pe_data->mm) {
+			if (link->arva) {
+				mmu_notifier_unregister(&pe_data->mmu_notifier,
+							pe_data->mm);
+				spin_lock(&link->atsd_lock);
+				pnv_ocxl_tlb_invalidate(&link->arva,
+							pe_data->mm->context.id,
+							0ull);
+				spin_unlock(&link->atsd_lock);
+			}
 			mm_context_remove_copro(pe_data->mm);
 			mmdrop(pe_data->mm);
 		}
Signature · 1 line
2.28.0

---end quoted text---

#10 lore raw

Re: [PATCH V2 4/5] ocxl: Add mmu notifier

Quoted 11 lines · Christoph Hellwig

drivers/misc/ocxl/link.c void ocxl_link_release(struct pci_dev *dev, void *link_handle) · line 496

Quoted 11 lines
drivers/misc/ocxl/link.c +7 -0
@@ -470,6 +487,26 @@ void ocxl_link_release(struct pci_dev *dev, void *link_handle)
 }
 EXPORT_SYMBOL_GPL(ocxl_link_release);
 
+static void invalidate_range(struct mmu_notifier *mn,
+			     struct mm_struct *mm,
+			     unsigned long start, unsigned long end)
+{
+	struct pe_data *pe_data = container_of(mn, struct pe_data, mmu_notifier);
+	struct ocxl_link *link = pe_data->link;
+	unsigned long addr, pid, page_size = PAGE_SIZE;

The page_size variable seems unnecessary

Quoted 23 lines
Quoted 23 lines
drivers/misc/ocxl/link.c +15 -0
+
+	pid = mm->context.id;
+
+	spin_lock(&link->atsd_lock);
+	for (addr = start; addr < end; addr += page_size)
+		pnv_ocxl_tlb_invalidate(&link->arva, pid, addr);
+	spin_unlock(&link->atsd_lock);
+}
+
+static const struct mmu_notifier_ops ocxl_mmu_notifier_ops = {
+	.invalidate_range = invalidate_range,
+};
+
 static u64 calculate_cfg_state(bool kernel)
 {
 	u64 state;
@@ -526,6 +563,8 @@ int ocxl_link_add_pe(void *link_handle, int pasid, u32 pidr, u32 tidr,
 	pe_data->mm = mm;
 	pe_data->xsl_err_cb = xsl_err_cb;
 	pe_data->xsl_err_data = xsl_err_data;
+	pe_data->link = link;
+	pe_data->mmu_notifier.ops = &ocxl_mmu_notifier_ops;
 
Quoted 14 lines
Quoted 4 lines
drivers/misc/ocxl/link.c +0 -0
 	memset(pe, 0, sizeof(struct ocxl_process_element));
 	pe->config_state = cpu_to_be64(calculate_cfg_state(pidr == 0));
@@ -542,8 +581,16 @@ int ocxl_link_add_pe(void *link_handle, int pasid, u32 pidr, u32 tidr,
 	 * by the nest MMU. If we have a kernel context, TLBIs are

drivers/misc/ocxl/link.c int ocxl_link_add_pe(void *link_handle, int pasid, u32 pidr, u32 tidr, · line 590

Quoted 10 lines
drivers/misc/ocxl/link.c +6 -1
 	 * already global.
 	 */
-	if (mm)
+	if (mm) {
 		mm_context_add_copro(mm);
+		if (link->arva) {
+			/* Use MMIO registers for the TLB Invalidate
+			 * operations.
+			 */
+			mmu_notifier_register(&pe_data->mmu_notifier, mm);

Every other place doing stuff like this is de-duplicating the notifier. If you have multiple clients this will do multiple redundant invalidations?

The notifier get/put API is designed to solve that problem, you'd get a single notifier for the mm and then add the impacted arva's to some list at the notifier.

Jason

#11 lore raw

Re: [PATCH V2 4/5] ocxl: Add mmu notifier

Le 24/11/2020 à 14:45, Jason Gunthorpe a écrit :

Quoted 47 lines · Christoph Hellwig
Quoted 5 lines
Quoted 5 lines
drivers/misc/ocxl/link.c +0 -0
@@ -470,6 +487,26 @@ void ocxl_link_release(struct pci_dev *dev, void *link_handle)
  }
  EXPORT_SYMBOL_GPL(ocxl_link_release);
 
Quoted 7 lines

drivers/misc/ocxl/link.c void ocxl_link_release(struct pci_dev *dev, void *link_handle) · line 496

Quoted 7 lines
drivers/misc/ocxl/link.c +7 -0
+static void invalidate_range(struct mmu_notifier *mn,
+			     struct mm_struct *mm,
+			     unsigned long start, unsigned long end)
+{
+	struct pe_data *pe_data = container_of(mn, struct pe_data, mmu_notifier);
+	struct ocxl_link *link = pe_data->link;
+	unsigned long addr, pid, page_size = PAGE_SIZE;

The page_size variable seems unnecessary

Quoted 32 lines
Quoted 32 lines
drivers/misc/ocxl/link.c +16 -1
+
+	pid = mm->context.id;
+
+	spin_lock(&link->atsd_lock);
+	for (addr = start; addr < end; addr += page_size)
+		pnv_ocxl_tlb_invalidate(&link->arva, pid, addr);
+	spin_unlock(&link->atsd_lock);
+}
+
+static const struct mmu_notifier_ops ocxl_mmu_notifier_ops = {
+	.invalidate_range = invalidate_range,
+};
+
  static u64 calculate_cfg_state(bool kernel)
  {
  	u64 state;
@@ -526,6 +563,8 @@ int ocxl_link_add_pe(void *link_handle, int pasid, u32 pidr, u32 tidr,
  	pe_data->mm = mm;
  	pe_data->xsl_err_cb = xsl_err_cb;
  	pe_data->xsl_err_data = xsl_err_data;
+	pe_data->link = link;
+	pe_data->mmu_notifier.ops = &ocxl_mmu_notifier_ops;
 
  	memset(pe, 0, sizeof(struct ocxl_process_element));
  	pe->config_state = cpu_to_be64(calculate_cfg_state(pidr == 0));
@@ -542,8 +581,16 @@ int ocxl_link_add_pe(void *link_handle, int pasid, u32 pidr, u32 tidr,
  	 * by the nest MMU. If we have a kernel context, TLBIs are
  	 * already global.
  	 */
-	if (mm)
+	if (mm) {
  		mm_context_add_copro(mm);
Quoted 8 lines
Quoted 5 lines

drivers/misc/ocxl/link.c int ocxl_link_add_pe(void *link_handle, int pasid, u32 pidr, u32 tidr, · line 590

Quoted 5 lines
drivers/misc/ocxl/link.c +5 -0
+		if (link->arva) {
+			/* Use MMIO registers for the TLB Invalidate
+			 * operations.
+			 */
+			mmu_notifier_register(&pe_data->mmu_notifier, mm);

Every other place doing stuff like this is de-duplicating the notifier. If you have multiple clients this will do multiple redundant invalidations?

We could have multiple clients, although not something that we have often. We have only one attach per process. But if not, we must still have invalidation for each.

Quoted 4 lines

The notifier get/put API is designed to solve that problem, you'd get a single notifier for the mm and then add the impacted arva's to some list at the notifier.

Thanks for the information.

Quoted 2 lines

Jason

#12 lore raw

PATCH v2 5/5 ocxl: Add new kernel traces

In reply to Christophe Lombard

Add specific kernel traces which provide information on mmu notifier and on pages range.

---
 drivers/misc/ocxl/link.c  |  4 +++
 drivers/misc/ocxl/trace.h | 64 +++++++++++++++++++++++++++++++++++++++
 2 files changed, 68 insertions(+)
drivers/misc/ocxl/link.c +4 -0
index 100bdfe9ec37..2e0eb46d032e 100644
@@ -496,6 +496,7 @@ static void invalidate_range(struct mmu_notifier *mn,
 	unsigned long addr, pid, page_size = PAGE_SIZE;
 
 	pid = mm->context.id;
+	trace_ocxl_mmu_notifier_range(start, end, pid);
 
 	spin_lock(&link->atsd_lock);
 	for (addr = start; addr < end; addr += page_size)
@@ -587,6 +588,7 @@ int ocxl_link_add_pe(void *link_handle, int pasid, u32 pidr, u32 tidr,
 			/* Use MMIO registers for the TLB Invalidate
 			 * operations.
 			 */
+			trace_ocxl_init_mmu_notifier(pasid, mm->context.id);
 			mmu_notifier_register(&pe_data->mmu_notifier, mm);
 		}
 	}
@@ -722,6 +724,8 @@ int ocxl_link_remove_pe(void *link_handle, int pasid)
 	} else {
 		if (pe_data->mm) {
 			if (link->arva) {
+				trace_ocxl_release_mmu_notifier(pasid,
+								pe_data->mm->context.id);
 				mmu_notifier_unregister(&pe_data->mmu_notifier,
 							pe_data->mm);
 				spin_lock(&link->atsd_lock);
drivers/misc/ocxl/trace.h +64 -0

70 lines not on this page · raw message

Signature · 1 line
2.28.0