On Wed, 2022-02-23 at 17:05 +1000, Nicholas Piggin wrote:

Excerpts from Haren Myneni's message of February 20, 2022 5:58 am:

quoted

The VAS window may not be active if the system looses credits and
the NX generates page fault when it receives request on unmap
paste address.

The kernel handles the fault by remap new paste address if the
window is active again, Otherwise return the paste instruction
failure if the executed instruction that caused the fault was
a paste.

Looks good, thanks for fixin the SIGBUS thing, was that my
fault? I vaguely remember writing some of this patch :P

Thanks for your reviews on all patches. 

No, it was my fault not handling the -EGAIN error. 

Thanks,
Nick

quoted

Signed-off-by: Nicholas Piggin <npiggin@gmail.com>
Signed-off-by: Haren Myneni <haren@linux.ibm.com>
---
 arch/powerpc/include/asm/ppc-opcode.h   |  2 +
 arch/powerpc/platforms/book3s/vas-api.c | 55
++++++++++++++++++++++++-
 2 files changed, 56 insertions(+), 1 deletion(-)

diff --git a/arch/powerpc/include/asm/ppc-opcode.h

b/arch/powerpc/include/asm/ppc-opcode.h
index 9675303b724e..82f1f0041c6f 100644

--- a/arch/powerpc/include/asm/ppc-opcode.h
+++ b/arch/powerpc/include/asm/ppc-opcode.h

@@ -262,6 +262,8 @@
 #define PPC_INST_MFSPR_PVR		0x7c1f42a6
 #define PPC_INST_MFSPR_PVR_MASK		0xfc1ffffe
 #define PPC_INST_MTMSRD			0x7c000164
+#define PPC_INST_PASTE			0x7c20070d
+#define PPC_INST_PASTE_MASK		0xfc2007ff
 #define PPC_INST_POPCNTB		0x7c0000f4
 #define PPC_INST_POPCNTB_MASK		0xfc0007fe
 #define PPC_INST_RFEBB			0x4c000124

diff --git a/arch/powerpc/platforms/book3s/vas-api.c

b/arch/powerpc/platforms/book3s/vas-api.c
index f359e7b2bf90..f3e421511ea6 100644

--- a/arch/powerpc/platforms/book3s/vas-api.c
+++ b/arch/powerpc/platforms/book3s/vas-api.c

@@ -351,6 +351,41 @@ static int coproc_release(struct inode *inode,

struct file *fp)
 	return 0;
 }
 
+/*
+ * If the executed instruction that caused the fault was a paste,
then
+ * clear regs CR0[EQ], advance NIP, and return 0. Else return
error code.
+ */
+static int do_fail_paste(void)
+{
+	struct pt_regs *regs = current->thread.regs;
+	u32 instword;
+
+	if (WARN_ON_ONCE(!regs))
+		return -EINVAL;
+
+	if (WARN_ON_ONCE(!user_mode(regs)))
+		return -EINVAL;
+
+	/*
+	 * If we couldn't translate the instruction, the driver should
+	 * return success without handling the fault, it will be
retried
+	 * or the instruction fetch will fault.
+	 */
+	if (get_user(instword, (u32 __user *)(regs->nip)))
+		return -EAGAIN;
+
+	/*
+	 * Not a paste instruction, driver may fail the fault.
+	 */
+	if ((instword & PPC_INST_PASTE_MASK) != PPC_INST_PASTE)
+		return -ENOENT;
+
+	regs->ccr &= ~0xe0000000;	/* Clear CR0[0-2] to fail paste */
+	regs_add_return_ip(regs, 4);	/* Emulate the paste */
+
+	return 0;
+}
+
 /*
  * This fault handler is invoked when the core generates page
fault on
  * the paste address. Happens if the kernel closes window in
hypervisor

@@ -408,9 +443,27 @@ static vm_fault_t vas_mmap_fault(struct

vm_fault *vmf)
 	}
 	mutex_unlock(&txwin->task_ref.mmap_mutex);
 
-	return VM_FAULT_SIGBUS;
+	/*
+	 * Received this fault due to closing the actual window.
+	 * It can happen during migration or lost credits.
+	 * Since no mapping, return the paste instruction failure
+	 * to the user space.
+	 */
+	ret = do_fail_paste();
+	/*
+	 * The user space can retry several times until success (needed
+	 * for migration) or should fallback to SW compression or
+	 * manage with the existing open windows if available.
+	 * Looking at sysfs interface, it can determine whether these
+	 * failures are coming during migration or core removal:
+	 * nr_used_credits > nr_total_credits when lost credits
+	 */
+	if (!ret || (ret == -EAGAIN))
+		return VM_FAULT_NOPAGE;
 
+	return VM_FAULT_SIGBUS;
 }
+
 static const struct vm_operations_struct vas_vm_ops = {
 	.fault = vas_mmap_fault,
 };
-- 
2.27.0