[PATCH v2 3/3] powerpc/pseries/mm: call H_BLOCK_REMOVE

[Laurent Dufour]

This hypervisor's call allows to remove up to 8 ptes with only call to
tlbie.

The virtual pages must be all within the same naturally aligned 8 pages
virtual address block and have the same page and segment size encodings.

Cc: "Aneesh Kumar K.V" <aneesh.kumar@xxxxxxxxxxxxx>
Cc: Nicholas Piggin <npiggin@xxxxxxxxx>
Cc: Michael Ellerman <mpe@xxxxxxxxxxxxxx>
Cc: Paul Mackerras <paulus@xxxxxxxxx>
Cc: Benjamin Herrenschmidt <benh@xxxxxxxxxxxxxxxxxxx>
Signed-off-by: Laurent Dufour <ldufour@xxxxxxxxxxxxxxxxxx>
---
 arch/powerpc/include/asm/hvcall.h     |   1 +
 arch/powerpc/platforms/pseries/lpar.c | 214 ++++++++++++++++++++++++++++++++--
 2 files changed, 207 insertions(+), 8 deletions(-)

diff --git a/arch/powerpc/include/asm/hvcall.h b/arch/powerpc/include/asm/hvcall.h
index a0b17f9f1ea4..c349d3960d63 100644
--- a/arch/powerpc/include/asm/hvcall.h
+++ b/arch/powerpc/include/asm/hvcall.h
@@ -278,6 +278,7 @@
 #define H_COP			0x304
 #define H_GET_MPP_X		0x314
 #define H_SET_MODE		0x31C
+#define H_BLOCK_REMOVE		0x328
 #define H_CLEAR_HPT		0x358
 #define H_REQUEST_VMC		0x360
 #define H_RESIZE_HPT_PREPARE	0x36C
diff --git a/arch/powerpc/platforms/pseries/lpar.c b/arch/powerpc/platforms/pseries/lpar.c
index ebc852e3607d..0b5081085a44 100644
--- a/arch/powerpc/platforms/pseries/lpar.c
+++ b/arch/powerpc/platforms/pseries/lpar.c
@@ -417,6 +417,79 @@ static void pSeries_lpar_hpte_invalidate(unsigned long slot, unsigned long vpn,
 	BUG_ON(lpar_rc != H_SUCCESS);
 }
 
+
+/*
+ * As defined in the PAPR's section 14.5.4.1.8
+ * The control mask doesn't include the returned reference and change bit from
+ * the processed PTE.
+ */
+#define HBLKR_AVPN		0x0100000000000000UL
+#define HBLKR_CTRL_MASK		0xf800000000000000UL
+#define HBLKR_CTRL_SUCCESS	0x8000000000000000UL
+#define HBLKR_CTRL_ERRNOTFOUND	0x8800000000000000UL
+#define HBLKR_CTRL_ERRBUSY	0xa000000000000000UL
+
+/**
+ * H_BLOCK_REMOVE caller.
+ * @idx should point to the latest @param entry set with a PTEX.
+ * If PTE cannot be processed because another CPUs has already locked that
+ * group, those entries are put back in @param starting at index 1.
+ * If entries has to be retried and @retry_busy is set to true, these entries
+ * are retried until success. If @retry_busy is set to false, the returned
+ * is the number of entries yet to process.
+ */
+static unsigned long call_block_remove(unsigned long idx, unsigned long *param,
+				       bool retry_busy)
+{
+	unsigned long i, rc, new_idx;
+	unsigned long retbuf[PLPAR_HCALL9_BUFSIZE];
+
+	if (idx < 2) {
+		pr_warn("Unexpected empty call to H_BLOCK_REMOVE");
+		return 0;
+	}
+again:
+	new_idx = 0;
+	if (idx > PLPAR_HCALL9_BUFSIZE) {
+		pr_err("Too many PTEs (%lu) for H_BLOCK_REMOVE", idx);
+		idx = PLPAR_HCALL9_BUFSIZE;
+	} else if (idx < PLPAR_HCALL9_BUFSIZE)
+		param[idx] = HBR_END;
+
+	rc = plpar_hcall9(H_BLOCK_REMOVE, retbuf,
+			  param[0], /* AVA */
+			  param[1],  param[2],  param[3],  param[4], /* TS0-7 */
+			  param[5],  param[6],  param[7],  param[8]);
+	if (rc == H_SUCCESS)
+		return 0;
+
+	BUG_ON(rc != H_PARTIAL);
+
+	/* Check that the unprocessed entries were 'not found' or 'busy' */
+	for (i = 0; i < idx-1; i++) {
+		unsigned long ctrl = retbuf[i] & HBLKR_CTRL_MASK;
+
+		if (ctrl == HBLKR_CTRL_ERRBUSY) {
+			param[++new_idx] = param[i+1];
+			continue;
+		}
+
+		BUG_ON(ctrl != HBLKR_CTRL_SUCCESS
+		       && ctrl != HBLKR_CTRL_ERRNOTFOUND);
+	}
+
+	/*
+	 * If there were entries found busy, retry these entries if requested,
+	 * of if all the entries have to be retried.
+	 */
+	if (new_idx && (retry_busy || new_idx == (PLPAR_HCALL9_BUFSIZE-1))) {
+		idx = new_idx + 1;
+		goto again;
+	}
+
+	return new_idx;
+}
+
 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
 /*
  * Limit iterations holding pSeries_lpar_tlbie_lock to 3. We also need
@@ -424,17 +497,57 @@ static void pSeries_lpar_hpte_invalidate(unsigned long slot, unsigned long vpn,
  */
 #define PPC64_HUGE_HPTE_BATCH 12
 
-static void __pSeries_lpar_hugepage_invalidate(unsigned long *slot,
-					     unsigned long *vpn, int count,
-					     int psize, int ssize)
+static void hugepage_block_invalidate(unsigned long *slot, unsigned long *vpn,
+				      int count, int psize, int ssize)
 {
 	unsigned long param[PLPAR_HCALL9_BUFSIZE];
-	int i = 0, pix = 0, rc;
-	unsigned long flags = 0;
-	int lock_tlbie = !mmu_has_feature(MMU_FTR_LOCKLESS_TLBIE);
+	unsigned long shift, current_vpgb, vpgb;
+	int i, pix = 0;
 
-	if (lock_tlbie)
-		spin_lock_irqsave(&pSeries_lpar_tlbie_lock, flags);
+	shift = mmu_psize_defs[psize].shift;
+
+	for (i = 0; i < count; i++) {
+		/*
+		 * Shifting 3 bits more on the right to get a
+		 * 8 pages aligned virtual addresse.
+		 */
+		vpgb = (vpn[i] >> (shift - VPN_SHIFT + 3));
+		if (!pix || vpgb != current_vpgb) {
+			/*
+			 * Need to start a new 8 pages block, flush
+			 * the current one if needed.
+			 */
+			if (pix)
+				(void)call_block_remove(pix, param, true);
+			current_vpgb = vpgb;
+			param[0] = hpte_encode_avpn(vpn[i], psize, ssize);
+			pix = 1;
+		}
+
+		param[pix++] = HBR_REQUEST | HBLKR_AVPN | slot[i];
+		if (pix == PLPAR_HCALL9_BUFSIZE) {
+			pix = call_block_remove(pix, param, false);
+			/*
+			 * pix = 0 means that all the entries were
+			 * removed, we can start a new block.
+			 * Otherwise, this means that there are entries
+			 * to retry, and pix points to latest one, so
+			 * we should increment it and try to continue
+			 * the same block.
+			 */
+			if (pix)
+				pix++;
+		}
+	}
+	if (pix)
+		(void)call_block_remove(pix, param, true);
+}
+
+static void hugepage_bulk_invalidate(unsigned long *slot, unsigned long *vpn,
+				     int count, int psize, int ssize)
+{
+	unsigned long param[PLPAR_HCALL9_BUFSIZE];
+	int i = 0, pix = 0, rc;
 
 	for (i = 0; i < count; i++) {
 
@@ -462,6 +575,23 @@ static void __pSeries_lpar_hugepage_invalidate(unsigned long *slot,
 				  param[6], param[7]);
 		BUG_ON(rc != H_SUCCESS);
 	}
+}
+
+static inline void __pSeries_lpar_hugepage_invalidate(unsigned long *slot,
+						      unsigned long *vpn,
+						      int count, int psize,
+						      int ssize)
+{
+	unsigned long flags = 0;
+	int lock_tlbie = !mmu_has_feature(MMU_FTR_LOCKLESS_TLBIE);
+
+	if (lock_tlbie)
+		spin_lock_irqsave(&pSeries_lpar_tlbie_lock, flags);
+
+	if (firmware_has_feature(FW_FEATURE_BLOCK_REMOVE))
+		hugepage_block_invalidate(slot, vpn, count, psize, ssize);
+	else
+		hugepage_bulk_invalidate(slot, vpn, count, psize, ssize);
 
 	if (lock_tlbie)
 		spin_unlock_irqrestore(&pSeries_lpar_tlbie_lock, flags);
@@ -564,6 +694,68 @@ static inline unsigned long compute_slot(real_pte_t pte,
 	return slot;
 }
 
+/**
+ * The hcall H_BLOCK_REMOVE implies that the virtual pages to processed are
+ * "all within the same naturally aligned 8 page virtual address block".
+ */
+static void do_block_remove(unsigned long number, struct ppc64_tlb_batch *batch,
+			    unsigned long *param)
+{
+	unsigned long vpn;
+	unsigned long i, pix = 0;
+	unsigned long index, shift, slot, current_vpgb, vpgb;
+	real_pte_t pte;
+	int psize, ssize;
+
+	psize = batch->psize;
+	ssize = batch->ssize;
+
+	for (i = 0; i < number; i++) {
+		vpn = batch->vpn[i];
+		pte = batch->pte[i];
+		pte_iterate_hashed_subpages(pte, psize, vpn, index, shift) {
+			/*
+			 * Shifting 3 bits more on the right to get a
+			 * 8 pages aligned virtual addresse.
+			 */
+			vpgb = (vpn >> (shift - VPN_SHIFT + 3));
+			if (!pix || vpgb != current_vpgb) {
+				/*
+				 * Need to start a new 8 pages block, flush
+				 * the current one if needed.
+				 */
+				if (pix)
+					(void)call_block_remove(pix, param,
+								true);
+				current_vpgb = vpgb;
+				param[0] = hpte_encode_avpn(vpn, psize,
+							    ssize);
+				pix = 1;
+			}
+
+			slot = compute_slot(pte, vpn, index, shift, ssize);
+			param[pix++] = HBR_REQUEST | HBLKR_AVPN | slot;
+
+			if (pix == PLPAR_HCALL9_BUFSIZE) {
+				pix = call_block_remove(pix, param, false);
+				/*
+				 * pix = 0 means that all the entries were
+				 * removed, we can start a new block.
+				 * Otherwise, this means that there are entries
+				 * to retry, and pix points to latest one, so
+				 * we should increment it and try to continue
+				 * the same block.
+				 */
+				if (pix)
+					pix++;
+			}
+		} pte_iterate_hashed_end();
+	}
+
+	if (pix)
+		(void)call_block_remove(pix, param, true);
+}
+
 /*
  * Take a spinlock around flushes to avoid bouncing the hypervisor tlbie
  * lock.
@@ -583,6 +775,11 @@ static void pSeries_lpar_flush_hash_range(unsigned long number, int local)
 	if (lock_tlbie)
 		spin_lock_irqsave(&pSeries_lpar_tlbie_lock, flags);
 
+	if (firmware_has_feature(FW_FEATURE_BLOCK_REMOVE)) {
+		do_block_remove(number, batch, param);
+		goto out;
+	}
+
 	psize = batch->psize;
 	ssize = batch->ssize;
 	pix = 0;
@@ -621,6 +818,7 @@ static void pSeries_lpar_flush_hash_range(unsigned long number, int local)
 		BUG_ON(rc != H_SUCCESS);
 	}
 
+out:
 	if (lock_tlbie)
 		spin_unlock_irqrestore(&pSeries_lpar_tlbie_lock, flags);
 }
-- 
2.7.4