ia64: Reduce stack usage by iterating over nodemask

GCC complains about sn2_global_tlb_purge() because of the large stack
required by the function,

arch/ia64/sn/kernel/sn2/sn2_smp.c: In function 'sn2_global_tlb_purge':
arch/ia64/sn/kernel/sn2/sn2_smp.c:319:1: warning: the frame size of 2176 bytes is larger than 2048 bytes [-Wframe-larger-than=]

2048 bytes of the stack are consumed by the node ID array 'nasids[]'.
But we don't actually need to put the ID array on the stack and can
use nodemask operations.

Cc: Tony Luck <[e-mail]>
Cc: Fenghua Yu <[e-mail]>
Cc: Bjorn Helgaas <[e-mail]>
Signed-off-by: Matt Fleming <[e-mail]>
Signed-off-by: Tony Luck <[e-mail]>

arch/ia64/sn/kernel/sn2/sn2_smp.c

···
 			       volatile unsigned long *, unsigned long,
 			       volatile unsigned long *, unsigned long);
 void
-sn2_ptc_deadlock_recovery(short *, short, short, int,
+sn2_ptc_deadlock_recovery(nodemask_t, short, short, int,
 			  volatile unsigned long *, unsigned long,
 			  volatile unsigned long *, unsigned long);
 
···
 	int use_cpu_ptcga;
 	volatile unsigned long *ptc0, *ptc1;
 	unsigned long itc, itc2, flags, data0 = 0, data1 = 0, rr_value, old_rr = 0;
-	short nasids[MAX_NUMNODES], nix;
+	short nix;
 	nodemask_t nodes_flushed;
 	int active, max_active, deadlock, flush_opt = sn2_flush_opt;
 
···
 	}
 
 	itc = ia64_get_itc();
-	nix = 0;
-	for_each_node_mask(cnode, nodes_flushed)
-		nasids[nix++] = cnodeid_to_nasid(cnode);
+	nix = nodes_weight(nodes_flushed);
 
 	rr_value = (mm->context << 3) | REGION_NUMBER(start);
 
···
 			data0 = (data0 & ~SH2_PTC_ADDR_MASK) | (start & SH2_PTC_ADDR_MASK);
 		deadlock = 0;
 		active = 0;
-		for (ibegin = 0, i = 0; i < nix; i++) {
-			nasid = nasids[i];
+		ibegin = 0;
+		i = 0;
+		for_each_node_mask(cnode, nodes_flushed) {
+			nasid = cnodeid_to_nasid(cnode);
 			if (use_cpu_ptcga && unlikely(nasid == mynasid)) {
 				ia64_ptcga(start, nbits << 2);
 				ia64_srlz_i();
···
 				if ((deadlock = wait_piowc())) {
 					if (flush_opt == 1)
 						goto done;
-					sn2_ptc_deadlock_recovery(nasids, ibegin, i, mynasid, ptc0, data0, ptc1, data1);
+					sn2_ptc_deadlock_recovery(nodes_flushed, ibegin, i, mynasid, ptc0, data0, ptc1, data1);
 					if (reset_max_active_on_deadlock())
 						max_active = 1;
 				}
 				active = 0;
 				ibegin = i + 1;
 			}
+			i++;
 		}
 		start += (1UL << nbits);
 	} while (start < end);
···
  */
 
 void
-sn2_ptc_deadlock_recovery(short *nasids, short ib, short ie, int mynasid,
+sn2_ptc_deadlock_recovery(nodemask_t nodes, short ib, short ie, int mynasid,
 			  volatile unsigned long *ptc0, unsigned long data0,
 			  volatile unsigned long *ptc1, unsigned long data1)
 {
 	short nasid, i;
+	int cnode;
 	unsigned long *piows, zeroval, n;
 
 	__this_cpu_inc(ptcstats.deadlocks);
···
 	piows = (unsigned long *) pda->pio_write_status_addr;
 	zeroval = pda->pio_write_status_val;
 
+	i = 0;
+	for_each_node_mask(cnode, nodes) {
+		if (i < ib)
+			goto next;
 
-	for (i=ib; i <= ie; i++) {
-		nasid = nasids[i];
+		if (i > ie)
+			break;
+
+		nasid = cnodeid_to_nasid(cnode);
 		if (local_node_uses_ptc_ga(is_shub1()) && nasid == mynasid)
-			continue;
+			goto next;
+
 		ptc0 = CHANGE_NASID(nasid, ptc0);
 		if (ptc1)
 			ptc1 = CHANGE_NASID(nasid, ptc1);
 
 		n = sn2_ptc_deadlock_recovery_core(ptc0, data0, ptc1, data1, piows, zeroval);
 		__this_cpu_add(ptcstats.deadlocks2, n);
+next:
+		i++;
 	}
 
 }