gru: preload tlb for bcopy instructions
Add anticipatory TLB dropins for GRU TLB misses that occur on BCOPY
instructions that copy large amounts of data.
Signed-off-by: Jack Steiner <steiner@sgi.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
diff --git a/drivers/misc/sgi-gru/grufault.c b/drivers/misc/sgi-gru/grufault.c
index 7d757e9..a1b3a1d 100644
--- a/drivers/misc/sgi-gru/grufault.c
+++ b/drivers/misc/sgi-gru/grufault.c
@@ -290,6 +290,61 @@
/*
+ * Flush a CBE from cache. The CBE is clean in the cache. Dirty the
+ * CBE cacheline so that the line will be written back to home agent.
+ * Otherwise the line may be silently dropped. This has no impact
+ * except on performance.
+ */
+static void gru_flush_cache_cbe(struct gru_control_block_extended *cbe)
+{
+ if (unlikely(cbe)) {
+ cbe->cbrexecstatus = 0; /* make CL dirty */
+ gru_flush_cache(cbe);
+ }
+}
+
+/*
+ * Preload the TLB with entries that may be required. Currently, preloading
+ * is implemented only for BCOPY. Preload <tlb_preload_count> pages OR to
+ * the end of the bcopy tranfer, whichever is smaller.
+ */
+static void gru_preload_tlb(struct gru_state *gru,
+ struct gru_thread_state *gts, int atomic,
+ unsigned long fault_vaddr, int asid, int write,
+ unsigned char tlb_preload_count,
+ struct gru_tlb_fault_handle *tfh,
+ struct gru_control_block_extended *cbe)
+{
+ unsigned long vaddr = 0, gpa;
+ int ret, pageshift;
+
+ if (cbe->opccpy != OP_BCOPY)
+ return;
+
+ if (fault_vaddr == cbe->cbe_baddr0)
+ vaddr = fault_vaddr + GRU_CACHE_LINE_BYTES * cbe->cbe_src_cl - 1;
+ else if (fault_vaddr == cbe->cbe_baddr1)
+ vaddr = fault_vaddr + (1 << cbe->xtypecpy) * cbe->cbe_nelemcur - 1;
+
+ fault_vaddr &= PAGE_MASK;
+ vaddr &= PAGE_MASK;
+ vaddr = min(vaddr, fault_vaddr + tlb_preload_count * PAGE_SIZE);
+
+ while (vaddr > fault_vaddr) {
+ ret = gru_vtop(gts, vaddr, write, atomic, &gpa, &pageshift);
+ if (ret || tfh_write_only(tfh, gpa, GAA_RAM, vaddr, asid, write,
+ GRU_PAGESIZE(pageshift)))
+ return;
+ gru_dbg(grudev,
+ "%s: gid %d, gts 0x%p, tfh 0x%p, vaddr 0x%lx, asid 0x%x, rw %d, ps %d, gpa 0x%lx\n",
+ atomic ? "atomic" : "non-atomic", gru->gs_gid, gts, tfh,
+ vaddr, asid, write, pageshift, gpa);
+ vaddr -= PAGE_SIZE;
+ STAT(tlb_preload_page);
+ }
+}
+
+/*
* Drop a TLB entry into the GRU. The fault is described by info in an TFH.
* Input:
* cb Address of user CBR. Null if not running in user context
@@ -303,6 +358,8 @@
struct gru_tlb_fault_handle *tfh,
struct gru_instruction_bits *cbk)
{
+ struct gru_control_block_extended *cbe = NULL;
+ unsigned char tlb_preload_count = gts->ts_tlb_preload_count;
int pageshift = 0, asid, write, ret, atomic = !cbk, indexway;
unsigned long gpa = 0, vaddr = 0;
@@ -314,6 +371,14 @@
*/
/*
+ * Prefetch the CBE if doing TLB preloading
+ */
+ if (unlikely(tlb_preload_count)) {
+ cbe = gru_tfh_to_cbe(tfh);
+ prefetchw(cbe);
+ }
+
+ /*
* Error if TFH state is IDLE or FMM mode & the user issuing a UPM call.
* Might be a hardware race OR a stupid user. Ignore FMM because FMM
* is a transient state.
@@ -359,6 +424,12 @@
goto failupm;
}
}
+
+ if (unlikely(cbe) && pageshift == PAGE_SHIFT) {
+ gru_preload_tlb(gts->ts_gru, gts, atomic, vaddr, asid, write, tlb_preload_count, tfh, cbe);
+ gru_flush_cache_cbe(cbe);
+ }
+
gru_cb_set_istatus_active(cbk);
tfh_write_restart(tfh, gpa, GAA_RAM, vaddr, asid, write,
GRU_PAGESIZE(pageshift));
@@ -378,11 +449,13 @@
tfh_user_polling_mode(tfh);
else
gru_flush_cache(tfh);
+ gru_flush_cache_cbe(cbe);
return -EAGAIN;
failupm:
/* Atomic failure switch CBR to UPM */
tfh_user_polling_mode(tfh);
+ gru_flush_cache_cbe(cbe);
STAT(tlb_dropin_fail_upm);
gru_dbg(grudev, "FAILED upm tfh: 0x%p, vaddr 0x%lx\n", tfh, vaddr);
return 1;
@@ -390,6 +463,7 @@
failfmm:
/* FMM state on UPM call */
gru_flush_cache(tfh);
+ gru_flush_cache_cbe(cbe);
STAT(tlb_dropin_fail_fmm);
gru_dbg(grudev, "FAILED fmm tfh: 0x%p, state %d\n", tfh, tfh->state);
return 0;
@@ -397,6 +471,7 @@
failnoexception:
/* TFH status did not show exception pending */
gru_flush_cache(tfh);
+ gru_flush_cache_cbe(cbe);
if (cbk)
gru_flush_cache(cbk);
STAT(tlb_dropin_fail_no_exception);
@@ -407,6 +482,7 @@
failidle:
/* TFH state was idle - no miss pending */
gru_flush_cache(tfh);
+ gru_flush_cache_cbe(cbe);
if (cbk)
gru_flush_cache(cbk);
STAT(tlb_dropin_fail_idle);
@@ -416,6 +492,7 @@
failinval:
/* All errors (atomic & non-atomic) switch CBR to EXCEPTION state */
tfh_exception(tfh);
+ gru_flush_cache_cbe(cbe);
STAT(tlb_dropin_fail_invalid);
gru_dbg(grudev, "FAILED inval tfh: 0x%p, vaddr 0x%lx\n", tfh, vaddr);
return -EFAULT;
@@ -426,6 +503,7 @@
tfh_user_polling_mode(tfh);
else
gru_flush_cache(tfh);
+ gru_flush_cache_cbe(cbe);
STAT(tlb_dropin_fail_range_active);
gru_dbg(grudev, "FAILED range active: tfh 0x%p, vaddr 0x%lx\n",
tfh, vaddr);
@@ -627,7 +705,7 @@
excdet.exceptdet1 = cbe->idef3upd;
excdet.cbrstate = cbe->cbrstate;
excdet.cbrexecstatus = cbe->cbrexecstatus;
- gru_flush_cache(cbe);
+ gru_flush_cache_cbe(cbe);
ret = 0;
} else {
ret = -EAGAIN;
@@ -770,9 +848,12 @@
return -EFAULT;
gru_dbg(grudev, "op %d, gseg 0x%lx, value1 0x%lx\n", req.op, req.gseg, req.val1);
- gts = gru_alloc_locked_gts(req.gseg);
- if (IS_ERR(gts))
- return PTR_ERR(gts);
+ gts = gru_find_lock_gts(req.gseg);
+ if (!gts) {
+ gts = gru_alloc_locked_gts(req.gseg);
+ if (IS_ERR(gts))
+ return PTR_ERR(gts);
+ }
switch (req.op) {
case sco_blade_chiplet: