@@ -350,14 +350,14 @@ static inline void arch_tlbbatch_flush(struct arch_tlbflush_unmap_batch *batch)
* entries one by one at the granularity of 'stride'. If the TLB
* range ops are supported, then:
*
- * 1. If 'pages' is odd, flush the first page through non-range
- * operations;
+ * 1. The minimum range granularity is decided by 'scale', so multiple range
+ * TLBI operations may be required. Start from scale = 3, flush the largest
+ * possible number of pages ((num+1)*2^(5*scale+1)) that fit into the
+ * requested range, then decrement scale and continue until one or zero pages
+ * are left.
*
- * 2. For remaining pages: the minimum range granularity is decided
- * by 'scale', so multiple range TLBI operations may be required.
- * Start from scale = 0, flush the corresponding number of pages
- * ((num+1)*2^(5*scale+1) starting from 'addr'), then increase it
- * until no pages left.
+ * 2. If there is 1 page remaining, flush it through non-range operations. Range
+ * operations can only span an even number of pages.
*
* Note that certain ranges can be represented by either num = 31 and
* scale or num = 0 and scale + 1. The loop below favours the latter
@@ -367,12 +367,12 @@ static inline void arch_tlbbatch_flush(struct arch_tlbflush_unmap_batch *batch)
asid, tlb_level, tlbi_user) \
do { \
int num = 0; \
- int scale = 0; \
+ int scale = 3; \
unsigned long addr; \
\
while (pages > 0) { \
if (!system_supports_tlb_range() || \
- pages % 2 == 1) { \
+ pages == 1) { \
addr = __TLBI_VADDR(start, asid); \
__tlbi_level(op, addr, tlb_level); \
if (tlbi_user) \
@@ -392,7 +392,7 @@ do { \
start += __TLBI_RANGE_PAGES(num, scale) << PAGE_SHIFT; \
pages -= __TLBI_RANGE_PAGES(num, scale); \
} \
- scale++; \
+ scale--; \
} \
} while (0)
In preparation for adding support for LPA2 to the tlb invalidation routines, modify the algorithm used by range-based tlbi to start at the highest 'scale' and decrement instead of starting at the lowest 'scale' and incrementing. This new approach makes it possible to maintain 64K alignment as we work through the range, until the last op (at scale=0). This is required when LPA2 is enabled. (This part will be added in a subsequent commit). This change is separated into its own patch because it will also impact non-LPA2 systems, and I want to make it easy to bisect in case it leads to performance regression (see below for benchmarks that suggest this should not be a problem). The original commit (d1d3aa98 "arm64: tlb: Use the TLBI RANGE feature in arm64") stated this as the reason for _incrementing_ scale: However, in most scenarios, the pages = 1 when flush_tlb_range() is called. Start from scale = 3 or other proper value (such as scale =ilog2(pages)), will incur extra overhead. So increase 'scale' from 0 to maximum. But pages=1 is already special cased by the non-range invalidation path, which will take care of it the first time through the loop (both in the original commit and in my change), so I don't think switching to decrement scale should have any extra performance impact after all. Indeed benchmarking kernel compilation, a TLBI-heavy workload, suggests that this new approach actually _improves_ performance slightly (using a virtual machine on Apple M2): Table shows time to execute kernel compilation workload with 8 jobs, relative to baseline without this patch (more negative number is bigger speedup). Repeated 9 times across 3 system reboots: | counter | mean | stdev | |:----------|-----------:|----------:| | real-time | -0.6% | 0.0% | | kern-time | -1.6% | 0.5% | | user-time | -0.4% | 0.1% | Signed-off-by: Ryan Roberts <ryan.roberts@arm.com> --- arch/arm64/include/asm/tlbflush.h | 20 ++++++++++---------- 1 file changed, 10 insertions(+), 10 deletions(-)