From: Petr Tesarik <petr.tesarik.ext@huawei.com>
Dynamic allocation of bounce buffers may introduce regression for
some workloads. The expected outcomes are bigger worst-case I/O
latency reduced performance for some workloads. Unfortunately,
real-world testing has been too unstable to draw any conclusion.
To stay on the safe side, make the feature disabled by default and
let people turn it on with "swiotlb=dynamic" if needed. Since this
option can be combined with "force", the parser must be modified to
allow multiple options separated by commas.
A new bool field is added to struct io_tlb_mem to tell whether
dynamic allocations are allowed. This field is always false for
DMA restricted pools. It is also false for other software IO
TLBs unless "swiotlb=dynamic" was specified.
Signed-off-by: Petr Tesarik <petr.tesarik.ext@huawei.com>
---
.../admin-guide/kernel-parameters.txt | 6 +++++-
include/linux/swiotlb.h | 3 ++-
kernel/dma/swiotlb.c | 19 ++++++++++++++-----
3 files changed, 21 insertions(+), 7 deletions(-)
@@ -6081,14 +6081,18 @@
Execution Facility on pSeries.
swiotlb= [ARM,IA-64,PPC,MIPS,X86]
- Format: { <int> [,<int>] | force | noforce }
+ Format: { <int> [,<int>] [,option-list] | option-list }
<int> -- Number of I/O TLB slabs
<int> -- Second integer after comma. Number of swiotlb
areas with their own lock. Will be rounded up
to a power of 2.
+ <option-list> -- Comma-separated list of options.
+
+ Available options:
force -- force using of bounce buffers even if they
wouldn't be automatically used by the kernel
noforce -- Never use bounce buffers (for debugging)
+ dynamic -- allow dynamic allocation of bounce buffers
switches= [HW,M68k]
@@ -101,6 +101,7 @@ struct io_tlb_mem {
bool late_alloc;
bool force_bounce;
bool for_alloc;
+ bool allow_dyn;
unsigned int nareas;
unsigned int area_nslabs;
struct io_tlb_area *areas;
@@ -123,7 +124,7 @@ static inline bool is_swiotlb_buffer(struct device *dev, phys_addr_t paddr)
return mem &&
(is_swiotlb_fixed(mem, paddr) ||
- is_swiotlb_dyn(mem, paddr));
+ (mem->allow_dyn && is_swiotlb_dyn(mem, paddr)));
}
static inline bool is_swiotlb_force_bounce(struct device *dev)
@@ -78,6 +78,7 @@ struct io_tlb_dyn_slot {
static bool swiotlb_force_bounce;
static bool swiotlb_force_disable;
+static bool swiotlb_dynamic;
struct io_tlb_mem io_tlb_default_mem;
@@ -159,10 +160,17 @@ setup_io_tlb_npages(char *str)
swiotlb_adjust_nareas(simple_strtoul(str, &str, 0));
if (*str == ',')
++str;
- if (!strcmp(str, "force"))
- swiotlb_force_bounce = true;
- else if (!strcmp(str, "noforce"))
- swiotlb_force_disable = true;
+ while (str && *str) {
+ char *opt = strsep(&str, ",");
+ if (!strcmp(opt, "force"))
+ swiotlb_force_bounce = true;
+ else if (!strcmp(opt, "noforce"))
+ swiotlb_force_disable = true;
+ else if (!strcmp(opt, "dynamic"))
+ swiotlb_dynamic = true;
+ else
+ pr_warn("Invalid swiotlb option: %s", opt);
+ }
return 0;
}
@@ -287,6 +295,7 @@ static void swiotlb_init_io_tlb_mem(struct io_tlb_mem *mem, phys_addr_t start,
mem->area_nslabs = nslabs / mem->nareas;
mem->force_bounce = swiotlb_force_bounce || (flags & SWIOTLB_FORCE);
+ mem->allow_dyn = swiotlb_dynamic;
for (i = 0; i < mem->nareas; i++) {
spin_lock_init(&mem->areas[i].lock);
@@ -930,7 +939,7 @@ phys_addr_t swiotlb_tbl_map_single(struct device *dev, phys_addr_t orig_addr,
}
tlb_addr = (phys_addr_t)DMA_MAPPING_ERROR;
- if (!is_swiotlb_for_alloc(dev))
+ if (mem->allow_dyn)
tlb_addr = swiotlb_dyn_map(dev, orig_addr, alloc_size,
alloc_align_mask, dir, attrs);
if (tlb_addr == (phys_addr_t)DMA_MAPPING_ERROR)