Commit Graph

4 Commits

Author SHA1 Message Date
Atish Patra
98ee15ca3a lib: provide a platform specific tlb range flush threshold
Currently, the tlb range flush threshold is fixed and set to 4k for
all platforms. However, it should be platform specific as it completely
depends upon how platform actually implements sfence instruction.

Define a platform feature that allows every individual platform to set
different values. If a platform doesn't define it, just use a page size as
the threshold.

Signed-off-by: Atish Patra <atish.patra@wdc.com>
Reviewed-by: Bin Meng <bmeng.cn@gmail.com>
Reviewed-by: Anup Patel <anup.patel@wdc.com>
2019-09-10 17:44:10 +05:30
Atish Patra
f2e82c3d79 lib: Change tlb range flush threshold to 4k page instead of 1G
In RISC-V, tlb flush happens at a page granularity. That's why OpenSBI
also have a tlb range flush limit which decides the which tlb flush
requests should be upgraded to full flush to avoid long delays.
Currently, this is set to 1G which would result in a many sfence.vma
execution in a tight loop for a large range.

Change the threshold to 4k to speed things up.

Signed-off-by: Atish Patra <atish.patra@wdc.com>
Reviewed-by: Bin Meng <bmeng.cn@gmail.com>
Reviewed-by: Anup Patel <anup.patel@wdc.com>
2019-09-10 17:44:07 +05:30
Atish Patra
897a97a6af lib: Fix race conditions in tlb fifo access.
Linux kernel expects tlb flush SBI call to be completely synchronous i.e.
the SBI call should only return once corresponding *fence* instruction is
executed.

OpenSBI manages the outstanding TLB flush requests by keeping them in a
per hart based fifo. However, there are few corner cases that may lead to
race conditions while updating the fifo.

Currently, the caller hart waits for IPI acknowledgement via clint
address which is not a very good method as synchronization on MMIO may not
be supported in every platform. Moreover, the waiter doesn't have any way of
identifying if the IPI is received for specific tlb flush request or any
other IPI. This may lead to unpredictable behavior in supervisor/user space.

Fix this by waiting on individual fifo entries rather than MMIO address.
Currently, a relaxed loop is being used because wfi again involves MMIO write
which would be slower compared to relaxed loop. To avoid deadlock, fifo
is processed every time a hart loops for fifo enqueue or fifo sync to consume
the tlb flush requests sent by other harts.

Signed-off-by: Anup Patel <anup.patel@wdc.com>
Signed-off-by: Atish Patra <atish.patra@wdc.com>
2019-08-16 08:42:55 +05:30
Anup Patel
95b7480ab4 lib: Factor-out TLB management from IPI management
This patch factor-out TLB management from IPI management to separate
sources sbi_tlb.c and sbi_tlb.h.

Signed-off-by: Anup Patel <anup.patel@wdc.com>
Reviewed-by: Atish Patra <atish.patra@wdc.com>
2019-05-24 08:22:47 +05:30