xlat v2: Flush xlat tables after being modified
During cold boot, the initial translation tables are created with data caches disabled, so all modifications go to memory directly. After the MMU is enabled and data cache is enabled, any modification to the tables goes to data cache, and eventually may get flushed to memory. If CPU0 modifies the tables while CPU1 is off, CPU0 will have the modified tables in its data cache. When CPU1 is powered on, the MMU is enabled, then it enables coherency, and then it enables the data cache. Until this is done, CPU1 isn't in coherency, and the translation tables it sees can be outdated if CPU0 still has some modified entries in its data cache. This can be a problem in some cases. For example, the warm boot code uses only the tables mapped during cold boot, which don't normally change. However, if they are modified (and a RO page is made RW, or a XN page is made executable) the CPU will see the old attributes and crash when it tries to access it. This doesn't happen in systems with HW_ASSISTED_COHERENCY or WARMBOOT_ENABLE_DCACHE_EARLY. In these systems, the data cache is enabled at the same time as the MMU. As soon as this happens, the CPU is in coherency. There was an attempt of a fix in psci_helpers.S, but it didn't solve the problem. That code has been deleted. The code was introduced in commit <26441030 > ("Invalidate TLB entries during warm boot"). Now, during a map or unmap operation, the memory associated to each modified table is flushed. Traversing a table will also flush it's memory, as there is no way to tell in the current implementation if the table that has been traversed has also been modified. Change-Id: I4b520bca27502f1018878061bc5fb82af740bb92 Signed-off-by: Antonio Nino Diaz <antonio.ninodiaz@arm.com>
Please register or sign in to comment