+/*
+ * addr64_t MapUserMemoryWindow(vm_map_t map, vm_offset_t va, size)
+ *
+ * map = the vm_map that we are mapping into the kernel
+ * va = start of the address range we are mapping
+ * Note that we do not test validty, we chose to trust our fellows...
+ *
+ * Maps a 512M slice of a user address space into a predefined kernel range
+ * on a per-thread basis. We map only the first 256M segment, allowing the
+ * second 256M segment to fault in as needed. This allows our clients to access
+ * an arbitrarily aligned operand up to 256M in size.
+ *
+ * In the future, the restriction of a predefined range may be loosened.
+ *
+ * Builds the proper linkage map to map the user range
+ * We will round this down to the previous segment boundary and calculate
+ * the relocation to the kernel slot
+ *
+ * We always make a segment table entry here if we need to. This is mainly because of
+ * copyin/out and if we don't, there will be multiple segment faults for
+ * each system call. I have seen upwards of 30000 per second.
+ *
+ * We do check, however, to see if the slice is already mapped and if so,
+ * we just exit. This is done for performance reasons. It was found that
+ * there was a considerable boost in copyin/out performance if we did not
+ * invalidate the segment at ReleaseUserAddressSpace time, so we dumped the
+ * restriction that you had to bracket MapUserMemoryWindow. Further, there
+ * is a yet further boost if you didn't need to map it each time. The theory
+ * behind this is that many times copies are to or from the same segment and
+ * done multiple times within the same system call. To take advantage of that,
+ * we check umwSpace and umwRelo to see if we've already got it.
+ *
+ * We also need to half-invalidate the slice when we context switch or go
+ * back to user state. A half-invalidate does not clear the actual mapping,
+ * but it does force the MapUserMemoryWindow function to reload the segment
+ * register/SLBE. If this is not done, we can end up some pretty severe
+ * performance penalties. If we map a slice, and the cached space/relocation is
+ * the same, we won't reload the segment registers. Howver, since we ran someone else,
+ * our SR is cleared and we will take a fault. This is reasonable if we block
+ * while copying (e.g., we took a page fault), but it is not reasonable when we
+ * just start. For this reason, we half-invalidate to make sure that the SR is
+ * explicitly reloaded.
+ *
+ * Note that we do not go to the trouble of making a pmap segment cache
+ * entry for these guys because they are very short term -- 99.99% of the time
+ * they will be unmapped before the next context switch.
+ *
+ */
+
+addr64_t MapUserMemoryWindow(
+ vm_map_t map,
+ addr64_t va) {
+
+ addr64_t baddrs, reladd;
+ thread_t thread;
+ mapping_t *mp;
projects / apple / xnu.git / blobdiff