On 2018/10/01 11:37, Sergey Senozhatsky wrote: > On (09/29/18 20:15), Tetsuo Handa wrote: >> >> Because there is no guarantee that memory information is dumped under the >> oom_lock mutex. The oom_lock is held when calling out_of_memory(), and it >> cannot be held when reporting GFP_ATOMIC memory allocation failures. > > IOW, static pr_line buffer needs additional synchronization for OOM. Correct?
Yes (assuming that your OOM refer to both out_of_memory() and warn_alloc()). And since warn_alloc() might be called from atomic/intrrupt contexts, we can't use locks for synchronization. > > If we are about to have a list of printk buffers then we probably can > define a list of NR_CPUS cont buffers. And we probably can reuse the > existing struct cont for buffered printk, having 2 different struct-s > for the same thing - struct cont and struct printk_buffer - is not very > cool. My plan is to remove "struct cont" after most of KERN_CONT users are converted to use buffered_printk(). There will be 2 different struct-s only during transition period. By the way, only up to two threads (the active printer thread and a thread which is marked as console_waiter) can stall inside printk(), doesn't it? Then, can you imagine a situation where 1024 (NR_CPUS) threads are stalling inside printk() waiting for flush? Such system is already dead. All callers but the two should release printk_buffer as soon as their printk() added their message to the log buffer. Maybe "struct printk_buffer" after all becomes identical to "struct cont". But I guess that even 16 printk_buffer-s is practically sufficient for 1024 CPUs system, and allocating NR_CPUS printk_buffer-s will be too wasteful. > >> But I don't want line buffered printk() API to truncate upon out of >> space for line buffered printk() API. > > All printk()-s are limited by LOG_LINE_MAX. Buffered printk() is not > special. I'm saying that I don't like discarding overflowed part because you are using seq_buf_vprintf() which just marks "overflowed" rather than "flush incomplete line" and "store the new data". DEFINE_PR_LINE(pr); pr_line(&pr, "1234567890123456789012345678901234567890123456789012345678901234567890"); pr_line(&pr, "1234567890abcde\n"); will discard "1234567890abcde\n" part, won't it? I think that getting 1234567890123456789012345678901234567890123456789012345678901234567890\n 1234567890abcde\n is better than getting 1234567890123456789012345678901234567890123456789012345678901234567890\n because we can still understand such output by prefixing caller information. Your DEFINE_PR_LINE() is limiting to far smaller than LOG_LINE_MAX. Since your version has to worry about "buffer full" (i.e. hitting seq_buf_set_overflow()) case, it might become a headache for API users.
